Skip to content

GitLab

Commit ff6aacaf authored by wangshaojie6's avatar wangshaojie6
Browse files

try using half_4 to do ds_read

parent 45440f17
Hide whitespace changes
Inline Side-by-side
Showing with 5 additions and 5 deletions
include/ck/config.hpp
View file @ ff6aacaf
...@@ -107,7 +107,7 @@ ...@@ -107,7 +107,7 @@
// experimental feature: use __builtin_memcpy instead of pointer cast to access a vector from // experimental feature: use __builtin_memcpy instead of pointer cast to access a vector from
// pointer of scalar // pointer of scalar
#define CK_EXPERIMENTAL_USE_MEMCPY_FOR_VECTOR_ACCESS 1 #define CK_EXPERIMENTAL_USE_MEMCPY_FOR_VECTOR_ACCESS 0
// experimental feature: use __builtin_memcpy instead of union to do bit_cast // experimental feature: use __builtin_memcpy instead of union to do bit_cast
#define CK_EXPERIMENTAL_USE_MEMCPY_FOR_BIT_CAST 1 #define CK_EXPERIMENTAL_USE_MEMCPY_FOR_BIT_CAST 1
......
include/ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp
View file @ ff6aacaf
...@@ -312,8 +312,8 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1 ...@@ -312,8 +312,8 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
static constexpr auto c_thread_desc_ = make_naive_tensor_descriptor_packed( static constexpr auto c_thread_desc_ = make_naive_tensor_descriptor_packed(
make_tuple(Number<MRepeat>{}, Number<NRepeat>{}, xdlops_gemm.GetRegSizePerXdlops())); make_tuple(Number<MRepeat>{}, Number<NRepeat>{}, xdlops_gemm.GetRegSizePerXdlops()));
//static constexpr index_t A_K1_vec = A_K1;// / 2; static constexpr index_t A_K1_vec = A_K1 / 2;
//static constexpr index_t B_K1_vec = B_K1;// / 2; static constexpr index_t B_K1_vec = B_K1 / 2;
using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB, using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB,
FloatAB, FloatAB,
...@@ -322,7 +322,7 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1 ...@@ -322,7 +322,7 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
Sequence<1, 1, 1, KPerBlock>, Sequence<1, 1, 1, KPerBlock>,
Sequence<0, 1, 2, 3>, Sequence<0, 1, 2, 3>,
3, 3,
A_K1, A_K1_vec,
A_K1>; A_K1>;
using BThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB, using BThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB,
...@@ -332,7 +332,7 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1 ...@@ -332,7 +332,7 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
Sequence<1, 1, 1, KPerBlock>, Sequence<1, 1, 1, KPerBlock>,
Sequence<0, 1, 2, 3>, Sequence<0, 1, 2, 3>,
3, 3,
B_K1, B_K1_vec,
B_K1>; B_K1>;
AThreadCopy a_thread_copy_{CalculateAThreadOriginDataIndex()}; AThreadCopy a_thread_copy_{CalculateAThreadOriginDataIndex()};
......
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