Skip to content

Commit

Permalink
gpu: jit: gemm: re-add missing ATS/DG2 TN igemv kernel
Browse files Browse the repository at this point in the history
  • Loading branch information
@petercad @karturov
petercad authored and karturov committed Oct 7, 2022
1 parent 9bc87e6 commit c8adc17
Showing 1 changed file with 4 additions and 2 deletions.
6 changes: 4 additions & 2 deletions src/gpu/jit/gemm/kernel.db
View file
Original file line number Diff line number Diff line change
Expand Up @@ -15,8 +15,8 @@
*******************************************************************************/

/*@kcatalog@*/
kcatalog::FlatCatalog<471> _CATALOG_
{1, 3408, 471, {
kcatalog::FlatCatalog<473> _CATALOG_
{1, 3408, 473, {
{{'9', "gemm", {"H", "H", "H"}, {"A", "B", "N"}}, {-1, -1, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {1, 1, 1}, ""}, "ab4x2 ab8x2 ab k16 acb bm8192 bn4096 bk1536", {8, (LoopType) 255, 128, {(LoopType) 0, (LoopType) 1, (LoopType) 255}, {8192, 4096, 1536}, {4096, 4096, 1536}, {32, 32, 16}, {2, 8, 1}, 1, (WGType) 0, false, false, false, 0, 0, {64, 64, 2}, {true, true, true}}, {'W', 1, {1024}}},
{{'9', "gemm", {"H", "H", "S"}, {"A", "B", "N"}}, {-1, -1, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {1, 1, 1}, ""}, "ab4 ab4 as k8 acb bm8192 bn4096 bk1536", {8, (LoopType) 255, 128, {(LoopType) 0, (LoopType) 1, (LoopType) 255}, {8192, 4096, 1536}, {4096, 4096, 1536}, {16, 32, 8}, {2, 8, 1}, 1, (WGType) 0, false, false, false, 0, 0, {32, 64, 4}, {true, true, true}}, {'W', 1, {512}}},
{{'9', "gemm", {"H", "H", "H"}, {"N", "N", "N"}}, {-1, -1, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {1, 1, 1}, ""}, "ab4x2 as8 ab k16 l4 acb", {8, (LoopType) 255, 128, {(LoopType) 0, (LoopType) 1, (LoopType) 255}, {4096, 4096, 2048}, {4096, 4096, 2048}, {32, 32, 16}, {2, 8, 1}, 1, (WGType) 0, true, false, false, 0, 0, {2, 2, 2}, {true, true, true}}, {'W', 1, {1024}}},
Expand Down Expand Up @@ -114,6 +114,7 @@ kcatalog::FlatCatalog<471> _CATALOG_
{{'D', "gemm", {"O", "O", "I"}, {"N", "N", "N"}}, {-1, -1, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {1, 1, 1}, "xyzno"}, "sb8 sb32 sb l4 cab1 wg 4x4 cs di pab", {8, (LoopType) 0, 128, {(LoopType) 0, (LoopType) 1, (LoopType) 255}, {8192, 8192, 2048}, {8192, 8192, 1024}, {32, 16, 32}, {4, 4, 1}, 1, (WGType) 1, true, false, false, 6144, 0, {1, 1, 4}, {false, false, false}}, {'W', 1, {512}}},
{{'D', "gemm", {"O", "O", "I"}, {"N", "N", "N"}}, {-1, -1, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {1, 1, 1}, "xyz"}, "sb8x2 su16x2 sb l4 ca1 wg 2x8 cs di", {8, (LoopType) 0, 128, {(LoopType) 0, (LoopType) 1, (LoopType) 255}, {8192, 8192, 2048}, {8192, 8192, 1024}, {32, 4, 32}, {2, 8, 1}, 1, (WGType) 1, true, false, false, 2048, 0, {1, 1, 4}, {false, false, false}}, {'W', 1, {128}}},
{{'D', "gemm", {"O", "O", "I"}, {"N", "T", "N"}}, {-1, -1, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {1, 1, 1}, "xyz"}, "sb8x2 sb8x2 sb l4 cab1 wg 4x4 cs di", {8, (LoopType) 0, 128, {(LoopType) 0, (LoopType) 1, (LoopType) 255}, {8192, 8192, 2048}, {8192, 8192, 1024}, {32, 16, 16}, {4, 4, 1}, 1, (WGType) 1, true, false, false, 3072, 0, {1, 1, 4}, {false, false, false}}, {'W', 1, {512}}},
{{'D', "gemm", {"O", "O", "I"}, {"T", "N", "N"}}, {-1, -1, {-1, -1, -1}, {-1, 1, -1}, {-1, -1, -1}, {-1, 1, -1}, {1, 1, 1}, "xyz"}, "su16 sb64x2 sb wg 2x1x16 l4 cs kr", {8, (LoopType) 0, 128, {(LoopType) 0, (LoopType) 1, (LoopType) 2}, {8192, 8192, 2048}, {8192, 8192, 1024}, {32, 1, 128}, {2, 1, 16}, 1, (WGType) 0, true, false, true, 0, 2048, {1, 1, 4}, {false, false, false}}, {'W', 1, {32}}},
{{'D', "gemm", {"O", "O", "I"}, {"T", "N", "N"}}, {-1, -1, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {1, 1, 1}, "xyz"}, "sb32 sb32 sb l4 cab1 wg 4x4 cs di", {8, (LoopType) 0, 128, {(LoopType) 0, (LoopType) 1, (LoopType) 255}, {8192, 8192, 2048}, {8192, 8192, 1024}, {16, 16, 32}, {4, 4, 1}, 1, (WGType) 1, true, false, false, 4096, 0, {1, 1, 4}, {false, false, false}}, {'W', 1, {256}}},
{{'D', "gemm", {"O", "O", "I"}, {"T", "T", "N"}}, {-1, -1, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {1, 1, 1}, "xyz"}, "sb32 sb32 sb l4 cab1 wg 4x4 cs di", {8, (LoopType) 0, 128, {(LoopType) 0, (LoopType) 1, (LoopType) 255}, {8192, 8192, 2048}, {8192, 8192, 1024}, {16, 16, 32}, {4, 4, 1}, 1, (WGType) 1, true, false, false, 4096, 0, {1, 1, 4}, {false, false, false}}, {'W', 1, {256}}},
{{'D', "gemm", {"S", "S", "S"}, {"A", "B", "N"}}, {-1, -1, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {1, 1, 1}, "xyz"}, "sb2x2 sb4x2 sb ca1x2 wg 2x8 cs di hi bk1536/1536", {8, (LoopType) 0, 128, {(LoopType) 144, (LoopType) 255, (LoopType) 255}, {524288, 262144, 1536}, {8192, 8192, 1536}, {32, 16, 16}, {2, 8, 1}, 1, (WGType) 1, false, false, false, 2048, 0, {128, 64, 4}, {false, false, false}}, {'W', 1, {512}}},
Expand Down Expand Up @@ -163,6 +164,7 @@ kcatalog::FlatCatalog<471> _CATALOG_
{{'E', "gemm", {"O", "O", "I"}, {"N", "N", "N"}}, {-1, -1, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {1, 1, 1}, "xyzno"}, "sb8 sb32 sb l4 cab1 wg 4x4 cs di pab", {8, (LoopType) 0, 128, {(LoopType) 0, (LoopType) 1, (LoopType) 255}, {8192, 8192, 2048}, {8192, 8192, 1024}, {32, 16, 32}, {4, 4, 1}, 1, (WGType) 1, true, false, false, 6144, 0, {1, 1, 4}, {false, false, false}}, {'W', 1, {512}}},
{{'E', "gemm", {"O", "O", "I"}, {"N", "N", "N"}}, {-1, -1, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {1, 1, 1}, "xyz"}, "sb32 sb32 sb l4 cab1 wg 4x4 cs di", {8, (LoopType) 0, 128, {(LoopType) 0, (LoopType) 1, (LoopType) 255}, {8192, 8192, 2048}, {8192, 8192, 1024}, {16, 16, 32}, {4, 4, 1}, 1, (WGType) 1, true, false, false, 4096, 0, {1, 1, 4}, {false, false, false}}, {'W', 1, {256}}},
{{'E', "gemm", {"O", "O", "I"}, {"N", "T", "N"}}, {-1, -1, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {1, 1, 1}, "xyz"}, "sb8x2 sb8x2 sb l4 cab1 wg 4x4 cs di", {8, (LoopType) 0, 128, {(LoopType) 0, (LoopType) 1, (LoopType) 255}, {8192, 8192, 2048}, {8192, 8192, 1024}, {32, 16, 16}, {4, 4, 1}, 1, (WGType) 1, true, false, false, 3072, 0, {1, 1, 4}, {false, false, false}}, {'W', 1, {512}}},
{{'E', "gemm", {"O", "O", "I"}, {"T", "N", "N"}}, {-1, -1, {-1, -1, -1}, {-1, 1, -1}, {-1, -1, -1}, {-1, 1, -1}, {1, 1, 1}, "xyz"}, "su16 sb64x2 sb wg 2x1x16 l4 cs kr", {8, (LoopType) 0, 128, {(LoopType) 0, (LoopType) 1, (LoopType) 2}, {8192, 8192, 2048}, {8192, 8192, 1024}, {32, 1, 128}, {2, 1, 16}, 1, (WGType) 0, true, false, true, 0, 2048, {1, 1, 4}, {false, false, false}}, {'W', 1, {32}}},
{{'E', "gemm", {"O", "O", "I"}, {"T", "N", "N"}}, {-1, -1, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {1, 1, 1}, "xyz"}, "sb32 sb32 sb l4 cab1 wg 4x4 cs di", {8, (LoopType) 0, 128, {(LoopType) 0, (LoopType) 1, (LoopType) 255}, {8192, 8192, 2048}, {8192, 8192, 1024}, {16, 16, 32}, {4, 4, 1}, 1, (WGType) 1, true, false, false, 4096, 0, {1, 1, 4}, {false, false, false}}, {'W', 1, {256}}},
{{'E', "gemm", {"O", "O", "I"}, {"T", "T", "N"}}, {-1, -1, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {1, 1, 1}, "xyz"}, "sb32 sb32 sb l4 cab1 wg 4x4 cs di", {8, (LoopType) 0, 128, {(LoopType) 0, (LoopType) 1, (LoopType) 255}, {8192, 8192, 2048}, {8192, 8192, 1024}, {16, 16, 32}, {4, 4, 1}, 1, (WGType) 1, true, false, false, 4096, 0, {1, 1, 4}, {false, false, false}}, {'W', 1, {256}}},
{{'E', "gemm", {"S", "S", "S"}, {"A", "B", "N"}}, {-1, -1, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {1, 1, 1}, "xyz"}, "sb2x2 sb4x2 sb ca1x2 wg 2x8 cs di hi bk1536/1536", {8, (LoopType) 0, 128, {(LoopType) 144, (LoopType) 255, (LoopType) 255}, {524288, 262144, 1536}, {8192, 8192, 1536}, {32, 16, 16}, {2, 8, 1}, 1, (WGType) 1, false, false, false, 2048, 0, {128, 64, 4}, {false, false, false}}, {'W', 1, {512}}},
Expand Down

0 comments on commit c8adc17

Please sign in to comment.