diff --git a/src/gpu/intel/gemm/jit/selector/db/kernel.db b/src/gpu/intel/gemm/jit/selector/db/kernel.db index 2768284cdc..a5e9c9b224 100644 --- a/src/gpu/intel/gemm/jit/selector/db/kernel.db +++ b/src/gpu/intel/gemm/jit/selector/db/kernel.db @@ -644,6 +644,7 @@ auto _CATALOG_ = kcatalog::toArray({ {{'F', "gemm", {"F", "H", "S"}, {"T", "N", "N"}}, {-1, -1, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {8, 8, 1}, "IABs"}, "at32+m32@64 am32+m32@64 aB wg 8x4 xaf st vav hi pt sr br sb64 bk0 sm sn grf256 sys kv afb", {16, (LoopType) 255, 256, {(LoopType) 208, (LoopType) 255, (LoopType) 255}, {524288, 917504, 16777216}, {524288, 917504, 32}, {32, 56, 32}, {8, 4, 1}, 1, (WGType) 1, 441, 0, 0, {8, 8, 4}, {true, true, true}}, {'E', 17, {871124, 600159, 0, 0, 5.72375e+06, 8.32307e+06, 0.555762, 0.650262, 0.789585, 1.21694, 0.00411938, 0.00411938, 0, 1, 1.69975, 1.10825, 2.61032e-12}}}, {{'F', "gemm", {"F", "H", "S"}, {"T", "N", "N"}}, {-1, -1, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {8, 8, 1}, "IABs"}, "at64x2+m64@64 am32+m32@64 aB wg 8x4 xaf fx vav hi pt sr br sb64 bk0 sm sn grf256 sys kv afb", {16, (LoopType) 255, 256, {(LoopType) 208, (LoopType) 255, (LoopType) 255}, {524288, 786432, 16777216}, {524288, 786432, 64}, {32, 48, 64}, {8, 4, 1}, 1, (WGType) 1, 441, 0, 0, {8, 8, 4}, {true, true, true}}, {'E', 17, {866537, 568292, 0, 0, 4.29425e+06, 7.31546e+06, 0.475793, 0.638791, 0.753591, 1.18019, 0.0041899, 0.0041899, 0, 1, 1.78867, 1.11209, 1.76449e-12}}}, {{'F', "gemm", {"F", "H", "S"}, {"T", "N", "N"}}, {-1, -1, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {8, 8, 1}, "ABI"}, "at64x2+m64@64 am32+m32@64 aB wg 8x4 xaf fx vav hi pt sr br sb64 bk0 sm sn grf256 sys kv afb", {16, (LoopType) 255, 256, {(LoopType) 208, (LoopType) 255, (LoopType) 255}, {524288, 524288, 16777216}, {524288, 524288, 64}, {32, 32, 64}, {8, 4, 1}, 1, (WGType) 1, 441, 0, 0, {8, 8, 4}, {true, true, true}}, {'E', 17, {1.00534e+06, 516328, 0, 0, 4.6121e+06, 5.65248e+06, 0.524475, 0.995728, 0.862826, 1.32399, 0.00510252, 0.00510252, 0, 1, 1.47564, 1.09581, 4.42262e-12}}}, +{{'F', "gemm", {"F", "H", "S"}, {"T", "N", "N"}}, {-1, -1, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {4, 8, 1}, "ABI"}, "at16x2+m32@48 am32+m16@64 aB wg 4x2x4 kr xaf st vav hi pt sr br sb64 bk0 sm sn grf256 sys kv afb", {16, (LoopType) 255, 256, {(LoopType) 208, (LoopType) 255, (LoopType) 2}, {262144, 262144, 16777216}, {262144, 262144, 32}, {16, 16, 32}, {4, 2, 4}, 1, (WGType) 1, 445, 0, 8192, {4, 8, 4}, {true, true, true}}, {'E', 17, {1.18986e+06, -136092, -44118.4, 227346, 2.90816e+06, 1.96608e+06, 0.463277, 0.703205, 0.870747, 1.51625, 0.0198219, 0.0158398, 0.00608028, 0.596299, 1.37888, 0.920771, 9.99051e-12}}}, {{'F', "gemm", {"F", "S", "S"}, {"N", "N", "N"}}, {-1, -1, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {1, 1, 1}, ""}, "aB8x2+B8@16 aS8+S8@24 aB wg 1x4 kc8 nse hi pt sr sb256 bk0 sn", {16, (LoopType) 255, 128, {(LoopType) 208, (LoopType) 255, (LoopType) 255}, {524288, 131072, 16777216}, {524288, 131072, 16777216}, {16, 8, 8}, {1, 4, 1}, 1, (WGType) 1, 257, 0, 0, {1, 4, 4}, {true, true, true}}, {'E', 17, {1.16538e+06, 40635.2, 0, 0, 0, 0, 1.30731, 1.53858, 0.584971, 1.42067, 0.0634061, 0.0581975, 0.0161667, 1, 1.44276, 1.00478, 2.34818e-11}}}, {{'F', "gemm", {"F", "S", "S"}, {"N", "N", "N"}}, {-1, -1, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {1, 1, 1}, ""}, "aB8x2+B16@8 aS8x2+S32@8 aB wg 2x8x2 kr kc8 nse hi pt sr kv sb256 bk0 sn grf256 afb", {16, (LoopType) 255, 256, {(LoopType) 208, (LoopType) 255, (LoopType) 2}, {1048576, 262144, 16777216}, {8192, 8192, 0}, {32, 16, 8}, {2, 8, 2}, 1, (WGType) 1, 413, 0, 65536, {4, 4, 4}, {true, true, true}}, {'E', 17, {1.12483e+06, -574330, -20527.6, 853244, 4.03046e+06, 0, 2.33007, 1.51079, 0.80632, 1.51643, 0.0625462, -1.54388e-05, 0.0629715, 0.425906, 1.06028, 0.229788, 1.06051e-11}}}, {{'F', "gemm", {"F", "S", "S"}, {"N", "N", "N"}}, {-1, -1, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {-1, 4, -1}, {1, 1, 1}, ""}, "am8+B8@8 at8x2 aB wg 4x1x16 kr kc8 nse li nmk pt sr bk0 sn kv afb sb32 l2d", {16, (LoopType) 255, 128, {(LoopType) 225, (LoopType) 255, (LoopType) 2}, {524288, 65536, 16777216}, {8192, 8192, 0}, {16, 4, 8}, {4, 1, 16}, 1, (WGType) 1, 413, 0, 2048, {4, 4, 4}, {true, true, true}}, {'E', 17, {1.21115e+06, -59718.2, -11600.6, 104303, 3.06381e+06, 0, 0.860506, 9.14342, 0.769527, 1.14843, 0.0733058, 0.0350639, 0.04512, 0.901895, 1.307, 0.986093, 2.06541e-11}}},