Skip to content
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.

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

fix the error message for the math dir, test=develop #27332

Merged
merged 2 commits into from
Sep 18, 2020
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
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
5 changes: 4 additions & 1 deletion paddle/fluid/operators/math/beam_search.cc
Original file line number Diff line number Diff line change
Expand Up @@ -87,7 +87,10 @@ class BeamSearchFunctor<platform::CPUDeviceContext, T> {
lod[0].assign(high_level.begin(), high_level.end());
lod[1].assign(low_level.begin(), low_level.end());
if (!framework::CheckLoD(lod)) {
PADDLE_THROW("lod %s is not right", framework::LoDToString(lod));
PADDLE_THROW(platform::errors::InvalidArgument(
"lod %s is not right in"
" beam_search, please check your code.",
framework::LoDToString(lod)));
}
selected_ids->set_lod(lod);
selected_scores->set_lod(lod);
Expand Down
5 changes: 4 additions & 1 deletion paddle/fluid/operators/math/beam_search.cu
Original file line number Diff line number Diff line change
Expand Up @@ -400,7 +400,10 @@ class BeamSearchFunctor<platform::CUDADeviceContext, T> {

context.Wait();
if (!framework::CheckLoD(selected_lod)) {
PADDLE_THROW("lod %s is not right", framework::LoDToString(selected_lod));
PADDLE_THROW(platform::errors::InvalidArgument(
"lod %s is not right in"
" beam_search, please check your code.",
framework::LoDToString(selected_lod)));
}

selected_ids->set_lod(selected_lod);
Expand Down
6 changes: 5 additions & 1 deletion paddle/fluid/operators/math/blas.cc
Original file line number Diff line number Diff line change
Expand Up @@ -20,7 +20,11 @@ namespace operators {
namespace math {
MatDescriptor CreateMatrixDescriptor(const framework::DDim &tensor_dim,
int num_flatten_cols, bool trans) {
PADDLE_ENFORCE_GT(tensor_dim.size(), 1);
PADDLE_ENFORCE_GT(
tensor_dim.size(), 1,
platform::errors::InvalidArgument("The tensor dim size should be greater "
"than 1, but reveived dim size is %d",
tensor_dim.size()));
MatDescriptor retv;
if (num_flatten_cols > 1) {
auto flatten_dim = framework::flatten_to_2d(tensor_dim, num_flatten_cols);
Expand Down
26 changes: 18 additions & 8 deletions paddle/fluid/operators/math/blas_impl.cu.h
Original file line number Diff line number Diff line change
Expand Up @@ -60,7 +60,8 @@ struct CUBlas<float> {
PADDLE_ENFORCE_CUDA_SUCCESS(
platform::dynload::cublasSgemmStridedBatched(args...));
#else
PADDLE_THROW("SgemmStridedBatched is not supported on cuda <= 7.5");
PADDLE_THROW(platform::errors::Unimplemented(
"SgemmStridedBatched is not supported on cuda <= 7.5"));
#endif
}

Expand All @@ -85,7 +86,8 @@ struct CUBlas<float> {
beta, C, Ctype, ldc));
});
#else
PADDLE_THROW("cublasSgemmEx is supported on cuda >= 8.0");
PADDLE_THROW(platform::errors::Unimplemented(
"cublasSgemmEx is not supported on cuda <= 7.5"));
#endif
}

Expand Down Expand Up @@ -146,13 +148,15 @@ struct CUBlas<double> {
PADDLE_ENFORCE_CUDA_SUCCESS(
platform::dynload::cublasDgemmStridedBatched(args...));
#else
PADDLE_THROW("DgemmStridedBatched is not supported on cuda <= 7.5");
PADDLE_THROW(platform::errors::Unimplemented(
"DgemmStridedBatched is not supported on cuda <= 7.5"));
#endif
}

template <typename... ARGS>
static void GEMM_EX(ARGS... args) {
PADDLE_THROW("Currently there are not cublasDgemmEx.");
PADDLE_THROW(platform::errors::Unimplemented(
"Currently there are not cublasDgemmEx."));
}

template <typename... ARGS>
Expand Down Expand Up @@ -216,7 +220,8 @@ struct CUBlas<platform::float16> {
reinterpret_cast<const __half *>(beta), reinterpret_cast<__half *>(C),
ldc, strideC, batchCount));
#else
PADDLE_THROW("HgemmStridedBatched is not supported on cuda <= 7.5");
PADDLE_THROW(platform::errors::Unimplemented(
"HgemmStridedBatched is not supported on cuda <= 7.5"));
#endif
}

Expand Down Expand Up @@ -247,7 +252,8 @@ struct CUBlas<platform::float16> {
beta, C, Ctype, ldc, computeType, algo));
});
#else
PADDLE_THROW("cublasGemmEx is supported on cuda >= 8.0");
PADDLE_THROW(platform::errors::Unimplemented(
"cublasGemmEx is not supported on cuda <= 7.5"));
#endif
}
};
Expand Down Expand Up @@ -302,8 +308,12 @@ inline void Blas<platform::CUDADeviceContext>::GEMM(
(transB == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T;

// TODO(kexinzhao): add processing code for compute capability < 53 case
PADDLE_ENFORCE_GE(context_.GetComputeCapability(), 53,
"cublas fp16 gemm requires GPU compute capability >= 53");
PADDLE_ENFORCE_GE(
context_.GetComputeCapability(), 53,
platform::errors::InvalidArgument(
"cublas fp16 gemm requires GPU compute capability >= 53,"
"but received %d",
context_.GetComputeCapability()));

float h_alpha = static_cast<float>(alpha);
float h_beta = static_cast<float>(beta);
Expand Down
148 changes: 115 additions & 33 deletions paddle/fluid/operators/math/blas_impl.h
Original file line number Diff line number Diff line change
Expand Up @@ -29,7 +29,8 @@ template <>
struct CBlas<int8_t> {
template <typename... ARGS>
static void VCOPY(ARGS... args) {
PADDLE_THROW("Blas VCOPY don't support int8_t");
PADDLE_THROW(platform::errors::Unimplemented(
"Blas VCOPY do not supported on CPU, please check your code"));
}
};

Expand Down Expand Up @@ -347,22 +348,47 @@ struct CBlas<double> {

template <>
struct CBlas<platform::float16> {
static void GEMM(...) { PADDLE_THROW("float16 GEMM not supported on CPU"); }
static void GEMM(...) {
PADDLE_THROW(platform::errors::Unimplemented(
"float16 GEMM not supported on CPU, please check your code"));
}

static void SMM_GEMM(...) {
PADDLE_THROW("float16 SMM_GEMM not supported on CPU");
PADDLE_THROW(platform::errors::Unimplemented(
"float16 SMM_GEMM not supported on CPU, please check your code"));
}
static void VMUL(...) { PADDLE_THROW("float16 VMUL not supported on CPU"); }
static void VEXP(...) { PADDLE_THROW("float16 VEXP not supported on CPU"); }
static void VSQUARE(...) {
PADDLE_THROW("float16 VSQUARE not supported on CPU");
static void VMUL(...) {
PADDLE_THROW(platform::errors::Unimplemented(
"float16 VMUL not supported on CPU, please check your code"));
}
static void VPOW(...) { PADDLE_THROW("float16 VPOW not supported on CPU"); }
static void DOT(...) { PADDLE_THROW("float16 DOT not supported on CPU"); };
static void SCAL(...) { PADDLE_THROW("float16 SCAL not supported on CPU"); };
static void ASUM(...) { PADDLE_THROW("float16 ASUM not supported on CPU"); };
static void VEXP(...) {
PADDLE_THROW(platform::errors::Unimplemented(
"float16 VEXP not supported on CPU, please check your code"));
}
static void VSQUARE(...) {
PADDLE_THROW(platform::errors::Unimplemented(
"float16 VSQUARE not supported on CPU, please check your code"));
}
static void VPOW(...) {
PADDLE_THROW(platform::errors::Unimplemented(
"float16 VPOW not supported on CPU, please check your code"));
}
static void DOT(...) {
PADDLE_THROW(platform::errors::Unimplemented(
"float16 DOT not supported on CPU, please check your code"));
};
static void SCAL(...) {
PADDLE_THROW(platform::errors::Unimplemented(
"float16 SCAL not supported on CPU, please check your code"));
};
static void ASUM(...) {
PADDLE_THROW(platform::errors::Unimplemented(
"float16 ASUM not supported on CPU, please check your code"));
};
#ifdef PADDLE_WITH_MKLML
static void GEMM_BATCH(...) {
PADDLE_THROW("float16 GEMM_BATCH not supported on CPU");
PADDLE_THROW(platform::errors::Unimplemented(
"float16 GEMM_BATCH not supported on CPU, please check your code"));
}
#endif
};
Expand Down Expand Up @@ -446,11 +472,18 @@ void Blas<DeviceContext>::MatMul(const framework::Tensor &mat_a, bool trans_a,
auto dim_a = mat_a.dims();
auto dim_b = mat_b.dims();
auto dim_out = mat_out->dims();
PADDLE_ENFORCE(dim_a.size() == 2 && dim_b.size() == 2 && dim_out.size() == 2,
"The input and output of matmul be matrix");
PADDLE_ENFORCE(
mat_a.place() == mat_b.place() && mat_a.place() == mat_out->place(),
"The places of matrices must be same");
PADDLE_ENFORCE_EQ(
dim_a.size() == 2 && dim_b.size() == 2 && dim_out.size() == 2, true,
platform::errors::InvalidArgument(
"The input and output of matmul should be matrix, the dim size must "
"be 2,"
"but received dim size input_a:%d, input_b:%d, output:%d",
dim_a.size(), dim_b.size(), dim_out.size()));
PADDLE_ENFORCE_EQ(
mat_a.place() == mat_b.place() && mat_a.place() == mat_out->place(), true,
platform::errors::InvalidArgument("The places of matrices in the matmul "
"should be same, please check your "
"code."));

int M = dim_out[0];
int N = dim_out[1];
Expand Down Expand Up @@ -715,7 +748,13 @@ void Blas<platform::CPUDeviceContext>::BatchedGEMMWithHead(
}

} else {
PADDLE_ENFORCE_EQ(W1, H2);
PADDLE_ENFORCE_EQ(
W1, H2,
platform::errors::InvalidArgument(
"The fisrt matrix width should be same as second matrix height,"
"but received fisrt matrix width %d"
", second matrix height %d",
W1, H2));
int ldc = W2 * head_number;
int sub_width = W1 / head_number;

Expand Down Expand Up @@ -785,20 +824,29 @@ void Blas<DeviceContext>::MatMul(const framework::Tensor &mat_a,
const framework::Tensor &mat_b,
const MatDescriptor &dim_b, T alpha,
framework::Tensor *mat_out, T beta) const {
PADDLE_ENFORCE_EQ(dim_a.width_, dim_b.height_);
PADDLE_ENFORCE_EQ(
dim_a.width_, dim_b.height_,
platform::errors::InvalidArgument(
"The fisrt matrix width should be same as second matrix height,"
"but received fisrt matrix width %d"
", second matrix height %d",
dim_a.width_, dim_b.height_));

CBLAS_TRANSPOSE transA = !dim_a.trans_ ? CblasNoTrans : CblasTrans;
CBLAS_TRANSPOSE transB = !dim_b.trans_ ? CblasNoTrans : CblasTrans;
if (dim_a.batch_size_ == 0 && dim_b.batch_size_ == 0) {
this->template GEMM<T>(transA, transB, dim_a.height_, dim_b.width_,
dim_a.width_, alpha, mat_a.data<T>(),
mat_b.data<T>(), beta, mat_out->data<T>());
} else {
PADDLE_ENFORCE(dim_a.batch_size_ == dim_b.batch_size_ ||
dim_a.batch_size_ == 0 || dim_b.batch_size_ == 0,
"dim_a.batch_size should be equal to dim_b.batch_size, or "
"one of dim_a.batch_size and dim_b.batch_size should be 0. "
"But got dim_a.batch_size = %d, dim_b.batch_size = %d.",
dim_a.batch_size_, dim_b.batch_size_);
PADDLE_ENFORCE_EQ(
dim_a.batch_size_ == dim_b.batch_size_ || dim_a.batch_size_ == 0 ||
dim_b.batch_size_ == 0,
true, platform::errors::InvalidArgument(
"dim_a.batch_size should be equal to dim_b.batch_size, or "
"one of dim_a.batch_size and dim_b.batch_size should be 0. "
"But got dim_a.batch_size = %d, dim_b.batch_size = %d.",
dim_a.batch_size_, dim_b.batch_size_));
this->template BatchedGEMM<T>(
transA, transB, dim_a.height_, dim_b.width_, dim_a.width_, alpha,
mat_a.data<T>(), mat_b.data<T>(), beta, mat_out->data<T>(),
Expand Down Expand Up @@ -834,15 +882,42 @@ void Blas<DeviceContext>::MatMulWithHead(const framework::Tensor &mat_a,
int head_number,
framework::Tensor *mat_out, T beta,
bool mat_b_split_vertical) const {
PADDLE_ENFORCE_EQ(dim_a.width_ % head_number, 0);
PADDLE_ENFORCE_GE(head_number, 1);
PADDLE_ENFORCE_LE(head_number, dim_a.width_);
PADDLE_ENFORCE_EQ(
dim_a.width_ % head_number, 0,
platform::errors::InvalidArgument(
"The first input width must be some times the head number"
"but received first input width %d"
", head_number %d",
dim_a.width_, head_number));
PADDLE_ENFORCE_GE(head_number, 1,
platform::errors::InvalidArgument(
"The head number should be greater equal 1,"
"but received head number %d",
head_number));
PADDLE_ENFORCE_LE(
head_number, dim_a.width_,
platform::errors::InvalidArgument(
"The head number should be less equal first input width,"
"but received first input width %d"
", head_number %d",
dim_a.width_, head_number));
CBLAS_TRANSPOSE transA = !dim_a.trans_ ? CblasNoTrans : CblasTrans;
CBLAS_TRANSPOSE transB = !dim_b.trans_ ? CblasNoTrans : CblasTrans;

if (mat_b_split_vertical) {
PADDLE_ENFORCE_EQ(dim_b.height_, dim_a.width_ / head_number);
PADDLE_ENFORCE_EQ(dim_b.width_ % head_number, 0);
PADDLE_ENFORCE_EQ(
dim_b.height_, dim_a.width_ / head_number,
platform::errors::InvalidArgument(
"The second input height should be equal than first input width,"
"but received second input height %d, first input width %d",
dim_b.height_, dim_a.width_ / head_number));
PADDLE_ENFORCE_EQ(
dim_a.width_ % head_number, 0,
platform::errors::InvalidArgument(
"The second input width should be some times the head number"
"but received second input width %d"
", head_number %d",
dim_b.width_, head_number));
}

if (dim_a.batch_size_ == 0 && dim_b.batch_size_ == 0) {
Expand Down Expand Up @@ -888,9 +963,16 @@ void Blas<DeviceContext>::MatMulWithHead(const framework::Tensor &mat_a,
mat_out->data<T>() + sub_matC_offset, ldc);
}
} else {
PADDLE_ENFORCE_EQ((dim_a.batch_size_ == dim_b.batch_size_ ||
dim_a.batch_size_ == 0 || dim_b.batch_size_ == 0),
true);
PADDLE_ENFORCE_EQ(
(dim_a.batch_size_ == dim_b.batch_size_ || dim_a.batch_size_ == 0 ||
dim_b.batch_size_ == 0),
true,
platform::errors::InvalidArgument(
"The first input batch size should be equal than second input,"
"either two input batch size is 0, but received first input batch "
"size"
" %d, second input batch size %d",
dim_a.batch_size_, dim_b.batch_size_));

this->template BatchedGEMMWithHead<T>(
transA, transB, dim_a.width_, dim_a.height_, dim_b.width_,
Expand Down