ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-12 01:10:17 +00:00
Code Issues Packages Projects Releases Wiki Activity

add

Browse Source
This commit is contained in:
Jing Zhang
2019-04-03 17:08:14 -05:00
parent b188c0d243
commit 6a3f3f951d
4 changed files with 17 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files

20
src/include/blockwise_gemm.hip.hpp
View File

@@ -404,10 +404,14 @@ struct BlockwiseGemmBlockABlockBThreadCTransANormalBNormalC_v2
outerProduct4x4(reg_a[1], reg_b[1], reg_c[9], reg_c[11], reg_c[13], reg_c[15]);
#else
int k = k_begin;
ds_read_b128(reg_a[0], a_loc, k * 512);
ds_read_b128(reg_b[0], b_loc, k * 256);
ds_read_b128(reg_b[1], b_loc, 128 + k * 256);
ds_read_b128(reg_a[1], a_loc, 256 + k * 512);
int lds_a_block_off = sizeof(Float) * M;
int lds_b_block_off = sizeof(Float) * N;
int lds_a_block_off_1 = MPerLevel1Cluster * sizeof(Float);
int lds_b_block_off_1 = NPerLevel1Cluster * sizeof(Float);
ds_read_b128(reg_a[0], a_loc, k * lds_a_block_off);
ds_read_b128(reg_b[0], b_loc, k * lds_b_block_off);
ds_read_b128(reg_b[1], b_loc, lds_b_block_off_1 + k * lds_b_block_off);
ds_read_b128(reg_a[1], a_loc, lds_a_block_off_1 + k * lds_a_block_off);
lgkmcnt(2);
outerProduct4x4(reg_a[0], reg_b[0], reg_c[0], reg_c[2], reg_c[4], reg_c[6]);
lgkmcnt(1);
@@ -416,12 +420,12 @@ struct BlockwiseGemmBlockABlockBThreadCTransANormalBNormalC_v2
for(int i = 0; i < k_chunk - 1; i++)
{
k = k + 1;
ds_read_b128(reg_a[0], a_loc, k * 512);
ds_read_b128(reg_a[0], a_loc, k * lds_a_block_off);
outerProduct4x4(reg_a[1], reg_b[0], reg_c[8], reg_c[10], reg_c[12], reg_c[14]);
ds_read_b128(reg_b[0], b_loc, k * 256);
ds_read_b128(reg_b[0], b_loc, k * lds_b_block_off);
outerProduct4x4(reg_a[1], reg_b[1], reg_c[9], reg_c[11], reg_c[13], reg_c[15]);
ds_read_b128(reg_b[1], b_loc, 128 + k * 256);
ds_read_b128(reg_a[1], a_loc, 256 + k * 512);
ds_read_b128(reg_b[1], b_loc, lds_b_block_off_1 + k * lds_b_block_off);
ds_read_b128(reg_a[1], a_loc, lds_a_block_off_1 + k * lds_a_block_off);
lgkmcnt(2);
outerProduct4x4(reg_a[0], reg_b[0], reg_c[0], reg_c[2], reg_c[4], reg_c[6]);
lgkmcnt(1);

3
src/include/gridwise_implicit_gemm_convolution_2_chwn_cyxk_khwn.hip.hpp
View File

@@ -297,8 +297,7 @@ class gridwise_implicit_gemm_convolution_2_chwn_cyxk_khwn
for(index_t c_block_data_begin = 0; c_block_data_begin < C; c_block_data_begin += CPerBlock,
p_in_global_block_offset += CPerBlock * in_cb_global_desc.GetStride(I0),
p_wei_global_block_offset += CPerBlock * wei_cyxk_global_desc.GetStride(I0),
__syncthreads())
p_wei_global_block_offset += CPerBlock * wei_cyxk_global_desc.GetStride(I0), __syncthreads())
{
// load data
blockwise_in_copy.Run(p_in_global_block_offset, p_in_block);