Skip to content

[Paddle Inference] Support depthwise_conv2d fp16. #44642

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
merged 4 commits into from
Jul 28, 2022
Merged

[Paddle Inference] Support depthwise_conv2d fp16. #44642

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
79343f5
depthwise_fp16
xiaoxiaohehe001 Jul 27, 2022
45fcbf7
depthwise_fp16
xiaoxiaohehe001 Jul 27, 2022
990a388
depthwise_fp16
xiaoxiaohehe001 Jul 27, 2022
af6af83
depthwise_fp16
xiaoxiaohehe001 Jul 28, 2022
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
59 changes: 37 additions & 22 deletions paddle/phi/kernels/gpu/depthwise_conv.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -153,7 +153,7 @@ __device__ __inline__ void KernelDepthwiseConvNCHW(

const int c_in = c_out / filter_multiplier;
const T* weight = filter_data + c_out * filter_height * filter_width;
T value = 0;
T value(0);
const int h_in_start = -padding_height + h_out * stride_height;
const int w_in_start = -padding_width + w_out * stride_width;
const int h_in_end = h_in_start + filter_height * dilate_height;
Expand All @@ -176,7 +176,7 @@ __device__ __inline__ void KernelDepthwiseConvNCHW(
int offset = in_offset + h_in * input_width + w_in;
T in_data = input_data[offset];
if (fuse_relu_before_conv) {
value += weight[weight_offset] * max(0.0f, in_data);
value += weight[weight_offset] * T(max(0.0f, double(in_data)));
} else {
value += weight[weight_offset] * in_data;
}
Expand Down Expand Up @@ -205,7 +205,7 @@ __device__ __inline__ void KernelDepthwiseConvNHWC(
const int batch = idx / output_width / output_height / output_channels;

const int c_in = c_out / filter_multiplier;
T value = 0;
T value(0);
const int h_in_start = -padding_height + h_out * stride_height;
const int w_in_start = -padding_width + w_out * stride_width;
const int h_in_end = h_in_start + filter_height * dilate_height;
Expand All @@ -228,7 +228,7 @@ __device__ __inline__ void KernelDepthwiseConvNHWC(
T in_data = input_data[offset];
const T* weight = filter_data + weight_offset * output_channels + c_out;
if (fuse_relu_before_conv) {
value += weight[0] * max(0.0f, in_data);
value += weight[0] * T(max(0.0f, double(in_data)));
} else {
value += weight[0] * in_data;
}
Expand Down Expand Up @@ -258,7 +258,7 @@ __device__ __inline__ void KernelDepthwiseConvCFilterNCHW(
const int c_out = blockIdx.x;

const int c_in = c_out / filter_multiplier;
T value = 0;
T value(0);
const int h_in_start = -padding_height + h_out * stride_height;
const int w_in_start = -padding_width + w_out * stride_width;
const int h_in_end = h_in_start + c_filter * dilate_height;
Expand All @@ -281,7 +281,7 @@ __device__ __inline__ void KernelDepthwiseConvCFilterNCHW(
int offset = in_offset + h_in * input_width + w_in;
if (fuse_relu_before_conv) {
value += r_weight[h_f * c_filter + w_f] *
max(0.0f, input_data[offset]);
T(max(0.0f, double(input_data[offset])));
} else {
value += r_weight[h_f * c_filter + w_f] * input_data[offset];
}
Expand Down Expand Up @@ -325,7 +325,7 @@ __device__ __inline__ void KernelDepthwiseConvCFilterNHWC(
if (w_out >= output_width) {
continue;
}
T value = 0;
T value(0);
const int w_in_start = -padding_width + w_out * stride_width;
for (int h_in = h_in_start, h_f = 0; h_f < c_filter;
h_in += dilate_height, h_f++) {
Expand All @@ -337,7 +337,7 @@ __device__ __inline__ void KernelDepthwiseConvCFilterNHWC(
in_offset + (h_in * input_width + w_in) * input_channels + c_in;
if (fuse_relu_before_conv) {
value += r_weight[h_f * c_filter + w_f] *
max(0.0f, input_data[offset]);
T(max(0.0, double(input_data[offset])));
} else {
value += r_weight[h_f * c_filter + w_f] * input_data[offset];
}
Expand Down Expand Up @@ -482,13 +482,13 @@ __device__ __inline__ void KernelDepthwiseConvInputGradNCHW(
w_in - (filter_width - 1) * dilate_width + padding_width;
int w_out_end = w_in + padding_width;

T value = 0;
T value(0);
int index =
((batch * gridDim.x + c_in) * input_height + h_in) * input_width +
w_in;

if (fuse_relu_before_conv) {
if (input_data[index] <= 0) {
if (input_data[index] <= T(0)) {
input_grad_data[index] = 0;
continue;
}
Expand Down Expand Up @@ -539,12 +539,12 @@ __device__ __inline__ void KernelDepthwiseConvInputGradNHWC(
int w_out_start =
w_in - (filter_width - 1) * dilate_width + padding_width;

T value = 0;
T value(0);
int index = ((batch * input_height + h_in) * input_width + w_in) *
input_channels +
c_in;
if (fuse_relu_before_conv) {
if (input_data[index] <= 0) {
if (input_data[index] <= T(0)) {
input_grad_data[index] = 0;
continue;
}
Expand Down Expand Up @@ -603,12 +603,12 @@ __device__ __inline__ void KernelDepthwiseConvInputGradCFilterNCHW(
int h_out_start = h_in - (c_filter - 1) * dilate_height + padding_height;
int w_out_start = w_in - (c_filter - 1) * dilate_width + padding_width;

T value = 0;
T value(0);
int index =
((batch * gridDim.x + c_in) * input_height + h_in) * input_width +
w_in;
if (fuse_relu_before_conv) {
if (input_data[index] <= 0) {
if (input_data[index] <= T(0)) {
input_grad_data[index] = 0;
continue;
}
Expand Down Expand Up @@ -676,12 +676,12 @@ __device__ __inline__ void KernelDepthwiseConvInputGradCFilterNHWC(
}
int w_out_start = w_in - (c_filter - 1) * dilate_width + padding_width;

T value = 0;
T value(0);
int index = ((batch * input_height + h_in) * input_width + w_in) *
input_channels +
c_in;
if (fuse_relu_before_conv) {
if (input_data[index] <= 0) {
if (input_data[index] <= T(0)) {
input_grad_data[index] = 0;
continue;
}
Expand Down Expand Up @@ -854,7 +854,7 @@ __device__ __inline__ void KernelDepthwiseConvFilterGradNCHW(
const int dilate_height,
const int dilate_width,
T* filter_grad_data) {
T s = 0;
T s(0);
int gbid = ((blockIdx.z * gridDim.y) + blockIdx.y) * gridDim.x + blockIdx.x;

for (int image_w = threadIdx.x; image_w < output_width;
Expand All @@ -880,7 +880,7 @@ __device__ __inline__ void KernelDepthwiseConvFilterGradNCHW(
image_wk;
if (fuse_relu_before_conv) {
s += output_grad_data[gaid(bid, kernel_id, image_h, image_w)] *
max(0.0f, input_data[input_id]);
T(max(0.0f, double(input_data[input_id])));
} else {
s += output_grad_data[gaid(bid, kernel_id, image_h, image_w)] *
input_data[input_id];
Expand Down Expand Up @@ -921,7 +921,7 @@ __device__ __inline__ void KernelDepthwiseConvFilterGradNHWC(
int kernel_ih = blockIdx.x / filter_width;
for (int kernel_id = threadIdx.x; kernel_id < output_channels;
kernel_id += blockDim.x) {
T s = 0;
T s(0);
int gbid =
((kernel_id * filter_height) + kernel_ih) * filter_width + kernel_iw;
for (int image_w = threadIdx.y; image_w < output_width;
Expand All @@ -941,7 +941,7 @@ __device__ __inline__ void KernelDepthwiseConvFilterGradNHWC(
kernel_id / filter_multiplier;
if (fuse_relu_before_conv) {
s += output_grad_data[gaid(bid, image_h, image_w, kernel_id)] *
max(0.0f, input_data[input_id]);
T(max(0.0f, double(input_data[input_id])));
} else {
s += output_grad_data[gaid(bid, image_h, image_w, kernel_id)] *
input_data[input_id];
Expand Down Expand Up @@ -1010,9 +1010,10 @@ __device__ __inline__ void KernelDepthwiseConvFilterGradCFilterNHWC(
((bid * output_height + image_h) * output_width + image_w) *
output_channels +
kernel_id;
T s = 0;
T s(0);
if (fuse_relu_before_conv) {
s = output_grad_data[output_id] * max(0.0f, input_data[input_id]);
s = output_grad_data[output_id] *
T(max(0.0f, double(input_data[input_id])));
} else {
s = output_grad_data[output_id] * input_data[input_id];
}
Expand Down Expand Up @@ -1672,21 +1673,35 @@ class DepthwiseConvFilterGradFunctor<phi::GPUContext,

template class DepthwiseConvFunctor<phi::GPUContext, float, false>;
template class DepthwiseConvFunctor<phi::GPUContext, double, false>;
template class DepthwiseConvFunctor<phi::GPUContext, platform::float16, false>;

template class DepthwiseConvInputGradFunctor<phi::GPUContext, float, false>;
template class DepthwiseConvInputGradFunctor<phi::GPUContext, double, false>;
template class DepthwiseConvInputGradFunctor<phi::GPUContext,
platform::float16,
false>;

template class DepthwiseConvFilterGradFunctor<phi::GPUContext, float, false>;
template class DepthwiseConvFilterGradFunctor<phi::GPUContext, double, false>;
template class DepthwiseConvFilterGradFunctor<phi::GPUContext,
platform::float16,
false>;

template class DepthwiseConvFunctor<phi::GPUContext, float, true>;
template class DepthwiseConvFunctor<phi::GPUContext, double, true>;
template class DepthwiseConvFunctor<phi::GPUContext, platform::float16, true>;

template class DepthwiseConvInputGradFunctor<phi::GPUContext, float, true>;
template class DepthwiseConvInputGradFunctor<phi::GPUContext, double, true>;
template class DepthwiseConvInputGradFunctor<phi::GPUContext,
platform::float16,
true>;

template class DepthwiseConvFilterGradFunctor<phi::GPUContext, float, true>;
template class DepthwiseConvFilterGradFunctor<phi::GPUContext, double, true>;
template class DepthwiseConvFilterGradFunctor<phi::GPUContext,
platform::float16,
true>;

} // namespace math
} // namespace operators
Expand Down
3 changes: 2 additions & 1 deletion paddle/phi/kernels/gpu/depthwise_conv_grad_kernel.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -139,4 +139,5 @@ PD_REGISTER_KERNEL(depthwise_conv2d_grad,
ALL_LAYOUT,
phi::DepthwiseConvGradKernel,
float,
double) {}
double,
phi::dtype::float16) {}
3 changes: 2 additions & 1 deletion paddle/phi/kernels/gpu/depthwise_conv_kernel.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -124,4 +124,5 @@ PD_REGISTER_KERNEL(depthwise_conv2d,
ALL_LAYOUT,
phi::DepthwiseConvKernel,
float,
double) {}
double,
phi::dtype::float16) {}