aboutsummaryrefslogtreecommitdiffstats
path: root/src/compiler
diff options
context:
space:
mode:
Diffstat (limited to 'src/compiler')
-rw-r--r--src/compiler/Makefile.nir.am1
-rw-r--r--src/compiler/Makefile.sources9
-rw-r--r--src/compiler/nir/nir_intrinsics.h3
-rw-r--r--src/compiler/spirv/GLSL.std.450.h127
-rw-r--r--src/compiler/spirv/nir_spirv.h54
-rw-r--r--src/compiler/spirv/spirv.h870
-rw-r--r--src/compiler/spirv/spirv_to_nir.c2710
-rw-r--r--src/compiler/spirv/vtn_alu.c464
-rw-r--r--src/compiler/spirv/vtn_cfg.c778
-rw-r--r--src/compiler/spirv/vtn_glsl450.c666
-rw-r--r--src/compiler/spirv/vtn_private.h484
-rw-r--r--src/compiler/spirv/vtn_variables.c1415
12 files changed, 7581 insertions, 0 deletions
diff --git a/src/compiler/Makefile.nir.am b/src/compiler/Makefile.nir.am
index e1acf8a29a2..45a9c809bf8 100644
--- a/src/compiler/Makefile.nir.am
+++ b/src/compiler/Makefile.nir.am
@@ -33,6 +33,7 @@ nir_libnir_la_LIBADD = \
nir_libnir_la_SOURCES = \
$(NIR_FILES) \
+ $(SPIRV_FILES) \
$(NIR_GENERATED_FILES)
PYTHON_GEN = $(AM_V_GEN)$(PYTHON2) $(PYTHON_FLAGS)
diff --git a/src/compiler/Makefile.sources b/src/compiler/Makefile.sources
index 6f09abf3d5c..19735339bca 100644
--- a/src/compiler/Makefile.sources
+++ b/src/compiler/Makefile.sources
@@ -233,3 +233,12 @@ NIR_FILES = \
nir/nir_vla.h \
nir/nir_worklist.c \
nir/nir_worklist.h
+
+SPIRV_FILES = \
+ spirv/nir_spirv.h \
+ spirv/spirv_to_nir.c \
+ spirv/vtn_alu.c \
+ spirv/vtn_cfg.c \
+ spirv/vtn_glsl450.c \
+ spirv/vtn_private.h \
+ spirv/vtn_variables.c
diff --git a/src/compiler/nir/nir_intrinsics.h b/src/compiler/nir/nir_intrinsics.h
index 66a169c019b..bd00fbbe6cc 100644
--- a/src/compiler/nir/nir_intrinsics.h
+++ b/src/compiler/nir/nir_intrinsics.h
@@ -345,6 +345,9 @@ LOAD(output, 1, 1, BASE, xx, xx, NIR_INTRINSIC_CAN_ELIMINATE)
LOAD(per_vertex_output, 2, 1, BASE, xx, xx, NIR_INTRINSIC_CAN_ELIMINATE)
/* src[] = { offset }. const_index[] = { base } */
LOAD(shared, 1, 1, BASE, xx, xx, NIR_INTRINSIC_CAN_ELIMINATE)
+/* src[] = { offset }. const_index[] = { base, range } */
+LOAD(push_constant, 1, 2, BASE, RANGE, xx,
+ NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER)
/*
* Stores work the same way as loads, except now the first source is the value
diff --git a/src/compiler/spirv/GLSL.std.450.h b/src/compiler/spirv/GLSL.std.450.h
new file mode 100644
index 00000000000..d1c9b5c1d44
--- /dev/null
+++ b/src/compiler/spirv/GLSL.std.450.h
@@ -0,0 +1,127 @@
+/*
+** Copyright (c) 2014-2015 The Khronos Group Inc.
+**
+** Permission is hereby granted, free of charge, to any person obtaining a copy
+** of this software and/or associated documentation files (the "Materials"),
+** to deal in the Materials without restriction, including without limitation
+** the rights to use, copy, modify, merge, publish, distribute, sublicense,
+** and/or sell copies of the Materials, and to permit persons to whom the
+** Materials are furnished to do so, subject to the following conditions:
+**
+** The above copyright notice and this permission notice shall be included in
+** all copies or substantial portions of the Materials.
+**
+** MODIFICATIONS TO THIS FILE MAY MEAN IT NO LONGER ACCURATELY REFLECTS KHRONOS
+** STANDARDS. THE UNMODIFIED, NORMATIVE VERSIONS OF KHRONOS SPECIFICATIONS AND
+** HEADER INFORMATION ARE LOCATED AT https://www.khronos.org/registry/
+**
+** THE MATERIALS ARE PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
+** OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+** FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
+** THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+** LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+** FROM,OUT OF OR IN CONNECTION WITH THE MATERIALS OR THE USE OR OTHER DEALINGS
+** IN THE MATERIALS.
+*/
+
+#ifndef GLSLstd450_H
+#define GLSLstd450_H
+
+const int GLSLstd450Version = 99;
+const int GLSLstd450Revision = 3;
+
+enum GLSLstd450 {
+ GLSLstd450Bad = 0, // Don't use
+
+ GLSLstd450Round = 1,
+ GLSLstd450RoundEven = 2,
+ GLSLstd450Trunc = 3,
+ GLSLstd450FAbs = 4,
+ GLSLstd450SAbs = 5,
+ GLSLstd450FSign = 6,
+ GLSLstd450SSign = 7,
+ GLSLstd450Floor = 8,
+ GLSLstd450Ceil = 9,
+ GLSLstd450Fract = 10,
+
+ GLSLstd450Radians = 11,
+ GLSLstd450Degrees = 12,
+ GLSLstd450Sin = 13,
+ GLSLstd450Cos = 14,
+ GLSLstd450Tan = 15,
+ GLSLstd450Asin = 16,
+ GLSLstd450Acos = 17,
+ GLSLstd450Atan = 18,
+ GLSLstd450Sinh = 19,
+ GLSLstd450Cosh = 20,
+ GLSLstd450Tanh = 21,
+ GLSLstd450Asinh = 22,
+ GLSLstd450Acosh = 23,
+ GLSLstd450Atanh = 24,
+ GLSLstd450Atan2 = 25,
+
+ GLSLstd450Pow = 26,
+ GLSLstd450Exp = 27,
+ GLSLstd450Log = 28,
+ GLSLstd450Exp2 = 29,
+ GLSLstd450Log2 = 30,
+ GLSLstd450Sqrt = 31,
+ GLSLstd450InverseSqrt = 32,
+
+ GLSLstd450Determinant = 33,
+ GLSLstd450MatrixInverse = 34,
+
+ GLSLstd450Modf = 35, // second operand needs an OpVariable to write to
+ GLSLstd450ModfStruct = 36, // no OpVariable operand
+ GLSLstd450FMin = 37,
+ GLSLstd450UMin = 38,
+ GLSLstd450SMin = 39,
+ GLSLstd450FMax = 40,
+ GLSLstd450UMax = 41,
+ GLSLstd450SMax = 42,
+ GLSLstd450FClamp = 43,
+ GLSLstd450UClamp = 44,
+ GLSLstd450SClamp = 45,
+ GLSLstd450FMix = 46,
+ GLSLstd450IMix = 47,
+ GLSLstd450Step = 48,
+ GLSLstd450SmoothStep = 49,
+
+ GLSLstd450Fma = 50,
+ GLSLstd450Frexp = 51, // second operand needs an OpVariable to write to
+ GLSLstd450FrexpStruct = 52, // no OpVariable operand
+ GLSLstd450Ldexp = 53,
+
+ GLSLstd450PackSnorm4x8 = 54,
+ GLSLstd450PackUnorm4x8 = 55,
+ GLSLstd450PackSnorm2x16 = 56,
+ GLSLstd450PackUnorm2x16 = 57,
+ GLSLstd450PackHalf2x16 = 58,
+ GLSLstd450PackDouble2x32 = 59,
+ GLSLstd450UnpackSnorm2x16 = 60,
+ GLSLstd450UnpackUnorm2x16 = 61,
+ GLSLstd450UnpackHalf2x16 = 62,
+ GLSLstd450UnpackSnorm4x8 = 63,
+ GLSLstd450UnpackUnorm4x8 = 64,
+ GLSLstd450UnpackDouble2x32 = 65,
+
+ GLSLstd450Length = 66,
+ GLSLstd450Distance = 67,
+ GLSLstd450Cross = 68,
+ GLSLstd450Normalize = 69,
+ GLSLstd450FaceForward = 70,
+ GLSLstd450Reflect = 71,
+ GLSLstd450Refract = 72,
+
+ GLSLstd450FindILsb = 73,
+ GLSLstd450FindSMsb = 74,
+ GLSLstd450FindUMsb = 75,
+
+ GLSLstd450InterpolateAtCentroid = 76,
+ GLSLstd450InterpolateAtSample = 77,
+ GLSLstd450InterpolateAtOffset = 78,
+
+ GLSLstd450Count
+};
+
+#endif // #ifndef GLSLstd450_H
diff --git a/src/compiler/spirv/nir_spirv.h b/src/compiler/spirv/nir_spirv.h
new file mode 100644
index 00000000000..500f2cb94df
--- /dev/null
+++ b/src/compiler/spirv/nir_spirv.h
@@ -0,0 +1,54 @@
+/*
+ * Copyright © 2015 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ *
+ * Authors:
+ * Jason Ekstrand ([email protected])
+ *
+ */
+
+#pragma once
+
+#ifndef _NIR_SPIRV_H_
+#define _NIR_SPIRV_H_
+
+#include "nir/nir.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct nir_spirv_specialization {
+ uint32_t id;
+ uint32_t data;
+};
+
+nir_function *spirv_to_nir(const uint32_t *words, size_t word_count,
+ struct nir_spirv_specialization *specializations,
+ unsigned num_specializations,
+ gl_shader_stage stage, const char *entry_point_name,
+ const nir_shader_compiler_options *options);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /* _NIR_SPIRV_H_ */
diff --git a/src/compiler/spirv/spirv.h b/src/compiler/spirv/spirv.h
new file mode 100644
index 00000000000..63bcb2f88dd
--- /dev/null
+++ b/src/compiler/spirv/spirv.h
@@ -0,0 +1,870 @@
+/*
+** Copyright (c) 2014-2015 The Khronos Group Inc.
+**
+** Permission is hereby granted, free of charge, to any person obtaining a copy
+** of this software and/or associated documentation files (the "Materials"),
+** to deal in the Materials without restriction, including without limitation
+** the rights to use, copy, modify, merge, publish, distribute, sublicense,
+** and/or sell copies of the Materials, and to permit persons to whom the
+** Materials are furnished to do so, subject to the following conditions:
+**
+** The above copyright notice and this permission notice shall be included in
+** all copies or substantial portions of the Materials.
+**
+** MODIFICATIONS TO THIS FILE MAY MEAN IT NO LONGER ACCURATELY REFLECTS KHRONOS
+** STANDARDS. THE UNMODIFIED, NORMATIVE VERSIONS OF KHRONOS SPECIFICATIONS AND
+** HEADER INFORMATION ARE LOCATED AT https://www.khronos.org/registry/
+**
+** THE MATERIALS ARE PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
+** OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+** FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
+** THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+** LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+** FROM,OUT OF OR IN CONNECTION WITH THE MATERIALS OR THE USE OR OTHER DEALINGS
+** IN THE MATERIALS.
+*/
+
+/*
+** This header is automatically generated by the same tool that creates
+** the Binary Section of the SPIR-V specification.
+*/
+
+/*
+** Enumeration tokens for SPIR-V, in various styles:
+** C, C++, C++11, JSON, Lua, Python
+**
+** - C will have tokens with a "Spv" prefix, e.g.: SpvSourceLanguageGLSL
+** - C++ will have tokens in the "spv" name space, e.g.: spv::SourceLanguageGLSL
+** - C++11 will use enum classes in the spv namespace, e.g.: spv::SourceLanguage::GLSL
+** - Lua will use tables, e.g.: spv.SourceLanguage.GLSL
+** - Python will use dictionaries, e.g.: spv['SourceLanguage']['GLSL']
+**
+** Some tokens act like mask values, which can be OR'd together,
+** while others are mutually exclusive. The mask-like ones have
+** "Mask" in their name, and a parallel enum that has the shift
+** amount (1 << x) for each corresponding enumerant.
+*/
+
+#ifndef spirv_H
+#define spirv_H
+
+typedef unsigned int SpvId;
+
+#define SPV_VERSION 0x10000
+#define SPV_REVISION 2
+
+static const unsigned int SpvMagicNumber = 0x07230203;
+static const unsigned int SpvVersion = 0x00010000;
+static const unsigned int SpvRevision = 2;
+static const unsigned int SpvOpCodeMask = 0xffff;
+static const unsigned int SpvWordCountShift = 16;
+
+typedef enum SpvSourceLanguage_ {
+ SpvSourceLanguageUnknown = 0,
+ SpvSourceLanguageESSL = 1,
+ SpvSourceLanguageGLSL = 2,
+ SpvSourceLanguageOpenCL_C = 3,
+ SpvSourceLanguageOpenCL_CPP = 4,
+} SpvSourceLanguage;
+
+typedef enum SpvExecutionModel_ {
+ SpvExecutionModelVertex = 0,
+ SpvExecutionModelTessellationControl = 1,
+ SpvExecutionModelTessellationEvaluation = 2,
+ SpvExecutionModelGeometry = 3,
+ SpvExecutionModelFragment = 4,
+ SpvExecutionModelGLCompute = 5,
+ SpvExecutionModelKernel = 6,
+} SpvExecutionModel;
+
+typedef enum SpvAddressingModel_ {
+ SpvAddressingModelLogical = 0,
+ SpvAddressingModelPhysical32 = 1,
+ SpvAddressingModelPhysical64 = 2,
+} SpvAddressingModel;
+
+typedef enum SpvMemoryModel_ {
+ SpvMemoryModelSimple = 0,
+ SpvMemoryModelGLSL450 = 1,
+ SpvMemoryModelOpenCL = 2,
+} SpvMemoryModel;
+
+typedef enum SpvExecutionMode_ {
+ SpvExecutionModeInvocations = 0,
+ SpvExecutionModeSpacingEqual = 1,
+ SpvExecutionModeSpacingFractionalEven = 2,
+ SpvExecutionModeSpacingFractionalOdd = 3,
+ SpvExecutionModeVertexOrderCw = 4,
+ SpvExecutionModeVertexOrderCcw = 5,
+ SpvExecutionModePixelCenterInteger = 6,
+ SpvExecutionModeOriginUpperLeft = 7,
+ SpvExecutionModeOriginLowerLeft = 8,
+ SpvExecutionModeEarlyFragmentTests = 9,
+ SpvExecutionModePointMode = 10,
+ SpvExecutionModeXfb = 11,
+ SpvExecutionModeDepthReplacing = 12,
+ SpvExecutionModeDepthGreater = 14,
+ SpvExecutionModeDepthLess = 15,
+ SpvExecutionModeDepthUnchanged = 16,
+ SpvExecutionModeLocalSize = 17,
+ SpvExecutionModeLocalSizeHint = 18,
+ SpvExecutionModeInputPoints = 19,
+ SpvExecutionModeInputLines = 20,
+ SpvExecutionModeInputLinesAdjacency = 21,
+ SpvExecutionModeTriangles = 22,
+ SpvExecutionModeInputTrianglesAdjacency = 23,
+ SpvExecutionModeQuads = 24,
+ SpvExecutionModeIsolines = 25,
+ SpvExecutionModeOutputVertices = 26,
+ SpvExecutionModeOutputPoints = 27,
+ SpvExecutionModeOutputLineStrip = 28,
+ SpvExecutionModeOutputTriangleStrip = 29,
+ SpvExecutionModeVecTypeHint = 30,
+ SpvExecutionModeContractionOff = 31,
+} SpvExecutionMode;
+
+typedef enum SpvStorageClass_ {
+ SpvStorageClassUniformConstant = 0,
+ SpvStorageClassInput = 1,
+ SpvStorageClassUniform = 2,
+ SpvStorageClassOutput = 3,
+ SpvStorageClassWorkgroup = 4,
+ SpvStorageClassCrossWorkgroup = 5,
+ SpvStorageClassPrivate = 6,
+ SpvStorageClassFunction = 7,
+ SpvStorageClassGeneric = 8,
+ SpvStorageClassPushConstant = 9,
+ SpvStorageClassAtomicCounter = 10,
+ SpvStorageClassImage = 11,
+} SpvStorageClass;
+
+typedef enum SpvDim_ {
+ SpvDim1D = 0,
+ SpvDim2D = 1,
+ SpvDim3D = 2,
+ SpvDimCube = 3,
+ SpvDimRect = 4,
+ SpvDimBuffer = 5,
+ SpvDimSubpassData = 6,
+} SpvDim;
+
+typedef enum SpvSamplerAddressingMode_ {
+ SpvSamplerAddressingModeNone = 0,
+ SpvSamplerAddressingModeClampToEdge = 1,
+ SpvSamplerAddressingModeClamp = 2,
+ SpvSamplerAddressingModeRepeat = 3,
+ SpvSamplerAddressingModeRepeatMirrored = 4,
+} SpvSamplerAddressingMode;
+
+typedef enum SpvSamplerFilterMode_ {
+ SpvSamplerFilterModeNearest = 0,
+ SpvSamplerFilterModeLinear = 1,
+} SpvSamplerFilterMode;
+
+typedef enum SpvImageFormat_ {
+ SpvImageFormatUnknown = 0,
+ SpvImageFormatRgba32f = 1,
+ SpvImageFormatRgba16f = 2,
+ SpvImageFormatR32f = 3,
+ SpvImageFormatRgba8 = 4,
+ SpvImageFormatRgba8Snorm = 5,
+ SpvImageFormatRg32f = 6,
+ SpvImageFormatRg16f = 7,
+ SpvImageFormatR11fG11fB10f = 8,
+ SpvImageFormatR16f = 9,
+ SpvImageFormatRgba16 = 10,
+ SpvImageFormatRgb10A2 = 11,
+ SpvImageFormatRg16 = 12,
+ SpvImageFormatRg8 = 13,
+ SpvImageFormatR16 = 14,
+ SpvImageFormatR8 = 15,
+ SpvImageFormatRgba16Snorm = 16,
+ SpvImageFormatRg16Snorm = 17,
+ SpvImageFormatRg8Snorm = 18,
+ SpvImageFormatR16Snorm = 19,
+ SpvImageFormatR8Snorm = 20,
+ SpvImageFormatRgba32i = 21,
+ SpvImageFormatRgba16i = 22,
+ SpvImageFormatRgba8i = 23,
+ SpvImageFormatR32i = 24,
+ SpvImageFormatRg32i = 25,
+ SpvImageFormatRg16i = 26,
+ SpvImageFormatRg8i = 27,
+ SpvImageFormatR16i = 28,
+ SpvImageFormatR8i = 29,
+ SpvImageFormatRgba32ui = 30,
+ SpvImageFormatRgba16ui = 31,
+ SpvImageFormatRgba8ui = 32,
+ SpvImageFormatR32ui = 33,
+ SpvImageFormatRgb10a2ui = 34,
+ SpvImageFormatRg32ui = 35,
+ SpvImageFormatRg16ui = 36,
+ SpvImageFormatRg8ui = 37,
+ SpvImageFormatR16ui = 38,
+ SpvImageFormatR8ui = 39,
+} SpvImageFormat;
+
+typedef enum SpvImageChannelOrder_ {
+ SpvImageChannelOrderR = 0,
+ SpvImageChannelOrderA = 1,
+ SpvImageChannelOrderRG = 2,
+ SpvImageChannelOrderRA = 3,
+ SpvImageChannelOrderRGB = 4,
+ SpvImageChannelOrderRGBA = 5,
+ SpvImageChannelOrderBGRA = 6,
+ SpvImageChannelOrderARGB = 7,
+ SpvImageChannelOrderIntensity = 8,
+ SpvImageChannelOrderLuminance = 9,
+ SpvImageChannelOrderRx = 10,
+ SpvImageChannelOrderRGx = 11,
+ SpvImageChannelOrderRGBx = 12,
+ SpvImageChannelOrderDepth = 13,
+ SpvImageChannelOrderDepthStencil = 14,
+ SpvImageChannelOrdersRGB = 15,
+ SpvImageChannelOrdersRGBx = 16,
+ SpvImageChannelOrdersRGBA = 17,
+ SpvImageChannelOrdersBGRA = 18,
+} SpvImageChannelOrder;
+
+typedef enum SpvImageChannelDataType_ {
+ SpvImageChannelDataTypeSnormInt8 = 0,
+ SpvImageChannelDataTypeSnormInt16 = 1,
+ SpvImageChannelDataTypeUnormInt8 = 2,
+ SpvImageChannelDataTypeUnormInt16 = 3,
+ SpvImageChannelDataTypeUnormShort565 = 4,
+ SpvImageChannelDataTypeUnormShort555 = 5,
+ SpvImageChannelDataTypeUnormInt101010 = 6,
+ SpvImageChannelDataTypeSignedInt8 = 7,
+ SpvImageChannelDataTypeSignedInt16 = 8,
+ SpvImageChannelDataTypeSignedInt32 = 9,
+ SpvImageChannelDataTypeUnsignedInt8 = 10,
+ SpvImageChannelDataTypeUnsignedInt16 = 11,
+ SpvImageChannelDataTypeUnsignedInt32 = 12,
+ SpvImageChannelDataTypeHalfFloat = 13,
+ SpvImageChannelDataTypeFloat = 14,
+ SpvImageChannelDataTypeUnormInt24 = 15,
+ SpvImageChannelDataTypeUnormInt101010_2 = 16,
+} SpvImageChannelDataType;
+
+typedef enum SpvImageOperandsShift_ {
+ SpvImageOperandsBiasShift = 0,
+ SpvImageOperandsLodShift = 1,
+ SpvImageOperandsGradShift = 2,
+ SpvImageOperandsConstOffsetShift = 3,
+ SpvImageOperandsOffsetShift = 4,
+ SpvImageOperandsConstOffsetsShift = 5,
+ SpvImageOperandsSampleShift = 6,
+ SpvImageOperandsMinLodShift = 7,
+} SpvImageOperandsShift;
+
+typedef enum SpvImageOperandsMask_ {
+ SpvImageOperandsMaskNone = 0,
+ SpvImageOperandsBiasMask = 0x00000001,
+ SpvImageOperandsLodMask = 0x00000002,
+ SpvImageOperandsGradMask = 0x00000004,
+ SpvImageOperandsConstOffsetMask = 0x00000008,
+ SpvImageOperandsOffsetMask = 0x00000010,
+ SpvImageOperandsConstOffsetsMask = 0x00000020,
+ SpvImageOperandsSampleMask = 0x00000040,
+ SpvImageOperandsMinLodMask = 0x00000080,
+} SpvImageOperandsMask;
+
+typedef enum SpvFPFastMathModeShift_ {
+ SpvFPFastMathModeNotNaNShift = 0,
+ SpvFPFastMathModeNotInfShift = 1,
+ SpvFPFastMathModeNSZShift = 2,
+ SpvFPFastMathModeAllowRecipShift = 3,
+ SpvFPFastMathModeFastShift = 4,
+} SpvFPFastMathModeShift;
+
+typedef enum SpvFPFastMathModeMask_ {
+ SpvFPFastMathModeMaskNone = 0,
+ SpvFPFastMathModeNotNaNMask = 0x00000001,
+ SpvFPFastMathModeNotInfMask = 0x00000002,
+ SpvFPFastMathModeNSZMask = 0x00000004,
+ SpvFPFastMathModeAllowRecipMask = 0x00000008,
+ SpvFPFastMathModeFastMask = 0x00000010,
+} SpvFPFastMathModeMask;
+
+typedef enum SpvFPRoundingMode_ {
+ SpvFPRoundingModeRTE = 0,
+ SpvFPRoundingModeRTZ = 1,
+ SpvFPRoundingModeRTP = 2,
+ SpvFPRoundingModeRTN = 3,
+} SpvFPRoundingMode;
+
+typedef enum SpvLinkageType_ {
+ SpvLinkageTypeExport = 0,
+ SpvLinkageTypeImport = 1,
+} SpvLinkageType;
+
+typedef enum SpvAccessQualifier_ {
+ SpvAccessQualifierReadOnly = 0,
+ SpvAccessQualifierWriteOnly = 1,
+ SpvAccessQualifierReadWrite = 2,
+} SpvAccessQualifier;
+
+typedef enum SpvFunctionParameterAttribute_ {
+ SpvFunctionParameterAttributeZext = 0,
+ SpvFunctionParameterAttributeSext = 1,
+ SpvFunctionParameterAttributeByVal = 2,
+ SpvFunctionParameterAttributeSret = 3,
+ SpvFunctionParameterAttributeNoAlias = 4,
+ SpvFunctionParameterAttributeNoCapture = 5,
+ SpvFunctionParameterAttributeNoWrite = 6,
+ SpvFunctionParameterAttributeNoReadWrite = 7,
+} SpvFunctionParameterAttribute;
+
+typedef enum SpvDecoration_ {
+ SpvDecorationRelaxedPrecision = 0,
+ SpvDecorationSpecId = 1,
+ SpvDecorationBlock = 2,
+ SpvDecorationBufferBlock = 3,
+ SpvDecorationRowMajor = 4,
+ SpvDecorationColMajor = 5,
+ SpvDecorationArrayStride = 6,
+ SpvDecorationMatrixStride = 7,
+ SpvDecorationGLSLShared = 8,
+ SpvDecorationGLSLPacked = 9,
+ SpvDecorationCPacked = 10,
+ SpvDecorationBuiltIn = 11,
+ SpvDecorationNoPerspective = 13,
+ SpvDecorationFlat = 14,
+ SpvDecorationPatch = 15,
+ SpvDecorationCentroid = 16,
+ SpvDecorationSample = 17,
+ SpvDecorationInvariant = 18,
+ SpvDecorationRestrict = 19,
+ SpvDecorationAliased = 20,
+ SpvDecorationVolatile = 21,
+ SpvDecorationConstant = 22,
+ SpvDecorationCoherent = 23,
+ SpvDecorationNonWritable = 24,
+ SpvDecorationNonReadable = 25,
+ SpvDecorationUniform = 26,
+ SpvDecorationSaturatedConversion = 28,
+ SpvDecorationStream = 29,
+ SpvDecorationLocation = 30,
+ SpvDecorationComponent = 31,
+ SpvDecorationIndex = 32,
+ SpvDecorationBinding = 33,
+ SpvDecorationDescriptorSet = 34,
+ SpvDecorationOffset = 35,
+ SpvDecorationXfbBuffer = 36,
+ SpvDecorationXfbStride = 37,
+ SpvDecorationFuncParamAttr = 38,
+ SpvDecorationFPRoundingMode = 39,
+ SpvDecorationFPFastMathMode = 40,
+ SpvDecorationLinkageAttributes = 41,
+ SpvDecorationNoContraction = 42,
+ SpvDecorationInputAttachmentIndex = 43,
+ SpvDecorationAlignment = 44,
+} SpvDecoration;
+
+typedef enum SpvBuiltIn_ {
+ SpvBuiltInPosition = 0,
+ SpvBuiltInPointSize = 1,
+ SpvBuiltInClipDistance = 3,
+ SpvBuiltInCullDistance = 4,
+ SpvBuiltInVertexId = 5,
+ SpvBuiltInInstanceId = 6,
+ SpvBuiltInPrimitiveId = 7,
+ SpvBuiltInInvocationId = 8,
+ SpvBuiltInLayer = 9,
+ SpvBuiltInViewportIndex = 10,
+ SpvBuiltInTessLevelOuter = 11,
+ SpvBuiltInTessLevelInner = 12,
+ SpvBuiltInTessCoord = 13,
+ SpvBuiltInPatchVertices = 14,
+ SpvBuiltInFragCoord = 15,
+ SpvBuiltInPointCoord = 16,
+ SpvBuiltInFrontFacing = 17,
+ SpvBuiltInSampleId = 18,
+ SpvBuiltInSamplePosition = 19,
+ SpvBuiltInSampleMask = 20,
+ SpvBuiltInFragDepth = 22,
+ SpvBuiltInHelperInvocation = 23,
+ SpvBuiltInNumWorkgroups = 24,
+ SpvBuiltInWorkgroupSize = 25,
+ SpvBuiltInWorkgroupId = 26,
+ SpvBuiltInLocalInvocationId = 27,
+ SpvBuiltInGlobalInvocationId = 28,
+ SpvBuiltInLocalInvocationIndex = 29,
+ SpvBuiltInWorkDim = 30,
+ SpvBuiltInGlobalSize = 31,
+ SpvBuiltInEnqueuedWorkgroupSize = 32,
+ SpvBuiltInGlobalOffset = 33,
+ SpvBuiltInGlobalLinearId = 34,
+ SpvBuiltInSubgroupSize = 36,
+ SpvBuiltInSubgroupMaxSize = 37,
+ SpvBuiltInNumSubgroups = 38,
+ SpvBuiltInNumEnqueuedSubgroups = 39,
+ SpvBuiltInSubgroupId = 40,
+ SpvBuiltInSubgroupLocalInvocationId = 41,
+ SpvBuiltInVertexIndex = 42,
+ SpvBuiltInInstanceIndex = 43,
+} SpvBuiltIn;
+
+typedef enum SpvSelectionControlShift_ {
+ SpvSelectionControlFlattenShift = 0,
+ SpvSelectionControlDontFlattenShift = 1,
+} SpvSelectionControlShift;
+
+typedef enum SpvSelectionControlMask_ {
+ SpvSelectionControlMaskNone = 0,
+ SpvSelectionControlFlattenMask = 0x00000001,
+ SpvSelectionControlDontFlattenMask = 0x00000002,
+} SpvSelectionControlMask;
+
+typedef enum SpvLoopControlShift_ {
+ SpvLoopControlUnrollShift = 0,
+ SpvLoopControlDontUnrollShift = 1,
+} SpvLoopControlShift;
+
+typedef enum SpvLoopControlMask_ {
+ SpvLoopControlMaskNone = 0,
+ SpvLoopControlUnrollMask = 0x00000001,
+ SpvLoopControlDontUnrollMask = 0x00000002,
+} SpvLoopControlMask;
+
+typedef enum SpvFunctionControlShift_ {
+ SpvFunctionControlInlineShift = 0,
+ SpvFunctionControlDontInlineShift = 1,
+ SpvFunctionControlPureShift = 2,
+ SpvFunctionControlConstShift = 3,
+} SpvFunctionControlShift;
+
+typedef enum SpvFunctionControlMask_ {
+ SpvFunctionControlMaskNone = 0,
+ SpvFunctionControlInlineMask = 0x00000001,
+ SpvFunctionControlDontInlineMask = 0x00000002,
+ SpvFunctionControlPureMask = 0x00000004,
+ SpvFunctionControlConstMask = 0x00000008,
+} SpvFunctionControlMask;
+
+typedef enum SpvMemorySemanticsShift_ {
+ SpvMemorySemanticsAcquireShift = 1,
+ SpvMemorySemanticsReleaseShift = 2,
+ SpvMemorySemanticsAcquireReleaseShift = 3,
+ SpvMemorySemanticsSequentiallyConsistentShift = 4,
+ SpvMemorySemanticsUniformMemoryShift = 6,
+ SpvMemorySemanticsSubgroupMemoryShift = 7,
+ SpvMemorySemanticsWorkgroupMemoryShift = 8,
+ SpvMemorySemanticsCrossWorkgroupMemoryShift = 9,
+ SpvMemorySemanticsAtomicCounterMemoryShift = 10,
+ SpvMemorySemanticsImageMemoryShift = 11,
+} SpvMemorySemanticsShift;
+
+typedef enum SpvMemorySemanticsMask_ {
+ SpvMemorySemanticsMaskNone = 0,
+ SpvMemorySemanticsAcquireMask = 0x00000002,
+ SpvMemorySemanticsReleaseMask = 0x00000004,
+ SpvMemorySemanticsAcquireReleaseMask = 0x00000008,
+ SpvMemorySemanticsSequentiallyConsistentMask = 0x00000010,
+ SpvMemorySemanticsUniformMemoryMask = 0x00000040,
+ SpvMemorySemanticsSubgroupMemoryMask = 0x00000080,
+ SpvMemorySemanticsWorkgroupMemoryMask = 0x00000100,
+ SpvMemorySemanticsCrossWorkgroupMemoryMask = 0x00000200,
+ SpvMemorySemanticsAtomicCounterMemoryMask = 0x00000400,
+ SpvMemorySemanticsImageMemoryMask = 0x00000800,
+} SpvMemorySemanticsMask;
+
+typedef enum SpvMemoryAccessShift_ {
+ SpvMemoryAccessVolatileShift = 0,
+ SpvMemoryAccessAlignedShift = 1,
+ SpvMemoryAccessNontemporalShift = 2,
+} SpvMemoryAccessShift;
+
+typedef enum SpvMemoryAccessMask_ {
+ SpvMemoryAccessMaskNone = 0,
+ SpvMemoryAccessVolatileMask = 0x00000001,
+ SpvMemoryAccessAlignedMask = 0x00000002,
+ SpvMemoryAccessNontemporalMask = 0x00000004,
+} SpvMemoryAccessMask;
+
+typedef enum SpvScope_ {
+ SpvScopeCrossDevice = 0,
+ SpvScopeDevice = 1,
+ SpvScopeWorkgroup = 2,
+ SpvScopeSubgroup = 3,
+ SpvScopeInvocation = 4,
+} SpvScope;
+
+typedef enum SpvGroupOperation_ {
+ SpvGroupOperationReduce = 0,
+ SpvGroupOperationInclusiveScan = 1,
+ SpvGroupOperationExclusiveScan = 2,
+} SpvGroupOperation;
+
+typedef enum SpvKernelEnqueueFlags_ {
+ SpvKernelEnqueueFlagsNoWait = 0,
+ SpvKernelEnqueueFlagsWaitKernel = 1,
+ SpvKernelEnqueueFlagsWaitWorkGroup = 2,
+} SpvKernelEnqueueFlags;
+
+typedef enum SpvKernelProfilingInfoShift_ {
+ SpvKernelProfilingInfoCmdExecTimeShift = 0,
+} SpvKernelProfilingInfoShift;
+
+typedef enum SpvKernelProfilingInfoMask_ {
+ SpvKernelProfilingInfoMaskNone = 0,
+ SpvKernelProfilingInfoCmdExecTimeMask = 0x00000001,
+} SpvKernelProfilingInfoMask;
+
+typedef enum SpvCapability_ {
+ SpvCapabilityMatrix = 0,
+ SpvCapabilityShader = 1,
+ SpvCapabilityGeometry = 2,
+ SpvCapabilityTessellation = 3,
+ SpvCapabilityAddresses = 4,
+ SpvCapabilityLinkage = 5,
+ SpvCapabilityKernel = 6,
+ SpvCapabilityVector16 = 7,
+ SpvCapabilityFloat16Buffer = 8,
+ SpvCapabilityFloat16 = 9,
+ SpvCapabilityFloat64 = 10,
+ SpvCapabilityInt64 = 11,
+ SpvCapabilityInt64Atomics = 12,
+ SpvCapabilityImageBasic = 13,
+ SpvCapabilityImageReadWrite = 14,
+ SpvCapabilityImageMipmap = 15,
+ SpvCapabilityPipes = 17,
+ SpvCapabilityGroups = 18,
+ SpvCapabilityDeviceEnqueue = 19,
+ SpvCapabilityLiteralSampler = 20,
+ SpvCapabilityAtomicStorage = 21,
+ SpvCapabilityInt16 = 22,
+ SpvCapabilityTessellationPointSize = 23,
+ SpvCapabilityGeometryPointSize = 24,
+ SpvCapabilityImageGatherExtended = 25,
+ SpvCapabilityStorageImageMultisample = 27,
+ SpvCapabilityUniformBufferArrayDynamicIndexing = 28,
+ SpvCapabilitySampledImageArrayDynamicIndexing = 29,
+ SpvCapabilityStorageBufferArrayDynamicIndexing = 30,
+ SpvCapabilityStorageImageArrayDynamicIndexing = 31,
+ SpvCapabilityClipDistance = 32,
+ SpvCapabilityCullDistance = 33,
+ SpvCapabilityImageCubeArray = 34,
+ SpvCapabilitySampleRateShading = 35,
+ SpvCapabilityImageRect = 36,
+ SpvCapabilitySampledRect = 37,
+ SpvCapabilityGenericPointer = 38,
+ SpvCapabilityInt8 = 39,
+ SpvCapabilityInputAttachment = 40,
+ SpvCapabilitySparseResidency = 41,
+ SpvCapabilityMinLod = 42,
+ SpvCapabilitySampled1D = 43,
+ SpvCapabilityImage1D = 44,
+ SpvCapabilitySampledCubeArray = 45,
+ SpvCapabilitySampledBuffer = 46,
+ SpvCapabilityImageBuffer = 47,
+ SpvCapabilityImageMSArray = 48,
+ SpvCapabilityStorageImageExtendedFormats = 49,
+ SpvCapabilityImageQuery = 50,
+ SpvCapabilityDerivativeControl = 51,
+ SpvCapabilityInterpolationFunction = 52,
+ SpvCapabilityTransformFeedback = 53,
+ SpvCapabilityGeometryStreams = 54,
+ SpvCapabilityStorageImageReadWithoutFormat = 55,
+ SpvCapabilityStorageImageWriteWithoutFormat = 56,
+ SpvCapabilityMultiViewport = 57,
+} SpvCapability;
+
+typedef enum SpvOp_ {
+ SpvOpNop = 0,
+ SpvOpUndef = 1,
+ SpvOpSourceContinued = 2,
+ SpvOpSource = 3,
+ SpvOpSourceExtension = 4,
+ SpvOpName = 5,
+ SpvOpMemberName = 6,
+ SpvOpString = 7,
+ SpvOpLine = 8,
+ SpvOpExtension = 10,
+ SpvOpExtInstImport = 11,
+ SpvOpExtInst = 12,
+ SpvOpMemoryModel = 14,
+ SpvOpEntryPoint = 15,
+ SpvOpExecutionMode = 16,
+ SpvOpCapability = 17,
+ SpvOpTypeVoid = 19,
+ SpvOpTypeBool = 20,
+ SpvOpTypeInt = 21,
+ SpvOpTypeFloat = 22,
+ SpvOpTypeVector = 23,
+ SpvOpTypeMatrix = 24,
+ SpvOpTypeImage = 25,
+ SpvOpTypeSampler = 26,
+ SpvOpTypeSampledImage = 27,
+ SpvOpTypeArray = 28,
+ SpvOpTypeRuntimeArray = 29,
+ SpvOpTypeStruct = 30,
+ SpvOpTypeOpaque = 31,
+ SpvOpTypePointer = 32,
+ SpvOpTypeFunction = 33,
+ SpvOpTypeEvent = 34,
+ SpvOpTypeDeviceEvent = 35,
+ SpvOpTypeReserveId = 36,
+ SpvOpTypeQueue = 37,
+ SpvOpTypePipe = 38,
+ SpvOpTypeForwardPointer = 39,
+ SpvOpConstantTrue = 41,
+ SpvOpConstantFalse = 42,
+ SpvOpConstant = 43,
+ SpvOpConstantComposite = 44,
+ SpvOpConstantSampler = 45,
+ SpvOpConstantNull = 46,
+ SpvOpSpecConstantTrue = 48,
+ SpvOpSpecConstantFalse = 49,
+ SpvOpSpecConstant = 50,
+ SpvOpSpecConstantComposite = 51,
+ SpvOpSpecConstantOp = 52,
+ SpvOpFunction = 54,
+ SpvOpFunctionParameter = 55,
+ SpvOpFunctionEnd = 56,
+ SpvOpFunctionCall = 57,
+ SpvOpVariable = 59,
+ SpvOpImageTexelPointer = 60,
+ SpvOpLoad = 61,
+ SpvOpStore = 62,
+ SpvOpCopyMemory = 63,
+ SpvOpCopyMemorySized = 64,
+ SpvOpAccessChain = 65,
+ SpvOpInBoundsAccessChain = 66,
+ SpvOpPtrAccessChain = 67,
+ SpvOpArrayLength = 68,
+ SpvOpGenericPtrMemSemantics = 69,
+ SpvOpInBoundsPtrAccessChain = 70,
+ SpvOpDecorate = 71,
+ SpvOpMemberDecorate = 72,
+ SpvOpDecorationGroup = 73,
+ SpvOpGroupDecorate = 74,
+ SpvOpGroupMemberDecorate = 75,
+ SpvOpVectorExtractDynamic = 77,
+ SpvOpVectorInsertDynamic = 78,
+ SpvOpVectorShuffle = 79,
+ SpvOpCompositeConstruct = 80,
+ SpvOpCompositeExtract = 81,
+ SpvOpCompositeInsert = 82,
+ SpvOpCopyObject = 83,
+ SpvOpTranspose = 84,
+ SpvOpSampledImage = 86,
+ SpvOpImageSampleImplicitLod = 87,
+ SpvOpImageSampleExplicitLod = 88,
+ SpvOpImageSampleDrefImplicitLod = 89,
+ SpvOpImageSampleDrefExplicitLod = 90,
+ SpvOpImageSampleProjImplicitLod = 91,
+ SpvOpImageSampleProjExplicitLod = 92,
+ SpvOpImageSampleProjDrefImplicitLod = 93,
+ SpvOpImageSampleProjDrefExplicitLod = 94,
+ SpvOpImageFetch = 95,
+ SpvOpImageGather = 96,
+ SpvOpImageDrefGather = 97,
+ SpvOpImageRead = 98,
+ SpvOpImageWrite = 99,
+ SpvOpImage = 100,
+ SpvOpImageQueryFormat = 101,
+ SpvOpImageQueryOrder = 102,
+ SpvOpImageQuerySizeLod = 103,
+ SpvOpImageQuerySize = 104,
+ SpvOpImageQueryLod = 105,
+ SpvOpImageQueryLevels = 106,
+ SpvOpImageQuerySamples = 107,
+ SpvOpConvertFToU = 109,
+ SpvOpConvertFToS = 110,
+ SpvOpConvertSToF = 111,
+ SpvOpConvertUToF = 112,
+ SpvOpUConvert = 113,
+ SpvOpSConvert = 114,
+ SpvOpFConvert = 115,
+ SpvOpQuantizeToF16 = 116,
+ SpvOpConvertPtrToU = 117,
+ SpvOpSatConvertSToU = 118,
+ SpvOpSatConvertUToS = 119,
+ SpvOpConvertUToPtr = 120,
+ SpvOpPtrCastToGeneric = 121,
+ SpvOpGenericCastToPtr = 122,
+ SpvOpGenericCastToPtrExplicit = 123,
+ SpvOpBitcast = 124,
+ SpvOpSNegate = 126,
+ SpvOpFNegate = 127,
+ SpvOpIAdd = 128,
+ SpvOpFAdd = 129,
+ SpvOpISub = 130,
+ SpvOpFSub = 131,
+ SpvOpIMul = 132,
+ SpvOpFMul = 133,
+ SpvOpUDiv = 134,
+ SpvOpSDiv = 135,
+ SpvOpFDiv = 136,
+ SpvOpUMod = 137,
+ SpvOpSRem = 138,
+ SpvOpSMod = 139,
+ SpvOpFRem = 140,
+ SpvOpFMod = 141,
+ SpvOpVectorTimesScalar = 142,
+ SpvOpMatrixTimesScalar = 143,
+ SpvOpVectorTimesMatrix = 144,
+ SpvOpMatrixTimesVector = 145,
+ SpvOpMatrixTimesMatrix = 146,
+ SpvOpOuterProduct = 147,
+ SpvOpDot = 148,
+ SpvOpIAddCarry = 149,
+ SpvOpISubBorrow = 150,
+ SpvOpUMulExtended = 151,
+ SpvOpSMulExtended = 152,
+ SpvOpAny = 154,
+ SpvOpAll = 155,
+ SpvOpIsNan = 156,
+ SpvOpIsInf = 157,
+ SpvOpIsFinite = 158,
+ SpvOpIsNormal = 159,
+ SpvOpSignBitSet = 160,
+ SpvOpLessOrGreater = 161,
+ SpvOpOrdered = 162,
+ SpvOpUnordered = 163,
+ SpvOpLogicalEqual = 164,
+ SpvOpLogicalNotEqual = 165,
+ SpvOpLogicalOr = 166,
+ SpvOpLogicalAnd = 167,
+ SpvOpLogicalNot = 168,
+ SpvOpSelect = 169,
+ SpvOpIEqual = 170,
+ SpvOpINotEqual = 171,
+ SpvOpUGreaterThan = 172,
+ SpvOpSGreaterThan = 173,
+ SpvOpUGreaterThanEqual = 174,
+ SpvOpSGreaterThanEqual = 175,
+ SpvOpULessThan = 176,
+ SpvOpSLessThan = 177,
+ SpvOpULessThanEqual = 178,
+ SpvOpSLessThanEqual = 179,
+ SpvOpFOrdEqual = 180,
+ SpvOpFUnordEqual = 181,
+ SpvOpFOrdNotEqual = 182,
+ SpvOpFUnordNotEqual = 183,
+ SpvOpFOrdLessThan = 184,
+ SpvOpFUnordLessThan = 185,
+ SpvOpFOrdGreaterThan = 186,
+ SpvOpFUnordGreaterThan = 187,
+ SpvOpFOrdLessThanEqual = 188,
+ SpvOpFUnordLessThanEqual = 189,
+ SpvOpFOrdGreaterThanEqual = 190,
+ SpvOpFUnordGreaterThanEqual = 191,
+ SpvOpShiftRightLogical = 194,
+ SpvOpShiftRightArithmetic = 195,
+ SpvOpShiftLeftLogical = 196,
+ SpvOpBitwiseOr = 197,
+ SpvOpBitwiseXor = 198,
+ SpvOpBitwiseAnd = 199,
+ SpvOpNot = 200,
+ SpvOpBitFieldInsert = 201,
+ SpvOpBitFieldSExtract = 202,
+ SpvOpBitFieldUExtract = 203,
+ SpvOpBitReverse = 204,
+ SpvOpBitCount = 205,
+ SpvOpDPdx = 207,
+ SpvOpDPdy = 208,
+ SpvOpFwidth = 209,
+ SpvOpDPdxFine = 210,
+ SpvOpDPdyFine = 211,
+ SpvOpFwidthFine = 212,
+ SpvOpDPdxCoarse = 213,
+ SpvOpDPdyCoarse = 214,
+ SpvOpFwidthCoarse = 215,
+ SpvOpEmitVertex = 218,
+ SpvOpEndPrimitive = 219,
+ SpvOpEmitStreamVertex = 220,
+ SpvOpEndStreamPrimitive = 221,
+ SpvOpControlBarrier = 224,
+ SpvOpMemoryBarrier = 225,
+ SpvOpAtomicLoad = 227,
+ SpvOpAtomicStore = 228,
+ SpvOpAtomicExchange = 229,
+ SpvOpAtomicCompareExchange = 230,
+ SpvOpAtomicCompareExchangeWeak = 231,
+ SpvOpAtomicIIncrement = 232,
+ SpvOpAtomicIDecrement = 233,
+ SpvOpAtomicIAdd = 234,
+ SpvOpAtomicISub = 235,
+ SpvOpAtomicSMin = 236,
+ SpvOpAtomicUMin = 237,
+ SpvOpAtomicSMax = 238,
+ SpvOpAtomicUMax = 239,
+ SpvOpAtomicAnd = 240,
+ SpvOpAtomicOr = 241,
+ SpvOpAtomicXor = 242,
+ SpvOpPhi = 245,
+ SpvOpLoopMerge = 246,
+ SpvOpSelectionMerge = 247,
+ SpvOpLabel = 248,
+ SpvOpBranch = 249,
+ SpvOpBranchConditional = 250,
+ SpvOpSwitch = 251,
+ SpvOpKill = 252,
+ SpvOpReturn = 253,
+ SpvOpReturnValue = 254,
+ SpvOpUnreachable = 255,
+ SpvOpLifetimeStart = 256,
+ SpvOpLifetimeStop = 257,
+ SpvOpGroupAsyncCopy = 259,
+ SpvOpGroupWaitEvents = 260,
+ SpvOpGroupAll = 261,
+ SpvOpGroupAny = 262,
+ SpvOpGroupBroadcast = 263,
+ SpvOpGroupIAdd = 264,
+ SpvOpGroupFAdd = 265,
+ SpvOpGroupFMin = 266,
+ SpvOpGroupUMin = 267,
+ SpvOpGroupSMin = 268,
+ SpvOpGroupFMax = 269,
+ SpvOpGroupUMax = 270,
+ SpvOpGroupSMax = 271,
+ SpvOpReadPipe = 274,
+ SpvOpWritePipe = 275,
+ SpvOpReservedReadPipe = 276,
+ SpvOpReservedWritePipe = 277,
+ SpvOpReserveReadPipePackets = 278,
+ SpvOpReserveWritePipePackets = 279,
+ SpvOpCommitReadPipe = 280,
+ SpvOpCommitWritePipe = 281,
+ SpvOpIsValidReserveId = 282,
+ SpvOpGetNumPipePackets = 283,
+ SpvOpGetMaxPipePackets = 284,
+ SpvOpGroupReserveReadPipePackets = 285,
+ SpvOpGroupReserveWritePipePackets = 286,
+ SpvOpGroupCommitReadPipe = 287,
+ SpvOpGroupCommitWritePipe = 288,
+ SpvOpEnqueueMarker = 291,
+ SpvOpEnqueueKernel = 292,
+ SpvOpGetKernelNDrangeSubGroupCount = 293,
+ SpvOpGetKernelNDrangeMaxSubGroupSize = 294,
+ SpvOpGetKernelWorkGroupSize = 295,
+ SpvOpGetKernelPreferredWorkGroupSizeMultiple = 296,
+ SpvOpRetainEvent = 297,
+ SpvOpReleaseEvent = 298,
+ SpvOpCreateUserEvent = 299,
+ SpvOpIsValidEvent = 300,
+ SpvOpSetUserEventStatus = 301,
+ SpvOpCaptureEventProfilingInfo = 302,
+ SpvOpGetDefaultQueue = 303,
+ SpvOpBuildNDRange = 304,
+ SpvOpImageSparseSampleImplicitLod = 305,
+ SpvOpImageSparseSampleExplicitLod = 306,
+ SpvOpImageSparseSampleDrefImplicitLod = 307,
+ SpvOpImageSparseSampleDrefExplicitLod = 308,
+ SpvOpImageSparseSampleProjImplicitLod = 309,
+ SpvOpImageSparseSampleProjExplicitLod = 310,
+ SpvOpImageSparseSampleProjDrefImplicitLod = 311,
+ SpvOpImageSparseSampleProjDrefExplicitLod = 312,
+ SpvOpImageSparseFetch = 313,
+ SpvOpImageSparseGather = 314,
+ SpvOpImageSparseDrefGather = 315,
+ SpvOpImageSparseTexelsResident = 316,
+ SpvOpNoLine = 317,
+ SpvOpAtomicFlagTestAndSet = 318,
+ SpvOpAtomicFlagClear = 319,
+} SpvOp;
+
+#endif // #ifndef spirv_H
+
diff --git a/src/compiler/spirv/spirv_to_nir.c b/src/compiler/spirv/spirv_to_nir.c
new file mode 100644
index 00000000000..99514b49650
--- /dev/null
+++ b/src/compiler/spirv/spirv_to_nir.c
@@ -0,0 +1,2710 @@
+/*
+ * Copyright © 2015 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ *
+ * Authors:
+ * Jason Ekstrand ([email protected])
+ *
+ */
+
+#include "vtn_private.h"
+#include "nir/nir_vla.h"
+#include "nir/nir_control_flow.h"
+#include "nir/nir_constant_expressions.h"
+
+static struct vtn_ssa_value *
+vtn_undef_ssa_value(struct vtn_builder *b, const struct glsl_type *type)
+{
+ struct vtn_ssa_value *val = rzalloc(b, struct vtn_ssa_value);
+ val->type = type;
+
+ if (glsl_type_is_vector_or_scalar(type)) {
+ unsigned num_components = glsl_get_vector_elements(val->type);
+ unsigned bit_size = glsl_get_bit_size(glsl_get_base_type(val->type));
+ val->def = nir_ssa_undef(&b->nb, num_components, bit_size);
+ } else {
+ unsigned elems = glsl_get_length(val->type);
+ val->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
+ if (glsl_type_is_matrix(type)) {
+ const struct glsl_type *elem_type =
+ glsl_vector_type(glsl_get_base_type(type),
+ glsl_get_vector_elements(type));
+
+ for (unsigned i = 0; i < elems; i++)
+ val->elems[i] = vtn_undef_ssa_value(b, elem_type);
+ } else if (glsl_type_is_array(type)) {
+ const struct glsl_type *elem_type = glsl_get_array_element(type);
+ for (unsigned i = 0; i < elems; i++)
+ val->elems[i] = vtn_undef_ssa_value(b, elem_type);
+ } else {
+ for (unsigned i = 0; i < elems; i++) {
+ const struct glsl_type *elem_type = glsl_get_struct_field(type, i);
+ val->elems[i] = vtn_undef_ssa_value(b, elem_type);
+ }
+ }
+ }
+
+ return val;
+}
+
+static struct vtn_ssa_value *
+vtn_const_ssa_value(struct vtn_builder *b, nir_constant *constant,
+ const struct glsl_type *type)
+{
+ struct hash_entry *entry = _mesa_hash_table_search(b->const_table, constant);
+
+ if (entry)
+ return entry->data;
+
+ struct vtn_ssa_value *val = rzalloc(b, struct vtn_ssa_value);
+ val->type = type;
+
+ switch (glsl_get_base_type(type)) {
+ case GLSL_TYPE_INT:
+ case GLSL_TYPE_UINT:
+ case GLSL_TYPE_BOOL:
+ case GLSL_TYPE_FLOAT:
+ case GLSL_TYPE_DOUBLE:
+ if (glsl_type_is_vector_or_scalar(type)) {
+ unsigned num_components = glsl_get_vector_elements(val->type);
+ nir_load_const_instr *load =
+ nir_load_const_instr_create(b->shader, num_components, 32);
+
+ for (unsigned i = 0; i < num_components; i++)
+ load->value.u32[i] = constant->value.u[i];
+
+ nir_instr_insert_before_cf_list(&b->impl->body, &load->instr);
+ val->def = &load->def;
+ } else {
+ assert(glsl_type_is_matrix(type));
+ unsigned rows = glsl_get_vector_elements(val->type);
+ unsigned columns = glsl_get_matrix_columns(val->type);
+ val->elems = ralloc_array(b, struct vtn_ssa_value *, columns);
+
+ for (unsigned i = 0; i < columns; i++) {
+ struct vtn_ssa_value *col_val = rzalloc(b, struct vtn_ssa_value);
+ col_val->type = glsl_get_column_type(val->type);
+ nir_load_const_instr *load =
+ nir_load_const_instr_create(b->shader, rows, 32);
+
+ for (unsigned j = 0; j < rows; j++)
+ load->value.u32[j] = constant->value.u[rows * i + j];
+
+ nir_instr_insert_before_cf_list(&b->impl->body, &load->instr);
+ col_val->def = &load->def;
+
+ val->elems[i] = col_val;
+ }
+ }
+ break;
+
+ case GLSL_TYPE_ARRAY: {
+ unsigned elems = glsl_get_length(val->type);
+ val->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
+ const struct glsl_type *elem_type = glsl_get_array_element(val->type);
+ for (unsigned i = 0; i < elems; i++)
+ val->elems[i] = vtn_const_ssa_value(b, constant->elements[i],
+ elem_type);
+ break;
+ }
+
+ case GLSL_TYPE_STRUCT: {
+ unsigned elems = glsl_get_length(val->type);
+ val->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
+ for (unsigned i = 0; i < elems; i++) {
+ const struct glsl_type *elem_type =
+ glsl_get_struct_field(val->type, i);
+ val->elems[i] = vtn_const_ssa_value(b, constant->elements[i],
+ elem_type);
+ }
+ break;
+ }
+
+ default:
+ unreachable("bad constant type");
+ }
+
+ return val;
+}
+
+struct vtn_ssa_value *
+vtn_ssa_value(struct vtn_builder *b, uint32_t value_id)
+{
+ struct vtn_value *val = vtn_untyped_value(b, value_id);
+ switch (val->value_type) {
+ case vtn_value_type_undef:
+ return vtn_undef_ssa_value(b, val->type->type);
+
+ case vtn_value_type_constant:
+ return vtn_const_ssa_value(b, val->constant, val->const_type);
+
+ case vtn_value_type_ssa:
+ return val->ssa;
+
+ case vtn_value_type_access_chain:
+ /* This is needed for function parameters */
+ return vtn_variable_load(b, val->access_chain);
+
+ default:
+ unreachable("Invalid type for an SSA value");
+ }
+}
+
+static char *
+vtn_string_literal(struct vtn_builder *b, const uint32_t *words,
+ unsigned word_count, unsigned *words_used)
+{
+ char *dup = ralloc_strndup(b, (char *)words, word_count * sizeof(*words));
+ if (words_used) {
+ /* Ammount of space taken by the string (including the null) */
+ unsigned len = strlen(dup) + 1;
+ *words_used = DIV_ROUND_UP(len, sizeof(*words));
+ }
+ return dup;
+}
+
+const uint32_t *
+vtn_foreach_instruction(struct vtn_builder *b, const uint32_t *start,
+ const uint32_t *end, vtn_instruction_handler handler)
+{
+ b->file = NULL;
+ b->line = -1;
+ b->col = -1;
+
+ const uint32_t *w = start;
+ while (w < end) {
+ SpvOp opcode = w[0] & SpvOpCodeMask;
+ unsigned count = w[0] >> SpvWordCountShift;
+ assert(count >= 1 && w + count <= end);
+
+ switch (opcode) {
+ case SpvOpNop:
+ break; /* Do nothing */
+
+ case SpvOpLine:
+ b->file = vtn_value(b, w[1], vtn_value_type_string)->str;
+ b->line = w[2];
+ b->col = w[3];
+ break;
+
+ case SpvOpNoLine:
+ b->file = NULL;
+ b->line = -1;
+ b->col = -1;
+ break;
+
+ default:
+ if (!handler(b, opcode, w, count))
+ return w;
+ break;
+ }
+
+ w += count;
+ }
+ assert(w == end);
+ return w;
+}
+
+static void
+vtn_handle_extension(struct vtn_builder *b, SpvOp opcode,
+ const uint32_t *w, unsigned count)
+{
+ switch (opcode) {
+ case SpvOpExtInstImport: {
+ struct vtn_value *val = vtn_push_value(b, w[1], vtn_value_type_extension);
+ if (strcmp((const char *)&w[2], "GLSL.std.450") == 0) {
+ val->ext_handler = vtn_handle_glsl450_instruction;
+ } else {
+ assert(!"Unsupported extension");
+ }
+ break;
+ }
+
+ case SpvOpExtInst: {
+ struct vtn_value *val = vtn_value(b, w[3], vtn_value_type_extension);
+ bool handled = val->ext_handler(b, w[4], w, count);
+ (void)handled;
+ assert(handled);
+ break;
+ }
+
+ default:
+ unreachable("Unhandled opcode");
+ }
+}
+
+static void
+_foreach_decoration_helper(struct vtn_builder *b,
+ struct vtn_value *base_value,
+ int parent_member,
+ struct vtn_value *value,
+ vtn_decoration_foreach_cb cb, void *data)
+{
+ for (struct vtn_decoration *dec = value->decoration; dec; dec = dec->next) {
+ int member;
+ if (dec->scope == VTN_DEC_DECORATION) {
+ member = parent_member;
+ } else if (dec->scope >= VTN_DEC_STRUCT_MEMBER0) {
+ assert(parent_member == -1);
+ member = dec->scope - VTN_DEC_STRUCT_MEMBER0;
+ } else {
+ /* Not a decoration */
+ continue;
+ }
+
+ if (dec->group) {
+ assert(dec->group->value_type == vtn_value_type_decoration_group);
+ _foreach_decoration_helper(b, base_value, member, dec->group,
+ cb, data);
+ } else {
+ cb(b, base_value, member, dec, data);
+ }
+ }
+}
+
+/** Iterates (recursively if needed) over all of the decorations on a value
+ *
+ * This function iterates over all of the decorations applied to a given
+ * value. If it encounters a decoration group, it recurses into the group
+ * and iterates over all of those decorations as well.
+ */
+void
+vtn_foreach_decoration(struct vtn_builder *b, struct vtn_value *value,
+ vtn_decoration_foreach_cb cb, void *data)
+{
+ _foreach_decoration_helper(b, value, -1, value, cb, data);
+}
+
+void
+vtn_foreach_execution_mode(struct vtn_builder *b, struct vtn_value *value,
+ vtn_execution_mode_foreach_cb cb, void *data)
+{
+ for (struct vtn_decoration *dec = value->decoration; dec; dec = dec->next) {
+ if (dec->scope != VTN_DEC_EXECUTION_MODE)
+ continue;
+
+ assert(dec->group == NULL);
+ cb(b, value, dec, data);
+ }
+}
+
+static void
+vtn_handle_decoration(struct vtn_builder *b, SpvOp opcode,
+ const uint32_t *w, unsigned count)
+{
+ const uint32_t *w_end = w + count;
+ const uint32_t target = w[1];
+ w += 2;
+
+ switch (opcode) {
+ case SpvOpDecorationGroup:
+ vtn_push_value(b, target, vtn_value_type_decoration_group);
+ break;
+
+ case SpvOpDecorate:
+ case SpvOpMemberDecorate:
+ case SpvOpExecutionMode: {
+ struct vtn_value *val = &b->values[target];
+
+ struct vtn_decoration *dec = rzalloc(b, struct vtn_decoration);
+ switch (opcode) {
+ case SpvOpDecorate:
+ dec->scope = VTN_DEC_DECORATION;
+ break;
+ case SpvOpMemberDecorate:
+ dec->scope = VTN_DEC_STRUCT_MEMBER0 + *(w++);
+ break;
+ case SpvOpExecutionMode:
+ dec->scope = VTN_DEC_EXECUTION_MODE;
+ break;
+ default:
+ unreachable("Invalid decoration opcode");
+ }
+ dec->decoration = *(w++);
+ dec->literals = w;
+
+ /* Link into the list */
+ dec->next = val->decoration;
+ val->decoration = dec;
+ break;
+ }
+
+ case SpvOpGroupMemberDecorate:
+ case SpvOpGroupDecorate: {
+ struct vtn_value *group =
+ vtn_value(b, target, vtn_value_type_decoration_group);
+
+ for (; w < w_end; w++) {
+ struct vtn_value *val = vtn_untyped_value(b, *w);
+ struct vtn_decoration *dec = rzalloc(b, struct vtn_decoration);
+
+ dec->group = group;
+ if (opcode == SpvOpGroupDecorate) {
+ dec->scope = VTN_DEC_DECORATION;
+ } else {
+ dec->scope = VTN_DEC_STRUCT_MEMBER0 + *(++w);
+ }
+
+ /* Link into the list */
+ dec->next = val->decoration;
+ val->decoration = dec;
+ }
+ break;
+ }
+
+ default:
+ unreachable("Unhandled opcode");
+ }
+}
+
+struct member_decoration_ctx {
+ unsigned num_fields;
+ struct glsl_struct_field *fields;
+ struct vtn_type *type;
+};
+
+/* does a shallow copy of a vtn_type */
+
+static struct vtn_type *
+vtn_type_copy(struct vtn_builder *b, struct vtn_type *src)
+{
+ struct vtn_type *dest = ralloc(b, struct vtn_type);
+ dest->type = src->type;
+ dest->is_builtin = src->is_builtin;
+ if (src->is_builtin)
+ dest->builtin = src->builtin;
+
+ if (!glsl_type_is_scalar(src->type)) {
+ switch (glsl_get_base_type(src->type)) {
+ case GLSL_TYPE_INT:
+ case GLSL_TYPE_UINT:
+ case GLSL_TYPE_BOOL:
+ case GLSL_TYPE_FLOAT:
+ case GLSL_TYPE_DOUBLE:
+ case GLSL_TYPE_ARRAY:
+ dest->row_major = src->row_major;
+ dest->stride = src->stride;
+ dest->array_element = src->array_element;
+ break;
+
+ case GLSL_TYPE_STRUCT: {
+ unsigned elems = glsl_get_length(src->type);
+
+ dest->members = ralloc_array(b, struct vtn_type *, elems);
+ memcpy(dest->members, src->members, elems * sizeof(struct vtn_type *));
+
+ dest->offsets = ralloc_array(b, unsigned, elems);
+ memcpy(dest->offsets, src->offsets, elems * sizeof(unsigned));
+ break;
+ }
+
+ default:
+ unreachable("unhandled type");
+ }
+ }
+
+ return dest;
+}
+
+static struct vtn_type *
+mutable_matrix_member(struct vtn_builder *b, struct vtn_type *type, int member)
+{
+ type->members[member] = vtn_type_copy(b, type->members[member]);
+ type = type->members[member];
+
+ /* We may have an array of matrices.... Oh, joy! */
+ while (glsl_type_is_array(type->type)) {
+ type->array_element = vtn_type_copy(b, type->array_element);
+ type = type->array_element;
+ }
+
+ assert(glsl_type_is_matrix(type->type));
+
+ return type;
+}
+
+static void
+struct_member_decoration_cb(struct vtn_builder *b,
+ struct vtn_value *val, int member,
+ const struct vtn_decoration *dec, void *void_ctx)
+{
+ struct member_decoration_ctx *ctx = void_ctx;
+
+ if (member < 0)
+ return;
+
+ assert(member < ctx->num_fields);
+
+ switch (dec->decoration) {
+ case SpvDecorationRelaxedPrecision:
+ break; /* FIXME: Do nothing with this for now. */
+ case SpvDecorationNoPerspective:
+ ctx->fields[member].interpolation = INTERP_QUALIFIER_NOPERSPECTIVE;
+ break;
+ case SpvDecorationFlat:
+ ctx->fields[member].interpolation = INTERP_QUALIFIER_FLAT;
+ break;
+ case SpvDecorationCentroid:
+ ctx->fields[member].centroid = true;
+ break;
+ case SpvDecorationSample:
+ ctx->fields[member].sample = true;
+ break;
+ case SpvDecorationLocation:
+ ctx->fields[member].location = dec->literals[0];
+ break;
+ case SpvDecorationBuiltIn:
+ ctx->type->members[member] = vtn_type_copy(b, ctx->type->members[member]);
+ ctx->type->members[member]->is_builtin = true;
+ ctx->type->members[member]->builtin = dec->literals[0];
+ ctx->type->builtin_block = true;
+ break;
+ case SpvDecorationOffset:
+ ctx->type->offsets[member] = dec->literals[0];
+ break;
+ case SpvDecorationMatrixStride:
+ mutable_matrix_member(b, ctx->type, member)->stride = dec->literals[0];
+ break;
+ case SpvDecorationColMajor:
+ break; /* Nothing to do here. Column-major is the default. */
+ case SpvDecorationRowMajor:
+ mutable_matrix_member(b, ctx->type, member)->row_major = true;
+ break;
+ default:
+ unreachable("Unhandled member decoration");
+ }
+}
+
+static void
+type_decoration_cb(struct vtn_builder *b,
+ struct vtn_value *val, int member,
+ const struct vtn_decoration *dec, void *ctx)
+{
+ struct vtn_type *type = val->type;
+
+ if (member != -1)
+ return;
+
+ switch (dec->decoration) {
+ case SpvDecorationArrayStride:
+ type->stride = dec->literals[0];
+ break;
+ case SpvDecorationBlock:
+ type->block = true;
+ break;
+ case SpvDecorationBufferBlock:
+ type->buffer_block = true;
+ break;
+ case SpvDecorationGLSLShared:
+ case SpvDecorationGLSLPacked:
+ /* Ignore these, since we get explicit offsets anyways */
+ break;
+
+ case SpvDecorationStream:
+ assert(dec->literals[0] == 0);
+ break;
+
+ default:
+ unreachable("Unhandled type decoration");
+ }
+}
+
+static unsigned
+translate_image_format(SpvImageFormat format)
+{
+ switch (format) {
+ case SpvImageFormatUnknown: return 0; /* GL_NONE */
+ case SpvImageFormatRgba32f: return 0x8814; /* GL_RGBA32F */
+ case SpvImageFormatRgba16f: return 0x881A; /* GL_RGBA16F */
+ case SpvImageFormatR32f: return 0x822E; /* GL_R32F */
+ case SpvImageFormatRgba8: return 0x8058; /* GL_RGBA8 */
+ case SpvImageFormatRgba8Snorm: return 0x8F97; /* GL_RGBA8_SNORM */
+ case SpvImageFormatRg32f: return 0x8230; /* GL_RG32F */
+ case SpvImageFormatRg16f: return 0x822F; /* GL_RG16F */
+ case SpvImageFormatR11fG11fB10f: return 0x8C3A; /* GL_R11F_G11F_B10F */
+ case SpvImageFormatR16f: return 0x822D; /* GL_R16F */
+ case SpvImageFormatRgba16: return 0x805B; /* GL_RGBA16 */
+ case SpvImageFormatRgb10A2: return 0x8059; /* GL_RGB10_A2 */
+ case SpvImageFormatRg16: return 0x822C; /* GL_RG16 */
+ case SpvImageFormatRg8: return 0x822B; /* GL_RG8 */
+ case SpvImageFormatR16: return 0x822A; /* GL_R16 */
+ case SpvImageFormatR8: return 0x8229; /* GL_R8 */
+ case SpvImageFormatRgba16Snorm: return 0x8F9B; /* GL_RGBA16_SNORM */
+ case SpvImageFormatRg16Snorm: return 0x8F99; /* GL_RG16_SNORM */
+ case SpvImageFormatRg8Snorm: return 0x8F95; /* GL_RG8_SNORM */
+ case SpvImageFormatR16Snorm: return 0x8F98; /* GL_R16_SNORM */
+ case SpvImageFormatR8Snorm: return 0x8F94; /* GL_R8_SNORM */
+ case SpvImageFormatRgba32i: return 0x8D82; /* GL_RGBA32I */
+ case SpvImageFormatRgba16i: return 0x8D88; /* GL_RGBA16I */
+ case SpvImageFormatRgba8i: return 0x8D8E; /* GL_RGBA8I */
+ case SpvImageFormatR32i: return 0x8235; /* GL_R32I */
+ case SpvImageFormatRg32i: return 0x823B; /* GL_RG32I */
+ case SpvImageFormatRg16i: return 0x8239; /* GL_RG16I */
+ case SpvImageFormatRg8i: return 0x8237; /* GL_RG8I */
+ case SpvImageFormatR16i: return 0x8233; /* GL_R16I */
+ case SpvImageFormatR8i: return 0x8231; /* GL_R8I */
+ case SpvImageFormatRgba32ui: return 0x8D70; /* GL_RGBA32UI */
+ case SpvImageFormatRgba16ui: return 0x8D76; /* GL_RGBA16UI */
+ case SpvImageFormatRgba8ui: return 0x8D7C; /* GL_RGBA8UI */
+ case SpvImageFormatR32ui: return 0x8236; /* GL_R32UI */
+ case SpvImageFormatRgb10a2ui: return 0x906F; /* GL_RGB10_A2UI */
+ case SpvImageFormatRg32ui: return 0x823C; /* GL_RG32UI */
+ case SpvImageFormatRg16ui: return 0x823A; /* GL_RG16UI */
+ case SpvImageFormatRg8ui: return 0x8238; /* GL_RG8UI */
+ case SpvImageFormatR16ui: return 0x823A; /* GL_RG16UI */
+ case SpvImageFormatR8ui: return 0x8232; /* GL_R8UI */
+ default:
+ assert(!"Invalid image format");
+ return 0;
+ }
+}
+
+static void
+vtn_handle_type(struct vtn_builder *b, SpvOp opcode,
+ const uint32_t *w, unsigned count)
+{
+ struct vtn_value *val = vtn_push_value(b, w[1], vtn_value_type_type);
+
+ val->type = rzalloc(b, struct vtn_type);
+ val->type->is_builtin = false;
+ val->type->val = val;
+
+ switch (opcode) {
+ case SpvOpTypeVoid:
+ val->type->type = glsl_void_type();
+ break;
+ case SpvOpTypeBool:
+ val->type->type = glsl_bool_type();
+ break;
+ case SpvOpTypeInt: {
+ const bool signedness = w[3];
+ val->type->type = (signedness ? glsl_int_type() : glsl_uint_type());
+ break;
+ }
+ case SpvOpTypeFloat:
+ val->type->type = glsl_float_type();
+ break;
+
+ case SpvOpTypeVector: {
+ struct vtn_type *base = vtn_value(b, w[2], vtn_value_type_type)->type;
+ unsigned elems = w[3];
+
+ assert(glsl_type_is_scalar(base->type));
+ val->type->type = glsl_vector_type(glsl_get_base_type(base->type), elems);
+
+ /* Vectors implicitly have sizeof(base_type) stride. For now, this
+ * is always 4 bytes. This will have to change if we want to start
+ * supporting doubles or half-floats.
+ */
+ val->type->stride = 4;
+ val->type->array_element = base;
+ break;
+ }
+
+ case SpvOpTypeMatrix: {
+ struct vtn_type *base = vtn_value(b, w[2], vtn_value_type_type)->type;
+ unsigned columns = w[3];
+
+ assert(glsl_type_is_vector(base->type));
+ val->type->type = glsl_matrix_type(glsl_get_base_type(base->type),
+ glsl_get_vector_elements(base->type),
+ columns);
+ assert(!glsl_type_is_error(val->type->type));
+ val->type->array_element = base;
+ val->type->row_major = false;
+ val->type->stride = 0;
+ break;
+ }
+
+ case SpvOpTypeRuntimeArray:
+ case SpvOpTypeArray: {
+ struct vtn_type *array_element =
+ vtn_value(b, w[2], vtn_value_type_type)->type;
+
+ unsigned length;
+ if (opcode == SpvOpTypeRuntimeArray) {
+ /* A length of 0 is used to denote unsized arrays */
+ length = 0;
+ } else {
+ length =
+ vtn_value(b, w[3], vtn_value_type_constant)->constant->value.u[0];
+ }
+
+ val->type->type = glsl_array_type(array_element->type, length);
+ val->type->array_element = array_element;
+ val->type->stride = 0;
+ break;
+ }
+
+ case SpvOpTypeStruct: {
+ unsigned num_fields = count - 2;
+ val->type->members = ralloc_array(b, struct vtn_type *, num_fields);
+ val->type->offsets = ralloc_array(b, unsigned, num_fields);
+
+ NIR_VLA(struct glsl_struct_field, fields, count);
+ for (unsigned i = 0; i < num_fields; i++) {
+ val->type->members[i] =
+ vtn_value(b, w[i + 2], vtn_value_type_type)->type;
+ fields[i] = (struct glsl_struct_field) {
+ .type = val->type->members[i]->type,
+ .name = ralloc_asprintf(b, "field%d", i),
+ .location = -1,
+ };
+ }
+
+ struct member_decoration_ctx ctx = {
+ .num_fields = num_fields,
+ .fields = fields,
+ .type = val->type
+ };
+
+ vtn_foreach_decoration(b, val, struct_member_decoration_cb, &ctx);
+
+ const char *name = val->name ? val->name : "struct";
+
+ val->type->type = glsl_struct_type(fields, num_fields, name);
+ break;
+ }
+
+ case SpvOpTypeFunction: {
+ const struct glsl_type *return_type =
+ vtn_value(b, w[2], vtn_value_type_type)->type->type;
+ NIR_VLA(struct glsl_function_param, params, count - 3);
+ for (unsigned i = 0; i < count - 3; i++) {
+ params[i].type = vtn_value(b, w[i + 3], vtn_value_type_type)->type->type;
+
+ /* FIXME: */
+ params[i].in = true;
+ params[i].out = true;
+ }
+ val->type->type = glsl_function_type(return_type, params, count - 3);
+ break;
+ }
+
+ case SpvOpTypePointer:
+ /* FIXME: For now, we'll just do the really lame thing and return
+ * the same type. The validator should ensure that the proper number
+ * of dereferences happen
+ */
+ val->type = vtn_value(b, w[3], vtn_value_type_type)->type;
+ break;
+
+ case SpvOpTypeImage: {
+ const struct glsl_type *sampled_type =
+ vtn_value(b, w[2], vtn_value_type_type)->type->type;
+
+ assert(glsl_type_is_vector_or_scalar(sampled_type));
+
+ enum glsl_sampler_dim dim;
+ switch ((SpvDim)w[3]) {
+ case SpvDim1D: dim = GLSL_SAMPLER_DIM_1D; break;
+ case SpvDim2D: dim = GLSL_SAMPLER_DIM_2D; break;
+ case SpvDim3D: dim = GLSL_SAMPLER_DIM_3D; break;
+ case SpvDimCube: dim = GLSL_SAMPLER_DIM_CUBE; break;
+ case SpvDimRect: dim = GLSL_SAMPLER_DIM_RECT; break;
+ case SpvDimBuffer: dim = GLSL_SAMPLER_DIM_BUF; break;
+ default:
+ unreachable("Invalid SPIR-V Sampler dimension");
+ }
+
+ bool is_shadow = w[4];
+ bool is_array = w[5];
+ bool multisampled = w[6];
+ unsigned sampled = w[7];
+ SpvImageFormat format = w[8];
+
+ if (count > 9)
+ val->type->access_qualifier = w[9];
+ else
+ val->type->access_qualifier = SpvAccessQualifierReadWrite;
+
+ if (multisampled) {
+ assert(dim == GLSL_SAMPLER_DIM_2D);
+ dim = GLSL_SAMPLER_DIM_MS;
+ }
+
+ val->type->image_format = translate_image_format(format);
+
+ if (sampled == 1) {
+ val->type->type = glsl_sampler_type(dim, is_shadow, is_array,
+ glsl_get_base_type(sampled_type));
+ } else if (sampled == 2) {
+ assert(format);
+ assert(!is_shadow);
+ val->type->type = glsl_image_type(dim, is_array,
+ glsl_get_base_type(sampled_type));
+ } else {
+ assert(!"We need to know if the image will be sampled");
+ }
+ break;
+ }
+
+ case SpvOpTypeSampledImage:
+ val->type = vtn_value(b, w[2], vtn_value_type_type)->type;
+ break;
+
+ case SpvOpTypeSampler:
+ /* The actual sampler type here doesn't really matter. It gets
+ * thrown away the moment you combine it with an image. What really
+ * matters is that it's a sampler type as opposed to an integer type
+ * so the backend knows what to do.
+ */
+ val->type->type = glsl_bare_sampler_type();
+ break;
+
+ case SpvOpTypeOpaque:
+ case SpvOpTypeEvent:
+ case SpvOpTypeDeviceEvent:
+ case SpvOpTypeReserveId:
+ case SpvOpTypeQueue:
+ case SpvOpTypePipe:
+ default:
+ unreachable("Unhandled opcode");
+ }
+
+ vtn_foreach_decoration(b, val, type_decoration_cb, NULL);
+}
+
+static nir_constant *
+vtn_null_constant(struct vtn_builder *b, const struct glsl_type *type)
+{
+ nir_constant *c = rzalloc(b, nir_constant);
+
+ switch (glsl_get_base_type(type)) {
+ case GLSL_TYPE_INT:
+ case GLSL_TYPE_UINT:
+ case GLSL_TYPE_BOOL:
+ case GLSL_TYPE_FLOAT:
+ case GLSL_TYPE_DOUBLE:
+ /* Nothing to do here. It's already initialized to zero */
+ break;
+
+ case GLSL_TYPE_ARRAY:
+ assert(glsl_get_length(type) > 0);
+ c->num_elements = glsl_get_length(type);
+ c->elements = ralloc_array(b, nir_constant *, c->num_elements);
+
+ c->elements[0] = vtn_null_constant(b, glsl_get_array_element(type));
+ for (unsigned i = 1; i < c->num_elements; i++)
+ c->elements[i] = c->elements[0];
+ break;
+
+ case GLSL_TYPE_STRUCT:
+ c->num_elements = glsl_get_length(type);
+ c->elements = ralloc_array(b, nir_constant *, c->num_elements);
+
+ for (unsigned i = 0; i < c->num_elements; i++) {
+ c->elements[i] = vtn_null_constant(b, glsl_get_struct_field(type, i));
+ }
+ break;
+
+ default:
+ unreachable("Invalid type for null constant");
+ }
+
+ return c;
+}
+
+static void
+spec_constant_deocoration_cb(struct vtn_builder *b, struct vtn_value *v,
+ int member, const struct vtn_decoration *dec,
+ void *data)
+{
+ assert(member == -1);
+ if (dec->decoration != SpvDecorationSpecId)
+ return;
+
+ uint32_t *const_value = data;
+
+ for (unsigned i = 0; i < b->num_specializations; i++) {
+ if (b->specializations[i].id == dec->literals[0]) {
+ *const_value = b->specializations[i].data;
+ return;
+ }
+ }
+}
+
+static uint32_t
+get_specialization(struct vtn_builder *b, struct vtn_value *val,
+ uint32_t const_value)
+{
+ vtn_foreach_decoration(b, val, spec_constant_deocoration_cb, &const_value);
+ return const_value;
+}
+
+static void
+vtn_handle_constant(struct vtn_builder *b, SpvOp opcode,
+ const uint32_t *w, unsigned count)
+{
+ struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_constant);
+ val->const_type = vtn_value(b, w[1], vtn_value_type_type)->type->type;
+ val->constant = rzalloc(b, nir_constant);
+ switch (opcode) {
+ case SpvOpConstantTrue:
+ assert(val->const_type == glsl_bool_type());
+ val->constant->value.u[0] = NIR_TRUE;
+ break;
+ case SpvOpConstantFalse:
+ assert(val->const_type == glsl_bool_type());
+ val->constant->value.u[0] = NIR_FALSE;
+ break;
+
+ case SpvOpSpecConstantTrue:
+ case SpvOpSpecConstantFalse: {
+ assert(val->const_type == glsl_bool_type());
+ uint32_t int_val =
+ get_specialization(b, val, (opcode == SpvOpSpecConstantTrue));
+ val->constant->value.u[0] = int_val ? NIR_TRUE : NIR_FALSE;
+ break;
+ }
+
+ case SpvOpConstant:
+ assert(glsl_type_is_scalar(val->const_type));
+ val->constant->value.u[0] = w[3];
+ break;
+ case SpvOpSpecConstant:
+ assert(glsl_type_is_scalar(val->const_type));
+ val->constant->value.u[0] = get_specialization(b, val, w[3]);
+ break;
+ case SpvOpSpecConstantComposite:
+ case SpvOpConstantComposite: {
+ unsigned elem_count = count - 3;
+ nir_constant **elems = ralloc_array(b, nir_constant *, elem_count);
+ for (unsigned i = 0; i < elem_count; i++)
+ elems[i] = vtn_value(b, w[i + 3], vtn_value_type_constant)->constant;
+
+ switch (glsl_get_base_type(val->const_type)) {
+ case GLSL_TYPE_UINT:
+ case GLSL_TYPE_INT:
+ case GLSL_TYPE_FLOAT:
+ case GLSL_TYPE_BOOL:
+ if (glsl_type_is_matrix(val->const_type)) {
+ unsigned rows = glsl_get_vector_elements(val->const_type);
+ assert(glsl_get_matrix_columns(val->const_type) == elem_count);
+ for (unsigned i = 0; i < elem_count; i++)
+ for (unsigned j = 0; j < rows; j++)
+ val->constant->value.u[rows * i + j] = elems[i]->value.u[j];
+ } else {
+ assert(glsl_type_is_vector(val->const_type));
+ assert(glsl_get_vector_elements(val->const_type) == elem_count);
+ for (unsigned i = 0; i < elem_count; i++)
+ val->constant->value.u[i] = elems[i]->value.u[0];
+ }
+ ralloc_free(elems);
+ break;
+
+ case GLSL_TYPE_STRUCT:
+ case GLSL_TYPE_ARRAY:
+ ralloc_steal(val->constant, elems);
+ val->constant->num_elements = elem_count;
+ val->constant->elements = elems;
+ break;
+
+ default:
+ unreachable("Unsupported type for constants");
+ }
+ break;
+ }
+
+ case SpvOpSpecConstantOp: {
+ SpvOp opcode = get_specialization(b, val, w[3]);
+ switch (opcode) {
+ case SpvOpVectorShuffle: {
+ struct vtn_value *v0 = vtn_value(b, w[4], vtn_value_type_constant);
+ struct vtn_value *v1 = vtn_value(b, w[5], vtn_value_type_constant);
+ unsigned len0 = glsl_get_vector_elements(v0->const_type);
+ unsigned len1 = glsl_get_vector_elements(v1->const_type);
+
+ uint32_t u[8];
+ for (unsigned i = 0; i < len0; i++)
+ u[i] = v0->constant->value.u[i];
+ for (unsigned i = 0; i < len1; i++)
+ u[len0 + i] = v1->constant->value.u[i];
+
+ for (unsigned i = 0; i < count - 6; i++) {
+ uint32_t comp = w[i + 6];
+ if (comp == (uint32_t)-1) {
+ val->constant->value.u[i] = 0xdeadbeef;
+ } else {
+ val->constant->value.u[i] = u[comp];
+ }
+ }
+ return;
+ }
+
+ case SpvOpCompositeExtract:
+ case SpvOpCompositeInsert: {
+ struct vtn_value *comp;
+ unsigned deref_start;
+ struct nir_constant **c;
+ if (opcode == SpvOpCompositeExtract) {
+ comp = vtn_value(b, w[4], vtn_value_type_constant);
+ deref_start = 5;
+ c = &comp->constant;
+ } else {
+ comp = vtn_value(b, w[5], vtn_value_type_constant);
+ deref_start = 6;
+ val->constant = nir_constant_clone(comp->constant,
+ (nir_variable *)b);
+ c = &val->constant;
+ }
+
+ int elem = -1;
+ const struct glsl_type *type = comp->const_type;
+ for (unsigned i = deref_start; i < count; i++) {
+ switch (glsl_get_base_type(type)) {
+ case GLSL_TYPE_UINT:
+ case GLSL_TYPE_INT:
+ case GLSL_TYPE_FLOAT:
+ case GLSL_TYPE_BOOL:
+ /* If we hit this granularity, we're picking off an element */
+ if (elem < 0)
+ elem = 0;
+
+ if (glsl_type_is_matrix(type)) {
+ elem += w[i] * glsl_get_vector_elements(type);
+ type = glsl_get_column_type(type);
+ } else {
+ assert(glsl_type_is_vector(type));
+ elem += w[i];
+ type = glsl_scalar_type(glsl_get_base_type(type));
+ }
+ continue;
+
+ case GLSL_TYPE_ARRAY:
+ c = &(*c)->elements[w[i]];
+ type = glsl_get_array_element(type);
+ continue;
+
+ case GLSL_TYPE_STRUCT:
+ c = &(*c)->elements[w[i]];
+ type = glsl_get_struct_field(type, w[i]);
+ continue;
+
+ default:
+ unreachable("Invalid constant type");
+ }
+ }
+
+ if (opcode == SpvOpCompositeExtract) {
+ if (elem == -1) {
+ val->constant = *c;
+ } else {
+ unsigned num_components = glsl_get_vector_elements(type);
+ for (unsigned i = 0; i < num_components; i++)
+ val->constant->value.u[i] = (*c)->value.u[elem + i];
+ }
+ } else {
+ struct vtn_value *insert =
+ vtn_value(b, w[4], vtn_value_type_constant);
+ assert(insert->const_type == type);
+ if (elem == -1) {
+ *c = insert->constant;
+ } else {
+ unsigned num_components = glsl_get_vector_elements(type);
+ for (unsigned i = 0; i < num_components; i++)
+ (*c)->value.u[elem + i] = insert->constant->value.u[i];
+ }
+ }
+ return;
+ }
+
+ default: {
+ bool swap;
+ nir_op op = vtn_nir_alu_op_for_spirv_opcode(opcode, &swap);
+
+ unsigned num_components = glsl_get_vector_elements(val->const_type);
+ unsigned bit_size =
+ glsl_get_bit_size(glsl_get_base_type(val->const_type));
+
+ nir_const_value src[3];
+ assert(count <= 7);
+ for (unsigned i = 0; i < count - 4; i++) {
+ nir_constant *c =
+ vtn_value(b, w[4 + i], vtn_value_type_constant)->constant;
+
+ unsigned j = swap ? 1 - i : i;
+ assert(bit_size == 32);
+ for (unsigned k = 0; k < num_components; k++)
+ src[j].u32[k] = c->value.u[k];
+ }
+
+ nir_const_value res = nir_eval_const_opcode(op, num_components,
+ bit_size, src);
+
+ for (unsigned k = 0; k < num_components; k++)
+ val->constant->value.u[k] = res.u32[k];
+
+ return;
+ } /* default */
+ }
+ }
+
+ case SpvOpConstantNull:
+ val->constant = vtn_null_constant(b, val->const_type);
+ break;
+
+ case SpvOpConstantSampler:
+ assert(!"OpConstantSampler requires Kernel Capability");
+ break;
+
+ default:
+ unreachable("Unhandled opcode");
+ }
+}
+
+static void
+vtn_handle_function_call(struct vtn_builder *b, SpvOp opcode,
+ const uint32_t *w, unsigned count)
+{
+ struct nir_function *callee =
+ vtn_value(b, w[3], vtn_value_type_function)->func->impl->function;
+
+ nir_call_instr *call = nir_call_instr_create(b->nb.shader, callee);
+ for (unsigned i = 0; i < call->num_params; i++) {
+ unsigned arg_id = w[4 + i];
+ struct vtn_value *arg = vtn_untyped_value(b, arg_id);
+ if (arg->value_type == vtn_value_type_access_chain) {
+ nir_deref_var *d = vtn_access_chain_to_deref(b, arg->access_chain);
+ call->params[i] = nir_deref_as_var(nir_copy_deref(call, &d->deref));
+ } else {
+ struct vtn_ssa_value *arg_ssa = vtn_ssa_value(b, arg_id);
+
+ /* Make a temporary to store the argument in */
+ nir_variable *tmp =
+ nir_local_variable_create(b->impl, arg_ssa->type, "arg_tmp");
+ call->params[i] = nir_deref_var_create(call, tmp);
+
+ vtn_local_store(b, arg_ssa, call->params[i]);
+ }
+ }
+
+ nir_variable *out_tmp = NULL;
+ if (!glsl_type_is_void(callee->return_type)) {
+ out_tmp = nir_local_variable_create(b->impl, callee->return_type,
+ "out_tmp");
+ call->return_deref = nir_deref_var_create(call, out_tmp);
+ }
+
+ nir_builder_instr_insert(&b->nb, &call->instr);
+
+ if (glsl_type_is_void(callee->return_type)) {
+ vtn_push_value(b, w[2], vtn_value_type_undef);
+ } else {
+ struct vtn_value *retval = vtn_push_value(b, w[2], vtn_value_type_ssa);
+ retval->ssa = vtn_local_load(b, call->return_deref);
+ }
+}
+
+struct vtn_ssa_value *
+vtn_create_ssa_value(struct vtn_builder *b, const struct glsl_type *type)
+{
+ struct vtn_ssa_value *val = rzalloc(b, struct vtn_ssa_value);
+ val->type = type;
+
+ if (!glsl_type_is_vector_or_scalar(type)) {
+ unsigned elems = glsl_get_length(type);
+ val->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
+ for (unsigned i = 0; i < elems; i++) {
+ const struct glsl_type *child_type;
+
+ switch (glsl_get_base_type(type)) {
+ case GLSL_TYPE_INT:
+ case GLSL_TYPE_UINT:
+ case GLSL_TYPE_BOOL:
+ case GLSL_TYPE_FLOAT:
+ case GLSL_TYPE_DOUBLE:
+ child_type = glsl_get_column_type(type);
+ break;
+ case GLSL_TYPE_ARRAY:
+ child_type = glsl_get_array_element(type);
+ break;
+ case GLSL_TYPE_STRUCT:
+ child_type = glsl_get_struct_field(type, i);
+ break;
+ default:
+ unreachable("unkown base type");
+ }
+
+ val->elems[i] = vtn_create_ssa_value(b, child_type);
+ }
+ }
+
+ return val;
+}
+
+static nir_tex_src
+vtn_tex_src(struct vtn_builder *b, unsigned index, nir_tex_src_type type)
+{
+ nir_tex_src src;
+ src.src = nir_src_for_ssa(vtn_ssa_value(b, index)->def);
+ src.src_type = type;
+ return src;
+}
+
+static void
+vtn_handle_texture(struct vtn_builder *b, SpvOp opcode,
+ const uint32_t *w, unsigned count)
+{
+ if (opcode == SpvOpSampledImage) {
+ struct vtn_value *val =
+ vtn_push_value(b, w[2], vtn_value_type_sampled_image);
+ val->sampled_image = ralloc(b, struct vtn_sampled_image);
+ val->sampled_image->image =
+ vtn_value(b, w[3], vtn_value_type_access_chain)->access_chain;
+ val->sampled_image->sampler =
+ vtn_value(b, w[4], vtn_value_type_access_chain)->access_chain;
+ return;
+ } else if (opcode == SpvOpImage) {
+ struct vtn_value *val =
+ vtn_push_value(b, w[2], vtn_value_type_access_chain);
+ struct vtn_value *src_val = vtn_untyped_value(b, w[3]);
+ if (src_val->value_type == vtn_value_type_sampled_image) {
+ val->access_chain = src_val->sampled_image->image;
+ } else {
+ assert(src_val->value_type == vtn_value_type_access_chain);
+ val->access_chain = src_val->access_chain;
+ }
+ return;
+ }
+
+ struct vtn_type *ret_type = vtn_value(b, w[1], vtn_value_type_type)->type;
+ struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
+
+ struct vtn_sampled_image sampled;
+ struct vtn_value *sampled_val = vtn_untyped_value(b, w[3]);
+ if (sampled_val->value_type == vtn_value_type_sampled_image) {
+ sampled = *sampled_val->sampled_image;
+ } else {
+ assert(sampled_val->value_type == vtn_value_type_access_chain);
+ sampled.image = NULL;
+ sampled.sampler = sampled_val->access_chain;
+ }
+
+ const struct glsl_type *image_type;
+ if (sampled.image) {
+ image_type = sampled.image->var->var->interface_type;
+ } else {
+ image_type = sampled.sampler->var->var->interface_type;
+ }
+
+ nir_tex_src srcs[8]; /* 8 should be enough */
+ nir_tex_src *p = srcs;
+
+ unsigned idx = 4;
+
+ bool has_coord = false;
+ switch (opcode) {
+ case SpvOpImageSampleImplicitLod:
+ case SpvOpImageSampleExplicitLod:
+ case SpvOpImageSampleDrefImplicitLod:
+ case SpvOpImageSampleDrefExplicitLod:
+ case SpvOpImageSampleProjImplicitLod:
+ case SpvOpImageSampleProjExplicitLod:
+ case SpvOpImageSampleProjDrefImplicitLod:
+ case SpvOpImageSampleProjDrefExplicitLod:
+ case SpvOpImageFetch:
+ case SpvOpImageGather:
+ case SpvOpImageDrefGather:
+ case SpvOpImageQueryLod: {
+ /* All these types have the coordinate as their first real argument */
+ struct vtn_ssa_value *coord = vtn_ssa_value(b, w[idx++]);
+ has_coord = true;
+ p->src = nir_src_for_ssa(coord->def);
+ p->src_type = nir_tex_src_coord;
+ p++;
+ break;
+ }
+
+ default:
+ break;
+ }
+
+ /* These all have an explicit depth value as their next source */
+ switch (opcode) {
+ case SpvOpImageSampleDrefImplicitLod:
+ case SpvOpImageSampleDrefExplicitLod:
+ case SpvOpImageSampleProjDrefImplicitLod:
+ case SpvOpImageSampleProjDrefExplicitLod:
+ (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_comparitor);
+ break;
+ default:
+ break;
+ }
+
+ /* For OpImageQuerySizeLod, we always have an LOD */
+ if (opcode == SpvOpImageQuerySizeLod)
+ (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_lod);
+
+ /* Figure out the base texture operation */
+ nir_texop texop;
+ switch (opcode) {
+ case SpvOpImageSampleImplicitLod:
+ case SpvOpImageSampleDrefImplicitLod:
+ case SpvOpImageSampleProjImplicitLod:
+ case SpvOpImageSampleProjDrefImplicitLod:
+ texop = nir_texop_tex;
+ break;
+
+ case SpvOpImageSampleExplicitLod:
+ case SpvOpImageSampleDrefExplicitLod:
+ case SpvOpImageSampleProjExplicitLod:
+ case SpvOpImageSampleProjDrefExplicitLod:
+ texop = nir_texop_txl;
+ break;
+
+ case SpvOpImageFetch:
+ if (glsl_get_sampler_dim(image_type) == GLSL_SAMPLER_DIM_MS) {
+ texop = nir_texop_txf_ms;
+ } else {
+ texop = nir_texop_txf;
+ }
+ break;
+
+ case SpvOpImageGather:
+ case SpvOpImageDrefGather:
+ texop = nir_texop_tg4;
+ break;
+
+ case SpvOpImageQuerySizeLod:
+ case SpvOpImageQuerySize:
+ texop = nir_texop_txs;
+ break;
+
+ case SpvOpImageQueryLod:
+ texop = nir_texop_lod;
+ break;
+
+ case SpvOpImageQueryLevels:
+ texop = nir_texop_query_levels;
+ break;
+
+ case SpvOpImageQuerySamples:
+ default:
+ unreachable("Unhandled opcode");
+ }
+
+ /* Now we need to handle some number of optional arguments */
+ if (idx < count) {
+ uint32_t operands = w[idx++];
+
+ if (operands & SpvImageOperandsBiasMask) {
+ assert(texop == nir_texop_tex);
+ texop = nir_texop_txb;
+ (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_bias);
+ }
+
+ if (operands & SpvImageOperandsLodMask) {
+ assert(texop == nir_texop_txl || texop == nir_texop_txf ||
+ texop == nir_texop_txf_ms || texop == nir_texop_txs);
+ (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_lod);
+ }
+
+ if (operands & SpvImageOperandsGradMask) {
+ assert(texop == nir_texop_tex);
+ texop = nir_texop_txd;
+ (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_ddx);
+ (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_ddy);
+ }
+
+ if (operands & SpvImageOperandsOffsetMask ||
+ operands & SpvImageOperandsConstOffsetMask)
+ (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_offset);
+
+ if (operands & SpvImageOperandsConstOffsetsMask)
+ assert(!"Constant offsets to texture gather not yet implemented");
+
+ if (operands & SpvImageOperandsSampleMask) {
+ assert(texop == nir_texop_txf_ms);
+ texop = nir_texop_txf_ms;
+ (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_ms_index);
+ }
+ }
+ /* We should have now consumed exactly all of the arguments */
+ assert(idx == count);
+
+ nir_tex_instr *instr = nir_tex_instr_create(b->shader, p - srcs);
+ instr->op = texop;
+
+ memcpy(instr->src, srcs, instr->num_srcs * sizeof(*instr->src));
+
+ instr->sampler_dim = glsl_get_sampler_dim(image_type);
+ instr->is_array = glsl_sampler_type_is_array(image_type);
+ instr->is_shadow = glsl_sampler_type_is_shadow(image_type);
+ instr->is_new_style_shadow = instr->is_shadow;
+
+ if (has_coord) {
+ switch (instr->sampler_dim) {
+ case GLSL_SAMPLER_DIM_1D:
+ case GLSL_SAMPLER_DIM_BUF:
+ instr->coord_components = 1;
+ break;
+ case GLSL_SAMPLER_DIM_2D:
+ case GLSL_SAMPLER_DIM_RECT:
+ case GLSL_SAMPLER_DIM_MS:
+ instr->coord_components = 2;
+ break;
+ case GLSL_SAMPLER_DIM_3D:
+ case GLSL_SAMPLER_DIM_CUBE:
+ instr->coord_components = 3;
+ break;
+ default:
+ assert("Invalid sampler type");
+ }
+
+ if (instr->is_array)
+ instr->coord_components++;
+ } else {
+ instr->coord_components = 0;
+ }
+
+ switch (glsl_get_sampler_result_type(image_type)) {
+ case GLSL_TYPE_FLOAT: instr->dest_type = nir_type_float; break;
+ case GLSL_TYPE_INT: instr->dest_type = nir_type_int; break;
+ case GLSL_TYPE_UINT: instr->dest_type = nir_type_uint; break;
+ case GLSL_TYPE_BOOL: instr->dest_type = nir_type_bool; break;
+ default:
+ unreachable("Invalid base type for sampler result");
+ }
+
+ nir_deref_var *sampler = vtn_access_chain_to_deref(b, sampled.sampler);
+ if (sampled.image) {
+ nir_deref_var *image = vtn_access_chain_to_deref(b, sampled.image);
+ instr->texture = nir_deref_as_var(nir_copy_deref(instr, &image->deref));
+ } else {
+ instr->texture = nir_deref_as_var(nir_copy_deref(instr, &sampler->deref));
+ }
+
+ switch (instr->op) {
+ case nir_texop_tex:
+ case nir_texop_txb:
+ case nir_texop_txl:
+ case nir_texop_txd:
+ /* These operations require a sampler */
+ instr->sampler = nir_deref_as_var(nir_copy_deref(instr, &sampler->deref));
+ break;
+ case nir_texop_txf:
+ case nir_texop_txf_ms:
+ case nir_texop_txs:
+ case nir_texop_lod:
+ case nir_texop_tg4:
+ case nir_texop_query_levels:
+ case nir_texop_texture_samples:
+ case nir_texop_samples_identical:
+ /* These don't */
+ instr->sampler = NULL;
+ break;
+ }
+
+ nir_ssa_dest_init(&instr->instr, &instr->dest,
+ nir_tex_instr_dest_size(instr), 32, NULL);
+
+ assert(glsl_get_vector_elements(ret_type->type) ==
+ nir_tex_instr_dest_size(instr));
+
+ val->ssa = vtn_create_ssa_value(b, ret_type->type);
+ val->ssa->def = &instr->dest.ssa;
+
+ nir_builder_instr_insert(&b->nb, &instr->instr);
+}
+
+static nir_ssa_def *
+get_image_coord(struct vtn_builder *b, uint32_t value)
+{
+ struct vtn_ssa_value *coord = vtn_ssa_value(b, value);
+
+ /* The image_load_store intrinsics assume a 4-dim coordinate */
+ unsigned dim = glsl_get_vector_elements(coord->type);
+ unsigned swizzle[4];
+ for (unsigned i = 0; i < 4; i++)
+ swizzle[i] = MIN2(i, dim - 1);
+
+ return nir_swizzle(&b->nb, coord->def, swizzle, 4, false);
+}
+
+static void
+vtn_handle_image(struct vtn_builder *b, SpvOp opcode,
+ const uint32_t *w, unsigned count)
+{
+ /* Just get this one out of the way */
+ if (opcode == SpvOpImageTexelPointer) {
+ struct vtn_value *val =
+ vtn_push_value(b, w[2], vtn_value_type_image_pointer);
+ val->image = ralloc(b, struct vtn_image_pointer);
+
+ val->image->image =
+ vtn_value(b, w[3], vtn_value_type_access_chain)->access_chain;
+ val->image->coord = get_image_coord(b, w[4]);
+ val->image->sample = vtn_ssa_value(b, w[5])->def;
+ return;
+ }
+
+ struct vtn_image_pointer image;
+
+ switch (opcode) {
+ case SpvOpAtomicExchange:
+ case SpvOpAtomicCompareExchange:
+ case SpvOpAtomicCompareExchangeWeak:
+ case SpvOpAtomicIIncrement:
+ case SpvOpAtomicIDecrement:
+ case SpvOpAtomicIAdd:
+ case SpvOpAtomicISub:
+ case SpvOpAtomicSMin:
+ case SpvOpAtomicUMin:
+ case SpvOpAtomicSMax:
+ case SpvOpAtomicUMax:
+ case SpvOpAtomicAnd:
+ case SpvOpAtomicOr:
+ case SpvOpAtomicXor:
+ image = *vtn_value(b, w[3], vtn_value_type_image_pointer)->image;
+ break;
+
+ case SpvOpImageQuerySize:
+ image.image =
+ vtn_value(b, w[3], vtn_value_type_access_chain)->access_chain;
+ image.coord = NULL;
+ image.sample = NULL;
+ break;
+
+ case SpvOpImageRead:
+ image.image =
+ vtn_value(b, w[3], vtn_value_type_access_chain)->access_chain;
+ image.coord = get_image_coord(b, w[4]);
+
+ if (count > 5 && (w[5] & SpvImageOperandsSampleMask)) {
+ assert(w[5] == SpvImageOperandsSampleMask);
+ image.sample = vtn_ssa_value(b, w[6])->def;
+ } else {
+ image.sample = nir_ssa_undef(&b->nb, 1, 32);
+ }
+ break;
+
+ case SpvOpImageWrite:
+ image.image =
+ vtn_value(b, w[1], vtn_value_type_access_chain)->access_chain;
+ image.coord = get_image_coord(b, w[2]);
+
+ /* texel = w[3] */
+
+ if (count > 4 && (w[4] & SpvImageOperandsSampleMask)) {
+ assert(w[4] == SpvImageOperandsSampleMask);
+ image.sample = vtn_ssa_value(b, w[5])->def;
+ } else {
+ image.sample = nir_ssa_undef(&b->nb, 1, 32);
+ }
+ break;
+
+ default:
+ unreachable("Invalid image opcode");
+ }
+
+ nir_intrinsic_op op;
+ switch (opcode) {
+#define OP(S, N) case SpvOp##S: op = nir_intrinsic_image_##N; break;
+ OP(ImageQuerySize, size)
+ OP(ImageRead, load)
+ OP(ImageWrite, store)
+ OP(AtomicExchange, atomic_exchange)
+ OP(AtomicCompareExchange, atomic_comp_swap)
+ OP(AtomicIIncrement, atomic_add)
+ OP(AtomicIDecrement, atomic_add)
+ OP(AtomicIAdd, atomic_add)
+ OP(AtomicISub, atomic_add)
+ OP(AtomicSMin, atomic_min)
+ OP(AtomicUMin, atomic_min)
+ OP(AtomicSMax, atomic_max)
+ OP(AtomicUMax, atomic_max)
+ OP(AtomicAnd, atomic_and)
+ OP(AtomicOr, atomic_or)
+ OP(AtomicXor, atomic_xor)
+#undef OP
+ default:
+ unreachable("Invalid image opcode");
+ }
+
+ nir_intrinsic_instr *intrin = nir_intrinsic_instr_create(b->shader, op);
+
+ nir_deref_var *image_deref = vtn_access_chain_to_deref(b, image.image);
+ intrin->variables[0] =
+ nir_deref_as_var(nir_copy_deref(&intrin->instr, &image_deref->deref));
+
+ /* ImageQuerySize doesn't take any extra parameters */
+ if (opcode != SpvOpImageQuerySize) {
+ /* The image coordinate is always 4 components but we may not have that
+ * many. Swizzle to compensate.
+ */
+ unsigned swiz[4];
+ for (unsigned i = 0; i < 4; i++)
+ swiz[i] = i < image.coord->num_components ? i : 0;
+ intrin->src[0] = nir_src_for_ssa(nir_swizzle(&b->nb, image.coord,
+ swiz, 4, false));
+ intrin->src[1] = nir_src_for_ssa(image.sample);
+ }
+
+ switch (opcode) {
+ case SpvOpImageQuerySize:
+ case SpvOpImageRead:
+ break;
+ case SpvOpImageWrite:
+ intrin->src[2] = nir_src_for_ssa(vtn_ssa_value(b, w[3])->def);
+ break;
+ case SpvOpAtomicIIncrement:
+ intrin->src[2] = nir_src_for_ssa(nir_imm_int(&b->nb, 1));
+ break;
+ case SpvOpAtomicIDecrement:
+ intrin->src[2] = nir_src_for_ssa(nir_imm_int(&b->nb, -1));
+ break;
+
+ case SpvOpAtomicExchange:
+ case SpvOpAtomicIAdd:
+ case SpvOpAtomicSMin:
+ case SpvOpAtomicUMin:
+ case SpvOpAtomicSMax:
+ case SpvOpAtomicUMax:
+ case SpvOpAtomicAnd:
+ case SpvOpAtomicOr:
+ case SpvOpAtomicXor:
+ intrin->src[2] = nir_src_for_ssa(vtn_ssa_value(b, w[6])->def);
+ break;
+
+ case SpvOpAtomicCompareExchange:
+ intrin->src[2] = nir_src_for_ssa(vtn_ssa_value(b, w[7])->def);
+ intrin->src[3] = nir_src_for_ssa(vtn_ssa_value(b, w[6])->def);
+ break;
+
+ case SpvOpAtomicISub:
+ intrin->src[2] = nir_src_for_ssa(nir_ineg(&b->nb, vtn_ssa_value(b, w[6])->def));
+ break;
+
+ default:
+ unreachable("Invalid image opcode");
+ }
+
+ if (opcode != SpvOpImageWrite) {
+ struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
+ struct vtn_type *type = vtn_value(b, w[1], vtn_value_type_type)->type;
+ nir_ssa_dest_init(&intrin->instr, &intrin->dest, 4, 32, NULL);
+
+ nir_builder_instr_insert(&b->nb, &intrin->instr);
+
+ /* The image intrinsics always return 4 channels but we may not want
+ * that many. Emit a mov to trim it down.
+ */
+ unsigned swiz[4] = {0, 1, 2, 3};
+ val->ssa = vtn_create_ssa_value(b, type->type);
+ val->ssa->def = nir_swizzle(&b->nb, &intrin->dest.ssa, swiz,
+ glsl_get_vector_elements(type->type), false);
+ } else {
+ nir_builder_instr_insert(&b->nb, &intrin->instr);
+ }
+}
+
+static nir_intrinsic_op
+get_ssbo_nir_atomic_op(SpvOp opcode)
+{
+ switch (opcode) {
+#define OP(S, N) case SpvOp##S: return nir_intrinsic_ssbo_##N;
+ OP(AtomicExchange, atomic_exchange)
+ OP(AtomicCompareExchange, atomic_comp_swap)
+ OP(AtomicIIncrement, atomic_add)
+ OP(AtomicIDecrement, atomic_add)
+ OP(AtomicIAdd, atomic_add)
+ OP(AtomicISub, atomic_add)
+ OP(AtomicSMin, atomic_imin)
+ OP(AtomicUMin, atomic_umin)
+ OP(AtomicSMax, atomic_imax)
+ OP(AtomicUMax, atomic_umax)
+ OP(AtomicAnd, atomic_and)
+ OP(AtomicOr, atomic_or)
+ OP(AtomicXor, atomic_xor)
+#undef OP
+ default:
+ unreachable("Invalid SSBO atomic");
+ }
+}
+
+static nir_intrinsic_op
+get_shared_nir_atomic_op(SpvOp opcode)
+{
+ switch (opcode) {
+#define OP(S, N) case SpvOp##S: return nir_intrinsic_var_##N;
+ OP(AtomicExchange, atomic_exchange)
+ OP(AtomicCompareExchange, atomic_comp_swap)
+ OP(AtomicIIncrement, atomic_add)
+ OP(AtomicIDecrement, atomic_add)
+ OP(AtomicIAdd, atomic_add)
+ OP(AtomicISub, atomic_add)
+ OP(AtomicSMin, atomic_imin)
+ OP(AtomicUMin, atomic_umin)
+ OP(AtomicSMax, atomic_imax)
+ OP(AtomicUMax, atomic_umax)
+ OP(AtomicAnd, atomic_and)
+ OP(AtomicOr, atomic_or)
+ OP(AtomicXor, atomic_xor)
+#undef OP
+ default:
+ unreachable("Invalid shared atomic");
+ }
+}
+
+static void
+fill_common_atomic_sources(struct vtn_builder *b, SpvOp opcode,
+ const uint32_t *w, nir_src *src)
+{
+ switch (opcode) {
+ case SpvOpAtomicIIncrement:
+ src[0] = nir_src_for_ssa(nir_imm_int(&b->nb, 1));
+ break;
+
+ case SpvOpAtomicIDecrement:
+ src[0] = nir_src_for_ssa(nir_imm_int(&b->nb, -1));
+ break;
+
+ case SpvOpAtomicISub:
+ src[0] =
+ nir_src_for_ssa(nir_ineg(&b->nb, vtn_ssa_value(b, w[6])->def));
+ break;
+
+ case SpvOpAtomicCompareExchange:
+ src[0] = nir_src_for_ssa(vtn_ssa_value(b, w[7])->def);
+ src[1] = nir_src_for_ssa(vtn_ssa_value(b, w[8])->def);
+ break;
+ /* Fall through */
+
+ case SpvOpAtomicExchange:
+ case SpvOpAtomicIAdd:
+ case SpvOpAtomicSMin:
+ case SpvOpAtomicUMin:
+ case SpvOpAtomicSMax:
+ case SpvOpAtomicUMax:
+ case SpvOpAtomicAnd:
+ case SpvOpAtomicOr:
+ case SpvOpAtomicXor:
+ src[0] = nir_src_for_ssa(vtn_ssa_value(b, w[6])->def);
+ break;
+
+ default:
+ unreachable("Invalid SPIR-V atomic");
+ }
+}
+
+static void
+vtn_handle_ssbo_or_shared_atomic(struct vtn_builder *b, SpvOp opcode,
+ const uint32_t *w, unsigned count)
+{
+ struct vtn_access_chain *chain =
+ vtn_value(b, w[3], vtn_value_type_access_chain)->access_chain;
+ nir_intrinsic_instr *atomic;
+
+ /*
+ SpvScope scope = w[4];
+ SpvMemorySemanticsMask semantics = w[5];
+ */
+
+ if (chain->var->mode == vtn_variable_mode_workgroup) {
+ nir_deref *deref = &vtn_access_chain_to_deref(b, chain)->deref;
+ nir_intrinsic_op op = get_shared_nir_atomic_op(opcode);
+ atomic = nir_intrinsic_instr_create(b->nb.shader, op);
+ atomic->variables[0] = nir_deref_as_var(nir_copy_deref(atomic, deref));
+ fill_common_atomic_sources(b, opcode, w, &atomic->src[0]);
+ } else {
+ assert(chain->var->mode == vtn_variable_mode_ssbo);
+ struct vtn_type *type;
+ nir_ssa_def *offset, *index;
+ offset = vtn_access_chain_to_offset(b, chain, &index, &type, NULL, false);
+
+ nir_intrinsic_op op = get_ssbo_nir_atomic_op(opcode);
+
+ atomic = nir_intrinsic_instr_create(b->nb.shader, op);
+ atomic->src[0] = nir_src_for_ssa(index);
+ atomic->src[1] = nir_src_for_ssa(offset);
+ fill_common_atomic_sources(b, opcode, w, &atomic->src[2]);
+ }
+
+ nir_ssa_dest_init(&atomic->instr, &atomic->dest, 1, 32, NULL);
+
+ struct vtn_type *type = vtn_value(b, w[1], vtn_value_type_type)->type;
+ struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
+ val->ssa = rzalloc(b, struct vtn_ssa_value);
+ val->ssa->def = &atomic->dest.ssa;
+ val->ssa->type = type->type;
+
+ nir_builder_instr_insert(&b->nb, &atomic->instr);
+}
+
+static nir_alu_instr *
+create_vec(nir_shader *shader, unsigned num_components, unsigned bit_size)
+{
+ nir_op op;
+ switch (num_components) {
+ case 1: op = nir_op_fmov; break;
+ case 2: op = nir_op_vec2; break;
+ case 3: op = nir_op_vec3; break;
+ case 4: op = nir_op_vec4; break;
+ default: unreachable("bad vector size");
+ }
+
+ nir_alu_instr *vec = nir_alu_instr_create(shader, op);
+ nir_ssa_dest_init(&vec->instr, &vec->dest.dest, num_components,
+ bit_size, NULL);
+ vec->dest.write_mask = (1 << num_components) - 1;
+
+ return vec;
+}
+
+struct vtn_ssa_value *
+vtn_ssa_transpose(struct vtn_builder *b, struct vtn_ssa_value *src)
+{
+ if (src->transposed)
+ return src->transposed;
+
+ struct vtn_ssa_value *dest =
+ vtn_create_ssa_value(b, glsl_transposed_type(src->type));
+
+ for (unsigned i = 0; i < glsl_get_matrix_columns(dest->type); i++) {
+ nir_alu_instr *vec = create_vec(b->shader,
+ glsl_get_matrix_columns(src->type),
+ glsl_get_bit_size(glsl_get_base_type(src->type)));
+ if (glsl_type_is_vector_or_scalar(src->type)) {
+ vec->src[0].src = nir_src_for_ssa(src->def);
+ vec->src[0].swizzle[0] = i;
+ } else {
+ for (unsigned j = 0; j < glsl_get_matrix_columns(src->type); j++) {
+ vec->src[j].src = nir_src_for_ssa(src->elems[j]->def);
+ vec->src[j].swizzle[0] = i;
+ }
+ }
+ nir_builder_instr_insert(&b->nb, &vec->instr);
+ dest->elems[i]->def = &vec->dest.dest.ssa;
+ }
+
+ dest->transposed = src;
+
+ return dest;
+}
+
+nir_ssa_def *
+vtn_vector_extract(struct vtn_builder *b, nir_ssa_def *src, unsigned index)
+{
+ unsigned swiz[4] = { index };
+ return nir_swizzle(&b->nb, src, swiz, 1, true);
+}
+
+nir_ssa_def *
+vtn_vector_insert(struct vtn_builder *b, nir_ssa_def *src, nir_ssa_def *insert,
+ unsigned index)
+{
+ nir_alu_instr *vec = create_vec(b->shader, src->num_components,
+ src->bit_size);
+
+ for (unsigned i = 0; i < src->num_components; i++) {
+ if (i == index) {
+ vec->src[i].src = nir_src_for_ssa(insert);
+ } else {
+ vec->src[i].src = nir_src_for_ssa(src);
+ vec->src[i].swizzle[0] = i;
+ }
+ }
+
+ nir_builder_instr_insert(&b->nb, &vec->instr);
+
+ return &vec->dest.dest.ssa;
+}
+
+nir_ssa_def *
+vtn_vector_extract_dynamic(struct vtn_builder *b, nir_ssa_def *src,
+ nir_ssa_def *index)
+{
+ nir_ssa_def *dest = vtn_vector_extract(b, src, 0);
+ for (unsigned i = 1; i < src->num_components; i++)
+ dest = nir_bcsel(&b->nb, nir_ieq(&b->nb, index, nir_imm_int(&b->nb, i)),
+ vtn_vector_extract(b, src, i), dest);
+
+ return dest;
+}
+
+nir_ssa_def *
+vtn_vector_insert_dynamic(struct vtn_builder *b, nir_ssa_def *src,
+ nir_ssa_def *insert, nir_ssa_def *index)
+{
+ nir_ssa_def *dest = vtn_vector_insert(b, src, insert, 0);
+ for (unsigned i = 1; i < src->num_components; i++)
+ dest = nir_bcsel(&b->nb, nir_ieq(&b->nb, index, nir_imm_int(&b->nb, i)),
+ vtn_vector_insert(b, src, insert, i), dest);
+
+ return dest;
+}
+
+static nir_ssa_def *
+vtn_vector_shuffle(struct vtn_builder *b, unsigned num_components,
+ nir_ssa_def *src0, nir_ssa_def *src1,
+ const uint32_t *indices)
+{
+ nir_alu_instr *vec = create_vec(b->shader, num_components, src0->bit_size);
+
+ for (unsigned i = 0; i < num_components; i++) {
+ uint32_t index = indices[i];
+ if (index == 0xffffffff) {
+ vec->src[i].src =
+ nir_src_for_ssa(nir_ssa_undef(&b->nb, 1, src0->bit_size));
+ } else if (index < src0->num_components) {
+ vec->src[i].src = nir_src_for_ssa(src0);
+ vec->src[i].swizzle[0] = index;
+ } else {
+ vec->src[i].src = nir_src_for_ssa(src1);
+ vec->src[i].swizzle[0] = index - src0->num_components;
+ }
+ }
+
+ nir_builder_instr_insert(&b->nb, &vec->instr);
+
+ return &vec->dest.dest.ssa;
+}
+
+/*
+ * Concatentates a number of vectors/scalars together to produce a vector
+ */
+static nir_ssa_def *
+vtn_vector_construct(struct vtn_builder *b, unsigned num_components,
+ unsigned num_srcs, nir_ssa_def **srcs)
+{
+ nir_alu_instr *vec = create_vec(b->shader, num_components,
+ srcs[0]->bit_size);
+
+ unsigned dest_idx = 0;
+ for (unsigned i = 0; i < num_srcs; i++) {
+ nir_ssa_def *src = srcs[i];
+ for (unsigned j = 0; j < src->num_components; j++) {
+ vec->src[dest_idx].src = nir_src_for_ssa(src);
+ vec->src[dest_idx].swizzle[0] = j;
+ dest_idx++;
+ }
+ }
+
+ nir_builder_instr_insert(&b->nb, &vec->instr);
+
+ return &vec->dest.dest.ssa;
+}
+
+static struct vtn_ssa_value *
+vtn_composite_copy(void *mem_ctx, struct vtn_ssa_value *src)
+{
+ struct vtn_ssa_value *dest = rzalloc(mem_ctx, struct vtn_ssa_value);
+ dest->type = src->type;
+
+ if (glsl_type_is_vector_or_scalar(src->type)) {
+ dest->def = src->def;
+ } else {
+ unsigned elems = glsl_get_length(src->type);
+
+ dest->elems = ralloc_array(mem_ctx, struct vtn_ssa_value *, elems);
+ for (unsigned i = 0; i < elems; i++)
+ dest->elems[i] = vtn_composite_copy(mem_ctx, src->elems[i]);
+ }
+
+ return dest;
+}
+
+static struct vtn_ssa_value *
+vtn_composite_insert(struct vtn_builder *b, struct vtn_ssa_value *src,
+ struct vtn_ssa_value *insert, const uint32_t *indices,
+ unsigned num_indices)
+{
+ struct vtn_ssa_value *dest = vtn_composite_copy(b, src);
+
+ struct vtn_ssa_value *cur = dest;
+ unsigned i;
+ for (i = 0; i < num_indices - 1; i++) {
+ cur = cur->elems[indices[i]];
+ }
+
+ if (glsl_type_is_vector_or_scalar(cur->type)) {
+ /* According to the SPIR-V spec, OpCompositeInsert may work down to
+ * the component granularity. In that case, the last index will be
+ * the index to insert the scalar into the vector.
+ */
+
+ cur->def = vtn_vector_insert(b, cur->def, insert->def, indices[i]);
+ } else {
+ cur->elems[indices[i]] = insert;
+ }
+
+ return dest;
+}
+
+static struct vtn_ssa_value *
+vtn_composite_extract(struct vtn_builder *b, struct vtn_ssa_value *src,
+ const uint32_t *indices, unsigned num_indices)
+{
+ struct vtn_ssa_value *cur = src;
+ for (unsigned i = 0; i < num_indices; i++) {
+ if (glsl_type_is_vector_or_scalar(cur->type)) {
+ assert(i == num_indices - 1);
+ /* According to the SPIR-V spec, OpCompositeExtract may work down to
+ * the component granularity. The last index will be the index of the
+ * vector to extract.
+ */
+
+ struct vtn_ssa_value *ret = rzalloc(b, struct vtn_ssa_value);
+ ret->type = glsl_scalar_type(glsl_get_base_type(cur->type));
+ ret->def = vtn_vector_extract(b, cur->def, indices[i]);
+ return ret;
+ } else {
+ cur = cur->elems[indices[i]];
+ }
+ }
+
+ return cur;
+}
+
+static void
+vtn_handle_composite(struct vtn_builder *b, SpvOp opcode,
+ const uint32_t *w, unsigned count)
+{
+ struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
+ const struct glsl_type *type =
+ vtn_value(b, w[1], vtn_value_type_type)->type->type;
+ val->ssa = vtn_create_ssa_value(b, type);
+
+ switch (opcode) {
+ case SpvOpVectorExtractDynamic:
+ val->ssa->def = vtn_vector_extract_dynamic(b, vtn_ssa_value(b, w[3])->def,
+ vtn_ssa_value(b, w[4])->def);
+ break;
+
+ case SpvOpVectorInsertDynamic:
+ val->ssa->def = vtn_vector_insert_dynamic(b, vtn_ssa_value(b, w[3])->def,
+ vtn_ssa_value(b, w[4])->def,
+ vtn_ssa_value(b, w[5])->def);
+ break;
+
+ case SpvOpVectorShuffle:
+ val->ssa->def = vtn_vector_shuffle(b, glsl_get_vector_elements(type),
+ vtn_ssa_value(b, w[3])->def,
+ vtn_ssa_value(b, w[4])->def,
+ w + 5);
+ break;
+
+ case SpvOpCompositeConstruct: {
+ unsigned elems = count - 3;
+ if (glsl_type_is_vector_or_scalar(type)) {
+ nir_ssa_def *srcs[4];
+ for (unsigned i = 0; i < elems; i++)
+ srcs[i] = vtn_ssa_value(b, w[3 + i])->def;
+ val->ssa->def =
+ vtn_vector_construct(b, glsl_get_vector_elements(type),
+ elems, srcs);
+ } else {
+ val->ssa->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
+ for (unsigned i = 0; i < elems; i++)
+ val->ssa->elems[i] = vtn_ssa_value(b, w[3 + i]);
+ }
+ break;
+ }
+ case SpvOpCompositeExtract:
+ val->ssa = vtn_composite_extract(b, vtn_ssa_value(b, w[3]),
+ w + 4, count - 4);
+ break;
+
+ case SpvOpCompositeInsert:
+ val->ssa = vtn_composite_insert(b, vtn_ssa_value(b, w[4]),
+ vtn_ssa_value(b, w[3]),
+ w + 5, count - 5);
+ break;
+
+ case SpvOpCopyObject:
+ val->ssa = vtn_composite_copy(b, vtn_ssa_value(b, w[3]));
+ break;
+
+ default:
+ unreachable("unknown composite operation");
+ }
+}
+
+static void
+vtn_handle_barrier(struct vtn_builder *b, SpvOp opcode,
+ const uint32_t *w, unsigned count)
+{
+ nir_intrinsic_op intrinsic_op;
+ switch (opcode) {
+ case SpvOpEmitVertex:
+ case SpvOpEmitStreamVertex:
+ intrinsic_op = nir_intrinsic_emit_vertex;
+ break;
+ case SpvOpEndPrimitive:
+ case SpvOpEndStreamPrimitive:
+ intrinsic_op = nir_intrinsic_end_primitive;
+ break;
+ case SpvOpMemoryBarrier:
+ intrinsic_op = nir_intrinsic_memory_barrier;
+ break;
+ case SpvOpControlBarrier:
+ intrinsic_op = nir_intrinsic_barrier;
+ break;
+ default:
+ unreachable("unknown barrier instruction");
+ }
+
+ nir_intrinsic_instr *intrin =
+ nir_intrinsic_instr_create(b->shader, intrinsic_op);
+
+ if (opcode == SpvOpEmitStreamVertex || opcode == SpvOpEndStreamPrimitive)
+ nir_intrinsic_set_stream_id(intrin, w[1]);
+
+ nir_builder_instr_insert(&b->nb, &intrin->instr);
+}
+
+static unsigned
+gl_primitive_from_spv_execution_mode(SpvExecutionMode mode)
+{
+ switch (mode) {
+ case SpvExecutionModeInputPoints:
+ case SpvExecutionModeOutputPoints:
+ return 0; /* GL_POINTS */
+ case SpvExecutionModeInputLines:
+ return 1; /* GL_LINES */
+ case SpvExecutionModeInputLinesAdjacency:
+ return 0x000A; /* GL_LINE_STRIP_ADJACENCY_ARB */
+ case SpvExecutionModeTriangles:
+ return 4; /* GL_TRIANGLES */
+ case SpvExecutionModeInputTrianglesAdjacency:
+ return 0x000C; /* GL_TRIANGLES_ADJACENCY_ARB */
+ case SpvExecutionModeQuads:
+ return 7; /* GL_QUADS */
+ case SpvExecutionModeIsolines:
+ return 0x8E7A; /* GL_ISOLINES */
+ case SpvExecutionModeOutputLineStrip:
+ return 3; /* GL_LINE_STRIP */
+ case SpvExecutionModeOutputTriangleStrip:
+ return 5; /* GL_TRIANGLE_STRIP */
+ default:
+ assert(!"Invalid primitive type");
+ return 4;
+ }
+}
+
+static unsigned
+vertices_in_from_spv_execution_mode(SpvExecutionMode mode)
+{
+ switch (mode) {
+ case SpvExecutionModeInputPoints:
+ return 1;
+ case SpvExecutionModeInputLines:
+ return 2;
+ case SpvExecutionModeInputLinesAdjacency:
+ return 4;
+ case SpvExecutionModeTriangles:
+ return 3;
+ case SpvExecutionModeInputTrianglesAdjacency:
+ return 6;
+ default:
+ assert(!"Invalid GS input mode");
+ return 0;
+ }
+}
+
+static gl_shader_stage
+stage_for_execution_model(SpvExecutionModel model)
+{
+ switch (model) {
+ case SpvExecutionModelVertex:
+ return MESA_SHADER_VERTEX;
+ case SpvExecutionModelTessellationControl:
+ return MESA_SHADER_TESS_CTRL;
+ case SpvExecutionModelTessellationEvaluation:
+ return MESA_SHADER_TESS_EVAL;
+ case SpvExecutionModelGeometry:
+ return MESA_SHADER_GEOMETRY;
+ case SpvExecutionModelFragment:
+ return MESA_SHADER_FRAGMENT;
+ case SpvExecutionModelGLCompute:
+ return MESA_SHADER_COMPUTE;
+ default:
+ unreachable("Unsupported execution model");
+ }
+}
+
+static bool
+vtn_handle_preamble_instruction(struct vtn_builder *b, SpvOp opcode,
+ const uint32_t *w, unsigned count)
+{
+ switch (opcode) {
+ case SpvOpSource:
+ case SpvOpSourceExtension:
+ case SpvOpSourceContinued:
+ case SpvOpExtension:
+ /* Unhandled, but these are for debug so that's ok. */
+ break;
+
+ case SpvOpCapability: {
+ SpvCapability cap = w[1];
+ switch (cap) {
+ case SpvCapabilityMatrix:
+ case SpvCapabilityShader:
+ case SpvCapabilityGeometry:
+ case SpvCapabilityTessellationPointSize:
+ case SpvCapabilityGeometryPointSize:
+ case SpvCapabilityUniformBufferArrayDynamicIndexing:
+ case SpvCapabilitySampledImageArrayDynamicIndexing:
+ case SpvCapabilityStorageBufferArrayDynamicIndexing:
+ case SpvCapabilityStorageImageArrayDynamicIndexing:
+ case SpvCapabilityImageRect:
+ case SpvCapabilitySampledRect:
+ case SpvCapabilitySampled1D:
+ case SpvCapabilityImage1D:
+ case SpvCapabilitySampledCubeArray:
+ case SpvCapabilitySampledBuffer:
+ case SpvCapabilityImageBuffer:
+ case SpvCapabilityImageQuery:
+ break;
+ case SpvCapabilityClipDistance:
+ case SpvCapabilityCullDistance:
+ case SpvCapabilityGeometryStreams:
+ fprintf(stderr, "WARNING: Unsupported SPIR-V Capability\n");
+ break;
+ default:
+ assert(!"Unsupported capability");
+ }
+ break;
+ }
+
+ case SpvOpExtInstImport:
+ vtn_handle_extension(b, opcode, w, count);
+ break;
+
+ case SpvOpMemoryModel:
+ assert(w[1] == SpvAddressingModelLogical);
+ assert(w[2] == SpvMemoryModelGLSL450);
+ break;
+
+ case SpvOpEntryPoint: {
+ struct vtn_value *entry_point = &b->values[w[2]];
+ /* Let this be a name label regardless */
+ unsigned name_words;
+ entry_point->name = vtn_string_literal(b, &w[3], count - 3, &name_words);
+
+ if (strcmp(entry_point->name, b->entry_point_name) != 0 ||
+ stage_for_execution_model(w[1]) != b->entry_point_stage)
+ break;
+
+ assert(b->entry_point == NULL);
+ b->entry_point = entry_point;
+ break;
+ }
+
+ case SpvOpString:
+ vtn_push_value(b, w[1], vtn_value_type_string)->str =
+ vtn_string_literal(b, &w[2], count - 2, NULL);
+ break;
+
+ case SpvOpName:
+ b->values[w[1]].name = vtn_string_literal(b, &w[2], count - 2, NULL);
+ break;
+
+ case SpvOpMemberName:
+ /* TODO */
+ break;
+
+ case SpvOpExecutionMode:
+ case SpvOpDecorationGroup:
+ case SpvOpDecorate:
+ case SpvOpMemberDecorate:
+ case SpvOpGroupDecorate:
+ case SpvOpGroupMemberDecorate:
+ vtn_handle_decoration(b, opcode, w, count);
+ break;
+
+ default:
+ return false; /* End of preamble */
+ }
+
+ return true;
+}
+
+static void
+vtn_handle_execution_mode(struct vtn_builder *b, struct vtn_value *entry_point,
+ const struct vtn_decoration *mode, void *data)
+{
+ assert(b->entry_point == entry_point);
+
+ switch(mode->exec_mode) {
+ case SpvExecutionModeOriginUpperLeft:
+ case SpvExecutionModeOriginLowerLeft:
+ b->origin_upper_left =
+ (mode->exec_mode == SpvExecutionModeOriginUpperLeft);
+ break;
+
+ case SpvExecutionModeEarlyFragmentTests:
+ assert(b->shader->stage == MESA_SHADER_FRAGMENT);
+ b->shader->info.fs.early_fragment_tests = true;
+ break;
+
+ case SpvExecutionModeInvocations:
+ assert(b->shader->stage == MESA_SHADER_GEOMETRY);
+ b->shader->info.gs.invocations = MAX2(1, mode->literals[0]);
+ break;
+
+ case SpvExecutionModeDepthReplacing:
+ assert(b->shader->stage == MESA_SHADER_FRAGMENT);
+ b->shader->info.fs.depth_layout = FRAG_DEPTH_LAYOUT_ANY;
+ break;
+ case SpvExecutionModeDepthGreater:
+ assert(b->shader->stage == MESA_SHADER_FRAGMENT);
+ b->shader->info.fs.depth_layout = FRAG_DEPTH_LAYOUT_GREATER;
+ break;
+ case SpvExecutionModeDepthLess:
+ assert(b->shader->stage == MESA_SHADER_FRAGMENT);
+ b->shader->info.fs.depth_layout = FRAG_DEPTH_LAYOUT_LESS;
+ break;
+ case SpvExecutionModeDepthUnchanged:
+ assert(b->shader->stage == MESA_SHADER_FRAGMENT);
+ b->shader->info.fs.depth_layout = FRAG_DEPTH_LAYOUT_UNCHANGED;
+ break;
+
+ case SpvExecutionModeLocalSize:
+ assert(b->shader->stage == MESA_SHADER_COMPUTE);
+ b->shader->info.cs.local_size[0] = mode->literals[0];
+ b->shader->info.cs.local_size[1] = mode->literals[1];
+ b->shader->info.cs.local_size[2] = mode->literals[2];
+ break;
+ case SpvExecutionModeLocalSizeHint:
+ break; /* Nothing do do with this */
+
+ case SpvExecutionModeOutputVertices:
+ assert(b->shader->stage == MESA_SHADER_GEOMETRY);
+ b->shader->info.gs.vertices_out = mode->literals[0];
+ break;
+
+ case SpvExecutionModeInputPoints:
+ case SpvExecutionModeInputLines:
+ case SpvExecutionModeInputLinesAdjacency:
+ case SpvExecutionModeTriangles:
+ case SpvExecutionModeInputTrianglesAdjacency:
+ case SpvExecutionModeQuads:
+ case SpvExecutionModeIsolines:
+ if (b->shader->stage == MESA_SHADER_GEOMETRY) {
+ b->shader->info.gs.vertices_in =
+ vertices_in_from_spv_execution_mode(mode->exec_mode);
+ } else {
+ assert(!"Tesselation shaders not yet supported");
+ }
+ break;
+
+ case SpvExecutionModeOutputPoints:
+ case SpvExecutionModeOutputLineStrip:
+ case SpvExecutionModeOutputTriangleStrip:
+ assert(b->shader->stage == MESA_SHADER_GEOMETRY);
+ b->shader->info.gs.output_primitive =
+ gl_primitive_from_spv_execution_mode(mode->exec_mode);
+ break;
+
+ case SpvExecutionModeSpacingEqual:
+ case SpvExecutionModeSpacingFractionalEven:
+ case SpvExecutionModeSpacingFractionalOdd:
+ case SpvExecutionModeVertexOrderCw:
+ case SpvExecutionModeVertexOrderCcw:
+ case SpvExecutionModePointMode:
+ assert(!"TODO: Add tessellation metadata");
+ break;
+
+ case SpvExecutionModePixelCenterInteger:
+ case SpvExecutionModeXfb:
+ assert(!"Unhandled execution mode");
+ break;
+
+ case SpvExecutionModeVecTypeHint:
+ case SpvExecutionModeContractionOff:
+ break; /* OpenCL */
+ }
+}
+
+static bool
+vtn_handle_variable_or_type_instruction(struct vtn_builder *b, SpvOp opcode,
+ const uint32_t *w, unsigned count)
+{
+ switch (opcode) {
+ case SpvOpSource:
+ case SpvOpSourceContinued:
+ case SpvOpSourceExtension:
+ case SpvOpExtension:
+ case SpvOpCapability:
+ case SpvOpExtInstImport:
+ case SpvOpMemoryModel:
+ case SpvOpEntryPoint:
+ case SpvOpExecutionMode:
+ case SpvOpString:
+ case SpvOpName:
+ case SpvOpMemberName:
+ case SpvOpDecorationGroup:
+ case SpvOpDecorate:
+ case SpvOpMemberDecorate:
+ case SpvOpGroupDecorate:
+ case SpvOpGroupMemberDecorate:
+ assert(!"Invalid opcode types and variables section");
+ break;
+
+ case SpvOpTypeVoid:
+ case SpvOpTypeBool:
+ case SpvOpTypeInt:
+ case SpvOpTypeFloat:
+ case SpvOpTypeVector:
+ case SpvOpTypeMatrix:
+ case SpvOpTypeImage:
+ case SpvOpTypeSampler:
+ case SpvOpTypeSampledImage:
+ case SpvOpTypeArray:
+ case SpvOpTypeRuntimeArray:
+ case SpvOpTypeStruct:
+ case SpvOpTypeOpaque:
+ case SpvOpTypePointer:
+ case SpvOpTypeFunction:
+ case SpvOpTypeEvent:
+ case SpvOpTypeDeviceEvent:
+ case SpvOpTypeReserveId:
+ case SpvOpTypeQueue:
+ case SpvOpTypePipe:
+ vtn_handle_type(b, opcode, w, count);
+ break;
+
+ case SpvOpConstantTrue:
+ case SpvOpConstantFalse:
+ case SpvOpConstant:
+ case SpvOpConstantComposite:
+ case SpvOpConstantSampler:
+ case SpvOpConstantNull:
+ case SpvOpSpecConstantTrue:
+ case SpvOpSpecConstantFalse:
+ case SpvOpSpecConstant:
+ case SpvOpSpecConstantComposite:
+ case SpvOpSpecConstantOp:
+ vtn_handle_constant(b, opcode, w, count);
+ break;
+
+ case SpvOpVariable:
+ vtn_handle_variables(b, opcode, w, count);
+ break;
+
+ default:
+ return false; /* End of preamble */
+ }
+
+ return true;
+}
+
+static bool
+vtn_handle_body_instruction(struct vtn_builder *b, SpvOp opcode,
+ const uint32_t *w, unsigned count)
+{
+ switch (opcode) {
+ case SpvOpLabel:
+ break;
+
+ case SpvOpLoopMerge:
+ case SpvOpSelectionMerge:
+ /* This is handled by cfg pre-pass and walk_blocks */
+ break;
+
+ case SpvOpUndef: {
+ struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_undef);
+ val->type = vtn_value(b, w[1], vtn_value_type_type)->type;
+ break;
+ }
+
+ case SpvOpExtInst:
+ vtn_handle_extension(b, opcode, w, count);
+ break;
+
+ case SpvOpVariable:
+ case SpvOpLoad:
+ case SpvOpStore:
+ case SpvOpCopyMemory:
+ case SpvOpCopyMemorySized:
+ case SpvOpAccessChain:
+ case SpvOpInBoundsAccessChain:
+ case SpvOpArrayLength:
+ vtn_handle_variables(b, opcode, w, count);
+ break;
+
+ case SpvOpFunctionCall:
+ vtn_handle_function_call(b, opcode, w, count);
+ break;
+
+ case SpvOpSampledImage:
+ case SpvOpImage:
+ case SpvOpImageSampleImplicitLod:
+ case SpvOpImageSampleExplicitLod:
+ case SpvOpImageSampleDrefImplicitLod:
+ case SpvOpImageSampleDrefExplicitLod:
+ case SpvOpImageSampleProjImplicitLod:
+ case SpvOpImageSampleProjExplicitLod:
+ case SpvOpImageSampleProjDrefImplicitLod:
+ case SpvOpImageSampleProjDrefExplicitLod:
+ case SpvOpImageFetch:
+ case SpvOpImageGather:
+ case SpvOpImageDrefGather:
+ case SpvOpImageQuerySizeLod:
+ case SpvOpImageQueryLod:
+ case SpvOpImageQueryLevels:
+ case SpvOpImageQuerySamples:
+ vtn_handle_texture(b, opcode, w, count);
+ break;
+
+ case SpvOpImageRead:
+ case SpvOpImageWrite:
+ case SpvOpImageTexelPointer:
+ vtn_handle_image(b, opcode, w, count);
+ break;
+
+ case SpvOpImageQuerySize: {
+ struct vtn_access_chain *image =
+ vtn_value(b, w[3], vtn_value_type_access_chain)->access_chain;
+ if (glsl_type_is_image(image->var->var->interface_type)) {
+ vtn_handle_image(b, opcode, w, count);
+ } else {
+ vtn_handle_texture(b, opcode, w, count);
+ }
+ break;
+ }
+
+ case SpvOpAtomicExchange:
+ case SpvOpAtomicCompareExchange:
+ case SpvOpAtomicCompareExchangeWeak:
+ case SpvOpAtomicIIncrement:
+ case SpvOpAtomicIDecrement:
+ case SpvOpAtomicIAdd:
+ case SpvOpAtomicISub:
+ case SpvOpAtomicSMin:
+ case SpvOpAtomicUMin:
+ case SpvOpAtomicSMax:
+ case SpvOpAtomicUMax:
+ case SpvOpAtomicAnd:
+ case SpvOpAtomicOr:
+ case SpvOpAtomicXor: {
+ struct vtn_value *pointer = vtn_untyped_value(b, w[3]);
+ if (pointer->value_type == vtn_value_type_image_pointer) {
+ vtn_handle_image(b, opcode, w, count);
+ } else {
+ assert(pointer->value_type == vtn_value_type_access_chain);
+ vtn_handle_ssbo_or_shared_atomic(b, opcode, w, count);
+ }
+ break;
+ }
+
+ case SpvOpSNegate:
+ case SpvOpFNegate:
+ case SpvOpNot:
+ case SpvOpAny:
+ case SpvOpAll:
+ case SpvOpConvertFToU:
+ case SpvOpConvertFToS:
+ case SpvOpConvertSToF:
+ case SpvOpConvertUToF:
+ case SpvOpUConvert:
+ case SpvOpSConvert:
+ case SpvOpFConvert:
+ case SpvOpQuantizeToF16:
+ case SpvOpConvertPtrToU:
+ case SpvOpConvertUToPtr:
+ case SpvOpPtrCastToGeneric:
+ case SpvOpGenericCastToPtr:
+ case SpvOpBitcast:
+ case SpvOpIsNan:
+ case SpvOpIsInf:
+ case SpvOpIsFinite:
+ case SpvOpIsNormal:
+ case SpvOpSignBitSet:
+ case SpvOpLessOrGreater:
+ case SpvOpOrdered:
+ case SpvOpUnordered:
+ case SpvOpIAdd:
+ case SpvOpFAdd:
+ case SpvOpISub:
+ case SpvOpFSub:
+ case SpvOpIMul:
+ case SpvOpFMul:
+ case SpvOpUDiv:
+ case SpvOpSDiv:
+ case SpvOpFDiv:
+ case SpvOpUMod:
+ case SpvOpSRem:
+ case SpvOpSMod:
+ case SpvOpFRem:
+ case SpvOpFMod:
+ case SpvOpVectorTimesScalar:
+ case SpvOpDot:
+ case SpvOpIAddCarry:
+ case SpvOpISubBorrow:
+ case SpvOpUMulExtended:
+ case SpvOpSMulExtended:
+ case SpvOpShiftRightLogical:
+ case SpvOpShiftRightArithmetic:
+ case SpvOpShiftLeftLogical:
+ case SpvOpLogicalEqual:
+ case SpvOpLogicalNotEqual:
+ case SpvOpLogicalOr:
+ case SpvOpLogicalAnd:
+ case SpvOpLogicalNot:
+ case SpvOpBitwiseOr:
+ case SpvOpBitwiseXor:
+ case SpvOpBitwiseAnd:
+ case SpvOpSelect:
+ case SpvOpIEqual:
+ case SpvOpFOrdEqual:
+ case SpvOpFUnordEqual:
+ case SpvOpINotEqual:
+ case SpvOpFOrdNotEqual:
+ case SpvOpFUnordNotEqual:
+ case SpvOpULessThan:
+ case SpvOpSLessThan:
+ case SpvOpFOrdLessThan:
+ case SpvOpFUnordLessThan:
+ case SpvOpUGreaterThan:
+ case SpvOpSGreaterThan:
+ case SpvOpFOrdGreaterThan:
+ case SpvOpFUnordGreaterThan:
+ case SpvOpULessThanEqual:
+ case SpvOpSLessThanEqual:
+ case SpvOpFOrdLessThanEqual:
+ case SpvOpFUnordLessThanEqual:
+ case SpvOpUGreaterThanEqual:
+ case SpvOpSGreaterThanEqual:
+ case SpvOpFOrdGreaterThanEqual:
+ case SpvOpFUnordGreaterThanEqual:
+ case SpvOpDPdx:
+ case SpvOpDPdy:
+ case SpvOpFwidth:
+ case SpvOpDPdxFine:
+ case SpvOpDPdyFine:
+ case SpvOpFwidthFine:
+ case SpvOpDPdxCoarse:
+ case SpvOpDPdyCoarse:
+ case SpvOpFwidthCoarse:
+ case SpvOpBitFieldInsert:
+ case SpvOpBitFieldSExtract:
+ case SpvOpBitFieldUExtract:
+ case SpvOpBitReverse:
+ case SpvOpBitCount:
+ case SpvOpTranspose:
+ case SpvOpOuterProduct:
+ case SpvOpMatrixTimesScalar:
+ case SpvOpVectorTimesMatrix:
+ case SpvOpMatrixTimesVector:
+ case SpvOpMatrixTimesMatrix:
+ vtn_handle_alu(b, opcode, w, count);
+ break;
+
+ case SpvOpVectorExtractDynamic:
+ case SpvOpVectorInsertDynamic:
+ case SpvOpVectorShuffle:
+ case SpvOpCompositeConstruct:
+ case SpvOpCompositeExtract:
+ case SpvOpCompositeInsert:
+ case SpvOpCopyObject:
+ vtn_handle_composite(b, opcode, w, count);
+ break;
+
+ case SpvOpEmitVertex:
+ case SpvOpEndPrimitive:
+ case SpvOpEmitStreamVertex:
+ case SpvOpEndStreamPrimitive:
+ case SpvOpControlBarrier:
+ case SpvOpMemoryBarrier:
+ vtn_handle_barrier(b, opcode, w, count);
+ break;
+
+ default:
+ unreachable("Unhandled opcode");
+ }
+
+ return true;
+}
+
+nir_function *
+spirv_to_nir(const uint32_t *words, size_t word_count,
+ struct nir_spirv_specialization *spec, unsigned num_spec,
+ gl_shader_stage stage, const char *entry_point_name,
+ const nir_shader_compiler_options *options)
+{
+ const uint32_t *word_end = words + word_count;
+
+ /* Handle the SPIR-V header (first 4 dwords) */
+ assert(word_count > 5);
+
+ assert(words[0] == SpvMagicNumber);
+ assert(words[1] >= 0x10000);
+ /* words[2] == generator magic */
+ unsigned value_id_bound = words[3];
+ assert(words[4] == 0);
+
+ words+= 5;
+
+ /* Initialize the stn_builder object */
+ struct vtn_builder *b = rzalloc(NULL, struct vtn_builder);
+ b->value_id_bound = value_id_bound;
+ b->values = rzalloc_array(b, struct vtn_value, value_id_bound);
+ exec_list_make_empty(&b->functions);
+ b->entry_point_stage = stage;
+ b->entry_point_name = entry_point_name;
+
+ /* Handle all the preamble instructions */
+ words = vtn_foreach_instruction(b, words, word_end,
+ vtn_handle_preamble_instruction);
+
+ if (b->entry_point == NULL) {
+ assert(!"Entry point not found");
+ ralloc_free(b);
+ return NULL;
+ }
+
+ b->shader = nir_shader_create(NULL, stage, options);
+
+ /* Set shader info defaults */
+ b->shader->info.gs.invocations = 1;
+
+ /* Parse execution modes */
+ vtn_foreach_execution_mode(b, b->entry_point,
+ vtn_handle_execution_mode, NULL);
+
+ b->specializations = spec;
+ b->num_specializations = num_spec;
+
+ /* Handle all variable, type, and constant instructions */
+ words = vtn_foreach_instruction(b, words, word_end,
+ vtn_handle_variable_or_type_instruction);
+
+ vtn_build_cfg(b, words, word_end);
+
+ foreach_list_typed(struct vtn_function, func, node, &b->functions) {
+ b->impl = func->impl;
+ b->const_table = _mesa_hash_table_create(b, _mesa_hash_pointer,
+ _mesa_key_pointer_equal);
+
+ vtn_function_emit(b, func, vtn_handle_body_instruction);
+ }
+
+ assert(b->entry_point->value_type == vtn_value_type_function);
+ nir_function *entry_point = b->entry_point->func->impl->function;
+ assert(entry_point);
+
+ ralloc_free(b);
+
+ return entry_point;
+}
diff --git a/src/compiler/spirv/vtn_alu.c b/src/compiler/spirv/vtn_alu.c
new file mode 100644
index 00000000000..8b9a63ce760
--- /dev/null
+++ b/src/compiler/spirv/vtn_alu.c
@@ -0,0 +1,464 @@
+/*
+ * Copyright © 2016 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#include "vtn_private.h"
+
+/*
+ * Normally, column vectors in SPIR-V correspond to a single NIR SSA
+ * definition. But for matrix multiplies, we want to do one routine for
+ * multiplying a matrix by a matrix and then pretend that vectors are matrices
+ * with one column. So we "wrap" these things, and unwrap the result before we
+ * send it off.
+ */
+
+static struct vtn_ssa_value *
+wrap_matrix(struct vtn_builder *b, struct vtn_ssa_value *val)
+{
+ if (val == NULL)
+ return NULL;
+
+ if (glsl_type_is_matrix(val->type))
+ return val;
+
+ struct vtn_ssa_value *dest = rzalloc(b, struct vtn_ssa_value);
+ dest->type = val->type;
+ dest->elems = ralloc_array(b, struct vtn_ssa_value *, 1);
+ dest->elems[0] = val;
+
+ return dest;
+}
+
+static struct vtn_ssa_value *
+unwrap_matrix(struct vtn_ssa_value *val)
+{
+ if (glsl_type_is_matrix(val->type))
+ return val;
+
+ return val->elems[0];
+}
+
+static struct vtn_ssa_value *
+matrix_multiply(struct vtn_builder *b,
+ struct vtn_ssa_value *_src0, struct vtn_ssa_value *_src1)
+{
+
+ struct vtn_ssa_value *src0 = wrap_matrix(b, _src0);
+ struct vtn_ssa_value *src1 = wrap_matrix(b, _src1);
+ struct vtn_ssa_value *src0_transpose = wrap_matrix(b, _src0->transposed);
+ struct vtn_ssa_value *src1_transpose = wrap_matrix(b, _src1->transposed);
+
+ unsigned src0_rows = glsl_get_vector_elements(src0->type);
+ unsigned src0_columns = glsl_get_matrix_columns(src0->type);
+ unsigned src1_columns = glsl_get_matrix_columns(src1->type);
+
+ const struct glsl_type *dest_type;
+ if (src1_columns > 1) {
+ dest_type = glsl_matrix_type(glsl_get_base_type(src0->type),
+ src0_rows, src1_columns);
+ } else {
+ dest_type = glsl_vector_type(glsl_get_base_type(src0->type), src0_rows);
+ }
+ struct vtn_ssa_value *dest = vtn_create_ssa_value(b, dest_type);
+
+ dest = wrap_matrix(b, dest);
+
+ bool transpose_result = false;
+ if (src0_transpose && src1_transpose) {
+ /* transpose(A) * transpose(B) = transpose(B * A) */
+ src1 = src0_transpose;
+ src0 = src1_transpose;
+ src0_transpose = NULL;
+ src1_transpose = NULL;
+ transpose_result = true;
+ }
+
+ if (src0_transpose && !src1_transpose &&
+ glsl_get_base_type(src0->type) == GLSL_TYPE_FLOAT) {
+ /* We already have the rows of src0 and the columns of src1 available,
+ * so we can just take the dot product of each row with each column to
+ * get the result.
+ */
+
+ for (unsigned i = 0; i < src1_columns; i++) {
+ nir_ssa_def *vec_src[4];
+ for (unsigned j = 0; j < src0_rows; j++) {
+ vec_src[j] = nir_fdot(&b->nb, src0_transpose->elems[j]->def,
+ src1->elems[i]->def);
+ }
+ dest->elems[i]->def = nir_vec(&b->nb, vec_src, src0_rows);
+ }
+ } else {
+ /* We don't handle the case where src1 is transposed but not src0, since
+ * the general case only uses individual components of src1 so the
+ * optimizer should chew through the transpose we emitted for src1.
+ */
+
+ for (unsigned i = 0; i < src1_columns; i++) {
+ /* dest[i] = sum(src0[j] * src1[i][j] for all j) */
+ dest->elems[i]->def =
+ nir_fmul(&b->nb, src0->elems[0]->def,
+ nir_channel(&b->nb, src1->elems[i]->def, 0));
+ for (unsigned j = 1; j < src0_columns; j++) {
+ dest->elems[i]->def =
+ nir_fadd(&b->nb, dest->elems[i]->def,
+ nir_fmul(&b->nb, src0->elems[j]->def,
+ nir_channel(&b->nb, src1->elems[i]->def, j)));
+ }
+ }
+ }
+
+ dest = unwrap_matrix(dest);
+
+ if (transpose_result)
+ dest = vtn_ssa_transpose(b, dest);
+
+ return dest;
+}
+
+static struct vtn_ssa_value *
+mat_times_scalar(struct vtn_builder *b,
+ struct vtn_ssa_value *mat,
+ nir_ssa_def *scalar)
+{
+ struct vtn_ssa_value *dest = vtn_create_ssa_value(b, mat->type);
+ for (unsigned i = 0; i < glsl_get_matrix_columns(mat->type); i++) {
+ if (glsl_get_base_type(mat->type) == GLSL_TYPE_FLOAT)
+ dest->elems[i]->def = nir_fmul(&b->nb, mat->elems[i]->def, scalar);
+ else
+ dest->elems[i]->def = nir_imul(&b->nb, mat->elems[i]->def, scalar);
+ }
+
+ return dest;
+}
+
+static void
+vtn_handle_matrix_alu(struct vtn_builder *b, SpvOp opcode,
+ struct vtn_value *dest,
+ struct vtn_ssa_value *src0, struct vtn_ssa_value *src1)
+{
+ switch (opcode) {
+ case SpvOpFNegate: {
+ dest->ssa = vtn_create_ssa_value(b, src0->type);
+ unsigned cols = glsl_get_matrix_columns(src0->type);
+ for (unsigned i = 0; i < cols; i++)
+ dest->ssa->elems[i]->def = nir_fneg(&b->nb, src0->elems[i]->def);
+ break;
+ }
+
+ case SpvOpFAdd: {
+ dest->ssa = vtn_create_ssa_value(b, src0->type);
+ unsigned cols = glsl_get_matrix_columns(src0->type);
+ for (unsigned i = 0; i < cols; i++)
+ dest->ssa->elems[i]->def =
+ nir_fadd(&b->nb, src0->elems[i]->def, src1->elems[i]->def);
+ break;
+ }
+
+ case SpvOpFSub: {
+ dest->ssa = vtn_create_ssa_value(b, src0->type);
+ unsigned cols = glsl_get_matrix_columns(src0->type);
+ for (unsigned i = 0; i < cols; i++)
+ dest->ssa->elems[i]->def =
+ nir_fsub(&b->nb, src0->elems[i]->def, src1->elems[i]->def);
+ break;
+ }
+
+ case SpvOpTranspose:
+ dest->ssa = vtn_ssa_transpose(b, src0);
+ break;
+
+ case SpvOpMatrixTimesScalar:
+ if (src0->transposed) {
+ dest->ssa = vtn_ssa_transpose(b, mat_times_scalar(b, src0->transposed,
+ src1->def));
+ } else {
+ dest->ssa = mat_times_scalar(b, src0, src1->def);
+ }
+ break;
+
+ case SpvOpVectorTimesMatrix:
+ case SpvOpMatrixTimesVector:
+ case SpvOpMatrixTimesMatrix:
+ if (opcode == SpvOpVectorTimesMatrix) {
+ dest->ssa = matrix_multiply(b, vtn_ssa_transpose(b, src1), src0);
+ } else {
+ dest->ssa = matrix_multiply(b, src0, src1);
+ }
+ break;
+
+ default: unreachable("unknown matrix opcode");
+ }
+}
+
+nir_op
+vtn_nir_alu_op_for_spirv_opcode(SpvOp opcode, bool *swap)
+{
+ /* Indicates that the first two arguments should be swapped. This is
+ * used for implementing greater-than and less-than-or-equal.
+ */
+ *swap = false;
+
+ switch (opcode) {
+ case SpvOpSNegate: return nir_op_ineg;
+ case SpvOpFNegate: return nir_op_fneg;
+ case SpvOpNot: return nir_op_inot;
+ case SpvOpIAdd: return nir_op_iadd;
+ case SpvOpFAdd: return nir_op_fadd;
+ case SpvOpISub: return nir_op_isub;
+ case SpvOpFSub: return nir_op_fsub;
+ case SpvOpIMul: return nir_op_imul;
+ case SpvOpFMul: return nir_op_fmul;
+ case SpvOpUDiv: return nir_op_udiv;
+ case SpvOpSDiv: return nir_op_idiv;
+ case SpvOpFDiv: return nir_op_fdiv;
+ case SpvOpUMod: return nir_op_umod;
+ case SpvOpSMod: return nir_op_imod;
+ case SpvOpFMod: return nir_op_fmod;
+ case SpvOpSRem: return nir_op_irem;
+ case SpvOpFRem: return nir_op_frem;
+
+ case SpvOpShiftRightLogical: return nir_op_ushr;
+ case SpvOpShiftRightArithmetic: return nir_op_ishr;
+ case SpvOpShiftLeftLogical: return nir_op_ishl;
+ case SpvOpLogicalOr: return nir_op_ior;
+ case SpvOpLogicalEqual: return nir_op_ieq;
+ case SpvOpLogicalNotEqual: return nir_op_ine;
+ case SpvOpLogicalAnd: return nir_op_iand;
+ case SpvOpLogicalNot: return nir_op_inot;
+ case SpvOpBitwiseOr: return nir_op_ior;
+ case SpvOpBitwiseXor: return nir_op_ixor;
+ case SpvOpBitwiseAnd: return nir_op_iand;
+ case SpvOpSelect: return nir_op_bcsel;
+ case SpvOpIEqual: return nir_op_ieq;
+
+ case SpvOpBitFieldInsert: return nir_op_bitfield_insert;
+ case SpvOpBitFieldSExtract: return nir_op_ibitfield_extract;
+ case SpvOpBitFieldUExtract: return nir_op_ubitfield_extract;
+ case SpvOpBitReverse: return nir_op_bitfield_reverse;
+ case SpvOpBitCount: return nir_op_bit_count;
+
+ /* Comparisons: (TODO: How do we want to handled ordered/unordered?) */
+ case SpvOpFOrdEqual: return nir_op_feq;
+ case SpvOpFUnordEqual: return nir_op_feq;
+ case SpvOpINotEqual: return nir_op_ine;
+ case SpvOpFOrdNotEqual: return nir_op_fne;
+ case SpvOpFUnordNotEqual: return nir_op_fne;
+ case SpvOpULessThan: return nir_op_ult;
+ case SpvOpSLessThan: return nir_op_ilt;
+ case SpvOpFOrdLessThan: return nir_op_flt;
+ case SpvOpFUnordLessThan: return nir_op_flt;
+ case SpvOpUGreaterThan: *swap = true; return nir_op_ult;
+ case SpvOpSGreaterThan: *swap = true; return nir_op_ilt;
+ case SpvOpFOrdGreaterThan: *swap = true; return nir_op_flt;
+ case SpvOpFUnordGreaterThan: *swap = true; return nir_op_flt;
+ case SpvOpULessThanEqual: *swap = true; return nir_op_uge;
+ case SpvOpSLessThanEqual: *swap = true; return nir_op_ige;
+ case SpvOpFOrdLessThanEqual: *swap = true; return nir_op_fge;
+ case SpvOpFUnordLessThanEqual: *swap = true; return nir_op_fge;
+ case SpvOpUGreaterThanEqual: return nir_op_uge;
+ case SpvOpSGreaterThanEqual: return nir_op_ige;
+ case SpvOpFOrdGreaterThanEqual: return nir_op_fge;
+ case SpvOpFUnordGreaterThanEqual: return nir_op_fge;
+
+ /* Conversions: */
+ case SpvOpConvertFToU: return nir_op_f2u;
+ case SpvOpConvertFToS: return nir_op_f2i;
+ case SpvOpConvertSToF: return nir_op_i2f;
+ case SpvOpConvertUToF: return nir_op_u2f;
+ case SpvOpBitcast: return nir_op_imov;
+ case SpvOpUConvert:
+ case SpvOpQuantizeToF16: return nir_op_fquantize2f16;
+ /* TODO: NIR is 32-bit only; these are no-ops. */
+ case SpvOpSConvert: return nir_op_imov;
+ case SpvOpFConvert: return nir_op_fmov;
+
+ /* Derivatives: */
+ case SpvOpDPdx: return nir_op_fddx;
+ case SpvOpDPdy: return nir_op_fddy;
+ case SpvOpDPdxFine: return nir_op_fddx_fine;
+ case SpvOpDPdyFine: return nir_op_fddy_fine;
+ case SpvOpDPdxCoarse: return nir_op_fddx_coarse;
+ case SpvOpDPdyCoarse: return nir_op_fddy_coarse;
+
+ default:
+ unreachable("No NIR equivalent");
+ }
+}
+
+static void
+handle_no_contraction(struct vtn_builder *b, struct vtn_value *val, int member,
+ const struct vtn_decoration *dec, void *_void)
+{
+ assert(dec->scope == VTN_DEC_DECORATION);
+ if (dec->decoration != SpvDecorationNoContraction)
+ return;
+
+ b->nb.exact = true;
+}
+
+void
+vtn_handle_alu(struct vtn_builder *b, SpvOp opcode,
+ const uint32_t *w, unsigned count)
+{
+ struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
+ const struct glsl_type *type =
+ vtn_value(b, w[1], vtn_value_type_type)->type->type;
+
+ vtn_foreach_decoration(b, val, handle_no_contraction, NULL);
+
+ /* Collect the various SSA sources */
+ const unsigned num_inputs = count - 3;
+ struct vtn_ssa_value *vtn_src[4] = { NULL, };
+ for (unsigned i = 0; i < num_inputs; i++)
+ vtn_src[i] = vtn_ssa_value(b, w[i + 3]);
+
+ if (glsl_type_is_matrix(vtn_src[0]->type) ||
+ (num_inputs >= 2 && glsl_type_is_matrix(vtn_src[1]->type))) {
+ vtn_handle_matrix_alu(b, opcode, val, vtn_src[0], vtn_src[1]);
+ b->nb.exact = false;
+ return;
+ }
+
+ val->ssa = vtn_create_ssa_value(b, type);
+ nir_ssa_def *src[4] = { NULL, };
+ for (unsigned i = 0; i < num_inputs; i++) {
+ assert(glsl_type_is_vector_or_scalar(vtn_src[i]->type));
+ src[i] = vtn_src[i]->def;
+ }
+
+ switch (opcode) {
+ case SpvOpAny:
+ if (src[0]->num_components == 1) {
+ val->ssa->def = nir_imov(&b->nb, src[0]);
+ } else {
+ nir_op op;
+ switch (src[0]->num_components) {
+ case 2: op = nir_op_bany_inequal2; break;
+ case 3: op = nir_op_bany_inequal3; break;
+ case 4: op = nir_op_bany_inequal4; break;
+ }
+ val->ssa->def = nir_build_alu(&b->nb, op, src[0],
+ nir_imm_int(&b->nb, NIR_FALSE),
+ NULL, NULL);
+ }
+ break;
+
+ case SpvOpAll:
+ if (src[0]->num_components == 1) {
+ val->ssa->def = nir_imov(&b->nb, src[0]);
+ } else {
+ nir_op op;
+ switch (src[0]->num_components) {
+ case 2: op = nir_op_ball_iequal2; break;
+ case 3: op = nir_op_ball_iequal3; break;
+ case 4: op = nir_op_ball_iequal4; break;
+ }
+ val->ssa->def = nir_build_alu(&b->nb, op, src[0],
+ nir_imm_int(&b->nb, NIR_TRUE),
+ NULL, NULL);
+ }
+ break;
+
+ case SpvOpOuterProduct: {
+ for (unsigned i = 0; i < src[1]->num_components; i++) {
+ val->ssa->elems[i]->def =
+ nir_fmul(&b->nb, src[0], nir_channel(&b->nb, src[1], i));
+ }
+ break;
+ }
+
+ case SpvOpDot:
+ val->ssa->def = nir_fdot(&b->nb, src[0], src[1]);
+ break;
+
+ case SpvOpIAddCarry:
+ assert(glsl_type_is_struct(val->ssa->type));
+ val->ssa->elems[0]->def = nir_iadd(&b->nb, src[0], src[1]);
+ val->ssa->elems[1]->def = nir_uadd_carry(&b->nb, src[0], src[1]);
+ break;
+
+ case SpvOpISubBorrow:
+ assert(glsl_type_is_struct(val->ssa->type));
+ val->ssa->elems[0]->def = nir_isub(&b->nb, src[0], src[1]);
+ val->ssa->elems[1]->def = nir_usub_borrow(&b->nb, src[0], src[1]);
+ break;
+
+ case SpvOpUMulExtended:
+ assert(glsl_type_is_struct(val->ssa->type));
+ val->ssa->elems[0]->def = nir_imul(&b->nb, src[0], src[1]);
+ val->ssa->elems[1]->def = nir_umul_high(&b->nb, src[0], src[1]);
+ break;
+
+ case SpvOpSMulExtended:
+ assert(glsl_type_is_struct(val->ssa->type));
+ val->ssa->elems[0]->def = nir_imul(&b->nb, src[0], src[1]);
+ val->ssa->elems[1]->def = nir_imul_high(&b->nb, src[0], src[1]);
+ break;
+
+ case SpvOpFwidth:
+ val->ssa->def = nir_fadd(&b->nb,
+ nir_fabs(&b->nb, nir_fddx(&b->nb, src[0])),
+ nir_fabs(&b->nb, nir_fddy(&b->nb, src[0])));
+ break;
+ case SpvOpFwidthFine:
+ val->ssa->def = nir_fadd(&b->nb,
+ nir_fabs(&b->nb, nir_fddx_fine(&b->nb, src[0])),
+ nir_fabs(&b->nb, nir_fddy_fine(&b->nb, src[0])));
+ break;
+ case SpvOpFwidthCoarse:
+ val->ssa->def = nir_fadd(&b->nb,
+ nir_fabs(&b->nb, nir_fddx_coarse(&b->nb, src[0])),
+ nir_fabs(&b->nb, nir_fddy_coarse(&b->nb, src[0])));
+ break;
+
+ case SpvOpVectorTimesScalar:
+ /* The builder will take care of splatting for us. */
+ val->ssa->def = nir_fmul(&b->nb, src[0], src[1]);
+ break;
+
+ case SpvOpIsNan:
+ val->ssa->def = nir_fne(&b->nb, src[0], src[0]);
+ break;
+
+ case SpvOpIsInf:
+ val->ssa->def = nir_feq(&b->nb, nir_fabs(&b->nb, src[0]),
+ nir_imm_float(&b->nb, INFINITY));
+ break;
+
+ default: {
+ bool swap;
+ nir_op op = vtn_nir_alu_op_for_spirv_opcode(opcode, &swap);
+
+ if (swap) {
+ nir_ssa_def *tmp = src[0];
+ src[0] = src[1];
+ src[1] = tmp;
+ }
+
+ val->ssa->def = nir_build_alu(&b->nb, op, src[0], src[1], src[2], src[3]);
+ break;
+ } /* default */
+ }
+
+ b->nb.exact = false;
+}
diff --git a/src/compiler/spirv/vtn_cfg.c b/src/compiler/spirv/vtn_cfg.c
new file mode 100644
index 00000000000..6a43ef8b2dd
--- /dev/null
+++ b/src/compiler/spirv/vtn_cfg.c
@@ -0,0 +1,778 @@
+/*
+ * Copyright © 2015 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#include "vtn_private.h"
+#include "nir/nir_vla.h"
+
+static bool
+vtn_cfg_handle_prepass_instruction(struct vtn_builder *b, SpvOp opcode,
+ const uint32_t *w, unsigned count)
+{
+ switch (opcode) {
+ case SpvOpFunction: {
+ assert(b->func == NULL);
+ b->func = rzalloc(b, struct vtn_function);
+
+ list_inithead(&b->func->body);
+ b->func->control = w[3];
+
+ const struct glsl_type *result_type =
+ vtn_value(b, w[1], vtn_value_type_type)->type->type;
+ struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_function);
+ val->func = b->func;
+
+ const struct glsl_type *func_type =
+ vtn_value(b, w[4], vtn_value_type_type)->type->type;
+
+ assert(glsl_get_function_return_type(func_type) == result_type);
+
+ nir_function *func =
+ nir_function_create(b->shader, ralloc_strdup(b->shader, val->name));
+
+ func->num_params = glsl_get_length(func_type);
+ func->params = ralloc_array(b->shader, nir_parameter, func->num_params);
+ for (unsigned i = 0; i < func->num_params; i++) {
+ const struct glsl_function_param *param =
+ glsl_get_function_param(func_type, i);
+ func->params[i].type = param->type;
+ if (param->in) {
+ if (param->out) {
+ func->params[i].param_type = nir_parameter_inout;
+ } else {
+ func->params[i].param_type = nir_parameter_in;
+ }
+ } else {
+ if (param->out) {
+ func->params[i].param_type = nir_parameter_out;
+ } else {
+ assert(!"Parameter is neither in nor out");
+ }
+ }
+ }
+
+ func->return_type = glsl_get_function_return_type(func_type);
+
+ b->func->impl = nir_function_impl_create(func);
+
+ b->func_param_idx = 0;
+ break;
+ }
+
+ case SpvOpFunctionEnd:
+ b->func->end = w;
+ b->func = NULL;
+ break;
+
+ case SpvOpFunctionParameter: {
+ struct vtn_value *val =
+ vtn_push_value(b, w[2], vtn_value_type_access_chain);
+
+ struct vtn_type *type = vtn_value(b, w[1], vtn_value_type_type)->type;
+
+ assert(b->func_param_idx < b->func->impl->num_params);
+ nir_variable *param = b->func->impl->params[b->func_param_idx++];
+
+ assert(param->type == type->type);
+
+ /* Name the parameter so it shows up nicely in NIR */
+ param->name = ralloc_strdup(param, val->name);
+
+ struct vtn_variable *vtn_var = rzalloc(b, struct vtn_variable);
+ vtn_var->type = type;
+ vtn_var->var = param;
+ vtn_var->chain.var = vtn_var;
+ vtn_var->chain.length = 0;
+
+ struct vtn_type *without_array = type;
+ while(glsl_type_is_array(without_array->type))
+ without_array = without_array->array_element;
+
+ if (glsl_type_is_image(without_array->type)) {
+ vtn_var->mode = vtn_variable_mode_image;
+ param->interface_type = without_array->type;
+ } else if (glsl_type_is_sampler(without_array->type)) {
+ vtn_var->mode = vtn_variable_mode_sampler;
+ param->interface_type = without_array->type;
+ } else {
+ vtn_var->mode = vtn_variable_mode_param;
+ }
+
+ val->access_chain = &vtn_var->chain;
+ break;
+ }
+
+ case SpvOpLabel: {
+ assert(b->block == NULL);
+ b->block = rzalloc(b, struct vtn_block);
+ b->block->node.type = vtn_cf_node_type_block;
+ b->block->label = w;
+ vtn_push_value(b, w[1], vtn_value_type_block)->block = b->block;
+
+ if (b->func->start_block == NULL) {
+ /* This is the first block encountered for this function. In this
+ * case, we set the start block and add it to the list of
+ * implemented functions that we'll walk later.
+ */
+ b->func->start_block = b->block;
+ exec_list_push_tail(&b->functions, &b->func->node);
+ }
+ break;
+ }
+
+ case SpvOpSelectionMerge:
+ case SpvOpLoopMerge:
+ assert(b->block && b->block->merge == NULL);
+ b->block->merge = w;
+ break;
+
+ case SpvOpBranch:
+ case SpvOpBranchConditional:
+ case SpvOpSwitch:
+ case SpvOpKill:
+ case SpvOpReturn:
+ case SpvOpReturnValue:
+ case SpvOpUnreachable:
+ assert(b->block && b->block->branch == NULL);
+ b->block->branch = w;
+ b->block = NULL;
+ break;
+
+ default:
+ /* Continue on as per normal */
+ return true;
+ }
+
+ return true;
+}
+
+static void
+vtn_add_case(struct vtn_builder *b, struct vtn_switch *swtch,
+ struct vtn_block *break_block,
+ uint32_t block_id, uint32_t val, bool is_default)
+{
+ struct vtn_block *case_block =
+ vtn_value(b, block_id, vtn_value_type_block)->block;
+
+ /* Don't create dummy cases that just break */
+ if (case_block == break_block)
+ return;
+
+ if (case_block->switch_case == NULL) {
+ struct vtn_case *c = ralloc(b, struct vtn_case);
+
+ list_inithead(&c->body);
+ c->start_block = case_block;
+ c->fallthrough = NULL;
+ nir_array_init(&c->values, b);
+ c->is_default = false;
+ c->visited = false;
+
+ list_addtail(&c->link, &swtch->cases);
+
+ case_block->switch_case = c;
+ }
+
+ if (is_default) {
+ case_block->switch_case->is_default = true;
+ } else {
+ nir_array_add(&case_block->switch_case->values, uint32_t, val);
+ }
+}
+
+/* This function performs a depth-first search of the cases and puts them
+ * in fall-through order.
+ */
+static void
+vtn_order_case(struct vtn_switch *swtch, struct vtn_case *cse)
+{
+ if (cse->visited)
+ return;
+
+ cse->visited = true;
+
+ list_del(&cse->link);
+
+ if (cse->fallthrough) {
+ vtn_order_case(swtch, cse->fallthrough);
+
+ /* If we have a fall-through, place this case right before the case it
+ * falls through to. This ensures that fallthroughs come one after
+ * the other. These two can never get separated because that would
+ * imply something else falling through to the same case. Also, this
+ * can't break ordering because the DFS ensures that this case is
+ * visited before anything that falls through to it.
+ */
+ list_addtail(&cse->link, &cse->fallthrough->link);
+ } else {
+ list_add(&cse->link, &swtch->cases);
+ }
+}
+
+static enum vtn_branch_type
+vtn_get_branch_type(struct vtn_block *block,
+ struct vtn_case *swcase, struct vtn_block *switch_break,
+ struct vtn_block *loop_break, struct vtn_block *loop_cont)
+{
+ if (block->switch_case) {
+ /* This branch is actually a fallthrough */
+ assert(swcase->fallthrough == NULL ||
+ swcase->fallthrough == block->switch_case);
+ swcase->fallthrough = block->switch_case;
+ return vtn_branch_type_switch_fallthrough;
+ } else if (block == switch_break) {
+ return vtn_branch_type_switch_break;
+ } else if (block == loop_break) {
+ return vtn_branch_type_loop_break;
+ } else if (block == loop_cont) {
+ return vtn_branch_type_loop_continue;
+ } else {
+ return vtn_branch_type_none;
+ }
+}
+
+static void
+vtn_cfg_walk_blocks(struct vtn_builder *b, struct list_head *cf_list,
+ struct vtn_block *start, struct vtn_case *switch_case,
+ struct vtn_block *switch_break,
+ struct vtn_block *loop_break, struct vtn_block *loop_cont,
+ struct vtn_block *end)
+{
+ struct vtn_block *block = start;
+ while (block != end) {
+ if (block->merge && (*block->merge & SpvOpCodeMask) == SpvOpLoopMerge &&
+ !block->loop) {
+ struct vtn_loop *loop = ralloc(b, struct vtn_loop);
+
+ loop->node.type = vtn_cf_node_type_loop;
+ list_inithead(&loop->body);
+ list_inithead(&loop->cont_body);
+ loop->control = block->merge[3];
+
+ list_addtail(&loop->node.link, cf_list);
+ block->loop = loop;
+
+ struct vtn_block *new_loop_break =
+ vtn_value(b, block->merge[1], vtn_value_type_block)->block;
+ struct vtn_block *new_loop_cont =
+ vtn_value(b, block->merge[2], vtn_value_type_block)->block;
+
+ /* Note: This recursive call will start with the current block as
+ * its start block. If we weren't careful, we would get here
+ * again and end up in infinite recursion. This is why we set
+ * block->loop above and check for it before creating one. This
+ * way, we only create the loop once and the second call that
+ * tries to handle this loop goes to the cases below and gets
+ * handled as a regular block.
+ *
+ * Note: When we make the recursive walk calls, we pass NULL for
+ * the switch break since you have to break out of the loop first.
+ * We do, however, still pass the current switch case because it's
+ * possible that the merge block for the loop is the start of
+ * another case.
+ */
+ vtn_cfg_walk_blocks(b, &loop->body, block, switch_case, NULL,
+ new_loop_break, new_loop_cont, NULL );
+ vtn_cfg_walk_blocks(b, &loop->cont_body, new_loop_cont, NULL, NULL,
+ new_loop_break, NULL, block);
+
+ block = new_loop_break;
+ continue;
+ }
+
+ assert(block->node.link.next == NULL);
+ list_addtail(&block->node.link, cf_list);
+
+ switch (*block->branch & SpvOpCodeMask) {
+ case SpvOpBranch: {
+ struct vtn_block *branch_block =
+ vtn_value(b, block->branch[1], vtn_value_type_block)->block;
+
+ block->branch_type = vtn_get_branch_type(branch_block,
+ switch_case, switch_break,
+ loop_break, loop_cont);
+
+ if (block->branch_type != vtn_branch_type_none)
+ return;
+
+ block = branch_block;
+ continue;
+ }
+
+ case SpvOpReturn:
+ case SpvOpReturnValue:
+ block->branch_type = vtn_branch_type_return;
+ return;
+
+ case SpvOpKill:
+ block->branch_type = vtn_branch_type_discard;
+ return;
+
+ case SpvOpBranchConditional: {
+ struct vtn_block *then_block =
+ vtn_value(b, block->branch[2], vtn_value_type_block)->block;
+ struct vtn_block *else_block =
+ vtn_value(b, block->branch[3], vtn_value_type_block)->block;
+
+ struct vtn_if *if_stmt = ralloc(b, struct vtn_if);
+
+ if_stmt->node.type = vtn_cf_node_type_if;
+ if_stmt->condition = block->branch[1];
+ list_inithead(&if_stmt->then_body);
+ list_inithead(&if_stmt->else_body);
+
+ list_addtail(&if_stmt->node.link, cf_list);
+
+ if (block->merge &&
+ (*block->merge & SpvOpCodeMask) == SpvOpSelectionMerge) {
+ if_stmt->control = block->merge[2];
+ }
+
+ if_stmt->then_type = vtn_get_branch_type(then_block,
+ switch_case, switch_break,
+ loop_break, loop_cont);
+ if_stmt->else_type = vtn_get_branch_type(else_block,
+ switch_case, switch_break,
+ loop_break, loop_cont);
+
+ if (if_stmt->then_type == vtn_branch_type_none &&
+ if_stmt->else_type == vtn_branch_type_none) {
+ /* Neither side of the if is something we can short-circuit. */
+ assert((*block->merge & SpvOpCodeMask) == SpvOpSelectionMerge);
+ struct vtn_block *merge_block =
+ vtn_value(b, block->merge[1], vtn_value_type_block)->block;
+
+ vtn_cfg_walk_blocks(b, &if_stmt->then_body, then_block,
+ switch_case, switch_break,
+ loop_break, loop_cont, merge_block);
+ vtn_cfg_walk_blocks(b, &if_stmt->else_body, else_block,
+ switch_case, switch_break,
+ loop_break, loop_cont, merge_block);
+
+ enum vtn_branch_type merge_type =
+ vtn_get_branch_type(merge_block, switch_case, switch_break,
+ loop_break, loop_cont);
+ if (merge_type == vtn_branch_type_none) {
+ block = merge_block;
+ continue;
+ } else {
+ return;
+ }
+ } else if (if_stmt->then_type != vtn_branch_type_none &&
+ if_stmt->else_type != vtn_branch_type_none) {
+ /* Both sides were short-circuited. We're done here. */
+ return;
+ } else {
+ /* Exeactly one side of the branch could be short-circuited.
+ * We set the branch up as a predicated break/continue and we
+ * continue on with the other side as if it were what comes
+ * after the if.
+ */
+ if (if_stmt->then_type == vtn_branch_type_none) {
+ block = then_block;
+ } else {
+ block = else_block;
+ }
+ continue;
+ }
+ unreachable("Should have returned or continued");
+ }
+
+ case SpvOpSwitch: {
+ assert((*block->merge & SpvOpCodeMask) == SpvOpSelectionMerge);
+ struct vtn_block *break_block =
+ vtn_value(b, block->merge[1], vtn_value_type_block)->block;
+
+ struct vtn_switch *swtch = ralloc(b, struct vtn_switch);
+
+ swtch->node.type = vtn_cf_node_type_switch;
+ swtch->selector = block->branch[1];
+ list_inithead(&swtch->cases);
+
+ list_addtail(&swtch->node.link, cf_list);
+
+ /* First, we go through and record all of the cases. */
+ const uint32_t *branch_end =
+ block->branch + (block->branch[0] >> SpvWordCountShift);
+
+ vtn_add_case(b, swtch, break_block, block->branch[2], 0, true);
+ for (const uint32_t *w = block->branch + 3; w < branch_end; w += 2)
+ vtn_add_case(b, swtch, break_block, w[1], w[0], false);
+
+ /* Now, we go through and walk the blocks. While we walk through
+ * the blocks, we also gather the much-needed fall-through
+ * information.
+ */
+ list_for_each_entry(struct vtn_case, cse, &swtch->cases, link) {
+ assert(cse->start_block != break_block);
+ vtn_cfg_walk_blocks(b, &cse->body, cse->start_block, cse,
+ break_block, NULL, loop_cont, NULL);
+ }
+
+ /* Finally, we walk over all of the cases one more time and put
+ * them in fall-through order.
+ */
+ for (const uint32_t *w = block->branch + 2; w < branch_end; w += 2) {
+ struct vtn_block *case_block =
+ vtn_value(b, *w, vtn_value_type_block)->block;
+
+ if (case_block == break_block)
+ continue;
+
+ assert(case_block->switch_case);
+
+ vtn_order_case(swtch, case_block->switch_case);
+ }
+
+ block = break_block;
+ continue;
+ }
+
+ case SpvOpUnreachable:
+ return;
+
+ default:
+ unreachable("Unhandled opcode");
+ }
+ }
+}
+
+void
+vtn_build_cfg(struct vtn_builder *b, const uint32_t *words, const uint32_t *end)
+{
+ vtn_foreach_instruction(b, words, end,
+ vtn_cfg_handle_prepass_instruction);
+
+ foreach_list_typed(struct vtn_function, func, node, &b->functions) {
+ vtn_cfg_walk_blocks(b, &func->body, func->start_block,
+ NULL, NULL, NULL, NULL, NULL);
+ }
+}
+
+static bool
+vtn_handle_phis_first_pass(struct vtn_builder *b, SpvOp opcode,
+ const uint32_t *w, unsigned count)
+{
+ if (opcode == SpvOpLabel)
+ return true; /* Nothing to do */
+
+ /* If this isn't a phi node, stop. */
+ if (opcode != SpvOpPhi)
+ return false;
+
+ /* For handling phi nodes, we do a poor-man's out-of-ssa on the spot.
+ * For each phi, we create a variable with the appropreate type and
+ * do a load from that variable. Then, in a second pass, we add
+ * stores to that variable to each of the predecessor blocks.
+ *
+ * We could do something more intelligent here. However, in order to
+ * handle loops and things properly, we really need dominance
+ * information. It would end up basically being the into-SSA
+ * algorithm all over again. It's easier if we just let
+ * lower_vars_to_ssa do that for us instead of repeating it here.
+ */
+ struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
+
+ struct vtn_type *type = vtn_value(b, w[1], vtn_value_type_type)->type;
+ nir_variable *phi_var =
+ nir_local_variable_create(b->nb.impl, type->type, "phi");
+ _mesa_hash_table_insert(b->phi_table, w, phi_var);
+
+ val->ssa = vtn_local_load(b, nir_deref_var_create(b, phi_var));
+
+ return true;
+}
+
+static bool
+vtn_handle_phi_second_pass(struct vtn_builder *b, SpvOp opcode,
+ const uint32_t *w, unsigned count)
+{
+ if (opcode != SpvOpPhi)
+ return true;
+
+ struct hash_entry *phi_entry = _mesa_hash_table_search(b->phi_table, w);
+ assert(phi_entry);
+ nir_variable *phi_var = phi_entry->data;
+
+ for (unsigned i = 3; i < count; i += 2) {
+ struct vtn_ssa_value *src = vtn_ssa_value(b, w[i]);
+ struct vtn_block *pred =
+ vtn_value(b, w[i + 1], vtn_value_type_block)->block;
+
+ b->nb.cursor = nir_after_block_before_jump(pred->end_block);
+
+ vtn_local_store(b, src, nir_deref_var_create(b, phi_var));
+ }
+
+ return true;
+}
+
+static void
+vtn_emit_branch(struct vtn_builder *b, enum vtn_branch_type branch_type,
+ nir_variable *switch_fall_var, bool *has_switch_break)
+{
+ switch (branch_type) {
+ case vtn_branch_type_switch_break:
+ nir_store_var(&b->nb, switch_fall_var, nir_imm_int(&b->nb, NIR_FALSE), 1);
+ *has_switch_break = true;
+ break;
+ case vtn_branch_type_switch_fallthrough:
+ break; /* Nothing to do */
+ case vtn_branch_type_loop_break:
+ nir_jump(&b->nb, nir_jump_break);
+ break;
+ case vtn_branch_type_loop_continue:
+ nir_jump(&b->nb, nir_jump_continue);
+ break;
+ case vtn_branch_type_return:
+ nir_jump(&b->nb, nir_jump_return);
+ break;
+ case vtn_branch_type_discard: {
+ nir_intrinsic_instr *discard =
+ nir_intrinsic_instr_create(b->nb.shader, nir_intrinsic_discard);
+ nir_builder_instr_insert(&b->nb, &discard->instr);
+ break;
+ }
+ default:
+ unreachable("Invalid branch type");
+ }
+}
+
+static void
+vtn_emit_cf_list(struct vtn_builder *b, struct list_head *cf_list,
+ nir_variable *switch_fall_var, bool *has_switch_break,
+ vtn_instruction_handler handler)
+{
+ list_for_each_entry(struct vtn_cf_node, node, cf_list, link) {
+ switch (node->type) {
+ case vtn_cf_node_type_block: {
+ struct vtn_block *block = (struct vtn_block *)node;
+
+ const uint32_t *block_start = block->label;
+ const uint32_t *block_end = block->merge ? block->merge :
+ block->branch;
+
+ block_start = vtn_foreach_instruction(b, block_start, block_end,
+ vtn_handle_phis_first_pass);
+
+ vtn_foreach_instruction(b, block_start, block_end, handler);
+
+ block->end_block = nir_cursor_current_block(b->nb.cursor);
+
+ if ((*block->branch & SpvOpCodeMask) == SpvOpReturnValue) {
+ struct vtn_ssa_value *src = vtn_ssa_value(b, block->branch[1]);
+ vtn_local_store(b, src,
+ nir_deref_var_create(b, b->impl->return_var));
+ }
+
+ if (block->branch_type != vtn_branch_type_none) {
+ vtn_emit_branch(b, block->branch_type,
+ switch_fall_var, has_switch_break);
+ }
+
+ break;
+ }
+
+ case vtn_cf_node_type_if: {
+ struct vtn_if *vtn_if = (struct vtn_if *)node;
+
+ nir_if *if_stmt = nir_if_create(b->shader);
+ if_stmt->condition =
+ nir_src_for_ssa(vtn_ssa_value(b, vtn_if->condition)->def);
+ nir_cf_node_insert(b->nb.cursor, &if_stmt->cf_node);
+
+ bool sw_break = false;
+
+ b->nb.cursor = nir_after_cf_list(&if_stmt->then_list);
+ if (vtn_if->then_type == vtn_branch_type_none) {
+ vtn_emit_cf_list(b, &vtn_if->then_body,
+ switch_fall_var, &sw_break, handler);
+ } else {
+ vtn_emit_branch(b, vtn_if->then_type, switch_fall_var, &sw_break);
+ }
+
+ b->nb.cursor = nir_after_cf_list(&if_stmt->else_list);
+ if (vtn_if->else_type == vtn_branch_type_none) {
+ vtn_emit_cf_list(b, &vtn_if->else_body,
+ switch_fall_var, &sw_break, handler);
+ } else {
+ vtn_emit_branch(b, vtn_if->else_type, switch_fall_var, &sw_break);
+ }
+
+ b->nb.cursor = nir_after_cf_node(&if_stmt->cf_node);
+
+ /* If we encountered a switch break somewhere inside of the if,
+ * then it would have been handled correctly by calling
+ * emit_cf_list or emit_branch for the interrior. However, we
+ * need to predicate everything following on wether or not we're
+ * still going.
+ */
+ if (sw_break) {
+ *has_switch_break = true;
+
+ nir_if *switch_if = nir_if_create(b->shader);
+ switch_if->condition =
+ nir_src_for_ssa(nir_load_var(&b->nb, switch_fall_var));
+ nir_cf_node_insert(b->nb.cursor, &switch_if->cf_node);
+
+ b->nb.cursor = nir_after_cf_list(&if_stmt->then_list);
+ }
+ break;
+ }
+
+ case vtn_cf_node_type_loop: {
+ struct vtn_loop *vtn_loop = (struct vtn_loop *)node;
+
+ nir_loop *loop = nir_loop_create(b->shader);
+ nir_cf_node_insert(b->nb.cursor, &loop->cf_node);
+
+ b->nb.cursor = nir_after_cf_list(&loop->body);
+ vtn_emit_cf_list(b, &vtn_loop->body, NULL, NULL, handler);
+
+ if (!list_empty(&vtn_loop->cont_body)) {
+ /* If we have a non-trivial continue body then we need to put
+ * it at the beginning of the loop with a flag to ensure that
+ * it doesn't get executed in the first iteration.
+ */
+ nir_variable *do_cont =
+ nir_local_variable_create(b->nb.impl, glsl_bool_type(), "cont");
+
+ b->nb.cursor = nir_before_cf_node(&loop->cf_node);
+ nir_store_var(&b->nb, do_cont, nir_imm_int(&b->nb, NIR_FALSE), 1);
+
+ b->nb.cursor = nir_before_cf_list(&loop->body);
+ nir_if *cont_if = nir_if_create(b->shader);
+ cont_if->condition = nir_src_for_ssa(nir_load_var(&b->nb, do_cont));
+ nir_cf_node_insert(b->nb.cursor, &cont_if->cf_node);
+
+ b->nb.cursor = nir_after_cf_list(&cont_if->then_list);
+ vtn_emit_cf_list(b, &vtn_loop->cont_body, NULL, NULL, handler);
+
+ b->nb.cursor = nir_after_cf_node(&cont_if->cf_node);
+ nir_store_var(&b->nb, do_cont, nir_imm_int(&b->nb, NIR_TRUE), 1);
+
+ b->has_loop_continue = true;
+ }
+
+ b->nb.cursor = nir_after_cf_node(&loop->cf_node);
+ break;
+ }
+
+ case vtn_cf_node_type_switch: {
+ struct vtn_switch *vtn_switch = (struct vtn_switch *)node;
+
+ /* First, we create a variable to keep track of whether or not the
+ * switch is still going at any given point. Any switch breaks
+ * will set this variable to false.
+ */
+ nir_variable *fall_var =
+ nir_local_variable_create(b->nb.impl, glsl_bool_type(), "fall");
+ nir_store_var(&b->nb, fall_var, nir_imm_int(&b->nb, NIR_FALSE), 1);
+
+ /* Next, we gather up all of the conditions. We have to do this
+ * up-front because we also need to build an "any" condition so
+ * that we can use !any for default.
+ */
+ const int num_cases = list_length(&vtn_switch->cases);
+ NIR_VLA(nir_ssa_def *, conditions, num_cases);
+
+ nir_ssa_def *sel = vtn_ssa_value(b, vtn_switch->selector)->def;
+ /* An accumulation of all conditions. Used for the default */
+ nir_ssa_def *any = NULL;
+
+ int i = 0;
+ list_for_each_entry(struct vtn_case, cse, &vtn_switch->cases, link) {
+ if (cse->is_default) {
+ conditions[i++] = NULL;
+ continue;
+ }
+
+ nir_ssa_def *cond = NULL;
+ nir_array_foreach(&cse->values, uint32_t, val) {
+ nir_ssa_def *is_val =
+ nir_ieq(&b->nb, sel, nir_imm_int(&b->nb, *val));
+
+ cond = cond ? nir_ior(&b->nb, cond, is_val) : is_val;
+ }
+
+ any = any ? nir_ior(&b->nb, any, cond) : cond;
+ conditions[i++] = cond;
+ }
+ assert(i == num_cases);
+
+ /* Now we can walk the list of cases and actually emit code */
+ i = 0;
+ list_for_each_entry(struct vtn_case, cse, &vtn_switch->cases, link) {
+ /* Figure out the condition */
+ nir_ssa_def *cond = conditions[i++];
+ if (cse->is_default) {
+ assert(cond == NULL);
+ cond = nir_inot(&b->nb, any);
+ }
+ /* Take fallthrough into account */
+ cond = nir_ior(&b->nb, cond, nir_load_var(&b->nb, fall_var));
+
+ nir_if *case_if = nir_if_create(b->nb.shader);
+ case_if->condition = nir_src_for_ssa(cond);
+ nir_cf_node_insert(b->nb.cursor, &case_if->cf_node);
+
+ bool has_break = false;
+ b->nb.cursor = nir_after_cf_list(&case_if->then_list);
+ nir_store_var(&b->nb, fall_var, nir_imm_int(&b->nb, NIR_TRUE), 1);
+ vtn_emit_cf_list(b, &cse->body, fall_var, &has_break, handler);
+ (void)has_break; /* We don't care */
+
+ b->nb.cursor = nir_after_cf_node(&case_if->cf_node);
+ }
+ assert(i == num_cases);
+
+ break;
+ }
+
+ default:
+ unreachable("Invalid CF node type");
+ }
+ }
+}
+
+void
+vtn_function_emit(struct vtn_builder *b, struct vtn_function *func,
+ vtn_instruction_handler instruction_handler)
+{
+ nir_builder_init(&b->nb, func->impl);
+ b->nb.cursor = nir_after_cf_list(&func->impl->body);
+ b->has_loop_continue = false;
+ b->phi_table = _mesa_hash_table_create(b, _mesa_hash_pointer,
+ _mesa_key_pointer_equal);
+
+ vtn_emit_cf_list(b, &func->body, NULL, NULL, instruction_handler);
+
+ vtn_foreach_instruction(b, func->start_block->label, func->end,
+ vtn_handle_phi_second_pass);
+
+ /* Continue blocks for loops get inserted before the body of the loop
+ * but instructions in the continue may use SSA defs in the loop body.
+ * Therefore, we need to repair SSA to insert the needed phi nodes.
+ */
+ if (b->has_loop_continue)
+ nir_repair_ssa_impl(func->impl);
+}
diff --git a/src/compiler/spirv/vtn_glsl450.c b/src/compiler/spirv/vtn_glsl450.c
new file mode 100644
index 00000000000..e05d28ffede
--- /dev/null
+++ b/src/compiler/spirv/vtn_glsl450.c
@@ -0,0 +1,666 @@
+/*
+ * Copyright © 2015 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ *
+ * Authors:
+ * Jason Ekstrand ([email protected])
+ *
+ */
+
+#include "vtn_private.h"
+#include "GLSL.std.450.h"
+
+#define M_PIf ((float) M_PI)
+#define M_PI_2f ((float) M_PI_2)
+#define M_PI_4f ((float) M_PI_4)
+
+static nir_ssa_def *
+build_mat2_det(nir_builder *b, nir_ssa_def *col[2])
+{
+ unsigned swiz[4] = {1, 0, 0, 0};
+ nir_ssa_def *p = nir_fmul(b, col[0], nir_swizzle(b, col[1], swiz, 2, true));
+ return nir_fsub(b, nir_channel(b, p, 0), nir_channel(b, p, 1));
+}
+
+static nir_ssa_def *
+build_mat3_det(nir_builder *b, nir_ssa_def *col[3])
+{
+ unsigned yzx[4] = {1, 2, 0, 0};
+ unsigned zxy[4] = {2, 0, 1, 0};
+
+ nir_ssa_def *prod0 =
+ nir_fmul(b, col[0],
+ nir_fmul(b, nir_swizzle(b, col[1], yzx, 3, true),
+ nir_swizzle(b, col[2], zxy, 3, true)));
+ nir_ssa_def *prod1 =
+ nir_fmul(b, col[0],
+ nir_fmul(b, nir_swizzle(b, col[1], zxy, 3, true),
+ nir_swizzle(b, col[2], yzx, 3, true)));
+
+ nir_ssa_def *diff = nir_fsub(b, prod0, prod1);
+
+ return nir_fadd(b, nir_channel(b, diff, 0),
+ nir_fadd(b, nir_channel(b, diff, 1),
+ nir_channel(b, diff, 2)));
+}
+
+static nir_ssa_def *
+build_mat4_det(nir_builder *b, nir_ssa_def **col)
+{
+ nir_ssa_def *subdet[4];
+ for (unsigned i = 0; i < 4; i++) {
+ unsigned swiz[3];
+ for (unsigned j = 0; j < 3; j++)
+ swiz[j] = j + (j >= i);
+
+ nir_ssa_def *subcol[3];
+ subcol[0] = nir_swizzle(b, col[1], swiz, 3, true);
+ subcol[1] = nir_swizzle(b, col[2], swiz, 3, true);
+ subcol[2] = nir_swizzle(b, col[3], swiz, 3, true);
+
+ subdet[i] = build_mat3_det(b, subcol);
+ }
+
+ nir_ssa_def *prod = nir_fmul(b, col[0], nir_vec(b, subdet, 4));
+
+ return nir_fadd(b, nir_fsub(b, nir_channel(b, prod, 0),
+ nir_channel(b, prod, 1)),
+ nir_fsub(b, nir_channel(b, prod, 2),
+ nir_channel(b, prod, 3)));
+}
+
+static nir_ssa_def *
+build_mat_det(struct vtn_builder *b, struct vtn_ssa_value *src)
+{
+ unsigned size = glsl_get_vector_elements(src->type);
+
+ nir_ssa_def *cols[4];
+ for (unsigned i = 0; i < size; i++)
+ cols[i] = src->elems[i]->def;
+
+ switch(size) {
+ case 2: return build_mat2_det(&b->nb, cols);
+ case 3: return build_mat3_det(&b->nb, cols);
+ case 4: return build_mat4_det(&b->nb, cols);
+ default:
+ unreachable("Invalid matrix size");
+ }
+}
+
+/* Computes the determinate of the submatrix given by taking src and
+ * removing the specified row and column.
+ */
+static nir_ssa_def *
+build_mat_subdet(struct nir_builder *b, struct vtn_ssa_value *src,
+ unsigned size, unsigned row, unsigned col)
+{
+ assert(row < size && col < size);
+ if (size == 2) {
+ return nir_channel(b, src->elems[1 - col]->def, 1 - row);
+ } else {
+ /* Swizzle to get all but the specified row */
+ unsigned swiz[3];
+ for (unsigned j = 0; j < 3; j++)
+ swiz[j] = j + (j >= row);
+
+ /* Grab all but the specified column */
+ nir_ssa_def *subcol[3];
+ for (unsigned j = 0; j < size; j++) {
+ if (j != col) {
+ subcol[j - (j > col)] = nir_swizzle(b, src->elems[j]->def,
+ swiz, size - 1, true);
+ }
+ }
+
+ if (size == 3) {
+ return build_mat2_det(b, subcol);
+ } else {
+ assert(size == 4);
+ return build_mat3_det(b, subcol);
+ }
+ }
+}
+
+static struct vtn_ssa_value *
+matrix_inverse(struct vtn_builder *b, struct vtn_ssa_value *src)
+{
+ nir_ssa_def *adj_col[4];
+ unsigned size = glsl_get_vector_elements(src->type);
+
+ /* Build up an adjugate matrix */
+ for (unsigned c = 0; c < size; c++) {
+ nir_ssa_def *elem[4];
+ for (unsigned r = 0; r < size; r++) {
+ elem[r] = build_mat_subdet(&b->nb, src, size, c, r);
+
+ if ((r + c) % 2)
+ elem[r] = nir_fneg(&b->nb, elem[r]);
+ }
+
+ adj_col[c] = nir_vec(&b->nb, elem, size);
+ }
+
+ nir_ssa_def *det_inv = nir_frcp(&b->nb, build_mat_det(b, src));
+
+ struct vtn_ssa_value *val = vtn_create_ssa_value(b, src->type);
+ for (unsigned i = 0; i < size; i++)
+ val->elems[i]->def = nir_fmul(&b->nb, adj_col[i], det_inv);
+
+ return val;
+}
+
+static nir_ssa_def*
+build_length(nir_builder *b, nir_ssa_def *vec)
+{
+ switch (vec->num_components) {
+ case 1: return nir_fsqrt(b, nir_fmul(b, vec, vec));
+ case 2: return nir_fsqrt(b, nir_fdot2(b, vec, vec));
+ case 3: return nir_fsqrt(b, nir_fdot3(b, vec, vec));
+ case 4: return nir_fsqrt(b, nir_fdot4(b, vec, vec));
+ default:
+ unreachable("Invalid number of components");
+ }
+}
+
+static inline nir_ssa_def *
+build_fclamp(nir_builder *b,
+ nir_ssa_def *x, nir_ssa_def *min_val, nir_ssa_def *max_val)
+{
+ return nir_fmin(b, nir_fmax(b, x, min_val), max_val);
+}
+
+/**
+ * Return e^x.
+ */
+static nir_ssa_def *
+build_exp(nir_builder *b, nir_ssa_def *x)
+{
+ return nir_fexp2(b, nir_fmul(b, x, nir_imm_float(b, M_LOG2E)));
+}
+
+/**
+ * Return ln(x) - the natural logarithm of x.
+ */
+static nir_ssa_def *
+build_log(nir_builder *b, nir_ssa_def *x)
+{
+ return nir_fmul(b, nir_flog2(b, x), nir_imm_float(b, 1.0 / M_LOG2E));
+}
+
+/**
+ * Approximate asin(x) by the formula:
+ * asin~(x) = sign(x) * (pi/2 - sqrt(1 - |x|) * (pi/2 + |x|(pi/4 - 1 + |x|(p0 + |x|p1))))
+ *
+ * which is correct to first order at x=0 and x=±1 regardless of the p
+ * coefficients but can be made second-order correct at both ends by selecting
+ * the fit coefficients appropriately. Different p coefficients can be used
+ * in the asin and acos implementation to minimize some relative error metric
+ * in each case.
+ */
+static nir_ssa_def *
+build_asin(nir_builder *b, nir_ssa_def *x, float p0, float p1)
+{
+ nir_ssa_def *abs_x = nir_fabs(b, x);
+ return nir_fmul(b, nir_fsign(b, x),
+ nir_fsub(b, nir_imm_float(b, M_PI_2f),
+ nir_fmul(b, nir_fsqrt(b, nir_fsub(b, nir_imm_float(b, 1.0f), abs_x)),
+ nir_fadd(b, nir_imm_float(b, M_PI_2f),
+ nir_fmul(b, abs_x,
+ nir_fadd(b, nir_imm_float(b, M_PI_4f - 1.0f),
+ nir_fmul(b, abs_x,
+ nir_fadd(b, nir_imm_float(b, p0),
+ nir_fmul(b, abs_x,
+ nir_imm_float(b, p1))))))))));
+}
+
+/**
+ * Compute xs[0] + xs[1] + xs[2] + ... using fadd.
+ */
+static nir_ssa_def *
+build_fsum(nir_builder *b, nir_ssa_def **xs, int terms)
+{
+ nir_ssa_def *accum = xs[0];
+
+ for (int i = 1; i < terms; i++)
+ accum = nir_fadd(b, accum, xs[i]);
+
+ return accum;
+}
+
+static nir_ssa_def *
+build_atan(nir_builder *b, nir_ssa_def *y_over_x)
+{
+ nir_ssa_def *abs_y_over_x = nir_fabs(b, y_over_x);
+ nir_ssa_def *one = nir_imm_float(b, 1.0f);
+
+ /*
+ * range-reduction, first step:
+ *
+ * / y_over_x if |y_over_x| <= 1.0;
+ * x = <
+ * \ 1.0 / y_over_x otherwise
+ */
+ nir_ssa_def *x = nir_fdiv(b, nir_fmin(b, abs_y_over_x, one),
+ nir_fmax(b, abs_y_over_x, one));
+
+ /*
+ * approximate atan by evaluating polynomial:
+ *
+ * x * 0.9999793128310355 - x^3 * 0.3326756418091246 +
+ * x^5 * 0.1938924977115610 - x^7 * 0.1173503194786851 +
+ * x^9 * 0.0536813784310406 - x^11 * 0.0121323213173444
+ */
+ nir_ssa_def *x_2 = nir_fmul(b, x, x);
+ nir_ssa_def *x_3 = nir_fmul(b, x_2, x);
+ nir_ssa_def *x_5 = nir_fmul(b, x_3, x_2);
+ nir_ssa_def *x_7 = nir_fmul(b, x_5, x_2);
+ nir_ssa_def *x_9 = nir_fmul(b, x_7, x_2);
+ nir_ssa_def *x_11 = nir_fmul(b, x_9, x_2);
+
+ nir_ssa_def *polynomial_terms[] = {
+ nir_fmul(b, x, nir_imm_float(b, 0.9999793128310355f)),
+ nir_fmul(b, x_3, nir_imm_float(b, -0.3326756418091246f)),
+ nir_fmul(b, x_5, nir_imm_float(b, 0.1938924977115610f)),
+ nir_fmul(b, x_7, nir_imm_float(b, -0.1173503194786851f)),
+ nir_fmul(b, x_9, nir_imm_float(b, 0.0536813784310406f)),
+ nir_fmul(b, x_11, nir_imm_float(b, -0.0121323213173444f)),
+ };
+
+ nir_ssa_def *tmp =
+ build_fsum(b, polynomial_terms, ARRAY_SIZE(polynomial_terms));
+
+ /* range-reduction fixup */
+ tmp = nir_fadd(b, tmp,
+ nir_fmul(b,
+ nir_b2f(b, nir_flt(b, one, abs_y_over_x)),
+ nir_fadd(b, nir_fmul(b, tmp,
+ nir_imm_float(b, -2.0f)),
+ nir_imm_float(b, M_PI_2f))));
+
+ /* sign fixup */
+ return nir_fmul(b, tmp, nir_fsign(b, y_over_x));
+}
+
+static nir_ssa_def *
+build_atan2(nir_builder *b, nir_ssa_def *y, nir_ssa_def *x)
+{
+ nir_ssa_def *zero = nir_imm_float(b, 0.0f);
+
+ /* If |x| >= 1.0e-8 * |y|: */
+ nir_ssa_def *condition =
+ nir_fge(b, nir_fabs(b, x),
+ nir_fmul(b, nir_imm_float(b, 1.0e-8f), nir_fabs(b, y)));
+
+ /* Then...call atan(y/x) and fix it up: */
+ nir_ssa_def *atan1 = build_atan(b, nir_fdiv(b, y, x));
+ nir_ssa_def *r_then =
+ nir_bcsel(b, nir_flt(b, x, zero),
+ nir_fadd(b, atan1,
+ nir_bcsel(b, nir_fge(b, y, zero),
+ nir_imm_float(b, M_PIf),
+ nir_imm_float(b, -M_PIf))),
+ atan1);
+
+ /* Else... */
+ nir_ssa_def *r_else =
+ nir_fmul(b, nir_fsign(b, y), nir_imm_float(b, M_PI_2f));
+
+ return nir_bcsel(b, condition, r_then, r_else);
+}
+
+static nir_ssa_def *
+build_frexp(nir_builder *b, nir_ssa_def *x, nir_ssa_def **exponent)
+{
+ nir_ssa_def *abs_x = nir_fabs(b, x);
+ nir_ssa_def *zero = nir_imm_float(b, 0.0f);
+
+ /* Single-precision floating-point values are stored as
+ * 1 sign bit;
+ * 8 exponent bits;
+ * 23 mantissa bits.
+ *
+ * An exponent shift of 23 will shift the mantissa out, leaving only the
+ * exponent and sign bit (which itself may be zero, if the absolute value
+ * was taken before the bitcast and shift.
+ */
+ nir_ssa_def *exponent_shift = nir_imm_int(b, 23);
+ nir_ssa_def *exponent_bias = nir_imm_int(b, -126);
+
+ nir_ssa_def *sign_mantissa_mask = nir_imm_int(b, 0x807fffffu);
+
+ /* Exponent of floating-point values in the range [0.5, 1.0). */
+ nir_ssa_def *exponent_value = nir_imm_int(b, 0x3f000000u);
+
+ nir_ssa_def *is_not_zero = nir_fne(b, abs_x, zero);
+
+ *exponent =
+ nir_iadd(b, nir_ushr(b, abs_x, exponent_shift),
+ nir_bcsel(b, is_not_zero, exponent_bias, zero));
+
+ return nir_ior(b, nir_iand(b, x, sign_mantissa_mask),
+ nir_bcsel(b, is_not_zero, exponent_value, zero));
+}
+
+static nir_op
+vtn_nir_alu_op_for_spirv_glsl_opcode(enum GLSLstd450 opcode)
+{
+ switch (opcode) {
+ case GLSLstd450Round: return nir_op_fround_even;
+ case GLSLstd450RoundEven: return nir_op_fround_even;
+ case GLSLstd450Trunc: return nir_op_ftrunc;
+ case GLSLstd450FAbs: return nir_op_fabs;
+ case GLSLstd450SAbs: return nir_op_iabs;
+ case GLSLstd450FSign: return nir_op_fsign;
+ case GLSLstd450SSign: return nir_op_isign;
+ case GLSLstd450Floor: return nir_op_ffloor;
+ case GLSLstd450Ceil: return nir_op_fceil;
+ case GLSLstd450Fract: return nir_op_ffract;
+ case GLSLstd450Sin: return nir_op_fsin;
+ case GLSLstd450Cos: return nir_op_fcos;
+ case GLSLstd450Pow: return nir_op_fpow;
+ case GLSLstd450Exp2: return nir_op_fexp2;
+ case GLSLstd450Log2: return nir_op_flog2;
+ case GLSLstd450Sqrt: return nir_op_fsqrt;
+ case GLSLstd450InverseSqrt: return nir_op_frsq;
+ case GLSLstd450FMin: return nir_op_fmin;
+ case GLSLstd450UMin: return nir_op_umin;
+ case GLSLstd450SMin: return nir_op_imin;
+ case GLSLstd450FMax: return nir_op_fmax;
+ case GLSLstd450UMax: return nir_op_umax;
+ case GLSLstd450SMax: return nir_op_imax;
+ case GLSLstd450FMix: return nir_op_flrp;
+ case GLSLstd450Fma: return nir_op_ffma;
+ case GLSLstd450Ldexp: return nir_op_ldexp;
+ case GLSLstd450FindILsb: return nir_op_find_lsb;
+ case GLSLstd450FindSMsb: return nir_op_ifind_msb;
+ case GLSLstd450FindUMsb: return nir_op_ufind_msb;
+
+ /* Packing/Unpacking functions */
+ case GLSLstd450PackSnorm4x8: return nir_op_pack_snorm_4x8;
+ case GLSLstd450PackUnorm4x8: return nir_op_pack_unorm_4x8;
+ case GLSLstd450PackSnorm2x16: return nir_op_pack_snorm_2x16;
+ case GLSLstd450PackUnorm2x16: return nir_op_pack_unorm_2x16;
+ case GLSLstd450PackHalf2x16: return nir_op_pack_half_2x16;
+ case GLSLstd450UnpackSnorm4x8: return nir_op_unpack_snorm_4x8;
+ case GLSLstd450UnpackUnorm4x8: return nir_op_unpack_unorm_4x8;
+ case GLSLstd450UnpackSnorm2x16: return nir_op_unpack_snorm_2x16;
+ case GLSLstd450UnpackUnorm2x16: return nir_op_unpack_unorm_2x16;
+ case GLSLstd450UnpackHalf2x16: return nir_op_unpack_half_2x16;
+
+ default:
+ unreachable("No NIR equivalent");
+ }
+}
+
+static void
+handle_glsl450_alu(struct vtn_builder *b, enum GLSLstd450 entrypoint,
+ const uint32_t *w, unsigned count)
+{
+ struct nir_builder *nb = &b->nb;
+ const struct glsl_type *dest_type =
+ vtn_value(b, w[1], vtn_value_type_type)->type->type;
+
+ struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
+ val->ssa = vtn_create_ssa_value(b, dest_type);
+
+ /* Collect the various SSA sources */
+ unsigned num_inputs = count - 5;
+ nir_ssa_def *src[3] = { NULL, };
+ for (unsigned i = 0; i < num_inputs; i++)
+ src[i] = vtn_ssa_value(b, w[i + 5])->def;
+
+ switch (entrypoint) {
+ case GLSLstd450Radians:
+ val->ssa->def = nir_fmul(nb, src[0], nir_imm_float(nb, 0.01745329251));
+ return;
+ case GLSLstd450Degrees:
+ val->ssa->def = nir_fmul(nb, src[0], nir_imm_float(nb, 57.2957795131));
+ return;
+ case GLSLstd450Tan:
+ val->ssa->def = nir_fdiv(nb, nir_fsin(nb, src[0]),
+ nir_fcos(nb, src[0]));
+ return;
+
+ case GLSLstd450Modf: {
+ nir_ssa_def *sign = nir_fsign(nb, src[0]);
+ nir_ssa_def *abs = nir_fabs(nb, src[0]);
+ val->ssa->def = nir_fmul(nb, sign, nir_ffract(nb, abs));
+ nir_store_deref_var(nb, vtn_nir_deref(b, w[6]),
+ nir_fmul(nb, sign, nir_ffloor(nb, abs)), 0xf);
+ return;
+ }
+
+ case GLSLstd450ModfStruct: {
+ nir_ssa_def *sign = nir_fsign(nb, src[0]);
+ nir_ssa_def *abs = nir_fabs(nb, src[0]);
+ assert(glsl_type_is_struct(val->ssa->type));
+ val->ssa->elems[0]->def = nir_fmul(nb, sign, nir_ffract(nb, abs));
+ val->ssa->elems[1]->def = nir_fmul(nb, sign, nir_ffloor(nb, abs));
+ return;
+ }
+
+ case GLSLstd450Step:
+ val->ssa->def = nir_sge(nb, src[1], src[0]);
+ return;
+
+ case GLSLstd450Length:
+ val->ssa->def = build_length(nb, src[0]);
+ return;
+ case GLSLstd450Distance:
+ val->ssa->def = build_length(nb, nir_fsub(nb, src[0], src[1]));
+ return;
+ case GLSLstd450Normalize:
+ val->ssa->def = nir_fdiv(nb, src[0], build_length(nb, src[0]));
+ return;
+
+ case GLSLstd450Exp:
+ val->ssa->def = build_exp(nb, src[0]);
+ return;
+
+ case GLSLstd450Log:
+ val->ssa->def = build_log(nb, src[0]);
+ return;
+
+ case GLSLstd450FClamp:
+ val->ssa->def = build_fclamp(nb, src[0], src[1], src[2]);
+ return;
+ case GLSLstd450UClamp:
+ val->ssa->def = nir_umin(nb, nir_umax(nb, src[0], src[1]), src[2]);
+ return;
+ case GLSLstd450SClamp:
+ val->ssa->def = nir_imin(nb, nir_imax(nb, src[0], src[1]), src[2]);
+ return;
+
+ case GLSLstd450Cross: {
+ unsigned yzx[4] = { 1, 2, 0, 0 };
+ unsigned zxy[4] = { 2, 0, 1, 0 };
+ val->ssa->def =
+ nir_fsub(nb, nir_fmul(nb, nir_swizzle(nb, src[0], yzx, 3, true),
+ nir_swizzle(nb, src[1], zxy, 3, true)),
+ nir_fmul(nb, nir_swizzle(nb, src[0], zxy, 3, true),
+ nir_swizzle(nb, src[1], yzx, 3, true)));
+ return;
+ }
+
+ case GLSLstd450SmoothStep: {
+ /* t = clamp((x - edge0) / (edge1 - edge0), 0, 1) */
+ nir_ssa_def *t =
+ build_fclamp(nb, nir_fdiv(nb, nir_fsub(nb, src[2], src[0]),
+ nir_fsub(nb, src[1], src[0])),
+ nir_imm_float(nb, 0.0), nir_imm_float(nb, 1.0));
+ /* result = t * t * (3 - 2 * t) */
+ val->ssa->def =
+ nir_fmul(nb, t, nir_fmul(nb, t,
+ nir_fsub(nb, nir_imm_float(nb, 3.0),
+ nir_fmul(nb, nir_imm_float(nb, 2.0), t))));
+ return;
+ }
+
+ case GLSLstd450FaceForward:
+ val->ssa->def =
+ nir_bcsel(nb, nir_flt(nb, nir_fdot(nb, src[2], src[1]),
+ nir_imm_float(nb, 0.0)),
+ src[0], nir_fneg(nb, src[0]));
+ return;
+
+ case GLSLstd450Reflect:
+ /* I - 2 * dot(N, I) * N */
+ val->ssa->def =
+ nir_fsub(nb, src[0], nir_fmul(nb, nir_imm_float(nb, 2.0),
+ nir_fmul(nb, nir_fdot(nb, src[0], src[1]),
+ src[1])));
+ return;
+
+ case GLSLstd450Refract: {
+ nir_ssa_def *I = src[0];
+ nir_ssa_def *N = src[1];
+ nir_ssa_def *eta = src[2];
+ nir_ssa_def *n_dot_i = nir_fdot(nb, N, I);
+ nir_ssa_def *one = nir_imm_float(nb, 1.0);
+ nir_ssa_def *zero = nir_imm_float(nb, 0.0);
+ /* k = 1.0 - eta * eta * (1.0 - dot(N, I) * dot(N, I)) */
+ nir_ssa_def *k =
+ nir_fsub(nb, one, nir_fmul(nb, eta, nir_fmul(nb, eta,
+ nir_fsub(nb, one, nir_fmul(nb, n_dot_i, n_dot_i)))));
+ nir_ssa_def *result =
+ nir_fsub(nb, nir_fmul(nb, eta, I),
+ nir_fmul(nb, nir_fadd(nb, nir_fmul(nb, eta, n_dot_i),
+ nir_fsqrt(nb, k)), N));
+ /* XXX: bcsel, or if statement? */
+ val->ssa->def = nir_bcsel(nb, nir_flt(nb, k, zero), zero, result);
+ return;
+ }
+
+ case GLSLstd450Sinh:
+ /* 0.5 * (e^x - e^(-x)) */
+ val->ssa->def =
+ nir_fmul(nb, nir_imm_float(nb, 0.5f),
+ nir_fsub(nb, build_exp(nb, src[0]),
+ build_exp(nb, nir_fneg(nb, src[0]))));
+ return;
+
+ case GLSLstd450Cosh:
+ /* 0.5 * (e^x + e^(-x)) */
+ val->ssa->def =
+ nir_fmul(nb, nir_imm_float(nb, 0.5f),
+ nir_fadd(nb, build_exp(nb, src[0]),
+ build_exp(nb, nir_fneg(nb, src[0]))));
+ return;
+
+ case GLSLstd450Tanh:
+ /* (0.5 * (e^x - e^(-x))) / (0.5 * (e^x + e^(-x))) */
+ val->ssa->def =
+ nir_fdiv(nb, nir_fmul(nb, nir_imm_float(nb, 0.5f),
+ nir_fsub(nb, build_exp(nb, src[0]),
+ build_exp(nb, nir_fneg(nb, src[0])))),
+ nir_fmul(nb, nir_imm_float(nb, 0.5f),
+ nir_fadd(nb, build_exp(nb, src[0]),
+ build_exp(nb, nir_fneg(nb, src[0])))));
+ return;
+
+ case GLSLstd450Asinh:
+ val->ssa->def = nir_fmul(nb, nir_fsign(nb, src[0]),
+ build_log(nb, nir_fadd(nb, nir_fabs(nb, src[0]),
+ nir_fsqrt(nb, nir_fadd(nb, nir_fmul(nb, src[0], src[0]),
+ nir_imm_float(nb, 1.0f))))));
+ return;
+ case GLSLstd450Acosh:
+ val->ssa->def = build_log(nb, nir_fadd(nb, src[0],
+ nir_fsqrt(nb, nir_fsub(nb, nir_fmul(nb, src[0], src[0]),
+ nir_imm_float(nb, 1.0f)))));
+ return;
+ case GLSLstd450Atanh: {
+ nir_ssa_def *one = nir_imm_float(nb, 1.0);
+ val->ssa->def = nir_fmul(nb, nir_imm_float(nb, 0.5f),
+ build_log(nb, nir_fdiv(nb, nir_fadd(nb, one, src[0]),
+ nir_fsub(nb, one, src[0]))));
+ return;
+ }
+
+ case GLSLstd450Asin:
+ val->ssa->def = build_asin(nb, src[0], 0.086566724, -0.03102955);
+ return;
+
+ case GLSLstd450Acos:
+ val->ssa->def = nir_fsub(nb, nir_imm_float(nb, M_PI_2f),
+ build_asin(nb, src[0], 0.08132463, -0.02363318));
+ return;
+
+ case GLSLstd450Atan:
+ val->ssa->def = build_atan(nb, src[0]);
+ return;
+
+ case GLSLstd450Atan2:
+ val->ssa->def = build_atan2(nb, src[0], src[1]);
+ return;
+
+ case GLSLstd450Frexp: {
+ nir_ssa_def *exponent;
+ val->ssa->def = build_frexp(nb, src[0], &exponent);
+ nir_store_deref_var(nb, vtn_nir_deref(b, w[6]), exponent, 0xf);
+ return;
+ }
+
+ case GLSLstd450FrexpStruct: {
+ assert(glsl_type_is_struct(val->ssa->type));
+ val->ssa->elems[0]->def = build_frexp(nb, src[0],
+ &val->ssa->elems[1]->def);
+ return;
+ }
+
+ default:
+ val->ssa->def =
+ nir_build_alu(&b->nb, vtn_nir_alu_op_for_spirv_glsl_opcode(entrypoint),
+ src[0], src[1], src[2], NULL);
+ return;
+ }
+}
+
+bool
+vtn_handle_glsl450_instruction(struct vtn_builder *b, uint32_t ext_opcode,
+ const uint32_t *w, unsigned count)
+{
+ switch ((enum GLSLstd450)ext_opcode) {
+ case GLSLstd450Determinant: {
+ struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
+ val->ssa = rzalloc(b, struct vtn_ssa_value);
+ val->ssa->type = vtn_value(b, w[1], vtn_value_type_type)->type->type;
+ val->ssa->def = build_mat_det(b, vtn_ssa_value(b, w[5]));
+ break;
+ }
+
+ case GLSLstd450MatrixInverse: {
+ struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
+ val->ssa = matrix_inverse(b, vtn_ssa_value(b, w[5]));
+ break;
+ }
+
+ case GLSLstd450InterpolateAtCentroid:
+ case GLSLstd450InterpolateAtSample:
+ case GLSLstd450InterpolateAtOffset:
+ unreachable("Unhandled opcode");
+
+ default:
+ handle_glsl450_alu(b, (enum GLSLstd450)ext_opcode, w, count);
+ }
+
+ return true;
+}
diff --git a/src/compiler/spirv/vtn_private.h b/src/compiler/spirv/vtn_private.h
new file mode 100644
index 00000000000..3840d8c4b65
--- /dev/null
+++ b/src/compiler/spirv/vtn_private.h
@@ -0,0 +1,484 @@
+/*
+ * Copyright © 2015 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ *
+ * Authors:
+ * Jason Ekstrand ([email protected])
+ *
+ */
+
+#include "nir/nir.h"
+#include "nir/nir_builder.h"
+#include "nir/nir_array.h"
+#include "nir_spirv.h"
+#include "spirv.h"
+
+struct vtn_builder;
+struct vtn_decoration;
+
+enum vtn_value_type {
+ vtn_value_type_invalid = 0,
+ vtn_value_type_undef,
+ vtn_value_type_string,
+ vtn_value_type_decoration_group,
+ vtn_value_type_type,
+ vtn_value_type_constant,
+ vtn_value_type_access_chain,
+ vtn_value_type_function,
+ vtn_value_type_block,
+ vtn_value_type_ssa,
+ vtn_value_type_extension,
+ vtn_value_type_image_pointer,
+ vtn_value_type_sampled_image,
+};
+
+enum vtn_branch_type {
+ vtn_branch_type_none,
+ vtn_branch_type_switch_break,
+ vtn_branch_type_switch_fallthrough,
+ vtn_branch_type_loop_break,
+ vtn_branch_type_loop_continue,
+ vtn_branch_type_discard,
+ vtn_branch_type_return,
+};
+
+enum vtn_cf_node_type {
+ vtn_cf_node_type_block,
+ vtn_cf_node_type_if,
+ vtn_cf_node_type_loop,
+ vtn_cf_node_type_switch,
+};
+
+struct vtn_cf_node {
+ struct list_head link;
+ enum vtn_cf_node_type type;
+};
+
+struct vtn_loop {
+ struct vtn_cf_node node;
+
+ /* The main body of the loop */
+ struct list_head body;
+
+ /* The "continue" part of the loop. This gets executed after the body
+ * and is where you go when you hit a continue.
+ */
+ struct list_head cont_body;
+
+ SpvLoopControlMask control;
+};
+
+struct vtn_if {
+ struct vtn_cf_node node;
+
+ uint32_t condition;
+
+ enum vtn_branch_type then_type;
+ struct list_head then_body;
+
+ enum vtn_branch_type else_type;
+ struct list_head else_body;
+
+ SpvSelectionControlMask control;
+};
+
+struct vtn_case {
+ struct list_head link;
+
+ struct list_head body;
+
+ /* The block that starts this case */
+ struct vtn_block *start_block;
+
+ /* The fallthrough case, if any */
+ struct vtn_case *fallthrough;
+
+ /* The uint32_t values that map to this case */
+ nir_array values;
+
+ /* True if this is the default case */
+ bool is_default;
+
+ /* Initialized to false; used when sorting the list of cases */
+ bool visited;
+};
+
+struct vtn_switch {
+ struct vtn_cf_node node;
+
+ uint32_t selector;
+
+ struct list_head cases;
+};
+
+struct vtn_block {
+ struct vtn_cf_node node;
+
+ /** A pointer to the label instruction */
+ const uint32_t *label;
+
+ /** A pointer to the merge instruction (or NULL if non exists) */
+ const uint32_t *merge;
+
+ /** A pointer to the branch instruction that ends this block */
+ const uint32_t *branch;
+
+ enum vtn_branch_type branch_type;
+
+ /** Points to the loop that this block starts (if it starts a loop) */
+ struct vtn_loop *loop;
+
+ /** Points to the switch case started by this block (if any) */
+ struct vtn_case *switch_case;
+
+ /** The last block in this SPIR-V block. */
+ nir_block *end_block;
+};
+
+struct vtn_function {
+ struct exec_node node;
+
+ nir_function_impl *impl;
+ struct vtn_block *start_block;
+
+ struct list_head body;
+
+ const uint32_t *end;
+
+ SpvFunctionControlMask control;
+};
+
+typedef bool (*vtn_instruction_handler)(struct vtn_builder *, uint32_t,
+ const uint32_t *, unsigned);
+
+void vtn_build_cfg(struct vtn_builder *b, const uint32_t *words,
+ const uint32_t *end);
+void vtn_function_emit(struct vtn_builder *b, struct vtn_function *func,
+ vtn_instruction_handler instruction_handler);
+
+const uint32_t *
+vtn_foreach_instruction(struct vtn_builder *b, const uint32_t *start,
+ const uint32_t *end, vtn_instruction_handler handler);
+
+struct vtn_ssa_value {
+ union {
+ nir_ssa_def *def;
+ struct vtn_ssa_value **elems;
+ };
+
+ /* For matrices, if this is non-NULL, then this value is actually the
+ * transpose of some other value. The value that `transposed` points to
+ * always dominates this value.
+ */
+ struct vtn_ssa_value *transposed;
+
+ const struct glsl_type *type;
+};
+
+struct vtn_type {
+ const struct glsl_type *type;
+
+ /* The value that declares this type. Used for finding decorations */
+ struct vtn_value *val;
+
+ /* for matrices, whether the matrix is stored row-major */
+ bool row_major;
+
+ /* for structs, the offset of each member */
+ unsigned *offsets;
+
+ /* for structs, whether it was decorated as a "non-SSBO-like" block */
+ bool block;
+
+ /* for structs, whether it was decorated as an "SSBO-like" block */
+ bool buffer_block;
+
+ /* for structs with block == true, whether this is a builtin block (i.e. a
+ * block that contains only builtins).
+ */
+ bool builtin_block;
+
+ /* Image format for image_load_store type images */
+ unsigned image_format;
+
+ /* Access qualifier for storage images */
+ SpvAccessQualifier access_qualifier;
+
+ /* for arrays and matrices, the array stride */
+ unsigned stride;
+
+ /* for arrays, the vtn_type for the elements of the array */
+ struct vtn_type *array_element;
+
+ /* for structures, the vtn_type for each member */
+ struct vtn_type **members;
+
+ /* Whether this type, or a parent type, has been decorated as a builtin */
+ bool is_builtin;
+
+ SpvBuiltIn builtin;
+};
+
+struct vtn_variable;
+
+enum vtn_access_mode {
+ vtn_access_mode_id,
+ vtn_access_mode_literal,
+};
+
+struct vtn_access_link {
+ enum vtn_access_mode mode;
+ uint32_t id;
+};
+
+struct vtn_access_chain {
+ struct vtn_variable *var;
+
+ uint32_t length;
+
+ /* Struct elements and array offsets */
+ struct vtn_access_link link[0];
+};
+
+enum vtn_variable_mode {
+ vtn_variable_mode_local,
+ vtn_variable_mode_global,
+ vtn_variable_mode_param,
+ vtn_variable_mode_ubo,
+ vtn_variable_mode_ssbo,
+ vtn_variable_mode_push_constant,
+ vtn_variable_mode_image,
+ vtn_variable_mode_sampler,
+ vtn_variable_mode_workgroup,
+ vtn_variable_mode_input,
+ vtn_variable_mode_output,
+};
+
+struct vtn_variable {
+ enum vtn_variable_mode mode;
+
+ struct vtn_type *type;
+
+ unsigned descriptor_set;
+ unsigned binding;
+
+ nir_variable *var;
+ nir_variable **members;
+
+ struct vtn_access_chain chain;
+};
+
+struct vtn_image_pointer {
+ struct vtn_access_chain *image;
+ nir_ssa_def *coord;
+ nir_ssa_def *sample;
+};
+
+struct vtn_sampled_image {
+ struct vtn_access_chain *image; /* Image or array of images */
+ struct vtn_access_chain *sampler; /* Sampler */
+};
+
+struct vtn_value {
+ enum vtn_value_type value_type;
+ const char *name;
+ struct vtn_decoration *decoration;
+ union {
+ void *ptr;
+ char *str;
+ struct vtn_type *type;
+ struct {
+ nir_constant *constant;
+ const struct glsl_type *const_type;
+ };
+ struct vtn_access_chain *access_chain;
+ struct vtn_image_pointer *image;
+ struct vtn_sampled_image *sampled_image;
+ struct vtn_function *func;
+ struct vtn_block *block;
+ struct vtn_ssa_value *ssa;
+ vtn_instruction_handler ext_handler;
+ };
+};
+
+#define VTN_DEC_DECORATION -1
+#define VTN_DEC_EXECUTION_MODE -2
+#define VTN_DEC_STRUCT_MEMBER0 0
+
+struct vtn_decoration {
+ struct vtn_decoration *next;
+
+ /* Specifies how to apply this decoration. Negative values represent a
+ * decoration or execution mode. (See the VTN_DEC_ #defines above.)
+ * Non-negative values specify that it applies to a structure member.
+ */
+ int scope;
+
+ const uint32_t *literals;
+ struct vtn_value *group;
+
+ union {
+ SpvDecoration decoration;
+ SpvExecutionMode exec_mode;
+ };
+};
+
+struct vtn_builder {
+ nir_builder nb;
+
+ nir_shader *shader;
+ nir_function_impl *impl;
+ struct vtn_block *block;
+
+ /* Current file, line, and column. Useful for debugging. Set
+ * automatically by vtn_foreach_instruction.
+ */
+ char *file;
+ int line, col;
+
+ /*
+ * In SPIR-V, constants are global, whereas in NIR, the load_const
+ * instruction we use is per-function. So while we parse each function, we
+ * keep a hash table of constants we've resolved to nir_ssa_value's so
+ * far, and we lazily resolve them when we see them used in a function.
+ */
+ struct hash_table *const_table;
+
+ /*
+ * Map from phi instructions (pointer to the start of the instruction)
+ * to the variable corresponding to it.
+ */
+ struct hash_table *phi_table;
+
+ unsigned num_specializations;
+ struct nir_spirv_specialization *specializations;
+
+ unsigned value_id_bound;
+ struct vtn_value *values;
+
+ gl_shader_stage entry_point_stage;
+ const char *entry_point_name;
+ struct vtn_value *entry_point;
+ bool origin_upper_left;
+
+ struct vtn_function *func;
+ struct exec_list functions;
+
+ /* Current function parameter index */
+ unsigned func_param_idx;
+
+ bool has_loop_continue;
+};
+
+static inline struct vtn_value *
+vtn_push_value(struct vtn_builder *b, uint32_t value_id,
+ enum vtn_value_type value_type)
+{
+ assert(value_id < b->value_id_bound);
+ assert(b->values[value_id].value_type == vtn_value_type_invalid);
+
+ b->values[value_id].value_type = value_type;
+
+ return &b->values[value_id];
+}
+
+static inline struct vtn_value *
+vtn_untyped_value(struct vtn_builder *b, uint32_t value_id)
+{
+ assert(value_id < b->value_id_bound);
+ return &b->values[value_id];
+}
+
+static inline struct vtn_value *
+vtn_value(struct vtn_builder *b, uint32_t value_id,
+ enum vtn_value_type value_type)
+{
+ struct vtn_value *val = vtn_untyped_value(b, value_id);
+ assert(val->value_type == value_type);
+ return val;
+}
+
+struct vtn_ssa_value *vtn_ssa_value(struct vtn_builder *b, uint32_t value_id);
+
+struct vtn_ssa_value *vtn_create_ssa_value(struct vtn_builder *b,
+ const struct glsl_type *type);
+
+struct vtn_ssa_value *vtn_ssa_transpose(struct vtn_builder *b,
+ struct vtn_ssa_value *src);
+
+nir_ssa_def *vtn_vector_extract(struct vtn_builder *b, nir_ssa_def *src,
+ unsigned index);
+nir_ssa_def *vtn_vector_extract_dynamic(struct vtn_builder *b, nir_ssa_def *src,
+ nir_ssa_def *index);
+nir_ssa_def *vtn_vector_insert(struct vtn_builder *b, nir_ssa_def *src,
+ nir_ssa_def *insert, unsigned index);
+nir_ssa_def *vtn_vector_insert_dynamic(struct vtn_builder *b, nir_ssa_def *src,
+ nir_ssa_def *insert, nir_ssa_def *index);
+
+nir_deref_var *vtn_nir_deref(struct vtn_builder *b, uint32_t id);
+
+nir_deref_var *vtn_access_chain_to_deref(struct vtn_builder *b,
+ struct vtn_access_chain *chain);
+nir_ssa_def *
+vtn_access_chain_to_offset(struct vtn_builder *b,
+ struct vtn_access_chain *chain,
+ nir_ssa_def **index_out, struct vtn_type **type_out,
+ unsigned *end_idx_out, bool stop_at_matrix);
+
+struct vtn_ssa_value *vtn_local_load(struct vtn_builder *b, nir_deref_var *src);
+
+void vtn_local_store(struct vtn_builder *b, struct vtn_ssa_value *src,
+ nir_deref_var *dest);
+
+struct vtn_ssa_value *
+vtn_variable_load(struct vtn_builder *b, struct vtn_access_chain *src);
+
+void vtn_variable_store(struct vtn_builder *b, struct vtn_ssa_value *src,
+ struct vtn_access_chain *dest);
+
+void vtn_handle_variables(struct vtn_builder *b, SpvOp opcode,
+ const uint32_t *w, unsigned count);
+
+
+typedef void (*vtn_decoration_foreach_cb)(struct vtn_builder *,
+ struct vtn_value *,
+ int member,
+ const struct vtn_decoration *,
+ void *);
+
+void vtn_foreach_decoration(struct vtn_builder *b, struct vtn_value *value,
+ vtn_decoration_foreach_cb cb, void *data);
+
+typedef void (*vtn_execution_mode_foreach_cb)(struct vtn_builder *,
+ struct vtn_value *,
+ const struct vtn_decoration *,
+ void *);
+
+void vtn_foreach_execution_mode(struct vtn_builder *b, struct vtn_value *value,
+ vtn_execution_mode_foreach_cb cb, void *data);
+
+nir_op vtn_nir_alu_op_for_spirv_opcode(SpvOp opcode, bool *swap);
+
+void vtn_handle_alu(struct vtn_builder *b, SpvOp opcode,
+ const uint32_t *w, unsigned count);
+
+bool vtn_handle_glsl450_instruction(struct vtn_builder *b, uint32_t ext_opcode,
+ const uint32_t *words, unsigned count);
diff --git a/src/compiler/spirv/vtn_variables.c b/src/compiler/spirv/vtn_variables.c
new file mode 100644
index 00000000000..3cbac1e5da8
--- /dev/null
+++ b/src/compiler/spirv/vtn_variables.c
@@ -0,0 +1,1415 @@
+/*
+ * Copyright © 2015 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ *
+ * Authors:
+ * Jason Ekstrand ([email protected])
+ *
+ */
+
+#include "vtn_private.h"
+
+static struct vtn_access_chain *
+vtn_access_chain_extend(struct vtn_builder *b, struct vtn_access_chain *old,
+ unsigned new_ids)
+{
+ struct vtn_access_chain *chain;
+
+ unsigned new_len = old->length + new_ids;
+ chain = ralloc_size(b, sizeof(*chain) + new_len * sizeof(chain->link[0]));
+
+ chain->var = old->var;
+ chain->length = new_len;
+
+ for (unsigned i = 0; i < old->length; i++)
+ chain->link[i] = old->link[i];
+
+ return chain;
+}
+
+static nir_ssa_def *
+vtn_access_link_as_ssa(struct vtn_builder *b, struct vtn_access_link link,
+ unsigned stride)
+{
+ assert(stride > 0);
+ if (link.mode == vtn_access_mode_literal) {
+ return nir_imm_int(&b->nb, link.id * stride);
+ } else if (stride == 1) {
+ return vtn_ssa_value(b, link.id)->def;
+ } else {
+ return nir_imul(&b->nb, vtn_ssa_value(b, link.id)->def,
+ nir_imm_int(&b->nb, stride));
+ }
+}
+
+static struct vtn_type *
+vtn_access_chain_tail_type(struct vtn_builder *b,
+ struct vtn_access_chain *chain)
+{
+ struct vtn_type *type = chain->var->type;
+ for (unsigned i = 0; i < chain->length; i++) {
+ if (glsl_type_is_struct(type->type)) {
+ assert(chain->link[i].mode == vtn_access_mode_literal);
+ type = type->members[chain->link[i].id];
+ } else {
+ type = type->array_element;
+ }
+ }
+ return type;
+}
+
+/* Crawls a chain of array derefs and rewrites the types so that the
+ * lengths stay the same but the terminal type is the one given by
+ * tail_type. This is useful for split structures.
+ */
+static void
+rewrite_deref_types(nir_deref *deref, const struct glsl_type *type)
+{
+ deref->type = type;
+ if (deref->child) {
+ assert(deref->child->deref_type == nir_deref_type_array);
+ assert(glsl_type_is_array(deref->type));
+ rewrite_deref_types(deref->child, glsl_get_array_element(type));
+ }
+}
+
+nir_deref_var *
+vtn_access_chain_to_deref(struct vtn_builder *b, struct vtn_access_chain *chain)
+{
+ nir_deref_var *deref_var;
+ if (chain->var->var) {
+ deref_var = nir_deref_var_create(b, chain->var->var);
+ } else {
+ assert(chain->var->members);
+ /* Create the deref_var manually. It will get filled out later. */
+ deref_var = rzalloc(b, nir_deref_var);
+ deref_var->deref.deref_type = nir_deref_type_var;
+ }
+
+ struct vtn_type *deref_type = chain->var->type;
+ nir_deref *tail = &deref_var->deref;
+ nir_variable **members = chain->var->members;
+
+ for (unsigned i = 0; i < chain->length; i++) {
+ enum glsl_base_type base_type = glsl_get_base_type(deref_type->type);
+ switch (base_type) {
+ case GLSL_TYPE_UINT:
+ case GLSL_TYPE_INT:
+ case GLSL_TYPE_FLOAT:
+ case GLSL_TYPE_DOUBLE:
+ case GLSL_TYPE_BOOL:
+ case GLSL_TYPE_ARRAY: {
+ deref_type = deref_type->array_element;
+
+ nir_deref_array *deref_arr = nir_deref_array_create(b);
+ deref_arr->deref.type = deref_type->type;
+
+ if (chain->link[i].mode == vtn_access_mode_literal) {
+ deref_arr->deref_array_type = nir_deref_array_type_direct;
+ deref_arr->base_offset = chain->link[i].id;
+ } else {
+ assert(chain->link[i].mode == vtn_access_mode_id);
+ deref_arr->deref_array_type = nir_deref_array_type_indirect;
+ deref_arr->base_offset = 0;
+ deref_arr->indirect =
+ nir_src_for_ssa(vtn_ssa_value(b, chain->link[i].id)->def);
+ }
+ tail->child = &deref_arr->deref;
+ tail = tail->child;
+ break;
+ }
+
+ case GLSL_TYPE_STRUCT: {
+ assert(chain->link[i].mode == vtn_access_mode_literal);
+ unsigned idx = chain->link[i].id;
+ deref_type = deref_type->members[idx];
+ if (members) {
+ /* This is a pre-split structure. */
+ deref_var->var = members[idx];
+ rewrite_deref_types(&deref_var->deref, members[idx]->type);
+ assert(tail->type == deref_type->type);
+ members = NULL;
+ } else {
+ nir_deref_struct *deref_struct = nir_deref_struct_create(b, idx);
+ deref_struct->deref.type = deref_type->type;
+ tail->child = &deref_struct->deref;
+ tail = tail->child;
+ }
+ break;
+ }
+ default:
+ unreachable("Invalid type for deref");
+ }
+ }
+
+ assert(members == NULL);
+ return deref_var;
+}
+
+static void
+_vtn_local_load_store(struct vtn_builder *b, bool load, nir_deref_var *deref,
+ nir_deref *tail, struct vtn_ssa_value *inout)
+{
+ /* The deref tail may contain a deref to select a component of a vector (in
+ * other words, it might not be an actual tail) so we have to save it away
+ * here since we overwrite it later.
+ */
+ nir_deref *old_child = tail->child;
+
+ if (glsl_type_is_vector_or_scalar(tail->type)) {
+ /* Terminate the deref chain in case there is one more link to pick
+ * off a component of the vector.
+ */
+ tail->child = NULL;
+
+ nir_intrinsic_op op = load ? nir_intrinsic_load_var :
+ nir_intrinsic_store_var;
+
+ nir_intrinsic_instr *intrin = nir_intrinsic_instr_create(b->shader, op);
+ intrin->variables[0] =
+ nir_deref_as_var(nir_copy_deref(intrin, &deref->deref));
+ intrin->num_components = glsl_get_vector_elements(tail->type);
+
+ if (load) {
+ nir_ssa_dest_init(&intrin->instr, &intrin->dest,
+ intrin->num_components,
+ glsl_get_bit_size(glsl_get_base_type(tail->type)),
+ NULL);
+ inout->def = &intrin->dest.ssa;
+ } else {
+ nir_intrinsic_set_write_mask(intrin, (1 << intrin->num_components) - 1);
+ intrin->src[0] = nir_src_for_ssa(inout->def);
+ }
+
+ nir_builder_instr_insert(&b->nb, &intrin->instr);
+ } else if (glsl_get_base_type(tail->type) == GLSL_TYPE_ARRAY ||
+ glsl_type_is_matrix(tail->type)) {
+ unsigned elems = glsl_get_length(tail->type);
+ nir_deref_array *deref_arr = nir_deref_array_create(b);
+ deref_arr->deref_array_type = nir_deref_array_type_direct;
+ deref_arr->deref.type = glsl_get_array_element(tail->type);
+ tail->child = &deref_arr->deref;
+ for (unsigned i = 0; i < elems; i++) {
+ deref_arr->base_offset = i;
+ _vtn_local_load_store(b, load, deref, tail->child, inout->elems[i]);
+ }
+ } else {
+ assert(glsl_get_base_type(tail->type) == GLSL_TYPE_STRUCT);
+ unsigned elems = glsl_get_length(tail->type);
+ nir_deref_struct *deref_struct = nir_deref_struct_create(b, 0);
+ tail->child = &deref_struct->deref;
+ for (unsigned i = 0; i < elems; i++) {
+ deref_struct->index = i;
+ deref_struct->deref.type = glsl_get_struct_field(tail->type, i);
+ _vtn_local_load_store(b, load, deref, tail->child, inout->elems[i]);
+ }
+ }
+
+ tail->child = old_child;
+}
+
+nir_deref_var *
+vtn_nir_deref(struct vtn_builder *b, uint32_t id)
+{
+ struct vtn_access_chain *chain =
+ vtn_value(b, id, vtn_value_type_access_chain)->access_chain;
+
+ return vtn_access_chain_to_deref(b, chain);
+}
+
+/*
+ * Gets the NIR-level deref tail, which may have as a child an array deref
+ * selecting which component due to OpAccessChain supporting per-component
+ * indexing in SPIR-V.
+ */
+static nir_deref *
+get_deref_tail(nir_deref_var *deref)
+{
+ nir_deref *cur = &deref->deref;
+ while (!glsl_type_is_vector_or_scalar(cur->type) && cur->child)
+ cur = cur->child;
+
+ return cur;
+}
+
+struct vtn_ssa_value *
+vtn_local_load(struct vtn_builder *b, nir_deref_var *src)
+{
+ nir_deref *src_tail = get_deref_tail(src);
+ struct vtn_ssa_value *val = vtn_create_ssa_value(b, src_tail->type);
+ _vtn_local_load_store(b, true, src, src_tail, val);
+
+ if (src_tail->child) {
+ nir_deref_array *vec_deref = nir_deref_as_array(src_tail->child);
+ assert(vec_deref->deref.child == NULL);
+ val->type = vec_deref->deref.type;
+ if (vec_deref->deref_array_type == nir_deref_array_type_direct)
+ val->def = vtn_vector_extract(b, val->def, vec_deref->base_offset);
+ else
+ val->def = vtn_vector_extract_dynamic(b, val->def,
+ vec_deref->indirect.ssa);
+ }
+
+ return val;
+}
+
+void
+vtn_local_store(struct vtn_builder *b, struct vtn_ssa_value *src,
+ nir_deref_var *dest)
+{
+ nir_deref *dest_tail = get_deref_tail(dest);
+
+ if (dest_tail->child) {
+ struct vtn_ssa_value *val = vtn_create_ssa_value(b, dest_tail->type);
+ _vtn_local_load_store(b, true, dest, dest_tail, val);
+ nir_deref_array *deref = nir_deref_as_array(dest_tail->child);
+ assert(deref->deref.child == NULL);
+ if (deref->deref_array_type == nir_deref_array_type_direct)
+ val->def = vtn_vector_insert(b, val->def, src->def,
+ deref->base_offset);
+ else
+ val->def = vtn_vector_insert_dynamic(b, val->def, src->def,
+ deref->indirect.ssa);
+ _vtn_local_load_store(b, false, dest, dest_tail, val);
+ } else {
+ _vtn_local_load_store(b, false, dest, dest_tail, src);
+ }
+}
+
+static nir_ssa_def *
+get_vulkan_resource_index(struct vtn_builder *b, struct vtn_access_chain *chain,
+ struct vtn_type **type, unsigned *chain_idx)
+{
+ /* Push constants have no explicit binding */
+ if (chain->var->mode == vtn_variable_mode_push_constant) {
+ *chain_idx = 0;
+ *type = chain->var->type;
+ return NULL;
+ }
+
+ nir_ssa_def *array_index;
+ if (glsl_type_is_array(chain->var->type->type)) {
+ assert(chain->length > 0);
+ array_index = vtn_access_link_as_ssa(b, chain->link[0], 1);
+ *chain_idx = 1;
+ *type = chain->var->type->array_element;
+ } else {
+ array_index = nir_imm_int(&b->nb, 0);
+ *chain_idx = 0;
+ *type = chain->var->type;
+ }
+
+ nir_intrinsic_instr *instr =
+ nir_intrinsic_instr_create(b->nb.shader,
+ nir_intrinsic_vulkan_resource_index);
+ instr->src[0] = nir_src_for_ssa(array_index);
+ nir_intrinsic_set_desc_set(instr, chain->var->descriptor_set);
+ nir_intrinsic_set_binding(instr, chain->var->binding);
+
+ nir_ssa_dest_init(&instr->instr, &instr->dest, 1, 32, NULL);
+ nir_builder_instr_insert(&b->nb, &instr->instr);
+
+ return &instr->dest.ssa;
+}
+
+nir_ssa_def *
+vtn_access_chain_to_offset(struct vtn_builder *b,
+ struct vtn_access_chain *chain,
+ nir_ssa_def **index_out, struct vtn_type **type_out,
+ unsigned *end_idx_out, bool stop_at_matrix)
+{
+ unsigned idx = 0;
+ struct vtn_type *type;
+ *index_out = get_vulkan_resource_index(b, chain, &type, &idx);
+
+ nir_ssa_def *offset = nir_imm_int(&b->nb, 0);
+ for (; idx < chain->length; idx++) {
+ enum glsl_base_type base_type = glsl_get_base_type(type->type);
+ switch (base_type) {
+ case GLSL_TYPE_UINT:
+ case GLSL_TYPE_INT:
+ case GLSL_TYPE_FLOAT:
+ case GLSL_TYPE_DOUBLE:
+ case GLSL_TYPE_BOOL:
+ /* Some users may not want matrix or vector derefs */
+ if (stop_at_matrix)
+ goto end;
+ /* Fall through */
+
+ case GLSL_TYPE_ARRAY:
+ offset = nir_iadd(&b->nb, offset,
+ vtn_access_link_as_ssa(b, chain->link[idx],
+ type->stride));
+
+ type = type->array_element;
+ break;
+
+ case GLSL_TYPE_STRUCT: {
+ assert(chain->link[idx].mode == vtn_access_mode_literal);
+ unsigned member = chain->link[idx].id;
+ offset = nir_iadd(&b->nb, offset,
+ nir_imm_int(&b->nb, type->offsets[member]));
+ type = type->members[member];
+ break;
+ }
+
+ default:
+ unreachable("Invalid type for deref");
+ }
+ }
+
+end:
+ *type_out = type;
+ if (end_idx_out)
+ *end_idx_out = idx;
+
+ return offset;
+}
+
+static void
+_vtn_load_store_tail(struct vtn_builder *b, nir_intrinsic_op op, bool load,
+ nir_ssa_def *index, nir_ssa_def *offset,
+ struct vtn_ssa_value **inout, const struct glsl_type *type)
+{
+ nir_intrinsic_instr *instr = nir_intrinsic_instr_create(b->nb.shader, op);
+ instr->num_components = glsl_get_vector_elements(type);
+
+ int src = 0;
+ if (!load) {
+ nir_intrinsic_set_write_mask(instr, (1 << instr->num_components) - 1);
+ instr->src[src++] = nir_src_for_ssa((*inout)->def);
+ }
+
+ /* We set the base and size for push constant load to the entire push
+ * constant block for now.
+ */
+ if (op == nir_intrinsic_load_push_constant) {
+ nir_intrinsic_set_base(instr, 0);
+ nir_intrinsic_set_range(instr, 128);
+ }
+
+ if (index)
+ instr->src[src++] = nir_src_for_ssa(index);
+
+ instr->src[src++] = nir_src_for_ssa(offset);
+
+ if (load) {
+ nir_ssa_dest_init(&instr->instr, &instr->dest,
+ instr->num_components,
+ glsl_get_bit_size(glsl_get_base_type(type)), NULL);
+ (*inout)->def = &instr->dest.ssa;
+ }
+
+ nir_builder_instr_insert(&b->nb, &instr->instr);
+
+ if (load && glsl_get_base_type(type) == GLSL_TYPE_BOOL)
+ (*inout)->def = nir_ine(&b->nb, (*inout)->def, nir_imm_int(&b->nb, 0));
+}
+
+static void
+_vtn_block_load_store(struct vtn_builder *b, nir_intrinsic_op op, bool load,
+ nir_ssa_def *index, nir_ssa_def *offset,
+ struct vtn_access_chain *chain, unsigned chain_idx,
+ struct vtn_type *type, struct vtn_ssa_value **inout)
+{
+ if (chain && chain_idx >= chain->length)
+ chain = NULL;
+
+ if (load && chain == NULL && *inout == NULL)
+ *inout = vtn_create_ssa_value(b, type->type);
+
+ enum glsl_base_type base_type = glsl_get_base_type(type->type);
+ switch (base_type) {
+ case GLSL_TYPE_UINT:
+ case GLSL_TYPE_INT:
+ case GLSL_TYPE_FLOAT:
+ case GLSL_TYPE_BOOL:
+ /* This is where things get interesting. At this point, we've hit
+ * a vector, a scalar, or a matrix.
+ */
+ if (glsl_type_is_matrix(type->type)) {
+ if (chain == NULL) {
+ /* Loading the whole matrix */
+ struct vtn_ssa_value *transpose;
+ unsigned num_ops, vec_width;
+ if (type->row_major) {
+ num_ops = glsl_get_vector_elements(type->type);
+ vec_width = glsl_get_matrix_columns(type->type);
+ if (load) {
+ const struct glsl_type *transpose_type =
+ glsl_matrix_type(base_type, vec_width, num_ops);
+ *inout = vtn_create_ssa_value(b, transpose_type);
+ } else {
+ transpose = vtn_ssa_transpose(b, *inout);
+ inout = &transpose;
+ }
+ } else {
+ num_ops = glsl_get_matrix_columns(type->type);
+ vec_width = glsl_get_vector_elements(type->type);
+ }
+
+ for (unsigned i = 0; i < num_ops; i++) {
+ nir_ssa_def *elem_offset =
+ nir_iadd(&b->nb, offset,
+ nir_imm_int(&b->nb, i * type->stride));
+ _vtn_load_store_tail(b, op, load, index, elem_offset,
+ &(*inout)->elems[i],
+ glsl_vector_type(base_type, vec_width));
+ }
+
+ if (load && type->row_major)
+ *inout = vtn_ssa_transpose(b, *inout);
+ } else if (type->row_major) {
+ /* Row-major but with an access chiain. */
+ nir_ssa_def *col_offset =
+ vtn_access_link_as_ssa(b, chain->link[chain_idx],
+ type->array_element->stride);
+ offset = nir_iadd(&b->nb, offset, col_offset);
+
+ if (chain_idx + 1 < chain->length) {
+ /* Picking off a single element */
+ nir_ssa_def *row_offset =
+ vtn_access_link_as_ssa(b, chain->link[chain_idx + 1],
+ type->stride);
+ offset = nir_iadd(&b->nb, offset, row_offset);
+ if (load)
+ *inout = vtn_create_ssa_value(b, glsl_scalar_type(base_type));
+ _vtn_load_store_tail(b, op, load, index, offset, inout,
+ glsl_scalar_type(base_type));
+ } else {
+ /* Grabbing a column; picking one element off each row */
+ unsigned num_comps = glsl_get_vector_elements(type->type);
+ const struct glsl_type *column_type =
+ glsl_get_column_type(type->type);
+
+ nir_ssa_def *comps[4];
+ for (unsigned i = 0; i < num_comps; i++) {
+ nir_ssa_def *elem_offset =
+ nir_iadd(&b->nb, offset,
+ nir_imm_int(&b->nb, i * type->stride));
+
+ struct vtn_ssa_value *comp, temp_val;
+ if (!load) {
+ temp_val.def = nir_channel(&b->nb, (*inout)->def, i);
+ temp_val.type = glsl_scalar_type(base_type);
+ }
+ comp = &temp_val;
+ _vtn_load_store_tail(b, op, load, index, elem_offset,
+ &comp, glsl_scalar_type(base_type));
+ comps[i] = comp->def;
+ }
+
+ if (load) {
+ if (*inout == NULL)
+ *inout = vtn_create_ssa_value(b, column_type);
+
+ (*inout)->def = nir_vec(&b->nb, comps, num_comps);
+ }
+ }
+ } else {
+ /* Column-major with a deref. Fall through to array case. */
+ nir_ssa_def *col_offset =
+ vtn_access_link_as_ssa(b, chain->link[chain_idx], type->stride);
+ offset = nir_iadd(&b->nb, offset, col_offset);
+
+ _vtn_block_load_store(b, op, load, index, offset,
+ chain, chain_idx + 1,
+ type->array_element, inout);
+ }
+ } else if (chain == NULL) {
+ /* Single whole vector */
+ assert(glsl_type_is_vector_or_scalar(type->type));
+ _vtn_load_store_tail(b, op, load, index, offset, inout, type->type);
+ } else {
+ /* Single component of a vector. Fall through to array case. */
+ nir_ssa_def *elem_offset =
+ vtn_access_link_as_ssa(b, chain->link[chain_idx], type->stride);
+ offset = nir_iadd(&b->nb, offset, elem_offset);
+
+ _vtn_block_load_store(b, op, load, index, offset, NULL, 0,
+ type->array_element, inout);
+ }
+ return;
+
+ case GLSL_TYPE_ARRAY: {
+ unsigned elems = glsl_get_length(type->type);
+ for (unsigned i = 0; i < elems; i++) {
+ nir_ssa_def *elem_off =
+ nir_iadd(&b->nb, offset, nir_imm_int(&b->nb, i * type->stride));
+ _vtn_block_load_store(b, op, load, index, elem_off, NULL, 0,
+ type->array_element, &(*inout)->elems[i]);
+ }
+ return;
+ }
+
+ case GLSL_TYPE_STRUCT: {
+ unsigned elems = glsl_get_length(type->type);
+ for (unsigned i = 0; i < elems; i++) {
+ nir_ssa_def *elem_off =
+ nir_iadd(&b->nb, offset, nir_imm_int(&b->nb, type->offsets[i]));
+ _vtn_block_load_store(b, op, load, index, elem_off, NULL, 0,
+ type->members[i], &(*inout)->elems[i]);
+ }
+ return;
+ }
+
+ default:
+ unreachable("Invalid block member type");
+ }
+}
+
+static struct vtn_ssa_value *
+vtn_block_load(struct vtn_builder *b, struct vtn_access_chain *src)
+{
+ nir_intrinsic_op op;
+ switch (src->var->mode) {
+ case vtn_variable_mode_ubo:
+ op = nir_intrinsic_load_ubo;
+ break;
+ case vtn_variable_mode_ssbo:
+ op = nir_intrinsic_load_ssbo;
+ break;
+ case vtn_variable_mode_push_constant:
+ op = nir_intrinsic_load_push_constant;
+ break;
+ default:
+ assert(!"Invalid block variable mode");
+ }
+
+ nir_ssa_def *offset, *index = NULL;
+ struct vtn_type *type;
+ unsigned chain_idx;
+ offset = vtn_access_chain_to_offset(b, src, &index, &type, &chain_idx, true);
+
+ struct vtn_ssa_value *value = NULL;
+ _vtn_block_load_store(b, op, true, index, offset,
+ src, chain_idx, type, &value);
+ return value;
+}
+
+static void
+vtn_block_store(struct vtn_builder *b, struct vtn_ssa_value *src,
+ struct vtn_access_chain *dst)
+{
+ nir_ssa_def *offset, *index = NULL;
+ struct vtn_type *type;
+ unsigned chain_idx;
+ offset = vtn_access_chain_to_offset(b, dst, &index, &type, &chain_idx, true);
+
+ _vtn_block_load_store(b, nir_intrinsic_store_ssbo, false, index, offset,
+ dst, chain_idx, type, &src);
+}
+
+static bool
+vtn_variable_is_external_block(struct vtn_variable *var)
+{
+ return var->mode == vtn_variable_mode_ssbo ||
+ var->mode == vtn_variable_mode_ubo ||
+ var->mode == vtn_variable_mode_push_constant;
+}
+
+static void
+_vtn_variable_load_store(struct vtn_builder *b, bool load,
+ struct vtn_access_chain *chain,
+ struct vtn_type *tail_type,
+ struct vtn_ssa_value **inout)
+{
+ enum glsl_base_type base_type = glsl_get_base_type(tail_type->type);
+ switch (base_type) {
+ case GLSL_TYPE_UINT:
+ case GLSL_TYPE_INT:
+ case GLSL_TYPE_FLOAT:
+ case GLSL_TYPE_BOOL:
+ /* At this point, we have a scalar, vector, or matrix so we know that
+ * there cannot be any structure splitting still in the way. By
+ * stopping at the matrix level rather than the vector level, we
+ * ensure that matrices get loaded in the optimal way even if they
+ * are storred row-major in a UBO.
+ */
+ if (load) {
+ *inout = vtn_local_load(b, vtn_access_chain_to_deref(b, chain));
+ } else {
+ vtn_local_store(b, *inout, vtn_access_chain_to_deref(b, chain));
+ }
+ return;
+
+ case GLSL_TYPE_ARRAY:
+ case GLSL_TYPE_STRUCT: {
+ struct vtn_access_chain *new_chain =
+ vtn_access_chain_extend(b, chain, 1);
+ new_chain->link[chain->length].mode = vtn_access_mode_literal;
+ unsigned elems = glsl_get_length(tail_type->type);
+ if (load) {
+ assert(*inout == NULL);
+ *inout = rzalloc(b, struct vtn_ssa_value);
+ (*inout)->type = tail_type->type;
+ (*inout)->elems = rzalloc_array(b, struct vtn_ssa_value *, elems);
+ }
+ for (unsigned i = 0; i < elems; i++) {
+ new_chain->link[chain->length].id = i;
+ struct vtn_type *elem_type = base_type == GLSL_TYPE_ARRAY ?
+ tail_type->array_element : tail_type->members[i];
+ _vtn_variable_load_store(b, load, new_chain, elem_type,
+ &(*inout)->elems[i]);
+ }
+ return;
+ }
+
+ default:
+ unreachable("Invalid access chain type");
+ }
+}
+
+struct vtn_ssa_value *
+vtn_variable_load(struct vtn_builder *b, struct vtn_access_chain *src)
+{
+ if (vtn_variable_is_external_block(src->var)) {
+ return vtn_block_load(b, src);
+ } else {
+ struct vtn_type *tail_type = vtn_access_chain_tail_type(b, src);
+ struct vtn_ssa_value *val = NULL;
+ _vtn_variable_load_store(b, true, src, tail_type, &val);
+ return val;
+ }
+}
+
+void
+vtn_variable_store(struct vtn_builder *b, struct vtn_ssa_value *src,
+ struct vtn_access_chain *dest)
+{
+ if (vtn_variable_is_external_block(dest->var)) {
+ assert(dest->var->mode == vtn_variable_mode_ssbo);
+ vtn_block_store(b, src, dest);
+ } else {
+ struct vtn_type *tail_type = vtn_access_chain_tail_type(b, dest);
+ _vtn_variable_load_store(b, false, dest, tail_type, &src);
+ }
+}
+
+static void
+_vtn_variable_copy(struct vtn_builder *b, struct vtn_access_chain *dest,
+ struct vtn_access_chain *src, struct vtn_type *tail_type)
+{
+ enum glsl_base_type base_type = glsl_get_base_type(tail_type->type);
+ switch (base_type) {
+ case GLSL_TYPE_UINT:
+ case GLSL_TYPE_INT:
+ case GLSL_TYPE_FLOAT:
+ case GLSL_TYPE_BOOL:
+ /* At this point, we have a scalar, vector, or matrix so we know that
+ * there cannot be any structure splitting still in the way. By
+ * stopping at the matrix level rather than the vector level, we
+ * ensure that matrices get loaded in the optimal way even if they
+ * are storred row-major in a UBO.
+ */
+ vtn_variable_store(b, vtn_variable_load(b, src), dest);
+ return;
+
+ case GLSL_TYPE_ARRAY:
+ case GLSL_TYPE_STRUCT: {
+ struct vtn_access_chain *new_src, *new_dest;
+ new_src = vtn_access_chain_extend(b, src, 1);
+ new_dest = vtn_access_chain_extend(b, dest, 1);
+ new_src->link[src->length].mode = vtn_access_mode_literal;
+ new_dest->link[dest->length].mode = vtn_access_mode_literal;
+ unsigned elems = glsl_get_length(tail_type->type);
+ for (unsigned i = 0; i < elems; i++) {
+ new_src->link[src->length].id = i;
+ new_dest->link[dest->length].id = i;
+ struct vtn_type *elem_type = base_type == GLSL_TYPE_ARRAY ?
+ tail_type->array_element : tail_type->members[i];
+ _vtn_variable_copy(b, new_dest, new_src, elem_type);
+ }
+ return;
+ }
+
+ default:
+ unreachable("Invalid access chain type");
+ }
+}
+
+static void
+vtn_variable_copy(struct vtn_builder *b, struct vtn_access_chain *dest,
+ struct vtn_access_chain *src)
+{
+ struct vtn_type *tail_type = vtn_access_chain_tail_type(b, src);
+ assert(vtn_access_chain_tail_type(b, dest)->type == tail_type->type);
+
+ /* TODO: At some point, we should add a special-case for when we can
+ * just emit a copy_var intrinsic.
+ */
+ _vtn_variable_copy(b, dest, src, tail_type);
+}
+
+static void
+set_mode_system_value(nir_variable_mode *mode)
+{
+ assert(*mode == nir_var_system_value || *mode == nir_var_shader_in);
+ *mode = nir_var_system_value;
+}
+
+static void
+vtn_get_builtin_location(struct vtn_builder *b,
+ SpvBuiltIn builtin, int *location,
+ nir_variable_mode *mode)
+{
+ switch (builtin) {
+ case SpvBuiltInPosition:
+ *location = VARYING_SLOT_POS;
+ break;
+ case SpvBuiltInPointSize:
+ *location = VARYING_SLOT_PSIZ;
+ break;
+ case SpvBuiltInClipDistance:
+ *location = VARYING_SLOT_CLIP_DIST0; /* XXX CLIP_DIST1? */
+ break;
+ case SpvBuiltInCullDistance:
+ /* XXX figure this out */
+ break;
+ case SpvBuiltInVertexIndex:
+ *location = SYSTEM_VALUE_VERTEX_ID;
+ set_mode_system_value(mode);
+ break;
+ case SpvBuiltInVertexId:
+ /* Vulkan defines VertexID to be zero-based and reserves the new
+ * builtin keyword VertexIndex to indicate the non-zero-based value.
+ */
+ *location = SYSTEM_VALUE_VERTEX_ID_ZERO_BASE;
+ set_mode_system_value(mode);
+ break;
+ case SpvBuiltInInstanceIndex:
+ *location = SYSTEM_VALUE_INSTANCE_INDEX;
+ set_mode_system_value(mode);
+ break;
+ case SpvBuiltInInstanceId:
+ *location = SYSTEM_VALUE_INSTANCE_ID;
+ set_mode_system_value(mode);
+ break;
+ case SpvBuiltInPrimitiveId:
+ *location = VARYING_SLOT_PRIMITIVE_ID;
+ *mode = nir_var_shader_out;
+ break;
+ case SpvBuiltInInvocationId:
+ *location = SYSTEM_VALUE_INVOCATION_ID;
+ set_mode_system_value(mode);
+ break;
+ case SpvBuiltInLayer:
+ *location = VARYING_SLOT_LAYER;
+ *mode = nir_var_shader_out;
+ break;
+ case SpvBuiltInViewportIndex:
+ *location = VARYING_SLOT_VIEWPORT;
+ if (b->shader->stage == MESA_SHADER_GEOMETRY)
+ *mode = nir_var_shader_out;
+ else if (b->shader->stage == MESA_SHADER_FRAGMENT)
+ *mode = nir_var_shader_in;
+ else
+ unreachable("invalid stage for SpvBuiltInViewportIndex");
+ break;
+ case SpvBuiltInTessLevelOuter:
+ case SpvBuiltInTessLevelInner:
+ case SpvBuiltInTessCoord:
+ case SpvBuiltInPatchVertices:
+ unreachable("no tessellation support");
+ case SpvBuiltInFragCoord:
+ *location = VARYING_SLOT_POS;
+ assert(*mode == nir_var_shader_in);
+ break;
+ case SpvBuiltInPointCoord:
+ *location = VARYING_SLOT_PNTC;
+ assert(*mode == nir_var_shader_in);
+ break;
+ case SpvBuiltInFrontFacing:
+ *location = VARYING_SLOT_FACE;
+ assert(*mode == nir_var_shader_in);
+ break;
+ case SpvBuiltInSampleId:
+ *location = SYSTEM_VALUE_SAMPLE_ID;
+ set_mode_system_value(mode);
+ break;
+ case SpvBuiltInSamplePosition:
+ *location = SYSTEM_VALUE_SAMPLE_POS;
+ set_mode_system_value(mode);
+ break;
+ case SpvBuiltInSampleMask:
+ *location = SYSTEM_VALUE_SAMPLE_MASK_IN; /* XXX out? */
+ set_mode_system_value(mode);
+ break;
+ case SpvBuiltInFragDepth:
+ *location = FRAG_RESULT_DEPTH;
+ assert(*mode == nir_var_shader_out);
+ break;
+ case SpvBuiltInNumWorkgroups:
+ *location = SYSTEM_VALUE_NUM_WORK_GROUPS;
+ set_mode_system_value(mode);
+ break;
+ case SpvBuiltInWorkgroupSize:
+ /* This should already be handled */
+ unreachable("unsupported builtin");
+ break;
+ case SpvBuiltInWorkgroupId:
+ *location = SYSTEM_VALUE_WORK_GROUP_ID;
+ set_mode_system_value(mode);
+ break;
+ case SpvBuiltInLocalInvocationId:
+ *location = SYSTEM_VALUE_LOCAL_INVOCATION_ID;
+ set_mode_system_value(mode);
+ break;
+ case SpvBuiltInLocalInvocationIndex:
+ *location = SYSTEM_VALUE_LOCAL_INVOCATION_INDEX;
+ set_mode_system_value(mode);
+ break;
+ case SpvBuiltInGlobalInvocationId:
+ *location = SYSTEM_VALUE_GLOBAL_INVOCATION_ID;
+ set_mode_system_value(mode);
+ break;
+ case SpvBuiltInHelperInvocation:
+ default:
+ unreachable("unsupported builtin");
+ }
+}
+
+static void
+var_decoration_cb(struct vtn_builder *b, struct vtn_value *val, int member,
+ const struct vtn_decoration *dec, void *void_var)
+{
+ struct vtn_variable *vtn_var = void_var;
+
+ /* Handle decorations that apply to a vtn_variable as a whole */
+ switch (dec->decoration) {
+ case SpvDecorationBinding:
+ vtn_var->binding = dec->literals[0];
+ return;
+ case SpvDecorationDescriptorSet:
+ vtn_var->descriptor_set = dec->literals[0];
+ return;
+
+ case SpvDecorationLocation: {
+ unsigned location = dec->literals[0];
+ bool is_vertex_input;
+ if (b->shader->stage == MESA_SHADER_FRAGMENT &&
+ vtn_var->mode == vtn_variable_mode_output) {
+ is_vertex_input = false;
+ location += FRAG_RESULT_DATA0;
+ } else if (b->shader->stage == MESA_SHADER_VERTEX &&
+ vtn_var->mode == vtn_variable_mode_input) {
+ is_vertex_input = true;
+ location += VERT_ATTRIB_GENERIC0;
+ } else if (vtn_var->mode == vtn_variable_mode_input ||
+ vtn_var->mode == vtn_variable_mode_output) {
+ is_vertex_input = false;
+ location += VARYING_SLOT_VAR0;
+ } else {
+ assert(!"Location must be on input or output variable");
+ }
+
+ if (vtn_var->var) {
+ vtn_var->var->data.location = location;
+ vtn_var->var->data.explicit_location = true;
+ } else {
+ assert(vtn_var->members);
+ unsigned length = glsl_get_length(vtn_var->type->type);
+ for (unsigned i = 0; i < length; i++) {
+ vtn_var->members[i]->data.location = location;
+ vtn_var->members[i]->data.explicit_location = true;
+ location +=
+ glsl_count_attribute_slots(vtn_var->members[i]->interface_type,
+ is_vertex_input);
+ }
+ }
+ return;
+ }
+
+ default:
+ break;
+ }
+
+ /* Now we handle decorations that apply to a particular nir_variable */
+ nir_variable *nir_var = vtn_var->var;
+ if (val->value_type == vtn_value_type_access_chain) {
+ assert(val->access_chain->length == 0);
+ assert(val->access_chain->var == void_var);
+ assert(member == -1);
+ } else {
+ assert(val->value_type == vtn_value_type_type);
+ if (member != -1)
+ nir_var = vtn_var->members[member];
+ }
+
+ if (nir_var == NULL)
+ return;
+
+ switch (dec->decoration) {
+ case SpvDecorationRelaxedPrecision:
+ break; /* FIXME: Do nothing with this for now. */
+ case SpvDecorationNoPerspective:
+ nir_var->data.interpolation = INTERP_QUALIFIER_NOPERSPECTIVE;
+ break;
+ case SpvDecorationFlat:
+ nir_var->data.interpolation = INTERP_QUALIFIER_FLAT;
+ break;
+ case SpvDecorationCentroid:
+ nir_var->data.centroid = true;
+ break;
+ case SpvDecorationSample:
+ nir_var->data.sample = true;
+ break;
+ case SpvDecorationInvariant:
+ nir_var->data.invariant = true;
+ break;
+ case SpvDecorationConstant:
+ assert(nir_var->constant_initializer != NULL);
+ nir_var->data.read_only = true;
+ break;
+ case SpvDecorationNonWritable:
+ nir_var->data.read_only = true;
+ break;
+ case SpvDecorationComponent:
+ nir_var->data.location_frac = dec->literals[0];
+ break;
+ case SpvDecorationIndex:
+ nir_var->data.explicit_index = true;
+ nir_var->data.index = dec->literals[0];
+ break;
+ case SpvDecorationBuiltIn: {
+ SpvBuiltIn builtin = dec->literals[0];
+
+ if (builtin == SpvBuiltInWorkgroupSize) {
+ /* This shouldn't be a builtin. It's actually a constant. */
+ nir_var->data.mode = nir_var_global;
+ nir_var->data.read_only = true;
+
+ nir_constant *c = rzalloc(nir_var, nir_constant);
+ c->value.u[0] = b->shader->info.cs.local_size[0];
+ c->value.u[1] = b->shader->info.cs.local_size[1];
+ c->value.u[2] = b->shader->info.cs.local_size[2];
+ nir_var->constant_initializer = c;
+ break;
+ }
+
+ nir_variable_mode mode = nir_var->data.mode;
+ vtn_get_builtin_location(b, builtin, &nir_var->data.location, &mode);
+ nir_var->data.explicit_location = true;
+ nir_var->data.mode = mode;
+
+ if (builtin == SpvBuiltInFragCoord || builtin == SpvBuiltInSamplePosition)
+ nir_var->data.origin_upper_left = b->origin_upper_left;
+ break;
+ }
+ case SpvDecorationRowMajor:
+ case SpvDecorationColMajor:
+ case SpvDecorationGLSLShared:
+ case SpvDecorationPatch:
+ case SpvDecorationRestrict:
+ case SpvDecorationAliased:
+ case SpvDecorationVolatile:
+ case SpvDecorationCoherent:
+ case SpvDecorationNonReadable:
+ case SpvDecorationUniform:
+ /* This is really nice but we have no use for it right now. */
+ case SpvDecorationCPacked:
+ case SpvDecorationSaturatedConversion:
+ case SpvDecorationStream:
+ case SpvDecorationOffset:
+ case SpvDecorationXfbBuffer:
+ case SpvDecorationFuncParamAttr:
+ case SpvDecorationFPRoundingMode:
+ case SpvDecorationFPFastMathMode:
+ case SpvDecorationLinkageAttributes:
+ case SpvDecorationSpecId:
+ break;
+ default:
+ unreachable("Unhandled variable decoration");
+ }
+}
+
+/* Tries to compute the size of an interface block based on the strides and
+ * offsets that are provided to us in the SPIR-V source.
+ */
+static unsigned
+vtn_type_block_size(struct vtn_type *type)
+{
+ enum glsl_base_type base_type = glsl_get_base_type(type->type);
+ switch (base_type) {
+ case GLSL_TYPE_UINT:
+ case GLSL_TYPE_INT:
+ case GLSL_TYPE_FLOAT:
+ case GLSL_TYPE_BOOL:
+ case GLSL_TYPE_DOUBLE: {
+ unsigned cols = type->row_major ? glsl_get_vector_elements(type->type) :
+ glsl_get_matrix_columns(type->type);
+ if (cols > 1) {
+ assert(type->stride > 0);
+ return type->stride * cols;
+ } else if (base_type == GLSL_TYPE_DOUBLE) {
+ return glsl_get_vector_elements(type->type) * 8;
+ } else {
+ return glsl_get_vector_elements(type->type) * 4;
+ }
+ }
+
+ case GLSL_TYPE_STRUCT:
+ case GLSL_TYPE_INTERFACE: {
+ unsigned size = 0;
+ unsigned num_fields = glsl_get_length(type->type);
+ for (unsigned f = 0; f < num_fields; f++) {
+ unsigned field_end = type->offsets[f] +
+ vtn_type_block_size(type->members[f]);
+ size = MAX2(size, field_end);
+ }
+ return size;
+ }
+
+ case GLSL_TYPE_ARRAY:
+ assert(type->stride > 0);
+ assert(glsl_get_length(type->type) > 0);
+ return type->stride * glsl_get_length(type->type);
+
+ default:
+ assert(!"Invalid block type");
+ return 0;
+ }
+}
+
+void
+vtn_handle_variables(struct vtn_builder *b, SpvOp opcode,
+ const uint32_t *w, unsigned count)
+{
+ switch (opcode) {
+ case SpvOpVariable: {
+ struct vtn_variable *var = rzalloc(b, struct vtn_variable);
+ var->type = vtn_value(b, w[1], vtn_value_type_type)->type;
+
+ var->chain.var = var;
+ var->chain.length = 0;
+
+ struct vtn_value *val =
+ vtn_push_value(b, w[2], vtn_value_type_access_chain);
+ val->access_chain = &var->chain;
+
+ struct vtn_type *without_array = var->type;
+ while(glsl_type_is_array(without_array->type))
+ without_array = without_array->array_element;
+
+ nir_variable_mode nir_mode;
+ switch ((SpvStorageClass)w[3]) {
+ case SpvStorageClassUniform:
+ case SpvStorageClassUniformConstant:
+ if (without_array->block) {
+ var->mode = vtn_variable_mode_ubo;
+ b->shader->info.num_ubos++;
+ } else if (without_array->buffer_block) {
+ var->mode = vtn_variable_mode_ssbo;
+ b->shader->info.num_ssbos++;
+ } else if (glsl_type_is_image(without_array->type)) {
+ var->mode = vtn_variable_mode_image;
+ nir_mode = nir_var_uniform;
+ b->shader->info.num_images++;
+ } else if (glsl_type_is_sampler(without_array->type)) {
+ var->mode = vtn_variable_mode_sampler;
+ nir_mode = nir_var_uniform;
+ b->shader->info.num_textures++;
+ } else {
+ assert(!"Invalid uniform variable type");
+ }
+ break;
+ case SpvStorageClassPushConstant:
+ var->mode = vtn_variable_mode_push_constant;
+ assert(b->shader->num_uniforms == 0);
+ b->shader->num_uniforms = vtn_type_block_size(var->type) * 4;
+ break;
+ case SpvStorageClassInput:
+ var->mode = vtn_variable_mode_input;
+ nir_mode = nir_var_shader_in;
+ break;
+ case SpvStorageClassOutput:
+ var->mode = vtn_variable_mode_output;
+ nir_mode = nir_var_shader_out;
+ break;
+ case SpvStorageClassPrivate:
+ var->mode = vtn_variable_mode_global;
+ nir_mode = nir_var_global;
+ break;
+ case SpvStorageClassFunction:
+ var->mode = vtn_variable_mode_local;
+ nir_mode = nir_var_local;
+ break;
+ case SpvStorageClassWorkgroup:
+ var->mode = vtn_variable_mode_workgroup;
+ nir_mode = nir_var_shared;
+ break;
+ case SpvStorageClassCrossWorkgroup:
+ case SpvStorageClassGeneric:
+ case SpvStorageClassAtomicCounter:
+ default:
+ unreachable("Unhandled variable storage class");
+ }
+
+ switch (var->mode) {
+ case vtn_variable_mode_local:
+ case vtn_variable_mode_global:
+ case vtn_variable_mode_image:
+ case vtn_variable_mode_sampler:
+ case vtn_variable_mode_workgroup:
+ /* For these, we create the variable normally */
+ var->var = rzalloc(b->shader, nir_variable);
+ var->var->name = ralloc_strdup(var->var, val->name);
+ var->var->type = var->type->type;
+ var->var->data.mode = nir_mode;
+
+ switch (var->mode) {
+ case vtn_variable_mode_image:
+ case vtn_variable_mode_sampler:
+ var->var->interface_type = without_array->type;
+ break;
+ default:
+ var->var->interface_type = NULL;
+ break;
+ }
+ break;
+
+ case vtn_variable_mode_input:
+ case vtn_variable_mode_output: {
+ /* For inputs and outputs, we immediately split structures. This
+ * is for a couple of reasons. For one, builtins may all come in
+ * a struct and we really want those split out into separate
+ * variables. For another, interpolation qualifiers can be
+ * applied to members of the top-level struct ane we need to be
+ * able to preserve that information.
+ */
+
+ int array_length = -1;
+ struct vtn_type *interface_type = var->type;
+ if (b->shader->stage == MESA_SHADER_GEOMETRY &&
+ glsl_type_is_array(var->type->type)) {
+ /* In Geometry shaders (and some tessellation), inputs come
+ * in per-vertex arrays. However, some builtins come in
+ * non-per-vertex, hence the need for the is_array check. In
+ * any case, there are no non-builtin arrays allowed so this
+ * check should be sufficient.
+ */
+ interface_type = var->type->array_element;
+ array_length = glsl_get_length(var->type->type);
+ }
+
+ if (glsl_type_is_struct(interface_type->type)) {
+ /* It's a struct. Split it. */
+ unsigned num_members = glsl_get_length(interface_type->type);
+ var->members = ralloc_array(b, nir_variable *, num_members);
+
+ for (unsigned i = 0; i < num_members; i++) {
+ const struct glsl_type *mtype = interface_type->members[i]->type;
+ if (array_length >= 0)
+ mtype = glsl_array_type(mtype, array_length);
+
+ var->members[i] = rzalloc(b->shader, nir_variable);
+ var->members[i]->name =
+ ralloc_asprintf(var->members[i], "%s.%d", val->name, i);
+ var->members[i]->type = mtype;
+ var->members[i]->interface_type =
+ interface_type->members[i]->type;
+ var->members[i]->data.mode = nir_mode;
+ }
+ } else {
+ var->var = rzalloc(b->shader, nir_variable);
+ var->var->name = ralloc_strdup(var->var, val->name);
+ var->var->type = var->type->type;
+ var->var->interface_type = interface_type->type;
+ var->var->data.mode = nir_mode;
+ }
+
+ /* For inputs and outputs, we need to grab locations and builtin
+ * information from the interface type.
+ */
+ vtn_foreach_decoration(b, interface_type->val, var_decoration_cb, var);
+ break;
+
+ case vtn_variable_mode_param:
+ unreachable("Not created through OpVariable");
+ }
+
+ case vtn_variable_mode_ubo:
+ case vtn_variable_mode_ssbo:
+ case vtn_variable_mode_push_constant:
+ /* These don't need actual variables. */
+ break;
+ }
+
+ if (count > 4) {
+ assert(count == 5);
+ nir_constant *constant =
+ vtn_value(b, w[4], vtn_value_type_constant)->constant;
+ var->var->constant_initializer =
+ nir_constant_clone(constant, var->var);
+ }
+
+ vtn_foreach_decoration(b, val, var_decoration_cb, var);
+
+ if (var->mode == vtn_variable_mode_image ||
+ var->mode == vtn_variable_mode_sampler) {
+ /* XXX: We still need the binding information in the nir_variable
+ * for these. We should fix that.
+ */
+ var->var->data.binding = var->binding;
+ var->var->data.descriptor_set = var->descriptor_set;
+
+ if (var->mode == vtn_variable_mode_image)
+ var->var->data.image.format = without_array->image_format;
+ }
+
+ if (var->mode == vtn_variable_mode_local) {
+ assert(var->members == NULL && var->var != NULL);
+ nir_function_impl_add_variable(b->impl, var->var);
+ } else if (var->var) {
+ nir_shader_add_variable(b->shader, var->var);
+ } else if (var->members) {
+ unsigned count = glsl_get_length(without_array->type);
+ for (unsigned i = 0; i < count; i++) {
+ assert(var->members[i]->data.mode != nir_var_local);
+ nir_shader_add_variable(b->shader, var->members[i]);
+ }
+ } else {
+ assert(var->mode == vtn_variable_mode_ubo ||
+ var->mode == vtn_variable_mode_ssbo ||
+ var->mode == vtn_variable_mode_push_constant);
+ }
+ break;
+ }
+
+ case SpvOpAccessChain:
+ case SpvOpInBoundsAccessChain: {
+ struct vtn_access_chain *base, *chain;
+ struct vtn_value *base_val = vtn_untyped_value(b, w[3]);
+ if (base_val->value_type == vtn_value_type_sampled_image) {
+ /* This is rather insane. SPIR-V allows you to use OpSampledImage
+ * to combine an array of images with a single sampler to get an
+ * array of sampled images that all share the same sampler.
+ * Fortunately, this means that we can more-or-less ignore the
+ * sampler when crawling the access chain, but it does leave us
+ * with this rather awkward little special-case.
+ */
+ base = base_val->sampled_image->image;
+ } else {
+ assert(base_val->value_type == vtn_value_type_access_chain);
+ base = base_val->access_chain;
+ }
+
+ chain = vtn_access_chain_extend(b, base, count - 4);
+
+ unsigned idx = base->length;
+ for (int i = 4; i < count; i++) {
+ struct vtn_value *link_val = vtn_untyped_value(b, w[i]);
+ if (link_val->value_type == vtn_value_type_constant) {
+ chain->link[idx].mode = vtn_access_mode_literal;
+ chain->link[idx].id = link_val->constant->value.u[0];
+ } else {
+ chain->link[idx].mode = vtn_access_mode_id;
+ chain->link[idx].id = w[i];
+ }
+ idx++;
+ }
+
+ if (base_val->value_type == vtn_value_type_sampled_image) {
+ struct vtn_value *val =
+ vtn_push_value(b, w[2], vtn_value_type_sampled_image);
+ val->sampled_image = ralloc(b, struct vtn_sampled_image);
+ val->sampled_image->image = chain;
+ val->sampled_image->sampler = base_val->sampled_image->sampler;
+ } else {
+ struct vtn_value *val =
+ vtn_push_value(b, w[2], vtn_value_type_access_chain);
+ val->access_chain = chain;
+ }
+ break;
+ }
+
+ case SpvOpCopyMemory: {
+ struct vtn_value *dest = vtn_value(b, w[1], vtn_value_type_access_chain);
+ struct vtn_value *src = vtn_value(b, w[2], vtn_value_type_access_chain);
+
+ vtn_variable_copy(b, dest->access_chain, src->access_chain);
+ break;
+ }
+
+ case SpvOpLoad: {
+ struct vtn_access_chain *src =
+ vtn_value(b, w[3], vtn_value_type_access_chain)->access_chain;
+
+ if (src->var->mode == vtn_variable_mode_image ||
+ src->var->mode == vtn_variable_mode_sampler) {
+ vtn_push_value(b, w[2], vtn_value_type_access_chain)->access_chain = src;
+ return;
+ }
+
+ struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
+ val->ssa = vtn_variable_load(b, src);
+ break;
+ }
+
+ case SpvOpStore: {
+ struct vtn_access_chain *dest =
+ vtn_value(b, w[1], vtn_value_type_access_chain)->access_chain;
+ struct vtn_ssa_value *src = vtn_ssa_value(b, w[2]);
+ vtn_variable_store(b, src, dest);
+ break;
+ }
+
+ case SpvOpArrayLength: {
+ struct vtn_access_chain *chain =
+ vtn_value(b, w[3], vtn_value_type_access_chain)->access_chain;
+
+ const uint32_t offset = chain->var->type->offsets[w[4]];
+ const uint32_t stride = chain->var->type->members[w[4]]->stride;
+
+ unsigned chain_idx;
+ struct vtn_type *type;
+ nir_ssa_def *index =
+ get_vulkan_resource_index(b, chain, &type, &chain_idx);
+
+ nir_intrinsic_instr *instr =
+ nir_intrinsic_instr_create(b->nb.shader,
+ nir_intrinsic_get_buffer_size);
+ instr->src[0] = nir_src_for_ssa(index);
+ nir_ssa_dest_init(&instr->instr, &instr->dest, 1, 32, NULL);
+ nir_builder_instr_insert(&b->nb, &instr->instr);
+ nir_ssa_def *buf_size = &instr->dest.ssa;
+
+ /* array_length = max(buffer_size - offset, 0) / stride */
+ nir_ssa_def *array_length =
+ nir_idiv(&b->nb,
+ nir_imax(&b->nb,
+ nir_isub(&b->nb,
+ buf_size,
+ nir_imm_int(&b->nb, offset)),
+ nir_imm_int(&b->nb, 0u)),
+ nir_imm_int(&b->nb, stride));
+
+ struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
+ val->ssa = vtn_create_ssa_value(b, glsl_uint_type());
+ val->ssa->def = array_length;
+ break;
+ }
+
+ case SpvOpCopyMemorySized:
+ default:
+ unreachable("Unhandled opcode");
+ }
+}
generated by cgit v1.2.3 (git 2.39.1) at 2024-11-30 05:25:50 +0000