nir: Add an ALU lowering pass for mul_high.

This is based on the glsl/lower_instructions.cpp implementation, but
should be much more readable.

Reviewed-by: Matt Turner <[email protected]>
Reviewed-by: Ian Romanick <[email protected]>

4 files changed, 170 insertions, 0 deletions

diff --git a/src/compiler/Makefile.sources b/src/compiler/Makefile.sources
index 3daa2c51334..d629c2b8ec8 100644
--- a/src/compiler/Makefile.sources
+++ b/src/compiler/Makefile.sources
@@ -214,6 +214,7 @@ NIR_FILES = \
 	nir/nir_loop_analyze.c \
 	nir/nir_loop_analyze.h \
 	nir/nir_lower_alpha_test.c \
+	nir/nir_lower_alu.c \
 	nir/nir_lower_alu_to_scalar.c \
 	nir/nir_lower_atomics_to_ssbo.c \
 	nir/nir_lower_bitmap.c \
diff --git a/src/compiler/nir/meson.build b/src/compiler/nir/meson.build
index 3fec363691d..598c68aff9f 100644
--- a/src/compiler/nir/meson.build
+++ b/src/compiler/nir/meson.build
@@ -104,6 +104,7 @@ files_libnir = files(
   'nir_liveness.c',
   'nir_loop_analyze.c',
   'nir_loop_analyze.h',
+  'nir_lower_alu.c',
   'nir_lower_alu_to_scalar.c',
   'nir_lower_alpha_test.c',
   'nir_lower_atomics_to_ssbo.c',
diff --git a/src/compiler/nir/nir.h b/src/compiler/nir/nir.h
index b9426f8eb4e..7d01eb23bc4 100644
--- a/src/compiler/nir/nir.h
+++ b/src/compiler/nir/nir.h
@@ -1919,6 +1919,8 @@ typedef struct nir_shader_compiler_options {
    bool lower_find_lsb;
    bool lower_uadd_carry;
    bool lower_usub_borrow;
+   /** Lowers imul_high/umul_high to 16-bit multiplies and carry operations. */
+   bool lower_mul_high;
    /** lowers fneg and ineg to fsub and isub. */
    bool lower_negate;
    /** lowers fsub and isub to fadd+fneg and iadd+ineg. */
@@ -2628,6 +2630,7 @@ bool nir_move_vec_src_uses_to_dest(nir_shader *shader);
 bool nir_lower_vec_to_movs(nir_shader *shader);
 void nir_lower_alpha_test(nir_shader *shader, enum compare_func func,
                           bool alpha_to_one);
+bool nir_lower_alu(nir_shader *shader);
 bool nir_lower_alu_to_scalar(nir_shader *shader);
 bool nir_lower_load_const_to_scalar(nir_shader *shader);
 bool nir_lower_read_invocation_to_scalar(nir_shader *shader);
diff --git a/src/compiler/nir/nir_lower_alu.c b/src/compiler/nir/nir_lower_alu.c
new file mode 100644
index 00000000000..28ecaf6badc
--- /dev/null
+++ b/src/compiler/nir/nir_lower_alu.c
@@ -0,0 +1,165 @@
+/*
+ * Copyright © 2010 Intel Corporation
+ * Copyright © 2018 Broadcom
+ *
+ * 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 "nir.h"
+#include "nir_builder.h"
+
+/** nir_lower_alu.c
+ *
+ * NIR's home for miscellaneous ALU operation lowering implementations.
+ *
+ * Most NIR ALU lowering occurs in nir_opt_algebraic.py, since it's generally
+ * easy to write them there.  However, if terms appear multiple times in the
+ * lowered code, it can get very verbose and cause a lot of work for CSE, so
+ * it may end up being easier to write out in C code.
+ *
+ * The shader must be in SSA for this pass.
+ */
+
+#define LOWER_MUL_HIGH (1 << 0)
+
+static bool
+lower_alu_instr(nir_alu_instr *instr, nir_builder *b)
+{
+   nir_ssa_def *lowered = NULL;
+
+   assert(instr->dest.dest.is_ssa);
+
+   b->cursor = nir_before_instr(&instr->instr);
+   b->exact = instr->exact;
+
+   switch (instr->op) {
+   case nir_op_imul_high:
+   case nir_op_umul_high:
+      if (b->shader->options->lower_mul_high) {
+         nir_ssa_def *c1 = nir_imm_int(b, 1);
+         nir_ssa_def *c16 = nir_imm_int(b, 16);
+
+         nir_ssa_def *src0 = nir_ssa_for_alu_src(b, instr, 0);
+         nir_ssa_def *src1 = nir_ssa_for_alu_src(b, instr, 1);
+         nir_ssa_def *different_signs = NULL;
+         if (instr->op == nir_op_imul_high) {
+            nir_ssa_def *c0 = nir_imm_int(b, 0);
+            different_signs = nir_ixor(b,
+                                       nir_ilt(b, src0, c0),
+                                       nir_ilt(b, src1, c0));
+            src0 = nir_iabs(b, src0);
+            src1 = nir_iabs(b, src1);
+         }
+
+         /*   ABCD
+          * * EFGH
+          * ======
+          * (GH * CD) + (GH * AB) << 16 + (EF * CD) << 16 + (EF * AB) << 32
+          *
+          * Start by splitting into the 4 multiplies.
+          */
+         nir_ssa_def *src0l = nir_iand(b, src0, nir_imm_int(b, 0xffff));
+         nir_ssa_def *src1l = nir_iand(b, src1, nir_imm_int(b, 0xffff));
+         nir_ssa_def *src0h = nir_ushr(b, src0, c16);
+         nir_ssa_def *src1h = nir_ushr(b, src1, c16);
+
+         nir_ssa_def *lo = nir_imul(b, src0l, src1l);
+         nir_ssa_def *m1 = nir_imul(b, src0l, src1h);
+         nir_ssa_def *m2 = nir_imul(b, src0h, src1l);
+         nir_ssa_def *hi = nir_imul(b, src0h, src1h);
+
+         nir_ssa_def *tmp;
+
+         tmp = nir_ishl(b, m1, c16);
+         hi = nir_iadd(b, hi, nir_iand(b, nir_uadd_carry(b, lo, tmp), c1));
+         lo = nir_iadd(b, lo, tmp);
+         hi = nir_iadd(b, hi, nir_ushr(b, m1, c16));
+
+         tmp = nir_ishl(b, m2, c16);
+         hi = nir_iadd(b, hi, nir_iand(b, nir_uadd_carry(b, lo, tmp), c1));
+         lo = nir_iadd(b, lo, tmp);
+         hi = nir_iadd(b, hi, nir_ushr(b, m2, c16));
+
+         if (instr->op == nir_op_imul_high) {
+            /* For channels where different_signs is set we have to perform a
+             * 64-bit negation.  This is *not* the same as just negating the
+             * high 32-bits.  Consider -3 * 2.  The high 32-bits is 0, but the
+             * desired result is -1, not -0!  Recall -x == ~x + 1.
+             */
+            hi = nir_bcsel(b, different_signs,
+                           nir_iadd(b,
+                                    nir_inot(b, hi),
+                                    nir_iand(b,
+                                             nir_uadd_carry(b,
+                                                            nir_inot(b, lo),
+                                                            c1),
+                                             nir_imm_int(b, 1))),
+                           hi);
+         }
+
+         lowered = hi;
+      }
+      break;
+
+   default:
+      break;
+   }
+
+   if (lowered) {
+      nir_ssa_def_rewrite_uses(&instr->dest.dest.ssa, nir_src_for_ssa(lowered));
+      nir_instr_remove(&instr->instr);
+      return true;
+   } else {
+      return false;
+   }
+}
+
+bool
+nir_lower_alu(nir_shader *shader)
+{
+   bool progress = false;
+
+   if (!shader->options->lower_mul_high)
+      return false;
+
+   nir_foreach_function(function, shader) {
+      if (function->impl) {
+         nir_builder builder;
+         nir_builder_init(&builder, function->impl);
+
+         nir_foreach_block(block, function->impl) {
+            nir_foreach_instr_safe(instr, block) {
+               if (instr->type == nir_instr_type_alu) {
+                  progress = lower_alu_instr(nir_instr_as_alu(instr),
+                                             &builder) || progress;
+               }
+            }
+         }
+
+         if (progress) {
+            nir_metadata_preserve(function->impl,
+                                  nir_metadata_block_index |
+                                  nir_metadata_dominance);
+         }
+      }
+   }
+
+   return progress;
+}