SPV_KHR_quad_control (#5547)

* SPV_KHR_quad_control 1. Add two new execute modes: RequireFullQuadsKHR and QuadDerivativesKHR 2. Add two opCodes: OpGroupNonUniformQuadAllKHR and OpGroupNonUniformQuadAnyKHR 3. Add one Capability: QuadControlKHR * update DEPS * Fixes * Build fixes * Formatting fixes * Test fixes * formatting --------- Co-authored-by: Alan Baker <alanbaker@google.com>
2024-12-25 17:21:06 +00:00 · 2024-01-27 04:49:56 +08:00 · 2024-01-27 04:49:56 +08:00 · b951948eaa
commit b951948eaa
parent 69197ba90b
8 changed files with 232 additions and 8 deletions
--- a/2
+++ b/2
@ -13,7 +13,7 @@ vars = {
  'protobuf_revision': 'v21.12',

  're2_revision': '264e71e88e1c8a4b5ec326e70e9cf1d476f58a58',
-  'spirv_headers_revision': 'ae6a8b39717523d96683bc0d20b541944e28072f',
+  'spirv_headers_revision': '5aa1dd8a11182ea9a6a0eabd6a9edc639d5dbecd',
 }

 deps = {
--- a/source/opcode.cpp
+++ b/source/opcode.cpp
@ -534,6 +534,8 @@ bool spvOpcodeIsNonUniformGroupOperation(spv::Op opcode) {
    case spv::Op::OpGroupNonUniformQuadBroadcast:
    case spv::Op::OpGroupNonUniformQuadSwap:
    case spv::Op::OpGroupNonUniformRotateKHR:
+    case spv::Op::OpGroupNonUniformQuadAllKHR:
+    case spv::Op::OpGroupNonUniformQuadAnyKHR:
      return true;
    default:
      return false;
--- a/source/val/validate_mode_setting.cpp
+++ b/source/val/validate_mode_setting.cpp
@ -557,6 +557,17 @@ spv_result_t ValidateExecutionMode(ValidationState_t& _,
                  "model.";
      }
      break;
+    case spv::ExecutionMode::QuadDerivativesKHR:
+      if (!std::all_of(models->begin(), models->end(),
+                       [](const spv::ExecutionModel& model) {
+                         return (model == spv::ExecutionModel::Fragment ||
+                                 model == spv::ExecutionModel::GLCompute);
+                       })) {
+        return _.diag(SPV_ERROR_INVALID_DATA, inst)
+               << "Execution mode can only be used with the Fragment or "
+                  "GLCompute execution model.";
+      }
+      break;
    case spv::ExecutionMode::PixelCenterInteger:
    case spv::ExecutionMode::OriginUpperLeft:
    case spv::ExecutionMode::OriginLowerLeft:
@ -581,6 +592,7 @@ spv_result_t ValidateExecutionMode(ValidationState_t& _,
    case spv::ExecutionMode::StencilRefUnchangedBackAMD:
    case spv::ExecutionMode::StencilRefGreaterBackAMD:
    case spv::ExecutionMode::StencilRefLessBackAMD:
+    case spv::ExecutionMode::RequireFullQuadsKHR:
      if (!std::all_of(models->begin(), models->end(),
                       [](const spv::ExecutionModel& model) {
                         return model == spv::ExecutionModel::Fragment;
--- a/source/val/validate_non_uniform.cpp
+++ b/source/val/validate_non_uniform.cpp
@ -422,9 +422,14 @@ spv_result_t NonUniformPass(ValidationState_t& _, const Instruction* inst) {
  const spv::Op opcode = inst->opcode();

  if (spvOpcodeIsNonUniformGroupOperation(opcode)) {
-    const uint32_t execution_scope = inst->GetOperandAs<uint32_t>(2);
-    if (auto error = ValidateExecutionScope(_, inst, execution_scope)) {
-      return error;
+    // OpGroupNonUniformQuadAllKHR and OpGroupNonUniformQuadAnyKHR don't have
+    // scope paramter
+    if ((opcode != spv::Op::OpGroupNonUniformQuadAllKHR) &&
+        (opcode != spv::Op::OpGroupNonUniformQuadAnyKHR)) {
+      const uint32_t execution_scope = inst->GetOperandAs<uint32_t>(2);
+      if (auto error = ValidateExecutionScope(_, inst, execution_scope)) {
+        return error;
+      }
    }
  }

--- a/source/val/validate_scopes.cpp
+++ b/source/val/validate_scopes.cpp
@ -97,8 +97,10 @@ spv_result_t ValidateExecutionScope(ValidationState_t& _,
    // Vulkan 1.1 specific rules
    if (_.context()->target_env != SPV_ENV_VULKAN_1_0) {
      // Scope for Non Uniform Group Operations must be limited to Subgroup
-      if (spvOpcodeIsNonUniformGroupOperation(opcode) &&
-          value != spv::Scope::Subgroup) {
+      if ((spvOpcodeIsNonUniformGroupOperation(opcode) &&
+           (opcode != spv::Op::OpGroupNonUniformQuadAllKHR) &&
+           (opcode != spv::Op::OpGroupNonUniformQuadAnyKHR)) &&
+          (value != spv::Scope::Subgroup)) {
        return _.diag(SPV_ERROR_INVALID_DATA, inst)
               << _.VkErrorID(4642) << spvOpcodeString(opcode)
               << ": in Vulkan environment Execution scope is limited to "
@ -178,6 +180,8 @@ spv_result_t ValidateExecutionScope(ValidationState_t& _,
  // Scope for execution must be limited to Workgroup or Subgroup for
  // non-uniform operations
  if (spvOpcodeIsNonUniformGroupOperation(opcode) &&
+      opcode != spv::Op::OpGroupNonUniformQuadAllKHR &&
+      opcode != spv::Op::OpGroupNonUniformQuadAnyKHR &&
      value != spv::Scope::Subgroup && value != spv::Scope::Workgroup) {
    return _.diag(SPV_ERROR_INVALID_DATA, inst)
           << spvOpcodeString(opcode)
--- a/test/val/val_modes_test.cpp
+++ b/test/val/val_modes_test.cpp
@ -1874,6 +1874,184 @@ OpFunctionEnd
                        "with the FPFastMathDefault execution mode"));
 }

+TEST_F(ValidateMode, FragmentShaderRequireFullQuadsKHR) {
+  const std::string spirv = R"(
+OpCapability Shader
+OpCapability GroupNonUniform
+OpCapability GroupNonUniformVote
+OpCapability GroupNonUniformBallot
+OpCapability QuadControlKHR
+OpExtension "SPV_KHR_quad_control"
+%1 = OpExtInstImport "GLSL.std.450"
+OpMemoryModel Logical GLSL450
+OpEntryPoint GLCompute %4 "main"
+OpExecutionMode %4 OriginUpperLeft
+OpExecutionMode %4 RequireFullQuadsKHR
+OpDecorate %17 Location 0
+OpDecorate %31 BuiltIn HelperInvocation
+OpDecorate %40 Location 0
+OpDecorate %44 DescriptorSet 0
+OpDecorate %44 Binding 0
+%2 = OpTypeVoid
+%3 = OpTypeFunction %2
+%6 = OpTypeInt 32 0
+%7 = OpTypeVector %6 4
+%8 = OpTypePointer Function %7
+%10 = OpTypeBool
+%11 = OpConstantTrue %10
+%12 = OpConstant %6 7
+%14 = OpTypeFloat 32
+%15 = OpTypeVector %14 4
+%16 = OpTypePointer Output %15
+%17 = OpVariable %16 Output
+%18 = OpConstant %14 1
+%19 = OpConstant %14 0
+%20 = OpConstantComposite %15 %18 %19 %19 %18
+%23 = OpConstant %6 4
+%27 = OpConstant %6 1
+%28 = OpTypePointer Output %14
+%30 = OpTypePointer Input %10
+%31 = OpVariable %30 Input
+%36 = OpConstant %6 2
+%38 = OpTypeVector %14 2
+%39 = OpTypePointer Input %38
+%40 = OpVariable %39 Input
+%41 = OpTypeImage %14 2D 0 0 0 1 Unknown
+%42 = OpTypeSampledImage %41
+%43 = OpTypePointer UniformConstant %42
+%44 = OpVariable %43 UniformConstant
+%4 = OpFunction %2 None %3
+%5 = OpLabel
+%9 = OpVariable %8 Function
+%13 = OpGroupNonUniformBallot %7 %12 %11
+OpStore %9 %13
+OpStore %17 %20
+%21 = OpLoad %7 %9
+%22 = OpGroupNonUniformBallotBitCount %6 %12 Reduce %21
+%24 = OpIEqual %10 %22 %23
+OpSelectionMerge %26 None
+OpBranchConditional %24 %25 %26
+%25 = OpLabel
+%29 = OpAccessChain %28 %17 %27
+OpStore %29 %18
+OpBranch %26
+%26 = OpLabel
+%32 = OpLoad %10 %31
+%33 = OpGroupNonUniformAny %10 %12 %32
+OpSelectionMerge %35 None
+OpBranchConditional %33 %34 %35
+%34 = OpLabel
+%37 = OpAccessChain %28 %17 %36
+OpStore %37 %18
+OpBranch %35
+%35 = OpLabel
+OpReturn
+OpFunctionEnd
+)";
+
+  CompileSuccessfully(spirv, SPV_ENV_UNIVERSAL_1_3);
+  EXPECT_THAT(SPV_ERROR_INVALID_DATA,
+              ValidateInstructions(SPV_ENV_UNIVERSAL_1_3));
+  EXPECT_THAT(
+      getDiagnosticString(),
+      HasSubstr(
+          "Execution mode can only be used with the Fragment execution model"));
+}
+
+TEST_F(ValidateMode, FragmentShaderQuadDerivativesKHR) {
+  const std::string spirv = R"(
+OpCapability Shader
+OpCapability GroupNonUniform
+OpCapability GroupNonUniformVote
+OpCapability QuadControlKHR
+OpExtension "SPV_KHR_quad_control"
+%1 = OpExtInstImport "GLSL.std.450"
+OpMemoryModel Logical GLSL450
+OpEntryPoint GLCompute %4 "main"
+OpExecutionMode %4 OriginUpperLeft
+OpExecutionMode %4 QuadDerivativesKHR
+OpDecorate %12 BuiltIn FragCoord
+OpDecorate %41 Location 0
+OpDecorate %45 DescriptorSet 0
+OpDecorate %45 Binding 0
+OpDecorate %49 Location 0
+%2 = OpTypeVoid
+%3 = OpTypeFunction %2
+%6 = OpTypeBool
+%7 = OpTypePointer Function %6
+%9 = OpTypeFloat 32
+%10 = OpTypeVector %9 4
+%11 = OpTypePointer Input %10
+%12 = OpVariable %11 Input
+%13 = OpTypeInt 32 0
+%14 = OpConstant %13 1
+%15 = OpTypePointer Input %9
+%18 = OpConstant %9 8.5
+%21 = OpConstant %9 0.100000001
+%25 = OpConstant %13 0
+%28 = OpConstant %9 3.5
+%30 = OpConstant %9 6
+%36 = OpConstant %13 7
+%40 = OpTypePointer Output %10
+%41 = OpVariable %40 Output
+%42 = OpTypeImage %9 2D 0 0 0 1 Unknown
+%43 = OpTypeSampledImage %42
+%44 = OpTypePointer UniformConstant %43
+%45 = OpVariable %44 UniformConstant
+%47 = OpTypeVector %9 2
+%48 = OpTypePointer Input %47
+%49 = OpVariable %48 Input
+%53 = OpConstant %9 0.899999976
+%54 = OpConstant %9 0.200000003
+%55 = OpConstant %9 1
+%56 = OpConstantComposite %10 %53 %54 %54 %55
+%4 = OpFunction %2 None %3
+%5 = OpLabel
+%8 = OpVariable %7 Function
+%16 = OpAccessChain %15 %12 %14
+%17 = OpLoad %9 %16
+%19 = OpFSub %9 %17 %18
+%20 = OpExtInst %9 %1 FAbs %19
+%22 = OpFOrdLessThan %6 %20 %21
+OpSelectionMerge %24 None
+OpBranchConditional %22 %23 %24
+%23 = OpLabel
+%26 = OpAccessChain %15 %12 %25
+%27 = OpLoad %9 %26
+%29 = OpFSub %9 %27 %28
+%31 = OpFMod %9 %29 %30
+%33 = OpFOrdLessThan %6 %31 %21
+OpBranch %24
+%24 = OpLabel
+%34 = OpPhi %6 %22 %5 %33 %23
+OpStore %8 %34
+%35 = OpLoad %6 %8
+%37 = OpGroupNonUniformAny %6 %36 %35
+OpSelectionMerge %39 None
+OpBranchConditional %37 %38 %52
+%38 = OpLabel
+%46 = OpLoad %43 %45
+%50 = OpLoad %47 %49
+%51 = OpImageSampleImplicitLod %10 %46 %50
+OpStore %41 %51
+OpBranch %39
+%52 = OpLabel
+OpStore %41 %56
+OpBranch %39
+%39 = OpLabel
+OpReturn
+OpFunctionEnd
+)";
+
+  CompileSuccessfully(spirv, SPV_ENV_UNIVERSAL_1_3);
+  EXPECT_THAT(SPV_ERROR_INVALID_DATA,
+              ValidateInstructions(SPV_ENV_UNIVERSAL_1_3));
+  EXPECT_THAT(
+      getDiagnosticString(),
+      HasSubstr(
+          "Execution mode can only be used with the Fragment execution model"));
+}
+
 }  // namespace
 }  // namespace val
 }  // namespace spvtools
--- a/test/val/val_non_uniform_test.cpp
+++ b/test/val/val_non_uniform_test.cpp
@ -45,7 +45,9 @@ OpCapability GroupNonUniformArithmetic
 OpCapability GroupNonUniformClustered
 OpCapability GroupNonUniformQuad
 OpCapability GroupNonUniformPartitionedNV
+OpCapability QuadControlKHR
 OpExtension "SPV_NV_shader_subgroup_partitioned"
+OpExtension "SPV_KHR_quad_control"
 )";

  ss << capabilities_and_extensions;
@ -178,7 +180,10 @@ TEST_P(GroupNonUniform, Vulkan1p1) {
  std::ostringstream sstr;
  sstr << "%result = " << opcode << " ";
  sstr << type << " ";
-  sstr << ConvertScope(execution_scope) << " ";
+  if (opcode != "OpGroupNonUniformQuadAllKHR" &&
+      opcode != "OpGroupNonUniformQuadAnyKHR") {
+    sstr << ConvertScope(execution_scope) << " ";
+  }
  sstr << args << "\n";

  CompileSuccessfully(GenerateShaderCode(sstr.str()), SPV_ENV_VULKAN_1_1);
@ -218,7 +223,10 @@ TEST_P(GroupNonUniform, Spirv1p3) {
  std::ostringstream sstr;
  sstr << "%result = " << opcode << " ";
  sstr << type << " ";
-  sstr << ConvertScope(execution_scope) << " ";
+  if (opcode != "OpGroupNonUniformQuadAllKHR" &&
+      opcode != "OpGroupNonUniformQuadAnyKHR") {
+    sstr << ConvertScope(execution_scope) << " ";
+  }
  sstr << args << "\n";

  CompileSuccessfully(GenerateShaderCode(sstr.str()), SPV_ENV_UNIVERSAL_1_3);
@ -869,6 +877,18 @@ INSTANTIATE_TEST_SUITE_P(
            Values("ClusteredReduce match_res %u32_undef"),
            Values("ClusterSize must be a constant instruction")));

+// Subgroup scope is not actual parameter, but used for test expectations,
+INSTANTIATE_TEST_SUITE_P(GroupNonUniformQuadAllKHR, GroupNonUniform,
+                         Combine(Values("OpGroupNonUniformQuadAllKHR"),
+                                 Values("%bool"), Values(spv::Scope::Subgroup),
+                                 Values("%true"), Values("")));
+
+// Subgroup scope is not actual parameter, but used for test expectations,
+INSTANTIATE_TEST_SUITE_P(GroupNonUniformQuadAnyKHR, GroupNonUniform,
+                         Combine(Values("OpGroupNonUniformQuadAnyKHR"),
+                                 Values("%bool"), Values(spv::Scope::Subgroup),
+                                 Values("%true"), Values("")));
+
 TEST_F(ValidateGroupNonUniform, VulkanGroupNonUniformBallotBitCountOperation) {
  std::string test = R"(
 OpCapability Shader
--- a/tools/sva/src/spirv.data.js
+++ b/tools/sva/src/spirv.data.js
@ -4376,6 +4376,9 @@ export default {
        "ShaderClockKHR": {
          "value": 5055
        },
+        "QuadControlKHR": {
+          "value": 5087
+        },
        "FragmentFullyCoveredEXT": {
          "value": 5265
        },