[mlir][sparse][gpu] force 16-byte alignment on data structs for cuSparseLt

[aartbik]

Summary

Also makes some minor consistency edits in the cuSparseLt wrapper lib.

[Peiming]

mlir/lib/ExecutionEngine/CudaRuntimeWrappers.cpp

557–558

used "Assume " in the above comment

643

unnecessary change?

[aartbik]

addressed comments

[aartbik]

rebased with main

mlir/lib/Conversion/GPUCommon/GPUToLLVMConversion.cpp

LogicalResult ConvertCreateDnTensorOpToGpuRuntimeCallPattern::matchAndRewrite(
   // used, we require two separate Creation ops to be the correct logic. In
   // future, we may add support to using one handle in sparse tensor / GPU
   // dialect in both cusparse and cusparseLt. use the cusparseLt create call if
   // the dnmat is used with spmat with 2:4 sparsity
   if (dims.size() == 2) {
     if (isSpMMCusparseLtOp(op.getDnTensor())) {
       auto handleSz = rewriter.create<LLVM::ConstantOp>(
           loc, getIndexType(), rewriter.getIndexAttr(11032));
-      handle = rewriter.create<LLVM::AllocaOp>(loc, llvmInt8PointerType,
-                                               llvmInt8Type, handleSz);
+      handle = rewriter.create<LLVM::AllocaOp>(
+          loc, llvmInt8PointerType, llvmInt8Type, handleSz, /*alignment=*/16);
       handle = rewriter.create<LLVM::BitcastOp>(loc, llvmPointerType, handle);
 
       createLtDnMatCallBuilder
           .create(loc, rewriter,
                   {handle, dims[0], dims[1], pTensor, dtp, stream})
           .getResult();
     } else {
       handle =
if (!getTypeConverter()->useOpaquePointers())
     pMat = rewriter.create<LLVM::BitcastOp>(loc, llvmPointerType, pMat);
   Type dType =
       llvm::cast<MemRefType>(op.getMemref().getType()).getElementType();
   auto dtp = genConstInt32From(rewriter, loc, getCuSparseDataTypeFrom(dType));
 
   // CUDA runner asserts the size is 44104 bytes.
   auto handleSz = rewriter.create<LLVM::ConstantOp>(
       loc, getIndexType(), rewriter.getIndexAttr(44104));
-  Value handle = rewriter.create<LLVM::AllocaOp>(loc, llvmInt8PointerType,
-                                                 llvmInt8Type, handleSz);
+  Value handle = rewriter.create<LLVM::AllocaOp>(
+      loc, llvmInt8PointerType, llvmInt8Type, handleSz, /*alignment=*/16);
   handle = rewriter.create<LLVM::BitcastOp>(loc, llvmPointerType, handle);
 
   create2To4SpMatCallBuilder
       .create(loc, rewriter,
               {handle, adaptor.getRows(), adaptor.getCols(), pMat, dtp, stream})
       .getResult();
   rewriter.replaceOp(op, {handle, stream});
   return success();
LogicalResult ConvertSpMMBufferSizeOpToGpuRuntimeCallPattern::matchAndRewrite(
   auto modeB = genConstInt32From(rewriter, loc, adaptor.getModeB());
   auto stream = adaptor.getAsyncDependencies().front();
   Value bufferSize;
   if (is2To4Sparsity(op.getSpmatA())) {
     auto computeType = genConstInt32From(
         rewriter, loc, getCuSparseLtDataTypeFrom(adaptor.getComputeType()));
     auto three = rewriter.create<LLVM::ConstantOp>(loc, getIndexType(),
                                                    rewriter.getIndexAttr(3));
-    auto bufferSize = rewriter.create<LLVM::AllocaOp>(loc, llvmInt64PointerType,
-                                                      llvmInt64Type, three);
+    auto bufferSize = rewriter.create<LLVM::AllocaOp>(
+        loc, llvmInt64PointerType, llvmInt64Type, three, /*alignment=*/16);
     createCuSparseLtSpMMBufferSizeBuilder
         .create(loc, rewriter,
                 {bufferSize, modeA, modeB, adaptor.getSpmatA(),
                  adaptor.getDnmatB(), adaptor.getDnmatC(), computeType, stream})
         .getResult();
 
     auto bufferSizePtr1 = rewriter.create<LLVM::GEPOp>(
         loc, llvmInt64PointerType, llvmInt64PointerType, bufferSize,

mlir/lib/ExecutionEngine/CudaRuntimeWrappers.cpp

 static_assert(sizeof(cusparseLtDnMatHandleAndData) == 11032,
               "Unexpected cusparseLt dense matrix handle size");
 
 extern "C" MLIR_CUDA_WRAPPERS_EXPORT void mgpuCreateSparseLtEnv() {
   // ScopedContext is for cuda initialization.
   ScopedContext scopedContext;
   assert(!cusparseLt_initiated &&
          "client called mgpuCreateSparseLtEnv() twice");
-  // Note that cuSparseLt still uses cusparseStatus_t
+  // Note that cuSparseLt still uses cusparseStatus_t.
   CUSPARSE_REPORT_IF_ERROR(cusparseLtInit(&cusparseLt_env));
   cusparseLt_initiated = true;
 }
 
 extern "C" MLIR_CUDA_WRAPPERS_EXPORT void mgpuDestroySparseLtEnv() {
   assert(cusparseLt_initiated && "client did not call mgpuCreateSparseLtEnv()");
   CUSPARSE_REPORT_IF_ERROR(cusparseLtDestroy(&cusparseLt_env));
   cusparseLt_initiated = false;
 }
 
 extern "C" MLIR_CUDA_WRAPPERS_EXPORT void
 mgpuCreateCuSparseLtDnMat(void *dh, intptr_t rows, intptr_t cols, void *values,
                           int32_t dtp, CUstream /*stream*/) {
   assert(cusparseLt_initiated && "client did not call mgpuCreateSparseLtEnv()");
-  // CusparseLt expects the descriptors to be zero-initialized.
-  memset(dh, 0, sizeof(cusparseLtDnMatHandleAndData));
   auto dnmat_handle = reinterpret_cast<cusparseLtDnMatHandleAndData *>(dh);
+  // CusparseLt expects the descriptors to be zero-initialized.
+  memset(dnmat_handle, 0, sizeof(cusparseLtDnMatHandleAndData));
+  dnmat_handle->values = values;
   auto dTp = static_cast<cudaDataType_t>(dtp);
-  // assuming row-major when deciding lda
+  // Assume row-major when deciding lda.
+  const uint32_t alignment = 16;
   CUSPARSE_REPORT_IF_ERROR(cusparseLtDenseDescriptorInit(
       &cusparseLt_env, &(dnmat_handle->mat), rows, cols, /*lda=*/cols,
-      /*alignment=*/16, dTp, CUSPARSE_ORDER_ROW))
-  dnmat_handle->values = values;
-}
-
-// This can be used to destroy both dense matrices and sparse matrices in
-// cusparseLt
-extern "C" MLIR_CUDA_WRAPPERS_EXPORT void
-mgpuDestroyCuSparseLtSpMat(void *m, CUstream /*stream*/) {
-  auto matAndData = reinterpret_cast<cusparseLtSpMatHandleAndData *>(m);
-  CUSPARSE_REPORT_IF_ERROR(cusparseLtMatDescriptorDestroy(&(matAndData->mat)))
+      alignment, dTp, CUSPARSE_ORDER_ROW))
 }
 
 extern "C" MLIR_CUDA_WRAPPERS_EXPORT void
-mgpuDestroyCuSparseLtDnMat(void *m, CUstream /*stream*/) {
-  auto matAndData = reinterpret_cast<cusparseLtDnMatHandleAndData *>(m);
-  CUSPARSE_REPORT_IF_ERROR(cusparseLtMatDescriptorDestroy(&(matAndData->mat)))
+mgpuDestroyCuSparseLtDnMat(void *dh, CUstream /*stream*/) {
+  auto dnmat_handle = reinterpret_cast<cusparseLtDnMatHandleAndData *>(dh);
+  CUSPARSE_REPORT_IF_ERROR(cusparseLtMatDescriptorDestroy(&(dnmat_handle->mat)))
 }
 
 extern "C" MLIR_CUDA_WRAPPERS_EXPORT void
 mgpuCusparseLtCreate2To4SpMat(void *sh, intptr_t rows, intptr_t cols,
                               void *values, int32_t dtp, CUstream /*stream*/) {
   assert(cusparseLt_initiated && "client did not call mgpuCreateSparseLtEnv()");
   auto spmat_handle = reinterpret_cast<cusparseLtSpMatHandleAndData *>(sh);
   // CusparseLt expects the descriptors to be zero-initialized.
   memset(spmat_handle, 0, sizeof(cusparseLtSpMatHandleAndData));
   spmat_handle->values = values;
   auto dTp = static_cast<cudaDataType_t>(dtp);
-  // assuming row-major when deciding lda
+  // Assume row-major when deciding lda.
+  const uint32_t alignment = 16;
   CUSPARSE_REPORT_IF_ERROR(cusparseLtStructuredDescriptorInit(
-      &cusparseLt_env, &(spmat_handle->mat), rows, cols, /*ld=*/cols,
-      /*alignment=*/16, dTp, CUSPARSE_ORDER_ROW,
-      CUSPARSELT_SPARSITY_50_PERCENT))
+      &cusparseLt_env, &(spmat_handle->mat), rows, cols, /*ld=*/cols, alignment,
+      dTp, CUSPARSE_ORDER_ROW, CUSPARSELT_SPARSITY_50_PERCENT))
+}
+
+extern "C" MLIR_CUDA_WRAPPERS_EXPORT void
+mgpuDestroyCuSparseLtSpMat(void *sh, CUstream /*stream*/) {
+  auto spmat_handle = reinterpret_cast<cusparseLtSpMatHandleAndData *>(sh);
+  CUSPARSE_REPORT_IF_ERROR(cusparseLtMatDescriptorDestroy(&(spmat_handle->mat)))
 }
 
 // Several things are being done in this stage, algorithm selection, planning,
 // and returning workspace and compressed matrices data buffer sizes.
 extern "C" MLIR_CUDA_WRAPPERS_EXPORT void
 mgpuCuSparseLtSpMMBufferSize(void *bs, int32_t ma, int32_t mb, void *a, void *b,
                              void *c, int32_t ctp, CUstream /*stream*/) {
   assert(cusparseLt_initiated && "client did not call mgpuCreateSparseLtEnv()");
CUSPARSE_REPORT_IF_ERROR(cusparseLtMatmulPlanInit(
       &cusparseLt_env, &(matA->plan), &(matA->matmul), &(matA->alg_sel)))
 
   CUSPARSE_REPORT_IF_ERROR(cusparseLtMatmulGetWorkspace(
       &cusparseLt_env, &(matA->plan), &workspace_size_))
   CUSPARSE_REPORT_IF_ERROR(cusparseLtSpMMACompressedSize(
       &cusparseLt_env, &(matA->plan), &compressed_size_,
       &compressed_buffer_size_))
 
-  // avoid zero-alloc
+  // Avoid zero-allocation.
   *workspace_size = (workspace_size_ == 0 ? 1 : workspace_size_);
   *compressed_size = (compressed_size_ == 0 ? 1 : compressed_size_);
   *compressed_buffer_size =
       (compressed_buffer_size_ == 0 ? 1 : compressed_buffer_size_);
 }
 
 extern "C" MLIR_CUDA_WRAPPERS_EXPORT void
 mgpuCuSparseLtSpMM(void *a, void *b, void *c, void *d_workspace,
CUSPARSE_REPORT_IF_ERROR(
                        /*dD*/ matC->values, d_workspace, nullptr, 0))
 
   CUSPARSE_REPORT_IF_ERROR(cusparseLtMatDescriptorDestroy(&(matA->mat)))
   // destroy the plan associated with the sparse matrix
   CUSPARSE_REPORT_IF_ERROR(cusparseLtMatmulPlanDestroy(&(matA->plan)))
 }
 
 #endif // MLIR_ENABLE_CUDA_CUSPARSELT
 #endif // MLIR_ENABLE_CUDA_CUSPARSE