ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-11 08:50:17 +00:00
Code Issues Packages Projects Releases Wiki Activity

cleaning up

Browse Source
This commit is contained in:
Chao Liu
2019-08-09 22:48:28 -05:00
parent 4908fe3fdc
commit 1c4ef23cff
3 changed files with 22 additions and 21 deletions
Download Patch File Download Diff File Expand all files Collapse all files

23
composable_kernel/include/kernel_algorithm/gridwise_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw_lds_double_buffer.hpp
View File

@@ -1,5 +1,5 @@
#ifndef CK_GRIDWISE_CONVOLUTION_IMPLICIT_GEMM_V4R1_NCHW_KCYX_NKHW_LDS_DOUBLE_BUFFER
#define CK_GRIDWISE_CONVOLUTION_IMPLICIT_GEMM_V4R1_NCHW_KCYX_NKHW_LDS_DOUBLE_BUFFER
#ifndef CK_GRIDWISE_CONVOLUTION_IMPLICIT_GEMM_V4R1_NCHW_KCYX_NKHW_LDS_DOUBLE_BUFFER_HPP
#define CK_GRIDWISE_CONVOLUTION_IMPLICIT_GEMM_V4R1_NCHW_KCYX_NKHW_LDS_DOUBLE_BUFFER_HPP
#include "common_header.hpp"
#include "ConstantTensorDescriptor.hpp"
@@ -23,8 +23,7 @@ template <index_t GridSize,
index_t BPerBlock,
index_t KPerBlock,
index_t EPerBlock,
index_t N1,
index_t N2,
index_t GemmNRepeat,
index_t GemmMPerThreadSubC,
index_t GemmNPerThreadSubC,
index_t GemmMLevel0Cluster,
@@ -47,17 +46,19 @@ template <index_t GridSize,
class WeiBlockCopySrcAccessOrder,
class WeiBlockCopyDstAccessOrder,
index_t WeiBlockCopySrcDataPerRead_E,
index_t WeiBlockCopyDstDataPerWrite_K,
index_t OutThreadCopyDataPerAccess_W>
index_t WeiBlockCopyDstDataPerWrite_K>
struct GridwiseConvolutionImplicitGemm_v4r1_nchw_kcyx_nkhw_lds_double_buffer
{
__device__ void Run(const Float* const __restrict__ p_in_global,
const Float* const __restrict__ p_wei_global,
Float* const __restrict__ p_out_global) const
__device__ void __launch_bounds__(BlockSize, 2)
Run(const Float* const __restrict__ p_in_global,
const Float* const __restrict__ p_wei_global,
Float* const __restrict__ p_out_global) const
{
// this is a mess
// TODO: find more elegent way of specifying (or calculating) performance parameters
static_assert(N2 == GemmNPerThreadSubC, "wrong!");
constexpr index_t N1 = GemmNRepeat;
constexpr index_t N2 = GemmNPerThreadSubC;
static_assert((N1 * N2 * BPerBlock) %
(GemmNPerThreadSubC * GemmNLevel0Cluster * GemmNLevel1Cluster) ==
0,
@@ -464,4 +465,4 @@ struct GridwiseConvolutionImplicitGemm_v4r1_nchw_kcyx_nkhw_lds_double_buffer
};
} // namespace ck
#endif
#endif // CK_GRIDWISE_CONVOLUTION_IMPLICIT_GEMM_V4R1_NCHW_KCYX_NKHW_LDS_DOUBLE_BUFFER_HPP

18
driver/include/device_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw.hpp
View File

@@ -54,11 +54,6 @@ void device_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw(InDesc,
wei_kcyx_device_buf.ToDevice(wei_kcyx.mData.data());
out_nkhw_device_buf.ToDevice(out_nkhw.mData.data());
constexpr index_t N1 = 2;
constexpr index_t N2 = 4;
constexpr index_t B = (N * Ho * Wo) / (N1 * N2);
#if 1
// each thread hold 64 data
constexpr index_t BlockSize = 256;
@@ -67,6 +62,8 @@ void device_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw(InDesc,
constexpr index_t KPerBlock = 128;
constexpr index_t EPerBlock = 8;
constexpr index_t GemmNRepeat = 2;
constexpr index_t GemmMPerThreadSubC = 4;
constexpr index_t GemmNPerThreadSubC = 4;
constexpr index_t GemmMLevel0Cluster = 4;
@@ -168,6 +165,11 @@ void device_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw(InDesc,
constexpr index_t WeiBlockCopyDstDataPerWrite_K = 1;
#endif
constexpr index_t N1 = GemmNRepeat;
constexpr index_t N2 = GemmNPerThreadSubC;
constexpr index_t B = (N * Ho * Wo) / (N1 * N2);
constexpr index_t GridSize =
((B + BPerBlock - 1) / BPerBlock) * ((K + KPerBlock - 1) / KPerBlock);
@@ -192,8 +194,7 @@ void device_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw(InDesc,
BPerBlock,
KPerBlock,
EPerBlock,
N1,
N2,
GemmNRepeat,
GemmMPerThreadSubC,
GemmNPerThreadSubC,
GemmMLevel0Cluster,
@@ -216,8 +217,7 @@ void device_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw(InDesc,
WeiBlockCopySrcAccessOrder,
WeiBlockCopyDstAccessOrder,
WeiBlockCopySrcDataPerRead_E,
WeiBlockCopyDstDataPerWrite_K,
OutThreadCopyDataPerAccess_W>{};
WeiBlockCopyDstDataPerWrite_K>{};
float time = launch_kernel(run_gridwise_convolution_kernel<decltype(gridwise_conv), T>,
dim3(GridSize),

2
driver/src/driver.cpp
View File

@@ -379,7 +379,7 @@ int main(int argc, char* argv[])
#elif 0
device_convolution_implicit_gemm_v3_nchw_cyxk_nkhw(
(in_nchw_desc, in_nchw, wei_kcyx_desc, wei_kcyx, out_nkhw_desc, out_nkhw_device, nrepeat);
#elif 0
#elif 1
device_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw(in_nchw_desc,
in_nchw,
wei_kcyx_desc,