Skip to content

[NPU] Improve performance #1431

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

Open
wants to merge 1 commit into
base: develop
Choose a base branch
from
Open

[NPU] Improve performance #1431

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
12 changes: 8 additions & 4 deletions backends/npu/kernels/sigmoid_cross_entropy_with_logits_kernel.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -51,12 +51,14 @@ void SigmoidCrossEntropyWithLogitsKernel(
phi::DenseTensor pos_weight_tensor;
phi::DenseTensorMeta weight_tensor_meta = {phi::DataType::FLOAT32, x.dims()};
weight_tensor.set_meta(weight_tensor_meta);
FillNpuTensorWithConstant<float>(&weight_tensor, dev_ctx, 1.0);
dev_ctx.template Alloc<float>(&weight_tensor);
EXEC_NPU_CMD(aclnnInplaceOne, dev_ctx, weight_tensor);
weight_tensor.Resize(x.dims());

if (pos_weight.get_ptr() == nullptr) {
pos_weight_tensor.set_meta(weight_tensor_meta);
FillNpuTensorWithConstant<float>(&pos_weight_tensor, dev_ctx, 1.0);
dev_ctx.template Alloc<float>(&pos_weight_tensor);
EXEC_NPU_CMD(aclnnInplaceOne, dev_ctx, pos_weight_tensor);
pos_weight_tensor.Resize(x.dims());
} else {
pos_weight_tensor = *pos_weight.get_ptr();
Expand Down Expand Up @@ -89,12 +91,14 @@ void SigmoidCrossEntropyWithLogitsGradKernel(
phi::DenseTensor pos_weight_tensor;
phi::DenseTensorMeta weight_tensor_meta = {phi::DataType::FLOAT32, x.dims()};
weight_tensor.set_meta(weight_tensor_meta);
FillNpuTensorWithConstant<float>(&weight_tensor, dev_ctx, 1.0);
dev_ctx.template Alloc<float>(&weight_tensor);
EXEC_NPU_CMD(aclnnInplaceOne, dev_ctx, weight_tensor);
weight_tensor.Resize(x.dims());

if (pos_weight.get_ptr() == nullptr) {
pos_weight_tensor.set_meta(weight_tensor_meta);
FillNpuTensorWithConstant<float>(&pos_weight_tensor, dev_ctx, 1.0);
dev_ctx.template Alloc<float>(&pos_weight_tensor);
EXEC_NPU_CMD(aclnnInplaceOne, dev_ctx, pos_weight_tensor);
pos_weight_tensor.Resize(x.dims());
} else {
pos_weight_tensor = *pos_weight.get_ptr();
Expand Down
48 changes: 37 additions & 11 deletions backends/npu/kernels/uniform_kernel.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -31,16 +31,16 @@ inline void UniformRealDistribution(T* data,
}

template <typename T, typename Context>
void UniformRawKernel(const Context& dev_ctx,
const phi::IntArray& shape,
phi::DataType dtype,
const phi::Scalar& min,
const phi::Scalar& max,
int seed,
int diag_num,
int diag_step,
float diag_val,
phi::DenseTensor* out) {
void UniformRawKernelCPU(const Context& dev_ctx,
const phi::IntArray& shape,
phi::DataType dtype,
const phi::Scalar& min,
const phi::Scalar& max,
int seed,
int diag_num,
int diag_step,
float diag_val,
phi::DenseTensor* out) {
out->Resize(phi::make_ddim(shape.GetData()));
VLOG(4) << out->dims();
T* data = dev_ctx.template Alloc<T>(out);
Expand Down Expand Up @@ -83,6 +83,32 @@ void UniformRawKernel(const Context& dev_ctx,
TensorCopy(dev_ctx, cpu_out, true, out);
}

template <typename T, typename Context>
void UniformRawKernel(const Context& dev_ctx,
const phi::IntArray& shape,
phi::DataType dtype,
const phi::Scalar& min,
const phi::Scalar& max,
int seed,
int diag_num,
int diag_step,
float diag_val,
phi::DenseTensor* out) {
DO_COMPATIBILITY(
aclnnInplaceUniform,
(custom_kernel::UniformRawKernelCPU<T, Context>(
dev_ctx, shape, dtype, min, max, seed, 0, 0, 0.0f, out)));

out->Resize(phi::make_ddim(shape.GetData()));
dev_ctx.template Alloc<T>(out);
double from = min.to<double>();
double to = max.to<double>();
uint64_t offset_int64 = 0;
uint64_t seed_int64 = static_cast<uint64_t>(seed);
EXEC_NPU_CMD(
aclnnInplaceUniform, dev_ctx, *out, from, to, seed_int64, offset_int64);
}

template <typename T, typename Context>
void UniformKernel(const Context& dev_ctx,
const phi::IntArray& shape,
Expand All @@ -91,7 +117,7 @@ void UniformKernel(const Context& dev_ctx,
const phi::Scalar& max,
int seed,
phi::DenseTensor* out) {
custom_kernel::UniformRawKernel<T>(
custom_kernel::UniformRawKernel<T, Context>(
dev_ctx, shape, dtype, min, max, seed, 0, 0, 0.0f, out);
}

Expand Down