[SM68] Add feature flag for SampleCmpBias/Grad and StartInstance/VertexLocation (#6164)

New feature flag ShaderFeatureInfo_SampleCmpGradientOrBias = 0x80000000 and ShaderFeatureInfo_ExtendedCommandInfo = 0x100000000 were added. They will be set when SampleCmpBias/Grad startInstance/VertexLocation exists. --------- Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
2024-01-19 20:01:56 -08:00 · 2024-01-19 20:01:56 -08:00 · ae9b762a30
--- a/include/dxc/DXIL/DxilConstants.h
+++ b/include/dxc/DXIL/DxilConstants.h
@ -1890,7 +1890,10 @@ const uint64_t ShaderFeatureInfo_WriteableMSAATextures = 0x40000000;
 // WaveMMA slots in between two SM 6.6 feature bits.
 const uint64_t ShaderFeatureInfo_WaveMMA = 0x8000000;

-const unsigned ShaderFeatureInfoCount = 31;
+const uint64_t ShaderFeatureInfo_SampleCmpGradientOrBias = 0x80000000;
+const uint64_t ShaderFeatureInfo_ExtendedCommandInfo = 0x100000000;
+
+const unsigned ShaderFeatureInfoCount = 33;

 // DxilSubobjectType must match D3D12_STATE_SUBOBJECT_TYPE, with
 // certain values reserved, since they cannot be used from Dxil.
--- a/include/dxc/DXIL/DxilShaderFlags.h
+++ b/include/dxc/DXIL/DxilShaderFlags.h
@ -197,6 +197,14 @@ public:
  void SetWaveMMA(bool flag) { m_bWaveMMA = flag; }
  bool GetWaveMMA() const { return m_bWaveMMA; }

+  void SetSampleCmpGradientOrBias(bool flag) {
+    m_bSampleCmpGradientOrBias = flag;
+  }
+  bool GetSampleCmpGradientOrBias() const { return m_bSampleCmpGradientOrBias; }
+
+  void SetExtendedCommandInfo(bool flag) { m_bExtendedCommandInfo = flag; }
+  bool GetExtendedCommandInfo() const { return m_bExtendedCommandInfo; }
+
 private:
  unsigned
      m_bDisableOptimizations : 1; // D3D11_1_SB_GLOBAL_FLAG_SKIP_OPTIMIZATION
@ -276,8 +284,10 @@ private:

  // SM 6.8+
  unsigned m_bWaveMMA : 1; // SHADER_FEATURE_WAVE_MMA
-
-  uint32_t m_align1 : 27; // align to 64 bit.
+  unsigned
+      m_bSampleCmpGradientOrBias : 1; // SHADER_FEATURE_SAMPLE_CMP_GRADIENT_OR_BIAS
+  unsigned m_bExtendedCommandInfo : 1; // SHADER_FEATURE_EXTENDED_COMMAND_INFO
+  uint32_t m_align1 : 25;              // align to 64 bit.
 };

 } // namespace hlsl
--- a/lib/DXIL/DxilShaderFlags.cpp
+++ b/lib/DXIL/DxilShaderFlags.cpp
@ -45,7 +45,8 @@ ShaderFlags::ShaderFlags()
      m_bSamplerDescriptorHeapIndexing(false),
      m_bAtomicInt64OnHeapResource(false), m_bResMayNotAlias(false),
      m_bAdvancedTextureOps(false), m_bWriteableMSAATextures(false),
-      m_bWaveMMA(false), m_align1(0) {
+      m_bWaveMMA(false), m_bSampleCmpGradientOrBias(false),
+      m_bExtendedCommandInfo(false), m_align1(0) {
  // Silence unused field warnings
  (void)m_align1;
 }
@ -125,6 +126,14 @@ uint64_t ShaderFlags::GetFeatureInfo() const {

  Flags |= m_bWaveMMA ? hlsl::DXIL::ShaderFeatureInfo_WaveMMA : 0;

+  Flags |= m_bSampleCmpGradientOrBias
+               ? hlsl::DXIL::ShaderFeatureInfo_SampleCmpGradientOrBias
+               : 0;
+
+  Flags |= m_bExtendedCommandInfo
+               ? hlsl::DXIL::ShaderFeatureInfo_ExtendedCommandInfo
+               : 0;
+
  return Flags;
 }

@ -185,6 +194,8 @@ uint64_t ShaderFlags::GetShaderFlagsRawForCollection() {
  Flags.SetAdvancedTextureOps(true);
  Flags.SetWriteableMSAATextures(true);
  Flags.SetWaveMMA(true);
+  Flags.SetSampleCmpGradientOrBias(true);
+  Flags.SetExtendedCommandInfo(true);
  return Flags.GetShaderFlagsRaw();
 }

@ -414,8 +425,10 @@ ShaderFlags ShaderFlags::CollectShaderFlags(const Function *F,

  bool hasAdvancedTextureOps = false;
  bool hasWriteableMSAATextures = false;
+  bool hasSampleCmpGradientOrBias = false;

  bool hasWaveMMA = false;
+  bool hasExtendedCommandInfo = false;

  // Try to maintain compatibility with a v1.0 validator if that's what we have.
  uint32_t valMajor, valMinor;
@ -590,6 +603,7 @@ ShaderFlags ShaderFlags::CollectShaderFlags(const Function *F,
        case DXIL::OpCode::SampleCmpGrad:
          hasAdvancedTextureOps |= hasNonConstantSampleOffsets(CI);
          hasLodClamp |= hasSampleClamp(CI);
+          hasSampleCmpGradientOrBias = dxilOp == DXIL::OpCode::SampleCmpGrad;
          break;
        case DXIL::OpCode::Sample:
        case DXIL::OpCode::SampleBias:
@ -597,6 +611,7 @@ ShaderFlags ShaderFlags::CollectShaderFlags(const Function *F,
        case DXIL::OpCode::SampleCmpBias:
          hasAdvancedTextureOps |= hasNonConstantSampleOffsets(CI);
          hasLodClamp |= hasSampleClamp(CI);
+          hasSampleCmpGradientOrBias = dxilOp == DXIL::OpCode::SampleCmpBias;
          LLVM_FALLTHROUGH;
        case DXIL::OpCode::DerivFineX:
        case DXIL::OpCode::DerivFineY:
@ -644,6 +659,10 @@ ShaderFlags ShaderFlags::CollectShaderFlags(const Function *F,
        case DXIL::OpCode::WaveMatrix_SumAccumulate:
          hasWaveMMA = true;
          break;
+        case DXIL::OpCode::StartVertexLocation:
+        case DXIL::OpCode::StartInstanceLocation:
+          hasExtendedCommandInfo = true;
+          break;
        default:
          // Normal opcodes.
          break;
@ -761,10 +780,11 @@ ShaderFlags ShaderFlags::CollectShaderFlags(const Function *F,
  flag.SetAdvancedTextureOps(hasAdvancedTextureOps);
  flag.SetWriteableMSAATextures(hasWriteableMSAATextures);
  flag.SetWaveMMA(hasWaveMMA);
-
  // Only bother setting the flag when there are UAVs.
  flag.SetResMayNotAlias(canSetResMayNotAlias && hasUAVs &&
                         !M->GetResMayAlias());
+  flag.SetSampleCmpGradientOrBias(hasSampleCmpGradientOrBias);
+  flag.SetExtendedCommandInfo(hasExtendedCommandInfo);

  return flag;
 }
--- a/tools/clang/test/HLSLFileCheck/hlsl/objects/Texture/SampleCmpBias.hlsl
+++ b/tools/clang/test/HLSLFileCheck/hlsl/objects/Texture/SampleCmpBias.hlsl
@ -1,5 +1,6 @@
 // RUN: %dxc -Tps_6_8 %s | FileCheck %s

+// CHECK: SampleCmp with gradient or bias
 SamplerComparisonState samp1;
 Texture1D<float4> tex1d;
 Texture1DArray<float4> tex1d_array;
--- a/tools/clang/test/HLSLFileCheck/hlsl/objects/Texture/SampleCmpGrad.hlsl
+++ b/tools/clang/test/HLSLFileCheck/hlsl/objects/Texture/SampleCmpGrad.hlsl
@ -1,5 +1,6 @@
 // RUN: %dxc -Tps_6_8 %s | FileCheck %s

+// CHECK: SampleCmp with gradient or bias
 SamplerComparisonState samp1;

 Texture2D<float4> tex2d;
--- a/tools/clang/test/HLSLFileCheck/hlsl/objects/Texture/has_lod_clmap.hlsl
+++ b/tools/clang/test/HLSLFileCheck/hlsl/objects/Texture/has_lod_clmap.hlsl
@ -1,14 +1,15 @@
-// RUN: %dxilver 1.8 | %dxc -E test_sample -T ps_6_8 %s | FileCheck %s
-// RUN: %dxilver 1.8 | %dxc -E test_sampleb -T ps_6_8 %s | FileCheck %s
-// RUN: %dxilver 1.8 | %dxc -E test_sampleg -T ps_6_8 %s | FileCheck %s
-// RUN: %dxilver 1.8 | %dxc -E test_samplec -T ps_6_8 %s | FileCheck %s
-// RUN: %dxilver 1.8 | %dxc -E test_samplecb -T ps_6_8 %s | FileCheck %s
-// RUN: %dxilver 1.8 | %dxc -E test_samplecg -T ps_6_8 %s | FileCheck %s
+// RUN: %dxilver 1.8 | %dxc -E test_sample -T ps_6_8 %s | FileCheck %s -check-prefixes=NRM,CHECK
+// RUN: %dxilver 1.8 | %dxc -E test_sampleb -T ps_6_8 %s | FileCheck %s -check-prefixes=NRM,CHECK
+// RUN: %dxilver 1.8 | %dxc -E test_sampleg -T ps_6_8 %s | FileCheck %s -check-prefixes=NRM,CHECK
+// RUN: %dxilver 1.8 | %dxc -E test_samplec -T ps_6_8 %s | FileCheck %s -check-prefixes=NRM,CHECK
+// RUN: %dxilver 1.8 | %dxc -E test_samplecb -T ps_6_8 %s | FileCheck %s -check-prefixes=CMPBG,CHECK
+// RUN: %dxilver 1.8 | %dxc -E test_samplecg -T ps_6_8 %s | FileCheck %s -check-prefixes=CMPBG,CHECK

 // LOD clamp requires TiledResources feature
 // From DXC disassembly comment:
 // CHECK: Note: shader requires additional functionality:
 // CHECK-NEXT: Tiled resources
+// CMPBG-NEXT: SampleCmp with gradient or bias

 // CHECK:define void @[[name:[a-z_]+]]()

@ -16,7 +17,10 @@
 // CHECK: ![[entryPoints]] = !{void ()* @[[name]], !"[[name]]", !{{[0-9]+}}, !{{[0-9]+}}, ![[extAttr:[0-9]+]]}

 // tag 0: ShaderFlags, 4096 = Tiled resources
-// CHECK: ![[extAttr]] = !{i32 0, i64 4096}
+// NRM: ![[extAttr]] = !{i32 0, i64 4096}
+
+// tag 0: ShaderFlags, 137438957568 = Tiled resources and SampleCmpGradientOrBias
+// CMPBG: ![[extAttr]] = !{i32 0, i64 137438957568}

 Texture2D T2D;
 SamplerState S;
--- a/tools/clang/test/HLSLFileCheck/hlsl/objects/Texture/no_lod_clmap.hlsl
+++ b/tools/clang/test/HLSLFileCheck/hlsl/objects/Texture/no_lod_clmap.hlsl
@ -1,27 +1,32 @@
-// RUN: %dxilver 1.8 | %dxc -E test_sample -T ps_6_8 %s | FileCheck %s
-// RUN: %dxilver 1.8 | %dxc -E test_sampleb -T ps_6_8 %s | FileCheck %s
-// RUN: %dxilver 1.8 | %dxc -E test_sampleg -T ps_6_8 %s | FileCheck %s
-// RUN: %dxilver 1.8 | %dxc -E test_samplec -T ps_6_8 %s | FileCheck %s
-// RUN: %dxilver 1.8 | %dxc -E test_samplecb -T ps_6_8 %s | FileCheck %s
-// RUN: %dxilver 1.8 | %dxc -E test_samplecg -T ps_6_8 %s | FileCheck %s
-// RUN: %dxilver 1.8 | %dxc -E test_sample_zero -T ps_6_8 %s | FileCheck %s
-// RUN: %dxilver 1.8 | %dxc -E test_sampleb_zero -T ps_6_8 %s | FileCheck %s
-// RUN: %dxilver 1.8 | %dxc -E test_sampleg_zero -T ps_6_8 %s | FileCheck %s
-// RUN: %dxilver 1.8 | %dxc -E test_samplec_zero -T ps_6_8 %s | FileCheck %s
-// RUN: %dxilver 1.8 | %dxc -E test_samplecb_zero -T ps_6_8 %s | FileCheck %s
-// RUN: %dxilver 1.8 | %dxc -E test_samplecg_zero -T ps_6_8 %s | FileCheck %s
+// RUN: %dxilver 1.8 | %dxc -E test_sample -T ps_6_8 %s | FileCheck %s -check-prefixes=NRM,CHECK
+// RUN: %dxilver 1.8 | %dxc -E test_sampleb -T ps_6_8 %s | FileCheck %s -check-prefixes=NRM,CHECK
+// RUN: %dxilver 1.8 | %dxc -E test_sampleg -T ps_6_8 %s | FileCheck %s -check-prefixes=NRM,CHECK
+// RUN: %dxilver 1.8 | %dxc -E test_samplec -T ps_6_8 %s | FileCheck %s -check-prefixes=NRM,CHECK
+// RUN: %dxilver 1.8 | %dxc -E test_samplecb -T ps_6_8 %s | FileCheck %s -check-prefixes=CMPBG,CHECK
+// RUN: %dxilver 1.8 | %dxc -E test_samplecg -T ps_6_8 %s | FileCheck %s -check-prefixes=CMPBG,CHECK
+// RUN: %dxilver 1.8 | %dxc -E test_sample_zero -T ps_6_8 %s | FileCheck %s -check-prefixes=NRM,CHECK
+// RUN: %dxilver 1.8 | %dxc -E test_sampleb_zero -T ps_6_8 %s | FileCheck %s -check-prefixes=NRM,CHECK
+// RUN: %dxilver 1.8 | %dxc -E test_sampleg_zero -T ps_6_8 %s | FileCheck %s -check-prefixes=NRM,CHECK
+// RUN: %dxilver 1.8 | %dxc -E test_samplec_zero -T ps_6_8 %s | FileCheck %s -check-prefixes=NRM,CHECK
+// RUN: %dxilver 1.8 | %dxc -E test_samplecb_zero -T ps_6_8 %s | FileCheck %s -check-prefixes=CMPBG,CHECK
+// RUN: %dxilver 1.8 | %dxc -E test_samplecg_zero -T ps_6_8 %s | FileCheck %s -check-prefixes=CMPBG,CHECK

 // Make sure no tile resources when no lod clamp or clamp is 0.

 // CHECK-NOT: Tiled resources
+// CMPBG: SampleCmp with gradient or bias

 // CHECK:define void @[[name:[a-z_]+]]()

 // CHECK: !dx.entryPoints = !{![[entryPoints:[0-9]+]]}
-// CHECK: ![[entryPoints]] = !{void ()* @[[name]], !"[[name]]", !{{[0-9]+}}, !{{[0-9]+}}, null}
+// NRM: ![[entryPoints]] = !{void ()* @[[name]], !"[[name]]", !{{[0-9]+}}, !{{[0-9]+}}, null}
+// CMPBG: ![[entryPoints]] = !{void ()* @[[name]], !"[[name]]", !{{[0-9]+}}, !{{[0-9]+}}, ![[extAttr:[0-9]+]]}

 // tag 0: ShaderFlags, 4096 = Tiled resources
-// CHECK-NOT:!{i32 0, i64 4096}
+// NRM-NOT:!{i32 0, i64 4096}
+
+// tag 0: ShaderFlags, 137438953472 = SampleCmpGradientOrBias
+// CMPBG: ![[extAttr]] = !{i32 0, i64 137438953472}

 Texture2D T2D;
 SamplerState S;
--- a/tools/clang/test/HLSLFileCheck/hlsl/semantics/ExtendedCommandInformation/no_in_sig_input.hlsl
+++ b/tools/clang/test/HLSLFileCheck/hlsl/semantics/ExtendedCommandInformation/no_in_sig_input.hlsl
@ -1,15 +1,19 @@
 // RUN: %dxc -E main -T vs_6_8 %s | FileCheck %s

+// CHECK: Extended command info
 // CHECK: @main

 // CHECK: call i32 @dx.op.startInstanceLocation.i32(i32 257)
 // CHECK: call i32 @dx.op.startVertexLocation.i32(i32 256)

 // Make sure no input element is generated for the entry point.
-// CHECK: !{void ()* @main, !"main", ![[SIG:[0-9]+]], null, null}
+// CHECK: !{void ()* @main, !"main", ![[SIG:[0-9]+]], null, ![[extAttr:[0-9]+]]}
 // The input should be null
 // CHECK: ![[SIG]] = !{null,

+// tag 0: ShaderFlags, 274877906944 = SampleCmpGradientOrBias
+// CHECK: ![[extAttr]] = !{i32 0, i64 274877906944}
+
 float4 main(int loc : SV_StartVertexLocation
           , uint loc2 : SV_StartInstanceLocation
           ) : SV_Position
--- a/tools/clang/tools/dxcompiler/dxcdisassembler.cpp
+++ b/tools/clang/tools/dxcompiler/dxcdisassembler.cpp
@ -347,6 +347,8 @@ PCSTR g_pFeatureInfoNames[] = {
    "64-bit Atomics on Heap Resources",
    "Advanced Texture Ops",
    "Writeable MSAA Textures",
+    "SampleCmp with gradient or bias",
+    "Extended command info",
 };
 static_assert(_countof(g_pFeatureInfoNames) == ShaderFeatureInfoCount,
              "g_pFeatureInfoNames needs to be updated");