gallivm: Cleanups and bugfixes to aos format translation.

3 files changed, 126 insertions, 259 deletions

diff --git a/src/gallium/auxiliary/gallivm/lp_bld_format.h b/src/gallium/auxiliary/gallivm/lp_bld_format.h
index 13d0de875d7..ecf2cfd62c0 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_format.h
+++ b/src/gallium/auxiliary/gallivm/lp_bld_format.h
@@ -51,14 +51,6 @@ lp_build_unpack_rgba_aos(LLVMBuilderRef builder,
                          const struct util_format_description *desc,
                          LLVMValueRef packed);
 
-
-LLVMValueRef
-lp_build_unpack_rgba8_aos(LLVMBuilderRef builder,
-                          const struct util_format_description *desc,
-                          struct lp_type type,
-                          LLVMValueRef packed);
-
-
 LLVMValueRef
 lp_build_pack_rgba_aos(LLVMBuilderRef builder,
                        const struct util_format_description *desc,
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_format_aos.c b/src/gallium/auxiliary/gallivm/lp_bld_format_aos.c
index 4cf285efb89..191562d460d 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_format_aos.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_format_aos.c
@@ -33,8 +33,8 @@
  */
 
 
-#include "util/u_cpu_detect.h"
 #include "util/u_format.h"
+#include "util/u_math.h"
 
 #include "lp_bld_type.h"
 #include "lp_bld_const.h"
@@ -63,10 +63,11 @@ lp_build_unpack_rgba_aos(LLVMBuilderRef builder,
    LLVMValueRef aux[4];
    bool normalized;
    int empty_channel;
+   bool needs_uitofp;
    unsigned shift;
    unsigned i;
 
-   /* FIXME: Support more formats */
+   /* TODO: Support more formats */
    assert(desc->layout == UTIL_FORMAT_LAYOUT_PLAIN);
    assert(desc->block.width == 1);
    assert(desc->block.height == 1);
@@ -93,6 +94,7 @@ lp_build_unpack_rgba_aos(LLVMBuilderRef builder,
 
    /* Initialize vector constants */
    normalized = FALSE;
+   needs_uitofp = FALSE;
    empty_channel = -1;
    shift = 0;
    for (i = 0; i < 4; ++i) {
@@ -105,10 +107,13 @@ lp_build_unpack_rgba_aos(LLVMBuilderRef builder,
          empty_channel = i;
       }
       else {
-         unsigned mask = (1 << bits) - 1;
+         unsigned long long mask = (1ULL << bits) - 1;
 
          assert(desc->channel[i].type == UTIL_FORMAT_TYPE_UNSIGNED);
-         assert(bits < 32);
+
+         if (bits == 32) {
+            needs_uitofp = TRUE;
+         }
 
          shifts[i] = LLVMConstInt(LLVMInt32Type(), shift, 0);
          masks[i] = LLVMConstInt(LLVMInt32Type(), mask, 0);
@@ -126,8 +131,12 @@ lp_build_unpack_rgba_aos(LLVMBuilderRef builder,
 
    shifted = LLVMBuildLShr(builder, packed, LLVMConstVector(shifts, 4), "");
    masked = LLVMBuildAnd(builder, shifted, LLVMConstVector(masks, 4), "");
-   /* UIToFP can't be expressed in SSE2 */
-   casted = LLVMBuildSIToFP(builder, masked, LLVMVectorType(LLVMFloatType(), 4), "");
+   if (!needs_uitofp) {
+      /* UIToFP can't be expressed in SSE2 */
+      casted = LLVMBuildSIToFP(builder, masked, LLVMVectorType(LLVMFloatType(), 4), "");
+   } else {
+      casted = LLVMBuildUIToFP(builder, masked, LLVMVectorType(LLVMFloatType(), 4), "");
+   }
 
    if (normalized)
       scaled = LLVMBuildMul(builder, casted, LLVMConstVector(scales, 4), "");
@@ -138,7 +147,22 @@ lp_build_unpack_rgba_aos(LLVMBuilderRef builder,
       aux[i] = LLVMGetUndef(LLVMFloatType());
 
    for (i = 0; i < 4; ++i) {
-      enum util_format_swizzle swizzle = desc->swizzle[i];
+      enum util_format_swizzle swizzle;
+
+      if (desc->colorspace == UTIL_FORMAT_COLORSPACE_ZS) {
+         /*
+          * For ZS formats do RGBA = ZZZ1
+          */
+         if (i == 3) {
+            swizzle = UTIL_FORMAT_SWIZZLE_1;
+         } else if (desc->swizzle[0] == UTIL_FORMAT_SWIZZLE_NONE) {
+            swizzle = UTIL_FORMAT_SWIZZLE_0;
+         } else {
+            swizzle = desc->swizzle[0];
+         }
+      } else {
+         swizzle = desc->swizzle[i];
+      }
 
       switch (swizzle) {
       case UTIL_FORMAT_SWIZZLE_X:
@@ -167,117 +191,6 @@ lp_build_unpack_rgba_aos(LLVMBuilderRef builder,
 
 
 /**
- * Take a vector with packed pixels and unpack into a rgba8 vector.
- *
- * Formats with bit depth smaller than 32bits are accepted, but they must be
- * padded to 32bits.
- */
-LLVMValueRef
-lp_build_unpack_rgba8_aos(LLVMBuilderRef builder,
-                          const struct util_format_description *desc,
-                          struct lp_type type,
-                          LLVMValueRef packed)
-{
-   struct lp_build_context bld;
-   bool rgba8;
-   LLVMValueRef res;
-   unsigned i;
-
-   lp_build_context_init(&bld, builder, type);
-
-   /* FIXME: Support more formats */
-   assert(desc->layout == UTIL_FORMAT_LAYOUT_PLAIN);
-   assert(desc->block.width == 1);
-   assert(desc->block.height == 1);
-   assert(desc->block.bits <= 32);
-
-   assert(!type.floating);
-   assert(!type.fixed);
-   assert(type.norm);
-   assert(type.width == 8);
-   assert(type.length % 4 == 0);
-
-   rgba8 = TRUE;
-   for(i = 0; i < 4; ++i) {
-      assert(desc->channel[i].type == UTIL_FORMAT_TYPE_UNSIGNED ||
-             desc->channel[i].type == UTIL_FORMAT_TYPE_VOID);
-      if(desc->channel[0].size != 8)
-         rgba8 = FALSE;
-   }
-
-   if(rgba8) {
-      /*
-       * The pixel is already in a rgba8 format variant. All it is necessary
-       * is to swizzle the channels.
-       */
-
-      unsigned char swizzles[4];
-      boolean zeros[4]; /* bitwise AND mask */
-      boolean ones[4]; /* bitwise OR mask */
-      boolean swizzles_needed = FALSE;
-      boolean zeros_needed = FALSE;
-      boolean ones_needed = FALSE;
-
-      for(i = 0; i < 4; ++i) {
-         enum util_format_swizzle swizzle = desc->swizzle[i];
-
-         /* Initialize with the no-op case */
-         swizzles[i] = util_cpu_caps.little_endian ? 3 - i : i;
-         zeros[i] = TRUE;
-         ones[i] = FALSE;
-
-         switch (swizzle) {
-         case UTIL_FORMAT_SWIZZLE_X:
-         case UTIL_FORMAT_SWIZZLE_Y:
-         case UTIL_FORMAT_SWIZZLE_Z:
-         case UTIL_FORMAT_SWIZZLE_W:
-            if(swizzle != swizzles[i]) {
-               swizzles[i] = swizzle;
-               swizzles_needed = TRUE;
-            }
-            break;
-         case UTIL_FORMAT_SWIZZLE_0:
-            zeros[i] = FALSE;
-            zeros_needed = TRUE;
-            break;
-         case UTIL_FORMAT_SWIZZLE_1:
-            ones[i] = TRUE;
-            ones_needed = TRUE;
-            break;
-         case UTIL_FORMAT_SWIZZLE_NONE:
-            assert(0);
-            break;
-         }
-      }
-
-      res = packed;
-
-      if(swizzles_needed)
-         res = lp_build_swizzle1_aos(&bld, res, swizzles);
-
-      if(zeros_needed) {
-         /* Mask out zero channels */
-         LLVMValueRef mask = lp_build_const_mask_aos(type, zeros);
-         res = LLVMBuildAnd(builder, res, mask, "");
-      }
-
-      if(ones_needed) {
-         /* Or one channels */
-         LLVMValueRef mask = lp_build_const_mask_aos(type, ones);
-         res = LLVMBuildOr(builder, res, mask, "");
-      }
-   }
-   else {
-      /* FIXME */
-      assert(0);
-      res = lp_build_undef(type);
-   }
-
-   return res;
-}
-
-
-/**
  * Pack a single pixel.
  *
  * @param rgba 4 float vector with the unpacked components.
@@ -393,13 +306,16 @@ lp_build_fetch_rgba_aos(LLVMBuilderRef builder,
         format_desc->colorspace == UTIL_FORMAT_COLORSPACE_ZS) &&
        format_desc->block.width == 1 &&
        format_desc->block.height == 1 &&
+       util_is_pot(format_desc->block.bits) &&
        format_desc->block.bits <= 32 &&
-       format_desc->channel[0].type != UTIL_FORMAT_TYPE_FLOAT)
+       format_desc->is_bitmask &&
+       !format_desc->is_mixed)
    {
       LLVMValueRef packed;
 
       ptr = LLVMBuildBitCast(builder, ptr,
-                             LLVMPointerType(LLVMIntType(format_desc->block.bits), 0) , "");
+                             LLVMPointerType(LLVMIntType(format_desc->block.bits), 0) ,
+                             "");
 
       packed = LLVMBuildLoad(builder, ptr, "packed");
 
diff --git a/src/gallium/drivers/llvmpipe/lp_test_format.c b/src/gallium/drivers/llvmpipe/lp_test_format.c
index fb595893bd0..13c3c3572d6 100644
--- a/src/gallium/drivers/llvmpipe/lp_test_format.c
+++ b/src/gallium/drivers/llvmpipe/lp_test_format.c
@@ -37,59 +37,13 @@
 
 #include "util/u_cpu_detect.h"
 #include "util/u_format.h"
+#include "util/u_format_tests.h"
+#include "util/u_format_s3tc.h"
 
 #include "gallivm/lp_bld_format.h"
 #include "lp_test.h"
 
 
-struct pixel_test_case
-{
-   enum pipe_format format;
-   uint32_t packed;
-   double unpacked[4];
-};
-
-
-struct pixel_test_case test_cases[] =
-{
-   {PIPE_FORMAT_B5G6R5_UNORM,   0x0000, {0.0, 0.0, 0.0, 1.0}},
-   {PIPE_FORMAT_B5G6R5_UNORM,   0x001f, {0.0, 0.0, 1.0, 1.0}},
-   {PIPE_FORMAT_B5G6R5_UNORM,   0x07e0, {0.0, 1.0, 0.0, 1.0}},
-   {PIPE_FORMAT_B5G6R5_UNORM,   0xf800, {1.0, 0.0, 0.0, 1.0}},
-   {PIPE_FORMAT_B5G6R5_UNORM,   0xffff, {1.0, 1.0, 1.0, 1.0}},
-
-   {PIPE_FORMAT_B5G5R5A1_UNORM, 0x0000, {0.0, 0.0, 0.0, 0.0}},
-   {PIPE_FORMAT_B5G5R5A1_UNORM, 0x001f, {0.0, 0.0, 1.0, 0.0}},
-   {PIPE_FORMAT_B5G5R5A1_UNORM, 0x03e0, {0.0, 1.0, 0.0, 0.0}},
-   {PIPE_FORMAT_B5G5R5A1_UNORM, 0x7c00, {1.0, 0.0, 0.0, 0.0}},
-   {PIPE_FORMAT_B5G5R5A1_UNORM, 0x8000, {0.0, 0.0, 0.0, 1.0}},
-   {PIPE_FORMAT_B5G5R5A1_UNORM, 0xffff, {1.0, 1.0, 1.0, 1.0}},
-
-   {PIPE_FORMAT_B8G8R8A8_UNORM, 0x00000000, {0.0, 0.0, 0.0, 0.0}},
-   {PIPE_FORMAT_B8G8R8A8_UNORM, 0x000000ff, {0.0, 0.0, 1.0, 0.0}},
-   {PIPE_FORMAT_B8G8R8A8_UNORM, 0x0000ff00, {0.0, 1.0, 0.0, 0.0}},
-   {PIPE_FORMAT_B8G8R8A8_UNORM, 0x00ff0000, {1.0, 0.0, 0.0, 0.0}},
-   {PIPE_FORMAT_B8G8R8A8_UNORM, 0xff000000, {0.0, 0.0, 0.0, 1.0}},
-   {PIPE_FORMAT_B8G8R8A8_UNORM, 0xffffffff, {1.0, 1.0, 1.0, 1.0}},
-
-#if 0
-   {PIPE_FORMAT_R8G8B8A8_UNORM, 0x00000000, {0.0, 0.0, 0.0, 0.0}},
-   {PIPE_FORMAT_R8G8B8A8_UNORM, 0x000000ff, {0.0, 0.0, 0.0, 1.0}},
-   {PIPE_FORMAT_R8G8B8A8_UNORM, 0x0000ff00, {0.0, 0.0, 1.0, 0.0}},
-   {PIPE_FORMAT_R8G8B8A8_UNORM, 0x00ff0000, {0.0, 1.0, 0.0, 0.0}},
-   {PIPE_FORMAT_R8G8B8A8_UNORM, 0xff000000, {1.0, 0.0, 0.0, 0.0}},
-   {PIPE_FORMAT_R8G8B8A8_UNORM, 0xffffffff, {1.0, 1.0, 1.0, 1.0}},
-#endif
-
-   {PIPE_FORMAT_A8R8G8B8_UNORM, 0x00000000, {0.0, 0.0, 0.0, 0.0}},
-   {PIPE_FORMAT_A8R8G8B8_UNORM, 0x000000ff, {0.0, 0.0, 0.0, 1.0}},
-   {PIPE_FORMAT_A8R8G8B8_UNORM, 0x0000ff00, {1.0, 0.0, 0.0, 0.0}},
-   {PIPE_FORMAT_A8R8G8B8_UNORM, 0x00ff0000, {0.0, 1.0, 0.0, 0.0}},
-   {PIPE_FORMAT_A8R8G8B8_UNORM, 0xff000000, {0.0, 0.0, 1.0, 0.0}},
-   {PIPE_FORMAT_A8R8G8B8_UNORM, 0xffffffff, {1.0, 1.0, 1.0, 1.0}},
-};
-
-
 void
 write_tsv_header(FILE *fp)
 {
@@ -114,52 +68,11 @@ write_tsv_row(FILE *fp,
 }
 
 
-typedef void (*load_ptr_t)(const uint32_t packed, float *);
+typedef void (*fetch_ptr_t)(const void *packed, float *);
 
 
 static LLVMValueRef
-add_load_rgba_test(LLVMModuleRef module,
-                   const struct util_format_description *desc)
-{
-   LLVMTypeRef args[2];
-   LLVMValueRef func;
-   LLVMValueRef packed;
-   LLVMValueRef rgba_ptr;
-   LLVMBasicBlockRef block;
-   LLVMBuilderRef builder;
-   LLVMValueRef rgba;
-
-   args[0] = LLVMInt32Type();
-   args[1] = LLVMPointerType(LLVMVectorType(LLVMFloatType(), 4), 0);
-
-   func = LLVMAddFunction(module, "load", LLVMFunctionType(LLVMVoidType(), args, 2, 0));
-   LLVMSetFunctionCallConv(func, LLVMCCallConv);
-   packed = LLVMGetParam(func, 0);
-   rgba_ptr = LLVMGetParam(func, 1);
-
-   block = LLVMAppendBasicBlock(func, "entry");
-   builder = LLVMCreateBuilder();
-   LLVMPositionBuilderAtEnd(builder, block);
-
-   if(desc->block.bits < 32)
-      packed = LLVMBuildTrunc(builder, packed, LLVMIntType(desc->block.bits), "");
-
-   rgba = lp_build_unpack_rgba_aos(builder, desc, packed);
-
-   LLVMBuildStore(builder, rgba, rgba_ptr);
-
-   LLVMBuildRetVoid(builder);
-
-   LLVMDisposeBuilder(builder);
-   return func;
-}
-
-
-typedef void (*store_ptr_t)(uint32_t *, const float *);
-
-
-static LLVMValueRef
-add_store_rgba_test(LLVMModuleRef module,
+add_fetch_rgba_test(LLVMModuleRef module,
                     const struct util_format_description *desc)
 {
    LLVMTypeRef args[2];
@@ -169,12 +82,11 @@ add_store_rgba_test(LLVMModuleRef module,
    LLVMBasicBlockRef block;
    LLVMBuilderRef builder;
    LLVMValueRef rgba;
-   LLVMValueRef packed;
 
-   args[0] = LLVMPointerType(LLVMInt32Type(), 0);
+   args[0] = LLVMPointerType(LLVMInt8Type(), 0);
    args[1] = LLVMPointerType(LLVMVectorType(LLVMFloatType(), 4), 0);
 
-   func = LLVMAddFunction(module, "store", LLVMFunctionType(LLVMVoidType(), args, 2, 0));
+   func = LLVMAddFunction(module, "fetch", LLVMFunctionType(LLVMVoidType(), args, 2, 0));
    LLVMSetFunctionCallConv(func, LLVMCCallConv);
    packed_ptr = LLVMGetParam(func, 0);
    rgba_ptr = LLVMGetParam(func, 1);
@@ -183,14 +95,9 @@ add_store_rgba_test(LLVMModuleRef module,
    builder = LLVMCreateBuilder();
    LLVMPositionBuilderAtEnd(builder, block);
 
-   rgba = LLVMBuildLoad(builder, rgba_ptr, "");
-
-   packed = lp_build_pack_rgba_aos(builder, desc, rgba);
+   rgba = lp_build_fetch_rgba_aos(builder, desc, packed_ptr);
 
-   if(desc->block.bits < 32)
-      packed = LLVMBuildZExt(builder, packed, LLVMInt32Type(), "");
-
-   LLVMBuildStore(builder, packed, packed_ptr);
+   LLVMBuildStore(builder, rgba, rgba_ptr);
 
    LLVMBuildRetVoid(builder);
 
@@ -201,30 +108,24 @@ add_store_rgba_test(LLVMModuleRef module,
 
 PIPE_ALIGN_STACK
 static boolean
-test_format(unsigned verbose, FILE *fp, const struct pixel_test_case *test)
+test_format(unsigned verbose, FILE *fp,
+            const struct util_format_description *desc,
+            const struct util_format_test_case *test)
 {
    LLVMModuleRef module = NULL;
-   LLVMValueRef load = NULL;
-   LLVMValueRef store = NULL;
+   LLVMValueRef fetch = NULL;
    LLVMExecutionEngineRef engine = NULL;
    LLVMModuleProviderRef provider = NULL;
    LLVMPassManagerRef pass = NULL;
    char *error = NULL;
-   const struct util_format_description *desc;
-   load_ptr_t load_ptr;
-   store_ptr_t store_ptr;
+   fetch_ptr_t fetch_ptr;
    float unpacked[4];
-   unsigned packed;
    boolean success;
    unsigned i;
 
-   desc = util_format_description(test->format);
-   fprintf(stderr, "%s\n", desc->name);
-
    module = LLVMModuleCreateWithName("test");
 
-   load = add_load_rgba_test(module, desc);
-   store = add_store_rgba_test(module, desc);
+   fetch = add_fetch_rgba_test(module, desc);
 
    if(LLVMVerifyModule(module, LLVMPrintMessageAction, &error)) {
       LLVMDumpModule(module);
@@ -254,33 +155,32 @@ test_format(unsigned verbose, FILE *fp, const struct pixel_test_case *test)
    (void)pass;
 #endif
 
-   load_ptr  = (load_ptr_t) LLVMGetPointerToGlobal(engine, load);
-   store_ptr = (store_ptr_t)LLVMGetPointerToGlobal(engine, store);
+   fetch_ptr  = (fetch_ptr_t) LLVMGetPointerToGlobal(engine, fetch);
 
    memset(unpacked, 0, sizeof unpacked);
-   packed = 0;
 
-   load_ptr(test->packed, unpacked);
-   store_ptr(&packed, unpacked);
+   fetch_ptr(test->packed, unpacked);
 
    success = TRUE;
-   if(test->packed != packed)
-      success = FALSE;
    for(i = 0; i < 4; ++i)
-      if(test->unpacked[i] != unpacked[i])
+      if(test->unpacked[0][0][i] != unpacked[i])
          success = FALSE;
 
    if (!success) {
       printf("FAILED\n");
-      printf("  Packed: %08x\n", test->packed);
-      printf("          %08x\n", packed);
-      printf("  Unpacked: %f %f %f %f\n", unpacked[0], unpacked[1], unpacked[2], unpacked[3]);
-      printf("            %f %f %f %f\n", test->unpacked[0], test->unpacked[1], test->unpacked[2], test->unpacked[3]);
+      printf("  Packed: %02x %02x %02x %02x\n",
+             test->packed[0], test->packed[1], test->packed[2], test->packed[3]);
+      printf("  Unpacked: %f %f %f %f obtained\n",
+             unpacked[0], unpacked[1], unpacked[2], unpacked[3]);
+      printf("            %f %f %f %f expected\n",
+             test->unpacked[0][0][0],
+             test->unpacked[0][0][1],
+             test->unpacked[0][0][2],
+             test->unpacked[0][0][3]);
       LLVMDumpModule(module);
    }
 
-   LLVMFreeMachineCodeForFunction(engine, store);
-   LLVMFreeMachineCodeForFunction(engine, load);
+   LLVMFreeMachineCodeForFunction(engine, fetch);
 
    LLVMDisposeExecutionEngine(engine);
    if(pass)
@@ -293,15 +193,74 @@ test_format(unsigned verbose, FILE *fp, const struct pixel_test_case *test)
 }
 
 
+
+static boolean
+test_one(unsigned verbose, FILE *fp,
+         const struct util_format_description *format_desc)
+{
+   unsigned i;
+   bool success = TRUE;
+
+   printf("Testing %s ...\n",
+          format_desc->name);
+
+   for (i = 0; i < util_format_nr_test_cases; ++i) {
+      const struct util_format_test_case *test = &util_format_test_cases[i];
+
+      if (test->format == format_desc->format) {
+
+         if (!test_format(verbose, fp, format_desc, test)) {
+           success = FALSE;
+         }
+
+      }
+   }
+
+   return success;
+}
+
+
 boolean
 test_all(unsigned verbose, FILE *fp)
 {
-   unsigned i;
+   enum pipe_format format;
    bool success = TRUE;
 
-   for (i = 0; i < sizeof(test_cases)/sizeof(test_cases[0]); ++i)
-      if(!test_format(verbose, fp, &test_cases[i]))
-        success = FALSE;
+   for (format = 1; format < PIPE_FORMAT_COUNT; ++format) {
+      const struct util_format_description *format_desc;
+
+      format_desc = util_format_description(format);
+      if (!format_desc) {
+         continue;
+      }
+
+      /*
+       * XXX: copied from lp_build_fetch_rgba_aos()
+       * TODO: test more
+       */
+
+      if (!(format_desc->layout == UTIL_FORMAT_LAYOUT_PLAIN &&
+            format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB &&
+            format_desc->block.width == 1 &&
+            format_desc->block.height == 1 &&
+            util_is_pot(format_desc->block.bits) &&
+            format_desc->block.bits <= 32 &&
+            format_desc->is_bitmask &&
+            !format_desc->is_mixed &&
+            (format_desc->channel[0].type == UTIL_FORMAT_TYPE_UNSIGNED ||
+             format_desc->channel[1].type == UTIL_FORMAT_TYPE_UNSIGNED))) {
+         continue;
+      }
+
+      if (format_desc->layout == UTIL_FORMAT_LAYOUT_S3TC &&
+          !util_format_s3tc_enabled) {
+         continue;
+      }
+
+      if (!test_one(verbose, fp, format_desc)) {
+           success = FALSE;
+      }
+   }
 
    return success;
 }