Add O1experimental SPIR-V fast compile mode

Author: AnastaZIuk

docs/SPIR-V.rst

@@ -518,12 +518,22 @@ generate valid SPIR-V for Vulkan.
 Optimization
 ~~~~~~~~~~~~
 
-Optimization is also delegated to SPIRV-Tools. Right now there are no difference
-between optimization levels greater than zero; they will all invoke the same
-optimization recipe. That is, the recipe behind ``spirv-opt -O``.  If you want to
-run a custom optimization recipe, you can do so using the command line option
-``-Oconfig=`` and specifying a comma-separated list of your desired passes.
-The passes are invoked in the specified order.
+Optimization is also delegated to SPIRV-Tools. There are two built-in
+optimization recipes for SPIR-V code generation:
+
+* ``-O1experimental``: a development-oriented recipe that favors lower
+  optimizer wall time. This profile is opt-in, leaves the default ``-O0``,
+  ``-O1``, ``-O2``, ``-O3`` and ``-Oconfig`` behavior unchanged, and may emit
+  ``VariablePointers`` for modules that need the fast path. On Vulkan targets,
+  use at least ``-fspv-target-env=vulkan1.1`` and ensure the target device
+  supports the required variable-pointer features. Vulkan 1.4 guarantees that
+  support.
+* ``-O1``, ``-O2``, and ``-O3``: the performance-oriented recipe behind
+  ``spirv-opt -O``.
+
+If you want to run a custom optimization recipe, you can do so using the
+command line option ``-Oconfig=`` and specifying a comma-separated list of
+your desired passes. The passes are invoked in the specified order.
 
 For example, you can specify ``-Oconfig=--loop-unroll,--scalar-replacement=300,--eliminate-dead-code-aggressive``
 to firstly invoke loop unrolling, then invoke scalar replacement of aggregates,

external/SPIRV-Tools

@@ -96,6 +96,9 @@ def O0 : Flag<["-", "/"], "O0">, Group<hlsloptz_Group>, Flags<[CoreOption]>,
     HelpText<"Optimization Level 0">;
 def O1 : Flag<["-", "/"], "O1">, Group<hlsloptz_Group>, Flags<[CoreOption]>,
     HelpText<"Optimization Level 1">;
+def O1experimental : Flag<["-"], "O1experimental">, Group<spirv_Group>,
+    Flags<[CoreOption, DriverOption]>,
+    HelpText<"Enable the experimental SPIR-V fast-compile profile; may emit VariablePointers and requires target support for them when used">;
 def O2 : Flag<["-", "/"], "O2">, Group<hlsloptz_Group>, Flags<[CoreOption]>,
     HelpText<"Optimization Level 2">;
 def O3 : Flag<["-", "/"], "O3">, Group<hlsloptz_Group>, Flags<[CoreOption]>,

include/dxc/Support/HLSLOptions.td

@@ -60,6 +60,7 @@ struct SpirvCodeGenOptions {
   bool noWarnIgnoredFeatures = false;
   bool preserveBindings = false;
   bool preserveInterface = false;
+  bool o1ExperimentalFastCompile = false;
   bool useDxLayout = false;
   bool useGlLayout = false;
   bool useLegacyBufferMatrixOrder = false;

include/dxc/Support/SPIRVOptions.h

@@ -792,11 +792,15 @@ int ReadDxcOpts(const OptTable *optionTable, unsigned flagsToInclude,
   }
 
   opts.DisableOptimizations = false;
-  if (Arg *A = Args.getLastArg(OPT_O0, OPT_O1, OPT_O2, OPT_O3, OPT_Od)) {
+  if (Arg *A =
+          Args.getLastArg(OPT_O0, OPT_O1, OPT_O1experimental, OPT_O2, OPT_O3,
+                          OPT_Od)) {
     if (A->getOption().matches(OPT_O0))
       opts.OptLevel = 0;
     if (A->getOption().matches(OPT_O1))
       opts.OptLevel = 1;
+    if (A->getOption().matches(OPT_O1experimental))
+      opts.OptLevel = 1;
     if (A->getOption().matches(OPT_O2))
       opts.OptLevel = 2;
     if (A->getOption().matches(OPT_O3))
@@ -1098,6 +1102,8 @@ int ReadDxcOpts(const OptTable *optionTable, unsigned flagsToInclude,
   opts.SpirvOptions.disableHLSLIntrinsics =
       Args.hasFlag(OPT_fvk_disable_hlsl_intrinsics, OPT_INVALID, false);
   opts.GenSPIRV = Args.hasFlag(OPT_spirv, OPT_INVALID, false);
+  opts.SpirvOptions.o1ExperimentalFastCompile =
+      Args.hasFlag(OPT_O1experimental, OPT_INVALID, false);
   opts.SpirvOptions.invertY =
       Args.hasFlag(OPT_fvk_invert_y, OPT_INVALID, false);
   opts.SpirvOptions.invertW =
@@ -1263,7 +1269,8 @@ int ReadDxcOpts(const OptTable *optionTable, unsigned flagsToInclude,
       errors << "-Oconfig should not be specified more than once";
       return 1;
     }
-    if (Args.getLastArg(OPT_O0, OPT_O1, OPT_O2, OPT_O3)) {
+    if (Args.getLastArg(OPT_O0, OPT_O1, OPT_O1experimental, OPT_O2,
+                        OPT_O3)) {
       errors << "-Oconfig should not be used together with -O";
       return 1;
     }
@@ -1275,6 +1282,11 @@ int ReadDxcOpts(const OptTable *optionTable, unsigned flagsToInclude,
   opts.SpirvOptions.entrypointName =
       Args.getLastArgValue(OPT_fspv_entrypoint_name_EQ);
 
+  if (opts.SpirvOptions.o1ExperimentalFastCompile && !opts.GenSPIRV) {
+    errors << "-O1experimental requires -spirv";
+    return 1;
+  }
+
   // Check for use of options not implemented in the SPIR-V backend.
   if (Args.hasFlag(OPT_spirv, OPT_INVALID, false) &&
       hasUnsupportedSpirvOption(Args, errors))
@@ -1297,6 +1309,7 @@ int ReadDxcOpts(const OptTable *optionTable, unsigned flagsToInclude,
       Args.hasFlag(OPT_fspv_reflect, OPT_INVALID, false) ||
       Args.hasFlag(OPT_fspv_fix_func_call_arguments, OPT_INVALID, false) ||
       Args.hasFlag(OPT_fspv_print_all, OPT_INVALID, false) ||
+      Args.hasFlag(OPT_O1experimental, OPT_INVALID, false) ||
       Args.hasFlag(OPT_Wno_vk_ignored_features, OPT_INVALID, false) ||
       Args.hasFlag(OPT_Wno_vk_emulated_features, OPT_INVALID, false) ||
       Args.hasFlag(OPT_fvk_auto_shift_bindings, OPT_INVALID, false) ||

lib/DxcSupport/HLSLOptions.cpp

@@ -595,6 +595,7 @@ SpirvEmitter::SpirvEmitter(CompilerInstance &ci)
       isSpecConstantMode(false), needsLegalization(false),
       needsLegalizationLoopUnroll(false),
       needsLegalizationSsaRewrite(false),
+      sawExplicitUnrollHint(false),
       beforeHlslLegalization(false), mainSourceFile(nullptr) {
 
   // Get ShaderModel from command line hlsl profile option.
@@ -920,6 +921,20 @@ void SpirvEmitter::HandleTranslationUnit(ASTContext &context) {
     }
   }
 
+  if (useSpirvFastCompileProfile() &&
+      (needsLegalizationLoopUnroll || sawExplicitUnrollHint)) {
+    if (featureManager.isTargetEnvVulkan() &&
+        !featureManager.isTargetEnvVulkan1p1OrAbove()) {
+      emitFatalError(
+          "-O1experimental requires -fspv-target-env=vulkan1.1 or above "
+          "when the generated module needs VariablePointers",
+          {});
+      return;
+    }
+
+    spvBuilder.requireCapability(spv::Capability::VariablePointers);
+  }
+
   // Output the constructed module.
   std::vector<uint32_t> m = spvBuilder.takeModule();
   if (context.getDiagnostics().hasErrorOccurred())
@@ -2289,6 +2304,7 @@ spv::LoopControlMask SpirvEmitter::translateLoopAttribute(const Stmt *stmt,
   case attr::HLSLFastOpt:
     return spv::LoopControlMask::DontUnroll;
   case attr::HLSLUnroll:
+    sawExplicitUnrollHint = true;
     return spv::LoopControlMask::Unroll;
   case attr::HLSLAllowUAVCondition:
     if (!spirvOptions.noWarnIgnoredFeatures) {
@@ -16651,8 +16667,12 @@ bool SpirvEmitter::spirvToolsOptimize(std::vector<uint32_t> *mod,
   options.set_max_id_bound(spirvOptions.maxId);
 
   if (spirvOptions.optConfig.empty()) {
-    // Add performance passes.
-    optimizer.RegisterPerformancePasses(spirvOptions.preserveInterface);
+    if (useSpirvFastCompileProfile()) {
+      optimizer.RegisterPerformancePassesFastCompile(
+          spirvOptions.preserveInterface);
+    } else {
+      optimizer.RegisterPerformancePasses(spirvOptions.preserveInterface);
+    }
 
     // Add propagation of volatile semantics passes.
     optimizer.RegisterPass(spvtools::CreateSpreadVolatileSemanticsPass());
@@ -16672,6 +16692,11 @@ bool SpirvEmitter::spirvToolsOptimize(std::vector<uint32_t> *mod,
   return optimizer.Run(mod->data(), mod->size(), mod, options);
 }
 
+bool SpirvEmitter::useSpirvFastCompileProfile() const {
+  return spirvOptions.o1ExperimentalFastCompile &&
+         spirvOptions.optConfig.empty();
+}
+
 bool SpirvEmitter::spirvToolsLegalize(std::vector<uint32_t> *mod,
                                       std::string *messages,
                                       const std::vector<DescriptorSetAndBinding>
@@ -16697,15 +16722,19 @@ bool SpirvEmitter::spirvToolsLegalize(std::vector<uint32_t> *mod,
     optimizer.RegisterPass(
         spvtools::CreateInterfaceVariableScalarReplacementPass());
   }
-  auto legalizationSsaRewriteMode = spvtools::SSARewriteMode::None;
-  if (needsLegalizationLoopUnroll) {
-    legalizationSsaRewriteMode = spvtools::SSARewriteMode::All;
-  } else if (needsLegalizationSsaRewrite) {
-    legalizationSsaRewriteMode = spvtools::SSARewriteMode::OpaqueOnly;
+  if (useSpirvFastCompileProfile()) {
+    auto legalizationSsaRewriteMode = spvtools::SSARewriteMode::None;
+    if (needsLegalizationLoopUnroll) {
+      legalizationSsaRewriteMode = spvtools::SSARewriteMode::All;
+    } else if (needsLegalizationSsaRewrite) {
+      legalizationSsaRewriteMode = spvtools::SSARewriteMode::OpaqueOnly;
+    }
+    optimizer.RegisterLegalizationPasses(
+        spirvOptions.preserveInterface, needsLegalizationLoopUnroll,
+        legalizationSsaRewriteMode);
+  } else {
+    optimizer.RegisterLegalizationPasses(spirvOptions.preserveInterface);
   }
-  optimizer.RegisterLegalizationPasses(
-      spirvOptions.preserveInterface, needsLegalizationLoopUnroll,
-      legalizationSsaRewriteMode);
   // Add flattening of resources if needed.
   if (spirvOptions.flattenResourceArrays) {
     optimizer.RegisterPass(

tools/clang/lib/SPIRV/SpirvEmitter.cpp

@@ -1339,6 +1339,7 @@ class SpirvEmitter : public ASTConsumer {
   /// gets the info/warning/error messages via |messages|.
   /// Returns true on success and false otherwise.
   bool spirvToolsOptimize(std::vector<uint32_t> *mod, std::string *messages);
+  bool useSpirvFastCompileProfile() const;
 
   // \brief Runs the pass represented by the given pass token on the module.
   // Returns true if the pass was successfully run. Any messages from the
@@ -1580,6 +1581,7 @@ class SpirvEmitter : public ASTConsumer {
   bool needsLegalization;
   bool needsLegalizationLoopUnroll;
   bool needsLegalizationSsaRewrite;
+  bool sawExplicitUnrollHint;
 
   /// Whether the translated SPIR-V binary passes --before-hlsl-legalization
   /// option to spirv-val because of illegal function parameter scope.

tools/clang/lib/SPIRV/SpirvEmitter.h

@@ -1,12 +1,17 @@
-// RUN: %dxc -T cs_6_0 -E main -O3 -Vd %s -spirv | FileCheck %s
-
-// CHECK: OpLoopMerge {{%[0-9]+}} {{%[0-9]+}} Unroll
-// CHECK: [[which:%[0-9]+]] = OpSelect %_ptr_Uniform_type_StructuredBuffer_S {{%[0-9]+}} %gSBuffer1 %gSBuffer2
-// CHECK: [[idx:%[0-9]+]] = OpBitcast %uint {{%[0-9]+}}
-// CHECK: [[src:%[0-9]+]] = OpAccessChain %_ptr_Uniform_S [[which]] %int_0 [[idx]]
-// CHECK: [[val:%[0-9]+]] = OpLoad %S [[src]]
-// CHECK: [[dst:%[0-9]+]] = OpAccessChain %_ptr_Uniform_S %gRWSBuffer %int_0 [[idx]]
-// CHECK: OpStore [[dst]] [[val]]
+// RUN: %dxc -T cs_6_0 -E main -O3 -Vd -fspv-target-env=vulkan1.3 %s -spirv | FileCheck %s --check-prefix=CHECK --check-prefix=CHECK-O3
+// RUN: %dxc -T cs_6_0 -E main -O1experimental -Vd -fspv-target-env=vulkan1.3 %s -spirv | FileCheck %s --check-prefix=CHECK --check-prefix=CHECK-O1EXP
+
+// CHECK-O1EXP: OpCapability VariablePointers
+
+// CHECK:      [[ptr:%[0-9]+]] = OpAccessChain %_ptr_Uniform_S %gSBuffer2
+// CHECK-NEXT: [[val:%[0-9]+]] = OpLoad %S [[ptr]]
+// CHECK-NEXT: [[ptr:%[0-9]+]] = OpAccessChain %_ptr_Uniform_S %gRWSBuffer
+// CHECK-NEXT:                OpStore [[ptr]] [[val]]
+
+// CHECK:      [[ptr:%[0-9]+]] = OpAccessChain %_ptr_Uniform_S %gSBuffer2
+// CHECK-NEXT: [[val:%[0-9]+]] = OpLoad %S [[ptr]]
+// CHECK-NEXT: [[ptr:%[0-9]+]] = OpAccessChain %_ptr_Uniform_S %gRWSBuffer
+// CHECK-NEXT:                OpStore [[ptr]] [[val]]
 
 struct S {
   float4 f;