ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-14 10:09:41 +00:00
Code Issues Packages Projects Releases Wiki Activity

Add read_first_lane function for int64 (#1347)

Browse Source 
[ROCm/composable_kernel commit: 8faec23cb4]
This commit is contained in:
Bartłomiej Kocot
2024-06-18 22:05:30 +02:00
committed by GitHub
parent 92853de60e
commit 6935a2481c
10 changed files with 105 additions and 44 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp
View File

@@ -93,9 +93,12 @@ __global__ void
__builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
const long_index_t a_batch_offset = compute_ptr_offset_of_batch.GetAPtrOffset(g_idx);
const long_index_t b_batch_offset = compute_ptr_offset_of_batch.GetBPtrOffset(g_idx);
const long_index_t e_batch_offset = compute_ptr_offset_of_batch.GetEPtrOffset(g_idx);
const long_index_t a_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx));
const long_index_t b_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx));
const long_index_t e_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_batch.GetEPtrOffset(g_idx));
const auto ds_batch_offset = compute_ptr_offset_of_batch.GetDsPtrOffset(g_idx);

9
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_dl.hpp
View File

@@ -54,9 +54,12 @@ __global__ void
__builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
const long_index_t a_batch_offset = compute_ptr_offset_of_batch.GetAPtrOffset(g_idx);
const long_index_t b_batch_offset = compute_ptr_offset_of_batch.GetBPtrOffset(g_idx);
const long_index_t c_batch_offset = compute_ptr_offset_of_batch.GetCPtrOffset(g_idx);
const long_index_t a_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx));
const long_index_t b_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx));
const long_index_t c_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_batch.GetCPtrOffset(g_idx));
__shared__ FloatAB p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(FloatAB)];

9
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_multiple_d_xdl_cshuffle.hpp
View File

@@ -66,9 +66,12 @@ __global__ void
__builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
const long_index_t a_batch_offset = compute_ptr_offset_of_batch.GetAPtrOffset(g_idx);
const long_index_t b_batch_offset = compute_ptr_offset_of_batch.GetBPtrOffset(g_idx);
const long_index_t c_batch_offset = compute_ptr_offset_of_batch.GetCPtrOffset(g_idx);
const long_index_t a_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx));
const long_index_t b_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx));
const long_index_t c_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_batch.GetCPtrOffset(g_idx));
__shared__ FloatA p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(FloatA)];

18
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_two_stage_xdl_cshuffle.hpp
View File

@@ -59,9 +59,12 @@ __global__ void
const index_t g_idx = __builtin_amdgcn_readfirstlane(blockIdx.z * NumBatchToMerge);
const index_t k_idx = __builtin_amdgcn_readfirstlane(blockIdx.y * num_k_per_block);
const long_index_t a_batch_offset = compute_ptr_offset_of_batch.GetAPtrOffset(g_idx);
const long_index_t b_batch_offset = compute_ptr_offset_of_batch.GetBPtrOffset(g_idx);
const long_index_t e_batch_offset = compute_ptr_offset_of_batch.GetEPtrOffset(g_idx);
const long_index_t a_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx));
const long_index_t b_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx));
const long_index_t e_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_batch.GetEPtrOffset(g_idx));
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
@@ -113,9 +116,12 @@ __global__ void
const index_t g_idx = __builtin_amdgcn_readfirstlane(blockIdx.z * NumBatchToMerge);
const index_t k_idx = __builtin_amdgcn_readfirstlane(blockIdx.y * num_k_per_block);
const long_index_t a_batch_offset = compute_ptr_offset_of_batch.GetAPtrOffset(g_idx);
const long_index_t b_batch_offset = compute_ptr_offset_of_batch.GetBPtrOffset(g_idx);
const long_index_t e_batch_offset = compute_ptr_offset_of_batch.GetEPtrOffset(g_idx);
const long_index_t a_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx));
const long_index_t b_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx));
const long_index_t e_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_batch.GetEPtrOffset(g_idx));
// Pass two lds pointer is the key to tell compiler that ds_read/write
// operate on different lds chunk at same time without order dependecy

9
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp
View File

@@ -97,9 +97,12 @@ __global__ void
__builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
const long_index_t a_batch_offset = compute_ptr_offset_of_batch.GetAPtrOffset(g_idx);
const long_index_t b_batch_offset = compute_ptr_offset_of_batch.GetBPtrOffset(g_idx);
const long_index_t c_batch_offset = compute_ptr_offset_of_batch.GetEPtrOffset(g_idx);
const long_index_t a_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx));
const long_index_t b_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx));
const long_index_t c_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_batch.GetEPtrOffset(g_idx));
const auto ds_batch_offset = compute_ptr_offset_of_batch.GetDsPtrOffset(g_idx);

17
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_abd_xdl_cshuffle.hpp
View File

@@ -106,10 +106,12 @@ __global__ void
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 = compute_ptr_offset_of_groups.GetEPtrOffset(g_idx);
const auto& ds_batch_offset = compute_ptr_offset_of_groups.GetDsPtrOffset(g_idx);
const long_index_t e_batch_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 long_index_t e_n_offset = compute_ptr_offset_of_n.GetEPtrOffset(n_idx);
const long_index_t e_n_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_n.GetEPtrOffset(n_idx));
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
@@ -170,10 +172,13 @@ __global__ void
}
else
{
const long_index_t a_batch_offset = compute_ptr_offset_of_groups.GetAPtrOffset(g_idx);
const long_index_t b_batch_offset = compute_ptr_offset_of_groups.GetBPtrOffset(g_idx);
const long_index_t a_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_groups.GetAPtrOffset(g_idx));
const long_index_t b_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_groups.GetBPtrOffset(g_idx));
const long_index_t a_n_offset = compute_ptr_offset_of_n.GetAPtrOffset(n_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,

30
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_abd_xdl_cshuffle_v3.hpp
View File

@@ -85,12 +85,17 @@ __global__ void
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 a_batch_offset = compute_ptr_offset_of_groups.GetAPtrOffset(g_idx);
const long_index_t b_batch_offset = compute_ptr_offset_of_groups.GetBPtrOffset(g_idx);
const long_index_t e_batch_offset = compute_ptr_offset_of_groups.GetEPtrOffset(g_idx);
const long_index_t a_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_groups.GetAPtrOffset(g_idx));
const long_index_t b_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_groups.GetBPtrOffset(g_idx));
const long_index_t e_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_groups.GetEPtrOffset(g_idx));
const long_index_t a_n_offset = compute_ptr_offset_of_n.GetAPtrOffset(n_idx);
const long_index_t e_n_offset = compute_ptr_offset_of_n.GetEPtrOffset(n_idx);
const long_index_t a_n_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_n.GetAPtrOffset(n_idx));
const long_index_t e_n_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_n.GetEPtrOffset(n_idx));
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
@@ -142,12 +147,17 @@ __global__ void
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 a_batch_offset = compute_ptr_offset_of_groups.GetAPtrOffset(g_idx);
const long_index_t b_batch_offset = compute_ptr_offset_of_groups.GetBPtrOffset(g_idx);
const long_index_t e_batch_offset = compute_ptr_offset_of_groups.GetEPtrOffset(g_idx);
const long_index_t a_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_groups.GetAPtrOffset(g_idx));
const long_index_t b_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_groups.GetBPtrOffset(g_idx));
const long_index_t e_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_groups.GetEPtrOffset(g_idx));
const long_index_t a_n_offset = compute_ptr_offset_of_n.GetAPtrOffset(n_idx);
const long_index_t e_n_offset = compute_ptr_offset_of_n.GetEPtrOffset(n_idx);
const long_index_t a_n_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_n.GetAPtrOffset(n_idx));
const long_index_t e_n_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_n.GetEPtrOffset(n_idx));
// Pass two lds pointer is the key to tell compiler that ds_read/write
// operate on different lds chunk at same time without order dependecy

6
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_multiple_r_xdl_cshuffle.hpp
View File

@@ -161,11 +161,11 @@ __global__ void
__builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
const long_index_t a_batch_offset = amd_wave_read_first_lane(
static_cast<long_index_t>(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx)));
const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
const long_index_t b_batch_offset = amd_wave_read_first_lane(
static_cast<long_index_t>(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx)));
const long_index_t e_batch_offset = __builtin_amdgcn_readfirstlane(
const long_index_t e_batch_offset = amd_wave_read_first_lane(
static_cast<long_index_t>(compute_ptr_offset_of_batch.GetEPtrOffset(g_idx)));
const auto ds_batch_offset = compute_ptr_offset_of_batch.GetDsPtrOffset(g_idx);

18
include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_wmma_cshuffle.hpp
View File

@@ -60,9 +60,12 @@ __global__ void
__builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
const long_index_t a_batch_offset = compute_ptr_offset_of_batch.GetAPtrOffset(g_idx);
const long_index_t b_batch_offset = compute_ptr_offset_of_batch.GetBPtrOffset(g_idx);
const long_index_t e_batch_offset = compute_ptr_offset_of_batch.GetEPtrOffset(g_idx);
const long_index_t a_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx));
const long_index_t b_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx));
const long_index_t e_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_batch.GetEPtrOffset(g_idx));
const auto ds_batch_offset = compute_ptr_offset_of_batch.GetDsPtrOffset(g_idx);
@@ -152,9 +155,12 @@ __global__ void
__builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
const long_index_t a_batch_offset = compute_ptr_offset_of_batch.GetAPtrOffset(g_idx);
const long_index_t b_batch_offset = compute_ptr_offset_of_batch.GetBPtrOffset(g_idx);
const long_index_t e_batch_offset = compute_ptr_offset_of_batch.GetEPtrOffset(g_idx);
const long_index_t a_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx));
const long_index_t b_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx));
const long_index_t e_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_batch.GetEPtrOffset(g_idx));
const auto ds_batch_offset = compute_ptr_offset_of_batch.GetDsPtrOffset(g_idx);

24
include/ck/utility/amd_wave_read_first_lane.hpp
View File

@@ -1,5 +1,5 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
@@ -95,11 +95,33 @@ using get_carrier_t = typename get_carrier<SizeInBytes>::type;
} // namespace detail
__device__ inline uint32_t amd_wave_read_first_lane(uint32_t value)
{
return __builtin_amdgcn_readfirstlane(value);
}
__device__ inline int32_t amd_wave_read_first_lane(int32_t value)
{
return __builtin_amdgcn_readfirstlane(value);
}
__device__ inline int64_t amd_wave_read_first_lane(int64_t value)
{
constexpr unsigned object_size = sizeof(int64_t);
constexpr unsigned second_part_offset = object_size / 2;
auto* const from_obj = reinterpret_cast<const std::byte*>(&value);
alignas(int64_t) std::byte to_obj[object_size];
using Sgpr = uint32_t;
*reinterpret_cast<Sgpr*>(to_obj) =
amd_wave_read_first_lane(*reinterpret_cast<const Sgpr*>(from_obj));
*reinterpret_cast<Sgpr*>(to_obj + second_part_offset) =
amd_wave_read_first_lane(*reinterpret_cast<const Sgpr*>(from_obj + second_part_offset));
return *reinterpret_cast<int64_t*>(to_obj);
}
template <
typename Object,
typename = std::enable_if_t<std::is_class_v<Object> && std::is_trivially_copyable_v<Object>>>