Skip to content

GitLab

"vscode:/vscode.git/clone" did not exist on "c54f8bcc25ace7b8d9ee86ddeb72738c87f908bb"
Commit 804e6803 authored by Harisankar Sadasivan's avatar Harisankar Sadasivan
Browse files

files modified for 1s cold and warm runs

parent 87efbb63
Show whitespace changes
Inline Side-by-side
Showing with 442 additions and 431 deletions
include/ck/host_utility/kernel_launch.hpp
View file @ 804e6803
...@@ -20,6 +20,33 @@ float launch_and_time_kernel(const StreamConfig& stream_config, ...@@ -20,6 +20,33 @@ float launch_and_time_kernel(const StreamConfig& stream_config,
#if CK_TIME_KERNEL #if CK_TIME_KERNEL
if(stream_config.time_kernel_) if(stream_config.time_kernel_)
{ {
if(ck::get_device_name() == "gfx940" || ck::get_device_name() == "gfx941" || ck::get_device_name() == "gfx942")
{
hipEvent_t start, stop;
hip_check_error(hipEventCreate(&start));
hip_check_error(hipEventCreate(&stop));
hip_check_error(hipDeviceSynchronize());
hip_check_error(hipEventRecord(start, stream_config.stream_id_));
for(int i = 0; i < stream_config.nrepeat_; ++i)
{
kernel<<<grid_dim, block_dim, lds_byte, stream_config.stream_id_>>>(args...);
hip_check_error(hipGetLastError());
}
hip_check_error(hipEventRecord(stop, stream_config.stream_id_));
hip_check_error(hipEventSynchronize(stop));
float total_time = 0;
hip_check_error(hipEventElapsedTime(&total_time, start, stop));
total_time/=10;
stream_config.cold_niters_ = (1000.0 / total_time);//we need longer runtime to ramp up the clk on MI300s
stream_config.nrepeat_ = stream_config.cold_niters_;
}
#if DEBUG_LOG #if DEBUG_LOG
printf("%s: grid_dim {%d, %d, %d}, block_dim {%d, %d, %d} \n", printf("%s: grid_dim {%d, %d, %d}, block_dim {%d, %d, %d} \n",
__func__, __func__,
......
include/ck/stream_config.hpp
View file @ 804e6803
...@@ -11,6 +11,6 @@ struct StreamConfig ...@@ -11,6 +11,6 @@ struct StreamConfig
hipStream_t stream_id_ = nullptr; hipStream_t stream_id_ = nullptr;
bool time_kernel_ = false; bool time_kernel_ = false;
int log_level_ = 0; int log_level_ = 0;
int cold_niters_ = 5; mutable int cold_niters_ = 5;
int nrepeat_ = 50; mutable int nrepeat_ = 50;
}; };
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4r2.hpp
View file @ 804e6803
...@@ -669,25 +669,14 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2 ...@@ -669,25 +669,14 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
const BElementwiseOperation b_element_op = BElementwiseOperation{}, const BElementwiseOperation b_element_op = BElementwiseOperation{},
const CElementwiseOperation c_element_op = CElementwiseOperation{}) const CElementwiseOperation c_element_op = CElementwiseOperation{})
{ {
for(auto i = 0; i < 1500; i++)
{
const FloatA* p_a_grid = karg.p_a_grid; const FloatA* p_a_grid = karg.p_a_grid;
const FloatB* p_b_grid = karg.p_b_grid; const FloatB* p_b_grid = karg.p_b_grid;
FloatC* p_c_grid = karg.p_c_grid; FloatC* p_c_grid = karg.p_c_grid;
const auto a_b_k0_m_k1_grid_desc = MakeAGridDescriptor_KBatch_K0_M_K1(karg.M, const auto a_b_k0_m_k1_grid_desc = MakeAGridDescriptor_KBatch_K0_M_K1(
karg.MPadded, karg.M, karg.MPadded, karg.K, karg.StrideA, karg.k_batch, karg.K0Padded, karg.KPadded);
karg.K, const auto b_b_k0_n_k1_grid_desc = MakeBGridDescriptor_KBatch_K0_N_K1(
karg.StrideA, karg.K, karg.NPadded, karg.N, karg.StrideB, karg.k_batch, karg.K0Padded, karg.KPadded);
karg.k_batch,
karg.K0Padded,
karg.KPadded);
const auto b_b_k0_n_k1_grid_desc = MakeBGridDescriptor_KBatch_K0_N_K1(karg.K,
karg.NPadded,
karg.N,
karg.StrideB,
karg.k_batch,
karg.K0Padded,
karg.KPadded);
const auto c_grid_desc_m_n = MakeCGridDescriptor_M_N(karg.M, karg.N, karg.StrideC); const auto c_grid_desc_m_n = MakeCGridDescriptor_M_N(karg.M, karg.N, karg.StrideC);
const auto c_grid_desc_mblock_mperblock_nblock_nperblock = const auto c_grid_desc_mblock_mperblock_nblock_nperblock =
...@@ -877,8 +866,8 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2 ...@@ -877,8 +866,8 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
auto c_thread_buf = blockwise_gemm.GetCThreadBuffer(); auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
// LDS allocation for A and B: be careful of alignment // LDS allocation for A and B: be careful of alignment
constexpr auto a_block_space_size = math::integer_least_multiple( constexpr auto a_block_space_size =
a_k0_m_k1_block_desc.GetElementSpaceSize(), max_lds_align); math::integer_least_multiple(a_k0_m_k1_block_desc.GetElementSpaceSize(), max_lds_align);
auto p_a_block = reinterpret_cast<LDSTypeA*>(p_shared_block); auto p_a_block = reinterpret_cast<LDSTypeA*>(p_shared_block);
auto p_b_block = reinterpret_cast<LDSTypeB*>(p_a_block + a_block_space_size); auto p_b_block = reinterpret_cast<LDSTypeB*>(p_a_block + a_block_space_size);
...@@ -943,23 +932,20 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2 ...@@ -943,23 +932,20 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor( constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
c_block_desc_mblock_mperblock_nblock_nperblock, c_block_desc_mblock_mperblock_nblock_nperblock,
make_tuple(make_freeze_transform(I0), // freeze mblock make_tuple(
make_unmerge_transform( make_freeze_transform(I0), // freeze mblock
make_tuple(CShuffleMRepeatPerShuffle, make_unmerge_transform(make_tuple(CShuffleMRepeatPerShuffle,
M1, M1,
M2, M2,
M3, M3,
M4)), // M1 = MWave, M2 * M3 * M4 = MPerXDL M4)), // M1 = MWave, M2 * M3 * M4 = MPerXDL
make_freeze_transform(I0), // freeze nblock make_freeze_transform(I0), // freeze nblock
make_unmerge_transform( make_unmerge_transform(make_tuple(CShuffleNRepeatPerShuffle,
make_tuple(CShuffleNRepeatPerShuffle,
N1, N1,
N2))), // M1 = MWave, M2 * M3 * M4 = MPerXDL N2))), // M1 = MWave, M2 * M3 * M4 = MPerXDL
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}), make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(Sequence<>{}, make_tuple(
Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{}));
Sequence<>{},
Sequence<1, 3, 7>{}));
// calculate origin of thread output tensor on global memory // calculate origin of thread output tensor on global memory
// blockwise GEMM c matrix starting index // blockwise GEMM c matrix starting index
...@@ -990,8 +976,8 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2 ...@@ -990,8 +976,8 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
make_multi_index(n_thread_data_on_block)); make_multi_index(n_thread_data_on_block));
// VGPR to LDS // VGPR to LDS
auto c_thread_copy_vgpr_to_lds = ThreadwiseTensorSliceTransfer_v1r3< auto c_thread_copy_vgpr_to_lds =
FloatAcc, ThreadwiseTensorSliceTransfer_v1r3<FloatAcc,
FloatC, FloatC,
decltype(c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc), decltype(c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc),
decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2), decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
...@@ -1009,7 +995,8 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2 ...@@ -1009,7 +995,8 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
1, 1,
InMemoryDataOperationEnum::Set, InMemoryDataOperationEnum::Set,
1, 1,
true>{c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2, true>{
c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
make_multi_index(0, make_multi_index(0,
0, 0,
m_thread_data_on_block_idx[I1], m_thread_data_on_block_idx[I1],
...@@ -1082,8 +1069,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2 ...@@ -1082,8 +1069,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
block_sync_lds(); block_sync_lds();
// LDS to global // LDS to global
c_block_copy_lds_to_global.Run( c_block_copy_lds_to_global.Run(c_block_desc_mblock_mperblock_nblock_nperblock,
c_block_desc_mblock_mperblock_nblock_nperblock,
c_block_buf, c_block_buf,
c_grid_desc_mblock_mperblock_nblock_nperblock, c_grid_desc_mblock_mperblock_nblock_nperblock,
c_grid_buf); c_grid_buf);
...@@ -1108,13 +1094,11 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2 ...@@ -1108,13 +1094,11 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
if constexpr(mxdlperwave < MRepeat - CShuffleMRepeatPerShuffle) if constexpr(mxdlperwave < MRepeat - CShuffleMRepeatPerShuffle)
{ {
c_block_copy_lds_to_global.MoveDstSliceWindow( c_block_copy_lds_to_global.MoveDstSliceWindow(
c_grid_desc_mblock_mperblock_nblock_nperblock, c_grid_desc_mblock_mperblock_nblock_nperblock, mxdlperwave_forward_step);
mxdlperwave_forward_step);
} }
}); });
} }
} }
}
static std::string GetTypeString() static std::string GetTypeString()
{ {
......
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