Skip to content

GitLab

Commit dc536427 authored by carlushuang's avatar carlushuang
Browse files

add kyxck8 in bias_act_add example

parent ad0a4ce1
Hide whitespace changes
Inline Side-by-side
Showing with 1605 additions and 491 deletions
example/cpu_02_conv2d_fwd_bias_relu_add/cpu_conv2d_fwd_bias_relu_add.cpp
View file @ dc536427
#include <sstream> #include <sstream>
#include "config.hpp" #include "config.hpp"
#include "device.hpp" #include "device.hpp"
#include "host_tensor.hpp" #include "host_tensor.hpp"
#include "host_tensor_generator.hpp" #include "host_tensor_generator.hpp"
#include "tensor_layout.hpp" #include "tensor_layout.hpp"
#include "device_tensor.hpp" #include "device_tensor.hpp"
#include "device_convnd_fwd_bias_activation_add_avx2_nhwc_kyxc_nhwk.hpp" #include "device_convnd_fwd_bias_activation_add_avx2_nhwc_kyxc_nhwk.hpp"
#include "element_wise_operation_cpu.hpp" #include "element_wise_operation_cpu.hpp"
#include "reference_conv_fwd_bias_activation_add.hpp" #include "reference_conv_fwd_bias_activation_add.hpp"
#include "element_wise_operation_cpu.hpp" #include "element_wise_operation_cpu.hpp"
#include "dynamic_buffer_cpu.hpp" #include "dynamic_buffer_cpu.hpp"
#include <omp.h> #include <omp.h>
#define AVX2_DATA_ALIGNMENT 32 #define AVX2_DATA_ALIGNMENT 32
#define TEST_LAYOUT_NHWC_KYXC_NHWK 0 #define TEST_LAYOUT_NHWC_KYXC_NHWK 0
#define TEST_LAYOUT_NHWC_KYXCK8_NHWK 1 #define TEST_LAYOUT_NHWC_KYXCK8_NHWK 1
#define TEST_LAYOUT TEST_LAYOUT_NHWC_KYXC_NHWK #define TEST_LAYOUT TEST_LAYOUT_NHWC_KYXCK8_NHWK
using F32 = float; using F32 = float;
using F16 = ck::half_t; using F16 = ck::half_t;
namespace ck { namespace ck {
namespace tensor_operation { namespace tensor_operation {
namespace cpu { namespace cpu {
namespace device { namespace device {
namespace device_conv2d_fwd_bias_activation_add_avx2_instance { namespace device_conv2d_fwd_bias_activation_add_avx2_instance {
using PassThrough = ck::tensor_operation::cpu::element_wise::PassThrough; using PassThrough = ck::tensor_operation::cpu::element_wise::PassThrough;
using AddReluAdd = ck::tensor_operation::cpu::element_wise::AddReluAdd; using AddReluAdd = ck::tensor_operation::cpu::element_wise::AddReluAdd;
void add_device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxc_nhwk( void add_device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxc_nhwk(
std::vector<DeviceConvFwdBiasActivationAddPtr<PassThrough, PassThrough, AddReluAdd>>& std::vector<DeviceConvFwdBiasActivationAddPtr<PassThrough, PassThrough, AddReluAdd>>&
instances); instances);
void add_device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxc_nhwk_local_c( void add_device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxc_nhwk_local_c(
std::vector<DeviceConvFwdBiasActivationAddPtr<PassThrough, PassThrough, AddReluAdd>>& std::vector<DeviceConvFwdBiasActivationAddPtr<PassThrough, PassThrough, AddReluAdd>>&
instances); instances);
void add_device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxc_nhwk_mt( void add_device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxc_nhwk_mt(
std::vector<DeviceConvFwdBiasActivationAddPtr<PassThrough, PassThrough, AddReluAdd>>& std::vector<DeviceConvFwdBiasActivationAddPtr<PassThrough, PassThrough, AddReluAdd>>&
instances); instances);
} // namespace device_conv2d_fwd_bias_activation_add_avx2_instance void add_device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxck8_nhwk(
} // namespace device std::vector<DeviceConvFwdBiasActivationAddPtr<PassThrough, PassThrough, AddReluAdd>>&
} // namespace cpu instances);
} // namespace tensor_operation
} // namespace ck void add_device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxck8_nhwk_local_c(
std::vector<DeviceConvFwdBiasActivationAddPtr<PassThrough, PassThrough, AddReluAdd>>&
using InElementOp = ck::tensor_operation::cpu::element_wise::PassThrough; instances);
using WeiElementOp = ck::tensor_operation::cpu::element_wise::PassThrough;
using OutElementOp = ck::tensor_operation::cpu::element_wise::AddReluAdd; void add_device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxck8_nhwk_mt(
std::vector<DeviceConvFwdBiasActivationAddPtr<PassThrough, PassThrough, AddReluAdd>>&
template <typename T> instances);
static bool
check_out(const Tensor<T>& ref, const Tensor<T>& result, double nrms, int per_pixel_check = 0) } // namespace device_conv2d_fwd_bias_activation_add_avx2_instance
{ } // namespace device
int error_count = 0; } // namespace cpu
float max_diff = 1e-5; } // namespace tensor_operation
} // namespace ck
double square_difference = .0;
double mag1 = .0; using InElementOp = ck::tensor_operation::cpu::element_wise::PassThrough;
double mag2 = .0; using WeiElementOp = ck::tensor_operation::cpu::element_wise::PassThrough;
using OutElementOp = ck::tensor_operation::cpu::element_wise::AddReluAdd;
for(int i = 0; i < ref.mData.size(); ++i)
{ template <typename T>
double ri = (double)ref.mData[i]; static bool
double pi = (double)result.mData[i]; check_out(const Tensor<T>& ref, const Tensor<T>& result, double nrms, int per_pixel_check = 0)
double d = ri - pi; {
int error_count = 0;
if(per_pixel_check) float max_diff = 1e-5;
{
if(max_diff < std::abs(d)) double square_difference = .0;
{ double mag1 = .0;
error_count++; double mag2 = .0;
printf("idx:%3d, ref:%f, res:%f (diff:%f)\n",
i, for(int i = 0; i < ref.mData.size(); ++i)
double(ref.mData[i]), {
double(result.mData[i]), double ri = (double)ref.mData[i];
d); double pi = (double)result.mData[i];
} double d = ri - pi;
}
if(per_pixel_check)
square_difference += d * d; {
if(std::abs(mag1) < std::abs(ri)) if(max_diff < std::abs(d))
mag1 = ri; {
if(std::abs(mag2) < std::abs(pi)) error_count++;
mag2 = pi; printf("idx:%3d, ref:%f, res:%f (diff:%f)\n",
} i,
double(ref.mData[i]),
double mag = std::max({std::fabs(mag1), std::fabs(mag2), std::numeric_limits<double>::min()}); double(result.mData[i]),
double computed_nrms = std::sqrt(square_difference) / (std::sqrt(ref.mData.size()) * mag); d);
}
if(computed_nrms >= nrms) }
printf("nrms:%lf, mag1:%lf, mag2:%lf, expected_nrms is %1f\n",
computed_nrms, square_difference += d * d;
mag1, if(std::abs(mag1) < std::abs(ri))
mag2, mag1 = ri;
nrms); if(std::abs(mag2) < std::abs(pi))
mag2 = pi;
return computed_nrms < nrms && error_count == 0; }
}
double mag = std::max({std::fabs(mag1), std::fabs(mag2), std::numeric_limits<double>::min()});
float calculate_gflops() {} double computed_nrms = std::sqrt(square_difference) / (std::sqrt(ref.mData.size()) * mag);
template <typename T> if(computed_nrms >= nrms)
void transpose_kyxc_2_kyxc8k(Tensor<T>& dst, printf("nrms:%lf, mag1:%lf, mag2:%lf, expected_nrms is %1f\n",
const Tensor<T>& src, computed_nrms,
ck::index_t K, mag1,
ck::index_t Y, mag2,
ck::index_t X, nrms);
ck::index_t C)
{ return computed_nrms < nrms && error_count == 0;
ck::index_t batch = K / 8; }
ck::index_t row = 8;
ck::index_t col = C * Y * X; float calculate_gflops() {}
for(auto i_b = 0; i_b < batch; i_b++)
{ template <typename T>
for(auto i_r = 0; i_r < row; i_r++) void transpose_kyxc_2_kyxc8k(Tensor<T>& dst,
{ const Tensor<T>& src,
for(auto i_c = 0; i_c < col; i_c++) ck::index_t K,
{ ck::index_t Y,
ck::index_t src_idx = i_b * row * col + i_r * col + i_c; ck::index_t X,
ck::index_t dst_idx = i_b * col * row + i_c * row + i_r; ck::index_t C)
dst.mData[dst_idx] = src.mData[src_idx]; {
} ck::index_t batch = K / 8;
} ck::index_t row = 8;
} ck::index_t col = C * Y * X;
} for(auto i_b = 0; i_b < batch; i_b++)
{
int main(int argc, char* argv[]) for(auto i_r = 0; i_r < row; i_r++)
{ {
int data_type = 0; for(auto i_c = 0; i_c < col; i_c++)
int init_method = 0; {
ck::index_t src_idx = i_b * row * col + i_r * col + i_c;
// Conv shape ck::index_t dst_idx = i_b * col * row + i_c * row + i_r;
ck::index_t N = 2; dst.mData[dst_idx] = src.mData[src_idx];
ck::index_t K = 256; }
ck::index_t C = 192; }
ck::index_t Y = 3; }
ck::index_t X = 3; }
ck::index_t Hi = 71;
ck::index_t Wi = 71; int main(int argc, char* argv[])
ck::index_t conv_stride_h = 1; {
ck::index_t conv_stride_w = 1; int data_type = 0;
ck::index_t conv_dilation_h = 1; int init_method = 0;
ck::index_t conv_dilation_w = 1;
ck::index_t in_left_pad_h = 1; // Conv shape
ck::index_t in_left_pad_w = 1; ck::index_t N = 2;
ck::index_t in_right_pad_h = 1; ck::index_t K = 256;
ck::index_t in_right_pad_w = 1; ck::index_t C = 192;
ck::index_t Y = 3;
if(argc == 1) ck::index_t X = 3;
{ ck::index_t Hi = 71;
data_type = 0; ck::index_t Wi = 71;
init_method = 1; ck::index_t conv_stride_h = 1;
} ck::index_t conv_stride_w = 1;
else if(argc == 3) ck::index_t conv_dilation_h = 1;
{ ck::index_t conv_dilation_w = 1;
data_type = std::stoi(argv[1]); ck::index_t in_left_pad_h = 1;
init_method = std::stoi(argv[2]); ck::index_t in_left_pad_w = 1;
} ck::index_t in_right_pad_h = 1;
else if(argc == 18) ck::index_t in_right_pad_w = 1;
{
data_type = std::stoi(argv[1]); if(argc == 1)
init_method = std::stoi(argv[2]); {
data_type = 0;
N = std::stoi(argv[3]); init_method = 1;
K = std::stoi(argv[4]); }
C = std::stoi(argv[5]); else if(argc == 3)
Y = std::stoi(argv[6]); {
X = std::stoi(argv[7]); data_type = std::stoi(argv[1]);
Hi = std::stoi(argv[8]); init_method = std::stoi(argv[2]);
Wi = std::stoi(argv[9]); }
conv_stride_h = std::stoi(argv[10]); else if(argc == 18)
conv_stride_w = std::stoi(argv[11]); {
conv_dilation_h = std::stoi(argv[12]); data_type = std::stoi(argv[1]);
conv_dilation_w = std::stoi(argv[13]); init_method = std::stoi(argv[2]);
in_left_pad_h = std::stoi(argv[14]);
in_left_pad_w = std::stoi(argv[15]); N = std::stoi(argv[3]);
in_right_pad_h = std::stoi(argv[16]); K = std::stoi(argv[4]);
in_right_pad_w = std::stoi(argv[17]); C = std::stoi(argv[5]);
} Y = std::stoi(argv[6]);
else X = std::stoi(argv[7]);
{ Hi = std::stoi(argv[8]);
printf("arg1: data type (0=fp32, 1=fp16)\n"); Wi = std::stoi(argv[9]);
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"); conv_stride_h = std::stoi(argv[10]);
printf("arg3 to 17: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, " conv_stride_w = std::stoi(argv[11]);
"RightPx\n"); conv_dilation_h = std::stoi(argv[12]);
exit(1); conv_dilation_w = std::stoi(argv[13]);
} in_left_pad_h = std::stoi(argv[14]);
in_left_pad_w = std::stoi(argv[15]);
auto Run = [&](auto input_type, auto wei_type, auto out_type) { in_right_pad_h = std::stoi(argv[16]);
using InDataType = decltype(input_type); in_right_pad_w = std::stoi(argv[17]);
using WeiDataType = decltype(wei_type); }
using OutDataType = decltype(out_type); else
{
using ReferenceConvFwdInstance = printf("arg1: data type (0=fp32, 1=fp16)\n");
ck::tensor_operation::host::ReferenceConvFwd_Bias_Activation_Add<InDataType, printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
WeiDataType, printf("arg3 to 17: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, "
OutDataType, "RightPx\n");
InElementOp, exit(1);
WeiElementOp, }
OutElementOp>;
auto Run = [&](auto input_type, auto wei_type, auto out_type) {
const ck::index_t YEff = (Y - 1) * conv_dilation_h + 1; using InDataType = decltype(input_type);
const ck::index_t XEff = (X - 1) * conv_dilation_w + 1; using WeiDataType = decltype(wei_type);
using OutDataType = decltype(out_type);
const ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - YEff) / conv_stride_h + 1;
const ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - XEff) / conv_stride_w + 1; using ReferenceConvFwdInstance =
ck::tensor_operation::host::ReferenceConvFwd_Bias_Activation_Add<InDataType,
const std::vector<ck::index_t> input_spatial_lengths{{Hi, Wi}}; WeiDataType,
const std::vector<ck::index_t> filter_spatial_lengths{{Y, X}}; OutDataType,
const std::vector<ck::index_t> output_spatial_lengths{{Ho, Wo}}; InElementOp,
const std::vector<ck::index_t> conv_filter_strides{{conv_stride_h, conv_stride_w}}; WeiElementOp,
const std::vector<ck::index_t> conv_filter_dilations{{conv_dilation_h, conv_dilation_w}}; OutElementOp>;
const std::vector<ck::index_t> input_left_pads{{in_left_pad_h, in_left_pad_w}};
const std::vector<ck::index_t> input_right_pads{{in_right_pad_h, in_right_pad_w}}; const ck::index_t YEff = (Y - 1) * conv_dilation_h + 1;
const ck::index_t XEff = (X - 1) * conv_dilation_w + 1;
auto f_host_tensor_descriptor = [](std::size_t N_,
std::size_t C_, const ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - YEff) / conv_stride_h + 1;
std::size_t H_, const ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - XEff) / conv_stride_w + 1;
std::size_t W_) {
return HostTensorDescriptor(std::vector<std::size_t>({N_, C_, H_, W_}), const std::vector<ck::index_t> input_spatial_lengths{{Hi, Wi}};
std::vector<std::size_t>({C_ * H_ * W_, 1, W_ * C_, C_})); const std::vector<ck::index_t> filter_spatial_lengths{{Y, X}};
}; const std::vector<ck::index_t> output_spatial_lengths{{Ho, Wo}};
const std::vector<ck::index_t> conv_filter_strides{{conv_stride_h, conv_stride_w}};
Tensor<InDataType> in_n_c_hi_wi(f_host_tensor_descriptor(N, C, Hi, Wi)); const std::vector<ck::index_t> conv_filter_dilations{{conv_dilation_h, conv_dilation_w}};
Tensor<WeiDataType> wei_k_c_y_x(f_host_tensor_descriptor(K, C, Y, X)); const std::vector<ck::index_t> input_left_pads{{in_left_pad_h, in_left_pad_w}};
#if TEST_LAYOUT == TEST_LAYOUT_NHWC_KYXCK8_NHWK const std::vector<ck::index_t> input_right_pads{{in_right_pad_h, in_right_pad_w}};
Tensor<WeiDataType> wei_k_c_y_x_k8(
f_host_tensor_descriptor(K, C, Y, X)); // TODO: This is only to hold data auto f_host_tensor_descriptor = [](std::size_t N_,
#endif std::size_t C_,
Tensor<OutDataType> out_n_k_ho_wo_host_result(f_host_tensor_descriptor(N, K, Ho, Wo)); std::size_t H_,
Tensor<OutDataType> out_n_k_ho_wo_device_result(f_host_tensor_descriptor(N, K, Ho, Wo)); std::size_t W_) {
return HostTensorDescriptor(std::vector<std::size_t>({N_, C_, H_, W_}),
// bias: assume contiguous 1d vector std::vector<std::size_t>({C_ * H_ * W_, 1, W_ * C_, C_}));
Tensor<OutDataType> bias( };
HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(K)})));
Tensor<InDataType> in_n_c_hi_wi(f_host_tensor_descriptor(N, C, Hi, Wi));
// residual: assume same layout as output tensor Tensor<WeiDataType> wei_k_c_y_x(f_host_tensor_descriptor(K, C, Y, X));
Tensor<OutDataType> residual(f_host_tensor_descriptor(N, K, Ho, Wo)); #if TEST_LAYOUT == TEST_LAYOUT_NHWC_KYXCK8_NHWK
Tensor<WeiDataType> wei_k_c_y_x_k8(
std::cout << "in (N, C, Hi, Wi): " << in_n_c_hi_wi.mDesc << std::endl; f_host_tensor_descriptor(K, C, Y, X)); // TODO: This is only to hold data
std::cout << "wei(K, C, Y, X): " << wei_k_c_y_x.mDesc << std::endl; #endif
std::cout << "out(N, K, Ho, Wo): " << out_n_k_ho_wo_host_result.mDesc << std::endl; Tensor<OutDataType> out_n_k_ho_wo_host_result(f_host_tensor_descriptor(N, K, Ho, Wo));
std::cout << "bias: " << bias.mDesc << std::endl; Tensor<OutDataType> out_n_k_ho_wo_device_result(f_host_tensor_descriptor(N, K, Ho, Wo));
std::cout << "residual: " << residual.mDesc << std::endl;
std::cout << "LPad(H, W):" << in_left_pad_h << "," << in_left_pad_w // bias: assume contiguous 1d vector
<< ", RPad(H, W):" << in_right_pad_h << "," << in_right_pad_w Tensor<OutDataType> bias(
<< ", Stride(H, W):" << conv_stride_h << ", " << conv_stride_w HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(K)})));
<< ", Dilation(H, W):" << conv_dilation_h << ", " << conv_dilation_w
<< ", Threads:" << omp_get_max_threads() << std::endl; // residual: assume same layout as output tensor
Tensor<OutDataType> residual(f_host_tensor_descriptor(N, K, Ho, Wo));
int per_pixel_check = 0;
switch(init_method) std::cout << "in (N, C, Hi, Wi): " << in_n_c_hi_wi.mDesc << std::endl;
{ std::cout << "wei(K, C, Y, X): " << wei_k_c_y_x.mDesc << std::endl;
case 0: std::cout << "out(N, K, Ho, Wo): " << out_n_k_ho_wo_host_result.mDesc << std::endl;
in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_1<InDataType>{}); std::cout << "bias: " << bias.mDesc << std::endl;
wei_k_c_y_x.GenerateTensorValue(GeneratorTensor_1<WeiDataType>{}); std::cout << "residual: " << residual.mDesc << std::endl;
bias.GenerateTensorValue(GeneratorTensor_1<WeiDataType>{}); std::cout << "LPad(H, W):" << in_left_pad_h << "," << in_left_pad_w
residual.GenerateTensorValue(GeneratorTensor_1<WeiDataType>{}); << ", RPad(H, W):" << in_right_pad_h << "," << in_right_pad_w
per_pixel_check = 1; << ", Stride(H, W):" << conv_stride_h << ", " << conv_stride_w
break; << ", Dilation(H, W):" << conv_dilation_h << ", " << conv_dilation_w
case 1: << ", Threads:" << omp_get_max_threads() << std::endl;
in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5}); int per_pixel_check = 0;
// in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_1<InDataType>{}); switch(init_method)
wei_k_c_y_x.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5}); {
// wei_k_c_y_x.GenerateTensorValue(GeneratorTensor_1<WeiDataType>{}); case 0:
bias.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5}); in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_1<InDataType>{});
residual.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5}); wei_k_c_y_x.GenerateTensorValue(GeneratorTensor_1<WeiDataType>{});
per_pixel_check = 1; bias.GenerateTensorValue(GeneratorTensor_1<WeiDataType>{});
break; residual.GenerateTensorValue(GeneratorTensor_1<WeiDataType>{});
per_pixel_check = 1;
case 2: break;
in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0}); case 1:
wei_k_c_y_x.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5});
bias.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0}); in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
residual.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0}); // in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_1<InDataType>{});
break; wei_k_c_y_x.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
default: // wei_k_c_y_x.GenerateTensorValue(GeneratorTensor_1<WeiDataType>{});
in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_3<InDataType>{0, 1}); bias.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
wei_k_c_y_x.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-1, 1}); residual.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
bias.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0}); per_pixel_check = 1;
residual.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0}); break;
}
case 2:
DeviceAlignedMemCPU in_device_buf(sizeof(InDataType) * in_n_c_hi_wi.mDesc.GetElementSpace(), in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0});
AVX2_DATA_ALIGNMENT); wei_k_c_y_x.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5});
DeviceAlignedMemCPU wei_device_buf( bias.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0});
sizeof(WeiDataType) * wei_k_c_y_x.mDesc.GetElementSpace(), AVX2_DATA_ALIGNMENT); residual.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0});
DeviceAlignedMemCPU out_device_buf(sizeof(OutDataType) * break;
out_n_k_ho_wo_host_result.mDesc.GetElementSpace(), default:
AVX2_DATA_ALIGNMENT); in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_3<InDataType>{0, 1});
wei_k_c_y_x.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-1, 1});
DeviceAlignedMemCPU bias_device_buf(sizeof(OutDataType) * bias.mDesc.GetElementSpace(), bias.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0});
AVX2_DATA_ALIGNMENT); residual.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0});
DeviceAlignedMemCPU resi_device_buf(sizeof(OutDataType) * residual.mDesc.GetElementSpace(), }
AVX2_DATA_ALIGNMENT);
DeviceAlignedMemCPU in_device_buf(sizeof(InDataType) * in_n_c_hi_wi.mDesc.GetElementSpace(),
in_device_buf.ToDevice(in_n_c_hi_wi.mData.data()); AVX2_DATA_ALIGNMENT);
#if TEST_LAYOUT == TEST_LAYOUT_NHWC_KYXC_NHWK DeviceAlignedMemCPU wei_device_buf(
wei_device_buf.ToDevice(wei_k_c_y_x.mData.data()); sizeof(WeiDataType) * wei_k_c_y_x.mDesc.GetElementSpace(), AVX2_DATA_ALIGNMENT);
#endif DeviceAlignedMemCPU out_device_buf(sizeof(OutDataType) *
#if TEST_LAYOUT == TEST_LAYOUT_NHWC_KYXCK8_NHWK out_n_k_ho_wo_host_result.mDesc.GetElementSpace(),
transpose_kyxc_2_kyxc8k(wei_k_c_y_x_k8, wei_k_c_y_x, K, Y, X, C); AVX2_DATA_ALIGNMENT);
wei_device_buf.ToDevice(wei_k_c_y_x_k8.mData.data());
#endif DeviceAlignedMemCPU bias_device_buf(sizeof(OutDataType) * bias.mDesc.GetElementSpace(),
bias_device_buf.ToDevice(bias.mData.data()); AVX2_DATA_ALIGNMENT);
resi_device_buf.ToDevice(residual.mData.data()); DeviceAlignedMemCPU resi_device_buf(sizeof(OutDataType) * residual.mDesc.GetElementSpace(),
AVX2_DATA_ALIGNMENT);
// get host result
{ in_device_buf.ToDevice(in_n_c_hi_wi.mData.data());
auto ref_conv = ReferenceConvFwdInstance{}; #if TEST_LAYOUT == TEST_LAYOUT_NHWC_KYXC_NHWK
auto ref_invoker = ref_conv.MakeInvoker(); wei_device_buf.ToDevice(wei_k_c_y_x.mData.data());
#endif
auto ref_argument = ref_conv.MakeArgument(in_n_c_hi_wi, #if TEST_LAYOUT == TEST_LAYOUT_NHWC_KYXCK8_NHWK
wei_k_c_y_x, transpose_kyxc_2_kyxc8k(wei_k_c_y_x_k8, wei_k_c_y_x, K, Y, X, C);
out_n_k_ho_wo_host_result, wei_device_buf.ToDevice(wei_k_c_y_x_k8.mData.data());
bias, #endif
residual, bias_device_buf.ToDevice(bias.mData.data());
conv_filter_strides, resi_device_buf.ToDevice(residual.mData.data());
conv_filter_dilations,
input_left_pads, // get host result
input_right_pads, {
InElementOp{}, auto ref_conv = ReferenceConvFwdInstance{};
WeiElementOp{}, auto ref_invoker = ref_conv.MakeInvoker();
OutElementOp{});
ref_invoker.Run(ref_argument); auto ref_argument = ref_conv.MakeArgument(in_n_c_hi_wi,
} wei_k_c_y_x,
out_n_k_ho_wo_host_result,
using PassThrough = ck::tensor_operation::cpu::element_wise::PassThrough; bias,
using AddReluAdd = ck::tensor_operation::cpu::element_wise::AddReluAdd; residual,
conv_filter_strides,
using DeviceConvFwdNoOpPtr = ck::tensor_operation::cpu::device:: conv_filter_dilations,
DeviceConvFwdBiasActivationAddPtr<PassThrough, PassThrough, AddReluAdd>; input_left_pads,
input_right_pads,
// add device Conv instances InElementOp{},
std::vector<DeviceConvFwdNoOpPtr> conv_ptrs; WeiElementOp{},
OutElementOp{});
if constexpr(ck::is_same_v<ck::remove_cv_t<InDataType>, float> && ref_invoker.Run(ref_argument);
ck::is_same_v<ck::remove_cv_t<WeiDataType>, float> && }
ck::is_same_v<ck::remove_cv_t<OutDataType>, float>)
{ using PassThrough = ck::tensor_operation::cpu::element_wise::PassThrough;
#if TEST_LAYOUT == TEST_LAYOUT_NHWC_KYXC_NHWK using AddReluAdd = ck::tensor_operation::cpu::element_wise::AddReluAdd;
if(omp_get_max_threads() > 1)
{ using DeviceConvFwdNoOpPtr = ck::tensor_operation::cpu::device::
ck::tensor_operation::cpu::device:: DeviceConvFwdBiasActivationAddPtr<PassThrough, PassThrough, AddReluAdd>;
device_conv2d_fwd_bias_activation_add_avx2_instance::
add_device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxc_nhwk_mt(conv_ptrs); // add device Conv instances
ck::tensor_operation::cpu::device:: std::vector<DeviceConvFwdNoOpPtr> conv_ptrs;
device_conv2d_fwd_bias_activation_add_avx2_instance::
add_device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxc_nhwk(conv_ptrs); if constexpr(ck::is_same_v<ck::remove_cv_t<InDataType>, float> &&
} ck::is_same_v<ck::remove_cv_t<WeiDataType>, float> &&
else ck::is_same_v<ck::remove_cv_t<OutDataType>, float>)
{ {
if(K % 8 == 0) #if TEST_LAYOUT == TEST_LAYOUT_NHWC_KYXC_NHWK
ck::tensor_operation::cpu::device:: if(omp_get_max_threads() > 1)
device_conv2d_fwd_bias_activation_add_avx2_instance:: {
add_device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxc_nhwk( ck::tensor_operation::cpu::device::
conv_ptrs); device_conv2d_fwd_bias_activation_add_avx2_instance::
else add_device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxc_nhwk_mt(conv_ptrs);
ck::tensor_operation::cpu::device:: ck::tensor_operation::cpu::device::
device_conv2d_fwd_bias_activation_add_avx2_instance:: device_conv2d_fwd_bias_activation_add_avx2_instance::
add_device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxc_nhwk_local_c( add_device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxc_nhwk(conv_ptrs);
conv_ptrs); }
} else
#endif {
#if TEST_LAYOUT == TEST_LAYOUT_NHWC_KYXCK8_NHWK if(K % 8 == 0)
if(omp_get_max_threads() > 1) ck::tensor_operation::cpu::device::
{ device_conv2d_fwd_bias_activation_add_avx2_instance::
ck::tensor_operation::cpu::device:: add_device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxc_nhwk(
device_conv2d_fwd_bias_activation_add_avx2_instance:: conv_ptrs);
add_device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_mt(conv_ptrs); else
ck::tensor_operation::cpu::device:: ck::tensor_operation::cpu::device::
device_conv2d_fwd_bias_activation_add_avx2_instance:: device_conv2d_fwd_bias_activation_add_avx2_instance::
add_device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk(conv_ptrs); add_device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxc_nhwk_local_c(
} conv_ptrs);
else }
{ #endif
if(K % 8 == 0) #if TEST_LAYOUT == TEST_LAYOUT_NHWC_KYXCK8_NHWK
ck::tensor_operation::cpu::device:: if(omp_get_max_threads() > 1)
device_conv2d_fwd_bias_activation_add_avx2_instance:: {
add_device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk(conv_ptrs); ck::tensor_operation::cpu::device::
else device_conv2d_fwd_bias_activation_add_avx2_instance::
ck::tensor_operation::cpu::device:: add_device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxck8_nhwk_mt(
device_conv2d_fwd_bias_activation_add_avx2_instance:: conv_ptrs);
add_device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_local_c(conv_ptrs); ck::tensor_operation::cpu::device::
} device_conv2d_fwd_bias_activation_add_avx2_instance::
#endif add_device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxck8_nhwk(conv_ptrs);
} }
else
if(conv_ptrs.size() <= 0) {
{ if(K % 8 == 0)
throw std::runtime_error("wrong! no device Conv instance found"); ck::tensor_operation::cpu::device::
} device_conv2d_fwd_bias_activation_add_avx2_instance::
add_device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxck8_nhwk(
// profile device Conv instances conv_ptrs);
bool success = true; else
double fastest_kernel_time = std::numeric_limits<double>::max(); ck::tensor_operation::cpu::device::
std::string fastest_kernel_name = ""; device_conv2d_fwd_bias_activation_add_avx2_instance::
double fastest_kernel_gflops = 0; add_device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxck8_nhwk_local_c(
for(auto& conv_ptr : conv_ptrs) conv_ptrs);
{ }
auto argument_ptr = conv_ptr->MakeArgumentPointer( #endif
static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()), }
static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()), if(conv_ptrs.size() <= 0)
static_cast<const OutDataType*>(bias_device_buf.GetDeviceBuffer()), {
static_cast<const OutDataType*>(resi_device_buf.GetDeviceBuffer()), throw std::runtime_error("wrong! no device Conv instance found");
N, }
K,
C, // profile device Conv instances
input_spatial_lengths, bool success = true;
filter_spatial_lengths, double fastest_kernel_time = std::numeric_limits<double>::max();
output_spatial_lengths, std::string fastest_kernel_name = "";
conv_filter_strides, double fastest_kernel_gflops = 0;
conv_filter_dilations, for(auto& conv_ptr : conv_ptrs)
input_left_pads, {
input_right_pads, auto argument_ptr = conv_ptr->MakeArgumentPointer(
InElementOp{}, static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
WeiElementOp{}, static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
OutElementOp{}); static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
static_cast<const OutDataType*>(bias_device_buf.GetDeviceBuffer()),
if(conv_ptr->IsSupportedArgument(argument_ptr.get())) static_cast<const OutDataType*>(resi_device_buf.GetDeviceBuffer()),
{ N,
auto invoker_ptr = conv_ptr->MakeInvokerPointer(); K,
double time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{}, 10); C,
input_spatial_lengths,
double total_flop = static_cast<double>(2) * N * C * Ho * Wo * K * Y * X; filter_spatial_lengths,
output_spatial_lengths,
double gflops = (total_flop * 1e-6) / time; conv_filter_strides,
conv_filter_dilations,
out_device_buf.FromDevice(out_n_k_ho_wo_device_result.mData.data()); input_left_pads,
input_right_pads,
if(!check_out(out_n_k_ho_wo_host_result, InElementOp{},
out_n_k_ho_wo_device_result, WeiElementOp{},
1e-6, OutElementOp{});
per_pixel_check))
{ if(conv_ptr->IsSupportedArgument(argument_ptr.get()))
std::cout << "Fail Info: " << conv_ptr->GetTypeString() << std::endl; {
success = false; auto invoker_ptr = conv_ptr->MakeInvokerPointer();
} double time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{}, 10);
else
{ double total_flop = static_cast<double>(2) * N * C * Ho * Wo * K * Y * X;
std::cout << "Pass Info: " << conv_ptr->GetTypeString() << ", Time:" << time
<< "ms, Gflops:" << gflops << std::endl; double gflops = (total_flop * 1e-6) / time;
if(time < fastest_kernel_time) out_device_buf.FromDevice(out_n_k_ho_wo_device_result.mData.data());
{
fastest_kernel_time = time; if(!check_out(out_n_k_ho_wo_host_result,
fastest_kernel_name = conv_ptr->GetTypeString(); out_n_k_ho_wo_device_result,
fastest_kernel_gflops = gflops; 1e-6,
} per_pixel_check))
} {
} std::cout << "Fail Info: " << conv_ptr->GetTypeString() << std::endl;
else success = false;
{ }
std::cout << "Not support Info: " << conv_ptr->GetTypeString() << std::endl; else
} {
} std::cout << "Pass Info: " << conv_ptr->GetTypeString() << ", Time:" << time
<< "ms, Gflops:" << gflops << std::endl;
if(fastest_kernel_time != std::numeric_limits<double>::max())
{ if(time < fastest_kernel_time)
std::cout << " fastest:" << fastest_kernel_name << ", time:" << fastest_kernel_time {
<< "ms, Gflops:" << fastest_kernel_gflops << std::endl; fastest_kernel_time = time;
} fastest_kernel_name = conv_ptr->GetTypeString();
return 0; fastest_kernel_gflops = gflops;
// if(success) }
// { }
// std::cout << "test conv2d fwd cpu : Pass" << std::endl; }
// return 0; else
// } {
// else std::cout << "Not support Info: " << conv_ptr->GetTypeString() << std::endl;
// { }
// std::cout << "test conv2d fwd cpu: Fail " << std::endl; }
// return -1;
// } if(fastest_kernel_time != std::numeric_limits<double>::max())
}; {
std::cout << " fastest:" << fastest_kernel_name << ", time:" << fastest_kernel_time
if(data_type == 0) << "ms, Gflops:" << fastest_kernel_gflops << std::endl;
{ }
return Run(F32(), F32(), F32()); return 0;
} // if(success)
else // {
{ // std::cout << "test conv2d fwd cpu : Pass" << std::endl;
return 1; // return 0;
} // }
} // else
// {
// std::cout << "test conv2d fwd cpu: Fail " << std::endl;
// return -1;
// }
};
if(data_type == 0)
{
return Run(F32(), F32(), F32());
}
else
{
return 1;
}
}
include/ck/tensor_operation/cpu/device/device_convnd_fwd_bias_activation_add_avx2_nhwc_kyxck8_nhwk.hpp 0 → 100644
View file @ dc536427
#ifndef DEVICE_CONV2D_FWD_BIAS_ACTIVATION_ADD_AVX2_NHWC_KYXCK8_NHWK_HPP
#define DEVICE_CONV2D_FWD_BIAS_ACTIVATION_ADD_AVX2_NHWC_KYXCK8_NHWK_HPP
#include <iostream>
#include <sstream>
#include <numeric>
#include "device.hpp"
#include "device_base_cpu.hpp"
#include "device_conv_fwd_cpu.hpp"
#include "convolution_forward_specialization_cpu.hpp"
#include "common_header.hpp"
#include "../../gpu/device/tensor_layout.hpp"
#include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp"
#include "gridwise_gemm_bias_activation_add_avx2.hpp"
#include "threadwise_gemm_avx2.hpp"
#include "threadwise_tensor_slice_transfer_avx2_specialization.hpp"
namespace ck {
namespace tensor_operation {
namespace cpu {
namespace device {
// out[N, Ho, Wo, K] = in[N, Hi, Wi, C] * wei[K, Y, X, C]
template <typename InDataType,
typename WeiDataType,
typename OutDataType,
typename BiasDataType,
typename AddDataType,
typename InElementwiseOperation,
typename WeiElementwiseOperation,
typename OutElementwiseOperation,
ConvolutionForwardSpecialization_t ConvForwardSpecialization,
ConvolutionForwardGemmKSpecialization_t GemmKSpecialization,
ConvolutionForwardBlockLoopOverSpecialization_t BlockLoopOverSpecialization,
ck::index_t NumDimSpatial,
ck::index_t MPerBlock, // block means data are designed to fit in cache (L1/L2/L3)
ck::index_t NPerBlock,
ck::index_t KPerBlock,
ck::index_t MPerThread,
ck::index_t NPerThread,
bool UseALocalBuffer,
bool UseBLocalBuffer,
bool UseCLocalBuffer,
bool BiasAlongGemmM>
struct DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K
: public DeviceConvFwdBiasActivationAdd<InElementwiseOperation,
WeiElementwiseOperation,
OutElementwiseOperation>
{
using DeviceOp =
DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K;
using ADataType = InDataType;
using BDataType = WeiDataType;
using CDataType = OutDataType;
using C0DataType = BiasDataType;
using C1DataType = AddDataType;
using AElementwiseOperation = InElementwiseOperation;
using BElementwiseOperation = WeiElementwiseOperation;
using CElementwiseOperation = OutElementwiseOperation;
// TODO make A/B datatype different
using ABDataType = InDataType;
static constexpr index_t NDimSpatial = NumDimSpatial;
static constexpr auto I0 = Number<0>{};
static constexpr auto I1 = Number<1>{};
static constexpr auto I2 = Number<2>{};
static constexpr auto I3 = Number<3>{};
static constexpr bool NonTemporalStore = false;
static constexpr auto GetBlockMNKAccessOrder()
{
if constexpr(BlockLoopOverSpecialization == DefaultBlockLoopOver ||
BlockLoopOverSpecialization == LoopOver_MNK)
return ck::Sequence<0, 1, 2>{};
else if constexpr(BlockLoopOverSpecialization == LoopOver_MKN)
return ck::Sequence<0, 2, 1>{};
}
using BlockMNKAccessOrder = decltype(GetBlockMNKAccessOrder());
static constexpr auto GetThreadwiseGemm_Dispatch()
{
if constexpr(MPerThread == 4 && NPerThread == 24)
{
return ck::cpu::ThreadwiseGemmAvx2_MxN_4x24_Dispatch<
InDataType,
WeiDataType,
OutDataType,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::ColumnMajor,
NonTemporalStore>{};
}
else if constexpr(MPerThread == 6 && NPerThread == 16)
{
return ck::cpu::ThreadwiseGemmAvx2_MxN_6x16_Dispatch<
InDataType,
WeiDataType,
OutDataType,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::ColumnMajor,
NonTemporalStore>{};
}
else
{
// static_assert(false, "invalid Mr/Nr");
}
}
using ThreadwiseGemm_Dispatch = decltype(GetThreadwiseGemm_Dispatch());
static constexpr auto GetInputBlockDescriptor()
{
return make_naive_tensor_descriptor_packed(make_tuple(MPerBlock, KPerBlock));
}
static constexpr auto GetWeightBlockDescriptor()
{
return make_naive_tensor_descriptor_packed(make_tuple(
math::integer_divide_ceil(NPerBlock, ThreadwiseGemm_Dispatch::MatrixBMinVectorSize),
KPerBlock,
ThreadwiseGemm_Dispatch::MatrixBMinVectorSize));
}
static constexpr auto GetOutputBlockDescriptor()
{
return make_naive_tensor_descriptor_packed(make_tuple(MPerBlock, NPerBlock));
}
static auto GetWeightTensorDescriptor(ck::index_t gemm_k, ck::index_t gemm_n)
{
return make_naive_tensor_descriptor_packed(make_tuple(gemm_n / 8, gemm_k, 8));
}
static auto GetOutputTensorDescriptor(ck::index_t gemm_m, ck::index_t gemm_n)
{
const auto out_gemm_m_n_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(gemm_m, gemm_n));
return out_gemm_m_n_grid_desc;
}
static auto MakeBiasTensorDescriptor(ck::index_t gemm_m, ck::index_t gemm_n)
{
if constexpr(BiasAlongGemmM)
{
return make_naive_tensor_descriptor_packed(make_tuple(gemm_m));
}
else
{
return make_naive_tensor_descriptor_packed(make_tuple(gemm_n));
}
}
template <ck::index_t NDim, typename std::enable_if<NDim == 1, bool>::type = false>
static auto GetInputTensorDescriptor(ck::index_t N,
ck::index_t C,
ck::index_t gemm_m,
ck::index_t gemm_k,
const std::vector<ck::index_t>& input_spatial_lengths,
const std::vector<ck::index_t>& filter_spatial_lengths,
const std::vector<ck::index_t>& output_spatial_lengths,
const std::vector<ck::index_t>& conv_filter_strides,
const std::vector<ck::index_t>& conv_filter_dilations,
const std::vector<ck::index_t>& input_left_pads,
const std::vector<ck::index_t>& input_right_pads)
{
const index_t Wi = input_spatial_lengths[0];
const index_t Wo = output_spatial_lengths[0];
const index_t ConvStrideW = conv_filter_strides[0];
if constexpr(ConvForwardSpecialization ==
ConvolutionForwardSpecialization_t::Filter1x1Stride1Pad0)
{
const auto in_gemm_m_k_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(gemm_m, gemm_k));
return in_gemm_m_k_grid_desc;
}
else if constexpr(ConvForwardSpecialization ==
ConvolutionForwardSpecialization_t::Filter1x1Pad0)
{
const auto in_n_wi_c_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(N, Wi, C));
const auto in_n_wo_c_grid_desc = transform_tensor_descriptor(
in_n_wi_c_grid_desc,
make_tuple(make_pass_through_transform(N),
make_embed_transform(make_tuple(Wo), make_tuple(ConvStrideW)),
make_pass_through_transform(C)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
const auto in_gemm_m_k_grid_desc = transform_tensor_descriptor(
in_n_wo_c_grid_desc,
make_tuple(make_merge_transform(make_tuple(N, Wo)), make_pass_through_transform(C)),
make_tuple(Sequence<0, 1>{}, Sequence<2>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
return in_gemm_m_k_grid_desc;
}
else
{
const index_t X = filter_spatial_lengths[0];
const index_t ConvDilationW = conv_filter_dilations[0];
const index_t InLeftPadW = input_left_pads[0];
const index_t InRightPadW = input_right_pads[0];
const auto in_n_wi_c_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(N, Wi, C));
const auto in_n_wip_c_grid_desc = transform_tensor_descriptor(
in_n_wi_c_grid_desc,
make_tuple(make_pass_through_transform(N),
make_pad_transform(Wi, InLeftPadW, InRightPadW),
make_pass_through_transform(C)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
const auto in_n_x_wo_c_grid_desc = transform_tensor_descriptor(
in_n_wip_c_grid_desc,
make_tuple(
make_pass_through_transform(N),
make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
make_pass_through_transform(C)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
const auto in_gemm_m_k_grid_desc =
transform_tensor_descriptor(in_n_x_wo_c_grid_desc,
make_tuple(make_merge_transform(make_tuple(N, Wo)),
make_merge_transform(make_tuple(X, C))),
make_tuple(Sequence<0, 2>{}, Sequence<1, 3>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
return in_gemm_m_k_grid_desc;
}
}
template <ck::index_t NDim, typename std::enable_if<NDim == 2, bool>::type = false>
static auto GetInputTensorDescriptor(ck::index_t N,
ck::index_t C,
ck::index_t gemm_m,
ck::index_t gemm_k,
const std::vector<ck::index_t>& input_spatial_lengths,
const std::vector<ck::index_t>& filter_spatial_lengths,
const std::vector<ck::index_t>& output_spatial_lengths,
const std::vector<ck::index_t>& conv_filter_strides,
const std::vector<ck::index_t>& conv_filter_dilations,
const std::vector<ck::index_t>& input_left_pads,
const std::vector<ck::index_t>& input_right_pads)
{
const index_t Hi = input_spatial_lengths[0];
const index_t Wi = input_spatial_lengths[1];
const index_t Ho = output_spatial_lengths[0];
const index_t Wo = output_spatial_lengths[1];
const index_t ConvStrideH = conv_filter_strides[0];
const index_t ConvStrideW = conv_filter_strides[1];
if constexpr(ConvForwardSpecialization ==
ConvolutionForwardSpecialization_t::Filter1x1Stride1Pad0)
{
const auto in_gemm_m_k_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(gemm_m, gemm_k));
return in_gemm_m_k_grid_desc;
}
else if constexpr(ConvForwardSpecialization ==
ConvolutionForwardSpecialization_t::Filter1x1Pad0)
{
const auto in_n_hi_wi_c_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
const auto in_n_ho_wo_c_grid_desc = transform_tensor_descriptor(
in_n_hi_wi_c_grid_desc,
make_tuple(make_pass_through_transform(N),
make_embed_transform(make_tuple(Ho), make_tuple(ConvStrideH)),
make_embed_transform(make_tuple(Wo), make_tuple(ConvStrideW)),
make_pass_through_transform(C)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
const auto in_gemm_m_k_grid_desc =
transform_tensor_descriptor(in_n_ho_wo_c_grid_desc,
make_tuple(make_merge_transform(make_tuple(N, Ho, Wo)),
make_pass_through_transform(C)),
make_tuple(Sequence<0, 1, 2>{}, Sequence<3>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
return in_gemm_m_k_grid_desc;
}
else
{
const index_t Y = filter_spatial_lengths[0];
const index_t X = filter_spatial_lengths[1];
const index_t ConvDilationH = conv_filter_dilations[0];
const index_t ConvDilationW = conv_filter_dilations[1];
const index_t InLeftPadH = input_left_pads[0];
const index_t InLeftPadW = input_left_pads[1];
const index_t InRightPadH = input_right_pads[0];
const index_t InRightPadW = input_right_pads[1];
const auto in_n_hi_wi_c_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
in_n_hi_wi_c_grid_desc,
make_tuple(make_pass_through_transform(N),
make_pad_transform(Hi, InLeftPadH, InRightPadH),
make_pad_transform(Wi, InLeftPadW, InRightPadW),
make_pass_through_transform(C)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
const auto in_n_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
in_n_hip_wip_c_grid_desc,
make_tuple(
make_pass_through_transform(N),
make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
make_pass_through_transform(C)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
const auto in_gemm_m_k_grid_desc =
transform_tensor_descriptor(in_n_y_ho_x_wo_c_grid_desc,
make_tuple(make_merge_transform(make_tuple(N, Ho, Wo)),
make_merge_transform(make_tuple(Y, X, C))),
make_tuple(Sequence<0, 2, 4>{}, Sequence<1, 3, 5>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
return in_gemm_m_k_grid_desc;
}
}
template <ck::index_t NDim, typename std::enable_if<NDim == 3, bool>::type = false>
static auto GetInputTensorDescriptor(ck::index_t N,
ck::index_t C,
ck::index_t gemm_m,
ck::index_t gemm_k,
ck::index_t gemm_m_pad,
const std::vector<ck::index_t>& input_spatial_lengths,
const std::vector<ck::index_t>& filter_spatial_lengths,
const std::vector<ck::index_t>& output_spatial_lengths,
const std::vector<ck::index_t>& conv_filter_strides,
const std::vector<ck::index_t>& conv_filter_dilations,
const std::vector<ck::index_t>& input_left_pads,
const std::vector<ck::index_t>& input_right_pads)
{
const index_t Di = input_spatial_lengths[0];
const index_t Hi = input_spatial_lengths[1];
const index_t Wi = input_spatial_lengths[2];
const index_t Do = output_spatial_lengths[0];
const index_t Ho = output_spatial_lengths[1];
const index_t Wo = output_spatial_lengths[2];
const index_t ConvStrideD = conv_filter_strides[0];
const index_t ConvStrideH = conv_filter_strides[1];
const index_t ConvStrideW = conv_filter_strides[2];
if constexpr(ConvForwardSpecialization ==
ConvolutionForwardSpecialization_t::Filter1x1Stride1Pad0)
{
const auto in_gemm_m_k_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(gemm_m, gemm_k));
return in_gemm_m_k_grid_desc;
}
else if constexpr(ConvForwardSpecialization ==
ConvolutionForwardSpecialization_t::Filter1x1Pad0)
{
const auto in_n_di_hi_wi_c_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(N, Di, Hi, Wi, C));
const auto in_n_do_ho_wo_c_grid_desc = transform_tensor_descriptor(
in_n_di_hi_wi_c_grid_desc,
make_tuple(make_pass_through_transform(N),
make_embed_transform(make_tuple(Do), make_tuple(ConvStrideD)),
make_embed_transform(make_tuple(Ho), make_tuple(ConvStrideH)),
make_embed_transform(make_tuple(Wo), make_tuple(ConvStrideW)),
make_pass_through_transform(C)),
make_tuple(
Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
make_tuple(
Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}));
const auto in_gemm_m_k_grid_desc = transform_tensor_descriptor(
in_n_do_ho_wo_c_grid_desc,
make_tuple(make_merge_transform(make_tuple(N, Do, Ho, Wo)),
make_pass_through_transform(C)),
make_tuple(Sequence<0, 1, 2, 3>{}, Sequence<4>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
return in_gemm_m_k_grid_desc;
}
else
{
const index_t Z = filter_spatial_lengths[0];
const index_t Y = filter_spatial_lengths[1];
const index_t X = filter_spatial_lengths[2];
const index_t ConvDilationD = conv_filter_dilations[0];
const index_t ConvDilationH = conv_filter_dilations[1];
const index_t ConvDilationW = conv_filter_dilations[2];
const index_t InLeftPadD = input_left_pads[0];
const index_t InLeftPadH = input_left_pads[1];
const index_t InLeftPadW = input_left_pads[2];
const index_t InRightPadD = input_right_pads[0];
const index_t InRightPadH = input_right_pads[1];
const index_t InRightPadW = input_right_pads[2];
const auto in_n_di_hi_wi_c_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(N, Di, Hi, Wi, C));
const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
in_n_di_hi_wi_c_grid_desc,
make_tuple(make_pass_through_transform(N),
make_pad_transform(Di, InLeftPadD, InRightPadD),
make_pad_transform(Hi, InLeftPadH, InRightPadH),
make_pad_transform(Wi, InLeftPadW, InRightPadW),
make_pass_through_transform(C)),
make_tuple(
Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
make_tuple(
Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}));
const auto in_n_z_do_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
in_n_hip_wip_c_grid_desc,
make_tuple(
make_pass_through_transform(N),
make_embed_transform(make_tuple(Z, Do), make_tuple(ConvDilationD, ConvStrideD)),
make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
make_pass_through_transform(C)),
make_tuple(
Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
make_tuple(Sequence<0>{},
Sequence<1, 2>{},
Sequence<3, 4>{},
Sequence<5, 6>{},
Sequence<7>{}));
const auto in_gemm_m_k_grid_desc = transform_tensor_descriptor(
in_n_z_do_y_ho_x_wo_c_grid_desc,
make_tuple(make_merge_transform(make_tuple(N, Do, Ho, Wo)),
make_merge_transform(make_tuple(Z, Y, X, C))),
make_tuple(Sequence<0, 2, 4, 6>{}, Sequence<1, 3, 5, 7>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
return in_gemm_m_k_grid_desc;
}
}
static index_t GetGemmM(ck::index_t N, const std::vector<ck::index_t>& output_spatial_lengths)
{
return N * std::accumulate(std::begin(output_spatial_lengths),
std::end(output_spatial_lengths),
1,
std::multiplies<ck::index_t>());
}
static index_t GetGemmK(ck::index_t C, const std::vector<ck::index_t>& filter_spatial_lengths)
{
return C * std::accumulate(std::begin(filter_spatial_lengths),
std::end(filter_spatial_lengths),
1,
std::multiplies<ck::index_t>());
}
static index_t GetGemmN(ck::index_t K)
{
// return ck::math::integer_least_multiple(K,
// ThreadwiseGemm_Dispatch::MatrixBMinVectorSize);
return K;
}
static auto MakeABCGridDescriptor(ck::index_t N,
ck::index_t K,
ck::index_t C,
std::vector<ck::index_t> input_spatial_lengths,
std::vector<ck::index_t> filter_spatial_lengths,
std::vector<ck::index_t> output_spatial_lengths,
std::vector<ck::index_t> conv_filter_strides,
std::vector<ck::index_t> conv_filter_dilations,
std::vector<ck::index_t> input_left_pads,
std::vector<ck::index_t> input_right_pads)
{
using namespace ck;
const index_t GemmM = GetGemmM(N, output_spatial_lengths);
const index_t GemmN = GetGemmN(K);
const index_t GemmK = GetGemmK(C, filter_spatial_lengths);
// A:
const auto in_gemm_m_k_grid_desc =
GetInputTensorDescriptor<NumDimSpatial>(N,
C,
GemmM,
GemmK,
input_spatial_lengths,
filter_spatial_lengths,
output_spatial_lengths,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads);
// B:
const auto wei_gemm_n0_k_n1_grid_desc = GetWeightTensorDescriptor(GemmK, GemmN);
// C:
const auto out_gemm_m_n_grid_desc = GetOutputTensorDescriptor(GemmM, GemmN);
return make_tuple(
in_gemm_m_k_grid_desc, wei_gemm_n0_k_n1_grid_desc, out_gemm_m_n_grid_desc);
}
template <ck::index_t NDim, typename std::enable_if<NDim == 1, bool>::type = false>
static auto GetABCGridDesc()
{
return MakeABCGridDescriptor(1, 1, 1, {1}, {1}, {1}, {1}, {1}, {1}, {1});
}
template <ck::index_t NDim, typename std::enable_if<NDim == 2, bool>::type = false>
static auto GetABCGridDesc()
{
return MakeABCGridDescriptor(
1, 1, 1, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1});
}
template <ck::index_t NDim, typename std::enable_if<NDim == 3, bool>::type = false>
static auto GetABCGridDesc()
{
return MakeABCGridDescriptor(
1, 1, 1, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1});
}
using ABCGridDescs = decltype(GetABCGridDesc<NumDimSpatial>());
using AGridDesc = remove_cvref_t<decltype(ABCGridDescs{}[I0])>;
using BGridDesc = remove_cvref_t<decltype(ABCGridDescs{}[I1])>;
using CGridDesc = remove_cvref_t<decltype(ABCGridDescs{}[I2])>;
using C0GridDesc = remove_cvref_t<decltype(MakeBiasTensorDescriptor(1, 1))>;
using C1GridDesc = CGridDesc;
// static constexpr bool UseCLocalBuffer = false;
using AThreadwiseCopy =
ck::cpu::ThreadwiseTensorSliceTransferAvx2Specialization_ConvFwd_In_NHWC<
ADataType,
ADataType,
AGridDesc,
decltype(GetInputBlockDescriptor()),
InElementwiseOperation,
false,
ConvForwardSpecialization,
GemmKSpecialization>;
using BThreadwiseCopy =
ck::cpu::ThreadwiseTensorSliceTransferAvx2Specialization_ConvFwd_Wei_KYXCK8<
BDataType,
BDataType,
BGridDesc,
decltype(GetWeightBlockDescriptor()),
WeiElementwiseOperation,
false,
ConvForwardSpecialization,
GemmKSpecialization>;
using CThreadwiseCopy =
ck::cpu::ThreadwiseTensorSliceTransferAvx2Specialization_MatC_Store_Bias_Residual_MxN<
CDataType,
C0DataType,
C1DataType,
CDataType,
CGridDesc,
C0GridDesc,
C1GridDesc,
decltype(GetOutputBlockDescriptor()),
OutElementwiseOperation,
!UseCLocalBuffer,
BiasAlongGemmM>;
using GridwiseGemm = ck::cpu::GridwiseGemmBiasActivationAddAvx2_MxN<
ADataType, // InDataType,
BDataType, // WeiDataType,
CDataType, // OutDataType,
C0DataType, // C0DataType
C1DataType, // C1DataType
AGridDesc, // AGridDesc,
BGridDesc, // BGridDesc,
CGridDesc, // CGridDesc,
C0GridDesc, // C0GridDesc,
C1GridDesc, // C1GridDesc,
AElementwiseOperation, // AElementwiseOperation,
BElementwiseOperation, // BElementwiseOperation,
CElementwiseOperation, // CElementwiseOperation,
MPerBlock, // MPerBlock,
NPerBlock, // NPerBlock,
KPerBlock, // KPerBlock,
ThreadwiseGemm_Dispatch, // ThreadwiseGemm_Dispatch,
AThreadwiseCopy, // AThreadwiseCopy
BThreadwiseCopy, // BThreadwiseCopy
CThreadwiseCopy, // CThreadwiseCopy
BlockMNKAccessOrder, // BlockMNKAccessOrder,
ck::Sequence<0, 1>, // ThreadMNAccessOrder
UseALocalBuffer, // UseALocalBuffer
UseBLocalBuffer, // UseBLocalBuffer
UseCLocalBuffer // UseCLocalBuffer
>;
// Argument
struct Argument : public BaseArgument
{
Argument(const InDataType* p_in_grid,
const WeiDataType* p_wei_grid,
OutDataType* p_out_grid,
const BiasDataType* p_bias_grid,
const AddDataType* p_add_grid,
ck::index_t N,
ck::index_t K,
ck::index_t C,
std::vector<ck::index_t> input_spatial_lengths,
std::vector<ck::index_t> filter_spatial_lengths,
std::vector<ck::index_t> output_spatial_lengths,
std::vector<ck::index_t> conv_filter_strides,
std::vector<ck::index_t> conv_filter_dilations,
std::vector<ck::index_t> input_left_pads,
std::vector<ck::index_t> input_right_pads,
InElementwiseOperation in_element_op,
WeiElementwiseOperation wei_element_op,
OutElementwiseOperation out_element_op)
: p_a_grid_{p_in_grid},
p_b_grid_{p_wei_grid},
p_c_grid_{p_out_grid},
p_c0_grid_{p_bias_grid},
p_c1_grid_{p_add_grid},
a_grid_desc_{},
b_grid_desc_{},
c_grid_desc_{},
c0_grid_desc_{},
c1_grid_desc_{},
a_element_op_{in_element_op},
b_element_op_{wei_element_op},
c_element_op_{out_element_op},
Conv_N_{N},
Conv_K_{K},
Conv_C_{C},
filter_spatial_lengths_{filter_spatial_lengths},
conv_filter_strides_{conv_filter_strides},
input_left_pads_{input_left_pads},
input_right_pads_{input_right_pads}
{
const auto descs = DeviceOp::MakeABCGridDescriptor(N,
K,
C,
input_spatial_lengths,
filter_spatial_lengths,
output_spatial_lengths,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads);
a_grid_desc_ = descs[I0];
b_grid_desc_ = descs[I1];
c_grid_desc_ = descs[I2];
c0_grid_desc_ = DeviceOp::MakeBiasTensorDescriptor(GetGemmM(N, output_spatial_lengths),
GetGemmN(K));
c1_grid_desc_ = descs[I2];
}
// private:
const ADataType* p_a_grid_;
const BDataType* p_b_grid_;
CDataType* p_c_grid_;
const C0DataType* p_c0_grid_;
const C1DataType* p_c1_grid_;
AGridDesc a_grid_desc_;
BGridDesc b_grid_desc_;
CGridDesc c_grid_desc_;
C0GridDesc c0_grid_desc_;
C1GridDesc c1_grid_desc_;
AElementwiseOperation a_element_op_;
BElementwiseOperation b_element_op_;
CElementwiseOperation c_element_op_;
// for checking IsSupportedArgument()
index_t Conv_N_;
index_t Conv_K_;
index_t Conv_C_;
std::vector<index_t> filter_spatial_lengths_;
std::vector<index_t> conv_filter_strides_;
std::vector<index_t> input_left_pads_;
std::vector<index_t> input_right_pads_;
};
// Invoker
struct Invoker : public BaseInvoker
{
using Argument = DeviceOp::Argument;
float Run(const Argument& arg,
const StreamConfig& stream_config = StreamConfig{},
int nrepeat = 1)
{
if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_, arg.b_grid_desc_, arg.c_grid_desc_))
{
throw std::runtime_error("wrong! GridwiseGemmAvx2_MxN has invalid setting");
}
memset(arg.p_c_grid_, 0, arg.c_grid_desc_.GetElementSpaceSize());
const auto kernel =
ck::cpu::kernel_gemm_bias_activation_add_avx_mxn<GridwiseGemm,
ADataType,
BDataType,
CDataType,
C0DataType,
C1DataType,
AGridDesc,
BGridDesc,
CGridDesc,
C0GridDesc,
C1GridDesc,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation>;
float ave_time = 0;
if(nrepeat != 1)
ave_time = launch_and_time_cpu_kernel(kernel,
nrepeat,
arg.p_a_grid_,
arg.p_b_grid_,
arg.p_c_grid_,
arg.p_c0_grid_,
arg.p_c1_grid_,
arg.a_grid_desc_,
arg.b_grid_desc_,
arg.c_grid_desc_,
arg.c0_grid_desc_,
arg.c1_grid_desc_,
arg.a_element_op_,
arg.b_element_op_,
arg.c_element_op_);
// TODO: this is for benchmark purpose, so last time we clear c buffer and calculate the
// result
memset(arg.p_c_grid_, 0, arg.c_grid_desc_.GetElementSpaceSize());
launch_cpu_kernel(kernel,
arg.p_a_grid_,
arg.p_b_grid_,
arg.p_c_grid_,
arg.p_c0_grid_,
arg.p_c1_grid_,
arg.a_grid_desc_,
arg.b_grid_desc_,
arg.c_grid_desc_,
arg.c0_grid_desc_,
arg.c1_grid_desc_,
arg.a_element_op_,
arg.b_element_op_,
arg.c_element_op_);
return ave_time;
}
float Run(const BaseArgument* p_arg,
const StreamConfig& stream_config = StreamConfig{},
int nrepeat = 1) override
{
return Run(*dynamic_cast<const Argument*>(p_arg), stream_config, nrepeat);
}
};
static constexpr bool IsValidCompilationParameter()
{
// TODO: properly implement this check
return true;
}
static bool IsSupportedArgument(const Argument& arg)
{
if constexpr(ConvForwardSpecialization ==
ConvolutionForwardSpecialization_t::Filter1x1Stride1Pad0)
{
// check if it's 1x1, stride=1 conv
if(!(arg.filter_spatial_lengths_[0] == 1 && arg.filter_spatial_lengths_[1] == 1 &&
arg.conv_filter_strides_[0] == 1 && arg.conv_filter_strides_[1] == 1 &&
arg.input_left_pads_[0] == 0 && arg.input_left_pads_[1] == 0 &&
arg.input_right_pads_[0] == 0 && arg.input_right_pads_[1] == 0))
{
return false;
}
}
else if constexpr(ConvForwardSpecialization ==
ConvolutionForwardSpecialization_t::Filter1x1Pad0)
{
// check if it's 1x1 conv
if(!(arg.filter_spatial_lengths_[0] == 1 && arg.filter_spatial_lengths_[1] == 1 &&
arg.input_left_pads_[0] == 0 && arg.input_left_pads_[1] == 0 &&
arg.input_right_pads_[0] == 0 && arg.input_right_pads_[1] == 0))
{
return false;
}
}
if constexpr(GemmKSpecialization ==
ConvolutionForwardGemmKSpecialization_t::NHWC_GemmKLoopOverC)
{
if(!(arg.Conv_C_ % KPerBlock == 0))
return false;
}
if(!(arg.Conv_K_ % 8 == 0))
return false;
// Gridwise GEMM size
return GridwiseGemm::CheckValidity(arg.a_grid_desc_, arg.b_grid_desc_, arg.c_grid_desc_);
}
bool IsSupportedArgument(const BaseArgument* p_arg) override
{
return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
}
static auto MakeArgument(const InDataType* p_in_grid,
const WeiDataType* p_wei_grid,
OutDataType* p_out_grid,
const BiasDataType* p_bias_grid,
const AddDataType* p_add_grid,
ck::index_t N,
ck::index_t K,
ck::index_t C,
std::vector<ck::index_t> input_spatial_lengths,
std::vector<ck::index_t> filter_spatial_lengths,
std::vector<ck::index_t> output_spatial_lengths,
std::vector<ck::index_t> conv_filter_strides,
std::vector<ck::index_t> conv_filter_dilations,
std::vector<ck::index_t> input_left_pads,
std::vector<ck::index_t> input_right_pads,
InElementwiseOperation in_element_op,
WeiElementwiseOperation wei_element_op,
OutElementwiseOperation out_element_op)
{
return Argument{p_in_grid,
p_wei_grid,
p_out_grid,
p_bias_grid,
p_add_grid,
N,
K,
C,
input_spatial_lengths,
filter_spatial_lengths,
output_spatial_lengths,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
in_element_op,
wei_element_op,
out_element_op};
}
static auto MakeInvoker() { return Invoker{}; }
std::unique_ptr<BaseArgument>
MakeArgumentPointer(const void* p_in_grid,
const void* p_wei_grid,
void* p_out_grid,
const void* p_bias_grid,
const void* p_add_grid,
ck::index_t N,
ck::index_t K,
ck::index_t C,
std::vector<ck::index_t> input_spatial_lengths,
std::vector<ck::index_t> filter_spatial_lengths,
std::vector<ck::index_t> output_spatial_lengths,
std::vector<ck::index_t> conv_filter_strides,
std::vector<ck::index_t> conv_filter_dilations,
std::vector<ck::index_t> input_left_pads,
std::vector<ck::index_t> input_right_pads,
InElementwiseOperation in_element_op,
WeiElementwiseOperation wei_element_op,
OutElementwiseOperation out_element_op) override
{
return std::make_unique<Argument>(static_cast<const InDataType*>(p_in_grid),
static_cast<const WeiDataType*>(p_wei_grid),
static_cast<OutDataType*>(p_out_grid),
static_cast<const BiasDataType*>(p_bias_grid),
static_cast<const AddDataType*>(p_add_grid),
N,
K,
C,
input_spatial_lengths,
filter_spatial_lengths,
output_spatial_lengths,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
in_element_op,
wei_element_op,
out_element_op);
}
std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
{
return std::make_unique<Invoker>(Invoker{});
}
std::string GetTypeString() const override
{
auto str = std::stringstream();
auto string_local_buffer = [](bool is_local_buffer) {
if(is_local_buffer)
return "L";
else
return "G";
};
// clang-format off
str << "DeviceConv" << std::to_string(NumDimSpatial)
<< "DFwd_BAA_Avx2_NHWC_KYXCK8"
<<"_FS"<< static_cast<int>(ConvForwardSpecialization)
<<"_KS"<< static_cast<int>(GemmKSpecialization)
<<"_BS"<< static_cast<int>(BlockLoopOverSpecialization)
<< "_BT" << MPerBlock << "x" << NPerBlock << "x" << KPerBlock
<< "_TT" << MPerThread << "x" << NPerThread
<< "_A" << string_local_buffer(UseALocalBuffer)
<< "_B" << string_local_buffer(UseBLocalBuffer)
<< "_C" << string_local_buffer(UseCLocalBuffer)
;
if constexpr (!std::is_same<OutElementwiseOperation,
ck::tensor_operation::cpu::element_wise::PassThrough>::value)
{
str << "_" << OutElementwiseOperation::Name();
}
// clang-format on
return str.str();
}
};
} // namespace device
} // namespace cpu
} // namespace tensor_operation
} // namespace ck
#endif
library/src/tensor_operation_instance/cpu/conv2d_fwd_bias_activation_add/CMakeLists.txt
View file @ dc536427
# device_conv2d_fwd_bias_activation_add_cpu_instance # device_conv2d_fwd_bias_activation_add_cpu_instance
set(DEVICE_CONV2D_FWD_CPU_INSTANCE_SOURCE set(DEVICE_CONV2D_FWD_CPU_INSTANCE_SOURCE
device_conv2d_bias_activation_add_avx2_nhwc_kyxc_nhwk_instance.cpp device_conv2d_bias_activation_add_avx2_nhwc_kyxc_nhwk_instance.cpp
device_conv2d_bias_activation_add_avx2_nhwc_kyxck8_nhwk_instance.cpp
) )
add_library(device_conv2d_fwd_bias_activation_add_cpu_instance SHARED ${DEVICE_CONV2D_FWD_CPU_INSTANCE_SOURCE}) add_library(device_conv2d_fwd_bias_activation_add_cpu_instance SHARED ${DEVICE_CONV2D_FWD_CPU_INSTANCE_SOURCE})
target_compile_features(device_conv2d_fwd_bias_activation_add_cpu_instance PUBLIC) target_compile_features(device_conv2d_fwd_bias_activation_add_cpu_instance PUBLIC)
......
library/src/tensor_operation_instance/cpu/conv2d_fwd_bias_activation_add/device_conv2d_bias_activation_add_avx2_nhwc_kyxck8_nhwk_instance.cpp 0 → 100644
View file @ dc536427
#include <stdlib.h>
#include "config.hpp"
#include "convolution_forward_specialization_cpu.hpp"
#include "device_convnd_fwd_bias_activation_add_avx2_nhwc_kyxck8_nhwk.hpp"
#include "element_wise_operation_cpu.hpp"
#include "device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace cpu {
namespace device {
namespace device_conv2d_fwd_bias_activation_add_avx2_instance {
using InType = float;
using WeiType = float;
using OutType = float;
using AccType = float;
using InLayout = ck::tensor_layout::gemm::RowMajor; // NHWC
using WeiLayout = ck::tensor_layout::gemm::ColumnMajor; // KYXCK8
static constexpr bool NonTemporalStore = false;
using PT = ck::tensor_operation::cpu::element_wise::PassThrough;
using AddReluAdd = ck::tensor_operation::cpu::element_wise::AddReluAdd;
static constexpr auto ConvFwdDefault =
ck::tensor_operation::cpu::device::ConvolutionForwardSpecialization_t::Default;
static constexpr auto ConvFwd1x1P0 =
ck::tensor_operation::cpu::device::ConvolutionForwardSpecialization_t::Filter1x1Pad0;
static constexpr auto ConvFwd1x1S1P0 =
ck::tensor_operation::cpu::device::ConvolutionForwardSpecialization_t::Filter1x1Stride1Pad0;
static constexpr auto DefaultGemmKLoop =
ck::tensor_operation::cpu::device::ConvolutionForwardGemmKSpecialization_t::DefaultGemmKLoop;
static constexpr auto GemmKLoopOverC =
ck::tensor_operation::cpu::device::ConvolutionForwardGemmKSpecialization_t::NHWC_GemmKLoopOverC;
static constexpr auto LoopOver_MNK = ck::tensor_operation::cpu::device::LoopOver_MNK;
static constexpr auto LoopOver_MKN = ck::tensor_operation::cpu::device::LoopOver_MKN;
// clang-format off
#define DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXCK8_NHWK_F32(a_elem_op, b_elem_op, c_elem_op, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, a_local_buf, b_local_buf, c_local_buf, bias_along_m) \
DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K<float , float , float, float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwdDefault, GemmKLoopOverC , LoopOver_MNK, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, a_local_buf, b_local_buf, c_local_buf, bias_along_m>, \
DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K<float , float , float, float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1S1P0, GemmKLoopOverC , LoopOver_MNK, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, a_local_buf, b_local_buf, c_local_buf, bias_along_m>, \
DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K<float , float , float, float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwdDefault, DefaultGemmKLoop, LoopOver_MNK, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, a_local_buf, b_local_buf, c_local_buf, bias_along_m>, \
DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K<float , float , float, float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1S1P0, DefaultGemmKLoop, LoopOver_MNK, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, a_local_buf, b_local_buf, c_local_buf, bias_along_m>, \
DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K<float , float , float, float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwdDefault, GemmKLoopOverC , LoopOver_MKN, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, a_local_buf, b_local_buf, c_local_buf, bias_along_m>, \
DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K<float , float , float, float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1S1P0, GemmKLoopOverC , LoopOver_MKN, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, a_local_buf, b_local_buf, c_local_buf, bias_along_m>, \
DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K<float , float , float, float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwdDefault, DefaultGemmKLoop, LoopOver_MKN, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, a_local_buf, b_local_buf, c_local_buf, bias_along_m>, \
DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K<float , float , float, float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1S1P0, DefaultGemmKLoop, LoopOver_MKN, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, a_local_buf, b_local_buf, c_local_buf, bias_along_m>
// clang-format on
using device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxck8_nhwk_f32_instances = std::tuple<
// clang-format off
DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, AddReluAdd, 256, 128, 64, 6, 16, true, true, false, false),
DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, AddReluAdd, 256, 128, 128, 6, 16, true, true, false, false),
DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, AddReluAdd, 128, 256, 128, 6, 16, true, true, false, false),
DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, AddReluAdd, 512, 240, 128, 4, 24, true, true, false, false),
DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, AddReluAdd, 512, 256, 128, 6, 16, true, true, false, false),
DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, AddReluAdd, 768, 320, 128, 6, 16, true, true, false, false),
DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, AddReluAdd, 896, 352, 128, 6, 16, true, true, false, false),
DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, AddReluAdd, 1024, 416, 128, 6, 16, true, true, false, false)>;
// clang-format on
// use this in single thread, but gemm_n is not multiple of 8
using device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxck8_nhwk_f32_local_c_instances =
std::tuple<
// clang-format off
DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, AddReluAdd, 256, 128, 64, 6, 16, true, true, true, false),
DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, AddReluAdd, 256, 128, 128, 6, 16, true, true, true, false),
DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, AddReluAdd, 128, 256, 128, 6, 16, true, true, true, false),
DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, AddReluAdd, 512, 240, 128, 4, 24, true, true, true, false),
DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, AddReluAdd, 512, 256, 128, 6, 16, true, true, true, false),
DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, AddReluAdd, 768, 320, 128, 6, 16, true, true, true, false),
DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, AddReluAdd, 896, 352, 128, 6, 16, true, true, true, false),
DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, AddReluAdd, 1024, 416, 128, 6, 16, true, true, true, false)>;
// clang-format on
// use this in multi thread environment (need local C buffer to avoid cache coherence, although some
// time no local c is better...)
using device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxck8_nhwk_f32_mt_instances = std::tuple<
// clang-format off
DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, AddReluAdd, 48, 24, 128, 4, 24, true, true, true, false),
DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, AddReluAdd, 72, 16, 128, 6, 16, true, true, true, false),
DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, AddReluAdd, 72, 32, 128, 6, 16, true, true, true, false),
DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, AddReluAdd, 96, 32, 128, 6, 16, true, true, true, false),
DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, AddReluAdd, 96, 64, 128, 6, 16, true, true, true, false),
DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, AddReluAdd, 120, 32, 128, 6, 16, true, true, true, false),
DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, AddReluAdd, 120, 64, 128, 6, 16, true, true, true, false),
// DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, PT, 256, 128, 64, 6, 16, true, true, true),
DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, AddReluAdd, 256, 128, 128, 6, 16, true, true, true, false),
DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, AddReluAdd, 128, 256, 128, 6, 16, true, true, true, false),
DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, AddReluAdd, 512, 240, 128, 4, 24, true, true, true, false),
DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, AddReluAdd, 512, 256, 128, 6, 16, true, true, true, false),
DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, AddReluAdd, 768, 320, 128, 6, 16, true, true, true, false),
DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, AddReluAdd, 896, 352, 128, 6, 16, true, true, true, false),
DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, AddReluAdd, 1024, 416, 128, 6, 16, true, true, true, false)>;
// clang-format on
void add_device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxck8_nhwk(
std::vector<DeviceConvFwdBiasActivationAddPtr<PT, PT, AddReluAdd>>& instances)
{
ck::tensor_operation::device::add_device_operation_instances(
instances, device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxck8_nhwk_f32_instances{});
}
void add_device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxck8_nhwk_local_c(
std::vector<DeviceConvFwdBiasActivationAddPtr<PT, PT, AddReluAdd>>& instances)
{
ck::tensor_operation::device::add_device_operation_instances(
instances,
device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxck8_nhwk_f32_local_c_instances{});
}
void add_device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxck8_nhwk_mt(
std::vector<DeviceConvFwdBiasActivationAddPtr<PT, PT, AddReluAdd>>& instances)
{
ck::tensor_operation::device::add_device_operation_instances(
instances, device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxck8_nhwk_f32_mt_instances{});
}
} // namespace device_conv2d_fwd_bias_activation_add_avx2_instance
} // namespace device
} // namespace cpu
} // namespace tensor_operation
} // namespace ck
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment