Skip to content

GitLab

"vscode:/vscode.git/clone" did not exist on "50ba6ab041684ef820d07fb6538e1ce2d34e1e67"
Commit ad2fddf4 authored by Jing Zhang's avatar Jing Zhang
Browse files

Merge remote-tracking branch 'origin/develop' into grouped_gemm_args_simplify

parents fa649421 1ee99dca
Expand all Show whitespace changes
Inline Side-by-side
Showing with 1234 additions and 87 deletions
include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_first_half.hpp
View file @ ad2fddf4
...@@ -161,7 +161,7 @@ struct GridwiseMultiblockWelfordFirstHalf ...@@ -161,7 +161,7 @@ struct GridwiseMultiblockWelfordFirstHalf
PassThroughOp, PassThroughOp,
ThreadBufferLengths_M_1, ThreadBufferLengths_M_1,
Sequence<0, 1>, Sequence<0, 1>,
1, 0,
1, 1,
InMemoryDataOperationEnum::Set, InMemoryDataOperationEnum::Set,
1, 1,
...@@ -180,7 +180,7 @@ struct GridwiseMultiblockWelfordFirstHalf ...@@ -180,7 +180,7 @@ struct GridwiseMultiblockWelfordFirstHalf
PassThroughOp, PassThroughOp,
ThreadBufferLengths_M_1, ThreadBufferLengths_M_1,
Sequence<0, 1>, Sequence<0, 1>,
1, 0,
1, 1,
InMemoryDataOperationEnum::Set, InMemoryDataOperationEnum::Set,
1, 1,
......
include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_second_half_batchnorm_forward_final.hpp → include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_second_half_batchnorm_forward_final_obsolete.hpp
View file @ ad2fddf4
...@@ -33,7 +33,6 @@ __global__ void kernel_welford_second_half_batchnorm_forward_final( ...@@ -33,7 +33,6 @@ __global__ void kernel_welford_second_half_batchnorm_forward_final(
const MeanVarGridDesc_M mean_var_grid_desc_m, const MeanVarGridDesc_M mean_var_grid_desc_m,
index_t blkgroup_size, index_t blkgroup_size,
index_t num_xy_k_block_tile_iteration, index_t num_xy_k_block_tile_iteration,
index_t num_mean_var_count_k_block_tile_iteration,
AccDataType epsilon, AccDataType epsilon,
const MeanVarDataType* const __restrict__ p_in_welford_mean, const MeanVarDataType* const __restrict__ p_in_welford_mean,
const MeanVarDataType* const __restrict__ p_in_welford_variance, const MeanVarDataType* const __restrict__ p_in_welford_variance,
...@@ -59,7 +58,6 @@ __global__ void kernel_welford_second_half_batchnorm_forward_final( ...@@ -59,7 +58,6 @@ __global__ void kernel_welford_second_half_batchnorm_forward_final(
mean_var_grid_desc_m, mean_var_grid_desc_m,
blkgroup_size, blkgroup_size,
num_xy_k_block_tile_iteration, num_xy_k_block_tile_iteration,
num_mean_var_count_k_block_tile_iteration,
epsilon, epsilon,
p_in_welford_mean, p_in_welford_mean,
p_in_welford_variance, p_in_welford_variance,
...@@ -152,7 +150,6 @@ struct GridwiseWelfordSecondHalfBatchNormForwardFinal ...@@ -152,7 +150,6 @@ struct GridwiseWelfordSecondHalfBatchNormForwardFinal
const MeanVarGridDesc_M& mean_var_grid_desc_m, const MeanVarGridDesc_M& mean_var_grid_desc_m,
index_t blkgroup_size, index_t blkgroup_size,
index_t num_xy_k_block_tile_iteration, index_t num_xy_k_block_tile_iteration,
index_t num_mean_var_count_k_block_tile_iteration,
AccDataType epsilon, AccDataType epsilon,
const MeanVarDataType* const __restrict__ p_in_welford_mean, const MeanVarDataType* const __restrict__ p_in_welford_mean,
const MeanVarDataType* const __restrict__ p_in_welford_variance, const MeanVarDataType* const __restrict__ p_in_welford_variance,
...@@ -223,7 +220,7 @@ struct GridwiseWelfordSecondHalfBatchNormForwardFinal ...@@ -223,7 +220,7 @@ struct GridwiseWelfordSecondHalfBatchNormForwardFinal
decltype(thread_buffer_desc_m_1), decltype(thread_buffer_desc_m_1),
ThreadBufferLengths_M_1, ThreadBufferLengths_M_1,
Sequence<0, 1>, Sequence<0, 1>,
1, 0,
1, 1,
1, 1,
true>( true>(
...@@ -239,7 +236,7 @@ struct GridwiseWelfordSecondHalfBatchNormForwardFinal ...@@ -239,7 +236,7 @@ struct GridwiseWelfordSecondHalfBatchNormForwardFinal
decltype(thread_buffer_desc_m_1), decltype(thread_buffer_desc_m_1),
ThreadBufferLengths_M_1, ThreadBufferLengths_M_1,
Sequence<0, 1>, Sequence<0, 1>,
1, 0,
1, 1,
1, 1,
true>( true>(
...@@ -257,9 +254,6 @@ struct GridwiseWelfordSecondHalfBatchNormForwardFinal ...@@ -257,9 +254,6 @@ struct GridwiseWelfordSecondHalfBatchNormForwardFinal
const auto welford_count_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>( const auto welford_count_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_in_welford_count, mean_var_count_grid_desc_m_k.GetElementSpaceSize()); p_in_welford_count, mean_var_count_grid_desc_m_k.GetElementSpaceSize());
constexpr auto mean_var_count_thread_copy_step_m_k =
make_multi_index(0, KThreadClusterSize * 1);
// Step 1: do final welford reduction to get mean and variance // Step 1: do final welford reduction to get mean and variance
static_for<0, MThreadSliceSize, 1>{}([&](auto I) { static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
...@@ -268,8 +262,11 @@ struct GridwiseWelfordSecondHalfBatchNormForwardFinal ...@@ -268,8 +262,11 @@ struct GridwiseWelfordSecondHalfBatchNormForwardFinal
welford_count_thread_buf(I) = 0; welford_count_thread_buf(I) = 0;
}); });
for(index_t reducedTiles = 0; reducedTiles < num_mean_var_count_k_block_tile_iteration; constexpr auto mean_var_count_thread_copy_step_m_k =
++reducedTiles) make_multi_index(0, KThreadClusterSize);
int32_t reducedSize = 0;
while(reducedSize < blkgroup_size)
{ {
threadwise_mean_var_load_m_k.Run(mean_var_count_grid_desc_m_k, threadwise_mean_var_load_m_k.Run(mean_var_count_grid_desc_m_k,
welford_mean_global_val_buf, welford_mean_global_val_buf,
...@@ -296,6 +293,8 @@ struct GridwiseWelfordSecondHalfBatchNormForwardFinal ...@@ -296,6 +293,8 @@ struct GridwiseWelfordSecondHalfBatchNormForwardFinal
welford_var_thread_buf, welford_var_thread_buf,
welford_count_thread_buf); welford_count_thread_buf);
reducedSize += KThreadClusterSize;
threadwise_mean_var_load_m_k.MoveSrcSliceWindow(mean_var_count_grid_desc_m_k, threadwise_mean_var_load_m_k.MoveSrcSliceWindow(mean_var_count_grid_desc_m_k,
mean_var_count_thread_copy_step_m_k); mean_var_count_thread_copy_step_m_k);
threadwise_count_load_m_k.MoveSrcSliceWindow(mean_var_count_grid_desc_m_k, threadwise_count_load_m_k.MoveSrcSliceWindow(mean_var_count_grid_desc_m_k,
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp
View file @ ad2fddf4
...@@ -3,6 +3,8 @@ ...@@ -3,6 +3,8 @@
#pragma once #pragma once
#include <iostream>
#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v1.hpp" #include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v1.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v2.hpp" #include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v2.hpp"
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v2.hpp
View file @ ad2fddf4
...@@ -79,6 +79,10 @@ struct GridwiseGemmPipeline_v2 ...@@ -79,6 +79,10 @@ struct GridwiseGemmPipeline_v2
do do
{ {
#if CK_EXPERIMENTAL_PIPELINE_V2_IGLP_OPT
__builtin_amdgcn_iglp_opt(CK_EXPERIMENTAL_PIPELINE_V2_IGLP_OPT);
#endif
block_sync_lds(); block_sync_lds();
// GEMM i // GEMM i
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r3.hpp
View file @ ad2fddf4
...@@ -27,6 +27,9 @@ template <typename GridwiseGemm, ...@@ -27,6 +27,9 @@ template <typename GridwiseGemm,
__global__ void __global__ void
#if CK_USE_LAUNCH_BOUNDS #if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU) __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
#endif
#if CK_USE_WAVES_PER_EU
__attribute__((amdgpu_waves_per_eu(CK_MIN_WAVES_PER_EU, CK_MAX_WAVES_PER_EU)))
#endif #endif
kernel_gemm_xdlops_v2r3(const FloatAB* __restrict__ p_a_grid, kernel_gemm_xdlops_v2r3(const FloatAB* __restrict__ p_a_grid,
const FloatAB* __restrict__ p_b_grid, const FloatAB* __restrict__ p_b_grid,
...@@ -60,6 +63,9 @@ template <typename GridwiseGemm, bool HasMainKBlockLoop> ...@@ -60,6 +63,9 @@ template <typename GridwiseGemm, bool HasMainKBlockLoop>
__global__ void __global__ void
#if CK_USE_LAUNCH_BOUNDS #if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU) __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
#endif
#if CK_USE_WAVES_PER_EU
__attribute__((amdgpu_waves_per_eu(CK_MIN_WAVES_PER_EU, CK_MAX_WAVES_PER_EU)))
#endif #endif
kernel_gemm_xdlops_v2r3(const typename GridwiseGemm::Argument karg) kernel_gemm_xdlops_v2r3(const typename GridwiseGemm::Argument karg)
{ {
......
include/ck/tensor_operation/gpu/warp/xdlops_gemm.hpp
View file @ ad2fddf4
...@@ -29,7 +29,9 @@ enum struct MfmaInstr ...@@ -29,7 +29,9 @@ enum struct MfmaInstr
mfma_i32_16x16x16i8, mfma_i32_16x16x16i8,
mfma_i32_32x32x16i8, mfma_i32_32x32x16i8,
mfma_i32_16x16x32i8, mfma_i32_16x16x32i8,
mfma_f64_16x16x4f64 mfma_f64_16x16x4f64,
mfma_f32_32x32x16f8f8,
mfma_f32_16x16x32f8f8
}; };
template <MfmaInstr instr> template <MfmaInstr instr>
...@@ -454,6 +456,50 @@ struct mfma_type<MfmaInstr::mfma_f64_16x16x4f64> ...@@ -454,6 +456,50 @@ struct mfma_type<MfmaInstr::mfma_f64_16x16x4f64>
} }
}; };
template <>
struct mfma_type<MfmaInstr::mfma_f32_32x32x16f8f8>
{
static constexpr index_t group_size = 4;
static constexpr index_t num_groups_per_blk = 4;
static constexpr index_t num_regs_per_blk = 16;
static constexpr index_t num_threads_per_blk = 32;
static constexpr index_t wave_size = 64;
static constexpr index_t num_input_blks = 2;
static constexpr index_t num_output_blks = 1;
static constexpr index_t m_per_blk = 32;
static constexpr index_t n_per_blk = 32;
static constexpr index_t k_per_blk = 8;
static constexpr bool is_k_reduction = true;
template <index_t MPerXdlops, index_t NPerXdlops, class FloatA, class FloatB, class FloatC>
__device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
{
intrin_mfma_f32_32x32x16f8f8<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
}
};
template <>
struct mfma_type<MfmaInstr::mfma_f32_16x16x32f8f8>
{
static constexpr index_t group_size = 4;
static constexpr index_t num_groups_per_blk = 1;
static constexpr index_t num_regs_per_blk = 4;
static constexpr index_t num_threads_per_blk = 16;
static constexpr index_t wave_size = 64;
static constexpr index_t num_input_blks = 4;
static constexpr index_t num_output_blks = 1;
static constexpr index_t m_per_blk = 16;
static constexpr index_t n_per_blk = 16;
static constexpr index_t k_per_blk = 8;
static constexpr bool is_k_reduction = true;
template <index_t MPerXdlops, index_t NPerXdlops, class FloatA, class FloatB, class FloatC>
__device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
{
intrin_mfma_f32_16x16x32f8f8<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
}
};
template <typename base_type, index_t MPerXdlops, index_t NPerXdlops> template <typename base_type, index_t MPerXdlops, index_t NPerXdlops>
struct MfmaSelector struct MfmaSelector
{ {
...@@ -594,6 +640,18 @@ struct MfmaSelector ...@@ -594,6 +640,18 @@ struct MfmaSelector
} }
#endif #endif
template <>
static constexpr auto GetMfma<f8_t, 32, 32>()
{
return MfmaInstr::mfma_f32_32x32x16f8f8;
}
template <>
static constexpr auto GetMfma<f8_t, 16, 16>()
{
return MfmaInstr::mfma_f32_16x16x32f8f8;
}
static constexpr auto selected_mfma = mfma_type<GetMfma<base_type, MPerXdlops, NPerXdlops>()>{}; static constexpr auto selected_mfma = mfma_type<GetMfma<base_type, MPerXdlops, NPerXdlops>()>{};
__host__ __device__ constexpr MfmaSelector() __host__ __device__ constexpr MfmaSelector()
...@@ -794,7 +852,7 @@ struct XdlopsGemm ...@@ -794,7 +852,7 @@ struct XdlopsGemm
{ {
static_assert(is_same<base_type, double>::value || is_same<base_type, float>::value || static_assert(is_same<base_type, double>::value || is_same<base_type, float>::value ||
is_same<base_type, half_t>::value || is_same<base_type, bhalf_t>::value || is_same<base_type, half_t>::value || is_same<base_type, bhalf_t>::value ||
is_same<base_type, int8_t>::value, is_same<base_type, int8_t>::value || is_same<base_type, f8_t>::value,
"base base_type must be double, float, half, bfloat16, and int8_t!"); "base base_type must be double, float, half, bfloat16, and int8_t!");
static_for<0, KPack / mfma_instr.k_per_blk, 1>{}([&](auto k) { static_for<0, KPack / mfma_instr.k_per_blk, 1>{}([&](auto k) {
......
include/ck/utility/amd_buffer_addressing.hpp
View file @ ad2fddf4
...@@ -1114,13 +1114,30 @@ amd_buffer_load_invalid_element_return_zero(const T* p_src_wave, ...@@ -1114,13 +1114,30 @@ amd_buffer_load_invalid_element_return_zero(const T* p_src_wave,
#if CK_EXPERIMENTAL_USE_BUFFER_LOAD_OOB_CHECK_OFFSET_TRICK #if CK_EXPERIMENTAL_USE_BUFFER_LOAD_OOB_CHECK_OFFSET_TRICK
uint32_t src_addr_shift = src_thread_element_valid ? 0 : 0x80000000; uint32_t src_addr_shift = src_thread_element_valid ? 0 : 0x80000000;
if constexpr(is_same<scalar_t, f8_t>::value)
{
auto tmp = amd_buffer_load_impl<int8_t, vector_size, coherence>(
src_wave_buffer_resource, src_addr_shift + src_thread_addr_offset, 0);
return bit_cast<vector_t>(tmp);
}
else
{
return amd_buffer_load_impl<scalar_t, vector_size, coherence>( return amd_buffer_load_impl<scalar_t, vector_size, coherence>(
src_wave_buffer_resource, src_addr_shift + src_thread_addr_offset, 0); src_wave_buffer_resource, src_addr_shift + src_thread_addr_offset, 0);
}
#else #else
if constexpr(is_same<scalar_t, f8_t>::value)
{
auto tmp = amd_buffer_load_impl<int8_t, vector_size, coherence>(
src_wave_buffer_resource, src_thread_addr_offset, 0);
return src_thread_element_valid ? bit_cast<vector_t>(tmp) : vector_t(0);
}
else
{
vector_t tmp = amd_buffer_load_impl<scalar_t, vector_size, coherence>( vector_t tmp = amd_buffer_load_impl<scalar_t, vector_size, coherence>(
src_wave_buffer_resource, src_thread_addr_offset, 0); src_wave_buffer_resource, src_thread_addr_offset, 0);
return src_thread_element_valid ? tmp : vector_t(0); return src_thread_element_valid ? tmp : vector_t(0);
}
#endif #endif
} }
...@@ -1179,14 +1196,34 @@ __device__ void amd_buffer_store(const typename vector_type_maker<T, N>::type::t ...@@ -1179,14 +1196,34 @@ __device__ void amd_buffer_store(const typename vector_type_maker<T, N>::type::t
#if CK_EXPERIMENTAL_USE_BUFFER_STORE_OOB_CHECK_OFFSET_TRICK #if CK_EXPERIMENTAL_USE_BUFFER_STORE_OOB_CHECK_OFFSET_TRICK
uint32_t dst_addr_shift = dst_thread_element_valid ? 0 : 0x80000000; uint32_t dst_addr_shift = dst_thread_element_valid ? 0 : 0x80000000;
if constexpr(is_same<scalar_t, f8_t>::value)
{
auto tmp =
bit_cast<typename vector_type_maker<int8_t, vector_size>::type::type>(src_thread_data);
amd_buffer_store_impl<int8_t, vector_size, coherence>(
tmp, dst_wave_buffer_resource, dst_addr_shift + dst_thread_addr_offset, 0);
}
else
{
amd_buffer_store_impl<scalar_t, vector_size, coherence>( amd_buffer_store_impl<scalar_t, vector_size, coherence>(
src_thread_data, dst_wave_buffer_resource, dst_addr_shift + dst_thread_addr_offset, 0); src_thread_data, dst_wave_buffer_resource, dst_addr_shift + dst_thread_addr_offset, 0);
}
#else #else
if(dst_thread_element_valid) if(dst_thread_element_valid)
{
if constexpr(is_same<scalar_t, f8_t>::value)
{
auto tmp = bit_cast<typename vector_type_maker<int8_t, vector_size>::type::type>(
src_thread_data);
amd_buffer_store_impl<int8_t, vector_size, coherence>(
tmp, dst_wave_buffer_resource, dst_thread_addr_offset, 0);
}
else
{ {
amd_buffer_store_impl<scalar_t, vector_size, coherence>( amd_buffer_store_impl<scalar_t, vector_size, coherence>(
src_thread_data, dst_wave_buffer_resource, dst_thread_addr_offset, 0); src_thread_data, dst_wave_buffer_resource, dst_thread_addr_offset, 0);
} }
}
#endif #endif
} }
......
include/ck/utility/amd_xdlops.hpp
View file @ ad2fddf4
...@@ -354,5 +354,68 @@ struct intrin_mfma_f64_16x16x4f64<16, 16> ...@@ -354,5 +354,68 @@ struct intrin_mfma_f64_16x16x4f64<16, 16>
#endif #endif
} }
}; };
template <index_t MPerWave, index_t NPerWave>
struct intrin_mfma_f32_32x32x16f8f8;
template <>
struct intrin_mfma_f32_32x32x16f8f8<32, 32>
{
template <class FloatC>
__device__ static void Run(const f8x8_t& reg_a, const f8x8_t& reg_b, FloatC& reg_c)
{
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
reg_c.template AsType<float16_t>()(Number<0>{}) =
__builtin_amdgcn_mfma_f32_32x32x16_fp8_fp8(
bit_cast<long>(reg_a),
bit_cast<long>(reg_b),
reg_c.template AsType<float16_t>()[Number<0>{}],
0,
0,
0);
#else
vector_type<f8_t, 8> reg_a_v(reg_a);
vector_type<f8_t, 8> reg_b_v(reg_b);
static_for<0, 8, 1>{}([&](auto k) {
float reg_a_f32 = type_convert<float>(reg_a_v.template AsType<f8_t>()[Number<k>{}]);
float reg_b_f32 = type_convert<float>(reg_b_v.template AsType<f8_t>()[Number<k>{}]);
intrin_mfma_f32_32x32x2f32<32, 32>::Run(reg_a_f32, reg_b_f32, reg_c);
});
#endif
}
};
template <index_t MPerWave, index_t NPerWave>
struct intrin_mfma_f32_16x16x32f8f8;
template <>
struct intrin_mfma_f32_16x16x32f8f8<16, 16>
{
template <class FloatC>
__device__ static void Run(const f8x8_t& reg_a, const f8x8_t& reg_b, FloatC& reg_c)
{
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
reg_c.template AsType<float4_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f32_16x16x32_fp8_fp8(
bit_cast<long>(reg_a),
bit_cast<long>(reg_b),
reg_c.template AsType<float4_t>()[Number<0>{}],
0,
0,
0);
#else
vector_type<f8_t, 8> reg_a_v(reg_a);
vector_type<f8_t, 8> reg_b_v(reg_b);
static_for<0, 8, 1>{}([&](auto k) {
float reg_a_f32 = type_convert<float>(reg_a_v.template AsType<f8_t>()[Number<k>{}]);
float reg_b_f32 = type_convert<float>(reg_b_v.template AsType<f8_t>()[Number<k>{}]);
intrin_mfma_f32_16x16x4f32<16, 16>::Run(reg_a_f32, reg_b_f32, reg_c);
});
#endif
}
};
} // namespace ck } // namespace ck
#endif #endif
include/ck/utility/get_shift.hpp 0 → 100644
View file @ ad2fddf4
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
namespace ck {
template <index_t N>
static constexpr __device__ index_t get_shift()
{
return (get_shift<N / 2>() + 1);
};
template <>
constexpr __device__ index_t get_shift<1>()
{
return (0);
}
} // namespace ck
include/ck/utility/reduction_common.hpp
View file @ ad2fddf4
...@@ -25,16 +25,4 @@ struct float_equal_zero ...@@ -25,16 +25,4 @@ struct float_equal_zero
}; };
}; };
template <index_t N>
static constexpr __device__ index_t get_shift()
{
return (get_shift<N / 2>() + 1);
};
template <>
constexpr __device__ index_t get_shift<1>()
{
return (0);
}
} // namespace ck } // namespace ck
include/ck/utility/workgroup_synchronization.hpp 0 → 100644
View file @ ad2fddf4
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/host_utility/hip_check_error.hpp"
namespace ck {
// Initialization flag of Barrier object, can be any value except for zero
static constexpr int BarrierInitFlag = 0x7856;
// 1) only the first thread-block in the synchronizaton group is supposed to call this function. It
// is the responsibility of the user to ensure the two integer values in p_control_bits are zeros
// before calling gms_init().
// 2) Aftercalling gms_reset(), the two integer values in p_control_bits will be zeros, so no
// repetitious initialization of p_control_bits buffer is required
static __device__ void gms_init(int NumWarps, int* p_control_bits)
{
union
{
int two32[2];
unsigned long one64;
} regs;
regs.two32[0] = BarrierInitFlag;
regs.two32[1] = NumWarps;
if(threadIdx.x == 0)
atomicCAS(reinterpret_cast<unsigned long*>(p_control_bits), 0, regs.one64);
};
// all the workgroups in the synchronization group is supposed to call this function
static __device__ void gms_barrier(int* p_control_bits)
{
constexpr int mask = warpSize - 1;
if((threadIdx.x & mask) == 0)
{
// ensure the barrier object is initialized
do
{
const int r0 = __atomic_load_n(&p_control_bits[0], __ATOMIC_RELAXED);
if(r0 == BarrierInitFlag)
break;
} while(true);
// go ahead toward the barrier line
atomicSub(&p_control_bits[1], 1);
// wait until all warps have arrived
do
{
const int r1 = __atomic_load_n(&p_control_bits[1], __ATOMIC_RELAXED);
if(r1 == 0)
break;
} while(true);
};
};
// 1) Only the first thread-block in the synchronizaton group is supposed to call this function.
// 2) Aftercalling gms_reset(), the two integer values in p_control_bits will be zeros, so no
// repetitious initialization of p_control_bits buffer is required
static __device__ void gms_reset(int* p_control_bits)
{
// reset the barrier object
if(threadIdx.x == 0)
(void)atomicCAS(&p_control_bits[0], BarrierInitFlag, 0);
};
} // namespace ck
library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_multi_d.hpp
View file @ ad2fddf4
// SPDX-License-Identifier: MIT // SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once #pragma once
......
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight.hpp
View file @ ad2fddf4
...@@ -91,6 +91,42 @@ void add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f32_instances( ...@@ -91,6 +91,42 @@ void add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f32_instances(
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
void add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_bf16_f32_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
NHWGC,
GKYXC,
NHWGK,
BF16,
F32,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
NHWGC,
GKYXC,
NHWGK,
F16,
F16,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
NHWGC,
GKYXC,
NHWGK,
F32,
F32,
F32,
PassThrough,
PassThrough,
PassThrough>>>& instances);
// conv3d backward weight // conv3d backward weight
void add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_bf16_f32_bf16_instances( void add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_bf16_f32_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3, std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
...@@ -162,8 +198,10 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -162,8 +198,10 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
{ {
std::vector<std::unique_ptr<DeviceOp>> op_ptrs; std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
if constexpr(NumDimSpatial == 1 && is_same_v<InLayout, GNWC> && if constexpr(NumDimSpatial == 1)
is_same_v<WeiLayout, GKXC> && is_same_v<OutLayout, GNWK>) {
if constexpr(is_same_v<InLayout, GNWC> && is_same_v<WeiLayout, GKXC> &&
is_same_v<OutLayout, GNWK>)
{ {
if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> && if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, float>) is_same_v<OutDataType, float>)
...@@ -175,35 +213,68 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -175,35 +213,68 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
{ {
add_device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f16_instances(op_ptrs); add_device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f16_instances(op_ptrs);
} }
else if constexpr(is_same_v<InDataType, ck::bhalf_t> && is_same_v<WeiDataType, float> && else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, ck::bhalf_t>) is_same_v<OutDataType, ck::bhalf_t>)
{ {
add_device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_bf16_f32_bf16_instances( add_device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_bf16_f32_bf16_instances(
op_ptrs); op_ptrs);
} }
} }
else if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, GNHWC> && }
is_same_v<WeiLayout, GKYXC> && is_same_v<OutLayout, GNHWK>) else if constexpr(NumDimSpatial == 2)
{
if constexpr(is_same_v<InLayout, GNHWC> && is_same_v<WeiLayout, GKYXC> &&
is_same_v<OutLayout, GNHWK>)
{ {
if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> && if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, float>) is_same_v<OutDataType, float>)
{ {
add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f32_instances(op_ptrs); add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f32_instances(
op_ptrs);
} }
else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> && else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t>) is_same_v<OutDataType, half_t>)
{ {
add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f16_instances(op_ptrs); add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f16_instances(
op_ptrs);
} }
else if constexpr(is_same_v<InDataType, ck::bhalf_t> && is_same_v<WeiDataType, float> && else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, ck::bhalf_t>) is_same_v<OutDataType, ck::bhalf_t>)
{ {
add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_bf16_f32_bf16_instances( add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_bf16_f32_bf16_instances(
op_ptrs); op_ptrs);
} }
} }
else if constexpr(NumDimSpatial == 3 && is_same_v<InLayout, GNDHWC> && else if constexpr(is_same_v<InLayout, NHWGC> && is_same_v<WeiLayout, GKYXC> &&
is_same_v<WeiLayout, GKZYXC> && is_same_v<OutLayout, GNDHWK>) is_same_v<OutLayout, NHWGK>)
{
if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, float>)
{
add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_instances(
op_ptrs);
}
else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t>)
{
add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f16_instances(
op_ptrs);
}
else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, ck::bhalf_t>)
{
add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_bf16_f32_bf16_instances(
op_ptrs);
}
}
}
else if constexpr(NumDimSpatial == 3)
{
if(is_same_v<InLayout, GNDHWC> && is_same_v<WeiLayout, GKZYXC> &&
is_same_v<OutLayout, GNDHWK>)
{ {
if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> && if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, float>) is_same_v<OutDataType, float>)
...@@ -217,13 +288,15 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -217,13 +288,15 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f16_instances( add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f16_instances(
op_ptrs); op_ptrs);
} }
else if constexpr(is_same_v<InDataType, ck::bhalf_t> && is_same_v<WeiDataType, float> && else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, ck::bhalf_t>) is_same_v<OutDataType, ck::bhalf_t>)
{ {
add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_bf16_f32_bf16_instances( add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_bf16_f32_bf16_instances(
op_ptrs); op_ptrs);
} }
} }
}
return op_ptrs; return op_ptrs;
} }
......
library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_instance.cpp
View file @ ad2fddf4
// SPDX-License-Identifier: MIT // SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib> #include <cstdlib>
......
library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_irregular_instance.cpp
View file @ ad2fddf4
// SPDX-License-Identifier: MIT // SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib> #include <cstdlib>
......
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