xehpg: jit: gemm: large-tile TN int8 compressed weights kernel

Author: petercad

src/gpu/intel/jit/gemm/kernel.db

@@ -15,8 +15,8 @@
 *******************************************************************************/
 
 /*@kcatalog@*/
-kcatalog::FlatCatalog<1108> _CATALOG_
-{1, 8310, 1108, {
+kcatalog::FlatCatalog<1109> _CATALOG_
+{1, 8311, 1109, {
 {{'9', "gemm", {"B", "B", "S"}, {"N", "N", "N"}}, {-1, -1, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {1, 1, 1}, ""}, "ab2x2 as8x2 ab l4 acb", {8, (LoopType) 255, 128, {(LoopType) 0, (LoopType) 1, (LoopType) 255}, {4096, 4096, 2048}, {4096, 4096, 2048}, {32, 16, 16}, {2, 8, 1}, 1, (WGType) 0, 1, 0, 0, {2, 2, 4}, {true, true, true}}, {'W', 1, {512}}},
 {{'9', "gemm", {"B", "B", "S"}, {"N", "T", "N"}}, {-1, -1, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {1, 1, 1}, ""}, "ab4 ab4x2 ab l4 acb nmk", {8, (LoopType) 255, 128, {(LoopType) 1, (LoopType) 0, (LoopType) 255}, {4096, 4096, 2048}, {4096, 4096, 2048}, {32, 16, 8}, {8, 2, 1}, 1, (WGType) 0, 1, 0, 0, {2, 2, 4}, {true, true, true}}, {'W', 1, {512}}},
 {{'9', "gemm", {"B", "B", "S"}, {"T", "N", "N"}}, {-1, -1, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {1, 1, 1}, ""}, "as8 as16 ab l4 acb", {8, (LoopType) 255, 128, {(LoopType) 0, (LoopType) 1, (LoopType) 255}, {4096, 4096, 1024}, {4096, 4096, 1024}, {16, 16, 16}, {2, 8, 1}, 1, (WGType) 0, 1, 0, 0, {2, 2, 4}, {true, true, true}}, {'W', 1, {256}}},
@@ -399,6 +399,7 @@ kcatalog::FlatCatalog<1108> _CATALOG_
 {{'E', "gemm", {"O", "H", "S"}, {"T", "N", "N"}}, {-1, -1, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {1, 1, 1}, "Ixy"}, "sB32 sB16 aB wg 2x4x4 kr cab4 ks32 af dw vav bo bk0 sm sn grf256 sys pab l4 sr", {8, (LoopType) 0, 256, {(LoopType) 128, (LoopType) 255, (LoopType) 2}, {262144, 131072, 16777216}, {8192, 8192, 16777216}, {16, 8, 32}, {2, 4, 4}, 1, (WGType) 1, 261, 65536, 16384, {2, 2, 4}, {false, false, true}}, {'E', 17, {1.00833e+06, 472991, 67249.4, 65461, 0, 0, 5.08333, 5.12749, 4.89837, 12.0468, 0.0787666, 0.065148, 0.0481217, 1, 1.20614, 1.20116, 2.71563e-15}}},
 {{'E', "gemm", {"O", "H", "S"}, {"T", "N", "N"}}, {-1, -1, {-1, -1, -1}, {-1, 16, -1}, {-1, -1, -1}, {-1, 16, -1}, {1, 1, 1}, "xyI"}, "sB64 sS16 aS wg 2x1x16 ikr af vav sr bk0 bm0 sys pab grf256 l4", {8, (LoopType) 0, 256, {(LoopType) 0, (LoopType) 1, (LoopType) 2}, {16777216, 8192, 16777216}, {8192, 8192, 16777216}, {8, 16, 64}, {2, 1, 16}, 1, (WGType) 0, 4357, 0, 1024, {1, 2, 4}, {false, false, true}}, {'W', 1, {128}}},
 {{'E', "gemm", {"O", "H", "S"}, {"T", "N", "N"}}, {-1, -1, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {1, 1, 1}, "ixy"}, "sB4 sB4 aB wg 4x8 kc4 cab4 ks8 nse bo sr bk0 sm sn l4 pab", {8, (LoopType) 0, 128, {(LoopType) 128, (LoopType) 255, (LoopType) 255}, {524288, 262144, 16777216}, {524288, 262144, 16777216}, {32, 16, 8}, {4, 8, 1}, 1, (WGType) 1, 257, 32768, 0, {1, 2, 4}, {false, false, true}}, {'W', 1, {512}}},
+{{'E', "gemm", {"O", "H", "S"}, {"T", "N", "N"}}, {-1, -1, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {4, 4, 1}, "xyI"}, "sS16x2 sB16 aB wg 16x2 cb4 ks16 xaf dw vav bo sr bk0 sn dm grf256 sys pab", {8, (LoopType) 0, 256, {(LoopType) 128, (LoopType) 255, (LoopType) 255}, {262144, 1048576, 16777216}, {262144, 1048576, 16777216}, {16, 64, 16}, {16, 2, 1}, 1, (WGType) 1, 257, 16384, 0, {4, 4, 4}, {false, false, true}}, {'E', 17, {1.02157e+06, 527328, 0, 0, 0, 0, 3.47453, 5.53847, 6.17805, 17.0853, 0.0227172, 0.00268985, 0.022493, 0.935291, 1.301, 1.18662, 5.43286e-13}}},
 {{'E', "gemm", {"O", "O", "I"}, {"A4#8,32", "B32", "N"}}, {-1, -1, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {1, 1, 1}, "xyzdsm"}, "sB16 sB16 sb fs wg 4x4 bo acb bk8192 sr", {8, (LoopType) 0, 256, {(LoopType) 128, (LoopType) 255, (LoopType) 255}, {524288, 786432, 8192}, {8192, 8192, 8192}, {32, 48, 32}, {4, 4, 1}, 1, (WGType) 1, 256, 32256, 0, {128, 128, 4}, {false, false, false}}, {'W', 1, {1536}}},
 {{'E', "gemm", {"O", "O", "I"}, {"A4#8,32", "B32", "N"}}, {-1, -1, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {1, 1, 1}, "xyzDsm"}, "sB16 sB16 sb fs wg 8x4 bo acb bk8192 sr", {8, (LoopType) 0, 256, {(LoopType) 128, (LoopType) 255, (LoopType) 255}, {524288, 786432, 8192}, {8192, 8192, 8192}, {32, 48, 32}, {8, 4, 1}, 1, (WGType) 1, 256, 61440, 0, {128, 128, 4}, {false, false, false}}, {'W', 1, {1536}}},
 {{'E', "gemm", {"O", "O", "I"}, {"A4#8,32", "B32", "N"}}, {-1, -1, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {1, 1, 1}, "sm"}, "ab16x3 ab16x3 ab fs sc bo acb bk4096 sr", {8, (LoopType) 0, 256, {(LoopType) 128, (LoopType) 255, (LoopType) 255}, {524288, 524288, 4096}, {8192, 8192, 4096}, {32, 32, 32}, {4, 4, 1}, 2, (WGType) 1, 256, 43008, 0, {128, 128, 4}, {true, true, true}}, {'W', 1, {1024}}},
@@ -1124,5 +1125,5 @@ kcatalog::FlatCatalog<1108> _CATALOG_
 {{'F', "gemm", {"[SB]", "[SB]", "S"}, {"T", "N", "N"}}, {-1, -1, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {1, 1, 1}, "I"}, "aS16+S32@48 aB32/16x2 aB wg 8x4 cb4x2 ks32 af vav di hi pt bk0 sm sn grf256 sys dm sr br", {16, (LoopType) 255, 256, {(LoopType) 208, (LoopType) 255, (LoopType) 255}, {262144, 262144, 16777216}, {8192, 8192, 16777216}, {16, 16, 32}, {8, 4, 1}, 1, (WGType) 1, 257, 16384, 0, {4, 4, 4}, {true, true, true}}, {'E', 17, {964729, 752109, 0, 0, 0, 0, 1.46062, 1.46554, 1.01724, 2.17327, 0.0239322, 0.0239322, 0, 1, 1.25665, 0.945813, 3.00962e-12}}},
 {{'F', "gemm", {"[SB]", "[SB]", "S"}, {"T", "N", "N"}}, {-1, -1, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {1, 1, 1}, "I"}, "aS16+S64@64 aB16x2 aB wg 4x8 cb4 ks64 af vav di hi pt bk0 sm sn grf256 sys dm sr br", {16, (LoopType) 255, 256, {(LoopType) 208, (LoopType) 255, (LoopType) 255}, {262144, 131072, 16777216}, {8192, 8192, 16777216}, {16, 8, 64}, {4, 8, 1}, 1, (WGType) 1, 257, 32768, 0, {4, 4, 4}, {true, true, true}}, {'E', 17, {1.00341e+06, 675337, 0, 0, 0, 0, 2.06693, 1.45219, 1.47193, 3.06315, 0.0374548, 0.0374548, 0, 1, 1.33339, 0.377183, 1.26633e-11}}},
 {{'G', "gemm", {"O", "H", "S"}, {"T", "N", "N"}}, {-1, -1, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {-1, 8, -1}, {16, 16, 1}, "IAB"}, "at32x2+m64@16 am64x2 aB wg 4x1x4 ikr af vav ar sb64 bm0 bk0 sys nmk np", {16, (LoopType) 255, 128, {(LoopType) 0, (LoopType) 1, (LoopType) 2}, {8192, 8192, 16777216}, {8192, 8192, 16777216}, {16, 8, 64}, {4, 1, 4}, 1, (WGType) 1, 261, 0, 2048, {16, 16, 4}, {true, true, true}}, {'W', 1, {128}}},
-{{'G', "gemm", {"O", "H", "S"}, {"T", "N", "N"}}, {-1, -1, {-1, -1, -1}, {-1, -1, -1}, {-1, 9, -1}, {-1, 64, -1}, {16, 16, 1}, "IAB"}, "at32x2+m64@16 am64x2 aB wg 4x1x4 ikr af vav ar sb64 bm0 bk0 sys nmk np", {16, (LoopType) 255, 128, {(LoopType) 0, (LoopType) 1, (LoopType) 2}, {8192, 8192, 16777216}, {8192, 8192, 16777216}, {16, 16, 64}, {4, 1, 4}, 1, (WGType) 1, 261, 0, 2048, {16, 16, 4}, {true, true, true}}, {'W', 1, {128}}},
+{{'G', "gemm", {"O", "H", "S"}, {"T", "N", "N"}}, {-1, -1, {-1, -1, -1}, {-1, -1, -1}, {-1, 9, -1}, {-1, 64, -1}, {16, 16, 1}, "IAB"}, "at32x2+m64@16 am64x2 aB wg 4x1x4 ikr af vav ar sb64 bm0 bk0 sys nmk np", {16, (LoopType) 255, 128, {(LoopType) 0, (LoopType) 1, (LoopType) 2}, {8192, 8192, 16777216}, {8192, 8192, 16777216}, {16, 16, 64}, {4, 1, 4}, 1, (WGType) 1, 261, 0, 2048, {16, 16, 4}, {true, true, true}}, {'W', 1, {128}}}
 }}