mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-11 17:00:18 +00:00
rename files, added header guard, added namespace
This commit is contained in:
Chao Liu
@@ -1,9 +1,11 @@
|
||||
#pragma once
|
||||
#include <unistd.h>
|
||||
#include "device.hpp"
|
||||
#include "gridwise_convolution_wrapper.hpp"
|
||||
#include "gridwise_convolution_kernel_wrapper.hpp"
|
||||
#include "gridwise_convolution_direct_v2_nchw_kcyx_nkhw.hpp"
|
||||
|
||||
using namespace ck;
|
||||
|
||||
template <class T, class InDesc, class WeiDesc, class OutDesc>
|
||||
void device_convolution_direct_v2_nchw_kcyx_nkhw(InDesc,
|
||||
const Tensor<T>& in,
|
||||
@@ -79,7 +81,7 @@ void device_convolution_direct_v2_nchw_kcyx_nkhw(InDesc,
|
||||
WoPerThread,
|
||||
InBlockCopyDataPerRead,
|
||||
WeiBlockCopyDataPerRead>;
|
||||
float time = launch_kernel(run_gridwise_convolution<gridwise_conv, T>,
|
||||
float time = launch_kernel(run_gridwise_convolution_kernel<gridwise_conv, T>,
|
||||
dim3(GridSize),
|
||||
dim3(BlockSize),
|
||||
0,
|
||||
|
||||
@@ -1,11 +1,13 @@
|
||||
#pragma once
|
||||
#include <unistd.h>
|
||||
#include "device.hpp"
|
||||
#include "gridwise_convolution_wrapper.hpp"
|
||||
#include "gridwise_convolution_kernel_wrapper.hpp"
|
||||
#include "gridwise_convolution_implicit_gemm_v1r1_chwn_cyxk_khwn.hpp"
|
||||
#include "gridwise_convolution_implicit_gemm_v1r2_chwn_cyxk_khwn.hpp"
|
||||
#include "gridwise_convolution_implicit_gemm_v1r3_chwn_cyxk_khwn.hpp"
|
||||
#include "gridwise_convolution_implicit_gemm_v1r3_lds_double_buffer_chwn_cyxk_khwn.hpp"
|
||||
#include "gridwise_convolution_implicit_gemm_v1r3_chwn_cyxk_khwn_lds_double_buffer.hpp"
|
||||
|
||||
using namespace ck;
|
||||
|
||||
template <class T, class InDesc, class WeiDesc, class OutDesc>
|
||||
void device_convolution_implicit_gemm_v1_chwn_cyxk_khwn(InDesc,
|
||||
@@ -478,7 +480,7 @@ void device_convolution_implicit_gemm_v1_chwn_cyxk_khwn(InDesc,
|
||||
#elif 0
|
||||
GridwiseConvolutionImplicitGemm_v1r3_chwn_cyxk_khwn
|
||||
#elif 1
|
||||
GridwiseConvolutionImplicitGemm_v1r3_lds_double_buffer_chwn_cyxk_khwn
|
||||
GridwiseConvolutionImplicitGemm_v1r3_chwn_cyxk_khwn_lds_double_buffer
|
||||
#endif
|
||||
<GridSize,
|
||||
BlockSize,
|
||||
@@ -509,7 +511,7 @@ void device_convolution_implicit_gemm_v1_chwn_cyxk_khwn(InDesc,
|
||||
WeiBlockCopyDataPerRead_K,
|
||||
OutThreadCopyDataPerWrite_N>{};
|
||||
|
||||
float time = launch_kernel(run_gridwise_convolution<decltype(gridwise_conv), T>,
|
||||
float time = launch_kernel(run_gridwise_convolution_kernel<decltype(gridwise_conv), T>,
|
||||
dim3(GridSize),
|
||||
dim3(BlockSize),
|
||||
0,
|
||||
|
||||
@@ -1,282 +0,0 @@
|
||||
#pragma once
|
||||
#include <unistd.h>
|
||||
#include "device.hpp"
|
||||
#include "gridwise_convolution_wrapper.hpp"
|
||||
#include "gridwise_convolution_implicit_gemm_v1r2_nchw_cyxk_khwn.hpp"
|
||||
#include "gridwise_convolution_implicit_gemm_v1r3_nchw_cyxk_khwn.hpp"
|
||||
#include "gridwise_convolution_implicit_gemm_v1r3_lds_double_buffer_nchw_cyxk_khwn.hpp"
|
||||
|
||||
template <class T, class InDesc, class WeiDesc, class OutDesc>
|
||||
void device_convolution_implicit_gemm_v1_nchw_cyxk_khwn(InDesc,
|
||||
const Tensor<T>& in_nchw,
|
||||
WeiDesc,
|
||||
const Tensor<T>& wei_kcyx,
|
||||
OutDesc,
|
||||
Tensor<T>& out_nkhw,
|
||||
index_t nrepeat)
|
||||
{
|
||||
constexpr auto I0 = Number<0>{};
|
||||
constexpr auto I1 = Number<1>{};
|
||||
constexpr auto I2 = Number<2>{};
|
||||
constexpr auto I3 = Number<3>{};
|
||||
|
||||
constexpr auto in_nchw_desc = InDesc{};
|
||||
constexpr auto wei_kcyx_desc = WeiDesc{};
|
||||
constexpr auto out_nkhw_desc = OutDesc{};
|
||||
|
||||
constexpr index_t Hi = in_nchw_desc.GetLength(I2);
|
||||
constexpr index_t Wi = in_nchw_desc.GetLength(I3);
|
||||
|
||||
constexpr index_t N = out_nkhw_desc.GetLength(I0);
|
||||
constexpr index_t Ho = out_nkhw_desc.GetLength(I2);
|
||||
constexpr index_t Wo = out_nkhw_desc.GetLength(I3);
|
||||
|
||||
constexpr index_t K = wei_kcyx_desc.GetLength(I0);
|
||||
constexpr index_t C = wei_kcyx_desc.GetLength(I1);
|
||||
constexpr index_t Y = wei_kcyx_desc.GetLength(I2);
|
||||
constexpr index_t X = wei_kcyx_desc.GetLength(I3);
|
||||
|
||||
// reorder weight
|
||||
auto wei_cyxk_desc = make_ConstantTensorDescriptor_packed(Sequence<C, Y, X, K>{});
|
||||
ostream_ConstantTensorDescriptor(wei_cyxk_desc, std::cout << "wei_cyxk_desc: ");
|
||||
|
||||
Tensor<T> wei_cyxk(make_TensorDescriptor(wei_cyxk_desc));
|
||||
|
||||
auto f_reorder_kcyx2cyxk = [&](auto k, auto c, auto y, auto x) {
|
||||
wei_cyxk(c, y, x, k) = wei_kcyx(k, c, y, x);
|
||||
};
|
||||
|
||||
make_ParallelTensorFunctor(f_reorder_kcyx2cyxk, K, C, Y, X)(
|
||||
std::thread::hardware_concurrency());
|
||||
|
||||
// output
|
||||
auto out_khwn_desc = make_ConstantTensorDescriptor_packed(Sequence<K, Ho, Wo, N>{});
|
||||
ostream_ConstantTensorDescriptor(out_khwn_desc, std::cout << "out_khwn_desc: ");
|
||||
|
||||
Tensor<T> out_khwn(make_TensorDescriptor(out_khwn_desc));
|
||||
|
||||
std::size_t data_sz = sizeof(T);
|
||||
DeviceMem in_nchw_device_buf(data_sz * in_nchw.mDesc.GetElementSpace());
|
||||
DeviceMem wei_cyxk_device_buf(data_sz * wei_cyxk.mDesc.GetElementSpace());
|
||||
DeviceMem out_khwn_device_buf(data_sz * out_khwn.mDesc.GetElementSpace());
|
||||
|
||||
in_nchw_device_buf.ToDevice(in_nchw.mData.data());
|
||||
wei_cyxk_device_buf.ToDevice(wei_cyxk.mData.data());
|
||||
out_khwn_device_buf.ToDevice(out_khwn.mData.data());
|
||||
|
||||
#if 1
|
||||
// for 3x3, 34x34, v1r3, Pascal
|
||||
constexpr index_t BlockSize = 128;
|
||||
|
||||
constexpr index_t NPerBlock = 2;
|
||||
constexpr index_t KPerBlock = 128;
|
||||
constexpr index_t CPerBlock = 8;
|
||||
constexpr index_t HoPerBlock = 2;
|
||||
constexpr index_t WoPerBlock = 16;
|
||||
|
||||
constexpr index_t NPerThread = 2;
|
||||
constexpr index_t KPerThread = 8;
|
||||
constexpr index_t HoPerThread = 1;
|
||||
constexpr index_t WoPerThread = 4;
|
||||
|
||||
constexpr index_t GemmMPerThreadSubC = 4;
|
||||
constexpr index_t GemmNPerThreadSubC = 4;
|
||||
constexpr index_t GemmMLevel0Cluster = 4;
|
||||
constexpr index_t GemmNLevel0Cluster = 2;
|
||||
constexpr index_t GemmMLevel1Cluster = 4;
|
||||
constexpr index_t GemmNLevel1Cluster = 2;
|
||||
constexpr index_t GemmKPerThreadLoop = 1;
|
||||
constexpr index_t GemmDataPerReadA = 4;
|
||||
constexpr index_t GemmDataPerReadB = 4;
|
||||
|
||||
using InBlockReorderSrcSubLengths_NCHW = Sequence<2, 1, 2, 1>;
|
||||
using InBlockReorderSrcClusterLengths_NCHW = Sequence<1, 8, 1, 16>;
|
||||
using InBlockReorderMapThreadCluster2SrcCluster_CHNW2NCHW = Sequence<1, 2, 0, 3>;
|
||||
constexpr index_t InBlockReorderDataPerRead_W = 1; // v1r3 cannot do vector load input for NCHW
|
||||
constexpr index_t InBlockReorderDataPerWrite_N = 2;
|
||||
|
||||
using WeiBlockCopyClusterLengths = void;
|
||||
constexpr index_t WeiBlockCopyDataPerRead_K = 4;
|
||||
|
||||
constexpr index_t OutThreadCopyDataPerWrite_N = 2;
|
||||
#elif 1
|
||||
// for 3x3, 34x34, v1r3, Vega 20
|
||||
constexpr index_t BlockSize = 256;
|
||||
|
||||
constexpr index_t NPerBlock = 2;
|
||||
constexpr index_t KPerBlock = 128;
|
||||
constexpr index_t CPerBlock = 8;
|
||||
constexpr index_t HoPerBlock = 4;
|
||||
constexpr index_t WoPerBlock = 16;
|
||||
|
||||
constexpr index_t NPerThread = 2;
|
||||
constexpr index_t KPerThread = 8;
|
||||
constexpr index_t HoPerThread = 1;
|
||||
constexpr index_t WoPerThread = 4;
|
||||
|
||||
constexpr index_t GemmMPerThreadSubC = 4;
|
||||
constexpr index_t GemmNPerThreadSubC = 4;
|
||||
constexpr index_t GemmMLevel0Cluster = 4;
|
||||
constexpr index_t GemmNLevel0Cluster = 2;
|
||||
constexpr index_t GemmMLevel1Cluster = 4;
|
||||
constexpr index_t GemmNLevel1Cluster = 2;
|
||||
constexpr index_t GemmKPerThreadLoop = 1;
|
||||
constexpr index_t GemmDataPerReadA = 4;
|
||||
constexpr index_t GemmDataPerReadB = 4;
|
||||
|
||||
using InBlockReorderSrcSubLengths_NCHW = Sequence<2, 1, 2, 1>;
|
||||
using InBlockReorderSrcClusterLengths_NCHW = Sequence<1, 8, 2, 16>;
|
||||
using InBlockReorderMapThreadCluster2SrcCluster_CHNW2NCHW = Sequence<1, 2, 0, 3>;
|
||||
constexpr index_t InBlockReorderDataPerRead_W = 1; // v1r3 cannot do vector load input for NCHW
|
||||
constexpr index_t InBlockReorderDataPerWrite_N = 2;
|
||||
|
||||
using WeiBlockCopyClusterLengths = void;
|
||||
constexpr index_t WeiBlockCopyDataPerRead_K = 4;
|
||||
|
||||
constexpr index_t OutThreadCopyDataPerWrite_N = 2;
|
||||
#elif 0
|
||||
// for 3x3, 28x28, v1r2, Pascal
|
||||
constexpr index_t BlockSize = 128;
|
||||
|
||||
constexpr index_t NPerBlock = 16;
|
||||
constexpr index_t KPerBlock = 128;
|
||||
constexpr index_t CPerBlock = 8;
|
||||
constexpr index_t HoPerBlock = 2;
|
||||
constexpr index_t WoPerBlock = 2;
|
||||
|
||||
constexpr index_t NPerThread = 4;
|
||||
constexpr index_t KPerThread = 8;
|
||||
constexpr index_t HoPerThread = 1;
|
||||
constexpr index_t WoPerThread = 2;
|
||||
|
||||
constexpr index_t GemmMPerThreadSubC = 4;
|
||||
constexpr index_t GemmNPerThreadSubC = 4;
|
||||
constexpr index_t GemmMLevel0Cluster = 4;
|
||||
constexpr index_t GemmNLevel0Cluster = 2;
|
||||
constexpr index_t GemmMLevel1Cluster = 4;
|
||||
constexpr index_t GemmNLevel1Cluster = 2;
|
||||
constexpr index_t GemmKPerThreadLoop = 1;
|
||||
constexpr index_t GemmDataPerReadA = 4;
|
||||
constexpr index_t GemmDataPerReadB = 4;
|
||||
|
||||
using InBlockReorderSrcSubLengths_NCHW = Sequence<4, 1, 1, 2>;
|
||||
using InBlockReorderSrcClusterLengths_NCHW = Sequence<4, 8, 2, 2>;
|
||||
using InBlockReorderMapThreadCluster2SrcCluster_CHNW2NCHW = Sequence<1, 2, 0, 3>;
|
||||
constexpr index_t InBlockReorderDataPerRead_W = 2;
|
||||
constexpr index_t InBlockReorderDataPerWrite_N = 4;
|
||||
|
||||
using WeiBlockCopyClusterLengths = Sequence<4, 1, 32>;
|
||||
constexpr index_t WeiBlockCopyDataPerRead_K = 4;
|
||||
|
||||
constexpr index_t OutThreadCopyDataPerWrite_N = 2;
|
||||
#elif 0
|
||||
// for 3x3, 28x28, v1r3, Pascal, bad
|
||||
constexpr index_t BlockSize = 128;
|
||||
|
||||
constexpr index_t NPerBlock = 16;
|
||||
constexpr index_t KPerBlock = 128;
|
||||
constexpr index_t CPerBlock = 8;
|
||||
constexpr index_t HoPerBlock = 2;
|
||||
constexpr index_t WoPerBlock = 2;
|
||||
|
||||
constexpr index_t NPerThread = 4;
|
||||
constexpr index_t KPerThread = 8;
|
||||
constexpr index_t HoPerThread = 1;
|
||||
constexpr index_t WoPerThread = 2;
|
||||
|
||||
constexpr index_t GemmMPerThreadSubC = 4;
|
||||
constexpr index_t GemmNPerThreadSubC = 4;
|
||||
constexpr index_t GemmMLevel0Cluster = 4;
|
||||
constexpr index_t GemmNLevel0Cluster = 2;
|
||||
constexpr index_t GemmMLevel1Cluster = 4;
|
||||
constexpr index_t GemmNLevel1Cluster = 2;
|
||||
constexpr index_t GemmKPerThreadLoop = 1;
|
||||
constexpr index_t GemmDataPerReadA = 4;
|
||||
constexpr index_t GemmDataPerReadB = 4;
|
||||
|
||||
using InBlockReorderSrcSubLengths_NCHW = Sequence<4, 1, 1, 1>;
|
||||
using InBlockReorderSrcClusterLengths_NCHW = Sequence<4, 8, 2, 2>;
|
||||
using InBlockReorderMapThreadCluster2SrcCluster_CHNW2NCHW = Sequence<1, 2, 0, 3>;
|
||||
constexpr index_t InBlockReorderDataPerRead_W = 1; // v1r3 cannot do vector load input for NCHW
|
||||
constexpr index_t InBlockReorderDataPerWrite_N = 1;
|
||||
|
||||
using WeiBlockCopyClusterLengths = void;
|
||||
constexpr index_t WeiBlockCopyDataPerRead_K = 4;
|
||||
|
||||
constexpr index_t OutThreadCopyDataPerWrite_N = 2;
|
||||
#endif
|
||||
|
||||
constexpr index_t GridSize =
|
||||
((N + NPerBlock - 1) / NPerBlock) * ((K + KPerBlock - 1) / KPerBlock) *
|
||||
((Ho + HoPerBlock - 1) / HoPerBlock) * ((Wo + WoPerBlock - 1) / WoPerBlock);
|
||||
|
||||
printf("%s: BlockSize %u, GridSize %u \n", __func__, BlockSize, GridSize);
|
||||
|
||||
for(index_t i = 0; i < nrepeat; ++i)
|
||||
{
|
||||
constexpr auto gridwise_conv =
|
||||
#if 0
|
||||
GridwiseConvolutionImplicitGemm_v1r2_nchw_cyxk_khwn
|
||||
#elif 0
|
||||
GridwiseConvolutionImplicitGemm_v1r3_nchw_cyxk_khwn
|
||||
#elif 1
|
||||
GridwiseConvolutionImplicitGemm_v1r3_lds_double_buffer_nchw_cyxk_khwn
|
||||
#endif
|
||||
<GridSize,
|
||||
BlockSize,
|
||||
T,
|
||||
decltype(in_nchw_desc),
|
||||
decltype(wei_cyxk_desc),
|
||||
decltype(out_khwn_desc),
|
||||
NPerBlock,
|
||||
KPerBlock,
|
||||
CPerBlock,
|
||||
HoPerBlock,
|
||||
WoPerBlock,
|
||||
NPerThread,
|
||||
KPerThread,
|
||||
HoPerThread,
|
||||
WoPerThread,
|
||||
GemmMPerThreadSubC,
|
||||
GemmNPerThreadSubC,
|
||||
GemmMLevel0Cluster,
|
||||
GemmNLevel0Cluster,
|
||||
GemmMLevel1Cluster,
|
||||
GemmNLevel1Cluster,
|
||||
GemmKPerThreadLoop,
|
||||
GemmDataPerReadA,
|
||||
GemmDataPerReadB,
|
||||
InBlockReorderSrcSubLengths_NCHW,
|
||||
InBlockReorderSrcClusterLengths_NCHW,
|
||||
InBlockReorderMapThreadCluster2SrcCluster_CHNW2NCHW,
|
||||
InBlockReorderDataPerRead_W,
|
||||
InBlockReorderDataPerWrite_N,
|
||||
WeiBlockCopyClusterLengths,
|
||||
WeiBlockCopyDataPerRead_K,
|
||||
OutThreadCopyDataPerWrite_N>{};
|
||||
|
||||
float time = launch_kernel(run_gridwise_convolution<decltype(gridwise_conv), T>,
|
||||
dim3(GridSize),
|
||||
dim3(BlockSize),
|
||||
0,
|
||||
static_cast<T*>(in_nchw_device_buf.GetDeviceBuffer()),
|
||||
static_cast<T*>(wei_cyxk_device_buf.GetDeviceBuffer()),
|
||||
static_cast<T*>(out_khwn_device_buf.GetDeviceBuffer()));
|
||||
|
||||
printf("Elapsed time : %f ms, %f TFlop/s\n",
|
||||
time,
|
||||
(float)calculate_convolution_flops(InDesc{}, WeiDesc{}, OutDesc{}) /
|
||||
(std::size_t(1000) * 1000 * 1000) / time);
|
||||
usleep(std::min(time * 1000, float(10000)));
|
||||
}
|
||||
|
||||
out_khwn_device_buf.FromDevice(out_khwn.mData.data());
|
||||
|
||||
// reorder output
|
||||
auto f_reorder_khwn2nkhw = [&](auto k, auto ho, auto wo, auto n) {
|
||||
out_nkhw(n, k, ho, wo) = out_khwn(k, ho, wo, n);
|
||||
};
|
||||
|
||||
make_ParallelTensorFunctor(f_reorder_khwn2nkhw, K, Ho, Wo, N)(
|
||||
std::thread::hardware_concurrency());
|
||||
}
|
||||
@@ -1,9 +1,11 @@
|
||||
#pragma once
|
||||
#include <unistd.h>
|
||||
#include "device.hpp"
|
||||
#include "gridwise_convolution_wrapper.hpp"
|
||||
#include "gridwise_convolution_kernel_wrapper.hpp"
|
||||
#include "gridwise_convolution_implicit_gemm_v1r3_nchw_cyxk_nkhw.hpp"
|
||||
#include "gridwise_convolution_implicit_gemm_v1r3_lds_double_buffer_nchw_cyxk_nkhw.hpp"
|
||||
#include "gridwise_convolution_implicit_gemm_v1r3_nchw_cyxk_nkhw_lds_double_buffer.hpp"
|
||||
|
||||
using namespace ck;
|
||||
|
||||
template <class T, class InDesc, class WeiDesc, class OutDesc>
|
||||
void device_convolution_implicit_gemm_v1_nchw_cyxk_nkhw(InDesc,
|
||||
@@ -313,10 +315,10 @@ void device_convolution_implicit_gemm_v1_nchw_cyxk_nkhw(InDesc,
|
||||
for(index_t i = 0; i < nrepeat; ++i)
|
||||
{
|
||||
constexpr auto gridwise_conv =
|
||||
#if 1
|
||||
#if 0
|
||||
GridwiseConvolutionImplicitGemm_v1r3_nchw_cyxk_nkhw
|
||||
#else
|
||||
GridwiseConvolutionImplicitGemm_v1r3_lds_double_buffer_nchw_cyxk_nkhw
|
||||
GridwiseConvolutionImplicitGemm_v1r3_nchw_cyxk_nkhw_lds_double_buffer
|
||||
#endif
|
||||
<GridSize,
|
||||
BlockSize,
|
||||
@@ -351,7 +353,7 @@ void device_convolution_implicit_gemm_v1_nchw_cyxk_nkhw(InDesc,
|
||||
WeiBlockCopyDataPerRead_K,
|
||||
OutThreadCopyDataPerWrite_W>{};
|
||||
|
||||
float time = launch_kernel(run_gridwise_convolution<decltype(gridwise_conv), T>,
|
||||
float time = launch_kernel(run_gridwise_convolution_kernel<decltype(gridwise_conv), T>,
|
||||
dim3(GridSize),
|
||||
dim3(BlockSize),
|
||||
0,
|
||||
|
||||
@@ -1,10 +1,12 @@
|
||||
#pragma once
|
||||
#include <unistd.h>
|
||||
#include "device.hpp"
|
||||
#include "gridwise_convolution_wrapper.hpp"
|
||||
#include "gridwise_convolution_kernel_wrapper.hpp"
|
||||
#include "gridwise_convolution_implicit_gemm_v2_chwn_cyxk_khwn.hpp"
|
||||
#include "gridwise_convolution_implicit_gemm_v2_chwn_cyxk_khwn_lds_double_buffer.hpp"
|
||||
|
||||
using namespace ck;
|
||||
|
||||
template <class T, class InDesc, class WeiDesc, class OutDesc>
|
||||
void device_convolution_implicit_gemm_v2_chwn_cyxk_khwn(InDesc,
|
||||
const Tensor<T>& in_nchw,
|
||||
@@ -303,7 +305,7 @@ void device_convolution_implicit_gemm_v2_chwn_cyxk_khwn(InDesc,
|
||||
WeiBlockCopyDataPerRead,
|
||||
OutThreadCopyDataPerWrite>{};
|
||||
|
||||
float time = launch_kernel(run_gridwise_convolution<decltype(gridwise_conv), T>,
|
||||
float time = launch_kernel(run_gridwise_convolution_kernel<decltype(gridwise_conv), T>,
|
||||
dim3(GridSize),
|
||||
dim3(BlockSize),
|
||||
0,
|
||||
|
||||
@@ -1,9 +1,11 @@
|
||||
#pragma once
|
||||
#include <unistd.h>
|
||||
#include "device.hpp"
|
||||
#include "gridwise_convolution_wrapper.hpp"
|
||||
#include "gridwise_convolution_kernel_wrapper.hpp"
|
||||
#include "gridwise_convolution_implicit_gemm_v3_nchw_cyxk_nkhw.hpp"
|
||||
#include "gridwise_convolution_implicit_gemm_v3_lds_double_buffer_nchw_cyxk_nkhw.hpp"
|
||||
#include "gridwise_convolution_implicit_gemm_v3_nchw_cyxk_nkhw_lds_double_buffer.hpp"
|
||||
|
||||
using namespace ck;
|
||||
|
||||
template <class T, class InDesc, class WeiDesc, class OutDesc>
|
||||
void device_convolution_implicit_gemm_v3_nchw_cyxk_nkhw(InDesc,
|
||||
@@ -102,7 +104,7 @@ void device_convolution_implicit_gemm_v3_nchw_cyxk_nkhw(InDesc,
|
||||
#if 0
|
||||
GridwiseConvolutionImplicitGemm_v3_nchw_cyxk_nkhw
|
||||
#else
|
||||
GridwiseConvolutionImplicitGemm_v3_lds_double_buffer_nchw_cyxk_nkhw
|
||||
GridwiseConvolutionImplicitGemm_v3_nchw_cyxk_nkhw_lds_double_buffer
|
||||
#endif
|
||||
<GridSize,
|
||||
BlockSize,
|
||||
@@ -133,7 +135,7 @@ void device_convolution_implicit_gemm_v3_nchw_cyxk_nkhw(InDesc,
|
||||
WeiBlockCopyDataPerAccess_K>{};
|
||||
|
||||
#if 1
|
||||
float time = launch_kernel(run_gridwise_convolution<decltype(gridwise_conv), T>,
|
||||
float time = launch_kernel(run_gridwise_convolution_kernel<decltype(gridwise_conv), T>,
|
||||
dim3(GridSize),
|
||||
dim3(BlockSize),
|
||||
0,
|
||||
|
||||
@@ -1,9 +1,11 @@
|
||||
#pragma once
|
||||
#include <unistd.h>
|
||||
#include "device.hpp"
|
||||
#include "gridwise_convolution_wrapper.hpp"
|
||||
#include "gridwise_convolution_kernel_wrapper.hpp"
|
||||
#include "gridwise_convolution_implicit_gemm_v4_nchw_kcyx_nkhw.hpp"
|
||||
#include "gridwise_convolution_implicit_gemm_v4_lds_double_buffer_nchw_kcyx_nkhw.hpp"
|
||||
#include "gridwise_convolution_implicit_gemm_v4_nchw_kcyx_nkhw_lds_double_buffer.hpp"
|
||||
|
||||
using namespace ck;
|
||||
|
||||
template <class T, class InDesc, class WeiDesc, class OutDesc>
|
||||
void device_convolution_implicit_gemm_v4_nchw_kcyx_nkhw(InDesc,
|
||||
@@ -96,7 +98,7 @@ void device_convolution_implicit_gemm_v4_nchw_kcyx_nkhw(InDesc,
|
||||
#if 0
|
||||
GridwiseConvolutionImplicitGemm_v4_nchw_kcyx_nkhw
|
||||
#else
|
||||
GridwiseConvolutionImplicitGemm_v4_lds_double_buffer_nchw_kcyx_nkhw
|
||||
GridwiseConvolutionImplicitGemm_v4_nchw_kcyx_nkhw_lds_double_buffer
|
||||
#endif
|
||||
<GridSize,
|
||||
BlockSize,
|
||||
@@ -133,7 +135,7 @@ void device_convolution_implicit_gemm_v4_nchw_kcyx_nkhw(InDesc,
|
||||
WeiBlockCopySrcDataPerRead_E,
|
||||
WeiBlockCopyDstDataPerWrite_K>{};
|
||||
|
||||
float time = launch_kernel(run_gridwise_convolution<decltype(gridwise_conv), T>,
|
||||
float time = launch_kernel(run_gridwise_convolution_kernel<decltype(gridwise_conv), T>,
|
||||
dim3(GridSize),
|
||||
dim3(BlockSize),
|
||||
0,
|
||||
|
||||
@@ -3,6 +3,8 @@
|
||||
#include "device.hpp"
|
||||
#include "gridwise_direct_convolution_2_vectorized_nchw_kcyx_nkhw.hpp"
|
||||
|
||||
using namespace ck;
|
||||
|
||||
template <class TInWei, class TOut, class InDesc, class WeiDesc, class OutDesc>
|
||||
void device_direct_convolution_2_vectorized_nchw_kcyx_nkhw(InDesc,
|
||||
const Tensor<TInWei>& in_nchw,
|
||||
|
||||
@@ -3,6 +3,8 @@
|
||||
#include "device.hpp"
|
||||
#include "gridwise_implicit_gemm_convolution_1_chwn_cyxk_khwn_padded.hpp"
|
||||
|
||||
using namespace ck;
|
||||
|
||||
template <class T, class InDesc, class WeiDesc, class OutDesc, class LowerPads, class UpperPads>
|
||||
void device_implicit_gemm_convolution_1_chwn_cyxk_khwn_padded(InDesc,
|
||||
const Tensor<T>& in_nchw,
|
||||
|
||||
@@ -3,19 +3,19 @@
|
||||
#include <initializer_list>
|
||||
#include <cstdlib>
|
||||
#include <stdlib.h>
|
||||
#include "config.h"
|
||||
#include "config.hpp"
|
||||
#include "tensor.hpp"
|
||||
#include "ConstantTensorDescriptor.hpp"
|
||||
#include "conv_common.hpp"
|
||||
#include "device_convolution_direct_v2_nchw_kcyx_nkhw.hpp"
|
||||
//#include "device_direct_convolution_2_vectorized_nchw_kcyx_nkhw.hpp"
|
||||
#include "device_convolution_implicit_gemm_v1_chwn_cyxk_khwn.hpp"
|
||||
#include "device_convolution_implicit_gemm_v1_nchw_cyxk_khwn.hpp"
|
||||
#include "device_convolution_implicit_gemm_v1_nchw_cyxk_nkhw.hpp"
|
||||
#include "device_convolution_implicit_gemm_v2_chwn_cyxk_khwn.hpp"
|
||||
#include "device_convolution_implicit_gemm_v3_nchw_cyxk_nkhw.hpp"
|
||||
#include "device_convolution_implicit_gemm_v4_nchw_kcyx_nkhw.hpp"
|
||||
|
||||
using namespace ck;
|
||||
|
||||
struct GeneratorTensor_1
|
||||
{
|
||||
template <class... Is>
|
||||
@@ -419,7 +419,7 @@ int main(int argc, char* argv[])
|
||||
|
||||
constexpr index_t HPad = 0;
|
||||
constexpr index_t WPad = 0;
|
||||
#elif 0
|
||||
#elif 1
|
||||
// 3x3, 34x34
|
||||
constexpr index_t N = 64;
|
||||
constexpr index_t C = 256;
|
||||
@@ -633,15 +633,9 @@ int main(int argc, char* argv[])
|
||||
|
||||
#if 1
|
||||
#if 0
|
||||
device_direct_convolution_1
|
||||
#elif 0
|
||||
device_convolution_direct_v2_nchw_kcyx_nkhw
|
||||
#elif 0
|
||||
device_direct_convolution_2_vectorized_nchw_kcyx_nkhw
|
||||
#elif 0
|
||||
device_convolution_implicit_gemm_v1_chwn_cyxk_khwn
|
||||
#elif 0
|
||||
device_convolution_implicit_gemm_v1_nchw_cyxk_khwn
|
||||
#elif 0
|
||||
device_convolution_implicit_gemm_v1_nchw_cyxk_nkhw
|
||||
#elif 0
|
||||
|
||||
Reference in New Issue
Block a user