Skip to content

Commit 4ea0a70

Browse files
co63occo63oc
committed
remove CUDA 11.0 message in sparse/gpu
1 parent ea557ba commit 4ea0a70

File tree

8 files changed

+16
-37
lines changed

8 files changed

+16
-37
lines changed

‎paddle/phi/kernels/fusion/gpu/fused_gemm_epilogue_kernel.cu‎

Lines changed: 1 addition & 7 deletions
Original file line numberDiff line numberDiff line change
@@ -77,13 +77,7 @@ void FusedGemmEpilogueKernel(const Context& dev_ctx,
7777
dev_ctx.template Alloc<T>(out);
7878
return;
7979
}
80-
#if defined(PADDLE_WITH_CUDA) && CUDA_VERSION < 11060
81-
PADDLE_THROW(common::errors::Unimplemented(
82-
"The fused_gemm_epilogue operator only support CUDA 11.6 "
83-
"or higher version."));
84-
#endif
85-
#if (defined(PADDLE_WITH_CUDA) && CUDA_VERSION >= 11060) || \
86-
defined(PADDLE_WITH_HIP)
80+
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
8781

8882
dev_ctx.template Alloc<T>(out, out->numel() * sizeof(T));
8983
// (M * K) * (K * N)

‎paddle/phi/kernels/fusion/gpu/fused_layernorm_kernel.cu‎

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -183,12 +183,12 @@ struct DefaultComputeType<half> {
183183
using type = float;
184184
};
185185

186-
#if CUDA_VERSION >= 11000
186+
#if defined(PADDLE_WITH_CUDA)
187187
template <>
188188
struct DefaultComputeType<nv_bfloat16> {
189189
using type = float;
190190
};
191-
#endif // CUDA_VERSION >= 11000
191+
#endif
192192

193193
template <typename T>
194194
class HasCanPackAs {

‎paddle/phi/kernels/sparse/gpu/addmm_kernel.cu‎

Lines changed: 1 addition & 5 deletions
Original file line numberDiff line numberDiff line change
@@ -34,7 +34,7 @@ void AddmmKernelImpl(const Context& dev_ctx,
3434
float beta,
3535
float alpha,
3636
DenseTensor* out) {
37-
#if CUDA_VERSION >= 11000
37+
#if defined(PADDLE_WITH_CUDA)
3838
std::vector<int64_t> input_dim = common::vectorize(input.dims());
3939
std::vector<int64_t> x_dim = common::vectorize(x.dims());
4040
std::vector<int64_t> y_dim = common::vectorize(y.dims());
@@ -95,10 +95,6 @@ void AddmmKernelImpl(const Context& dev_ctx,
9595
auto sparse_blas = funcs::sparse::GetSparseBlas<Context, T>(dev_ctx);
9696
sparse_blas.SPMM(
9797
false, false, static_cast<T>(alpha), x, y, static_cast<T>(beta), out);
98-
#else
99-
PADDLE_THROW(common::errors::Unimplemented(
100-
"forward of 'sparse.addmm' use cusparseSpMM, "
101-
"which is supported from CUDA 11.0"));
10298
#endif
10399
}
104100

‎paddle/phi/kernels/sparse/gpu/matmul_grad_kernel.cu‎

Lines changed: 4 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -37,7 +37,7 @@ void MatmulCooDenseGradKernel(const Context& dev_ctx,
3737
const DenseTensor& dout,
3838
SparseCooTensor* dx,
3939
DenseTensor* dy) {
40-
#if CUDA_VERSION >= 11030 || HIP_VERSION >= 403
40+
#if defined(PADDLE_WITH_CUDA) || HIP_VERSION >= 403
4141
auto sparse_blas = funcs::sparse::GetSparseBlas<Context, T>(dev_ctx);
4242

4343
// dx{SparseCoo} = dout{Dense} * y'{Dense}
@@ -84,7 +84,7 @@ void MatmulCsrDenseGradKernel(const Context& dev_ctx,
8484
const DenseTensor& dout,
8585
SparseCsrTensor* dx,
8686
DenseTensor* dy) {
87-
#if CUDA_VERSION >= 11030 || HIP_VERSION >= 403
87+
#if defined(PADDLE_WITH_CUDA) || HIP_VERSION >= 403
8888
auto sparse_blas = funcs::sparse::GetSparseBlas<Context, T>(dev_ctx);
8989

9090
// dx{SparseCsr} = dout{Dense} * y'{Dense}
@@ -123,7 +123,7 @@ void MatmulCsrCsrGradKernel(const Context& dev_ctx,
123123
const SparseCsrTensor& dout,
124124
SparseCsrTensor* dx,
125125
SparseCsrTensor* dy) {
126-
#if CUDA_VERSION >= 11000
126+
#if defined(PADDLE_WITH_CUDA)
127127
auto sparse_blas = funcs::sparse::GetSparseBlas<Context, T>(dev_ctx);
128128

129129
std::vector<int64_t> xdim_vec = phi::vectorize(x.dims());
@@ -187,7 +187,7 @@ void MaskedMatmulCsrGradKernel(const Context& dev_ctx,
187187
const SparseCsrTensor& dout,
188188
DenseTensor* dx,
189189
DenseTensor* dy) {
190-
#if CUDA_VERSION >= 11000
190+
#if defined(PADDLE_WITH_CUDA)
191191
auto sparse_blas = funcs::sparse::GetSparseBlas<Context, T>(dev_ctx);
192192

193193
// dx{Dense} = dout{SparseCsr} * y'{Dense}

‎paddle/phi/kernels/sparse/gpu/matmul_kernel.cu‎

Lines changed: 3 additions & 14 deletions
Original file line numberDiff line numberDiff line change
@@ -39,7 +39,7 @@ void MatmulKernelImpl(const Context& dev_ctx,
3939
const TensorType& x,
4040
const DenseTensor& y,
4141
DenseTensor* out) {
42-
#if CUDA_VERSION >= 11000 || HIP_VERSION >= 402
42+
#if defined(PADDLE_WITH_CUDA) || HIP_VERSION >= 402
4343
std::vector<int64_t> xdim_vec = common::vectorize(x.dims());
4444
std::vector<int64_t> ydim_vec = common::vectorize(y.dims());
4545
auto x_ndims = xdim_vec.size();
@@ -115,7 +115,7 @@ void MatmulCsrCsrKernel(const Context& dev_ctx,
115115
const SparseCsrTensor& x,
116116
const SparseCsrTensor& y,
117117
SparseCsrTensor* out) {
118-
#if CUDA_VERSION >= 11000
118+
#if defined(PADDLE_WITH_CUDA)
119119
std::vector<int64_t> xdim_vec = phi::vectorize(x.dims());
120120
std::vector<int64_t> ydim_vec = phi::vectorize(y.dims());
121121
auto x_ndims = xdim_vec.size();
@@ -152,13 +152,6 @@ void MatmulCsrCsrKernel(const Context& dev_ctx,
152152
auto sparse_blas = funcs::sparse::GetSparseBlas<Context, T>(dev_ctx);
153153
sparse_blas.SPGEMM(
154154
false, false, static_cast<T>(1), x, y, static_cast<T>(0), out);
155-
156-
#else
157-
#ifdef PADDLE_WITH_CUDA
158-
PADDLE_THROW(common::errors::Unimplemented(
159-
"forward of 'sparse.matmul' use cusparseSpGEMM, "
160-
"which is supported from CUDA 11.0"));
161-
#endif
162155
#endif
163156
}
164157

@@ -182,7 +175,7 @@ void MaskedMatmulCsrKernel(const Context& dev_ctx,
182175
const DenseTensor& y,
183176
const SparseCsrTensor& mask,
184177
SparseCsrTensor* out) {
185-
#if CUDA_VERSION >= 11030
178+
#if defined(PADDLE_WITH_CUDA)
186179
std::vector<int64_t> xdim_vec = common::vectorize(x.dims());
187180
std::vector<int64_t> ydim_vec = common::vectorize(y.dims());
188181
std::vector<int64_t> maskdim_vec = common::vectorize(mask.dims());
@@ -252,10 +245,6 @@ void MaskedMatmulCsrKernel(const Context& dev_ctx,
252245
auto sparse_blas = funcs::sparse::GetSparseBlas<Context, T>(dev_ctx);
253246
sparse_blas.SDDMM(
254247
false, false, static_cast<T>(1), x, y, static_cast<T>(0), out);
255-
#else
256-
PADDLE_THROW(common::errors::Unimplemented(
257-
"forward of 'sparse.masked_matmul' use cusparseSDDMM, which is supported "
258-
"from CUDA 11.3"));
259248
#endif
260249
}
261250

‎paddle/phi/kernels/sparse/gpu/mv_grad_kernel.cu‎

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -86,7 +86,7 @@ void MvCooGradKernel(const Context &dev_ctx,
8686

8787
// dvec{Dense} = x'{SparseCoo} * dout{Dense}
8888
if (dvec) {
89-
#if CUDA_VERSION >= 11000
89+
#if defined(PADDLE_WITH_CUDA)
9090
// InferMeta of DenseTensor 'dvec'
9191
dvec->Resize(vec.dims());
9292
dev_ctx.template Alloc<T>(dvec);
@@ -130,7 +130,7 @@ void MvCsrGradKernel(const Context &dev_ctx,
130130

131131
// dvec{Dense} = x'{SparseCsr} * dout{Dense}
132132
if (dvec) {
133-
#if CUDA_VERSION >= 11000
133+
#if defined(PADDLE_WITH_CUDA)
134134
// InferMeta of DenseTensor 'dvec'
135135
dvec->Resize(vec.dims());
136136
dev_ctx.template Alloc<T>(dvec);

‎paddle/phi/kernels/sparse/gpu/mv_kernel.cu‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -29,7 +29,7 @@ void MvKernelImpl(const Context& dev_ctx,
2929
const TensorType& x,
3030
const DenseTensor& vec,
3131
DenseTensor* out) {
32-
#if CUDA_VERSION >= 11000
32+
#if defined(PADDLE_WITH_CUDA)
3333
std::vector<int64_t> x_dim = common::vectorize(x.dims());
3434
std::vector<int64_t> vec_dim = common::vectorize(vec.dims());
3535
auto x_ndims = x_dim.size();

‎paddle/phi/kernels/sparse/gpu/sparse_attention_kernel.cu‎

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -460,7 +460,7 @@ void DotSdd(const phi::GPUContext& dev_ctx,
460460
const int num_cols,
461461
const bool a_transpose,
462462
const bool b_transpose) {
463-
#if defined(PADDLE_WITH_CUDA) && CUDA_VERSION >= 11030
463+
#if defined(PADDLE_WITH_CUDA)
464464
const T* a_data = a->data<T>();
465465
const T* b_data = b->data<T>();
466466
const int* c_offset_data = c_offset->data<int>();
@@ -560,7 +560,7 @@ void DotDsd(const phi::GPUContext& dev_ctx,
560560
const int num_cols,
561561
const bool a_transpose,
562562
const bool b_transpose) {
563-
#if defined(PADDLE_WITH_CUDA) && CUDA_VERSION >= 11000
563+
#if defined(PADDLE_WITH_CUDA)
564564
const int* a_offset_data = a_offset->data<int>();
565565
const int* a_columns_data = a_columns->data<int>();
566566
const T* a_value_data = a_value->data<T>();

0 commit comments

Comments
 (0)