ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-24 14:54:47 +00:00
Code Issues Packages Projects Releases Wiki Activity

Jing's ds_read inline asm

Browse Source
This commit is contained in:
Chao Liu
2019-03-28 19:46:29 -05:00
parent 766b0a9eaf
commit d6d9a8e4ce
6 changed files with 126 additions and 25 deletions
Download Patch File Download Diff File Expand all files Collapse all files

24
src/include/threadwise_gemm.hip.hpp
View File

@@ -10,7 +10,7 @@ __device__ void threadwise_matrix_copy(SrcMatrix,
constexpr auto src_mtx = SrcMatrix{};
constexpr auto dst_mtx = DstMatrix{};
#if 0
#if 1
for(index_t i = 0; i < NRow; ++i)
{
for(index_t j = 0; j < NCol; ++j)
@@ -21,7 +21,7 @@ __device__ void threadwise_matrix_copy(SrcMatrix,
p_dst[dst_index] = p_src[src_index];
}
}
#elif 1
#elif 0
static_assert(NCol == 4, "only for NCol == 4");
using vector_t = typename vector_type<Float, 4>::MemoryType;
@@ -31,15 +31,21 @@ __device__ void threadwise_matrix_copy(SrcMatrix,
const index_t src_index = src_mtx.Get1dIndex(i, 0);
const index_t dst_index = dst_mtx.Get1dIndex(i, 0);
#if 1
*(reinterpret_cast<vector_t*>(p_dst + dst_index)) =
*(reinterpret_cast<const vector_t*>(p_src + src_index));
#if 0
*(reinterpret_cast<vector_t*>(&p_dst[dst_index]) =
*(reinterpret_cast<const vector_t*>(&p_src[src_index]));
#elif 0
asm volatile("\n \
ds_read2_b64 %0, %1 offset1:1 \n \
s_waitcnt lgkmcnt(0)"
: "=v"(*(reinterpret_cast<vector_t*>(&p_dst[dst_index])))
: "v"(__to_local((void*)(&p_src[src_index]))));
#elif 1
asm volatile("\n \
ds_read_b128 %0, %1, offset:0 \n \
"
: "=v"(*(reinterpret_cast<vector_t*>(p_dst+dst_index)))
: "v"((uint32_t)(p_src + src_index)));
ds_read_b128 %0, %1 \n \
s_waitcnt lgkmcnt(0)"
: "=v"(*(reinterpret_cast<vector_t*>(&p_dst[dst_index])))
: "v"(__to_local((void*)(&p_src[src_index]))));
#endif
}
#endif