已合并
整改安装重名头文件：权威 TLA / attention/common / mc2/common / math 侧保留；改次要副本 #10499
chenyifan创建于 9 天前
整改安装重名头文件：权威 TLA / attention/common / mc2/common / math 侧保留；改次要副本 #10499
已合并
共 188 个文件变更+362-784
Rattention/block_sparse_attention/op_kernel/attn_infra/status.hpp→attention/block_sparse_attention/op_kernel/attn_infra/bsa_status.hpp+2-2
| @@ -8,8 +8,8 @@ | |||
| 8 | * See LICENSE in the root of the software repository for the full text of the License. | 8 | * See LICENSE in the root of the software repository for the full text of the License. |
| 9 | */ | 9 | */ |
| 10 | 10 | ||
| 11 | -#ifndef STATUS_HPP | 11 | +#ifndef BSA_STATUS_HPP |
| 12 | -#define STATUS_HPP | 12 | +#define BSA_STATUS_HPP |
| 13 | 13 | ||
| 14 | namespace NpuArch{ | 14 | namespace NpuArch{ |
| 15 | 15 | ||
| @@ -15,7 +15,7 @@ | |||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | 17 | ||
| 18 | -#include "catlass/epilogue/tile/tile_copy.hpp" | 18 | +#include "catlass/epilogue/tile/kda_epilogue_tile_copy.hpp" |
| 19 | 19 | ||
| 20 | namespace Catlass::Epilogue::Block { | 20 | namespace Catlass::Epilogue::Block { |
| 21 | 21 | ||
| @@ -15,7 +15,7 @@ | |||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | 17 | ||
| 18 | -#include "catlass/epilogue/tile/tile_copy.hpp" | 18 | +#include "catlass/epilogue/tile/kda_epilogue_tile_copy.hpp" |
| 19 | 19 | ||
| 20 | namespace Catlass::Epilogue::Block { | 20 | namespace Catlass::Epilogue::Block { |
| 21 | 21 | ||
| @@ -22,12 +22,12 @@ | |||
| 22 | 22 | ||
| 23 | 23 | ||
| 24 | 24 | ||
| 25 | -#include "catlass/gemm/dispatch_policy.hpp" | 25 | +#include "catlass/gemm/kda_gemm_dispatch_policy.hpp" |
| 26 | 26 | ||
| 27 | 27 | ||
| 28 | 28 | ||
| 29 | 29 | ||
| 30 | -#include "tla/layout.hpp" | 30 | +#include "tla/kda_tla_layout.hpp" |
| 31 | 31 | ||
| 32 | 32 | ||
| 33 | 33 | ||
| @@ -23,14 +23,14 @@ | |||
| 23 | 23 | ||
| 24 | 24 | ||
| 25 | 25 | ||
| 26 | -#include "catlass/gemm/dispatch_policy.hpp" | 26 | +#include "catlass/gemm/kda_gemm_dispatch_policy.hpp" |
| 27 | -#include "catlass/gemm/tile/tile_copy.hpp" | 27 | +#include "catlass/gemm/tile/kda_gemm_tile_copy.hpp" |
| 28 | 28 | ||
| 29 | 29 | ||
| 30 | 30 | ||
| 31 | 31 | ||
| 32 | 32 | ||
| 33 | -#include "tla/layout.hpp" | 33 | +#include "tla/kda_tla_layout.hpp" |
| 34 | 34 | ||
| 35 | 35 | ||
| 36 | using namespace AscendC; | 36 | using namespace AscendC; |
| @@ -13,7 +13,7 @@ | |||
| 13 | 13 | ||
| 14 | 14 | ||
| 15 | 15 | ||
| 16 | -#include "catlass/gemm/tile/tile_copy.hpp" | 16 | +#include "catlass/gemm/tile/kda_gemm_tile_copy.hpp" |
| 17 | 17 | ||
| 18 | 18 | ||
| 19 | 19 | ||
| @@ -23,8 +23,8 @@ | |||
| 23 | 23 | ||
| 24 | 24 | ||
| 25 | 25 | ||
| 26 | -#include "catlass/gemm/dispatch_policy.hpp" | 26 | +#include "catlass/gemm/kda_gemm_dispatch_policy.hpp" |
| 27 | -#include "catlass/gemm/tile/tile_copy.hpp" | 27 | +#include "catlass/gemm/tile/kda_gemm_tile_copy.hpp" |
| 28 | 28 | ||
| 29 | 29 | ||
| 30 | 30 | ||
| @@ -37,7 +37,7 @@ | |||
| 37 | 37 | ||
| 38 | 38 | ||
| 39 | 39 | ||
| 40 | -#include "tla/layout.hpp" | 40 | +#include "tla/kda_tla_layout.hpp" |
| 41 | 41 | ||
| 42 | 42 | ||
| 43 | using namespace AscendC; | 43 | using namespace AscendC; |
| @@ -23,8 +23,8 @@ | |||
| 23 | 23 | ||
| 24 | 24 | ||
| 25 | 25 | ||
| 26 | -#include "catlass/gemm/dispatch_policy.hpp" | 26 | +#include "catlass/gemm/kda_gemm_dispatch_policy.hpp" |
| 27 | -#include "catlass/gemm/tile/tile_copy.hpp" | 27 | +#include "catlass/gemm/tile/kda_gemm_tile_copy.hpp" |
| 28 | 28 | ||
| 29 | 29 | ||
| 30 | 30 | ||
| @@ -37,7 +37,7 @@ | |||
| 37 | 37 | ||
| 38 | 38 | ||
| 39 | 39 | ||
| 40 | -#include "tla/layout.hpp" | 40 | +#include "tla/kda_tla_layout.hpp" |
| 41 | 41 | ||
| 42 | 42 | ||
| 43 | 43 | ||
| @@ -15,7 +15,7 @@ | |||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | 17 | ||
| 18 | -#include "catlass/epilogue/tile/tile_copy.hpp" | 18 | +#include "catlass/epilogue/tile/kda_epilogue_tile_copy.hpp" |
| 19 | 19 | ||
| 20 | 20 | ||
| 21 | namespace Catlass::Epilogue::Block { | 21 | namespace Catlass::Epilogue::Block { |
| @@ -15,7 +15,7 @@ | |||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | 17 | ||
| 18 | -#include "catlass/epilogue/tile/tile_copy.hpp" | 18 | +#include "catlass/epilogue/tile/kda_epilogue_tile_copy.hpp" |
| 19 | 19 | ||
| 20 | 20 | ||
| 21 | namespace Catlass::Epilogue::Block { | 21 | namespace Catlass::Epilogue::Block { |
| @@ -24,12 +24,12 @@ | |||
| 24 | 24 | ||
| 25 | 25 | ||
| 26 | 26 | ||
| 27 | -#include "catlass/gemm/dispatch_policy.hpp" | 27 | +#include "catlass/gemm/kda_gemm_dispatch_policy.hpp" |
| 28 | 28 | ||
| 29 | 29 | ||
| 30 | 30 | ||
| 31 | 31 | ||
| 32 | -#include "tla/layout.hpp" | 32 | +#include "tla/kda_tla_layout.hpp" |
| 33 | 33 | ||
| 34 | 34 | ||
| 35 | using _0 = tla::Int<0>; | 35 | using _0 = tla::Int<0>; |
| @@ -12,7 +12,7 @@ | |||
| 12 | 12 | ||
| 13 | 13 | ||
| 14 | 14 | ||
| 15 | -#include "tla/layout.hpp" | 15 | +#include "tla/kda_tla_layout.hpp" |
| 16 | 16 | ||
| 17 | //////////////////////////////////////////////////////////////////////////////////////////////////// | 17 | //////////////////////////////////////////////////////////////////////////////////////////////////// |
| 18 | 18 | ||
| @@ -16,7 +16,7 @@ | |||
| 16 | 16 | ||
| 17 | 17 | ||
| 18 | 18 | ||
| 19 | -#include "catlass/epilogue/tile/tile_copy.hpp" | 19 | +#include "catlass/epilogue/tile/kda_epilogue_tile_copy.hpp" |
| 20 | 20 | ||
| 21 | namespace Catlass::Epilogue::Block { | 21 | namespace Catlass::Epilogue::Block { |
| 22 | 22 | ||
Mattention/chunk_kda_fwd/op_kernel/catlass/epilogue/block/block_epilogue_fa_rescale_o_ascend950.hpp+1-1
| @@ -17,7 +17,7 @@ | |||
| 17 | 17 | ||
| 18 | 18 | ||
| 19 | 19 | ||
| 20 | -#include "tla/layout.hpp" | 20 | +#include "tla/kda_tla_layout.hpp" |
| 21 | 21 | ||
| 22 | namespace Catlass::Epilogue::Block { | 22 | namespace Catlass::Epilogue::Block { |
| 23 | 23 | ||
Mattention/chunk_kda_fwd/op_kernel/catlass/epilogue/block/block_epilogue_fa_softmax_ascend950.hpp+1-1
| @@ -18,7 +18,7 @@ | |||
| 18 | 18 | ||
| 19 | 19 | ||
| 20 | 20 | ||
| 21 | -#include "tla/layout.hpp" | 21 | +#include "tla/kda_tla_layout.hpp" |
| 22 | 22 | ||
| 23 | namespace Catlass::Epilogue::Block { | 23 | namespace Catlass::Epilogue::Block { |
| 24 | 24 | ||
| @@ -16,7 +16,7 @@ | |||
| 16 | 16 | ||
| 17 | 17 | ||
| 18 | 18 | ||
| 19 | -#include "catlass/epilogue/tile/tile_copy.hpp" | 19 | +#include "catlass/epilogue/tile/kda_epilogue_tile_copy.hpp" |
| 20 | 20 | ||
| 21 | 21 | ||
| 22 | 22 | ||
| @@ -14,7 +14,7 @@ | |||
| 14 | 14 | ||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | -#include "catlass/epilogue/tile/tile_copy.hpp" | 17 | +#include "catlass/epilogue/tile/kda_epilogue_tile_copy.hpp" |
| 18 | 18 | ||
| 19 | 19 | ||
| 20 | 20 | ||
| @@ -14,7 +14,7 @@ | |||
| 14 | 14 | ||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | -#include "catlass/epilogue/tile/tile_copy.hpp" | 17 | +#include "catlass/epilogue/tile/kda_epilogue_tile_copy.hpp" |
| 18 | 18 | ||
| 19 | 19 | ||
| 20 | 20 | ||
| @@ -14,7 +14,7 @@ | |||
| 14 | 14 | ||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | -#include "catlass/epilogue/tile/tile_copy.hpp" | 17 | +#include "catlass/epilogue/tile/kda_epilogue_tile_copy.hpp" |
| 18 | 18 | ||
| 19 | 19 | ||
| 20 | 20 | ||
Mattention/chunk_kda_fwd/op_kernel/catlass/epilogue/block/block_epilogue_online_softmax_no_mask.hpp+1-1
| @@ -15,7 +15,7 @@ | |||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | 17 | ||
| 18 | -#include "catlass/epilogue/tile/tile_copy.hpp" | 18 | +#include "catlass/epilogue/tile/kda_epilogue_tile_copy.hpp" |
| 19 | 19 | ||
| 20 | 20 | ||
| 21 | 21 | ||
Mattention/chunk_kda_fwd/op_kernel/catlass/epilogue/block/block_epilogue_per_group_per_block.hpp+1-1
| @@ -17,7 +17,7 @@ | |||
| 17 | 17 | ||
| 18 | 18 | ||
| 19 | 19 | ||
| 20 | -#include "tla/layout.hpp" | 20 | +#include "tla/kda_tla_layout.hpp" |
| 21 | 21 | ||
| 22 | namespace Catlass::Epilogue::Block { | 22 | namespace Catlass::Epilogue::Block { |
| 23 | 23 | ||
Mattention/chunk_kda_fwd/op_kernel/catlass/epilogue/block/block_epilogue_per_token_dequant_tla.hpp+1-1
| @@ -19,7 +19,7 @@ | |||
| 19 | 19 | ||
| 20 | 20 | ||
| 21 | 21 | ||
| 22 | -#include "tla/layout.hpp" | 22 | +#include "tla/kda_tla_layout.hpp" |
| 23 | 23 | ||
| 24 | 24 | ||
| 25 | namespace Catlass::Epilogue::Block { | 25 | namespace Catlass::Epilogue::Block { |
Mattention/chunk_kda_fwd/op_kernel/catlass/epilogue/block/block_epilogue_rescale_o_no_split_row.hpp+1-1
| @@ -14,7 +14,7 @@ | |||
| 14 | 14 | ||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | -#include "catlass/epilogue/tile/tile_copy.hpp" | 17 | +#include "catlass/epilogue/tile/kda_epilogue_tile_copy.hpp" |
| 18 | 18 | ||
| 19 | 19 | ||
| 20 | 20 | ||
| @@ -17,7 +17,7 @@ | |||
| 17 | 17 | ||
| 18 | 18 | ||
| 19 | 19 | ||
| 20 | -#include "catlass/epilogue/tile/tile_copy.hpp" | 20 | +#include "catlass/epilogue/tile/kda_epilogue_tile_copy.hpp" |
| 21 | 21 | ||
| 22 | namespace Catlass::Epilogue::Block { | 22 | namespace Catlass::Epilogue::Block { |
| 23 | 23 | ||
| @@ -16,7 +16,7 @@ | |||
| 16 | 16 | ||
| 17 | 17 | ||
| 18 | 18 | ||
| 19 | -#include "catlass/epilogue/tile/tile_copy.hpp" | 19 | +#include "catlass/epilogue/tile/kda_epilogue_tile_copy.hpp" |
| 20 | 20 | ||
| 21 | 21 | ||
| 22 | namespace Catlass::Epilogue::Block { | 22 | namespace Catlass::Epilogue::Block { |
| @@ -15,7 +15,7 @@ | |||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | 17 | ||
| 18 | -#include "catlass/epilogue/tile/tile_copy.hpp" | 18 | +#include "catlass/epilogue/tile/kda_epilogue_tile_copy.hpp" |
| 19 | 19 | ||
| 20 | 20 | ||
| 21 | 21 | ||
Mattention/chunk_kda_fwd/op_kernel/catlass/epilogue/block/kda_block_epilogue_mla_tp1_softmax.hpp+1-1
| @@ -15,7 +15,7 @@ | |||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | 17 | ||
| 18 | -#include "catlass/epilogue/tile/tile_copy.hpp" | 18 | +#include "catlass/epilogue/tile/kda_epilogue_tile_copy.hpp" |
| 19 | 19 | ||
| 20 | 20 | ||
| 21 | 21 | ||
| @@ -14,7 +14,7 @@ | |||
| 14 | 14 | ||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | -#include "tla/layout.hpp" | 17 | +#include "tla/kda_tla_layout.hpp" |
| 18 | 18 | ||
| 19 | namespace Catlass::Epilogue::Tile { | 19 | namespace Catlass::Epilogue::Tile { |
| 20 | 20 | ||
| @@ -14,7 +14,7 @@ | |||
| 14 | 14 | ||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | -#include "tla/layout.hpp" | 17 | +#include "tla/kda_tla_layout.hpp" |
| 18 | 18 | ||
| 19 | namespace Catlass::Epilogue::Tile { | 19 | namespace Catlass::Epilogue::Tile { |
| 20 | 20 | ||
| @@ -14,7 +14,7 @@ | |||
| 14 | 14 | ||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | -#include "tla/layout.hpp" | 17 | +#include "tla/kda_tla_layout.hpp" |
| 18 | 18 | ||
| 19 | namespace Catlass::Epilogue::Tile { | 19 | namespace Catlass::Epilogue::Tile { |
| 20 | 20 | ||
Rattention/chunk_kda_fwd/op_kernel/catlass/epilogue/tile/tile_copy.hpp→attention/chunk_kda_fwd/op_kernel/catlass/epilogue/tile/kda_epilogue_tile_copy.hpp+3-3
| @@ -8,8 +8,8 @@ | |||
| 8 | * See LICENSE in the root of the software repository for the full text of the License. | 8 | * See LICENSE in the root of the software repository for the full text of the License. |
| 9 | */ | 9 | */ |
| 10 | 10 | ||
| 11 | -#ifndef CATLASS_EPILOGUE_TILE_TILE_COPY_HPP | 11 | +#ifndef CATLASS_KDA_EPILOGUE_TILE_COPY_HPP |
| 12 | -#define CATLASS_EPILOGUE_TILE_TILE_COPY_HPP | 12 | +#define CATLASS_KDA_EPILOGUE_TILE_COPY_HPP |
| 13 | 13 | ||
| 14 | 14 | ||
| 15 | 15 | ||
| @@ -189,4 +189,4 @@ struct TileCopyDequantTla { | |||
| 189 | 189 | ||
| 190 | } // namespace Catlass::Epilogue::Tile | 190 | } // namespace Catlass::Epilogue::Tile |
| 191 | 191 | ||
| 192 | -#endif // CATLASS_EPILOGUE_TILE_TILE_COPY_HPP | 192 | +#endif // CATLASS_KDA_EPILOGUE_TILE_COPY_HPP |
| @@ -15,9 +15,9 @@ | |||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | 17 | ||
| 18 | -#include "catlass/gemm/dispatch_policy.hpp" | 18 | +#include "catlass/gemm/kda_gemm_dispatch_policy.hpp" |
| 19 | 19 | ||
| 20 | -#include "catlass/gemm/tile/tile_copy.hpp" | 20 | +#include "catlass/gemm/tile/kda_gemm_tile_copy.hpp" |
| 21 | 21 | ||
| 22 | 22 | ||
| 23 | namespace Catlass::Gemm::Block { | 23 | namespace Catlass::Gemm::Block { |
| @@ -15,9 +15,9 @@ | |||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | 17 | ||
| 18 | -#include "catlass/gemm/dispatch_policy.hpp" | 18 | +#include "catlass/gemm/kda_gemm_dispatch_policy.hpp" |
| 19 | 19 | ||
| 20 | -#include "catlass/gemm/tile/tile_copy.hpp" | 20 | +#include "catlass/gemm/tile/kda_gemm_tile_copy.hpp" |
| 21 | 21 | ||
| 22 | 22 | ||
| 23 | namespace Catlass::Gemm::Block { | 23 | namespace Catlass::Gemm::Block { |
| @@ -16,9 +16,9 @@ | |||
| 16 | 16 | ||
| 17 | 17 | ||
| 18 | 18 | ||
| 19 | -#include "catlass/gemm/dispatch_policy.hpp" | 19 | +#include "catlass/gemm/kda_gemm_dispatch_policy.hpp" |
| 20 | 20 | ||
| 21 | -#include "catlass/gemm/tile/tile_copy.hpp" | 21 | +#include "catlass/gemm/tile/kda_gemm_tile_copy.hpp" |
| 22 | 22 | ||
| 23 | 23 | ||
| 24 | namespace Catlass::Gemm::Block { | 24 | namespace Catlass::Gemm::Block { |
| @@ -14,9 +14,9 @@ | |||
| 14 | 14 | ||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | -#include "catlass/gemm/dispatch_policy.hpp" | 17 | +#include "catlass/gemm/kda_gemm_dispatch_policy.hpp" |
| 18 | 18 | ||
| 19 | -#include "catlass/gemm/tile/tile_copy.hpp" | 19 | +#include "catlass/gemm/tile/kda_gemm_tile_copy.hpp" |
| 20 | 20 | ||
| 21 | 21 | ||
| 22 | 22 | ||
| @@ -15,9 +15,9 @@ | |||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | 17 | ||
| 18 | -#include "catlass/gemm/dispatch_policy.hpp" | 18 | +#include "catlass/gemm/kda_gemm_dispatch_policy.hpp" |
| 19 | 19 | ||
| 20 | -#include "catlass/gemm/tile/tile_copy.hpp" | 20 | +#include "catlass/gemm/tile/kda_gemm_tile_copy.hpp" |
| 21 | 21 | ||
| 22 | 22 | ||
| 23 | namespace Catlass::Gemm::Block { | 23 | namespace Catlass::Gemm::Block { |
| @@ -15,9 +15,9 @@ | |||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | 17 | ||
| 18 | -#include "catlass/gemm/dispatch_policy.hpp" | 18 | +#include "catlass/gemm/kda_gemm_dispatch_policy.hpp" |
| 19 | 19 | ||
| 20 | -#include "catlass/gemm/tile/tile_copy.hpp" | 20 | +#include "catlass/gemm/tile/kda_gemm_tile_copy.hpp" |
| 21 | 21 | ||
| 22 | 22 | ||
| 23 | namespace Catlass::Gemm::Block { | 23 | namespace Catlass::Gemm::Block { |
| @@ -14,10 +14,10 @@ | |||
| 14 | 14 | ||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | -#include "catlass/gemm/dispatch_policy.hpp" | 17 | +#include "catlass/gemm/kda_gemm_dispatch_policy.hpp" |
| 18 | 18 | ||
| 19 | 19 | ||
| 20 | -#include "catlass/gemm/tile/tile_copy.hpp" | 20 | +#include "catlass/gemm/tile/kda_gemm_tile_copy.hpp" |
| 21 | 21 | ||
| 22 | 22 | ||
| 23 | //////////////////////////////////////////////////////////////////// | 23 | //////////////////////////////////////////////////////////////////// |
| @@ -14,12 +14,12 @@ | |||
| 14 | 14 | ||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | -#include "catlass/gemm/dispatch_policy.hpp" | 17 | +#include "catlass/gemm/kda_gemm_dispatch_policy.hpp" |
| 18 | 18 | ||
| 19 | 19 | ||
| 20 | -#include "catlass/gemm/tile/tile_copy.hpp" | 20 | +#include "catlass/gemm/tile/kda_gemm_tile_copy.hpp" |
| 21 | 21 | ||
| 22 | -#include "tla/layout.hpp" | 22 | +#include "tla/kda_tla_layout.hpp" |
| 23 | 23 | ||
| 24 | 24 | ||
| 25 | //////////////////////////////////////////////////////////////////// | 25 | //////////////////////////////////////////////////////////////////// |
| @@ -14,10 +14,10 @@ | |||
| 14 | 14 | ||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | -#include "catlass/gemm/dispatch_policy.hpp" | 17 | +#include "catlass/gemm/kda_gemm_dispatch_policy.hpp" |
| 18 | 18 | ||
| 19 | 19 | ||
| 20 | -#include "catlass/gemm/tile/tile_copy.hpp" | 20 | +#include "catlass/gemm/tile/kda_gemm_tile_copy.hpp" |
| 21 | 21 | ||
| 22 | 22 | ||
| 23 | //////////////////////////////////////////////////////////////////// | 23 | //////////////////////////////////////////////////////////////////// |
| @@ -14,12 +14,12 @@ | |||
| 14 | 14 | ||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | -#include "catlass/gemm/dispatch_policy.hpp" | 17 | +#include "catlass/gemm/kda_gemm_dispatch_policy.hpp" |
| 18 | 18 | ||
| 19 | 19 | ||
| 20 | -#include "catlass/gemm/tile/tile_copy.hpp" | 20 | +#include "catlass/gemm/tile/kda_gemm_tile_copy.hpp" |
| 21 | 21 | ||
| 22 | -#include "tla/layout.hpp" | 22 | +#include "tla/kda_tla_layout.hpp" |
| 23 | 23 | ||
| 24 | 24 | ||
| 25 | //////////////////////////////////////////////////////////////////// | 25 | //////////////////////////////////////////////////////////////////// |
| @@ -13,13 +13,13 @@ | |||
| 13 | 13 | ||
| 14 | 14 | ||
| 15 | 15 | ||
| 16 | -#include "catlass/gemm/dispatch_policy.hpp" | 16 | +#include "catlass/gemm/kda_gemm_dispatch_policy.hpp" |
| 17 | 17 | ||
| 18 | 18 | ||
| 19 | -#include "catlass/gemm/tile/tile_copy.hpp" | 19 | +#include "catlass/gemm/tile/kda_gemm_tile_copy.hpp" |
| 20 | 20 | ||
| 21 | 21 | ||
| 22 | -#include "tla/layout.hpp" | 22 | +#include "tla/kda_tla_layout.hpp" |
| 23 | 23 | ||
| 24 | //////////////////////////////////////////////////////////////////// | 24 | //////////////////////////////////////////////////////////////////// |
| 25 | 25 | ||
| @@ -14,10 +14,10 @@ | |||
| 14 | 14 | ||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | -#include "catlass/gemm/dispatch_policy.hpp" | 17 | +#include "catlass/gemm/kda_gemm_dispatch_policy.hpp" |
| 18 | 18 | ||
| 19 | 19 | ||
| 20 | -#include "catlass/gemm/tile/tile_copy.hpp" | 20 | +#include "catlass/gemm/tile/kda_gemm_tile_copy.hpp" |
| 21 | 21 | ||
| 22 | 22 | ||
| 23 | //////////////////////////////////////////////////////////////////// | 23 | //////////////////////////////////////////////////////////////////// |
| @@ -14,12 +14,12 @@ | |||
| 14 | 14 | ||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | -#include "catlass/gemm/dispatch_policy.hpp" | 17 | +#include "catlass/gemm/kda_gemm_dispatch_policy.hpp" |
| 18 | 18 | ||
| 19 | 19 | ||
| 20 | -#include "catlass/gemm/tile/tile_copy.hpp" | 20 | +#include "catlass/gemm/tile/kda_gemm_tile_copy.hpp" |
| 21 | 21 | ||
| 22 | -#include "tla/layout.hpp" | 22 | +#include "tla/kda_tla_layout.hpp" |
| 23 | 23 | ||
| 24 | 24 | ||
| 25 | //////////////////////////////////////////////////////////////////// | 25 | //////////////////////////////////////////////////////////////////// |
| @@ -14,10 +14,10 @@ | |||
| 14 | 14 | ||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | -#include "catlass/gemm/dispatch_policy.hpp" | 17 | +#include "catlass/gemm/kda_gemm_dispatch_policy.hpp" |
| 18 | 18 | ||
| 19 | 19 | ||
| 20 | -#include "catlass/gemm/tile/tile_copy.hpp" | 20 | +#include "catlass/gemm/tile/kda_gemm_tile_copy.hpp" |
| 21 | 21 | ||
| 22 | 22 | ||
| 23 | //////////////////////////////////////////////////////////////////// | 23 | //////////////////////////////////////////////////////////////////// |
| @@ -14,12 +14,12 @@ | |||
| 14 | 14 | ||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | -#include "catlass/gemm/dispatch_policy.hpp" | 17 | +#include "catlass/gemm/kda_gemm_dispatch_policy.hpp" |
| 18 | 18 | ||
| 19 | 19 | ||
| 20 | -#include "catlass/gemm/tile/tile_copy.hpp" | 20 | +#include "catlass/gemm/tile/kda_gemm_tile_copy.hpp" |
| 21 | 21 | ||
| 22 | -#include "tla/layout.hpp" | 22 | +#include "tla/kda_tla_layout.hpp" |
| 23 | 23 | ||
| 24 | 24 | ||
| 25 | //////////////////////////////////////////////////////////////////// | 25 | //////////////////////////////////////////////////////////////////// |
| @@ -13,13 +13,13 @@ | |||
| 13 | 13 | ||
| 14 | 14 | ||
| 15 | 15 | ||
| 16 | -#include "catlass/gemm/dispatch_policy.hpp" | 16 | +#include "catlass/gemm/kda_gemm_dispatch_policy.hpp" |
| 17 | 17 | ||
| 18 | 18 | ||
| 19 | -#include "catlass/gemm/tile/tile_copy.hpp" | 19 | +#include "catlass/gemm/tile/kda_gemm_tile_copy.hpp" |
| 20 | 20 | ||
| 21 | 21 | ||
| 22 | -#include "tla/layout.hpp" | 22 | +#include "tla/kda_tla_layout.hpp" |
| 23 | 23 | ||
| 24 | //////////////////////////////////////////////////////////////////// | 24 | //////////////////////////////////////////////////////////////////// |
| 25 | 25 | ||
| @@ -15,9 +15,9 @@ | |||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | 17 | ||
| 18 | -#include "catlass/gemm/dispatch_policy.hpp" | 18 | +#include "catlass/gemm/kda_gemm_dispatch_policy.hpp" |
| 19 | 19 | ||
| 20 | -#include "tla/layout.hpp" | 20 | +#include "tla/kda_tla_layout.hpp" |
| 21 | 21 | ||
| 22 | 22 | ||
| 23 | namespace Catlass::Gemm::Block { | 23 | namespace Catlass::Gemm::Block { |
| @@ -15,9 +15,9 @@ | |||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | 17 | ||
| 18 | -#include "catlass/gemm/dispatch_policy.hpp" | 18 | +#include "catlass/gemm/kda_gemm_dispatch_policy.hpp" |
| 19 | 19 | ||
| 20 | -#include "catlass/gemm/tile/tile_copy.hpp" | 20 | +#include "catlass/gemm/tile/kda_gemm_tile_copy.hpp" |
| 21 | 21 | ||
| 22 | 22 | ||
| 23 | namespace Catlass::Gemm::Block { | 23 | namespace Catlass::Gemm::Block { |
| @@ -14,9 +14,9 @@ | |||
| 14 | 14 | ||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | -#include "catlass/gemm/dispatch_policy.hpp" | 17 | +#include "catlass/gemm/kda_gemm_dispatch_policy.hpp" |
| 18 | 18 | ||
| 19 | -#include "tla/layout.hpp" | 19 | +#include "tla/kda_tla_layout.hpp" |
| 20 | 20 | ||
| 21 | 21 | ||
| 22 | namespace Catlass::Gemm::Block { | 22 | namespace Catlass::Gemm::Block { |
Mattention/chunk_kda_fwd/op_kernel/catlass/gemm/block/block_mmad_pingpong_per_group_per_block_tla.hpp+3-3
| @@ -15,11 +15,11 @@ | |||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | 17 | ||
| 18 | -#include "catlass/gemm/dispatch_policy.hpp" | 18 | +#include "catlass/gemm/kda_gemm_dispatch_policy.hpp" |
| 19 | 19 | ||
| 20 | -#include "catlass/gemm/tile/tile_copy.hpp" | 20 | +#include "catlass/gemm/tile/kda_gemm_tile_copy.hpp" |
| 21 | 21 | ||
| 22 | -#include "tla/layout.hpp" | 22 | +#include "tla/kda_tla_layout.hpp" |
| 23 | 23 | ||
| 24 | 24 | ||
| 25 | namespace Catlass::Gemm::Block { | 25 | namespace Catlass::Gemm::Block { |
Mattention/chunk_kda_fwd/op_kernel/catlass/gemm/block/block_mmad_pingpong_slice_k_with_prologue.hpp+2-2
| @@ -15,9 +15,9 @@ | |||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | 17 | ||
| 18 | -#include "catlass/gemm/dispatch_policy.hpp" | 18 | +#include "catlass/gemm/kda_gemm_dispatch_policy.hpp" |
| 19 | 19 | ||
| 20 | -#include "catlass/gemm/tile/tile_copy.hpp" | 20 | +#include "catlass/gemm/tile/kda_gemm_tile_copy.hpp" |
| 21 | 21 | ||
| 22 | 22 | ||
| 23 | namespace Catlass::Gemm::Block { | 23 | namespace Catlass::Gemm::Block { |
| @@ -15,11 +15,11 @@ | |||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | 17 | ||
| 18 | -#include "catlass/gemm/dispatch_policy.hpp" | 18 | +#include "catlass/gemm/kda_gemm_dispatch_policy.hpp" |
| 19 | 19 | ||
| 20 | -#include "catlass/gemm/tile/tile_copy.hpp" | 20 | +#include "catlass/gemm/tile/kda_gemm_tile_copy.hpp" |
| 21 | 21 | ||
| 22 | -#include "tla/layout.hpp" | 22 | +#include "tla/kda_tla_layout.hpp" |
| 23 | 23 | ||
| 24 | 24 | ||
| 25 | namespace Catlass::Gemm::Block { | 25 | namespace Catlass::Gemm::Block { |
| @@ -15,11 +15,11 @@ | |||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | 17 | ||
| 18 | -#include "catlass/gemm/dispatch_policy.hpp" | 18 | +#include "catlass/gemm/kda_gemm_dispatch_policy.hpp" |
| 19 | 19 | ||
| 20 | -#include "catlass/gemm/tile/tile_copy.hpp" | 20 | +#include "catlass/gemm/tile/kda_gemm_tile_copy.hpp" |
| 21 | 21 | ||
| 22 | -#include "tla/layout.hpp" | 22 | +#include "tla/kda_tla_layout.hpp" |
| 23 | 23 | ||
| 24 | 24 | ||
| 25 | namespace Catlass::Gemm::Block { | 25 | namespace Catlass::Gemm::Block { |
| @@ -15,9 +15,9 @@ | |||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | 17 | ||
| 18 | -#include "catlass/gemm/dispatch_policy.hpp" | 18 | +#include "catlass/gemm/kda_gemm_dispatch_policy.hpp" |
| 19 | 19 | ||
| 20 | -#include "catlass/gemm/tile/tile_copy.hpp" | 20 | +#include "catlass/gemm/tile/kda_gemm_tile_copy.hpp" |
| 21 | 21 | ||
| 22 | 22 | ||
| 23 | namespace Catlass::Gemm::Block { | 23 | namespace Catlass::Gemm::Block { |
Mattention/chunk_kda_fwd/op_kernel/catlass/gemm/block/block_mmad_preload_async_with_callback_tla.hpp+2-2
| @@ -14,10 +14,10 @@ | |||
| 14 | 14 | ||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | -#include "catlass/gemm/dispatch_policy.hpp" | 17 | +#include "catlass/gemm/kda_gemm_dispatch_policy.hpp" |
| 18 | 18 | ||
| 19 | 19 | ||
| 20 | -#include "tla/layout.hpp" | 20 | +#include "tla/kda_tla_layout.hpp" |
| 21 | 21 | ||
| 22 | 22 | ||
| 23 | namespace Catlass::Gemm::Block { | 23 | namespace Catlass::Gemm::Block { |
| @@ -16,11 +16,11 @@ | |||
| 16 | 16 | ||
| 17 | 17 | ||
| 18 | 18 | ||
| 19 | -#include "catlass/gemm/dispatch_policy.hpp" | 19 | +#include "catlass/gemm/kda_gemm_dispatch_policy.hpp" |
| 20 | 20 | ||
| 21 | -#include "catlass/gemm/tile/tile_copy.hpp" | 21 | +#include "catlass/gemm/tile/kda_gemm_tile_copy.hpp" |
| 22 | 22 | ||
| 23 | -#include "tla/layout.hpp" | 23 | +#include "tla/kda_tla_layout.hpp" |
| 24 | 24 | ||
| 25 | 25 | ||
| 26 | namespace Catlass::Gemm::Block { | 26 | namespace Catlass::Gemm::Block { |
| @@ -15,9 +15,9 @@ | |||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | 17 | ||
| 18 | -#include "catlass/gemm/dispatch_policy.hpp" | 18 | +#include "catlass/gemm/kda_gemm_dispatch_policy.hpp" |
| 19 | 19 | ||
| 20 | -#include "catlass/gemm/tile/tile_copy.hpp" | 20 | +#include "catlass/gemm/tile/kda_gemm_tile_copy.hpp" |
| 21 | 21 | ||
| 22 | 22 | ||
| 23 | namespace Catlass::Gemm::Block { | 23 | namespace Catlass::Gemm::Block { |
| @@ -15,9 +15,9 @@ | |||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | 17 | ||
| 18 | -#include "catlass/gemm/dispatch_policy.hpp" | 18 | +#include "catlass/gemm/kda_gemm_dispatch_policy.hpp" |
| 19 | 19 | ||
| 20 | -#include "catlass/gemm/tile/tile_copy.hpp" | 20 | +#include "catlass/gemm/tile/kda_gemm_tile_copy.hpp" |
| 21 | 21 | ||
| 22 | 22 | ||
| 23 | namespace Catlass::Gemm::Block { | 23 | namespace Catlass::Gemm::Block { |
| @@ -14,11 +14,11 @@ | |||
| 14 | 14 | ||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | -#include "catlass/gemm/dispatch_policy.hpp" | 17 | +#include "catlass/gemm/kda_gemm_dispatch_policy.hpp" |
| 18 | 18 | ||
| 19 | -#include "catlass/gemm/tile/tile_copy.hpp" | 19 | +#include "catlass/gemm/tile/kda_gemm_tile_copy.hpp" |
| 20 | 20 | ||
| 21 | -#include "tla/layout.hpp" | 21 | +#include "tla/kda_tla_layout.hpp" |
| 22 | 22 | ||
| 23 | 23 | ||
| 24 | namespace Catlass::Gemm::Block { | 24 | namespace Catlass::Gemm::Block { |
| @@ -15,9 +15,9 @@ | |||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | 17 | ||
| 18 | -#include "catlass/gemm/dispatch_policy.hpp" | 18 | +#include "catlass/gemm/kda_gemm_dispatch_policy.hpp" |
| 19 | 19 | ||
| 20 | -#include "catlass/gemm/tile/tile_copy.hpp" | 20 | +#include "catlass/gemm/tile/kda_gemm_tile_copy.hpp" |
| 21 | 21 | ||
| 22 | 22 | ||
| 23 | namespace Catlass::Gemm::Block { | 23 | namespace Catlass::Gemm::Block { |
| @@ -14,9 +14,9 @@ | |||
| 14 | 14 | ||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | -#include "catlass/gemm/dispatch_policy.hpp" | 17 | +#include "catlass/gemm/kda_gemm_dispatch_policy.hpp" |
| 18 | 18 | ||
| 19 | -#include "catlass/gemm/tile/tile_copy.hpp" | 19 | +#include "catlass/gemm/tile/kda_gemm_tile_copy.hpp" |
| 20 | 20 | ||
| 21 | 21 | ||
| 22 | 22 | ||
| @@ -12,7 +12,7 @@ | |||
| 12 | 12 | ||
| 13 | 13 | ||
| 14 | 14 | ||
| 15 | -#include "catlass/gemm/tile/tile_copy.hpp" | 15 | +#include "catlass/gemm/tile/kda_gemm_tile_copy.hpp" |
| 16 | 16 | ||
| 17 | 17 | ||
| 18 | namespace Catlass::Gemm::Block { | 18 | namespace Catlass::Gemm::Block { |
| @@ -15,10 +15,10 @@ | |||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | 17 | ||
| 18 | -#include "catlass/gemm/dispatch_policy.hpp" | 18 | +#include "catlass/gemm/kda_gemm_dispatch_policy.hpp" |
| 19 | 19 | ||
| 20 | 20 | ||
| 21 | -#include "catlass/gemm/tile/tile_copy.hpp" | 21 | +#include "catlass/gemm/tile/kda_gemm_tile_copy.hpp" |
| 22 | 22 | ||
| 23 | 23 | ||
| 24 | //////////////////////////////////////////////////////////////////// | 24 | //////////////////////////////////////////////////////////////////// |
| @@ -14,10 +14,10 @@ | |||
| 14 | 14 | ||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | -#include "catlass/gemm/dispatch_policy.hpp" | 17 | +#include "catlass/gemm/kda_gemm_dispatch_policy.hpp" |
| 18 | 18 | ||
| 19 | 19 | ||
| 20 | -#include "catlass/gemm/tile/tile_copy.hpp" | 20 | +#include "catlass/gemm/tile/kda_gemm_tile_copy.hpp" |
| 21 | 21 | ||
| 22 | 22 | ||
| 23 | //////////////////////////////////////////////////////////////////// | 23 | //////////////////////////////////////////////////////////////////// |
| @@ -13,10 +13,10 @@ | |||
| 13 | 13 | ||
| 14 | 14 | ||
| 15 | 15 | ||
| 16 | -#include "catlass/gemm/tile/tile_copy.hpp" | 16 | +#include "catlass/gemm/tile/kda_gemm_tile_copy.hpp" |
| 17 | 17 | ||
| 18 | 18 | ||
| 19 | -#include "catlass/gemm/dispatch_policy.hpp" | 19 | +#include "catlass/gemm/kda_gemm_dispatch_policy.hpp" |
| 20 | 20 | ||
| 21 | 21 | ||
| 22 | namespace Catlass::Gemm::Block { | 22 | namespace Catlass::Gemm::Block { |
| @@ -15,10 +15,10 @@ | |||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | 17 | ||
| 18 | -#include "catlass/gemm/dispatch_policy.hpp" | 18 | +#include "catlass/gemm/kda_gemm_dispatch_policy.hpp" |
| 19 | 19 | ||
| 20 | 20 | ||
| 21 | -#include "catlass/gemm/tile/tile_copy.hpp" | 21 | +#include "catlass/gemm/tile/kda_gemm_tile_copy.hpp" |
| 22 | 22 | ||
| 23 | 23 | ||
| 24 | //////////////////////////////////////////////////////////////////// | 24 | //////////////////////////////////////////////////////////////////// |
| @@ -15,10 +15,10 @@ | |||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | 17 | ||
| 18 | -#include "catlass/gemm/dispatch_policy.hpp" | 18 | +#include "catlass/gemm/kda_gemm_dispatch_policy.hpp" |
| 19 | 19 | ||
| 20 | 20 | ||
| 21 | -#include "catlass/gemm/tile/tile_copy.hpp" | 21 | +#include "catlass/gemm/tile/kda_gemm_tile_copy.hpp" |
| 22 | 22 | ||
| 23 | 23 | ||
| 24 | //////////////////////////////////////////////////////////////////// | 24 | //////////////////////////////////////////////////////////////////// |
| @@ -14,10 +14,10 @@ | |||
| 14 | 14 | ||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | -#include "catlass/gemm/dispatch_policy.hpp" | 17 | +#include "catlass/gemm/kda_gemm_dispatch_policy.hpp" |
| 18 | 18 | ||
| 19 | 19 | ||
| 20 | -#include "catlass/gemm/tile/tile_copy.hpp" | 20 | +#include "catlass/gemm/tile/kda_gemm_tile_copy.hpp" |
| 21 | 21 | ||
| 22 | 22 | ||
| 23 | //////////////////////////////////////////////////////////////////// | 23 | //////////////////////////////////////////////////////////////////// |
| @@ -14,10 +14,10 @@ | |||
| 14 | 14 | ||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | -#include "catlass/gemm/dispatch_policy.hpp" | 17 | +#include "catlass/gemm/kda_gemm_dispatch_policy.hpp" |
| 18 | 18 | ||
| 19 | 19 | ||
| 20 | -#include "catlass/gemm/tile/tile_copy.hpp" | 20 | +#include "catlass/gemm/tile/kda_gemm_tile_copy.hpp" |
| 21 | 21 | ||
| 22 | 22 | ||
| 23 | //////////////////////////////////////////////////////////////////// | 23 | //////////////////////////////////////////////////////////////////// |
| @@ -15,9 +15,9 @@ | |||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | 17 | ||
| 18 | -#include "catlass/gemm/dispatch_policy.hpp" | 18 | +#include "catlass/gemm/kda_gemm_dispatch_policy.hpp" |
| 19 | 19 | ||
| 20 | -#include "catlass/gemm/tile/tile_copy.hpp" | 20 | +#include "catlass/gemm/tile/kda_gemm_tile_copy.hpp" |
| 21 | 21 | ||
| 22 | 22 | ||
| 23 | namespace Catlass::Gemm::Block { | 23 | namespace Catlass::Gemm::Block { |
| @@ -15,9 +15,9 @@ | |||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | 17 | ||
| 18 | -#include "catlass/gemm/dispatch_policy.hpp" | 18 | +#include "catlass/gemm/kda_gemm_dispatch_policy.hpp" |
| 19 | 19 | ||
| 20 | -#include "catlass/gemm/tile/tile_copy.hpp" | 20 | +#include "catlass/gemm/tile/kda_gemm_tile_copy.hpp" |
| 21 | 21 | ||
| 22 | 22 | ||
| 23 | namespace Catlass::Gemm::Block { | 23 | namespace Catlass::Gemm::Block { |
| @@ -15,9 +15,9 @@ | |||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | 17 | ||
| 18 | -#include "catlass/gemm/dispatch_policy.hpp" | 18 | +#include "catlass/gemm/kda_gemm_dispatch_policy.hpp" |
| 19 | 19 | ||
| 20 | -#include "catlass/gemm/tile/tile_copy.hpp" | 20 | +#include "catlass/gemm/tile/kda_gemm_tile_copy.hpp" |
| 21 | 21 | ||
| 22 | 22 | ||
| 23 | namespace Catlass::Gemm::Block { | 23 | namespace Catlass::Gemm::Block { |
| @@ -16,9 +16,9 @@ | |||
| 16 | 16 | ||
| 17 | 17 | ||
| 18 | 18 | ||
| 19 | -#include "catlass/gemm/dispatch_policy.hpp" | 19 | +#include "catlass/gemm/kda_gemm_dispatch_policy.hpp" |
| 20 | 20 | ||
| 21 | -#include "catlass/gemm/tile/tile_copy.hpp" | 21 | +#include "catlass/gemm/tile/kda_gemm_tile_copy.hpp" |
| 22 | 22 | ||
| 23 | 23 | ||
| 24 | namespace Catlass::Gemm::Block { | 24 | namespace Catlass::Gemm::Block { |
Mattention/chunk_kda_fwd/op_kernel/catlass/gemm/block/kda_block_mmad_preload_async_with_callback.hpp+2-2
| @@ -16,9 +16,9 @@ | |||
| 16 | 16 | ||
| 17 | 17 | ||
| 18 | 18 | ||
| 19 | -#include "catlass/gemm/dispatch_policy.hpp" | 19 | +#include "catlass/gemm/kda_gemm_dispatch_policy.hpp" |
| 20 | 20 | ||
| 21 | -#include "catlass/gemm/tile/tile_copy.hpp" | 21 | +#include "catlass/gemm/tile/kda_gemm_tile_copy.hpp" |
| 22 | 22 | ||
| 23 | 23 | ||
| 24 | namespace Catlass::Gemm::Block { | 24 | namespace Catlass::Gemm::Block { |
Rattention/chunk_kda_fwd/op_kernel/catlass/gemm/dispatch_policy.hpp→attention/chunk_kda_fwd/op_kernel/catlass/gemm/kda_gemm_dispatch_policy.hpp+3-3
| @@ -8,8 +8,8 @@ | |||
| 8 | * See LICENSE in the root of the software repository for the full text of the License. | 8 | * See LICENSE in the root of the software repository for the full text of the License. |
| 9 | */ | 9 | */ |
| 10 | 10 | ||
| 11 | -#ifndef CATLASS_GEMM_DISPATCH_POLICY_HPP | 11 | +#ifndef CATLASS_KDA_GEMM_DISPATCH_POLICY_HPP |
| 12 | -#define CATLASS_GEMM_DISPATCH_POLICY_HPP | 12 | +#define CATLASS_KDA_GEMM_DISPATCH_POLICY_HPP |
| 13 | 13 | ||
| 14 | 14 | ||
| 15 | 15 | ||
| @@ -370,4 +370,4 @@ struct MmadPingpongPertile : public MmadBase<ArchTag_, false> { | |||
| 370 | 370 | ||
| 371 | } // namespace Catlass::Gemm | 371 | } // namespace Catlass::Gemm |
| 372 | 372 | ||
| 373 | -#endif // CATLASS_GEMM_DISPATCH_POLICY_HPP | 373 | +#endif // CATLASS_KDA_GEMM_DISPATCH_POLICY_HPP |
| @@ -16,7 +16,7 @@ | |||
| 16 | 16 | ||
| 17 | 17 | ||
| 18 | 18 | ||
| 19 | -#include "tla/layout.hpp" | 19 | +#include "tla/kda_tla_layout.hpp" |
| 20 | 20 | ||
| 21 | namespace Catlass::Gemm::helper { | 21 | namespace Catlass::Gemm::helper { |
| 22 | 22 | ||
| @@ -15,7 +15,7 @@ | |||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | 17 | ||
| 18 | -#include "catlass/gemm/dispatch_policy.hpp" | 18 | +#include "catlass/gemm/kda_gemm_dispatch_policy.hpp" |
| 19 | 19 | ||
| 20 | 20 | ||
| 21 | namespace Catlass::Gemm::Tile { | 21 | namespace Catlass::Gemm::Tile { |
Rattention/chunk_kda_fwd/op_kernel/catlass/gemm/tile/atlasa2/copy_gm_to_ub.hpp→attention/chunk_kda_fwd/op_kernel/catlass/gemm/tile/atlasa2/kda_atlasa2_copy_gm_to_ub.hpp+3-3
| @@ -8,8 +8,8 @@ | |||
| 8 | * See LICENSE in the root of the software repository for the full text of the License. | 8 | * See LICENSE in the root of the software repository for the full text of the License. |
| 9 | */ | 9 | */ |
| 10 | 10 | ||
| 11 | -#ifndef CATLASS_GEMM_TILE_ATLASA2_COPY_GM_TO_UB_HPP | 11 | +#ifndef CATLASS_KDA_ATLASA2_COPY_GM_TO_UB_HPP |
| 12 | -#define CATLASS_GEMM_TILE_ATLASA2_COPY_GM_TO_UB_HPP | 12 | +#define CATLASS_KDA_ATLASA2_COPY_GM_TO_UB_HPP |
| 13 | 13 | ||
| 14 | 14 | ||
| 15 | 15 | ||
| @@ -96,4 +96,4 @@ struct CopyGm2Ub<Arch::AtlasA2, Gemm::GemmType<Element, layout::VectorLayout>> { | |||
| 96 | 96 | ||
| 97 | } // namespace Catlass::Gemm::Tile | 97 | } // namespace Catlass::Gemm::Tile |
| 98 | 98 | ||
| 99 | -#endif // CATLASS_GEMM_TILE_ATLASA2_COPY_GM_TO_UB_HPP | 99 | +#endif // CATLASS_KDA_ATLASA2_COPY_GM_TO_UB_HPP |
Rattention/chunk_kda_fwd/op_kernel/catlass/gemm/tile/atlasa2/copy_ub_to_gm.hpp→attention/chunk_kda_fwd/op_kernel/catlass/gemm/tile/atlasa2/kda_atlasa2_copy_ub_to_gm.hpp+3-3
| @@ -8,8 +8,8 @@ | |||
| 8 | * See LICENSE in the root of the software repository for the full text of the License. | 8 | * See LICENSE in the root of the software repository for the full text of the License. |
| 9 | */ | 9 | */ |
| 10 | 10 | ||
| 11 | -#ifndef CATLASS_GEMM_TILE_ATLASA2_COPY_UB_TO_GM_HPP | 11 | +#ifndef CATLASS_KDA_ATLASA2_COPY_UB_TO_GM_HPP |
| 12 | -#define CATLASS_GEMM_TILE_ATLASA2_COPY_UB_TO_GM_HPP | 12 | +#define CATLASS_KDA_ATLASA2_COPY_UB_TO_GM_HPP |
| 13 | 13 | ||
| 14 | 14 | ||
| 15 | 15 | ||
| @@ -128,4 +128,4 @@ struct CopyUb2Gm<Arch::AtlasA2, Gemm::GemmType<Element, layout::RowMajor>> { | |||
| 128 | 128 | ||
| 129 | } // namespace Catlass::Gemm::Tile | 129 | } // namespace Catlass::Gemm::Tile |
| 130 | 130 | ||
| 131 | -#endif // CATLASS_GEMM_TILE_ATLASA2_COPY_UB_TO_GM_HPP | 131 | +#endif // CATLASS_KDA_ATLASA2_COPY_UB_TO_GM_HPP |
| @@ -15,7 +15,7 @@ | |||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | 17 | ||
| 18 | -#include "catlass/gemm/dispatch_policy.hpp" | 18 | +#include "catlass/gemm/kda_gemm_dispatch_policy.hpp" |
| 19 | 19 | ||
| 20 | 20 | ||
| 21 | namespace Catlass::Gemm::Tile { | 21 | namespace Catlass::Gemm::Tile { |
| @@ -15,7 +15,7 @@ | |||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | 17 | ||
| 18 | -#include "catlass/gemm/dispatch_policy.hpp" | 18 | +#include "catlass/gemm/kda_gemm_dispatch_policy.hpp" |
| 19 | 19 | ||
| 20 | 20 | ||
| 21 | namespace Catlass::Gemm::Tile { | 21 | namespace Catlass::Gemm::Tile { |
Rattention/chunk_kda_fwd/op_kernel/catlass/gemm/tile/copy_gm_to_ub.hpp→attention/chunk_kda_fwd/op_kernel/catlass/gemm/tile/kda_copy_gm_to_ub.hpp+3-3
| @@ -8,11 +8,11 @@ | |||
| 8 | * See LICENSE in the root of the software repository for the full text of the License. | 8 | * See LICENSE in the root of the software repository for the full text of the License. |
| 9 | */ | 9 | */ |
| 10 | 10 | ||
| 11 | -#ifndef CATLASS_GEMM_TILE_COPY_GM_TO_UB_HPP | 11 | +#ifndef CATLASS_KDA_GEMM_COPY_GM_TO_UB_HPP |
| 12 | -#define CATLASS_GEMM_TILE_COPY_GM_TO_UB_HPP | 12 | +#define CATLASS_KDA_GEMM_COPY_GM_TO_UB_HPP |
| 13 | 13 | ||
| 14 | 14 | ||
| 15 | -#include "catlass/gemm/tile/atlasa2/copy_gm_to_ub.hpp" | 15 | +#include "catlass/gemm/tile/atlasa2/kda_atlasa2_copy_gm_to_ub.hpp" |
| 16 | 16 | ||
| 17 | 17 | ||
| 18 | 18 | ||
Rattention/chunk_kda_fwd/op_kernel/catlass/gemm/tile/copy_ub_to_gm.hpp→attention/chunk_kda_fwd/op_kernel/catlass/gemm/tile/kda_copy_ub_to_gm.hpp+3-3
| @@ -8,11 +8,11 @@ | |||
| 8 | * See LICENSE in the root of the software repository for the full text of the License. | 8 | * See LICENSE in the root of the software repository for the full text of the License. |
| 9 | */ | 9 | */ |
| 10 | 10 | ||
| 11 | -#ifndef CATLASS_GEMM_TILE_COPY_UB_TO_GM_HPP | 11 | +#ifndef CATLASS_KDA_GEMM_COPY_UB_TO_GM_HPP |
| 12 | -#define CATLASS_GEMM_TILE_COPY_UB_TO_GM_HPP | 12 | +#define CATLASS_KDA_GEMM_COPY_UB_TO_GM_HPP |
| 13 | 13 | ||
| 14 | 14 | ||
| 15 | -#include "catlass/gemm/tile/atlasa2/copy_ub_to_gm.hpp" | 15 | +#include "catlass/gemm/tile/atlasa2/kda_atlasa2_copy_ub_to_gm.hpp" |
| 16 | 16 | ||
| 17 | 17 | ||
| 18 | 18 | ||
Rattention/chunk_kda_fwd/op_kernel/catlass/gemm/tile/tile_copy.hpp→attention/chunk_kda_fwd/op_kernel/catlass/gemm/tile/kda_gemm_tile_copy.hpp+5-5
| @@ -8,8 +8,8 @@ | |||
| 8 | * See LICENSE in the root of the software repository for the full text of the License. | 8 | * See LICENSE in the root of the software repository for the full text of the License. |
| 9 | */ | 9 | */ |
| 10 | 10 | ||
| 11 | -#ifndef CATLASS_GEMM_TILE_TILE_COPY_HPP | 11 | +#ifndef CATLASS_KDA_GEMM_TILE_COPY_HPP |
| 12 | -#define CATLASS_GEMM_TILE_TILE_COPY_HPP | 12 | +#define CATLASS_KDA_GEMM_TILE_COPY_HPP |
| 13 | 13 | ||
| 14 | 14 | ||
| 15 | 15 | ||
| @@ -20,14 +20,14 @@ | |||
| 20 | 20 | ||
| 21 | 21 | ||
| 22 | 22 | ||
| 23 | -#include "catlass/gemm/tile/copy_gm_to_ub.hpp" | 23 | +#include "catlass/gemm/tile/kda_copy_gm_to_ub.hpp" |
| 24 | 24 | ||
| 25 | 25 | ||
| 26 | 26 | ||
| 27 | 27 | ||
| 28 | 28 | ||
| 29 | 29 | ||
| 30 | -#include "catlass/gemm/tile/copy_ub_to_gm.hpp" | 30 | +#include "catlass/gemm/tile/kda_copy_ub_to_gm.hpp" |
| 31 | 31 | ||
| 32 | 32 | ||
| 33 | 33 | ||
| @@ -517,4 +517,4 @@ struct PackedTileCopyTlaToUB | |||
| 517 | 517 | ||
| 518 | } // namespace Catlass::Gemm::Tile | 518 | } // namespace Catlass::Gemm::Tile |
| 519 | 519 | ||
| 520 | -#endif // CATLASS_GEMM_TILE_TILE_COPY_HPP | 520 | +#endif // CATLASS_KDA_GEMM_TILE_COPY_HPP |
| @@ -19,14 +19,14 @@ | |||
| 19 | 19 | ||
| 20 | 20 | ||
| 21 | 21 | ||
| 22 | -#include "catlass/gemm/dispatch_policy.hpp" | 22 | +#include "catlass/gemm/kda_gemm_dispatch_policy.hpp" |
| 23 | -#include "catlass/gemm/tile/tile_copy.hpp" | 23 | +#include "catlass/gemm/tile/kda_gemm_tile_copy.hpp" |
| 24 | 24 | ||
| 25 | 25 | ||
| 26 | 26 | ||
| 27 | 27 | ||
| 28 | 28 | ||
| 29 | -#include "tla/layout.hpp" | 29 | +#include "tla/kda_tla_layout.hpp" |
| 30 | 30 | ||
| 31 | 31 | ||
| 32 | using namespace AscendC; | 32 | using namespace AscendC; |
| @@ -19,15 +19,15 @@ | |||
| 19 | 19 | ||
| 20 | 20 | ||
| 21 | 21 | ||
| 22 | -#include "catlass/gemm/dispatch_policy.hpp" | 22 | +#include "catlass/gemm/kda_gemm_dispatch_policy.hpp" |
| 23 | -#include "catlass/gemm/tile/tile_copy.hpp" | 23 | +#include "catlass/gemm/tile/kda_gemm_tile_copy.hpp" |
| 24 | 24 | ||
| 25 | 25 | ||
| 26 | 26 | ||
| 27 | 27 | ||
| 28 | 28 | ||
| 29 | 29 | ||
| 30 | -#include "tla/layout.hpp" | 30 | +#include "tla/kda_tla_layout.hpp" |
| 31 | 31 | ||
| 32 | 32 | ||
| 33 | using namespace AscendC; | 33 | using namespace AscendC; |
| @@ -19,14 +19,14 @@ | |||
| 19 | 19 | ||
| 20 | 20 | ||
| 21 | 21 | ||
| 22 | -#include "catlass/gemm/dispatch_policy.hpp" | 22 | +#include "catlass/gemm/kda_gemm_dispatch_policy.hpp" |
| 23 | -#include "catlass/gemm/tile/tile_copy.hpp" | 23 | +#include "catlass/gemm/tile/kda_gemm_tile_copy.hpp" |
| 24 | 24 | ||
| 25 | 25 | ||
| 26 | 26 | ||
| 27 | 27 | ||
| 28 | 28 | ||
| 29 | -#include "tla/layout.hpp" | 29 | +#include "tla/kda_tla_layout.hpp" |
| 30 | 30 | ||
| 31 | 31 | ||
| 32 | using namespace AscendC; | 32 | using namespace AscendC; |
| @@ -15,12 +15,12 @@ | |||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | 17 | ||
| 18 | -#include "catlass/gemm/dispatch_policy.hpp" | 18 | +#include "catlass/gemm/kda_gemm_dispatch_policy.hpp" |
| 19 | 19 | ||
| 20 | -#include "catlass/gemm/tile/tile_copy.hpp" | 20 | +#include "catlass/gemm/tile/kda_gemm_tile_copy.hpp" |
| 21 | 21 | ||
| 22 | 22 | ||
| 23 | -#include "tla/layout.hpp" | 23 | +#include "tla/kda_tla_layout.hpp" |
| 24 | 24 | ||
| 25 | 25 | ||
| 26 | namespace Common { | 26 | namespace Common { |
| @@ -15,11 +15,11 @@ | |||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | 17 | ||
| 18 | -#include "catlass/gemm/dispatch_policy.hpp" | 18 | +#include "catlass/gemm/kda_gemm_dispatch_policy.hpp" |
| 19 | 19 | ||
| 20 | -#include "catlass/gemm/tile/tile_copy.hpp" | 20 | +#include "catlass/gemm/tile/kda_gemm_tile_copy.hpp" |
| 21 | 21 | ||
| 22 | -#include "tla/layout.hpp" | 22 | +#include "tla/kda_tla_layout.hpp" |
| 23 | 23 | ||
| 24 | 24 | ||
| 25 | namespace Catlass::Gemm { | 25 | namespace Catlass::Gemm { |
| @@ -15,11 +15,11 @@ | |||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | 17 | ||
| 18 | -#include "catlass/gemm/dispatch_policy.hpp" | 18 | +#include "catlass/gemm/kda_gemm_dispatch_policy.hpp" |
| 19 | 19 | ||
| 20 | -#include "catlass/gemm/tile/tile_copy.hpp" | 20 | +#include "catlass/gemm/tile/kda_gemm_tile_copy.hpp" |
| 21 | 21 | ||
| 22 | -#include "tla/layout.hpp" | 22 | +#include "tla/kda_tla_layout.hpp" |
| 23 | 23 | ||
| 24 | 24 | ||
| 25 | namespace Catlass::Gemm { | 25 | namespace Catlass::Gemm { |
| @@ -12,7 +12,7 @@ | |||
| 12 | 12 | ||
| 13 | 13 | ||
| 14 | 14 | ||
| 15 | -#include "tla/layout.hpp" // tla::Shape | 15 | +#include "tla/kda_tla_layout.hpp" // tla::Shape |
| 16 | 16 | ||
| 17 | 17 | ||
| 18 | 18 | ||
Rattention/chunk_kda_fwd/op_kernel/tla/layout.hpp→attention/chunk_kda_fwd/op_kernel/tla/kda_tla_layout.hpp+3-3
| @@ -8,8 +8,8 @@ | |||
| 8 | * See LICENSE in the root of the software repository for the full text of the License. | 8 | * See LICENSE in the root of the software repository for the full text of the License. |
| 9 | */ | 9 | */ |
| 10 | 10 | ||
| 11 | -#ifndef TLA_LAYOUT_HPP | 11 | +#ifndef KDA_TLA_LAYOUT_HPP |
| 12 | -#define TLA_LAYOUT_HPP | 12 | +#define KDA_TLA_LAYOUT_HPP |
| 13 | 13 | ||
| 14 | 14 | ||
| 15 | 15 | ||
| @@ -577,4 +577,4 @@ using MakeBatchedLayout_t = PrependDimLayout_t<Layout>; | |||
| 577 | 577 | ||
| 578 | } // end namespace tla | 578 | } // end namespace tla |
| 579 | 579 | ||
| 580 | -#endif // TLA_LAYOUT_HPP | 580 | +#endif // KDA_TLA_LAYOUT_HPP |
Rattention/sparse_flash_mla/op_kernel/arch22/sparse_flash_mla_metadata.h→attention/sparse_flash_mla/op_kernel/arch22/sparse_flash_mla_arch22_metadata.h+4-4
| @@ -9,12 +9,12 @@ | |||
| 9 | */ | 9 | */ |
| 10 | 10 | ||
| 11 | /*! | 11 | /*! |
| 12 | - * \file sparse_flash_mla_metadata.h | 12 | + * \file sparse_flash_mla_arch22_metadata.h |
| 13 | * \brief | 13 | * \brief |
| 14 | */ | 14 | */ |
| 15 | 15 | ||
| 16 | -#ifndef SPARSE_FLASH_MLA_METADATA_H | 16 | +#ifndef SPARSE_FLASH_MLA_ARCH22_METADATA_H |
| 17 | -#define SPARSE_FLASH_MLA_METADATA_H | 17 | +#define SPARSE_FLASH_MLA_ARCH22_METADATA_H |
| 18 | 18 | ||
| 19 | 19 | ||
| 20 | 20 | ||
| @@ -77,4 +77,4 @@ struct SasMetadata { | |||
| 77 | static_assert(SMLA_META_SIZE * sizeof(SMLA_METADATA_T) >= sizeof(detail::SasMetadata)); | 77 | static_assert(SMLA_META_SIZE * sizeof(SMLA_METADATA_T) >= sizeof(detail::SasMetadata)); |
| 78 | } // namespace optiling | 78 | } // namespace optiling |
| 79 | 79 | ||
| 80 | -#endif | 80 | +#endif // SPARSE_FLASH_MLA_ARCH22_METADATA_H |
| @@ -24,7 +24,7 @@ | |||
| 24 | 24 | ||
| 25 | 25 | ||
| 26 | 26 | ||
| 27 | -#include "sparse_flash_mla_metadata.h" | 27 | +#include "sparse_flash_mla_arch22_metadata.h" |
| 28 | 28 | ||
| 29 | namespace SMLAKernel { | 29 | namespace SMLAKernel { |
| 30 | using namespace matmul; | 30 | using namespace matmul; |
| @@ -24,7 +24,7 @@ | |||
| 24 | 24 | ||
| 25 | 25 | ||
| 26 | 26 | ||
| 27 | -#include "sparse_flash_mla_metadata.h" | 27 | +#include "sparse_flash_mla_arch22_metadata.h" |
| 28 | 28 | ||
| 29 | namespace SMLAKernel { | 29 | namespace SMLAKernel { |
| 30 | using namespace matmul; | 30 | using namespace matmul; |
| @@ -20,7 +20,7 @@ | |||
| 20 | 20 | ||
| 21 | 21 | ||
| 22 | 22 | ||
| 23 | -#include "../sparse_flash_mla_metadata.h" | 23 | +#include "../sparse_flash_mla_kernel_metadata.h" |
| 24 | 24 | ||
| 25 | 25 | ||
| 26 | 26 | ||
| @@ -20,7 +20,7 @@ | |||
| 20 | 20 | ||
| 21 | 21 | ||
| 22 | 22 | ||
| 23 | -#include "../sparse_flash_mla_metadata.h" | 23 | +#include "../sparse_flash_mla_kernel_metadata.h" |
| 24 | 24 | ||
| 25 | 25 | ||
| 26 | 26 | ||
| @@ -19,14 +19,14 @@ | |||
| 19 | 19 | ||
| 20 | 20 | ||
| 21 | 21 | ||
| 22 | -#include "sparse_flash_mla_metadata.h" | 22 | +#include "sparse_flash_mla_kernel_metadata.h" |
| 23 | 23 | ||
| 24 | 24 | ||
| 25 | 25 | ||
| 26 | 26 | ||
| 27 | 27 | ||
| 28 | 28 | ||
| 29 | -#include "arch22/sparse_flash_mla_metadata.h" | 29 | +#include "arch22/sparse_flash_mla_arch22_metadata.h" |
| 30 | 30 | ||
| 31 | 31 | ||
| 32 | using namespace AscendC; | 32 | using namespace AscendC; |
Rattention/sparse_flash_mla/op_kernel/sparse_flash_mla_metadata.h→attention/sparse_flash_mla/op_kernel/sparse_flash_mla_kernel_metadata.h+4-4
| @@ -9,12 +9,12 @@ | |||
| 9 | */ | 9 | */ |
| 10 | 10 | ||
| 11 | /*! | 11 | /*! |
| 12 | - * \file sparse_flash_mla_metadata.h | 12 | + * \file sparse_flash_mla_kernel_metadata.h |
| 13 | * \brief | 13 | * \brief |
| 14 | */ | 14 | */ |
| 15 | 15 | ||
| 16 | -#ifndef SPARSE_FLASH_MLA_METADATA_H | 16 | +#ifndef SPARSE_FLASH_MLA_KERNEL_METADATA_H |
| 17 | -#define SPARSE_FLASH_MLA_METADATA_H | 17 | +#define SPARSE_FLASH_MLA_KERNEL_METADATA_H |
| 18 | 18 | ||
| 19 | 19 | ||
| 20 | 20 | ||
| @@ -77,4 +77,4 @@ namespace detail { | |||
| 77 | static_assert(SMLA_METADATA_TOTAL_SIZE * sizeof(SMLA_METADATA_T) >= sizeof(detail::SmlaMetadata)); | 77 | static_assert(SMLA_METADATA_TOTAL_SIZE * sizeof(SMLA_METADATA_T) >= sizeof(detail::SmlaMetadata)); |
| 78 | }; | 78 | }; |
| 79 | 79 | ||
| 80 | -#endif // SPARSE_FLASH_MLA_METADATA_H | 80 | +#endif // SPARSE_FLASH_MLA_KERNEL_METADATA_H |
| @@ -20,7 +20,7 @@ | |||
| 20 | 20 | ||
| 21 | 21 | ||
| 22 | 22 | ||
| 23 | -#include "../../../op_kernel/arch22/sparse_flash_mla_metadata.h" | 23 | +#include "../../../op_kernel/arch22/sparse_flash_mla_arch22_metadata.h" |
| 24 | 24 | ||
| 25 | namespace smla_ut { | 25 | namespace smla_ut { |
| 26 | 26 | ||
| @@ -17,7 +17,7 @@ | |||
| 17 | 17 | ||
| 18 | 18 | ||
| 19 | 19 | ||
| 20 | -#include "../../sparse_flash_mla/op_kernel/sparse_flash_mla_metadata.h" | 20 | +#include "../../sparse_flash_mla/op_kernel/sparse_flash_mla_kernel_metadata.h" |
| 21 | 21 | ||
| 22 | 22 | ||
| 23 | 23 | ||
| @@ -22,7 +22,7 @@ | |||
| 22 | 22 | ||
| 23 | 23 | ||
| 24 | 24 | ||
| 25 | -#include "../../sparse_flash_mla/op_kernel/sparse_flash_mla_metadata.h" | 25 | +#include "../../sparse_flash_mla/op_kernel/sparse_flash_mla_kernel_metadata.h" |
| 26 | 26 | ||
| 27 | 27 | ||
| 28 | namespace aicpu { | 28 | namespace aicpu { |
Rgmm/grouped_matmul/op_kernel/a16w4_msd/tool.h→gmm/grouped_matmul/op_kernel/a16w4_msd/gmm_a16w4_tool.h+3-3
| @@ -9,11 +9,11 @@ | |||
| 9 | */ | 9 | */ |
| 10 | 10 | ||
| 11 | /* ! | 11 | /* ! |
| 12 | - * \file tool.h | 12 | + * \file gmm_a16w4_tool.h |
| 13 | * \brief | 13 | * \brief |
| 14 | */ | 14 | */ |
| 15 | -#ifndef GROUPED_MATMUL_WEIGHT_QUANT_TOOL_H | 15 | +#ifndef GMM_A16W4_TOOL_H |
| 16 | -#define GROUPED_MATMUL_WEIGHT_QUANT_TOOL_H | 16 | +#define GMM_A16W4_TOOL_H |
| 17 | 17 | ||
| 18 | 18 | ||
| 19 | 19 | ||
Rgmm/grouped_matmul/op_kernel/a16w4_msd/static_diag_constant.h→gmm/grouped_matmul/op_kernel/a16w4_msd/gmm_static_diag_constant.h+3-3
| @@ -9,11 +9,11 @@ | |||
| 9 | */ | 9 | */ |
| 10 | 10 | ||
| 11 | /*! | 11 | /*! |
| 12 | - * \file static_diag_constant.h | 12 | + * \file gmm_static_diag_constant.h |
| 13 | * \brief | 13 | * \brief |
| 14 | */ | 14 | */ |
| 15 | -#ifndef GROUPED_MATMUL_WEIGHT_QUANT_STATIC_DIAG_CONSTANT_H | 15 | +#ifndef GMM_STATIC_DIAG_CONSTANT_H |
| 16 | -#define GROUPED_MATMUL_WEIGHT_QUANT_STATIC_DIAG_CONSTANT_H | 16 | +#define GMM_STATIC_DIAG_CONSTANT_H |
| 17 | 17 | ||
| 18 | 18 | ||
| 19 | 19 | ||
| @@ -18,8 +18,8 @@ | |||
| 18 | 18 | ||
| 19 | 19 | ||
| 20 | 20 | ||
| 21 | -#include "static_diag_constant.h" | 21 | +#include "gmm_static_diag_constant.h" |
| 22 | -#include "tool.h" | 22 | +#include "gmm_a16w4_tool.h" |
| 23 | 23 | ||
| 24 | namespace GROUPED_MATMUL::A16W4Msd { | 24 | namespace GROUPED_MATMUL::A16W4Msd { |
| 25 | 25 | ||
| @@ -18,7 +18,7 @@ | |||
| 18 | 18 | ||
| 19 | 19 | ||
| 20 | 20 | ||
| 21 | -#include "tool.h" | 21 | +#include "gmm_a16w4_tool.h" |
| 22 | 22 | ||
| 23 | using AscendC::RoundMode; | 23 | using AscendC::RoundMode; |
| 24 | 24 | ||
Rgmm/grouped_matmul/op_kernel/gmm_infra/epilogue/tile/tile_broadcast_add.hpp→gmm/grouped_matmul/op_kernel/gmm_infra/epilogue/tile/gmm_tile_broadcast_add.hpp+2-2
| @@ -9,8 +9,8 @@ | |||
| 9 | */ | 9 | */ |
| 10 | 10 | ||
| 11 | 11 | ||
| 12 | -#ifndef GMM_EPILOGUE_TILE_TILE_BROADCAST_ADD_HPP | 12 | +#ifndef GMM_TILE_BROADCAST_ADD_HPP |
| 13 | -#define GMM_EPILOGUE_TILE_TILE_BROADCAST_ADD_HPP | 13 | +#define GMM_TILE_BROADCAST_ADD_HPP |
| 14 | 14 | ||
| 15 | 15 | ||
| 16 | 16 | ||
Rgmm/grouped_matmul/op_kernel/gmm_infra/gemm/tile/copy_gm_to_ub.hpp→gmm/grouped_matmul/op_kernel/gmm_infra/gemm/tile/gmm_copy_gm_to_ub.hpp+3-3
| @@ -8,8 +8,8 @@ | |||
| 8 | * See LICENSE in the root of the software repository for the full text of the License. | 8 | * See LICENSE in the root of the software repository for the full text of the License. |
| 9 | */ | 9 | */ |
| 10 | 10 | ||
| 11 | -#ifndef GMM_GEMM_TILE_COPY_GM_TO_UB_HPP | 11 | +#ifndef GMM_COPY_GM_TO_UB_HPP |
| 12 | -#define GMM_GEMM_TILE_COPY_GM_TO_UB_HPP | 12 | +#define GMM_COPY_GM_TO_UB_HPP |
| 13 | 13 | ||
| 14 | 14 | ||
| 15 | 15 | ||
| @@ -56,4 +56,4 @@ struct CopyGm2Ub<Arch::AtlasA2, Gemm::GemmType<Element, layout::VectorLayout>> { | |||
| 56 | 56 | ||
| 57 | } // Catlass::Gemm::Tile | 57 | } // Catlass::Gemm::Tile |
| 58 | 58 | ||
| 59 | -#endif // GMM_GEMM_TILE_COPY_GM_TO_UB_HPP | 59 | +#endif // GMM_COPY_GM_TO_UB_HPP |
Rgmm/grouped_matmul/op_kernel/gmm_infra/gemm/tile/copy_ub_to_gm.hpp→gmm/grouped_matmul/op_kernel/gmm_infra/gemm/tile/gmm_copy_ub_to_gm.hpp+3-3
| @@ -8,8 +8,8 @@ | |||
| 8 | * See LICENSE in the root of the software repository for the full text of the License. | 8 | * See LICENSE in the root of the software repository for the full text of the License. |
| 9 | */ | 9 | */ |
| 10 | 10 | ||
| 11 | -#ifndef GMM_GEMM_TILE_COPY_UB_TO_GM_HPP | 11 | +#ifndef GMM_COPY_UB_TO_GM_HPP |
| 12 | -#define GMM_GEMM_TILE_COPY_UB_TO_GM_HPP | 12 | +#define GMM_COPY_UB_TO_GM_HPP |
| 13 | 13 | ||
| 14 | 14 | ||
| 15 | 15 | ||
| @@ -57,4 +57,4 @@ struct CopyUb2Gm<Arch::AtlasA2, Gemm::GemmType<Element, layout::RowMajor>> { | |||
| 57 | 57 | ||
| 58 | } // Catlass::Gemm::Tile | 58 | } // Catlass::Gemm::Tile |
| 59 | 59 | ||
| 60 | -#endif // GMM_GEMM_TILE_COPY_UB_TO_GM_HPP | 60 | +#endif // GMM_COPY_UB_TO_GM_HPP |
| @@ -1,422 +0,0 @@ | |||
| 1 | -/** | ||
| 2 | - * Copyright (c) 2025 Huawei Technologies Co., Ltd. | ||
| 3 | - * This program is free software, you can redistribute it and/or modify it under the terms and conditions of | ||
| 4 | - * CANN Open Software License Agreement Version 2.0 (the "License"). | ||
| 5 | - * Please refer to the License for details. You may not use this file except in compliance with the License. | ||
| 6 | - * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED, | ||
| 7 | - * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE. | ||
| 8 | - * See LICENSE in the root of the software repository for the full text of the License. | ||
| 9 | - */ | ||
| 10 | - | ||
| 11 | -/*! | ||
| 12 | - * \file batch_mat_mul_nd2nz.h | ||
| 13 | - * \brief | ||
| 14 | - */ | ||
| 15 | - | ||
| 16 | - | ||
| 17 | - | ||
| 18 | - | ||
| 19 | - | ||
| 20 | -using namespace AscendC; | ||
| 21 | -using namespace matmul; | ||
| 22 | - | ||
| 23 | -using namespace std; | ||
| 24 | - | ||
| 25 | - | ||
| 26 | -template <class T> | ||
| 27 | -class Mc2KernelND2NZBMM { | ||
| 28 | -public: | ||
| 29 | - __aicore__ inline Mc2KernelND2NZBMM(){}; | ||
| 30 | - __aicore__ inline void CopyIn(uint64_t progress, LocalTensor<T> &dstLocal); | ||
| 31 | - __aicore__ inline bool SetBufBMM(); | ||
| 32 | - __aicore__ inline void Init(GM_ADDR dst, GM_ADDR src, uint32_t height, uint32_t width, uint32_t batch, | ||
| 33 | - TBuf<TPosition::VECCALC> &ubBuffer, uint32_t usedCoreNum); | ||
| 34 | - | ||
| 35 | - __aicore__ inline bool ProcessBMM(); | ||
| 36 | - | ||
| 37 | -private: | ||
| 38 | - __aicore__ inline void CopyOutDirect(uint64_t gmOutOffset, uint32_t startPad, uint16_t total, uint64_t progress); | ||
| 39 | - __aicore__ inline void CopyOutPageInit(uint64_t &gmOutOffset, uint32_t startPad, uint32_t &bufOffset); | ||
| 40 | - __aicore__ inline void CopyOutMakePage(uint32_t nLoop, uint32_t &bufOffset); | ||
| 41 | - __aicore__ inline void CopyOutPageMainImp(uint64_t &gmOutOffset, uint32_t nLoop, uint32_t &bufOffset); | ||
| 42 | - __aicore__ inline void CopyOutPageMain(uint64_t &gmOutOffset, uint32_t mPage, uint32_t startPad, uint32_t total, | ||
| 43 | - uint32_t &bufOffset, uint64_t progress); | ||
| 44 | - __aicore__ inline void CopyOutPageEnd(uint64_t gmOutOffset, uint32_t res, uint32_t &bufOffset); | ||
| 45 | - __aicore__ inline void CopyOutPage(uint64_t gmOutOffset, uint32_t mPage, uint32_t total, uint32_t startPad, | ||
| 46 | - uint64_t progress); | ||
| 47 | - __aicore__ inline void CopyOutBatchReform(uint64_t gmOutOffset, uint32_t mPage, uint32_t total, uint32_t startPad, | ||
| 48 | - uint64_t progress); | ||
| 49 | - __aicore__ inline void ComputeBMM(uint64_t progress); | ||
| 50 | - | ||
| 51 | -private: | ||
| 52 | - TBuf<TPosition::VECCALC> *ubPtr_; | ||
| 53 | - GlobalTensor<T> srcGM; | ||
| 54 | - GlobalTensor<T> dstGM; | ||
| 55 | - LocalTensor<T> inBuf_; | ||
| 56 | - LocalTensor<T> inBuf2_; | ||
| 57 | - LocalTensor<T> midBuf_; | ||
| 58 | - LocalTensor<T> outBuf_; | ||
| 59 | - LocalTensor<T> outBuf2_; | ||
| 60 | - LocalTensor<T> zeroBuf_; | ||
| 61 | - uint32_t padSize_; | ||
| 62 | - uint32_t height_; | ||
| 63 | - uint32_t hAligned_; | ||
| 64 | - uint32_t width_; | ||
| 65 | - uint32_t batch_; | ||
| 66 | - uint32_t wTail_; | ||
| 67 | - uint32_t hBuffer_; | ||
| 68 | - uint32_t nFullProgress_; | ||
| 69 | - uint32_t heightTotalTail_; | ||
| 70 | - uint16_t hPad_; | ||
| 71 | - uint32_t blockDim_; | ||
| 72 | - uint32_t blockIdx_; | ||
| 73 | - uint32_t hBlockNum_; | ||
| 74 | - uint32_t copyInSize_; | ||
| 75 | - uint64_t c0_; | ||
| 76 | - uint32_t copyInRepeat_; | ||
| 77 | - uint16_t widthBlockTotal_; | ||
| 78 | - bool noPadD_; | ||
| 79 | -}; | ||
| 80 | - | ||
| 81 | -template <class T> | ||
| 82 | -__aicore__ inline void Mc2KernelND2NZBMM<T>::CopyIn(uint64_t progress, LocalTensor<T> &dstLocal) | ||
| 83 | -{ | ||
| 84 | - uint64_t curCopyInSize = progress == nFullProgress_ ? heightTotalTail_ * width_ : copyInSize_; | ||
| 85 | - uint64_t gmInOffset = copyInSize_ * progress; | ||
| 86 | - DataCopyExtParams copyParams{DEFAULT_DATA_COPY_NBURST, static_cast<uint32_t>(curCopyInSize * sizeof(T)), | ||
| 87 | - DEFAULT_DATA_COPY_STRIDE, DEFAULT_DATA_COPY_STRIDE, 0}; | ||
| 88 | - DataCopyPadExtParams<T> padParams{false, 0, 0, 0}; | ||
| 89 | - DataCopyPad(dstLocal, srcGM[gmInOffset], copyParams, padParams); | ||
| 90 | -} | ||
| 91 | - | ||
| 92 | -template <class T> | ||
| 93 | -__aicore__ inline bool Mc2KernelND2NZBMM<T>::SetBufBMM() | ||
| 94 | -{ | ||
| 95 | - uint32_t hTotal = height_ * batch_; | ||
| 96 | - uint32_t wAligned = Align2(width_, c0_); | ||
| 97 | - | ||
| 98 | - uint32_t hMax = TOTAL_UB_SIZE / sizeof(T) / (width_ + width_ + width_ + wAligned); | ||
| 99 | - // hBlockNumEle表示最少要几行连续数据才能32B对齐 | ||
| 100 | - uint32_t hBlockNumEle = M_BLOCK_NUM_ELE_LIST[wTail_] * 2 / sizeof(T); | ||
| 101 | - hBlockNumEle = width_ == 1 ? 1 : hBlockNumEle; | ||
| 102 | - hBlockNumEle = hBlockNumEle == 0 ? 1 : hBlockNumEle; | ||
| 103 | - // gcd是c0_和width_的最大公约数 | ||
| 104 | - uint32_t gcd = GCD_LIST[wTail_]; | ||
| 105 | - if constexpr (sizeof(T) == sizeof(float)) { | ||
| 106 | - gcd = wTail_ == 0 ? 8 : gcd; | ||
| 107 | - } | ||
| 108 | - // hEle表示最小载入行数，为满足vnchwconv的要求，要乘个16 | ||
| 109 | - uint32_t hEle = hBlockNumEle * ALIGNED_H; | ||
| 110 | - // eleNum是在ub_buffer和外轴限制的基础上，最多可载入几倍的hEle | ||
| 111 | - uint32_t eleNum = (hTotal + hEle - 1) / hEle; | ||
| 112 | - uint32_t eleNumTmp = hMax / hEle; | ||
| 113 | - eleNum = min(eleNumTmp, eleNum); | ||
| 114 | - eleNum = eleNum * hBlockNumEle > REPEAT_TIMES_MAX ? (REPEAT_TIMES_MAX / hBlockNumEle) : eleNum; | ||
| 115 | - | ||
| 116 | - if (eleNum == 0) { | ||
| 117 | - return false; | ||
| 118 | - } | ||
| 119 | - | ||
| 120 | - copyInRepeat_ = eleNum * width_ / gcd; | ||
| 121 | - | ||
| 122 | - hBuffer_ = eleNum * hEle; | ||
| 123 | - copyInSize_ = hBuffer_ * width_; | ||
| 124 | - // 16 * (eleNum * hBlockNumEle)*width_，计算地址偏移时使用 | ||
| 125 | - hBlockNum_ = eleNum * hBlockNumEle; | ||
| 126 | - nFullProgress_ = hTotal / hBuffer_; | ||
| 127 | - heightTotalTail_ = hTotal % hBuffer_; | ||
| 128 | - | ||
| 129 | - midBuf_ = ubPtr_->Get<T>()[0]; | ||
| 130 | - zeroBuf_ = ubPtr_->Get<T>()[copyInSize_]; | ||
| 131 | - inBuf_ = ubPtr_->Get<T>()[copyInSize_ * 2]; | ||
| 132 | - outBuf_ = ubPtr_->Get<T>()[copyInSize_ * 3]; | ||
| 133 | - // 清零可以去掉，mad使用实际的大小计算，就不需要清零 | ||
| 134 | - Duplicate(zeroBuf_, T(0), copyInSize_); | ||
| 135 | - | ||
| 136 | - PipeBarrier<PIPE_ALL>(); | ||
| 137 | - return true; | ||
| 138 | -} | ||
| 139 | - | ||
| 140 | -template <class T> | ||
| 141 | -__aicore__ inline void Mc2KernelND2NZBMM<T>::Init(GM_ADDR dst, GM_ADDR src, uint32_t height, uint32_t width, | ||
| 142 | - uint32_t batch, TBuf<TPosition::VECCALC> &ubBuffer, | ||
| 143 | - uint32_t usedCoreNum) | ||
| 144 | -{ | ||
| 145 | - height_ = height; | ||
| 146 | - width_ = width; | ||
| 147 | - batch_ = batch; | ||
| 148 | - uint32_t hTotal = height_ * batch_; | ||
| 149 | - | ||
| 150 | - blockDim_ = usedCoreNum; | ||
| 151 | - blockIdx_ = GetBlockIdx(); | ||
| 152 | - | ||
| 153 | - c0_ = BLOCK_SIZE_BYTE / sizeof(T); | ||
| 154 | - | ||
| 155 | - srcGM.SetGlobalBuffer((__gm__ T *)src); | ||
| 156 | - dstGM.SetGlobalBuffer((__gm__ T *)dst); | ||
| 157 | - ubPtr_ = &ubBuffer; | ||
| 158 | - | ||
| 159 | - noPadD_ = (width_ == c0_); | ||
| 160 | - | ||
| 161 | - uint32_t batchTail = height_ % ALIGNED_H; | ||
| 162 | - hPad_ = batchTail == 0 ? 0 : ALIGNED_H - batchTail; | ||
| 163 | - | ||
| 164 | - padSize_ = hPad_ * c0_; | ||
| 165 | - | ||
| 166 | - hAligned_ = Align2(height_, ALIGNED_H); | ||
| 167 | - | ||
| 168 | - uint32_t widthBlock = width_ / c0_; | ||
| 169 | - wTail_ = width_ & (c0_ - 1); | ||
| 170 | - | ||
| 171 | - widthBlockTotal_ = wTail_ ? widthBlock + 1 : widthBlock; | ||
| 172 | -} | ||
| 173 | - | ||
| 174 | -template <class T> | ||
| 175 | -__aicore__ inline bool Mc2KernelND2NZBMM<T>::ProcessBMM() | ||
| 176 | -{ | ||
| 177 | - if (SetBufBMM()) { | ||
| 178 | - uint32_t nLoop = heightTotalTail_ ? nFullProgress_ + 1 : nFullProgress_; | ||
| 179 | - for (uint32_t i = blockIdx_; i < nLoop; i += blockDim_) { | ||
| 180 | - ComputeBMM(i); | ||
| 181 | - | ||
| 182 | - SetFlag<HardEvent::MTE3_MTE2>(EVENT_ID0); | ||
| 183 | - WaitFlag<HardEvent::MTE3_MTE2>(EVENT_ID0); | ||
| 184 | - } | ||
| 185 | - PipeBarrier<PIPE_ALL>(); | ||
| 186 | - return true; | ||
| 187 | - } | ||
| 188 | - return false; | ||
| 189 | -} | ||
| 190 | - | ||
| 191 | -template <class T> | ||
| 192 | -__aicore__ inline void Mc2KernelND2NZBMM<T>::CopyOutDirect(uint64_t gmOutOffset, uint32_t startPad, uint16_t total, | ||
| 193 | - uint64_t progress) | ||
| 194 | -{ | ||
| 195 | - uint32_t start = startPad - hPad_; | ||
| 196 | - SetFlag<HardEvent::V_MTE3>(EVENT_ID0); | ||
| 197 | - WaitFlag<HardEvent::V_MTE3>(EVENT_ID0); | ||
| 198 | - // 处理上个核心的尾部没完成的batch | ||
| 199 | - if (start > total) { | ||
| 200 | - if (hAligned_ - total <= UINT16_MAX) { | ||
| 201 | - DataCopy(dstGM[gmOutOffset], outBuf_, | ||
| 202 | - {widthBlockTotal_, total, 0, static_cast<uint16_t>(hAligned_ - total)}); | ||
| 203 | - } else { | ||
| 204 | - for (uint16_t i = 0; i < widthBlockTotal_; i++) { | ||
| 205 | - DataCopy(dstGM[gmOutOffset + hAligned_ * c0_ * i], outBuf_[total * c0_ * i], {1, total, 0, 0}); | ||
| 206 | - } | ||
| 207 | - } | ||
| 208 | - return; | ||
| 209 | - } else if (start == total) { | ||
| 210 | - DataCopy(dstGM[gmOutOffset], outBuf_, | ||
| 211 | - {widthBlockTotal_, uint16_t(start), uint16_t(hBuffer_ - start), uint16_t(hAligned_ - start)}); | ||
| 212 | - DataCopy(dstGM[gmOutOffset + start * c0_], zeroBuf_, {widthBlockTotal_, hPad_, 0, uint16_t(height_)}); | ||
| 213 | - return; | ||
| 214 | - } | ||
| 215 | - | ||
| 216 | - if (startPad != hAligned_) { | ||
| 217 | - DataCopy(dstGM[gmOutOffset], outBuf_, | ||
| 218 | - {widthBlockTotal_, uint16_t(start), uint16_t(hBuffer_ - start), uint16_t(hAligned_ - start)}); | ||
| 219 | - DataCopy(dstGM[gmOutOffset + start * c0_], zeroBuf_, {widthBlockTotal_, hPad_, 0, uint16_t(height_)}); | ||
| 220 | - gmOutOffset += startPad * c0_ + (widthBlockTotal_ - 1) * hAligned_ * c0_; | ||
| 221 | - | ||
| 222 | - } else { | ||
| 223 | - start = 0; | ||
| 224 | - } | ||
| 225 | - // 处理完整的batch | ||
| 226 | - uint32_t nLoop = (total - start) / height_; | ||
| 227 | - uint16_t res = (total - start) % height_; | ||
| 228 | - | ||
| 229 | - if (height_ <= total - start) { | ||
| 230 | - for (int i = 0; i < nLoop; i++) { | ||
| 231 | - DataCopy(dstGM[gmOutOffset], outBuf_[start * c0_ + height_ * c0_ * i], | ||
| 232 | - {widthBlockTotal_, uint16_t(height_), uint16_t(hBuffer_ - height_), hPad_}); | ||
| 233 | - DataCopy(dstGM[gmOutOffset + height_ * c0_], zeroBuf_, {widthBlockTotal_, hPad_, 0, uint16_t(height_)}); | ||
| 234 | - gmOutOffset += hAligned_ * c0_ * widthBlockTotal_; | ||
| 235 | - } | ||
| 236 | - } | ||
| 237 | - // 处理尾部余下的batch | ||
| 238 | - if (res) { | ||
| 239 | - if (hAligned_ - total <= UINT16_MAX) { | ||
| 240 | - DataCopy(dstGM[gmOutOffset], outBuf_[start * c0_ + height_ * c0_ * nLoop], | ||
| 241 | - {widthBlockTotal_, res, uint16_t(hBuffer_ - res), uint16_t(hAligned_ - res)}); | ||
| 242 | - } else { | ||
| 243 | - for (uint16_t i = 0; i < widthBlockTotal_; i++) { | ||
| 244 | - DataCopy(dstGM[gmOutOffset + hAligned_ * c0_ * i], | ||
| 245 | - outBuf_[start * c0_ + height_ * c0_ * nLoop + total * c0_ * i], {1, res, 0, 0}); | ||
| 246 | - } | ||
| 247 | - } | ||
| 248 | - } | ||
| 249 | -} | ||
| 250 | - | ||
| 251 | -template <class T> | ||
| 252 | -__aicore__ inline void Mc2KernelND2NZBMM<T>::CopyOutPageInit(uint64_t &gmOutOffset, uint32_t startPad, | ||
| 253 | - uint32_t &bufOffset) | ||
| 254 | -{ | ||
| 255 | - uint32_t start = startPad - hPad_; | ||
| 256 | - uint32_t startSize = start * c0_; | ||
| 257 | - | ||
| 258 | - for (int k = 0; k < widthBlockTotal_; k++) { | ||
| 259 | - if (start > 0) { | ||
| 260 | - Copy(midBuf_[startPad * c0_ * k], outBuf_[hBuffer_ * c0_ * k], startSize); | ||
| 261 | - Duplicate(midBuf_[startPad * c0_ * k + startSize], T(0), padSize_); | ||
| 262 | - } else { | ||
| 263 | - Duplicate(midBuf_[startPad * c0_ * k], T(0), startPad * c0_); | ||
| 264 | - } | ||
| 265 | - } | ||
| 266 | - SetFlag<HardEvent::V_MTE3>(EVENT_ID0); | ||
| 267 | - WaitFlag<HardEvent::V_MTE3>(EVENT_ID0); | ||
| 268 | - | ||
| 269 | - DataCopy(dstGM[gmOutOffset], midBuf_, {widthBlockTotal_, uint16_t(startPad), 0, uint16_t(hAligned_ - startPad)}); | ||
| 270 | - SetFlag<HardEvent::MTE3_V>(EVENT_ID0); | ||
| 271 | - WaitFlag<HardEvent::MTE3_V>(EVENT_ID0); | ||
| 272 | - | ||
| 273 | - bufOffset = startSize; | ||
| 274 | - gmOutOffset += startPad * c0_ + (widthBlockTotal_ - 1) * hAligned_ * c0_; | ||
| 275 | -} | ||
| 276 | - | ||
| 277 | -template <class T> | ||
| 278 | -__aicore__ inline void Mc2KernelND2NZBMM<T>::CopyOutMakePage(uint32_t nLoop, uint32_t &bufOffset) | ||
| 279 | -{ | ||
| 280 | - for (int j = 0; j < nLoop; j++) { | ||
| 281 | - for (int k = 0; k < widthBlockTotal_; k++) { | ||
| 282 | - Copy(midBuf_[c0_ * hAligned_ * (k + widthBlockTotal_ * j)], | ||
| 283 | - outBuf_[bufOffset + hBuffer_ * c0_ * k + c0_ * height_ * j], height_ * c0_); | ||
| 284 | - Duplicate(midBuf_[height_ * c0_ + c0_ * hAligned_ * (k + widthBlockTotal_ * j)], T(0), padSize_); | ||
| 285 | - } | ||
| 286 | - } | ||
| 287 | - bufOffset += c0_ * height_ * nLoop; | ||
| 288 | -} | ||
| 289 | - | ||
| 290 | -template <class T> | ||
| 291 | -__aicore__ inline void Mc2KernelND2NZBMM<T>::CopyOutPageMainImp(uint64_t &gmOutOffset, uint32_t nLoop, | ||
| 292 | - uint32_t &bufOffset) | ||
| 293 | -{ | ||
| 294 | - CopyOutMakePage(nLoop, bufOffset); | ||
| 295 | - SetFlag<HardEvent::V_MTE3>(EVENT_ID0); | ||
| 296 | - WaitFlag<HardEvent::V_MTE3>(EVENT_ID0); | ||
| 297 | - | ||
| 298 | - DataCopy(dstGM[gmOutOffset], midBuf_, hAligned_ * widthBlockTotal_ * c0_ * nLoop); | ||
| 299 | - | ||
| 300 | - gmOutOffset += hAligned_ * widthBlockTotal_ * c0_ * nLoop; | ||
| 301 | - SetFlag<HardEvent::MTE3_V>(EVENT_ID0); | ||
| 302 | - WaitFlag<HardEvent::MTE3_V>(EVENT_ID0); | ||
| 303 | -} | ||
| 304 | - | ||
| 305 | -template <class T> | ||
| 306 | -__aicore__ inline void Mc2KernelND2NZBMM<T>::CopyOutPageMain(uint64_t &gmOutOffset, uint32_t mPage, uint32_t startPad, | ||
| 307 | - uint32_t total, uint32_t &bufOffset, uint64_t progress) | ||
| 308 | -{ | ||
| 309 | - uint32_t mPage2 = mPage / widthBlockTotal_; | ||
| 310 | - uint32_t nLoopIn = mPage2 / hAligned_; | ||
| 311 | - uint32_t mFinal = ((total - startPad + hPad_) / height_ + 1) * hPad_ + total; | ||
| 312 | - uint32_t nFull = (startPad == hAligned_) ? mFinal / hAligned_ : (mFinal - startPad) / hAligned_; | ||
| 313 | - uint32_t nLoopOut = nFull / nLoopIn; | ||
| 314 | - | ||
| 315 | - for (int i = 0; i < nLoopOut; i++) { | ||
| 316 | - CopyOutPageMainImp(gmOutOffset, nLoopIn, bufOffset); | ||
| 317 | - } | ||
| 318 | - uint32_t nLoopTail = nFull % nLoopIn; | ||
| 319 | - | ||
| 320 | - CopyOutPageMainImp(gmOutOffset, nLoopTail, bufOffset); | ||
| 321 | -} | ||
| 322 | - | ||
| 323 | -template <class T> | ||
| 324 | -__aicore__ inline void Mc2KernelND2NZBMM<T>::CopyOutPageEnd(uint64_t gmOutOffset, uint32_t res, uint32_t &bufOffset) | ||
| 325 | -{ | ||
| 326 | - for (int k = 0; k < widthBlockTotal_; k++) { | ||
| 327 | - Copy(midBuf_[c0_ * res * k], outBuf_[bufOffset + hBuffer_ * c0_ * k], res * c0_); | ||
| 328 | - } | ||
| 329 | - SetFlag<HardEvent::V_MTE3>(EVENT_ID0); | ||
| 330 | - WaitFlag<HardEvent::V_MTE3>(EVENT_ID0); | ||
| 331 | - | ||
| 332 | - DataCopy(dstGM[gmOutOffset], midBuf_, {uint16_t(widthBlockTotal_), uint16_t(res), 0, uint16_t(hAligned_ - res)}); | ||
| 333 | -} | ||
| 334 | - | ||
| 335 | -template <class T> | ||
| 336 | -__aicore__ inline void Mc2KernelND2NZBMM<T>::CopyOutPage(uint64_t gmOutOffset, uint32_t mPage, uint32_t total, | ||
| 337 | - uint32_t startPad, uint64_t progress) | ||
| 338 | -{ | ||
| 339 | - uint32_t bufOffset = 0; | ||
| 340 | - uint32_t start = startPad - hPad_; | ||
| 341 | - | ||
| 342 | - if (startPad != hAligned_) { | ||
| 343 | - CopyOutPageInit(gmOutOffset, startPad, bufOffset); | ||
| 344 | - } else { | ||
| 345 | - start = 0; | ||
| 346 | - } | ||
| 347 | - | ||
| 348 | - uint32_t res = (total - start) % height_; | ||
| 349 | - | ||
| 350 | - CopyOutPageMain(gmOutOffset, mPage, startPad, total, bufOffset, progress); | ||
| 351 | - | ||
| 352 | - if (res) { | ||
| 353 | - CopyOutPageEnd(gmOutOffset, res, bufOffset); | ||
| 354 | - } | ||
| 355 | -} | ||
| 356 | - | ||
| 357 | -template <class T> | ||
| 358 | -__aicore__ inline void Mc2KernelND2NZBMM<T>::CopyOutBatchReform(uint64_t gmOutOffset, uint32_t mPage, uint32_t total, | ||
| 359 | - uint32_t startPad, uint64_t progress) | ||
| 360 | -{ | ||
| 361 | - uint32_t mPage2 = mPage / widthBlockTotal_; | ||
| 362 | - | ||
| 363 | - if (hAligned_ > mPage2) { | ||
| 364 | - CopyOutDirect(gmOutOffset, startPad, total, progress); | ||
| 365 | - return; | ||
| 366 | - } | ||
| 367 | - CopyOutPage(gmOutOffset, mPage, total, startPad, progress); | ||
| 368 | -} | ||
| 369 | - | ||
| 370 | -template <class T> | ||
| 371 | -__aicore__ inline void Mc2KernelND2NZBMM<T>::ComputeBMM(uint64_t progress) | ||
| 372 | -{ | ||
| 373 | - if (noPadD_) { | ||
| 374 | - CopyIn(progress, outBuf_); | ||
| 375 | - } else { | ||
| 376 | - CopyIn(progress, inBuf_); | ||
| 377 | - SetFlag<HardEvent::MTE2_V>(EVENT_ID0); | ||
| 378 | - WaitFlag<HardEvent::MTE2_V>(EVENT_ID0); | ||
| 379 | - if (wTail_ == 0) { // 内轴32B对齐，大块搬入，再重排，当前实现可能有问题 | ||
| 380 | - PadDAligned<T>(progress, outBuf_, inBuf_, 0, 0, width_, c0_, hBlockNum_, false); | ||
| 381 | - } else { | ||
| 382 | - PadDMain<T>(progress, outBuf_, inBuf_, midBuf_, zeroBuf_, 0, 0, width_, c0_, hBlockNum_, copyInRepeat_, | ||
| 383 | - hBuffer_, wTail_, false); | ||
| 384 | - } | ||
| 385 | - } | ||
| 386 | - PipeBarrier<PIPE_ALL>(); | ||
| 387 | - | ||
| 388 | - uint64_t gmOutOffset = | ||
| 389 | - (hBuffer_ * progress) / height_ * hAligned_ * widthBlockTotal_ * c0_ + ((hBuffer_ * progress) % height_) * c0_; | ||
| 390 | - uint32_t total = (progress == nFullProgress_) ? heightTotalTail_ : hBuffer_; | ||
| 391 | - uint32_t startPad = hAligned_ - (progress * hBuffer_) % height_; | ||
| 392 | - uint32_t mPage = (hBuffer_ * width_) / c0_ / ALIGNED_H * ALIGNED_H; | ||
| 393 | - | ||
| 394 | - CopyOutBatchReform(gmOutOffset, mPage, total, startPad, progress); | ||
| 395 | - SetFlag<HardEvent::MTE3_MTE2>(EVENT_ID0); | ||
| 396 | - WaitFlag<HardEvent::MTE3_MTE2>(EVENT_ID0); | ||
| 397 | -} | ||
| 398 | - | ||
| 399 | - | ||
| 400 | -template <class T> | ||
| 401 | -__aicore__ inline bool Mc2Nd2nzVnchwBMM(GlobalTensor<T> &dst, GlobalTensor<T> &src, uint32_t height, uint32_t width, | ||
| 402 | - uint32_t batch, TBuf<TPosition::VECCALC> &ubBuffer, uint32_t usedCoreNum) | ||
| 403 | -{ | ||
| 404 | - Mc2KernelND2NZBMM<T> op; | ||
| 405 | - op.Init((GM_ADDR)dst[0].GetPhyAddr(), (GM_ADDR)src[0].GetPhyAddr(), height, width, batch, ubBuffer, usedCoreNum); | ||
| 406 | - return op.ProcessBMM(); | ||
| 407 | -} | ||
| 408 | - | ||
| 409 | -template <> | ||
| 410 | -__aicore__ inline bool Mc2Nd2nzVnchwBMM(GlobalTensor<bfloat16_t> &dst, GlobalTensor<bfloat16_t> &src, uint32_t height, | ||
| 411 | - uint32_t width, uint32_t batch, TBuf<TPosition::VECCALC> &ubBuffer, | ||
| 412 | - uint32_t usedCoreNum) | ||
| 413 | -{ | ||
| 414 | - GlobalTensor<half> dstGlobalTrans; | ||
| 415 | - GlobalTensor<half> srcGlobalTrans; | ||
| 416 | - dstGlobalTrans.SetGlobalBuffer((__gm__ half *)dst.GetPhyAddr(0)); | ||
| 417 | - srcGlobalTrans.SetGlobalBuffer((__gm__ half *)src.GetPhyAddr(0)); | ||
| 418 | - return Mc2Nd2nzVnchwBMM(dstGlobalTrans, srcGlobalTrans, height, width, batch, ubBuffer, usedCoreNum); | ||
| 419 | -} | ||
| 420 | - | ||
| 421 | - | ||
| 422 | - | ||
| @@ -17,7 +17,7 @@ | |||
| 17 | 17 | ||
| 18 | 18 | ||
| 19 | 19 | ||
| 20 | -#include "batch_mat_mul_nd2nz.h" | 20 | +#include "../../batch_mat_mul_v3/op_kernel/batch_mat_mul_nd2nz.h" |
| 21 | 21 | ||
| 22 | using namespace AscendC; | 22 | using namespace AscendC; |
| 23 | using namespace matmul; | 23 | using namespace matmul; |
| @@ -10,7 +10,7 @@ | |||
| 10 | 10 | ||
| 11 | 11 | ||
| 12 | 12 | ||
| 13 | -#include "../utils/device_utils.h" | 13 | +#include "../utils/wq_device_utils.h" |
| 14 | namespace Mc2WeightQuantBatchMatmulV2::Arch35::Catlass { | 14 | namespace Mc2WeightQuantBatchMatmulV2::Arch35::Catlass { |
| 15 | template <size_t I, class Tuple> | 15 | template <size_t I, class Tuple> |
| 16 | struct deduce_optional_input { | 16 | struct deduce_optional_input { |
| @@ -10,7 +10,7 @@ | |||
| 10 | 10 | ||
| 11 | 11 | ||
| 12 | 12 | ||
| 13 | -#include "../dispatch_policy.h" | 13 | +#include "../wq_dispatch_policy.h" |
| 14 | 14 | ||
| 15 | 15 | ||
| 16 | 16 | ||
| @@ -16,7 +16,7 @@ | |||
| 16 | 16 | ||
| 17 | 17 | ||
| 18 | 18 | ||
| 19 | -#include "../utils/device_utils.h" | 19 | +#include "../utils/wq_device_utils.h" |
| 20 | 20 | ||
| 21 | 21 | ||
| 22 | 22 | ||
| @@ -20,7 +20,7 @@ | |||
| 20 | 20 | ||
| 21 | 21 | ||
| 22 | 22 | ||
| 23 | -#include "../utils/device_utils.h" | 23 | +#include "../utils/wq_device_utils.h" |
| 24 | 24 | ||
| 25 | 25 | ||
| 26 | 26 | ||
| @@ -10,7 +10,7 @@ | |||
| 10 | 10 | ||
| 11 | 11 | ||
| 12 | 12 | ||
| 13 | -#include "../utils/device_utils.h" | 13 | +#include "../utils/wq_device_utils.h" |
| 14 | 14 | ||
| 15 | namespace Mc2WeightQuantBatchMatmulV2::Arch35::Catlass { | 15 | namespace Mc2WeightQuantBatchMatmulV2::Arch35::Catlass { |
| 16 | /* | 16 | /* |
Mmc2/3rd/weight_quant_batch_matmul_v2/op_kernel/arch35/catlass/kernel/david_wqbmm_load_in_advance.h+1-1
| @@ -10,7 +10,7 @@ | |||
| 10 | 10 | ||
| 11 | 11 | ||
| 12 | 12 | ||
| 13 | -#include "../utils/device_utils.h" | 13 | +#include "../utils/wq_device_utils.h" |
| 14 | 14 | ||
| 15 | namespace Mc2WeightQuantBatchMatmulV2::Arch35::Catlass { | 15 | namespace Mc2WeightQuantBatchMatmulV2::Arch35::Catlass { |
| 16 | template <typename ProblemShape, typename BlockMainloop, typename TileScheduler> | 16 | template <typename ProblemShape, typename BlockMainloop, typename TileScheduler> |
Mmc2/3rd/weight_quant_batch_matmul_v2/op_kernel/arch35/catlass/pipeline/pipeline_stage_mixcore.h+2-2
| @@ -10,11 +10,11 @@ | |||
| 10 | 10 | ||
| 11 | 11 | ||
| 12 | 12 | ||
| 13 | -#include "../utils/device_utils.h" | 13 | +#include "../utils/wq_device_utils.h" |
| 14 | 14 | ||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | -#include "utils.h" | 17 | +#include "wq_pipeline_utils.h" |
| 18 | 18 | ||
| 19 | using AscendC::CrossCoreSetFlag; | 19 | using AscendC::CrossCoreSetFlag; |
| 20 | using AscendC::CrossCoreWaitFlag; | 20 | using AscendC::CrossCoreWaitFlag; |
Mmc2/3rd/weight_quant_batch_matmul_v2/op_kernel/arch35/catlass/pipeline/pipeline_stage_singlecore.h+2-2
| @@ -10,11 +10,11 @@ | |||
| 10 | 10 | ||
| 11 | 11 | ||
| 12 | 12 | ||
| 13 | -#include "../utils/device_utils.h" | 13 | +#include "../utils/wq_device_utils.h" |
| 14 | 14 | ||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | -#include "utils.h" | 17 | +#include "wq_pipeline_utils.h" |
| 18 | 18 | ||
| 19 | 19 | ||
| 20 | 20 | ||
| @@ -10,10 +10,10 @@ | |||
| 10 | 10 | ||
| 11 | 11 | ||
| 12 | 12 | ||
| 13 | -#include "../utils/device_utils.h" | 13 | +#include "../utils/wq_device_utils.h" |
| 14 | 14 | ||
| 15 | 15 | ||
| 16 | -#include "utils.h" | 16 | +#include "wq_pipeline_utils.h" |
| 17 | 17 | ||
| 18 | using AscendC::HardEvent; | 18 | using AscendC::HardEvent; |
| 19 | using AscendC::Hardware; | 19 | using AscendC::Hardware; |
| @@ -13,7 +13,7 @@ | |||
| 13 | 13 | ||
| 14 | 14 | ||
| 15 | 15 | ||
| 16 | -#include "utils.h" | 16 | +#include "wq_pipeline_utils.h" |
| 17 | 17 | ||
| 18 | using AscendC::HardEvent; | 18 | using AscendC::HardEvent; |
| 19 | using AscendC::Hardware; | 19 | using AscendC::Hardware; |
| @@ -12,7 +12,7 @@ | |||
| 12 | 12 | ||
| 13 | 13 | ||
| 14 | 14 | ||
| 15 | -#include "../utils/device_utils.h" | 15 | +#include "../utils/wq_device_utils.h" |
| 16 | 16 | ||
| 17 | namespace Mc2WeightQuantBatchMatmulV2::Arch35::Catlass { | 17 | namespace Mc2WeightQuantBatchMatmulV2::Arch35::Catlass { |
| 18 | 18 | ||
Rmc2/3rd/weight_quant_batch_matmul_v2/op_kernel/arch35/catlass/pipeline/utils.h→mc2/3rd/weight_quant_batch_matmul_v2/op_kernel/arch35/catlass/pipeline/wq_pipeline_utils.h+3-3
| @@ -7,10 +7,10 @@ | |||
| 7 | * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE. | 7 | * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE. |
| 8 | * See LICENSE in the root of the software repository for the full text of the License. | 8 | * See LICENSE in the root of the software repository for the full text of the License. |
| 9 | */ | 9 | */ |
| 10 | -#ifndef ARCH35_CATLASS_PIPELINE_UTILS_H | 10 | +#ifndef WQ_PIPELINE_UTILS_H |
| 11 | -#define ARCH35_CATLASS_PIPELINE_UTILS_H | 11 | +#define WQ_PIPELINE_UTILS_H |
| 12 | 12 | ||
| 13 | -#include "../utils/device_utils.h" | 13 | +#include "../utils/wq_device_utils.h" |
| 14 | 14 | ||
| 15 | 15 | ||
| 16 | using AscendC::HardEvent; | 16 | using AscendC::HardEvent; |
| @@ -12,7 +12,7 @@ | |||
| 12 | 12 | ||
| 13 | 13 | ||
| 14 | 14 | ||
| 15 | -#include "../utils/device_utils.h" | 15 | +#include "../utils/wq_device_utils.h" |
| 16 | 16 | ||
| 17 | 17 | ||
| 18 | 18 | ||
| @@ -10,7 +10,7 @@ | |||
| 10 | 10 | ||
| 11 | 11 | ||
| 12 | 12 | ||
| 13 | -#include "../utils/device_utils.h" | 13 | +#include "../utils/wq_device_utils.h" |
| 14 | 14 | ||
| 15 | 15 | ||
| 16 | using AscendC::IsSameType; | 16 | using AscendC::IsSameType; |
| @@ -10,7 +10,7 @@ | |||
| 10 | 10 | ||
| 11 | 11 | ||
| 12 | 12 | ||
| 13 | -#include "../utils/device_utils.h" | 13 | +#include "../utils/wq_device_utils.h" |
| 14 | 14 | ||
| 15 | namespace Mc2WeightQuantBatchMatmulV2::Arch35::Catlass { | 15 | namespace Mc2WeightQuantBatchMatmulV2::Arch35::Catlass { |
| 16 | using AscendC::DataCopyParams; | 16 | using AscendC::DataCopyParams; |
| @@ -10,7 +10,7 @@ | |||
| 10 | 10 | ||
| 11 | 11 | ||
| 12 | 12 | ||
| 13 | -#include "../utils/device_utils.h" | 13 | +#include "../utils/wq_device_utils.h" |
| 14 | 14 | ||
| 15 | namespace Mc2WeightQuantBatchMatmulV2::Arch35::Catlass { | 15 | namespace Mc2WeightQuantBatchMatmulV2::Arch35::Catlass { |
| 16 | using AscendC::GlobalTensor; | 16 | using AscendC::GlobalTensor; |
| @@ -10,7 +10,7 @@ | |||
| 10 | 10 | ||
| 11 | 11 | ||
| 12 | 12 | ||
| 13 | -#include "../utils/device_utils.h" | 13 | +#include "../utils/wq_device_utils.h" |
| 14 | 14 | ||
| 15 | namespace Mc2WeightQuantBatchMatmulV2::Arch35::Catlass { | 15 | namespace Mc2WeightQuantBatchMatmulV2::Arch35::Catlass { |
| 16 | using AscendC::DataCopyExtParams; | 16 | using AscendC::DataCopyExtParams; |
| @@ -10,7 +10,7 @@ | |||
| 10 | 10 | ||
| 11 | 11 | ||
| 12 | 12 | ||
| 13 | -#include "device_utils.h" | 13 | +#include "wq_device_utils.h" |
| 14 | 14 | ||
| 15 | namespace Mc2WeightQuantBatchMatmulV2::Arch35::Catlass { | 15 | namespace Mc2WeightQuantBatchMatmulV2::Arch35::Catlass { |
| 16 | template <typename T> | 16 | template <typename T> |
Rmc2/3rd/weight_quant_batch_matmul_v2/op_kernel/arch35/catlass/utils/device_utils.h→mc2/3rd/weight_quant_batch_matmul_v2/op_kernel/arch35/catlass/utils/wq_device_utils.h+2-2
| @@ -7,8 +7,8 @@ | |||
| 7 | * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE. | 7 | * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE. |
| 8 | * See LICENSE in the root of the software repository for the full text of the License. | 8 | * See LICENSE in the root of the software repository for the full text of the License. |
| 9 | */ | 9 | */ |
| 10 | -#ifndef ARCH35_CATLASS_UTILS_DEVICE_UTILS_H | 10 | +#ifndef WQ_DEVICE_UTILS_H |
| 11 | -#define ARCH35_CATLASS_UTILS_DEVICE_UTILS_H | 11 | +#define WQ_DEVICE_UTILS_H |
| 12 | 12 | ||
| 13 | 13 | ||
| 14 | 14 | ||
Rmc2/3rd/weight_quant_batch_matmul_v2/op_kernel/arch35/catlass/dispatch_policy.h→mc2/3rd/weight_quant_batch_matmul_v2/op_kernel/arch35/catlass/wq_dispatch_policy.h+2-2
| @@ -7,8 +7,8 @@ | |||
| 7 | * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE. | 7 | * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE. |
| 8 | * See LICENSE in the root of the software repository for the full text of the License. | 8 | * See LICENSE in the root of the software repository for the full text of the License. |
| 9 | */ | 9 | */ |
| 10 | -#ifndef ARCH35_CATLASS_DISPATCH_POLICY_H | 10 | +#ifndef WQ_DISPATCH_POLICY_H |
| 11 | -#define ARCH35_CATLASS_DISPATCH_POLICY_H | 11 | +#define WQ_DISPATCH_POLICY_H |
| 12 | 12 | ||
| 13 | namespace Mc2WeightQuantBatchMatmulV2::Arch35::Catlass { | 13 | namespace Mc2WeightQuantBatchMatmulV2::Arch35::Catlass { |
| 14 | struct KernelWqbmm {}; | 14 | struct KernelWqbmm {}; |
| @@ -12,7 +12,7 @@ | |||
| 12 | 12 | ||
| 13 | 13 | ||
| 14 | 14 | ||
| 15 | -#include "catlass/dispatch_policy.h" | 15 | +#include "catlass/wq_dispatch_policy.h" |
| 16 | 16 | ||
| 17 | 17 | ||
| 18 | 18 | ||
| @@ -23,7 +23,7 @@ | |||
| 23 | 23 | ||
| 24 | 24 | ||
| 25 | 25 | ||
| 26 | -#include "../../op_kernel/all_gather_matmul_tiling.h" | 26 | +#include "../../op_kernel/all_gather_matmul_op_tiling.h" |
| 27 | 27 | ||
| 28 | namespace optiling { | 28 | namespace optiling { |
| 29 | 29 | ||
| @@ -18,7 +18,7 @@ | |||
| 18 | 18 | ||
| 19 | 19 | ||
| 20 | 20 | ||
| 21 | -#include "all_gather_matmul_tiling.h" | 21 | +#include "all_gather_matmul_op_tiling.h" |
| 22 | 22 | ||
| 23 | 23 | ||
| 24 | 24 | ||
Rmc2/all_gather_matmul/op_kernel/all_gather_matmul_tiling.h→mc2/all_gather_matmul/op_kernel/all_gather_matmul_op_tiling.h+4-4
| @@ -9,12 +9,12 @@ | |||
| 9 | */ | 9 | */ |
| 10 | 10 | ||
| 11 | /*! | 11 | /*! |
| 12 | - * \file all_gather_matmul_tiling.h | 12 | + * \file all_gather_matmul_op_tiling.h |
| 13 | * \brief | 13 | * \brief |
| 14 | */ | 14 | */ |
| 15 | 15 | ||
| 16 | -#ifndef __ALL_GATHER_MATMUL_TILING_H__ | 16 | +#ifndef _ALL_GATHER_MATMUL_OP_TILING_H_ |
| 17 | -#define __ALL_GATHER_MATMUL_TILING_H__ | 17 | +#define _ALL_GATHER_MATMUL_OP_TILING_H_ |
| 18 | 18 | ||
| 19 | 19 | ||
| 20 | 20 | ||
| @@ -43,4 +43,4 @@ public: | |||
| 43 | }; | 43 | }; |
| 44 | } // namespace Mc2Tiling | 44 | } // namespace Mc2Tiling |
| 45 | 45 | ||
| 46 | -#endif //__ALL_GATHER_MATMUL_TILING_H__ | 46 | +#endif //_ALL_GATHER_MATMUL_OP_TILING_H_ |
| @@ -16,7 +16,7 @@ | |||
| 16 | 16 | ||
| 17 | 17 | ||
| 18 | 18 | ||
| 19 | -#include "../../../op_kernel/all_gather_matmul_tiling.h" | 19 | +#include "../../../op_kernel/all_gather_matmul_op_tiling.h" |
| 20 | 20 | ||
| 21 | constexpr uint16_t MAX_TENSOR_CONT = 256; | 21 | constexpr uint16_t MAX_TENSOR_CONT = 256; |
| 22 | constexpr uint16_t MAX_CORE_CONT = 64; | 22 | constexpr uint16_t MAX_CORE_CONT = 64; |
Rmc2/all_gather_matmul_v2/op_kernel/copy_gm_to_l1.h→mc2/all_gather_matmul_v2/op_kernel/agm_v2_copy_gm_to_l1.h+3-3
| @@ -9,12 +9,12 @@ | |||
| 9 | */ | 9 | */ |
| 10 | 10 | ||
| 11 | /*! | 11 | /*! |
| 12 | - * \file copy_gm_to_l1.h | 12 | + * \file agm_v2_copy_gm_to_l1.h |
| 13 | * \brief | 13 | * \brief |
| 14 | */ | 14 | */ |
| 15 | 15 | ||
| 16 | -#ifndef CATLASS_GEMM_TILE_COPY_GM_TO_L1_EXPAND_H | 16 | +#ifndef CATLASS_AGM_V2_COPY_GM_TO_L1_H |
| 17 | -#define CATLASS_GEMM_TILE_COPY_GM_TO_L1_EXPAND_H | 17 | +#define CATLASS_AGM_V2_COPY_GM_TO_L1_H |
| 18 | 18 | ||
| 19 | 19 | ||
| 20 | 20 | ||
| @@ -15,7 +15,7 @@ | |||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | 17 | ||
| 18 | -#include "copy_gm_to_l1.h" | 18 | +#include "agm_v2_copy_gm_to_l1.h" |
| 19 | 19 | ||
| 20 | 20 | ||
| 21 | 21 | ||
| @@ -23,8 +23,8 @@ | |||
| 23 | 23 | ||
| 24 | 24 | ||
| 25 | 25 | ||
| 26 | -#include "block_epilogue_dequant.hpp" | 26 | +#include "allto_all_matmul_block_epilogue_dequant.hpp" |
| 27 | -#include "tile_broadcast_add.hpp" | 27 | +#include "allto_all_matmul_tile_broadcast_add.hpp" |
| 28 | 28 | ||
| 29 | 29 | ||
| 30 | 30 | ||
Rmc2/allto_all_matmul/op_kernel/arch22/block_epilogue_dequant.hpp→mc2/allto_all_matmul/op_kernel/arch22/allto_all_matmul_block_epilogue_dequant.hpp+4-4
| @@ -9,12 +9,12 @@ | |||
| 9 | */ | 9 | */ |
| 10 | 10 | ||
| 11 | /*! | 11 | /*! |
| 12 | - * \file block_epilogue_dequant.hpp | 12 | + * \file allto_all_matmul_block_epilogue_dequant.hpp |
| 13 | * \brief | 13 | * \brief |
| 14 | */ | 14 | */ |
| 15 | 15 | ||
| 16 | -#ifndef CATLASS_EPILOGUE_BLOCK_EPILOGUE_DEQUANT_HPP | 16 | +#ifndef ALLTO_ALL_MATMUL_BLOCK_EPILOGUE_DEQUANT_HPP |
| 17 | -#define CATLASS_EPILOGUE_BLOCK_EPILOGUE_DEQUANT_HPP | 17 | +#define ALLTO_ALL_MATMUL_BLOCK_EPILOGUE_DEQUANT_HPP |
| 18 | 18 | ||
| 19 | 19 | ||
| 20 | 20 | ||
| @@ -364,4 +364,4 @@ private: | |||
| 364 | }; | 364 | }; |
| 365 | } // namespace Catlass::Epilogue::Block | 365 | } // namespace Catlass::Epilogue::Block |
| 366 | 366 | ||
| 367 | -#endif // CATLASS_EPILOGUE_BLOCK_EPILOGUE_DEQUANT_HPP | 367 | +#endif // ALLTO_ALL_MATMUL_BLOCK_EPILOGUE_DEQUANT_HPP |
Rmc2/allto_all_matmul/op_kernel/arch22/tile_broadcast_add.hpp→mc2/allto_all_matmul/op_kernel/arch22/allto_all_matmul_tile_broadcast_add.hpp+2-2
| @@ -8,8 +8,8 @@ | |||
| 8 | * See LICENSE in the root of the software repository for the full text of the License. | 8 | * See LICENSE in the root of the software repository for the full text of the License. |
| 9 | */ | 9 | */ |
| 10 | 10 | ||
| 11 | -#ifndef CATLASS_EPILOGUE_TILE_TILE_BROADCAST_ADD_HPP | 11 | +#ifndef CATLASS_ALLTO_ALL_MATMUL_TILE_BROADCAST_ADD_HPP |
| 12 | -#define CATLASS_EPILOGUE_TILE_TILE_BROADCAST_ADD_HPP | 12 | +#define CATLASS_ALLTO_ALL_MATMUL_TILE_BROADCAST_ADD_HPP |
| 13 | 13 | ||
| 14 | 14 | ||
| 15 | 15 | ||
| @@ -29,9 +29,9 @@ | |||
| 29 | 29 | ||
| 30 | 30 | ||
| 31 | 31 | ||
| 32 | -#include "utils.h" | 32 | +#include "apace_st_utils.h" |
| 33 | -#include "apace/utils/constant.h" | 33 | +#include "apace/utils/apace_constant.h" |
| 34 | -#include "apace/utils/common_utils.h" | 34 | +#include "apace/utils/apace_common_utils.h" |
| 35 | 35 | ||
| 36 | 36 | ||
| 37 | 37 | ||
| @@ -29,9 +29,9 @@ | |||
| 29 | 29 | ||
| 30 | 30 | ||
| 31 | 31 | ||
| 32 | -#include "utils.h" | 32 | +#include "apace_st_utils.h" |
| 33 | -#include "apace/utils/constant.h" | 33 | +#include "apace/utils/apace_constant.h" |
| 34 | -#include "apace/utils/common_utils.h" | 34 | +#include "apace/utils/apace_common_utils.h" |
| 35 | 35 | ||
| 36 | 36 | ||
| 37 | 37 | ||
| @@ -28,7 +28,7 @@ | |||
| 28 | 28 | ||
| 29 | 29 | ||
| 30 | 30 | ||
| 31 | -#include "utils.h" | 31 | +#include "apace_st_utils.h" |
| 32 | 32 | ||
| 33 | 33 | ||
| 34 | 34 | ||
| @@ -39,9 +39,9 @@ | |||
| 39 | 39 | ||
| 40 | 40 | ||
| 41 | 41 | ||
| 42 | -#include "utils.h" | 42 | +#include "apace_st_utils.h" |
| 43 | -#include "apace/utils/constant.h" | 43 | +#include "apace/utils/apace_constant.h" |
| 44 | -#include "apace/utils/common_utils.h" | 44 | +#include "apace/utils/apace_common_utils.h" |
| 45 | 45 | ||
| 46 | 46 | ||
| 47 | 47 | ||
Rmc2/common/op_kernel/apace/tests/st/utils/utils.h→mc2/common/op_kernel/apace/tests/st/utils/apace_st_utils.h+2-2
| @@ -7,8 +7,8 @@ | |||
| 7 | * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE. | 7 | * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE. |
| 8 | * See LICENSE in the root of the software repository for the full text of the License. | 8 | * See LICENSE in the root of the software repository for the full text of the License. |
| 9 | */ | 9 | */ |
| 10 | -#ifndef UTILS_H | 10 | +#ifndef APACE_ST_UTILS_H |
| 11 | -#define UTILS_H | 11 | +#define APACE_ST_UTILS_H |
| 12 | 12 | ||
| 13 | 13 | ||
| 14 | 14 | ||
| @@ -26,7 +26,7 @@ | |||
| 26 | 26 | ||
| 27 | 27 | ||
| 28 | 28 | ||
| 29 | -#include "utils.h" | 29 | +#include "apace_st_utils.h" |
| 30 | 30 | ||
| 31 | class RootInfoExchanger { | 31 | class RootInfoExchanger { |
| 32 | public: | 32 | public: |
| @@ -18,10 +18,10 @@ | |||
| 18 | 18 | ||
| 19 | 19 | ||
| 20 | 20 | ||
| 21 | -#include "../utils/common_utils.h" | 21 | +#include "../utils/apace_common_utils.h" |
| 22 | 22 | ||
| 23 | 23 | ||
| 24 | -#include "../utils/constant.h" | 24 | +#include "../utils/apace_constant.h" |
| 25 | 25 | ||
| 26 | template <mm::DataType aDataType, mm::DataType bDataType> | 26 | template <mm::DataType aDataType, mm::DataType bDataType> |
| 27 | class QuantMatmulTilingBase { | 27 | class QuantMatmulTilingBase { |
Rmc2/common/op_kernel/apace/utils/common_utils.h→mc2/common/op_kernel/apace/utils/apace_common_utils.h+1-1
| @@ -9,7 +9,7 @@ | |||
| 9 | */ | 9 | */ |
| 10 | 10 | ||
| 11 | /*! | 11 | /*! |
| 12 | - * \file common_utils.h | 12 | + * \file apace_common_utils.h |
| 13 | * \brief Host-side argument parsing, shape helpers, and error utilities for matmul examples. | 13 | * \brief Host-side argument parsing, shape helpers, and error utilities for matmul examples. |
| 14 | */ | 14 | */ |
| 15 | 15 | ||
| @@ -9,7 +9,7 @@ | |||
| 9 | */ | 9 | */ |
| 10 | 10 | ||
| 11 | /*! | 11 | /*! |
| 12 | - * \file constant.h | 12 | + * \file apace_constant.h |
| 13 | * \brief Shared constants and helper types for matmul. | 13 | * \brief Shared constants and helper types for matmul. |
| 14 | */ | 14 | */ |
| 15 | 15 | ||
| @@ -15,7 +15,7 @@ | |||
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | 17 | ||
| 18 | -#include "rms_norm_base.h" | 18 | +#include "mararn_rms_norm_base.h" |
| 19 | using namespace AscendC; | 19 | using namespace AscendC; |
| 20 | 20 | ||
| 21 | template <typename T> | 21 | template <typename T> |
| @@ -14,7 +14,7 @@ | |||
| 14 | */ | 14 | */ |
| 15 | 15 | ||
| 16 | 16 | ||
| 17 | -#include "rms_norm_base.h" | 17 | +#include "mararn_rms_norm_base.h" |
| 18 | using namespace AscendC; | 18 | using namespace AscendC; |
| 19 | 19 | ||
| 20 | template <typename T> | 20 | template <typename T> |
Rmc2/matmul_all_reduce_add_rms_norm/op_kernel/reduce_common.h→mc2/matmul_all_reduce_add_rms_norm/op_kernel/mararn_reduce_common.h+3-3
| @@ -9,11 +9,11 @@ | |||
| 9 | */ | 9 | */ |
| 10 | 10 | ||
| 11 | /*! | 11 | /*! |
| 12 | - * \file reduce_common.h | 12 | + * \file mararn_reduce_common.h |
| 13 | * \brief | 13 | * \brief |
| 14 | */ | 14 | */ |
| 15 | -#ifndef _REDUCE_COMMON_H_ | 15 | +#ifndef _MARARN_REDUCE_COMMON_H_ |
| 16 | -#define _REDUCE_COMMON_H_ | 16 | +#define _MARARN_REDUCE_COMMON_H_ |
| 17 | 17 | ||
| 18 | 18 | ||
| 19 | 19 | ||
Rmc2/matmul_all_reduce_add_rms_norm/op_kernel/rms_norm_base.h→mc2/matmul_all_reduce_add_rms_norm/op_kernel/mararn_rms_norm_base.h+4-4
| @@ -9,17 +9,17 @@ | |||
| 9 | */ | 9 | */ |
| 10 | 10 | ||
| 11 | /*! | 11 | /*! |
| 12 | - * \file rms_norm_base.h | 12 | + * \file mararn_rms_norm_base.h |
| 13 | * \brief | 13 | * \brief |
| 14 | */ | 14 | */ |
| 15 | -#ifndef RMS_NORM_BASE_H | 15 | +#ifndef MARARN_RMS_NORM_BASE_H |
| 16 | -#define RMS_NORM_BASE_H | 16 | +#define MARARN_RMS_NORM_BASE_H |
| 17 | 17 | ||
| 18 | 18 | ||
| 19 | 19 | ||
| 20 | 20 | ||
| 21 | 21 | ||
| 22 | -#include "reduce_common.h" | 22 | +#include "mararn_reduce_common.h" |
| 23 | 23 | ||
| 24 | using namespace AscendC; | 24 | using namespace AscendC; |
| 25 | 25 | ||
Rmc2/matmul_allto_all/op_kernel/arch22/block_epilogue_dequant.hpp→mc2/matmul_allto_all/op_kernel/arch22/matmul_allto_all_block_epilogue_dequant.hpp+3-3
| @@ -8,8 +8,8 @@ | |||
| 8 | * See LICENSE in the root of the software repository for the full text of the License. | 8 | * See LICENSE in the root of the software repository for the full text of the License. |
| 9 | */ | 9 | */ |
| 10 | 10 | ||
| 11 | -#ifndef CATLASS_EPILOGUE_BLOCK_EPILOGUE_DEQUANT_HPP | 11 | +#ifndef MATMUL_ALLTO_ALL_BLOCK_EPILOGUE_DEQUANT_HPP |
| 12 | -#define CATLASS_EPILOGUE_BLOCK_EPILOGUE_DEQUANT_HPP | 12 | +#define MATMUL_ALLTO_ALL_BLOCK_EPILOGUE_DEQUANT_HPP |
| 13 | 13 | ||
| 14 | 14 | ||
| 15 | 15 | ||
| @@ -347,4 +347,4 @@ private: | |||
| 347 | }; | 347 | }; |
| 348 | } // namespace Catlass::Epilogue::Block | 348 | } // namespace Catlass::Epilogue::Block |
| 349 | 349 | ||
| 350 | -#endif // CATLASS_EPILOGUE_BLOCK_EPILOGUE_DEQUANT_HPP | 350 | +#endif // MATMUL_ALLTO_ALL_BLOCK_EPILOGUE_DEQUANT_HPP |
Rmc2/matmul_allto_all/op_kernel/arch22/quant_matmul.hpp→mc2/matmul_allto_all/op_kernel/arch22/matmul_allto_all_quant_matmul.hpp+4-4
| @@ -9,12 +9,12 @@ | |||
| 9 | */ | 9 | */ |
| 10 | 10 | ||
| 11 | /*! | 11 | /*! |
| 12 | - * \file quant_matmul.hpp | 12 | + * \file matmul_allto_all_quant_matmul.hpp |
| 13 | * \brief | 13 | * \brief |
| 14 | */ | 14 | */ |
| 15 | 15 | ||
| 16 | -#ifndef CATLASS_GEMM_KERNEL_QUANT_MATMUL_HPP | 16 | +#ifndef CATLASS_MATMUL_ALLTO_ALL_QUANT_MATMUL_HPP |
| 17 | -#define CATLASS_GEMM_KERNEL_QUANT_MATMUL_HPP | 17 | +#define CATLASS_MATMUL_ALLTO_ALL_QUANT_MATMUL_HPP |
| 18 | 18 | ||
| 19 | 19 | ||
| 20 | 20 | ||
| @@ -411,4 +411,4 @@ private: | |||
| 411 | }; | 411 | }; |
| 412 | } // namespace Catlass::Gemm::Kernel | 412 | } // namespace Catlass::Gemm::Kernel |
| 413 | 413 | ||
| 414 | -#endif // CATLASS_GEMM_KERNEL_QUANT_MATMUL_HPP | 414 | +#endif // CATLASS_MATMUL_ALLTO_ALL_QUANT_MATMUL_HPP |
Rmc2/matmul_allto_all/op_kernel/arch22/tile_broadcast_add.hpp→mc2/matmul_allto_all/op_kernel/arch22/matmul_allto_all_tile_broadcast_add.hpp+2-2
| @@ -8,8 +8,8 @@ | |||
| 8 | * See LICENSE in the root of the software repository for the full text of the License. | 8 | * See LICENSE in the root of the software repository for the full text of the License. |
| 9 | */ | 9 | */ |
| 10 | 10 | ||
| 11 | -#ifndef CATLASS_EPILOGUE_TILE_TILE_BROADCAST_ADD_HPP | 11 | +#ifndef CATLASS_MATMUL_ALLTO_ALL_TILE_BROADCAST_ADD_HPP |
| 12 | -#define CATLASS_EPILOGUE_TILE_TILE_BROADCAST_ADD_HPP | 12 | +#define CATLASS_MATMUL_ALLTO_ALL_TILE_BROADCAST_ADD_HPP |
| 13 | 13 | ||
| 14 | 14 | ||
| 15 | 15 | ||
| @@ -32,7 +32,7 @@ | |||
| 32 | 32 | ||
| 33 | 33 | ||
| 34 | 34 | ||
| 35 | -#include "tile_broadcast_add.hpp" | 35 | +#include "mrs_v2_tile_broadcast_add.hpp" |
| 36 | 36 | ||
| 37 | using namespace matmulReduceScatterV2_util; | 37 | using namespace matmulReduceScatterV2_util; |
| 38 | namespace dequant { | 38 | namespace dequant { |
Rmc2/matmul_reduce_scatter_v2/op_kernel/block_mmad_preload_fixpipe.h→mc2/matmul_reduce_scatter_v2/op_kernel/mrs_v2_block_mmad_preload_fixpipe.h+6-6
| @@ -9,14 +9,14 @@ | |||
| 9 | */ | 9 | */ |
| 10 | 10 | ||
| 11 | /*! | 11 | /*! |
| 12 | - * \file block_mmad_preload_fixpipe.h | 12 | + * \file mrs_v2_block_mmad_preload_fixpipe.h |
| 13 | * \brief | 13 | * \brief |
| 14 | */ | 14 | */ |
| 15 | -#ifndef BLOCK_MMAD_PRELOAD_FIXPIPE_H | 15 | +#ifndef MRS_V2_BLOCK_MMAD_PRELOAD_FIXPIPE_H |
| 16 | -#define BLOCK_MMAD_PRELOAD_FIXPIPE_H | 16 | +#define MRS_V2_BLOCK_MMAD_PRELOAD_FIXPIPE_H |
| 17 | 17 | ||
| 18 | -#include "copy_gm_to_l1.h" | 18 | +#include "mrs_v2_copy_gm_to_l1.h" |
| 19 | -#include "copy_l0c_to_gm.h" | 19 | +#include "mrs_v2_copy_l0c_to_gm.h" |
| 20 | 20 | ||
| 21 | 21 | ||
| 22 | 22 | ||
| @@ -398,4 +398,4 @@ protected: | |||
| 398 | CopyL0CToGm copyL0CToGm; | 398 | CopyL0CToGm copyL0CToGm; |
| 399 | }; | 399 | }; |
| 400 | } // namespace Catlass::Gemm::Block | 400 | } // namespace Catlass::Gemm::Block |
| 401 | -#endif // BLOCK_MMAD_PRELOAD_FIXPIPE_H | 401 | +#endif // MRS_V2_BLOCK_MMAD_PRELOAD_FIXPIPE_H |
Rmc2/matmul_reduce_scatter_v2/op_kernel/copy_gm_to_l1.h→mc2/matmul_reduce_scatter_v2/op_kernel/mrs_v2_copy_gm_to_l1.h+3-3
| @@ -9,12 +9,12 @@ | |||
| 9 | */ | 9 | */ |
| 10 | 10 | ||
| 11 | /*! | 11 | /*! |
| 12 | - * \file copy_gm_to_l1.h | 12 | + * \file mrs_v2_copy_gm_to_l1.h |
| 13 | * \brief | 13 | * \brief |
| 14 | */ | 14 | */ |
| 15 | 15 | ||
| 16 | -#ifndef COPY_GM_TO_L1_H | 16 | +#ifndef MRS_V2_COPY_GM_TO_L1_H |
| 17 | -#define COPY_GM_TO_L1_H | 17 | +#define MRS_V2_COPY_GM_TO_L1_H |
| 18 | 18 | ||
| 19 | 19 | ||
| 20 | 20 | ||
Rmc2/matmul_reduce_scatter_v2/op_kernel/copy_l0c_to_gm.h→mc2/matmul_reduce_scatter_v2/op_kernel/mrs_v2_copy_l0c_to_gm.h+3-3
| @@ -9,12 +9,12 @@ | |||
| 9 | */ | 9 | */ |
| 10 | 10 | ||
| 11 | /*! | 11 | /*! |
| 12 | - * \file copy_l0c_to_gm.h | 12 | + * \file mrs_v2_copy_l0c_to_gm.h |
| 13 | * \brief | 13 | * \brief |
| 14 | */ | 14 | */ |
| 15 | 15 | ||
| 16 | -#ifndef COPY_L0C_TO_GM_H | 16 | +#ifndef MRS_V2_COPY_L0C_TO_GM_H |
| 17 | -#define COPY_L0C_TO_GM_H | 17 | +#define MRS_V2_COPY_L0C_TO_GM_H |
| 18 | 18 | ||
| 19 | 19 | ||
| 20 | 20 | ||
Rmc2/matmul_reduce_scatter_v2/op_kernel/tile_broadcast_add.hpp→mc2/matmul_reduce_scatter_v2/op_kernel/mrs_v2_tile_broadcast_add.hpp+2-2
| @@ -8,8 +8,8 @@ | |||
| 8 | * See LICENSE in the root of the software repository for the full text of the License. | 8 | * See LICENSE in the root of the software repository for the full text of the License. |
| 9 | */ | 9 | */ |
| 10 | 10 | ||
| 11 | -#ifndef CATLASS_EPILOGUE_TILE_TILE_BROADCAST_ADD_HPP | 11 | +#ifndef CATLASS_MRS_V2_TILE_BROADCAST_ADD_HPP |
| 12 | -#define CATLASS_EPILOGUE_TILE_TILE_BROADCAST_ADD_HPP | 12 | +#define CATLASS_MRS_V2_TILE_BROADCAST_ADD_HPP |
| 13 | 13 | ||
| 14 | 14 | ||
| 15 | 15 | ||
| @@ -27,7 +27,7 @@ | |||
| 27 | 27 | ||
| 28 | 28 | ||
| 29 | 29 | ||
| 30 | -#include "../moe_distribute_base.h" | 30 | +#include "../moe_distribute_combine_setup_base.h" |
| 31 | 31 | ||
| 32 | 32 | ||
| 33 | namespace MoeDistributeCombineSetupImpl { | 33 | namespace MoeDistributeCombineSetupImpl { |
Rmc2/moe_distribute_combine_setup/op_kernel/moe_distribute_base.h→mc2/moe_distribute_combine_setup/op_kernel/moe_distribute_combine_setup_base.h+3-3
| @@ -9,12 +9,12 @@ | |||
| 9 | */ | 9 | */ |
| 10 | 10 | ||
| 11 | /*! | 11 | /*! |
| 12 | - * \file moe_distribute_base.h | 12 | + * \file moe_distribute_combine_setup_base.h |
| 13 | * \brief | 13 | * \brief |
| 14 | */ | 14 | */ |
| 15 | 15 | ||
| 16 | -#ifndef MOE_DISTRIBUTE_BASE_H | 16 | +#ifndef MOE_DISTRIBUTE_COMBINE_SETUP_BASE_H |
| 17 | -#define MOE_DISTRIBUTE_BASE_H | 17 | +#define MOE_DISTRIBUTE_COMBINE_SETUP_BASE_H |
| 18 | 18 | ||
| 19 | 19 | ||
| 20 | 20 | ||
| @@ -21,10 +21,10 @@ | |||
| 21 | 21 | ||
| 22 | 22 | ||
| 23 | 23 | ||
| 24 | -#if __has_include("../../moe_distribute_combine_setup/moe_distribute_base.h") | 24 | +#if __has_include("../../moe_distribute_combine_setup/moe_distribute_combine_setup_base.h") |
| 25 | -#include "../../moe_distribute_combine_setup/moe_distribute_base.h" | 25 | +#include "../../moe_distribute_combine_setup/moe_distribute_combine_setup_base.h" |
| 26 | 26 | ||
| 27 | -#include "../../../moe_distribute_combine_setup/op_kernel/moe_distribute_base.h" | 27 | +#include "../../../moe_distribute_combine_setup/op_kernel/moe_distribute_combine_setup_base.h" |
| 28 | 28 | ||
| 29 | 29 | ||
| 30 | 30 | ||
| @@ -24,10 +24,10 @@ | |||
| 24 | 24 | ||
| 25 | 25 | ||
| 26 | 26 | ||
| 27 | -#if __has_include("../../moe_distribute_combine_setup/moe_distribute_base.h") | 27 | +#if __has_include("../../moe_distribute_combine_setup/moe_distribute_combine_setup_base.h") |
| 28 | -#include "../../moe_distribute_combine_setup/moe_distribute_base.h" | 28 | +#include "../../moe_distribute_combine_setup/moe_distribute_combine_setup_base.h" |
| 29 | 29 | ||
| 30 | -#include "../../../moe_distribute_combine_setup/op_kernel/moe_distribute_base.h" | 30 | +#include "../../../moe_distribute_combine_setup/op_kernel/moe_distribute_combine_setup_base.h" |
| 31 | 31 | ||
| 32 | 32 | ||
| 33 | namespace MoeDistributeCombineTeardownImpl { | 33 | namespace MoeDistributeCombineTeardownImpl { |
| @@ -22,7 +22,7 @@ | |||
| 22 | 22 | ||
| 23 | 23 | ||
| 24 | 24 | ||
| 25 | -#include "../moe_distribute_base.h" | 25 | +#include "../moe_distribute_dispatch_setup_base.h" |
| 26 | 26 | ||
| 27 | 27 | ||
| 28 | namespace Mc2Kernel { | 28 | namespace Mc2Kernel { |
| @@ -19,7 +19,7 @@ | |||
| 19 | 19 | ||
| 20 | 20 | ||
| 21 | 21 | ||
| 22 | -#include "../moe_distribute_base.h" | 22 | +#include "../moe_distribute_dispatch_setup_base.h" |
| 23 | 23 | ||
| 24 | 24 | ||
| 25 | 25 | ||
Rmc2/moe_distribute_dispatch_setup/op_kernel/moe_distribute_base.h→mc2/moe_distribute_dispatch_setup/op_kernel/moe_distribute_dispatch_setup_base.h+3-3
| @@ -9,12 +9,12 @@ | |||
| 9 | */ | 9 | */ |
| 10 | 10 | ||
| 11 | /*! | 11 | /*! |
| 12 | - * \file moe_distribute_base.h | 12 | + * \file moe_distribute_dispatch_setup_base.h |
| 13 | * \brief | 13 | * \brief |
| 14 | */ | 14 | */ |
| 15 | 15 | ||
| 16 | -#ifndef MOE_DISTRIBUTE_BASE_H | 16 | +#ifndef MOE_DISTRIBUTE_DISPATCH_SETUP_BASE_H |
| 17 | -#define MOE_DISTRIBUTE_BASE_H | 17 | +#define MOE_DISTRIBUTE_DISPATCH_SETUP_BASE_H |
| 18 | 18 | ||
| 19 | 19 | ||
| 20 | 20 | ||
Mmc2/moe_distribute_dispatch_teardown/op_kernel/arch22/moe_distribute_dispatch_teardown_arch22.h+3-3
| @@ -22,10 +22,10 @@ | |||
| 22 | 22 | ||
| 23 | 23 | ||
| 24 | 24 | ||
| 25 | -#if __has_include("../../moe_distribute_dispatch_setup/moe_distribute_base.h") | 25 | +#if __has_include("../../moe_distribute_dispatch_setup/moe_distribute_dispatch_setup_base.h") |
| 26 | -#include "../../moe_distribute_dispatch_setup/moe_distribute_base.h" | 26 | +#include "../../moe_distribute_dispatch_setup/moe_distribute_dispatch_setup_base.h" |
| 27 | 27 | ||
| 28 | -#include "../../../moe_distribute_dispatch_setup/op_kernel/moe_distribute_base.h" | 28 | +#include "../../../moe_distribute_dispatch_setup/op_kernel/moe_distribute_dispatch_setup_base.h" |
| 29 | 29 | ||
| 30 | 30 | ||
| 31 | 31 | ||
Mmc2/moe_distribute_dispatch_teardown/op_kernel/arch35/moe_distribute_dispatch_teardown_arch35.h+2-2
| @@ -20,11 +20,11 @@ | |||
| 20 | 20 | ||
| 21 | 21 | ||
| 22 | 22 | ||
| 23 | -#include "../../moe_distribute_dispatch_setup/moe_distribute_base.h" | 23 | +#include "../../moe_distribute_dispatch_setup/moe_distribute_dispatch_setup_base.h" |
| 24 | 24 | ||
| 25 | 25 | ||
| 26 | 26 | ||
| 27 | -#include "../../../moe_distribute_dispatch_setup/op_kernel/moe_distribute_base.h" | 27 | +#include "../../../moe_distribute_dispatch_setup/op_kernel/moe_distribute_dispatch_setup_base.h" |
| 28 | 28 | ||
| 29 | 29 | ||
| 30 | 30 | ||
| @@ -23,7 +23,7 @@ | |||
| 23 | 23 | ||
| 24 | 24 | ||
| 25 | 25 | ||
| 26 | -#include "../../quant_reduce_scatter/op_kernel/utils.h" | 26 | +#include "../../quant_reduce_scatter/op_kernel/quant_reduce_scatter_utils.h" |
| 27 | 27 | ||
| 28 | 28 | ||
| 29 | 29 | ||
| @@ -26,7 +26,7 @@ | |||
| 26 | 26 | ||
| 27 | 27 | ||
| 28 | 28 | ||
| 29 | -#include "utils.h" | 29 | +#include "quant_reduce_scatter_utils.h" |
| 30 | 30 | ||
| 31 | 31 | ||
| 32 | 32 | ||
Rmc2/quant_reduce_scatter/op_kernel/utils.h→mc2/quant_reduce_scatter/op_kernel/quant_reduce_scatter_utils.h+3-3
| @@ -9,12 +9,12 @@ | |||
| 9 | */ | 9 | */ |
| 10 | 10 | ||
| 11 | /*! | 11 | /*! |
| 12 | - * \file utils.h | 12 | + * \file quant_reduce_scatter_utils.h |
| 13 | * \brief | 13 | * \brief |
| 14 | */ | 14 | */ |
| 15 | 15 | ||
| 16 | -#ifndef UTILS_H | 16 | +#ifndef QUANT_REDUCE_SCATTER_UTILS_H |
| 17 | -#define UTILS_H | 17 | +#define QUANT_REDUCE_SCATTER_UTILS_H |
| 18 | 18 | ||
| 19 | 19 | ||
| 20 | namespace AscendC { | 20 | namespace AscendC { |
Rmoe/moe_token_permute_with_routing_map/op_kernel/masked_select_v3.h→moe/moe_token_permute_with_routing_map/op_kernel/moe_token_permute_masked_select_v3.h+4-4
| @@ -9,11 +9,11 @@ | |||
| 9 | */ | 9 | */ |
| 10 | 10 | ||
| 11 | /*! | 11 | /*! |
| 12 | - * \file masked_select_v3.h | 12 | + * \file moe_token_permute_masked_select_v3.h |
| 13 | * \brief | 13 | * \brief |
| 14 | */ | 14 | */ |
| 15 | -#ifndef MASKED_SELECT_V3_H_ | 15 | +#ifndef MOE_TOKEN_PERMUTE_MASKED_SELECT_V3_H |
| 16 | -#define MASKED_SELECT_V3_H_ | 16 | +#define MOE_TOKEN_PERMUTE_MASKED_SELECT_V3_H |
| 17 | 17 | ||
| 18 | 18 | ||
| 19 | 19 | ||
| @@ -682,4 +682,4 @@ private: | |||
| 682 | }; | 682 | }; |
| 683 | } // namespace AscendC | 683 | } // namespace AscendC |
| 684 | 684 | ||
| 685 | -#endif // MASKED_SELECT_V3_H_ | 685 | +#endif // MOE_TOKEN_PERMUTE_MASKED_SELECT_V3_H |
| @@ -19,7 +19,7 @@ | |||
| 19 | 19 | ||
| 20 | 20 | ||
| 21 | 21 | ||
| 22 | -#include "masked_select_v3.h" | 22 | +#include "moe_token_permute_masked_select_v3.h" |
| 23 | 23 | ||
| 24 | 24 | ||
| 25 | 25 | ||