Skip to content

feat: syrk cuda lowering #1703

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
avik-pal merged 8 commits into from
Dec 9, 2025
+196 −1
Merged

feat: syrk cuda lowering #1703

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
8aa4037
feat: syrk cuda lowering
avik-pal Dec 7, 2025
5588c69
fix: use jax convention
avik-pal Dec 8, 2025
c61407a
test: update
avik-pal Dec 8, 2025
47a52f6
fix: ??
avik-pal Dec 8, 2025
245a66b
feat: lower to our custom ffi
avik-pal Dec 8, 2025
42c68f9
feat: use alpha/beta attributes
avik-pal Dec 8, 2025
52eb8a3
test: update
avik-pal Dec 8, 2025
4dcc9a1
chore: run fmt
avik-pal Dec 8, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
134 changes: 133 additions & 1 deletion src/enzyme_ad/jax/Passes/LowerEnzymeXLABLAS.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -27,6 +27,50 @@ using namespace mlir::enzyme;
using namespace mlir::enzymexla;
using namespace mlir::stablehlo;

// Helper function to extract constant scalar value (real/imag parts)
static bool extractConstantScalar(Value val, double &realPart,
double &imagPart) {
DenseElementsAttr attr;
if (!matchPattern(val, m_Constant(&attr)))
return false;

auto valType = cast<RankedTensorType>(val.getType());
auto elemType = valType.getElementType();

if (auto complexType = dyn_cast<ComplexType>(elemType)) {
// Complex scalar
auto complexVal = attr.getSplatValue<std::complex<APFloat>>();
realPart = complexVal.real().convertToDouble();
imagPart = complexVal.imag().convertToDouble();
return true;
} else if (isa<FloatType>(elemType)) {
// Real scalar
realPart = attr.getSplatValue<APFloat>().convertToDouble();
imagPart = 0.0;
return true;
}
return false;
}

// Helper function to create operand and rank for scalar value
// Returns the operand to use and the rank (1 for empty placeholder, 0 for
// scalar)
static std::pair<Value, int64_t> createScalarOperand(PatternRewriter &rewriter,
Location loc,
Value originalVal,
bool useAttribute) {
if (useAttribute) {
// Create an empty 0-element tensor as placeholder
auto emptyType = RankedTensorType::get(
{0}, cast<RankedTensorType>(originalVal.getType()).getElementType());
auto emptyTensor = stablehlo::ConstantOp::create(
rewriter, loc,
DenseElementsAttr::get(emptyType, ArrayRef<Attribute>{}));
return {emptyTensor, 1};
}
return {originalVal, 0};
}

struct SyrkOpLowering : public OpRewritePattern<enzymexla::SyrkOp> {
using OpRewritePattern<enzymexla::SyrkOp>::OpRewritePattern;

Expand All @@ -39,6 +83,8 @@ struct SyrkOpLowering : public OpRewritePattern<enzymexla::SyrkOp> {
PatternRewriter &rewriter) const override {
if (backend == "cpu")
return matchAndRewriteCPU(op, rewriter);
if (backend == "cuda")
return matchAndRewriteCUDA(op, rewriter);

return matchAndRewriteFallback(op, rewriter);
}
Expand Down Expand Up @@ -248,7 +294,93 @@ struct SyrkOpLowering : public OpRewritePattern<enzymexla::SyrkOp> {
return success();
}

// TODO: gpu lowering after we register the cublas functions via XLA FFI
LogicalResult matchAndRewriteCUDA(enzymexla::SyrkOp op,
PatternRewriter &rewriter) const {
auto CType = cast<RankedTensorType>(op.getC().getType());
auto rank = CType.getRank();

bool isComplex = false;
if (auto complex_type = dyn_cast<ComplexType>(CType.getElementType())) {
isComplex = true;
}

if (isComplex && op.getTranspose() == enzymexla::LapackTranspose::adjoint) {
return rewriter.notifyMatchFailure(
op, "Complex matrix not supported for complex transpose");
}

bool transpose = op.getTranspose() != enzymexla::LapackTranspose::none;

// Try to extract alpha and beta as constants
double alphaReal = 0.0, alphaImag = 0.0;
double betaReal = 0.0, betaImag = 0.0;
bool useAlphaAttr =
extractConstantScalar(op.getAlpha(), alphaReal, alphaImag);
bool useBetaAttr = extractConstantScalar(op.getBeta(), betaReal, betaImag);

// Build operands list - use empty tensors for constant alpha/beta
SmallVector<Value> operands;
operands.push_back(op.getA());
operands.push_back(op.getC());

SmallVector<int64_t> operandRanks = {rank, rank};

auto [alphaOperand, alphaRank] =
createScalarOperand(rewriter, op.getLoc(), op.getAlpha(), useAlphaAttr);
operands.push_back(alphaOperand);
operandRanks.push_back(alphaRank);

auto [betaOperand, betaRank] =
createScalarOperand(rewriter, op.getLoc(), op.getBeta(), useBetaAttr);
operands.push_back(betaOperand);
operandRanks.push_back(betaRank);

auto customCall = stablehlo::CustomCallOp::create(
rewriter, op.getLoc(), TypeRange{CType}, operands,
rewriter.getStringAttr("reactant_cublas_syrk_ffi"),
/*has_side_effect*/ nullptr,
/*backend_config*/
rewriter.getDictionaryAttr({
rewriter.getNamedAttr("transpose", rewriter.getBoolAttr(transpose)),
rewriter.getNamedAttr(
"uplo",
rewriter.getBoolAttr(op.getUplo() == enzymexla::LapackUplo::U)),
rewriter.getNamedAttr("use_alpha_attribute",
rewriter.getBoolAttr(useAlphaAttr)),
rewriter.getNamedAttr("use_beta_attribute",
rewriter.getBoolAttr(useBetaAttr)),
rewriter.getNamedAttr("alpha_real",
rewriter.getF64FloatAttr(alphaReal)),
rewriter.getNamedAttr("alpha_imag",
rewriter.getF64FloatAttr(alphaImag)),
rewriter.getNamedAttr("beta_real",
rewriter.getF64FloatAttr(betaReal)),
rewriter.getNamedAttr("beta_imag",
rewriter.getF64FloatAttr(betaImag)),
}),
/*api_version*/
stablehlo::CustomCallApiVersionAttr::get(
rewriter.getContext(),
mlir::stablehlo::CustomCallApiVersion::API_VERSION_TYPED_FFI),
/*calledcomputations*/ nullptr,
/*operand_layouts*/
getSHLOLayout(rewriter, operandRanks, SmallVector<bool>(4, true), rank),
/*result_layouts*/
getSHLOLayout(rewriter, {rank}, SmallVector<bool>(rank, true), rank),
/*output_operand_aliases*/
rewriter.getArrayAttr({
stablehlo::OutputOperandAliasAttr::get(op.getContext(), {}, 1, {}),
}));

auto result = customCall.getResult(0);
if (op.getFill() || op.getUplo() == enzymexla::LapackUplo::L) {
result = stablehlo::copyTriangularPart(rewriter, result,
enzymexla::LapackUplo::U);
}
rewriter.replaceAllUsesWith(op.getResult(), result);

return success();
}

LogicalResult matchAndRewriteFallback(enzymexla::SyrkOp op,
PatternRewriter &rewriter) const {
Expand Down
63 changes: 63 additions & 0 deletions test/lit_tests/linalg/syrk.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,4 +1,5 @@
// RUN: enzymexlamlir-opt --lower-enzymexla-blas="backend=cpu" --enzyme-hlo-opt %s | FileCheck %s --check-prefix=CPU
// RUN: enzymexlamlir-opt --lower-enzymexla-blas="backend=cuda" --enzyme-hlo-opt %s | FileCheck %s --check-prefix=CUDA
// RUN: enzymexlamlir-opt --lower-enzymexla-blas="backend=tpu" --enzyme-hlo-opt %s | FileCheck %s --check-prefix=TPU

module {
Expand Down Expand Up @@ -36,6 +37,18 @@ module {
// CPU-NEXT: return %5 : tensor<64x64xf32>
// CPU-NEXT: }

// CUDA: func.func @main1(%arg0: tensor<64x32xf32>, %arg1: tensor<64x64xf32>) -> tensor<64x64xf32> {
// CUDA-NEXT: %0 = tensor.empty() : tensor<0xf32>
// CUDA-NEXT: %1 = tensor.empty() : tensor<0xf32>
// CUDA-NEXT: %2 = stablehlo.custom_call @reactant_cublas_syrk_ffi(%arg0, %arg1, %0, %1) {api_version = 4 : i32, backend_config = {alpha_imag = 0.000000e+00 : f64, alpha_real = 2.000000e+00 : f64, beta_imag = 0.000000e+00 : f64, beta_real = 3.000000e+00 : f64, transpose = false, uplo = true, use_alpha_attribute = true, use_beta_attribute = true}, enzymexla.symmetric_matrix = [#enzymexla<guaranteed NOTGUARANTEED>], operand_layouts = [dense<[0, 1]> : tensor<2xindex>, dense<[0, 1]> : tensor<2xindex>, dense<0> : tensor<1xindex>, dense<0> : tensor<1xindex>], output_operand_aliases = [#stablehlo.output_operand_alias<output_tuple_indices = [], operand_index = 1, operand_tuple_indices = []>], result_layouts = [dense<[0, 1]> : tensor<2xindex>]} : (tensor<64x32xf32>, tensor<64x64xf32>, tensor<0xf32>, tensor<0xf32>) -> tensor<64x64xf32>
// CUDA-NEXT: %3 = stablehlo.iota dim = 0 : tensor<64x64xi32>
// CUDA-NEXT: %4 = stablehlo.iota dim = 1 : tensor<64x64xi32>
// CUDA-NEXT: %5 = stablehlo.compare LT, %3, %4 : (tensor<64x64xi32>, tensor<64x64xi32>) -> tensor<64x64xi1>
// CUDA-NEXT: %6 = stablehlo.transpose %2, dims = [1, 0] : (tensor<64x64xf32>) -> tensor<64x64xf32>
// CUDA-NEXT: %7 = stablehlo.select %5, %2, %6 : tensor<64x64xi1>, tensor<64x64xf32>
// CUDA-NEXT: return %7 : tensor<64x64xf32>
// CUDA-NEXT: }

// TPU: func.func @main1(%arg0: tensor<64x32xf32>, %arg1: tensor<64x64xf32>) -> tensor<64x64xf32> {
// TPU-NEXT: %cst = stablehlo.constant dense<3.000000e+00> : tensor<64x64xf32>
// TPU-NEXT: %cst_0 = stablehlo.constant dense<2.000000e+00> : tensor<64x64xf32>
Expand Down Expand Up @@ -81,6 +94,18 @@ module {
// CPU-NEXT: return %0 : tensor<64x64xf32>
// CPU-NEXT: }

// CUDA: func.func @main2(%arg0: tensor<64x32xf32>, %arg1: tensor<64x64xf32>) -> tensor<64x64xf32> {
// CUDA-NEXT: %0 = tensor.empty() : tensor<0xf32>
// CUDA-NEXT: %1 = tensor.empty() : tensor<0xf32>
// CUDA-NEXT: %2 = stablehlo.custom_call @reactant_cublas_syrk_ffi(%arg0, %arg1, %0, %1) {api_version = 4 : i32, backend_config = {alpha_imag = 0.000000e+00 : f64, alpha_real = 2.000000e+00 : f64, beta_imag = 0.000000e+00 : f64, beta_real = 3.000000e+00 : f64, transpose = false, uplo = false, use_alpha_attribute = true, use_beta_attribute = true}, enzymexla.symmetric_matrix = [#enzymexla<guaranteed NOTGUARANTEED>], operand_layouts = [dense<[0, 1]> : tensor<2xindex>, dense<[0, 1]> : tensor<2xindex>, dense<0> : tensor<1xindex>, dense<0> : tensor<1xindex>], output_operand_aliases = [#stablehlo.output_operand_alias<output_tuple_indices = [], operand_index = 1, operand_tuple_indices = []>], result_layouts = [dense<[0, 1]> : tensor<2xindex>]} : (tensor<64x32xf32>, tensor<64x64xf32>, tensor<0xf32>, tensor<0xf32>) -> tensor<64x64xf32>
// CUDA-NEXT: %3 = stablehlo.iota dim = 0 : tensor<64x64xi32>
// CUDA-NEXT: %4 = stablehlo.iota dim = 1 : tensor<64x64xi32>
// CUDA-NEXT: %5 = stablehlo.compare LT, %3, %4 : (tensor<64x64xi32>, tensor<64x64xi32>) -> tensor<64x64xi1>
// CUDA-NEXT: %6 = stablehlo.transpose %2, dims = [1, 0] : (tensor<64x64xf32>) -> tensor<64x64xf32>
// CUDA-NEXT: %7 = stablehlo.select %5, %2, %6 : tensor<64x64xi1>, tensor<64x64xf32>
// CUDA-NEXT: return %7 : tensor<64x64xf32>
// CUDA-NEXT: }

// TPU: func.func @main2(%arg0: tensor<64x32xf32>, %arg1: tensor<64x64xf32>) -> tensor<64x64xf32> {
// TPU-NEXT: %cst = stablehlo.constant dense<3.000000e+00> : tensor<64x64xf32>
// TPU-NEXT: %cst_0 = stablehlo.constant dense<2.000000e+00> : tensor<64x64xf32>
Expand Down Expand Up @@ -126,6 +151,13 @@ module {
// CPU-NEXT: return %0 : tensor<64x64xf32>
// CPU-NEXT: }

// CUDA: func.func @main3(%arg0: tensor<64x32xf32>, %arg1: tensor<64x64xf32>) -> tensor<64x64xf32> {
// CUDA-NEXT: %0 = tensor.empty() : tensor<0xf32>
// CUDA-NEXT: %1 = tensor.empty() : tensor<0xf32>
// CUDA-NEXT: %2 = stablehlo.custom_call @reactant_cublas_syrk_ffi(%arg0, %arg1, %0, %1) {api_version = 4 : i32, backend_config = {alpha_imag = 0.000000e+00 : f64, alpha_real = 2.000000e+00 : f64, beta_imag = 0.000000e+00 : f64, beta_real = 3.000000e+00 : f64, transpose = false, uplo = false, use_alpha_attribute = true, use_beta_attribute = true}, operand_layouts = [dense<[0, 1]> : tensor<2xindex>, dense<[0, 1]> : tensor<2xindex>, dense<0> : tensor<1xindex>, dense<0> : tensor<1xindex>], output_operand_aliases = [#stablehlo.output_operand_alias<output_tuple_indices = [], operand_index = 1, operand_tuple_indices = []>], result_layouts = [dense<[0, 1]> : tensor<2xindex>]} : (tensor<64x32xf32>, tensor<64x64xf32>, tensor<0xf32>, tensor<0xf32>) -> tensor<64x64xf32>
// CUDA-NEXT: return %2 : tensor<64x64xf32>
// CUDA-NEXT: }

// TPU: func.func @main3(%arg0: tensor<64x32xf32>, %arg1: tensor<64x64xf32>) -> tensor<64x64xf32> {
// TPU-NEXT: %cst = stablehlo.constant dense<3.000000e+00> : tensor<64x64xf32>
// TPU-NEXT: %cst_0 = stablehlo.constant dense<2.000000e+00> : tensor<64x64xf32>
Expand Down Expand Up @@ -177,10 +209,41 @@ module {
// CPU-NEXT: return %4 : tensor<4x4xf32>
// CPU-NEXT: }

// CUDA: func.func @main4(%arg0: tensor<5x4xf32>, %arg1: tensor<4x4xf32>) -> tensor<4x4xf32> {
// CUDA-NEXT{LITERAL}: %c = stablehlo.constant dense<[[false, true, true, true], [false, false, true, true], [false, false, false, true], [false, false, false, false]]> : tensor<4x4xi1>
// CUDA-NEXT: %cst = stablehlo.constant dense<5.000000e+00> : tensor<4x4xf32>
// CUDA-NEXT: %cst_0 = stablehlo.constant dense<0.000000e+00> : tensor<4x4xf32>
// CUDA-NEXT: %0 = tensor.empty() : tensor<0xf32>
// CUDA-NEXT: %1 = tensor.empty() : tensor<0xf32>
// CUDA-NEXT: %2 = stablehlo.custom_call @reactant_cublas_syrk_ffi(%arg0, %cst_0, %0, %1) {api_version = 4 : i32, backend_config = {alpha_imag = 0.000000e+00 : f64, alpha_real = 1.000000e+00 : f64, beta_imag = 0.000000e+00 : f64, beta_real = 0.000000e+00 : f64, transpose = true, uplo = true, use_alpha_attribute = true, use_beta_attribute = true}, enzymexla.symmetric_matrix = [#enzymexla<guaranteed NOTGUARANTEED>], operand_layouts = [dense<[0, 1]> : tensor<2xindex>, dense<[0, 1]> : tensor<2xindex>, dense<0> : tensor<1xindex>, dense<0> : tensor<1xindex>], output_operand_aliases = [#stablehlo.output_operand_alias<output_tuple_indices = [], operand_index = 1, operand_tuple_indices = []>], result_layouts = [dense<[0, 1]> : tensor<2xindex>]} : (tensor<5x4xf32>, tensor<4x4xf32>, tensor<0xf32>, tensor<0xf32>) -> tensor<4x4xf32>
// CUDA-NEXT: %3 = stablehlo.transpose %2, dims = [1, 0] : (tensor<4x4xf32>) -> tensor<4x4xf32>
// CUDA-NEXT: %4 = stablehlo.select %c, %2, %3 : tensor<4x4xi1>, tensor<4x4xf32>
// CUDA-NEXT: %5 = stablehlo.multiply %cst, %arg1 : tensor<4x4xf32>
// CUDA-NEXT: %6 = stablehlo.add %4, %5 : tensor<4x4xf32>
// CUDA-NEXT: return %6 : tensor<4x4xf32>
// CUDA-NEXT: }

// TPU: func.func @main4(%arg0: tensor<5x4xf32>, %arg1: tensor<4x4xf32>) -> tensor<4x4xf32> {
// TPU-NEXT: %cst = stablehlo.constant dense<5.000000e+00> : tensor<4x4xf32>
// TPU-NEXT: %0 = stablehlo.dot_general %arg0, %arg0, contracting_dims = [0] x [0] : (tensor<5x4xf32>, tensor<5x4xf32>) -> tensor<4x4xf32>
// TPU-NEXT: %1 = stablehlo.multiply %cst, %arg1 : tensor<4x4xf32>
// TPU-NEXT: %2 = stablehlo.add %0, %1 : tensor<4x4xf32>
// TPU-NEXT: return %2 : tensor<4x4xf32>
// TPU-NEXT: }

module {
func.func @main(%arg0: tensor<64x32xf32>, %arg1: tensor<64x64xf32>, %alpha: tensor<f32>, %beta: tensor<f32>) -> tensor<64x64xf32> {
%0 = enzymexla.blas.syrk %arg0, %arg1, %alpha, %beta {fill, transpose = #enzymexla.transpose<none>, uplo = #enzymexla.uplo<U>} : (tensor<64x32xf32>, tensor<64x64xf32>, tensor<f32>, tensor<f32>) -> tensor<64x64xf32>
return %0 : tensor<64x64xf32>
}
}

// CUDA: func.func @main(%arg0: tensor<64x32xf32>, %arg1: tensor<64x64xf32>, %arg2: tensor<f32>, %arg3: tensor<f32>) -> tensor<64x64xf32> {
// CUDA-NEXT: %0 = stablehlo.custom_call @reactant_cublas_syrk_ffi(%arg0, %arg1, %arg2, %arg3) {api_version = 4 : i32, backend_config = {alpha_imag = 0.000000e+00 : f64, alpha_real = 0.000000e+00 : f64, beta_imag = 0.000000e+00 : f64, beta_real = 0.000000e+00 : f64, transpose = false, uplo = true, use_alpha_attribute = false, use_beta_attribute = false}, enzymexla.symmetric_matrix = [#enzymexla<guaranteed NOTGUARANTEED>], operand_layouts = [dense<[0, 1]> : tensor<2xindex>, dense<[0, 1]> : tensor<2xindex>, dense<> : tensor<0xindex>, dense<> : tensor<0xindex>], output_operand_aliases = [#stablehlo.output_operand_alias<output_tuple_indices = [], operand_index = 1, operand_tuple_indices = []>], result_layouts = [dense<[0, 1]> : tensor<2xindex>]} : (tensor<64x32xf32>, tensor<64x64xf32>, tensor<f32>, tensor<f32>) -> tensor<64x64xf32>
// CUDA-NEXT: %1 = stablehlo.iota dim = 0 : tensor<64x64xi32>
// CUDA-NEXT: %2 = stablehlo.iota dim = 1 : tensor<64x64xi32>
// CUDA-NEXT: %3 = stablehlo.compare LT, %1, %2 : (tensor<64x64xi32>, tensor<64x64xi32>) -> tensor<64x64xi1>
// CUDA-NEXT: %4 = stablehlo.transpose %0, dims = [1, 0] : (tensor<64x64xf32>) -> tensor<64x64xf32>
// CUDA-NEXT: %5 = stablehlo.select %3, %0, %4 : tensor<64x64xi1>, tensor<64x64xf32>
// CUDA-NEXT: return %5 : tensor<64x64xf32>
// CUDA-NEXT: }
Loading