Skip to content

GitLab

Commit a65d6459 authored by Po-Yen, Chen's avatar Po-Yen, Chen
Browse files

Remove redundant descriptor parameter

parent 542451d8
Hide whitespace changes
Inline Side-by-side
Showing with 25 additions and 36 deletions
include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl.hpp
View file @ a65d6459
......@@ -252,7 +252,6 @@ struct DeviceGemmXdl : public DeviceGemm<ALayout,
a_grid_desc_k0_m_k1_{},
b_grid_desc_k0_n_k1_{},
c_grid_desc_m_n_{},
c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_{},
block_2_ctile_map_{},
M01_{M01},
N01_{N01},
......@@ -264,15 +263,6 @@ struct DeviceGemmXdl : public DeviceGemm<ALayout,
block_2_ctile_map_ =
GridwiseGemm::MakeDefaultBlock2CTileMap(c_grid_desc_m_n_, M01, N01);
if(GridwiseGemm::CheckValidity(a_grid_desc_k0_m_k1_,
b_grid_desc_k0_n_k1_,
c_grid_desc_m_n_,
block_2_ctile_map_))
{
c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_ =
GridwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_grid_desc_m_n_);
}
}
// private:
......@@ -282,8 +272,6 @@ struct DeviceGemmXdl : public DeviceGemm<ALayout,
AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1_;
BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1_;
CGridDesc_M_N c_grid_desc_m_n_;
typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2
c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_;
typename GridwiseGemm::DefaultBlock2CTileMap block_2_ctile_map_;
index_t M01_;
index_t N01_;
......@@ -337,7 +325,7 @@ struct DeviceGemmXdl : public DeviceGemm<ALayout,
CDataType,
remove_reference_t<DeviceGemmXdl::AGridDesc_K0_M_K1>,
remove_reference_t<DeviceGemmXdl::BGridDesc_K0_N_K1>,
remove_reference_t<typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
remove_reference_t<DeviceGemmXdl::CGridDesc_M_N>,
remove_reference_t<typename GridwiseGemm::DefaultBlock2CTileMap>,
true>;
......@@ -351,7 +339,7 @@ struct DeviceGemmXdl : public DeviceGemm<ALayout,
arg.p_c_grid_,
arg.a_grid_desc_k0_m_k1_,
arg.b_grid_desc_k0_n_k1_,
arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_,
arg.c_grid_desc_m_n_,
arg.block_2_ctile_map_);
}
else
......@@ -362,7 +350,7 @@ struct DeviceGemmXdl : public DeviceGemm<ALayout,
CDataType,
remove_reference_t<DeviceGemmXdl::AGridDesc_K0_M_K1>,
remove_reference_t<DeviceGemmXdl::BGridDesc_K0_N_K1>,
remove_reference_t<typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
remove_reference_t<DeviceGemmXdl::CGridDesc_M_N>,
remove_reference_t<typename GridwiseGemm::DefaultBlock2CTileMap>,
false>;
......@@ -376,7 +364,7 @@ struct DeviceGemmXdl : public DeviceGemm<ALayout,
arg.p_c_grid_,
arg.a_grid_desc_k0_m_k1_,
arg.b_grid_desc_k0_n_k1_,
arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_,
arg.c_grid_desc_m_n_,
arg.block_2_ctile_map_);
}
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r3.hpp
View file @ a65d6459
......@@ -21,21 +21,20 @@ template <typename GridwiseGemm,
typename FloatC,
typename AGridDesc_K0_M_K1,
typename BGridDesc_K0_N_K1,
typename CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2,
typename CGridDesc_M_N,
typename Block2CTileMap,
bool HasMainKBlockLoop>
__global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
#endif
kernel_gemm_xdlops_v2r3(
const FloatAB* __restrict__ p_a_grid,
const FloatAB* __restrict__ p_b_grid,
FloatC* __restrict__ p_c_grid,
const AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1,
const BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1,
const CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2 c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2,
const Block2CTileMap block_2_ctile_map)
kernel_gemm_xdlops_v2r3(const FloatAB* __restrict__ p_a_grid,
const FloatAB* __restrict__ p_b_grid,
FloatC* __restrict__ p_c_grid,
const AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1,
const BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1,
const CGridDesc_M_N c_grid_desc_m_n,
const Block2CTileMap block_2_ctile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx940__))
......@@ -47,7 +46,7 @@ __global__ void
p_shared,
a_grid_desc_k0_m_k1,
b_grid_desc_k0_n_k1,
c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2,
c_grid_desc_m_n,
block_2_ctile_map);
#else
ignore = p_a_grid;
......@@ -55,7 +54,7 @@ __global__ void
ignore = p_c_grid;
ignore = a_grid_desc_k0_m_k1;
ignore = b_grid_desc_k0_n_k1;
ignore = c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2;
ignore = c_grid_desc_m_n;
ignore = block_2_ctile_map;
#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
}
......@@ -306,16 +305,18 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3
using DefaultBlock2CTileMap = decltype(MakeDefaultBlock2CTileMap(CGridDesc_M_N{}, 1, 1));
template <bool HasMainKBlockLoop, typename Block2CTileMap = DefaultBlock2CTileMap>
__device__ static void
Run(const FloatAB* __restrict__ p_a_grid,
const FloatAB* __restrict__ p_b_grid,
FloatC* __restrict__ p_c_grid,
void* __restrict__ p_shared,
const AGridDesc_K0_M_K1& a_grid_desc_k0_m_k1,
const BGridDesc_K0_N_K1& b_grid_desc_k0_n_k1,
const CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2& c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2,
const Block2CTileMap& block_2_ctile_map)
__device__ static void Run(const FloatAB* __restrict__ p_a_grid,
const FloatAB* __restrict__ p_b_grid,
FloatC* __restrict__ p_c_grid,
void* __restrict__ p_shared,
const AGridDesc_K0_M_K1& a_grid_desc_k0_m_k1,
const BGridDesc_K0_N_K1& b_grid_desc_k0_n_k1,
const CGridDesc_M_N& c_grid_desc_m_n,
const Block2CTileMap& block_2_ctile_map)
{
const auto c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_grid_desc_m_n);
const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_a_grid, a_grid_desc_k0_m_k1.GetElementSpaceSize());
const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
......
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