Skip to content

GitLab

Commit 1762f081 authored by Jun Liu's avatar Jun Liu
Browse files

Merge branch 'amd-develop-0725' into amd-master

parents ad65dfe7 be3fbf7f
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 1164 additions and 375 deletions
example/ck_tile/01_fmha/script/benchmark_bwd.sh 100644 → 100755
View file @ 1762f081
File mode changed from 100644 to 100755
example/ck_tile/01_fmha/script/smoke_test_bwd.sh 100644 → 100755
View file @ 1762f081
......@@ -8,7 +8,7 @@ export CK_WARMUP=0
export CK_REPEAT=1
COMMON_ARGS='-v=1'
set -x
for prec in "fp16" "bf16" ; do
for perm in 0 1 ; do
for hdim in 32 64 128 ; do
......@@ -31,3 +31,4 @@ done
done
done
done
set +x
example/ck_tile/01_fmha/script/smoke_test_fwd.sh
View file @ 1762f081
......@@ -10,7 +10,7 @@ export CK_REPEAT=1
COMMON_ARGS='-v=1 -warmup=0 -repeat=1'
# mode=0
# export HIP_VISIBLE_DEVICES=4
set -x
for prec in "fp16" "bf16" ; do
for mode in 1 0 ; do
for perm in 0 1 ; do
......@@ -40,6 +40,7 @@ done
done
done
for perm in 0 1 ; do
for bias in "n" "e" "a" ; do
for b in 1 2 ; do
......@@ -49,3 +50,4 @@ done
done
done
done
set +x
include/ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp
View file @ 1762f081
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
......@@ -15,6 +15,7 @@ enum struct ConvolutionForwardSpecialization
Filter1x1Pad0,
Filter1x1Stride1Pad0,
OddC,
Filter3x3,
};
inline std::string getConvForwardSpecializationString(const ConvolutionForwardSpecialization& s)
......@@ -25,6 +26,7 @@ inline std::string getConvForwardSpecializationString(const ConvolutionForwardSp
case ConvolutionForwardSpecialization::Filter1x1Pad0: return "Filter1x1Pad0";
case ConvolutionForwardSpecialization::Filter1x1Stride1Pad0: return "Filter1x1Stride1Pad0";
case ConvolutionForwardSpecialization::OddC: return "OddC";
case ConvolutionForwardSpecialization::Filter3x3: return "Filter3x3";
default: return "Unrecognized specialization!";
}
}
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_two_stage_xdl_cshuffle.hpp
View file @ 1762f081
......@@ -36,7 +36,7 @@ template <typename GridwiseGemm,
typename BGridDesc_BK0_N_K1,
typename CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
typename ComputePtrOffsetOfBatch,
index_t NumBatchToMerge,
index_t NumGroupsToMerge,
bool HasMainKBlockLoop,
InMemoryDataOperationEnum CGlobalMemoryDataOperation,
index_t MinimumOccupancy = 1,
......@@ -56,7 +56,7 @@ __global__ void
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx94__))
const index_t g_idx = __builtin_amdgcn_readfirstlane(blockIdx.z * NumBatchToMerge);
const index_t g_idx = __builtin_amdgcn_readfirstlane(blockIdx.z * NumGroupsToMerge);
const index_t k_idx = __builtin_amdgcn_readfirstlane(blockIdx.y * num_k_per_block);
const long_index_t a_batch_offset =
......@@ -92,7 +92,7 @@ template <typename GridwiseGemm,
typename BGridDesc_BK0_N_K1,
typename CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
typename ComputePtrOffsetOfBatch,
index_t NumBatchToMerge,
index_t NumGroupsToMerge,
bool HasMainKBlockLoop,
InMemoryDataOperationEnum CGlobalMemoryDataOperation,
index_t MinimumOccupancy = 1,
......@@ -113,7 +113,7 @@ __global__ void
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
// offset base pointer for each work-group
const index_t g_idx = __builtin_amdgcn_readfirstlane(blockIdx.z * NumBatchToMerge);
const index_t g_idx = __builtin_amdgcn_readfirstlane(blockIdx.z * NumGroupsToMerge);
const index_t k_idx = __builtin_amdgcn_readfirstlane(blockIdx.y * num_k_per_block);
const long_index_t a_batch_offset =
......@@ -189,7 +189,7 @@ template <ck::index_t NDimSpatial,
index_t CBlockTransferScalarPerVector_NWaveNPerXdl,
BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave,
BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1,
index_t NumBatchToMerge = 1,
index_t NumGroupsToMerge = 1,
typename ComputeTypeA = InDataType,
typename ComputeTypeB = ComputeTypeA>
struct DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle
......@@ -238,7 +238,7 @@ struct DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle
NPerBlock,
K1Number,
KPerBlock / K1Number,
NumBatchToMerge,
NumGroupsToMerge,
ConvBackwardWeightSpecialization>{};
static constexpr auto conv_to_gemm_transformer_v1 =
......@@ -638,7 +638,7 @@ struct DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle
index_t gdx, gdy, gdz;
std::tie(gdx, gdy, gdz) = GridwiseGemm::CalculateGridSize(
gemm_arg.M, gemm_arg.N, gemm_arg.KBatch, arg.Conv_G_ / NumBatchToMerge);
gemm_arg.M, gemm_arg.N, gemm_arg.KBatch, arg.Conv_G_ / NumGroupsToMerge);
float ave_time = 0;
......@@ -724,7 +724,7 @@ struct DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle
remove_reference_t<
DeviceOp::CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>,
ComputePtrOffsetOfStridedBatch<I1, I1, I0>,
NumBatchToMerge,
NumGroupsToMerge,
true,
InMemoryDataOperationEnum::AtomicAdd,
minimum_occupancy>;
......@@ -739,7 +739,7 @@ struct DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle
remove_reference_t<
DeviceOp::CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>,
ComputePtrOffsetOfStridedBatch<I1, I1, I0>,
NumBatchToMerge,
NumGroupsToMerge,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy>;
......@@ -760,7 +760,7 @@ struct DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle
remove_reference_t<
DeviceOp::CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>,
ComputePtrOffsetOfStridedBatch<I1, I1, I0>,
NumBatchToMerge,
NumGroupsToMerge,
true,
InMemoryDataOperationEnum::AtomicAdd,
minimum_occupancy,
......@@ -777,7 +777,7 @@ struct DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle
remove_reference_t<
DeviceOp::CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>,
ComputePtrOffsetOfStridedBatch<I1, I1, I0>,
NumBatchToMerge,
NumGroupsToMerge,
true,
InMemoryDataOperationEnum::AtomicAdd,
minimum_occupancy,
......@@ -796,7 +796,7 @@ struct DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle
remove_reference_t<
DeviceOp::CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>,
ComputePtrOffsetOfStridedBatch<I1, I1, I0>,
NumBatchToMerge,
NumGroupsToMerge,
true,
InMemoryDataOperationEnum::AtomicAdd,
minimum_occupancy,
......@@ -817,7 +817,7 @@ struct DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle
remove_reference_t<
DeviceOp::CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>,
ComputePtrOffsetOfStridedBatch<I1, I1, I0>,
NumBatchToMerge,
NumGroupsToMerge,
true,
InMemoryDataOperationEnum::AtomicAdd,
minimum_occupancy,
......@@ -838,7 +838,7 @@ struct DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle
remove_reference_t<
DeviceOp::CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>,
ComputePtrOffsetOfStridedBatch<I1, I1, I0>,
NumBatchToMerge,
NumGroupsToMerge,
true,
InMemoryDataOperationEnum::AtomicAdd,
minimum_occupancy,
......@@ -859,7 +859,7 @@ struct DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle
remove_reference_t<
DeviceOp::CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>,
ComputePtrOffsetOfStridedBatch<I1, I1, I0>,
NumBatchToMerge,
NumGroupsToMerge,
true,
InMemoryDataOperationEnum::AtomicAdd,
minimum_occupancy,
......@@ -879,7 +879,7 @@ struct DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle
remove_reference_t<
DeviceOp::CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>,
ComputePtrOffsetOfStridedBatch<I1, I1, I0>,
NumBatchToMerge,
NumGroupsToMerge,
true,
InMemoryDataOperationEnum::AtomicAdd,
minimum_occupancy,
......@@ -900,7 +900,7 @@ struct DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle
remove_reference_t<
DeviceOp::CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>,
ComputePtrOffsetOfStridedBatch<I1, I1, I0>,
NumBatchToMerge,
NumGroupsToMerge,
true,
InMemoryDataOperationEnum::AtomicAdd,
minimum_occupancy,
......@@ -920,7 +920,7 @@ struct DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle
remove_reference_t<
DeviceOp::CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>,
ComputePtrOffsetOfStridedBatch<I1, I1, I0>,
NumBatchToMerge,
NumGroupsToMerge,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
......@@ -937,7 +937,7 @@ struct DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle
remove_reference_t<
DeviceOp::CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>,
ComputePtrOffsetOfStridedBatch<I1, I1, I0>,
NumBatchToMerge,
NumGroupsToMerge,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
......@@ -956,7 +956,7 @@ struct DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle
remove_reference_t<
DeviceOp::CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>,
ComputePtrOffsetOfStridedBatch<I1, I1, I0>,
NumBatchToMerge,
NumGroupsToMerge,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
......@@ -977,7 +977,7 @@ struct DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle
remove_reference_t<
DeviceOp::CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>,
ComputePtrOffsetOfStridedBatch<I1, I1, I0>,
NumBatchToMerge,
NumGroupsToMerge,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
......@@ -998,7 +998,7 @@ struct DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle
remove_reference_t<
DeviceOp::CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>,
ComputePtrOffsetOfStridedBatch<I1, I1, I0>,
NumBatchToMerge,
NumGroupsToMerge,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
......@@ -1019,7 +1019,7 @@ struct DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle
remove_reference_t<
DeviceOp::CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>,
ComputePtrOffsetOfStridedBatch<I1, I1, I0>,
NumBatchToMerge,
NumGroupsToMerge,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
......@@ -1039,7 +1039,7 @@ struct DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle
remove_reference_t<
DeviceOp::CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>,
ComputePtrOffsetOfStridedBatch<I1, I1, I0>,
NumBatchToMerge,
NumGroupsToMerge,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
......@@ -1060,7 +1060,7 @@ struct DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle
remove_reference_t<
DeviceOp::CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>,
ComputePtrOffsetOfStridedBatch<I1, I1, I0>,
NumBatchToMerge,
NumGroupsToMerge,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
......@@ -1084,7 +1084,7 @@ struct DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle
remove_reference_t<
DeviceOp::CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>,
ComputePtrOffsetOfStridedBatch<I1, I1, I0>,
NumBatchToMerge,
NumGroupsToMerge,
true,
InMemoryDataOperationEnum::AtomicAdd,
minimum_occupancy,
......@@ -1100,7 +1100,7 @@ struct DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle
remove_reference_t<
DeviceOp::CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>,
ComputePtrOffsetOfStridedBatch<I1, I1, I0>,
NumBatchToMerge,
NumGroupsToMerge,
true,
InMemoryDataOperationEnum::AtomicAdd,
minimum_occupancy,
......@@ -1119,7 +1119,7 @@ struct DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle
remove_reference_t<
DeviceOp::CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>,
ComputePtrOffsetOfStridedBatch<I1, I1, I0>,
NumBatchToMerge,
NumGroupsToMerge,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
......@@ -1135,7 +1135,7 @@ struct DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle
remove_reference_t<
DeviceOp::CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>,
ComputePtrOffsetOfStridedBatch<I1, I1, I0>,
NumBatchToMerge,
NumGroupsToMerge,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
......@@ -1157,7 +1157,7 @@ struct DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle
remove_reference_t<
DeviceOp::CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>,
ComputePtrOffsetOfStridedBatch<I1, I1, I0>,
NumBatchToMerge,
NumGroupsToMerge,
true,
InMemoryDataOperationEnum::AtomicAdd,
minimum_occupancy,
......@@ -1173,7 +1173,7 @@ struct DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle
remove_reference_t<
DeviceOp::CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>,
ComputePtrOffsetOfStridedBatch<I1, I1, I0>,
NumBatchToMerge,
NumGroupsToMerge,
true,
InMemoryDataOperationEnum::AtomicAdd,
minimum_occupancy,
......@@ -1192,7 +1192,7 @@ struct DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle
remove_reference_t<
DeviceOp::CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>,
ComputePtrOffsetOfStridedBatch<I1, I1, I0>,
NumBatchToMerge,
NumGroupsToMerge,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
......@@ -1208,7 +1208,7 @@ struct DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle
remove_reference_t<
DeviceOp::CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>,
ComputePtrOffsetOfStridedBatch<I1, I1, I0>,
NumBatchToMerge,
NumGroupsToMerge,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
......@@ -1232,7 +1232,7 @@ struct DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle
remove_reference_t<
DeviceOp::CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>,
ComputePtrOffsetOfStridedBatch<I1, I1, I0>,
NumBatchToMerge,
NumGroupsToMerge,
false,
InMemoryDataOperationEnum::AtomicAdd,
minimum_occupancy>;
......@@ -1247,7 +1247,7 @@ struct DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle
remove_reference_t<
DeviceOp::CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>,
ComputePtrOffsetOfStridedBatch<I1, I1, I0>,
NumBatchToMerge,
NumGroupsToMerge,
false,
InMemoryDataOperationEnum::Set,
minimum_occupancy>;
......@@ -1389,7 +1389,7 @@ struct DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle
}
}
if constexpr(NumBatchToMerge > 1)
if constexpr(NumGroupsToMerge > 1)
{
// support only if whole M and N can be proccessed on one block
if(!(GemmM <= MPerBlock && GemmN <= NPerBlock))
......@@ -1400,7 +1400,7 @@ struct DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle
{
return false;
}
if(arg.Conv_G_ % NumBatchToMerge != 0)
if(arg.Conv_G_ % NumGroupsToMerge != 0)
{
return false;
}
......@@ -1563,7 +1563,7 @@ struct DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle
<< BlkGemmPipelineSchedulerToString[BlkGemmPipeSched] << ", "
<< "BlkGemmPipelineVersion: "
<< BlkGemmPipelineVersionToString[BlkGemmPipelineVer] << ", "
<< NumBatchToMerge
<< NumGroupsToMerge
<< ">";
// clang-format on
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_abd_xdl_cshuffle.hpp
View file @ 1762f081
......@@ -86,7 +86,6 @@ __global__ void
const AElementwiseOperation a_element_op,
const BElementwiseOperation b_element_op,
const CDEElementwiseOperation cde_element_op,
const index_t groups_count,
const AGridDesc_AK0_M_AK1 a_grid_desc_k0_m_k1,
const BGridDesc_BK0_N_BK1 b_grid_desc_k0_n_k1,
const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
......@@ -101,14 +100,11 @@ __global__ void
defined(__gfx94__))
// offset base pointer for each work-group
const index_t num_blocks_per_batch = __builtin_amdgcn_readfirstlane(gridDim.y / groups_count);
const index_t& num_blocks_per_n = groups_count;
const index_t g_idx = __builtin_amdgcn_readfirstlane(blockIdx.y / num_blocks_per_batch);
const index_t n_idx = __builtin_amdgcn_readfirstlane(blockIdx.y / num_blocks_per_n);
const long_index_t e_batch_offset =
const index_t g_idx = __builtin_amdgcn_readfirstlane(blockIdx.y);
const index_t n_idx = __builtin_amdgcn_readfirstlane(blockIdx.z);
const long_index_t e_group_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_groups.GetEPtrOffset(g_idx));
const auto& ds_batch_offset = compute_ptr_offset_of_groups.GetDsPtrOffset(g_idx);
const auto& ds_group_offset = compute_ptr_offset_of_groups.GetDsPtrOffset(g_idx);
const long_index_t e_n_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_n.GetEPtrOffset(n_idx));
......@@ -121,14 +117,14 @@ __global__ void
DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock::Size();
static_for<0, NumDTensor, 1>{}(
[&](auto i) { p_ds_grid_grp(i) = p_ds_grid[i] + ds_batch_offset[i]; });
[&](auto i) { p_ds_grid_grp(i) = p_ds_grid[i] + ds_group_offset[i]; });
if constexpr(isMultiA || isMultiB)
{
AsPointer p_as_grid_grp;
BsPointer p_bs_grid_grp;
const auto& as_batch_offset = compute_ptr_offset_of_groups.GetAsPtrOffset(g_idx);
const auto& as_group_offset = compute_ptr_offset_of_groups.GetAsPtrOffset(g_idx);
// compute_ptr_offset_of_n_ not need BatchStrideB so
// in case of MultiA is false but isMultiB is true
......@@ -139,27 +135,27 @@ __global__ void
static constexpr index_t NumATensor = AGridDesc_AK0_M_AK1::Size();
static_for<0, NumATensor, 1>{}([&](auto i) {
p_as_grid_grp(i) = p_as_grid[i] + as_batch_offset[i] + as_n_offset[i];
p_as_grid_grp(i) = p_as_grid[i] + as_group_offset[i] + as_n_offset[i];
});
}
else
{
const long_index_t a_n_offset = compute_ptr_offset_of_n.GetAPtrOffset(n_idx);
static_for<0, 1, 1>{}(
[&](auto i) { p_as_grid_grp(i) = p_as_grid[i] + as_batch_offset[i] + a_n_offset; });
[&](auto i) { p_as_grid_grp(i) = p_as_grid[i] + as_group_offset[i] + a_n_offset; });
}
const auto& bs_batch_offset = compute_ptr_offset_of_groups.GetBsPtrOffset(g_idx);
const auto& bs_group_offset = compute_ptr_offset_of_groups.GetBsPtrOffset(g_idx);
static constexpr index_t NumBTensor = BGridDesc_BK0_N_BK1::Size();
static_for<0, NumBTensor, 1>{}(
[&](auto i) { p_bs_grid_grp(i) = p_bs_grid[i] + bs_batch_offset[i]; });
[&](auto i) { p_bs_grid_grp(i) = p_bs_grid[i] + bs_group_offset[i]; });
GridwiseGemm::template Run<HasMainKBlockLoop>(
p_as_grid_grp,
p_bs_grid_grp,
p_ds_grid_grp,
p_e_grid + e_batch_offset + e_n_offset,
p_e_grid + e_group_offset + e_n_offset,
p_shared,
a_element_op,
b_element_op,
......@@ -172,19 +168,19 @@ __global__ void
}
else
{
const long_index_t a_batch_offset =
const long_index_t a_group_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_groups.GetAPtrOffset(g_idx));
const long_index_t b_batch_offset =
const long_index_t b_group_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_groups.GetBPtrOffset(g_idx));
const long_index_t a_n_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_n.GetAPtrOffset(n_idx));
GridwiseGemm::template Run<HasMainKBlockLoop>(
p_as_grid + a_batch_offset + a_n_offset,
p_bs_grid + b_batch_offset,
p_as_grid + a_group_offset + a_n_offset,
p_bs_grid + b_group_offset,
p_ds_grid_grp,
p_e_grid + e_batch_offset + e_n_offset,
p_e_grid + e_group_offset + e_n_offset,
p_shared,
a_element_op,
b_element_op,
......@@ -200,7 +196,6 @@ __global__ void
ignore = p_bs_grid;
ignore = p_ds_grid;
ignore = p_e_grid;
ignore = groups_count;
ignore = a_grid_desc_k0_m_k1;
ignore = b_grid_desc_k0_n_k1;
ignore = ds_grid_desc_mblock_mperblock_nblock_nperblock;
......@@ -287,7 +282,8 @@ template <index_t NDimSpatial,
// in tuple for MultiAB), unpack if tuple was
// passed
typename BComputeDataType = AComputeDataType,
LoopScheduler LoopSched = make_default_loop_scheduler()>
LoopScheduler LoopSched = make_default_loop_scheduler(),
index_t NumGroupsToMerge = 1>
struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
: public DeviceGroupedConvFwdMultipleABD<NDimSpatial,
ALayout,
......@@ -306,6 +302,8 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
{
using DeviceOp = DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle;
static_assert(NumGroupsToMerge >= 1);
static constexpr bool isMultiA = is_detected<is_tuple, ADataType>::value;
static constexpr bool isMultiB = is_detected<is_tuple, BDataType>::value;
......@@ -319,7 +317,7 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
static constexpr auto I3 = Number<3>{};
static constexpr auto conv_to_gemm_transformer =
TransformConvFwdToGemm<NDimSpatial, ConvForwardSpecialization>{};
TransformConvFwdToGemm<NDimSpatial, ConvForwardSpecialization, NumGroupsToMerge>{};
static constexpr auto matrix_padder =
MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, KPerBlock};
......@@ -550,7 +548,8 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
{
static_for<0, NumATensor, 1>{}([&](auto i) {
// Init compute_ptr_offset_of_groups_ for multiple AB
compute_ptr_offset_of_groups_.BatchStrideA_(i) = a_g_n_c_wis_strides[0];
compute_ptr_offset_of_groups_.BatchStrideA_(i) =
a_g_n_c_wis_strides[0] * NumGroupsToMerge;
// Use GemmADataType/GemmBDataType to iterate over tuple (even if passed data
// type is not tuple)
......@@ -578,7 +577,8 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
});
static_for<0, NumBTensor, 1>{}([&](auto i) {
// Init compute_ptr_offset_of_groups_ for multiple AB
compute_ptr_offset_of_groups_.BatchStrideB_(i) = b_g_k_c_xs_strides[0];
compute_ptr_offset_of_groups_.BatchStrideB_(i) =
b_g_k_c_xs_strides[0] * NumGroupsToMerge;
using DataType = remove_cvref_t<tuple_element_t<i.value, GemmBDataType>>;
// It is possible that one of the AB is a pointer and one is a tuple.
......@@ -598,8 +598,10 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
}
else
{
compute_ptr_offset_of_groups_.BatchStrideA_ = a_g_n_c_wis_strides[0];
compute_ptr_offset_of_groups_.BatchStrideB_ = b_g_k_c_xs_strides[0];
compute_ptr_offset_of_groups_.BatchStrideA_ =
a_g_n_c_wis_strides[0] * NumGroupsToMerge;
compute_ptr_offset_of_groups_.BatchStrideB_ =
b_g_k_c_xs_strides[0] * NumGroupsToMerge;
compute_ptr_offset_of_n_.BatchStrideA_ = a_g_n_c_wis_strides[1] * conv_N_per_block_;
// p_as and p_bs are pointers
......@@ -616,7 +618,8 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
p_ds_grid_(i) = static_cast<const DDataType*>(p_ds[i]);
// D batch stride
compute_ptr_offset_of_groups_.BatchStrideDs_(i) = ds_g_n_k_wos_strides[i][0];
compute_ptr_offset_of_groups_.BatchStrideDs_(i) =
ds_g_n_k_wos_strides[i][0] * NumGroupsToMerge;
compute_ptr_offset_of_n_.BatchStrideDs_(i) =
ds_g_n_k_wos_strides[i][1] * conv_N_per_block_;
......@@ -624,7 +627,7 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
ds_grid_desc_m_n_(i) = DeviceOp::MakeEGridDescriptor_M_N<DLayout>(
e_g_n_k_wos_lengths, ds_g_n_k_wos_strides[i], conv_N_per_block_);
});
compute_ptr_offset_of_groups_.BatchStrideE_ = e_g_n_k_wos_strides[0];
compute_ptr_offset_of_groups_.BatchStrideE_ = e_g_n_k_wos_strides[0] * NumGroupsToMerge;
compute_ptr_offset_of_n_.BatchStrideE_ = e_g_n_k_wos_strides[1] * conv_N_per_block_;
// populate desc for Ds/E
......@@ -745,8 +748,8 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
arg.a_g_n_c_wis_lengths_[I1] / arg.conv_N_per_block_;
const index_t gdx = arg.block_2_etile_map_.CalculateGridSize(arg.e_grid_desc_m_n_);
const index_t gdy = arg.num_group_ * num_workgroups_per_Conv_N;
const index_t gdz = 1;
const index_t gdy = arg.num_group_ / NumGroupsToMerge;
const index_t gdz = num_workgroups_per_Conv_N;
const auto K =
arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) * arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
......@@ -795,7 +798,6 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
arg.a_element_op_,
arg.b_element_op_,
arg.cde_element_op_,
arg.a_g_n_c_wis_lengths_[0], // Group count
as_grid_desc_ak0_m_ak1,
bs_grid_desc_bk0_n_bk1,
arg.ds_grid_desc_mblock_mperblock_nblock_nperblock_,
......@@ -839,7 +841,6 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
arg.a_element_op_,
arg.b_element_op_,
arg.cde_element_op_,
arg.a_g_n_c_wis_lengths_[0], // Group count
arg.a_grid_desc_ak0_m_ak1_,
arg.b_grid_desc_bk0_n_bk1_,
arg.ds_grid_desc_mblock_mperblock_nblock_nperblock_,
......@@ -871,6 +872,10 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
{
namespace ctc = tensor_layout::convolution;
const index_t G = arg.b_g_k_c_xs_lengths_[I0];
const index_t K = arg.b_g_k_c_xs_lengths_[I1];
const index_t C = arg.b_g_k_c_xs_lengths_[I2];
// check device
if(get_device_name() == "gfx908")
{
......@@ -919,6 +924,42 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
}
}
}
else if constexpr(ConvForwardSpecialization == ConvolutionForwardSpecialization::Filter3x3)
{
if(C != 1)
{
return false;
}
for(index_t i = 0; i < NDimSpatial; ++i)
{
const index_t filter_spatial_dim = arg.b_g_k_c_xs_lengths_[i + I3];
if(filter_spatial_dim != I3)
{
return false;
}
}
if constexpr(!is_NSpatialGK_GKSpatial_NSpatialGC<ALayout, BLayout, ELayout>())
{
return false;
}
}
if constexpr(NumGroupsToMerge > 1)
{
if(!(C == 1))
{
return false;
}
if(G % NumGroupsToMerge != 0)
{
return false;
}
if constexpr(!is_NSpatialGK_GKSpatial_NSpatialGC<ALayout, BLayout, ELayout>())
{
return false;
}
}
// check vector access of A
// FIXME: layout
......@@ -928,11 +969,16 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
is_same_v<ALayout, ctc::NWGC> || is_same_v<ALayout, ctc::NHWGC> ||
is_same_v<ALayout, ctc::NDHWGC>)
{
const index_t C = arg.a_g_n_c_wis_lengths_[2];
// Check access per C
if(!(ABlockTransferSrcVectorDim == 2 && C % ABlockTransferSrcScalarPerVector == 0))
{
return false;
// If not possible, check access per G
if(!(ABlockTransferSrcVectorDim == 1 && C == 1 &&
is_NSpatialGK_GKSpatial_NSpatialGC<ALayout, BLayout, ELayout>() &&
G % ABlockTransferSrcScalarPerVector == 0))
{
return false;
}
}
}
else
......@@ -949,8 +995,6 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
is_same_v<BLayout, ctc::KZYXGC>)
{
const index_t C = arg.b_g_k_c_xs_lengths_[2];
if(!(BBlockTransferSrcVectorDim == 2 && C % BBlockTransferSrcScalarPerVector == 0))
{
return false;
......@@ -974,8 +1018,6 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
is_same_v<DLayout, ctc::NWGK> || is_same_v<DLayout, ctc::NHWGK> ||
is_same_v<DLayout, ctc::NDHWGK> || is_same_v<DLayout, ctc::G_K>)
{
const index_t K = arg.ds_g_n_k_wos_lengths_[i][2];
if(!(K % CDEBlockTransferScalarPerVector_NPerBlock == 0))
{
valid = false;
......@@ -1020,8 +1062,6 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
is_same_v<ELayout, ctc::NWGK> || is_same_v<ELayout, ctc::NHWGK> ||
is_same_v<ELayout, ctc::NDHWGK>)
{
const index_t K = arg.e_g_n_k_wos_lengths_[2];
if(!(K % CDEBlockTransferScalarPerVector_NPerBlock == 0))
{
return false;
......@@ -1172,7 +1212,8 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
<< BBlockTransferSrcScalarPerVector << ", "
<< CDEBlockTransferScalarPerVector_NPerBlock << ", "
<< CShuffleMXdlPerWavePerShuffle << ", "
<< CShuffleNXdlPerWavePerShuffle
<< CShuffleNXdlPerWavePerShuffle << ", "
<< NumGroupsToMerge
<< ">";
// clang-format on
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_utils.hpp
View file @ 1762f081
......@@ -59,6 +59,22 @@ constexpr bool is_GNDHWK_GKZYXC_GNDHWC()
is_same_v<OutLayout, tensor_layout::convolution::GNDHWK>;
}
template <typename InLayout, typename WeiLayout, typename OutLayout>
constexpr bool is_NSpatialGK_GKSpatial_NSpatialGC()
{
return is_NWGK_GKXC_NWGC<InLayout, WeiLayout, OutLayout>() ||
is_NHWGK_GKYXC_NHWGC<InLayout, WeiLayout, OutLayout>() ||
is_NDHWGK_GKZYXC_NDHWGC<InLayout, WeiLayout, OutLayout>();
}
template <typename InLayout, typename WeiLayout, typename OutLayout>
constexpr bool is_GNSpatialK_GKSpatial_GNSpatialC()
{
return is_GNWK_GKXC_GNWC<InLayout, WeiLayout, OutLayout>() ||
is_GNHWK_GKYXC_GNHWC<InLayout, WeiLayout, OutLayout>() ||
is_GNDHWK_GKZYXC_GNDHWC<InLayout, WeiLayout, OutLayout>();
}
template <index_t NumATensor = 1, index_t NumBTensor = 1, index_t NumDTensor = 0, typename = void>
struct ComputePtrOffsetOfStridedBatch
{
......
include/ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp
View file @ 1762f081
......@@ -431,7 +431,7 @@ struct Relu
// https://paperswithcode.com/method/gelu
// y = 0.5*x*(1+tanh(sqrt(2/pi)*(x+0.044715*x^3)))
// host code use higher accuracy "exp" and "div"
// gpu code use lower accuracy "__expf" and "rcp" function
// gpu code use lower accuracy "_ocml_exp_f32" and "rcp" function
struct FastGelu
{
template <typename Y, typename X>
......@@ -451,7 +451,7 @@ struct FastGelu
y = x / (1.f + emu);
}
// device code, use lower precision "__expf" and "rcp"
// device code, use lower precision "__ocml_exp_f32" and "rcp"
template <>
__device__ void operator()<float, float>(float& y, const float& x) const
{
......@@ -459,7 +459,7 @@ struct FastGelu
const float c1 = -2.0 * 0.035677f;
const float c2 = -2.0 * 0.797885f;
const float u = x * (c1 * x * x + c2);
const float emu = __expf(u);
const float emu = __ocml_exp_f32(u);
y = x * ck::math::rcp(1.f + emu);
}
......@@ -1025,6 +1025,31 @@ struct ConvScale
float scale_out_;
};
struct ConvScaleRelu
{
__host__ __device__ ConvScaleRelu(float scale_in = 1.f,
float scale_wei = 1.f,
float scale_out = 1.f)
: scale_in_(scale_in), scale_wei_(scale_wei), scale_out_(scale_out)
{
}
template <typename E, typename C>
__host__ __device__ void operator()(E& e, const C& c) const;
template <>
__host__ __device__ void operator()<f8_t, float>(f8_t& e, const float& c) const
{
float x;
Relu{}.template operator()<float>(x, c * scale_in_ * scale_wei_);
e = type_convert<f8_t>(x * scale_out_);
};
float scale_in_;
float scale_wei_;
float scale_out_;
};
// support fastconvert of int8 to fp16
template <typename InputDataType, typename OutputDataType, index_t RegPackNumber>
......
include/ck/tensor_operation/operator_transform/transform_conv_bwd_weight_to_gemm_v2.hpp
View file @ 1762f081
......@@ -27,7 +27,7 @@ template <index_t NDimSpatial,
index_t NPerBlock,
index_t GemmK1Number,
index_t K0PerBlock,
index_t NumBatchToMerge,
index_t NumGroupsToMerge,
device::ConvolutionBackwardWeightSpecialization ConvBackwardWeightSpecialization>
struct TransformConvBwdWeightToGemmV2
{
......@@ -45,7 +45,7 @@ struct TransformConvBwdWeightToGemmV2
const index_t BatchStride = output_strides[0];
const index_t WoStride = output_strides[4];
const auto KStride = Number<1>{};
return make_naive_tensor_descriptor(make_tuple(N * Ho * Wo, NumBatchToMerge, K),
return make_naive_tensor_descriptor(make_tuple(N * Ho * Wo, NumGroupsToMerge, K),
make_tuple(WoStride, BatchStride, KStride));
}
......@@ -65,13 +65,13 @@ struct TransformConvBwdWeightToGemmV2
if constexpr(ConvBackwardWeightSpecialization ==
device::ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0)
{
return make_naive_tensor_descriptor(make_tuple(N * Hi * Wi, NumBatchToMerge, C),
return make_naive_tensor_descriptor(make_tuple(N * Hi * Wi, NumGroupsToMerge, C),
make_tuple(WiStride, BatchStride, CStride));
}
else
{
return make_naive_tensor_descriptor(
make_tuple(N, Hi, Wi, NumBatchToMerge, C),
make_tuple(N, Hi, Wi, NumGroupsToMerge, C),
make_tuple(NStride, HiStride, WiStride, BatchStride, CStride));
}
}
......@@ -88,30 +88,30 @@ struct TransformConvBwdWeightToGemmV2
const auto KStride = weights_strides[1];
const auto XStride = weights_strides[4];
const auto BatchStride = weights_strides[0];
// Add NumBatchToMerge for Batch+M dimension and, 1 as a placehorder
// Add NumGroupsToMerge for Batch+M dimension and, 1 as a placehorder
// for Batch+N dimension
const auto desc = make_naive_tensor_descriptor(
make_tuple(NumBatchToMerge, K, Y * X, 1, C),
make_tuple(NumGroupsToMerge, K, Y * X, 1, C),
make_tuple(BatchStride, KStride, XStride, BatchStride, CStride));
// Padd 1 to NumBatchToMerge
// Padd 1 to NumGroupsToMerge
const auto padded_desc = transform_tensor_descriptor(
desc,
make_tuple(make_pass_through_transform(NumBatchToMerge),
make_tuple(make_pass_through_transform(NumGroupsToMerge),
make_pass_through_transform(K),
make_pass_through_transform(Y * X),
make_pad_transform(1, 0, NumBatchToMerge - 1),
make_pad_transform(1, 0, NumGroupsToMerge - 1),
make_pass_through_transform(C)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}));
// We need only matrices from diagonal. Xor returns 0 for the same
// values. So if matrices is not on diagonal then it will be stored in padding.
// To avoid use of modulo after xor we assume that NumBatch to merge is power of 2.
static_assert(NumBatchToMerge == 1 || NumBatchToMerge == 2 || NumBatchToMerge == 4 ||
NumBatchToMerge == 8 || NumBatchToMerge == 16 || NumBatchToMerge == 32 ||
NumBatchToMerge == 64);
static_assert(NumGroupsToMerge == 1 || NumGroupsToMerge == 2 || NumGroupsToMerge == 4 ||
NumGroupsToMerge == 8 || NumGroupsToMerge == 16 || NumGroupsToMerge == 32 ||
NumGroupsToMerge == 64);
const auto unmerged_padded_desc = transform_tensor_descriptor(
padded_desc,
make_tuple(make_xor_transform(make_tuple(NumBatchToMerge, NumBatchToMerge)),
make_tuple(make_xor_transform(make_tuple(NumGroupsToMerge, NumGroupsToMerge)),
make_pass_through_transform(K),
make_pass_through_transform(Y * X),
make_pass_through_transform(C)),
......@@ -120,8 +120,8 @@ struct TransformConvBwdWeightToGemmV2
// Merge To M, N
return transform_tensor_descriptor(
unmerged_padded_desc,
make_tuple(make_merge_transform(make_tuple(NumBatchToMerge, K)),
make_merge_transform(make_tuple(Y * X, NumBatchToMerge, C))),
make_tuple(make_merge_transform(make_tuple(NumGroupsToMerge, K)),
make_merge_transform(make_tuple(Y * X, NumGroupsToMerge, C))),
make_tuple(Sequence<0, 1>{}, Sequence<2, 3, 4>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
}
......@@ -138,7 +138,7 @@ struct TransformConvBwdWeightToGemmV2
const index_t BatchStride = output_strides[0];
const index_t WoStride = output_strides[5];
const auto KStride = Number<1>{};
return make_naive_tensor_descriptor(make_tuple(N * Do * Ho * Wo, NumBatchToMerge, K),
return make_naive_tensor_descriptor(make_tuple(N * Do * Ho * Wo, NumGroupsToMerge, K),
make_tuple(WoStride, BatchStride, KStride));
}
......@@ -160,13 +160,13 @@ struct TransformConvBwdWeightToGemmV2
if constexpr(ConvBackwardWeightSpecialization ==
device::ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0)
{
return make_naive_tensor_descriptor(make_tuple(N * Di * Hi * Wi, NumBatchToMerge, C),
return make_naive_tensor_descriptor(make_tuple(N * Di * Hi * Wi, NumGroupsToMerge, C),
make_tuple(WiStride, BatchStride, CStride));
}
else
{
return make_naive_tensor_descriptor(
make_tuple(N, Di, Hi, Wi, NumBatchToMerge, C),
make_tuple(N, Di, Hi, Wi, NumGroupsToMerge, C),
make_tuple(NStride, DiStride, HiStride, WiStride, BatchStride, CStride));
}
}
......@@ -184,29 +184,29 @@ struct TransformConvBwdWeightToGemmV2
const auto KStride = weights_strides[1];
const auto XStride = weights_strides[5];
const auto BatchStride = weights_strides[0];
// Add NumBatchToMerge for Batch+M dimension and, 1 for placehord for Batch+N dimension
// Add NumGroupsToMerge for Batch+M dimension and, 1 for placehord for Batch+N dimension
const auto desc = make_naive_tensor_descriptor(
make_tuple(NumBatchToMerge, K, Z * Y * X, 1, C),
make_tuple(NumGroupsToMerge, K, Z * Y * X, 1, C),
make_tuple(BatchStride, KStride, XStride, BatchStride, CStride));
// Padd 1 to NumBatchToMerge
// Padd 1 to NumGroupsToMerge
const auto padded_desc = transform_tensor_descriptor(
desc,
make_tuple(make_pass_through_transform(NumBatchToMerge),
make_tuple(make_pass_through_transform(NumGroupsToMerge),
make_pass_through_transform(K),
make_pass_through_transform(Z * Y * X),
make_pad_transform(1, 0, NumBatchToMerge - 1),
make_pad_transform(1, 0, NumGroupsToMerge - 1),
make_pass_through_transform(C)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}));
// We need only matrices from diagonal. Xor returns 0 for the same
// values. So if matrices is not on diagonal then it will be stored in padding.
// To avoid use of modulo after xor we assume that NumBatch to merge is power of 2.
static_assert(NumBatchToMerge == 1 || NumBatchToMerge == 2 || NumBatchToMerge == 4 ||
NumBatchToMerge == 8 || NumBatchToMerge == 16 || NumBatchToMerge == 32 ||
NumBatchToMerge == 64);
static_assert(NumGroupsToMerge == 1 || NumGroupsToMerge == 2 || NumGroupsToMerge == 4 ||
NumGroupsToMerge == 8 || NumGroupsToMerge == 16 || NumGroupsToMerge == 32 ||
NumGroupsToMerge == 64);
const auto unmerged_padded_desc = transform_tensor_descriptor(
padded_desc,
make_tuple(make_xor_transform(make_tuple(NumBatchToMerge, NumBatchToMerge)),
make_tuple(make_xor_transform(make_tuple(NumGroupsToMerge, NumGroupsToMerge)),
make_pass_through_transform(K),
make_pass_through_transform(Z * Y * X),
make_pass_through_transform(C)),
......@@ -215,8 +215,8 @@ struct TransformConvBwdWeightToGemmV2
// Merge To M, N
return transform_tensor_descriptor(
unmerged_padded_desc,
make_tuple(make_merge_transform(make_tuple(NumBatchToMerge, K)),
make_merge_transform(make_tuple(Z * Y * X, NumBatchToMerge, C))),
make_tuple(make_merge_transform(make_tuple(NumGroupsToMerge, K)),
make_merge_transform(make_tuple(Z * Y * X, NumGroupsToMerge, C))),
make_tuple(Sequence<0, 1>{}, Sequence<2, 3, 4>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
}
......@@ -262,8 +262,8 @@ struct TransformConvBwdWeightToGemmV2
const index_t InRightPadW = input_right_pads[1];
const index_t GemmKTotal = N * Ho * Wo;
const index_t GemmM = K * NumBatchToMerge;
const index_t GemmN = C * X * Y * NumBatchToMerge;
const index_t GemmM = K * NumGroupsToMerge;
const index_t GemmN = C * X * Y * NumGroupsToMerge;
const auto PadGemmM = MPerBlock - GemmM % MPerBlock;
const auto PadGemmN = NPerBlock - GemmN % NPerBlock;
......@@ -286,7 +286,7 @@ struct TransformConvBwdWeightToGemmV2
out_grid_desc,
make_tuple(
make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
make_merge_transform(make_tuple(NumBatchToMerge, GemmM / NumBatchToMerge))),
make_merge_transform(make_tuple(NumGroupsToMerge, GemmM / NumGroupsToMerge))),
make_tuple(Sequence<0>{}, Sequence<1, 2>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
......@@ -302,7 +302,7 @@ struct TransformConvBwdWeightToGemmV2
in_grid_desc,
make_tuple(
make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
make_merge_transform(make_tuple(NumBatchToMerge, GemmN / NumBatchToMerge))),
make_merge_transform(make_tuple(NumGroupsToMerge, GemmN / NumGroupsToMerge))),
make_tuple(Sequence<0>{}, Sequence<1, 2>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
......@@ -324,7 +324,7 @@ struct TransformConvBwdWeightToGemmV2
out_grid_desc,
make_tuple(
make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
make_merge_transform(make_tuple(NumBatchToMerge, GemmM / NumBatchToMerge))),
make_merge_transform(make_tuple(NumGroupsToMerge, GemmM / NumGroupsToMerge))),
make_tuple(Sequence<0>{}, Sequence<1, 2>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
......@@ -341,7 +341,7 @@ struct TransformConvBwdWeightToGemmV2
make_tuple(make_pass_through_transform(N),
make_pad_transform(Hi, InLeftPadH, InRightPadH),
make_pad_transform(Wi, InLeftPadW, InRightPadW),
make_pass_through_transform(NumBatchToMerge),
make_pass_through_transform(NumGroupsToMerge),
make_pass_through_transform(C)),
make_tuple(
Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
......@@ -354,7 +354,7 @@ struct TransformConvBwdWeightToGemmV2
make_pass_through_transform(N),
make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
make_pass_through_transform(NumBatchToMerge),
make_pass_through_transform(NumGroupsToMerge),
make_pass_through_transform(C)),
make_tuple(
Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
......@@ -366,7 +366,7 @@ struct TransformConvBwdWeightToGemmV2
const auto in_gemmktotal_gemmn_grid_desc = transform_tensor_descriptor(
in_n_y_ho_x_wo_c_grid_desc,
make_tuple(make_merge_transform(make_tuple(Y, X, NumBatchToMerge, C)),
make_tuple(make_merge_transform(make_tuple(Y, X, NumGroupsToMerge, C)),
make_merge_transform(make_tuple(N, Ho, Wo))),
make_tuple(Sequence<1, 3, 5, 6>{}, Sequence<0, 2, 4>{}),
make_tuple(Sequence<1>{}, Sequence<0>{}));
......@@ -465,8 +465,8 @@ struct TransformConvBwdWeightToGemmV2
const index_t InRightPadW = input_right_pads[2];
const index_t GemmKTotal = N * Do * Ho * Wo;
const index_t GemmM = K * NumBatchToMerge;
const index_t GemmN = C * Z * X * Y * NumBatchToMerge;
const index_t GemmM = K * NumGroupsToMerge;
const index_t GemmN = C * Z * X * Y * NumGroupsToMerge;
const auto PadGemmM = MPerBlock - GemmM % MPerBlock;
const auto PadGemmN = NPerBlock - GemmN % NPerBlock;
......@@ -489,7 +489,7 @@ struct TransformConvBwdWeightToGemmV2
out_grid_desc,
make_tuple(
make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
make_merge_transform(make_tuple(NumBatchToMerge, GemmM / NumBatchToMerge))),
make_merge_transform(make_tuple(NumGroupsToMerge, GemmM / NumGroupsToMerge))),
make_tuple(Sequence<0>{}, Sequence<1, 2>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
......@@ -505,7 +505,7 @@ struct TransformConvBwdWeightToGemmV2
in_grid_desc,
make_tuple(
make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
make_merge_transform(make_tuple(NumBatchToMerge, GemmN / NumBatchToMerge))),
make_merge_transform(make_tuple(NumGroupsToMerge, GemmN / NumGroupsToMerge))),
make_tuple(Sequence<0>{}, Sequence<1, 2>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
......@@ -527,7 +527,7 @@ struct TransformConvBwdWeightToGemmV2
out_grid_desc,
make_tuple(
make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
make_merge_transform(make_tuple(NumBatchToMerge, GemmM / NumBatchToMerge))),
make_merge_transform(make_tuple(NumGroupsToMerge, GemmM / NumGroupsToMerge))),
make_tuple(Sequence<0>{}, Sequence<1, 2>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
......@@ -545,7 +545,7 @@ struct TransformConvBwdWeightToGemmV2
make_pad_transform(Di, InLeftPadD, InRightPadD),
make_pad_transform(Hi, InLeftPadH, InRightPadH),
make_pad_transform(Wi, InLeftPadW, InRightPadW),
make_pass_through_transform(NumBatchToMerge),
make_pass_through_transform(NumGroupsToMerge),
make_pass_through_transform(C)),
make_tuple(Sequence<0>{},
Sequence<1>{},
......@@ -567,7 +567,7 @@ struct TransformConvBwdWeightToGemmV2
make_embed_transform(make_tuple(Z, Do), make_tuple(ConvDilationD, ConvStrideD)),
make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
make_pass_through_transform(NumBatchToMerge),
make_pass_through_transform(NumGroupsToMerge),
make_pass_through_transform(C)),
make_tuple(Sequence<0>{},
Sequence<1>{},
......@@ -584,7 +584,7 @@ struct TransformConvBwdWeightToGemmV2
const auto in_gemmktotal_gemmn_grid_desc = transform_tensor_descriptor(
in_n_z_do_y_ho_x_wo_c_grid_desc,
make_tuple(make_merge_transform(make_tuple(Z, Y, X, NumBatchToMerge, C)),
make_tuple(make_merge_transform(make_tuple(Z, Y, X, NumGroupsToMerge, C)),
make_merge_transform(make_tuple(N, Do, Ho, Wo))),
make_tuple(Sequence<1, 3, 5, 7, 8>{}, Sequence<0, 2, 4, 6>{}),
make_tuple(Sequence<1>{}, Sequence<0>{}));
......
include/ck/utility/math_v2.hpp
View file @ 1762f081
......@@ -839,7 +839,7 @@ inline __device__ T rcp(T x)
template <typename T>
inline __device__ T exp(T x)
{
return ck::type_convert<T>(__expf(ck::type_convert<float>(x)));
return ck::type_convert<T>(__ocml_exp_f32(ck::type_convert<float>(x)));
};
template <>
......@@ -851,7 +851,7 @@ inline __device__ half_t exp<half_t>(half_t x)
template <>
inline __device__ float exp<float>(float x)
{
return __expf(x);
return __ocml_exp_f32(x);
};
template <>
......
include/ck_tile/core/numeric/bfloat16.hpp
View file @ 1762f081
......@@ -331,7 +331,10 @@ bfloat16_t sqrt(bfloat16_t x)
};
CK_TILE_DEVICE
bfloat16_t exp(bfloat16_t x) { return static_cast<bfloat16_t>(__expf(static_cast<float>(x))); };
bfloat16_t exp(bfloat16_t x)
{
return static_cast<bfloat16_t>(__ocml_exp_f32(static_cast<float>(x)));
};
CK_TILE_DEVICE
bfloat16_t exp2(bfloat16_t x) { return static_cast<bfloat16_t>(exp2f(static_cast<float>(x))); };
......
include/ck_tile/core/numeric/float8.hpp
View file @ 1762f081
......@@ -835,7 +835,7 @@ CK_TILE_DEVICE
fp8_t sqrt(fp8_t x) { return static_cast<fp8_t>(__builtin_amdgcn_sqrtf(static_cast<float>(x))); };
CK_TILE_DEVICE
fp8_t exp(fp8_t x) { return static_cast<fp8_t>(__expf(static_cast<float>(x))); };
fp8_t exp(fp8_t x) { return static_cast<fp8_t>(__ocml_exp_f32(static_cast<float>(x))); };
CK_TILE_DEVICE
fp8_t exp2(fp8_t x) { return static_cast<fp8_t>(exp2f(static_cast<float>(x))); };
......@@ -860,7 +860,7 @@ CK_TILE_DEVICE
bf8_t sqrt(bf8_t x) { return static_cast<bf8_t>(__builtin_amdgcn_sqrtf(static_cast<float>(x))); };
CK_TILE_DEVICE
bf8_t exp(bf8_t x) { return static_cast<bf8_t>(__expf(static_cast<float>(x))); };
bf8_t exp(bf8_t x) { return static_cast<bf8_t>(__ocml_exp_f32(static_cast<float>(x))); };
CK_TILE_DEVICE
bf8_t exp2(bf8_t x) { return static_cast<bf8_t>(exp2f(static_cast<float>(x))); };
......
include/ck_tile/core/numeric/half.hpp
View file @ 1762f081
......@@ -374,7 +374,7 @@ half_t sqrt(half_t x)
};
CK_TILE_DEVICE
half_t exp(half_t x) { return static_cast<half_t>(__expf(static_cast<float>(x))); };
half_t exp(half_t x) { return static_cast<half_t>(__ocml_exp_f32(static_cast<float>(x))); };
CK_TILE_DEVICE
half_t exp2(half_t x) { return static_cast<half_t>(exp2f(static_cast<float>(x))); };
......
include/ck_tile/core/numeric/math.hpp
View file @ 1762f081
......@@ -519,7 +519,7 @@ CK_TILE_DEVICE
double sqrt(double x) { return __builtin_amdgcn_sqrt(x); };
CK_TILE_DEVICE
float exp(float x) { return __expf(x); };
float exp(float x) { return __ocml_exp_f32(x); };
CK_TILE_HOST
float exp(float x) { return std::expf(x); }
......
library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_xdl_instance.hpp
View file @ 1762f081
......@@ -40,10 +40,10 @@ template <ck::index_t NDimSpatial,
BlockGemmPipelineVersion PipelineVersion>
using device_grouped_conv_bwd_weight_two_stage_xdl_c_shuffle_f16_instances = std::tuple<
// clang-format off
//#########################################| Num| InLayout| WeiLayout| OutLayout| InData| WeiData| OutData| AccData| In| Wei| Out| ConvBackward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransfer| CBlockTransfer| BlockGemm| BlockGemm| NumBatch|
//#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline| ToMerge|
//#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| |
//#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | |
//#########################################| Num| InLayout| WeiLayout| OutLayout| InData| WeiData| OutData| AccData| In| Wei| Out| ConvBackward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransfer| CBlockTransfer| BlockGemm| BlockGemm| NumGroups|
//#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline| ToMerge|
//#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| |
//#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | |
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1>,
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 32, 32, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2>,
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 32, 8, 32, 32, 1, 2, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 4>,
......
library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_merged_groups_instance.hpp 0 → 100644
View file @ 1762f081
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_abd_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using BF16 = ck::bhalf_t;
using F16 = ck::half_t;
using F32 = float;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using Empty_Tuple = ck::Tuple<>;
using namespace ck::tensor_layout::convolution;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto ConvFwdDefault =
ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
static constexpr auto ConvFwd3x3 = ConvolutionForwardSpecialization::Filter3x3;
static constexpr auto GemmMNKPadding = GemmSpecialization::MNKPadding;
template <index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
ConvolutionForwardSpecialization ConvSpec>
using device_grouped_conv_fwd_xdl_merged_groups_bf16_instances = std::tuple<
// clang-format off
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| ACompute| BCompute| BlockGemm| NumGroups|
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| Type| Type| Pipeline| ToMerge|
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| | | Scheduler| |
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// Instances with NumGroupsPerBatch > 1
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF16, BF16, F32, BF16, DsLayout, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 16, 16, 4, 4, 16, 16, 4, 1, S< 4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, BF16, BF16, LoopScheduler::Default, 8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF16, BF16, F32, BF16, DsLayout, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 16, 16, 4, 4, 16, 16, 4, 1, S< 4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, BF16, BF16, LoopScheduler::Default, 16>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF16, BF16, F32, BF16, DsLayout, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 16, 16, 4, 4, 16, 16, 4, 1, S< 4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, BF16, BF16, LoopScheduler::Default, 32>
// clang-format on
>;
template <index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
ConvolutionForwardSpecialization ConvSpec>
using device_grouped_conv_fwd_xdl_merged_groups_f16_instances = std::tuple<
// clang-format off
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// Instances with NumGroupsPerBatch > 1
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 16, 16, 4, 4, 16, 16, 4, 1, S< 4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, F16, F16, LoopScheduler::Default, 8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 16, 16, 4, 4, 16, 16, 4, 1, S< 4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, F16, F16, LoopScheduler::Default, 16>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 16, 16, 4, 4, 16, 16, 4, 1, S< 4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, F16, F16, LoopScheduler::Default, 32>
// clang-format on
>;
template <index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
ConvolutionForwardSpecialization ConvSpec>
using device_grouped_conv_fwd_xdl_merged_groups_f32_instances = std::tuple<
// clang-format off
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// Instances with NumGroupsPerBatch > 1
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsLayout, F32, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 16, 16, 4, 4, 16, 16, 4, 1, S< 4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, F32, F32, LoopScheduler::Default, 8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsLayout, F32, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 16, 16, 4, 4, 16, 16, 4, 1, S< 4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, F32, F32, LoopScheduler::Default, 16>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsLayout, F32, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 16, 16, 4, 4, 16, 16, 4, 1, S< 4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, F32, F32, LoopScheduler::Default, 32>
// clang-format on
>;
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_outelementop_instance.hpp
View file @ 1762f081
......@@ -147,6 +147,43 @@ using device_grouped_conv_fwd_xdl_outelementop_f8_bf8_instances = std::tuple<
// clang-format on
>;
template <index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
ConvolutionForwardSpecialization ConvSpec,
typename OutElementOp>
using device_grouped_conv_fwd_xdl_outelementop_bf8_f8_instances = std::tuple<
// clang-format off
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Compute| Compute|
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| TypeA| TypeB|
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| | |
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(CK_ENABLE_FP8) && defined(CK_ENABLE_BF8)
// generic instance
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF8, F8, F32, F32, Tuple<>, F8, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, BF8, F8>,
// instances for small conv.K and conv.C
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF8, F8, F32, F32, Tuple<>, F8, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, BF8, F8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF8, F8, F32, F32, Tuple<>, F8, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, BF8, F8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF8, F8, F32, F32, Tuple<>, F8, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, BF8, F8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF8, F8, F32, F32, Tuple<>, F8, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, BF8, F8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF8, F8, F32, F32, Tuple<>, F8, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, BF8, F8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF8, F8, F32, F32, Tuple<>, F8, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, BF8, F8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF8, F8, F32, F32, Tuple<>, F8, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF8, F8, F32, F32, Tuple<>, F8, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, BF8, F8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF8, F8, F32, F32, Tuple<>, F8, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF8, F8, F32, F32, Tuple<>, F8, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, BF8, F8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF8, F8, F32, F32, Tuple<>, F8, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, BF8, F8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF8, F8, F32, F32, Tuple<>, F8, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF8, F8, F32, F32, Tuple<>, F8, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, BF8, F8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF8, F8, F32, F32, Tuple<>, F8, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF8, F8, F32, F32, Tuple<>, F8, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8>
#endif
// clang-format on
>;
} // namespace instance
} // namespace device
} // namespace tensor_operation
......
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp
View file @ 1762f081
......@@ -17,6 +17,7 @@
#endif
#ifdef CK_USE_XDL
#include "grouped_convolution_forward_xdl.inc"
#include "grouped_convolution_forward_xdl_merged_groups.inc"
#include "grouped_convolution_forward_comp_xdl.inc"
#include "grouped_convolution_forward_mem_inter_xdl.inc"
#include "grouped_convolution_forward_mem_intra_xdl.inc"
......@@ -199,6 +200,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
is_same_v<BComputeType, float>)
{
add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instances(op_ptrs);
add_device_grouped_conv2d_fwd_xdl_merged_groups_nhwgc_gkyxc_nhwgk_f32_instances(
op_ptrs);
add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_comp_instances(op_ptrs);
add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_mem_intra_instances(
op_ptrs);
......@@ -212,6 +215,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
is_same_v<BComputeType, half_t>)
{
add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instances(op_ptrs);
add_device_grouped_conv2d_fwd_xdl_merged_groups_nhwgc_gkyxc_nhwgk_f16_instances(
op_ptrs);
add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_comp_instances(op_ptrs);
add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_mem_intra_instances(
op_ptrs);
......@@ -227,6 +232,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
is_same_v<BComputeType, ck::bhalf_t>)
{
add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_instances(op_ptrs);
add_device_grouped_conv2d_fwd_xdl_merged_groups_nhwgc_gkyxc_nhwgk_bf16_instances(
op_ptrs);
add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_instances(op_ptrs);
add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_mem_intra_instances(
op_ptrs);
......@@ -284,6 +291,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
is_same_v<BComputeType, float>)
{
add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instances(op_ptrs);
add_device_grouped_conv3d_fwd_xdl_merged_groups_ndhwgc_gkzyxc_ndhwgk_f32_instances(
op_ptrs);
add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_comp_instances(op_ptrs);
add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_mem_intra_instances(
op_ptrs);
......@@ -338,6 +347,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
is_same_v<BComputeType, half_t>)
{
add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instances(op_ptrs);
add_device_grouped_conv3d_fwd_xdl_merged_groups_ndhwgc_gkzyxc_ndhwgk_f16_instances(
op_ptrs);
add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_comp_instances(op_ptrs);
add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_mem_intra_instances(
op_ptrs);
......@@ -353,6 +364,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
is_same_v<BComputeType, ck::bhalf_t>)
{
add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instances(op_ptrs);
add_device_grouped_conv3d_fwd_xdl_merged_groups_ndhwgc_gkzyxc_ndhwgk_bf16_instances(
op_ptrs);
add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_comp_instances(op_ptrs);
add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_mem_intra_instances(
op_ptrs);
......
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