[mlir][spirv] Add convergent attribute to builtin #122131
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Add the `convergent` attribute to builtin functions and builtin function calls when lowering SPIR-V non-uniform group functions to LLVM dialect. Signed-off-by: Lukas Sommer <[email protected]>
|
@llvm/pr-subscribers-mlir @llvm/pr-subscribers-mlir-spirv Author: Lukas Sommer (sommerlukas) ChangesAdd the Patch is 21.49 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/122131.diff 2 Files Affected:
diff --git a/mlir/lib/Conversion/SPIRVToLLVM/SPIRVToLLVM.cpp b/mlir/lib/Conversion/SPIRVToLLVM/SPIRVToLLVM.cpp
index b11511f21d03d4..e79005f208c3d9 100644
--- a/mlir/lib/Conversion/SPIRVToLLVM/SPIRVToLLVM.cpp
+++ b/mlir/lib/Conversion/SPIRVToLLVM/SPIRVToLLVM.cpp
@@ -1028,8 +1028,7 @@ class ReturnValuePattern : public SPIRVToLLVMConversion<spirv::ReturnValueOp> {
static LLVM::LLVMFuncOp lookupOrCreateSPIRVFn(Operation *symbolTable,
StringRef name,
ArrayRef<Type> paramTypes,
- Type resultType,
- bool convergent = true) {
+ Type resultType) {
auto func = dyn_cast_or_null<LLVM::LLVMFuncOp>(
SymbolTable::lookupSymbolIn(symbolTable, name));
if (func)
@@ -1040,7 +1039,7 @@ static LLVM::LLVMFuncOp lookupOrCreateSPIRVFn(Operation *symbolTable,
symbolTable->getLoc(), name,
LLVM::LLVMFunctionType::get(resultType, paramTypes));
func.setCConv(LLVM::cconv::CConv::SPIR_FUNC);
- func.setConvergent(convergent);
+ func.setConvergent(true);
func.setNoUnwind(true);
func.setWillReturn(true);
return func;
@@ -1253,8 +1252,8 @@ class GroupReducePattern : public SPIRVToLLVMConversion<ReduceOp> {
Operation *symbolTable =
op->template getParentWithTrait<OpTrait::SymbolTable>();
- LLVM::LLVMFuncOp func = lookupOrCreateSPIRVFn(
- symbolTable, funcName, paramTypes, retTy, !NonUniform);
+ LLVM::LLVMFuncOp func =
+ lookupOrCreateSPIRVFn(symbolTable, funcName, paramTypes, retTy);
Location loc = op.getLoc();
Value scope = rewriter.create<LLVM::ConstantOp>(
diff --git a/mlir/test/Conversion/SPIRVToLLVM/non-uniform-ops-to-llvm.mlir b/mlir/test/Conversion/SPIRVToLLVM/non-uniform-ops-to-llvm.mlir
index e81048792c45de..ab174ba2b41e48 100644
--- a/mlir/test/Conversion/SPIRVToLLVM/non-uniform-ops-to-llvm.mlir
+++ b/mlir/test/Conversion/SPIRVToLLVM/non-uniform-ops-to-llvm.mlir
@@ -2,30 +2,30 @@
// NOTE: Assertions have been autogenerated by utils/generate-test-checks.py
-// CHECK-LABEL: llvm.func spir_funccc @_Z33__spirv_GroupNonUniformLogicalXoriib(i32, i32, i1) -> i1 attributes {no_unwind, will_return}
-// CHECK: llvm.func spir_funccc @_Z32__spirv_GroupNonUniformLogicalOriib(i32, i32, i1) -> i1 attributes {no_unwind, will_return}
-// CHECK: llvm.func spir_funccc @_Z33__spirv_GroupNonUniformLogicalAndiib(i32, i32, i1) -> i1 attributes {no_unwind, will_return}
-// CHECK: llvm.func spir_funccc @_Z33__spirv_GroupNonUniformBitwiseXoriij(i32, i32, i32) -> i32 attributes {no_unwind, will_return}
-// CHECK: llvm.func spir_funccc @_Z32__spirv_GroupNonUniformBitwiseOriij(i32, i32, i32) -> i32 attributes {no_unwind, will_return}
-// CHECK: llvm.func spir_funccc @_Z33__spirv_GroupNonUniformBitwiseAndiij(i32, i32, i32) -> i32 attributes {no_unwind, will_return}
-// CHECK: llvm.func spir_funccc @_Z27__spirv_GroupNonUniformSMaxiijj(i32, i32, i32, i32) -> i32 attributes {no_unwind, will_return}
-// CHECK: llvm.func spir_funccc @_Z27__spirv_GroupNonUniformFMaxiif(i32, i32, f32) -> f32 attributes {no_unwind, will_return}
-// CHECK: llvm.func spir_funccc @_Z27__spirv_GroupNonUniformUMaxiij(i32, i32, i32) -> i32 attributes {no_unwind, will_return}
-// CHECK: llvm.func spir_funccc @_Z27__spirv_GroupNonUniformSMaxiij(i32, i32, i32) -> i32 attributes {no_unwind, will_return}
-// CHECK: llvm.func spir_funccc @_Z27__spirv_GroupNonUniformFMiniifj(i32, i32, f32, i32) -> f32 attributes {no_unwind, will_return}
-// CHECK: llvm.func spir_funccc @_Z27__spirv_GroupNonUniformFMiniif(i32, i32, f32) -> f32 attributes {no_unwind, will_return}
-// CHECK: llvm.func spir_funccc @_Z27__spirv_GroupNonUniformUMiniij(i32, i32, i32) -> i32 attributes {no_unwind, will_return}
-// CHECK: llvm.func spir_funccc @_Z27__spirv_GroupNonUniformSMiniij(i32, i32, i32) -> i32 attributes {no_unwind, will_return}
-// CHECK: llvm.func spir_funccc @_Z27__spirv_GroupNonUniformFMuliif(i32, i32, f32) -> f32 attributes {no_unwind, will_return}
-// CHECK: llvm.func spir_funccc @_Z27__spirv_GroupNonUniformIMuliijj(i32, i32, i32, i32) -> i32 attributes {no_unwind, will_return}
-// CHECK: llvm.func spir_funccc @_Z27__spirv_GroupNonUniformFAddiifj(i32, i32, f32, i32) -> f32 attributes {no_unwind, will_return}
-// CHECK: llvm.func spir_funccc @_Z27__spirv_GroupNonUniformIAddiij(i32, i32, i32) -> i32 attributes {no_unwind, will_return}
+// CHECK-LABEL: llvm.func spir_funccc @_Z33__spirv_GroupNonUniformLogicalXoriib(i32, i32, i1) -> i1 attributes {convergent, no_unwind, will_return}
+// CHECK: llvm.func spir_funccc @_Z32__spirv_GroupNonUniformLogicalOriib(i32, i32, i1) -> i1 attributes {convergent, no_unwind, will_return}
+// CHECK: llvm.func spir_funccc @_Z33__spirv_GroupNonUniformLogicalAndiib(i32, i32, i1) -> i1 attributes {convergent, no_unwind, will_return}
+// CHECK: llvm.func spir_funccc @_Z33__spirv_GroupNonUniformBitwiseXoriij(i32, i32, i32) -> i32 attributes {convergent, no_unwind, will_return}
+// CHECK: llvm.func spir_funccc @_Z32__spirv_GroupNonUniformBitwiseOriij(i32, i32, i32) -> i32 attributes {convergent, no_unwind, will_return}
+// CHECK: llvm.func spir_funccc @_Z33__spirv_GroupNonUniformBitwiseAndiij(i32, i32, i32) -> i32 attributes {convergent, no_unwind, will_return}
+// CHECK: llvm.func spir_funccc @_Z27__spirv_GroupNonUniformSMaxiijj(i32, i32, i32, i32) -> i32 attributes {convergent, no_unwind, will_return}
+// CHECK: llvm.func spir_funccc @_Z27__spirv_GroupNonUniformFMaxiif(i32, i32, f32) -> f32 attributes {convergent, no_unwind, will_return}
+// CHECK: llvm.func spir_funccc @_Z27__spirv_GroupNonUniformUMaxiij(i32, i32, i32) -> i32 attributes {convergent, no_unwind, will_return}
+// CHECK: llvm.func spir_funccc @_Z27__spirv_GroupNonUniformSMaxiij(i32, i32, i32) -> i32 attributes {convergent, no_unwind, will_return}
+// CHECK: llvm.func spir_funccc @_Z27__spirv_GroupNonUniformFMiniifj(i32, i32, f32, i32) -> f32 attributes {convergent, no_unwind, will_return}
+// CHECK: llvm.func spir_funccc @_Z27__spirv_GroupNonUniformFMiniif(i32, i32, f32) -> f32 attributes {convergent, no_unwind, will_return}
+// CHECK: llvm.func spir_funccc @_Z27__spirv_GroupNonUniformUMiniij(i32, i32, i32) -> i32 attributes {convergent, no_unwind, will_return}
+// CHECK: llvm.func spir_funccc @_Z27__spirv_GroupNonUniformSMiniij(i32, i32, i32) -> i32 attributes {convergent, no_unwind, will_return}
+// CHECK: llvm.func spir_funccc @_Z27__spirv_GroupNonUniformFMuliif(i32, i32, f32) -> f32 attributes {convergent, no_unwind, will_return}
+// CHECK: llvm.func spir_funccc @_Z27__spirv_GroupNonUniformIMuliijj(i32, i32, i32, i32) -> i32 attributes {convergent, no_unwind, will_return}
+// CHECK: llvm.func spir_funccc @_Z27__spirv_GroupNonUniformFAddiifj(i32, i32, f32, i32) -> f32 attributes {convergent, no_unwind, will_return}
+// CHECK: llvm.func spir_funccc @_Z27__spirv_GroupNonUniformIAddiij(i32, i32, i32) -> i32 attributes {convergent, no_unwind, will_return}
// CHECK-LABEL: llvm.func @non_uniform_iadd(
// CHECK-SAME: %[[VAL_0:.*]]: i32) -> i32 {
// CHECK: %[[VAL_1:.*]] = llvm.mlir.constant(3 : i32) : i32
// CHECK: %[[VAL_2:.*]] = llvm.mlir.constant(0 : i32) : i32
-// CHECK: %[[VAL_3:.*]] = llvm.call spir_funccc @_Z27__spirv_GroupNonUniformIAddiij(%[[VAL_1]], %[[VAL_2]], %[[VAL_0]]) {no_unwind, will_return} : (i32, i32, i32) -> i32
+// CHECK: %[[VAL_3:.*]] = llvm.call spir_funccc @_Z27__spirv_GroupNonUniformIAddiij(%[[VAL_1]], %[[VAL_2]], %[[VAL_0]]) {convergent, no_unwind, will_return} : (i32, i32, i32) -> i32
// CHECK: llvm.return %[[VAL_3]] : i32
// CHECK: }
spirv.func @non_uniform_iadd(%arg0: i32) -> i32 "None" {
@@ -38,7 +38,7 @@ spirv.func @non_uniform_iadd(%arg0: i32) -> i32 "None" {
// CHECK: %[[VAL_1:.*]] = llvm.mlir.constant(16 : i32) : i32
// CHECK: %[[VAL_2:.*]] = llvm.mlir.constant(3 : i32) : i32
// CHECK: %[[VAL_3:.*]] = llvm.mlir.constant(3 : i32) : i32
-// CHECK: %[[VAL_4:.*]] = llvm.call spir_funccc @_Z27__spirv_GroupNonUniformFAddiifj(%[[VAL_2]], %[[VAL_3]], %[[VAL_0]], %[[VAL_1]]) {no_unwind, will_return} : (i32, i32, f32, i32) -> f32
+// CHECK: %[[VAL_4:.*]] = llvm.call spir_funccc @_Z27__spirv_GroupNonUniformFAddiifj(%[[VAL_2]], %[[VAL_3]], %[[VAL_0]], %[[VAL_1]]) {convergent, no_unwind, will_return} : (i32, i32, f32, i32) -> f32
// CHECK: llvm.return %[[VAL_4]] : f32
// CHECK: }
spirv.func @non_uniform_fadd(%arg0: f32) -> f32 "None" {
@@ -52,7 +52,7 @@ spirv.func @non_uniform_fadd(%arg0: f32) -> f32 "None" {
// CHECK: %[[VAL_1:.*]] = llvm.mlir.constant(16 : i32) : i32
// CHECK: %[[VAL_2:.*]] = llvm.mlir.constant(3 : i32) : i32
// CHECK: %[[VAL_3:.*]] = llvm.mlir.constant(3 : i32) : i32
-// CHECK: %[[VAL_4:.*]] = llvm.call spir_funccc @_Z27__spirv_GroupNonUniformIMuliijj(%[[VAL_2]], %[[VAL_3]], %[[VAL_0]], %[[VAL_1]]) {no_unwind, will_return} : (i32, i32, i32, i32) -> i32
+// CHECK: %[[VAL_4:.*]] = llvm.call spir_funccc @_Z27__spirv_GroupNonUniformIMuliijj(%[[VAL_2]], %[[VAL_3]], %[[VAL_0]], %[[VAL_1]]) {convergent, no_unwind, will_return} : (i32, i32, i32, i32) -> i32
// CHECK: llvm.return %[[VAL_4]] : i32
// CHECK: }
spirv.func @non_uniform_imul(%arg0: i32) -> i32 "None" {
@@ -65,7 +65,7 @@ spirv.func @non_uniform_imul(%arg0: i32) -> i32 "None" {
// CHECK-SAME: %[[VAL_0:.*]]: f32) -> f32 {
// CHECK: %[[VAL_1:.*]] = llvm.mlir.constant(3 : i32) : i32
// CHECK: %[[VAL_2:.*]] = llvm.mlir.constant(0 : i32) : i32
-// CHECK: %[[VAL_3:.*]] = llvm.call spir_funccc @_Z27__spirv_GroupNonUniformFMuliif(%[[VAL_1]], %[[VAL_2]], %[[VAL_0]]) {no_unwind, will_return} : (i32, i32, f32) -> f32
+// CHECK: %[[VAL_3:.*]] = llvm.call spir_funccc @_Z27__spirv_GroupNonUniformFMuliif(%[[VAL_1]], %[[VAL_2]], %[[VAL_0]]) {convergent, no_unwind, will_return} : (i32, i32, f32) -> f32
// CHECK: llvm.return %[[VAL_3]] : f32
// CHECK: }
spirv.func @non_uniform_fmul(%arg0: f32) -> f32 "None" {
@@ -77,7 +77,7 @@ spirv.func @non_uniform_fmul(%arg0: f32) -> f32 "None" {
// CHECK-SAME: %[[VAL_0:.*]]: i32) -> i32 {
// CHECK: %[[VAL_1:.*]] = llvm.mlir.constant(3 : i32) : i32
// CHECK: %[[VAL_2:.*]] = llvm.mlir.constant(0 : i32) : i32
-// CHECK: %[[VAL_3:.*]] = llvm.call spir_funccc @_Z27__spirv_GroupNonUniformSMiniij(%[[VAL_1]], %[[VAL_2]], %[[VAL_0]]) {no_unwind, will_return} : (i32, i32, i32) -> i32
+// CHECK: %[[VAL_3:.*]] = llvm.call spir_funccc @_Z27__spirv_GroupNonUniformSMiniij(%[[VAL_1]], %[[VAL_2]], %[[VAL_0]]) {convergent, no_unwind, will_return} : (i32, i32, i32) -> i32
// CHECK: llvm.return %[[VAL_3]] : i32
// CHECK: }
spirv.func @non_uniform_smin(%arg0: i32) -> i32 "None" {
@@ -89,7 +89,7 @@ spirv.func @non_uniform_smin(%arg0: i32) -> i32 "None" {
// CHECK-SAME: %[[VAL_0:.*]]: i32) -> i32 {
// CHECK: %[[VAL_1:.*]] = llvm.mlir.constant(3 : i32) : i32
// CHECK: %[[VAL_2:.*]] = llvm.mlir.constant(0 : i32) : i32
-// CHECK: %[[VAL_3:.*]] = llvm.call spir_funccc @_Z27__spirv_GroupNonUniformUMiniij(%[[VAL_1]], %[[VAL_2]], %[[VAL_0]]) {no_unwind, will_return} : (i32, i32, i32) -> i32
+// CHECK: %[[VAL_3:.*]] = llvm.call spir_funccc @_Z27__spirv_GroupNonUniformUMiniij(%[[VAL_1]], %[[VAL_2]], %[[VAL_0]]) {convergent, no_unwind, will_return} : (i32, i32, i32) -> i32
// CHECK: llvm.return %[[VAL_3]] : i32
// CHECK: }
spirv.func @non_uniform_umin(%arg0: i32) -> i32 "None" {
@@ -101,7 +101,7 @@ spirv.func @non_uniform_umin(%arg0: i32) -> i32 "None" {
// CHECK-SAME: %[[VAL_0:.*]]: f32) -> f32 {
// CHECK: %[[VAL_1:.*]] = llvm.mlir.constant(3 : i32) : i32
// CHECK: %[[VAL_2:.*]] = llvm.mlir.constant(0 : i32) : i32
-// CHECK: %[[VAL_3:.*]] = llvm.call spir_funccc @_Z27__spirv_GroupNonUniformFMiniif(%[[VAL_1]], %[[VAL_2]], %[[VAL_0]]) {no_unwind, will_return} : (i32, i32, f32) -> f32
+// CHECK: %[[VAL_3:.*]] = llvm.call spir_funccc @_Z27__spirv_GroupNonUniformFMiniif(%[[VAL_1]], %[[VAL_2]], %[[VAL_0]]) {convergent, no_unwind, will_return} : (i32, i32, f32) -> f32
// CHECK: llvm.return %[[VAL_3]] : f32
// CHECK: }
spirv.func @non_uniform_fmin(%arg0: f32) -> f32 "None" {
@@ -114,7 +114,7 @@ spirv.func @non_uniform_fmin(%arg0: f32) -> f32 "None" {
// CHECK: %[[VAL_1:.*]] = llvm.mlir.constant(16 : i32) : i32
// CHECK: %[[VAL_2:.*]] = llvm.mlir.constant(3 : i32) : i32
// CHECK: %[[VAL_3:.*]] = llvm.mlir.constant(3 : i32) : i32
-// CHECK: %[[VAL_4:.*]] = llvm.call spir_funccc @_Z27__spirv_GroupNonUniformFMiniifj(%[[VAL_2]], %[[VAL_3]], %[[VAL_0]], %[[VAL_1]]) {no_unwind, will_return} : (i32, i32, f32, i32) -> f32
+// CHECK: %[[VAL_4:.*]] = llvm.call spir_funccc @_Z27__spirv_GroupNonUniformFMiniifj(%[[VAL_2]], %[[VAL_3]], %[[VAL_0]], %[[VAL_1]]) {convergent, no_unwind, will_return} : (i32, i32, f32, i32) -> f32
// CHECK: llvm.return %[[VAL_4]] : f32
// CHECK: }
spirv.func @non_uniform_fmin_cluster(%arg0: f32) -> f32 "None" {
@@ -127,7 +127,7 @@ spirv.func @non_uniform_fmin_cluster(%arg0: f32) -> f32 "None" {
// CHECK-SAME: %[[VAL_0:.*]]: i32) -> i32 {
// CHECK: %[[VAL_1:.*]] = llvm.mlir.constant(3 : i32) : i32
// CHECK: %[[VAL_2:.*]] = llvm.mlir.constant(0 : i32) : i32
-// CHECK: %[[VAL_3:.*]] = llvm.call spir_funccc @_Z27__spirv_GroupNonUniformSMaxiij(%[[VAL_1]], %[[VAL_2]], %[[VAL_0]]) {no_unwind, will_return} : (i32, i32, i32) -> i32
+// CHECK: %[[VAL_3:.*]] = llvm.call spir_funccc @_Z27__spirv_GroupNonUniformSMaxiij(%[[VAL_1]], %[[VAL_2]], %[[VAL_0]]) {convergent, no_unwind, will_return} : (i32, i32, i32) -> i32
// CHECK: llvm.return %[[VAL_3]] : i32
// CHECK: }
spirv.func @non_uniform_smax(%arg0: i32) -> i32 "None" {
@@ -139,7 +139,7 @@ spirv.func @non_uniform_smax(%arg0: i32) -> i32 "None" {
// CHECK-SAME: %[[VAL_0:.*]]: i32) -> i32 {
// CHECK: %[[VAL_1:.*]] = llvm.mlir.constant(3 : i32) : i32
// CHECK: %[[VAL_2:.*]] = llvm.mlir.constant(0 : i32) : i32
-// CHECK: %[[VAL_3:.*]] = llvm.call spir_funccc @_Z27__spirv_GroupNonUniformUMaxiij(%[[VAL_1]], %[[VAL_2]], %[[VAL_0]]) {no_unwind, will_return} : (i32, i32, i32) -> i32
+// CHECK: %[[VAL_3:.*]] = llvm.call spir_funccc @_Z27__spirv_GroupNonUniformUMaxiij(%[[VAL_1]], %[[VAL_2]], %[[VAL_0]]) {convergent, no_unwind, will_return} : (i32, i32, i32) -> i32
// CHECK: llvm.return %[[VAL_3]] : i32
// CHECK: }
spirv.func @non_uniform_umax(%arg0: i32) -> i32 "None" {
@@ -151,7 +151,7 @@ spirv.func @non_uniform_umax(%arg0: i32) -> i32 "None" {
// CHECK-SAME: %[[VAL_0:.*]]: f32) -> f32 {
// CHECK: %[[VAL_1:.*]] = llvm.mlir.constant(3 : i32) : i32
// CHECK: %[[VAL_2:.*]] = llvm.mlir.constant(0 : i32) : i32
-// CHECK: %[[VAL_3:.*]] = llvm.call spir_funccc @_Z27__spirv_GroupNonUniformFMaxiif(%[[VAL_1]], %[[VAL_2]], %[[VAL_0]]) {no_unwind, will_return} : (i32, i32, f32) -> f32
+// CHECK: %[[VAL_3:.*]] = llvm.call spir_funccc @_Z27__spirv_GroupNonUniformFMaxiif(%[[VAL_1]], %[[VAL_2]], %[[VAL_0]]) {convergent, no_unwind, will_return} : (i32, i32, f32) -> f32
// CHECK: llvm.return %[[VAL_3]] : f32
// CHECK: }
spirv.func @non_uniform_fmax(%arg0: f32) -> f32 "None" {
@@ -164,7 +164,7 @@ spirv.func @non_uniform_fmax(%arg0: f32) -> f32 "None" {
// CHECK: %[[VAL_1:.*]] = llvm.mlir.constant(16 : i32) : i32
// CHECK: %[[VAL_2:.*]] = llvm.mlir.constant(3 : i32) : i32
// CHECK: %[[VAL_3:.*]] = llvm.mlir.constant(3 : i32) : i32
-// CHECK: %[[VAL_4:.*]] = llvm.call spir_funccc @_Z27__spirv_GroupNonUniformSMaxiijj(%[[VAL_2]], %[[VAL_3]], %[[VAL_0]], %[[VAL_1]]) {no_unwind, will_return} : (i32, i32, i32, i32) -> i32
+// CHECK: %[[VAL_4:.*]] = llvm.call spir_funccc @_Z27__spirv_GroupNonUniformSMaxiijj(%[[VAL_2]], %[[VAL_3]], %[[VAL_0]], %[[VAL_1]]) {convergent, no_unwind, will_return} : (i32, i32, i32, i32) -> i32
// CHECK: llvm.return %[[VAL_4]] : i32
// CHECK: }
spirv.func @non_uniform_smax_cluster(%arg0: i32) -> i32 "None" {
@@ -177,7 +177,7 @@ spirv.func @non_uniform_smax_cluster(%arg0: i32) -> i32 "None" {
// CHECK-SAME: %[[VAL_0:.*]]: i32) -> i32 {
// CHECK: %[[VAL_1:.*]] = llvm.mlir.constant(3 : i32) : i32
// CHECK: %[[VAL_2:.*]] = llvm.mlir.constant(0 : i32) : i32
-// CHECK: %[[VAL_3:.*]] = llvm.call spir_funccc @_Z33__spirv_GroupNonUniformBitwiseAndiij(%[[VAL_1]], %[[VAL_2]], %[[VAL_0]]) {no_unwind, will_return} : (i32, i32, i32) -> i32
+// CHECK: %[[VAL_3:.*]] = llvm.call spir_funccc @_Z33__spirv_GroupNonUniformBitwiseAndiij(%[[VAL_1]], %[[VAL_2]], %[[VAL_0]]) {convergent, no_unwind, will_return} : (i32, i32, i32) -> i32
// CHECK: llvm.return %[[VAL_3]] : i32
// CHECK: }
spirv.func @non_uniform_bitwise_and(%arg0: i32) -> i32 "None" {
@@ -189,7 +189,7 @@ spirv.func @non_uniform_bitwise_and(%arg0: i32) -> i32 "None" {
// CHECK-SAME: %[[VAL_0:.*]]: i32) -> i32 {
// CHECK: %[[VAL_1:.*]] = llvm.mlir.constant(3 : i32) : i32
// CHECK: %[[VAL_2:.*]] = llvm.mlir.constant(0 : i32) : i32
-// CHECK: %[[VAL_3:.*]] = llvm.call spir_funccc @_Z32__spirv_GroupNonUniformBitwiseOriij(%[[VAL_1]], %[[VAL_2]], %[[VAL_0]]) {no_unwind, will_return} : (i32, i32, i32) -> i32
+// CHECK: %[[VAL_3:.*]] = llvm.call spir_funccc @_Z32__spirv_GroupNonUniformBitwiseOriij(%[[VAL_1]], %[[VAL_2]], %[[VAL_0]]) {convergent, no_unwind, will_return} : (i32, i32, i32) -> i32
// CHECK: llvm.return %[[VAL_3]] : i32
// CHECK: }
spirv.func @non_uniform_bitwise_or(%arg0: i32) -> i32 "None" {
@@ -201,7 +201,7 @@ spirv.func @non_uniform_bitwise_or(%arg0: i32) -> i32 "None" {
// CHECK-SAME: %[[VAL_0:.*]]: i32) -> i32 {
// CHECK: %[[VAL_1:.*]] = llvm.mlir.constant(3 : i32) : i32
// CHECK: %[[VAL_2:.*]] = llvm.mlir.constant(0 : i32) : i32
-// CHECK: %[[VAL_3:.*]] = llvm.call spir_funccc @_Z33__spirv_GroupNonUniformBitwiseXoriij(%[[VAL_1]], %[[VAL_2]], %[[VAL_0]]) {no_unwind, will_return} : (i32, i32, i32) -> i32
+// CHECK: %[[VAL_3:.*]] = llvm.call spir_funccc @_Z33__spirv_GroupNonUniformBitwiseXoriij(%[[VAL_1]], %[[VAL_2]], %[[VAL_0]]) {convergent, no_unwind, will_return} : (i32, i32, i32) -> i32
// CHECK: llvm.return %[[VAL_3]] : i32
// CHECK: }
spirv.func @non_uniform_bitwise_xor(%arg0: i32) -> i32 "None" {
@@ -213,7 +213,7 @@ spirv.func @non_uniform_bitwise_xor(%arg0: i32) -> i32 "None" {
// CHECK-SAME: %[[VAL_0:.*]]: i1) -> i1 {
// CHECK: %[[VAL_1:.*]] = llvm.mlir.constant(3 : i32) : i32
// CHECK: %[[VAL_2:.*]] = llvm.mlir.constant(0 : i32) : i32
-// CHECK: %[[VAL_3:.*]] = llvm.call spir_funccc @_Z33__spirv_GroupNonUniformLogicalAndiib(%[[VAL_1]], %[[VAL_2]], %[[VAL_0]]) {no_unwind, will_return} : (i32, i32, i1) -> i1
+// CHECK: %[[VAL_3:.*]] = llvm.call spir_funccc @_Z33__spirv_GroupNonUniformLogicalAndiib(%[[VAL_1]], %[[VAL_2]], %[[VAL_0]]) {convergent, no_unwind, will_return} : (i32, i32, i1) -> i1
// CHECK: llvm.return %[[VAL_3]] : i1
// CHECK: }
spirv.func @non_uniform_logical_and(%arg0: i1) -> i1 "None" {
@@ -225,7 +225,7 @@ spirv.func @non_uniform_logical_and(%arg0: i1) -> i1 "None" {
// CHECK-SAME: ...
[truncated]
|
victor-eds
approved these changes
Jan 8, 2025
kuhar
reviewed
Jan 8, 2025
The overview defines a convergent operation as:
Based on my understanding, the non-uniform SPIR-V group operations fulfill this definition, because they (a) involve inter-thread communication (reduction) and (b) the set of threads participating is affected by control flow. |
|
I thought that this is what the old code used to do, no? Now |
I'm not sure I understand what you mean here. My understanding is that the non-uniform SPIR-V functions fulfill the definition of a convergent operation and should therefore be marked as such, which the new version of Is that what you mean? |
This I agree with.
This I don't follow. Looking at the previous implementation of this code, I'd think that this was already being handled by LLVM::LLVMFuncOp func = lookupOrCreateSPIRVFn(
symbolTable, funcName, paramTypes, retTy, !NonUniform);I'm not very familiar with this code, and it seems odd to me that a very generic-sounding function like |
|
I see what @kuhar means. Maybe simply doing: LLVM::LLVMFuncOp func = lookupOrCreateSPIRVFn(
symbolTable, funcName, paramTypes, retTy, /*isConvergent=*/true);and keeping |
Yeah, sorry for misunderstanding your point yesterday @kuhar. I had only introduced the The use of this function that I changed in this PR was the only use of that function that specifies a value for The only other use of of That's why I figured I could remove the However, if you and @victor-eds prefer to keep the parameter, I can bring it back, no problem. WDYT? |
I'd rather keep the option as we may have to bring it back again in the future. |
Signed-off-by: Lukas Sommer <[email protected]>
Ok, done in the latest commit. |
victor-eds
approved these changes
Jan 9, 2025
kuhar
approved these changes
Jan 9, 2025
BaiXilin
pushed a commit
to BaiXilin/llvm-fix-vnni-instr-types
that referenced
this pull request
Jan 12, 2025
Add the `convergent` attribute to builtin functions and builtin function calls when lowering SPIR-V non-uniform group functions to LLVM dialect. --------- Signed-off-by: Lukas Sommer <[email protected]>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Add the
convergentattribute to builtin functions and builtin function calls when lowering SPIR-V non-uniform group functions to LLVM dialect.