Translate .Load() to imageLoad() for Vulkan (#967)

* Translate .Load() to imageLoad() for Vulkan

We were already emitting calls to `imageLoad()` and `imageStore` when a `RWTexture*` was used with `operator[]`:

```hlsl
RWTexture2D<float> myTex;
...
float value = myTex[xy]; // becomes an imageLoad
myTex[xy] = value;       // becomes an imageStore
```

However, we were *not* correctly handling the translation of an explicit `.Load()` operation:

```hlsl
float value = myTex.Load(xy);
```

The `.Load()` operation was being translated to a GLSL `texelFetch` as it would be a for a `Texture2D`, and not to an `imageLoad()` as would make sense for a `RWTexture2D` (which becomes a GLSL `image2D`).

This fix is confined to the stdlib, and is mostly a matter of emitting either `texelFetch` or `imageLoad` as the GLSL function name depending on the "access" of the resource type. It is messy code, but straightforward.

One extra detail was that there had been logic to emit a `, 0` argument in the `texelFetch` calls in the non-read-only case, because `texelFetch` usualy requires an explicit mip-level argument and `.Load()` on a `RWTexture*` doesn't recieve an LOD parameter. This is a non-issue now that we are calling `imageLoad()` instead, because `imageLoad` doesn't need/want the extra argument.

* fixup: change test baseline based on recent GLSL output changes

* fixup: review feedback

6 files changed, 72 insertions, 20 deletions

diff --git a/source/slang/core.meta.slang b/source/slang/core.meta.slang
index a3d73f66a..bc9c9b50a 100644
--- a/source/slang/core.meta.slang
+++ b/source/slang/core.meta.slang
@@ -687,6 +687,8 @@ for (int tt = 0; tt < kBaseTextureTypeCount; ++tt)
 
                 int loadCoordCount = kBaseTextureTypes[tt].coordCount + isArray + (needsMipLevel?1:0);
 
+                char const* glslFuncName = (access == SLANG_RESOURCE_ACCESS_READ) ? "texelFetch" : "imageLoad";
+
                 // When translating to GLSL, we need to break apart the `location` argument.
                 //
                 // TODO: this should realy be handled by having this member actually get lowered!
@@ -703,19 +705,19 @@ for (int tt = 0; tt < kBaseTextureTypeCount; ++tt)
                 if (isMultisample)
                 {
                     sb << "__glsl_extension(GL_EXT_samplerless_texture_functions)";
-                    sb << "__target_intrinsic(glsl, \"$ctexelFetch($0, $1, $3)$z\")\n";
+                    sb << "__target_intrinsic(glsl, \"$c" << glslFuncName << "($0, $1, $3)$z\")\n";
                 }
                 else
                 {
                     sb << "__glsl_extension(GL_EXT_samplerless_texture_functions)";
-                    sb << "__target_intrinsic(glsl, \"$ctexelFetch($0, ";
+                    sb << "__target_intrinsic(glsl, \"$c" << glslFuncName << "($0, ";
                     if( needsMipLevel )
                     {
                         sb << "($1)." << kGLSLLoadCoordsSwizzle[loadCoordCount] << ", ($1)." << kGLSLLoadLODSwizzle[loadCoordCount];
                     }
                     else
                     {
-                        sb << "$1, 0";
+                        sb << "$1";
                     }
                     sb << ")$z\")\n";
                 }
@@ -730,12 +732,12 @@ for (int tt = 0; tt < kBaseTextureTypeCount; ++tt)
                 if (isMultisample)
                 {
                     sb << "__glsl_extension(GL_EXT_samplerless_texture_functions)";
-                    sb << "__target_intrinsic(glsl, \"$ctexelFetchOffset($0, $0, $1, $2)$z\")\n";
+                    sb << "__target_intrinsic(glsl, \"$c" << glslFuncName << "($0, $0, $1, $2)$z\")\n";
                 }
                 else
                 {
                     sb << "__glsl_extension(GL_EXT_samplerless_texture_functions)";
-                    sb << "__target_intrinsic(glsl, \"$ctexelFetch($0, ";
+                    sb << "__target_intrinsic(glsl, \"$c" << glslFuncName << "($0, ";
                     if( needsMipLevel )
                     {
                         sb << "($1)." << kGLSLLoadCoordsSwizzle[loadCoordCount] << ", ($1)." << kGLSLLoadLODSwizzle[loadCoordCount];
diff --git a/source/slang/core.meta.slang.h b/source/slang/core.meta.slang.h
index ba6636ffc..8362a2c4b 100644
--- a/source/slang/core.meta.slang.h
+++ b/source/slang/core.meta.slang.h
@@ -705,6 +705,8 @@ for (int tt = 0; tt < kBaseTextureTypeCount; ++tt)
 
                 int loadCoordCount = kBaseTextureTypes[tt].coordCount + isArray + (needsMipLevel?1:0);
 
+                char const* glslFuncName = (access == SLANG_RESOURCE_ACCESS_READ) ? "texelFetch" : "imageLoad";
+
                 // When translating to GLSL, we need to break apart the `location` argument.
                 //
                 // TODO: this should realy be handled by having this member actually get lowered!
@@ -721,19 +723,19 @@ for (int tt = 0; tt < kBaseTextureTypeCount; ++tt)
                 if (isMultisample)
                 {
                     sb << "__glsl_extension(GL_EXT_samplerless_texture_functions)";
-                    sb << "__target_intrinsic(glsl, \"$ctexelFetch($0, $1, $3)$z\")\n";
+                    sb << "__target_intrinsic(glsl, \"$c" << glslFuncName << "($0, $1, $3)$z\")\n";
                 }
                 else
                 {
                     sb << "__glsl_extension(GL_EXT_samplerless_texture_functions)";
-                    sb << "__target_intrinsic(glsl, \"$ctexelFetch($0, ";
+                    sb << "__target_intrinsic(glsl, \"$c" << glslFuncName << "($0, ";
                     if( needsMipLevel )
                     {
                         sb << "($1)." << kGLSLLoadCoordsSwizzle[loadCoordCount] << ", ($1)." << kGLSLLoadLODSwizzle[loadCoordCount];
                     }
                     else
                     {
-                        sb << "$1, 0";
+                        sb << "$1";
                     }
                     sb << ")$z\")\n";
                 }
@@ -748,12 +750,12 @@ for (int tt = 0; tt < kBaseTextureTypeCount; ++tt)
                 if (isMultisample)
                 {
                     sb << "__glsl_extension(GL_EXT_samplerless_texture_functions)";
-                    sb << "__target_intrinsic(glsl, \"$ctexelFetchOffset($0, $0, $1, $2)$z\")\n";
+                    sb << "__target_intrinsic(glsl, \"$c" << glslFuncName << "($0, $0, $1, $2)$z\")\n";
                 }
                 else
                 {
                     sb << "__glsl_extension(GL_EXT_samplerless_texture_functions)";
-                    sb << "__target_intrinsic(glsl, \"$ctexelFetch($0, ";
+                    sb << "__target_intrinsic(glsl, \"$c" << glslFuncName << "($0, ";
                     if( needsMipLevel )
                     {
                         sb << "($1)." << kGLSLLoadCoordsSwizzle[loadCoordCount] << ", ($1)." << kGLSLLoadLODSwizzle[loadCoordCount];
@@ -1205,7 +1207,7 @@ for (auto op : binaryOps)
         sb << "__intrinsic_op(" << int(op.opCode) << ") matrix<" << resultType << ",N,M> operator" << op.opName << "(" << leftQual << "matrix<" << leftType << ",N,M> left, " << rightType << " right);\n";
     }
 }
-SLANG_RAW("#line 1190 \"core.meta.slang\"")
+SLANG_RAW("#line 1192 \"core.meta.slang\"")
 SLANG_RAW("\n")
 SLANG_RAW("\n")
 SLANG_RAW("// Operators to apply to `enum` types\n")
diff --git a/source/slang/hlsl.meta.slang b/source/slang/hlsl.meta.slang
index 40bd02597..e2d79e073 100644
--- a/source/slang/hlsl.meta.slang
+++ b/source/slang/hlsl.meta.slang
@@ -1429,7 +1429,8 @@ static const int kBaseBufferAccessLevelCount = sizeof(kBaseBufferAccessLevels) /
 
 for (int aa = 0; aa < kBaseBufferAccessLevelCount; ++aa)
 {
-    auto flavor = TextureFlavor::create(TextureFlavor::Shape::ShapeBuffer, kBaseBufferAccessLevels[aa].access).flavor;
+    auto access = kBaseBufferAccessLevels[aa].access;
+    auto flavor = TextureFlavor::create(TextureFlavor::Shape::ShapeBuffer, access).flavor;
     sb << "__generic<T>\n";
     sb << "__magic_type(Texture," << int(flavor) << ")\n";
     sb << "__intrinsic_type(" << (kIROp_TextureType + (int(flavor) << kIROpMeta_OtherShift)) << ")\n";
@@ -1439,8 +1440,10 @@ for (int aa = 0; aa < kBaseBufferAccessLevelCount; ++aa)
 
     sb << "void GetDimensions(out uint dim);\n";
 
+    char const* glslLoadFuncName = (access == SLANG_RESOURCE_ACCESS_READ) ? "texelFetch" : "imageLoad";
+
     sb << "__glsl_extension(GL_EXT_samplerless_texture_functions)";
-    sb << "__target_intrinsic(glsl, \"texelFetch($0, $1)$z\")\n";
+    sb << "__target_intrinsic(glsl, \"" << glslLoadFuncName << "($0, $1)$z\")\n";
     sb << "T Load(int location);\n";
 
     sb << "T Load(int location, out uint status);\n";
@@ -1448,9 +1451,9 @@ for (int aa = 0; aa < kBaseBufferAccessLevelCount; ++aa)
     sb << "__subscript(uint index) -> T {\n";
 
     sb << "__glsl_extension(GL_EXT_samplerless_texture_functions)";
-    sb << "__target_intrinsic(glsl, \"texelFetch($0, int($1))$z\") get;\n";
+    sb << "__target_intrinsic(glsl, \"" << glslLoadFuncName << "($0, int($1))$z\") get;\n";
 
-    if (kBaseBufferAccessLevels[aa].access != SLANG_RESOURCE_ACCESS_READ)
+    if (access != SLANG_RESOURCE_ACCESS_READ)
     {
         sb << "ref;\n";
     }
diff --git a/source/slang/hlsl.meta.slang.h b/source/slang/hlsl.meta.slang.h
index b7bbc3985..fa018189c 100644
--- a/source/slang/hlsl.meta.slang.h
+++ b/source/slang/hlsl.meta.slang.h
@@ -1505,7 +1505,8 @@ static const int kBaseBufferAccessLevelCount = sizeof(kBaseBufferAccessLevels) /
 
 for (int aa = 0; aa < kBaseBufferAccessLevelCount; ++aa)
 {
-    auto flavor = TextureFlavor::create(TextureFlavor::Shape::ShapeBuffer, kBaseBufferAccessLevels[aa].access).flavor;
+    auto access = kBaseBufferAccessLevels[aa].access;
+    auto flavor = TextureFlavor::create(TextureFlavor::Shape::ShapeBuffer, access).flavor;
     sb << "__generic<T>\n";
     sb << "__magic_type(Texture," << int(flavor) << ")\n";
     sb << "__intrinsic_type(" << (kIROp_TextureType + (int(flavor) << kIROpMeta_OtherShift)) << ")\n";
@@ -1515,8 +1516,10 @@ for (int aa = 0; aa < kBaseBufferAccessLevelCount; ++aa)
 
     sb << "void GetDimensions(out uint dim);\n";
 
+    char const* glslLoadFuncName = (access == SLANG_RESOURCE_ACCESS_READ) ? "texelFetch" : "imageLoad";
+
     sb << "__glsl_extension(GL_EXT_samplerless_texture_functions)";
-    sb << "__target_intrinsic(glsl, \"texelFetch($0, $1)$z\")\n";
+    sb << "__target_intrinsic(glsl, \"" << glslLoadFuncName << "($0, $1)$z\")\n";
     sb << "T Load(int location);\n";
 
     sb << "T Load(int location, out uint status);\n";
@@ -1524,9 +1527,9 @@ for (int aa = 0; aa < kBaseBufferAccessLevelCount; ++aa)
     sb << "__subscript(uint index) -> T {\n";
 
     sb << "__glsl_extension(GL_EXT_samplerless_texture_functions)";
-    sb << "__target_intrinsic(glsl, \"texelFetch($0, int($1))$z\") get;\n";
+    sb << "__target_intrinsic(glsl, \"" << glslLoadFuncName << "($0, int($1))$z\") get;\n";
 
-    if (kBaseBufferAccessLevels[aa].access != SLANG_RESOURCE_ACCESS_READ)
+    if (access != SLANG_RESOURCE_ACCESS_READ)
     {
         sb << "ref;\n";
     }
@@ -1535,7 +1538,7 @@ for (int aa = 0; aa < kBaseBufferAccessLevelCount; ++aa)
 
     sb << "};\n";
 }
-SLANG_RAW("#line 1462 \"hlsl.meta.slang\"")
+SLANG_RAW("#line 1465 \"hlsl.meta.slang\"")
 SLANG_RAW("\n")
 SLANG_RAW("\n")
 SLANG_RAW("\n")
diff --git a/tests/cross-compile/image-load.slang b/tests/cross-compile/image-load.slang
new file mode 100644
index 000000000..7f125bd5a
--- /dev/null
+++ b/tests/cross-compile/image-load.slang
@@ -0,0 +1,19 @@
+// image-load.slang
+
+// This test confirms that `.Load()` on a `RWTexure*`
+// gets properly converted to a call to `imageLoad`
+// and not just `texelFetch` as it would for a `Texture*`.
+
+//TEST:CROSS_COMPILE:-target spirv-assembly -entry main -stage compute
+
+struct Params
+{
+	RWTexture2DArray<float> tex;
+}
+
+ParameterBlock<Params> gParams;
+
+void main(uint3 tid : SV_DispatchThreadID)
+{
+	float f = gParams.tex.Load(int3(tid.xy, tid.z));
+}
diff --git a/tests/cross-compile/image-load.slang.glsl b/tests/cross-compile/image-load.slang.glsl
new file mode 100644
index 000000000..c7233c38c
--- /dev/null
+++ b/tests/cross-compile/image-load.slang.glsl
@@ -0,0 +1,23 @@
+// image-load.slang.glsl
+//TEST_IGNORE_FILE:
+
+#version 450
+
+#extension GL_EXT_samplerless_texture_functions : require
+
+layout(r32f)
+layout(binding = 0)
+uniform image2DArray gParams_tex_0;
+
+layout(local_size_x = 1, local_size_y = 1, local_size_z = 1) in;
+
+void main()
+{
+    float _S1 = imageLoad(
+    	gParams_tex_0,
+    	ivec3(
+    		ivec2(gl_GlobalInvocationID.xy),
+    		int(gl_GlobalInvocationID.z))).x;
+
+    return;
+}