xe: gemm: add Xe2 HHS TNN strategy for gemv (n=1)

Author: skazakov1

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

@@ -1206,6 +1206,7 @@ auto _CATALOG_ = kcatalog::toArray({
 {{'G', "gemm", {"H", "H", "S"}, {"T", "N", "N"}}, {-1, -1, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {-1, 128, -1}, {16, 16, 1}, "ABI"}, "at32+m16@48 am16x2+m32@32 aB wg 2x4x4 kr xaf st vav hi pt sr br sb64 bk0 sm sn grf256 sys", {16, (LoopType) 255, 256, {(LoopType) 208, (LoopType) 255, (LoopType) 2}, {262144, 131072, 16777216}, {262144, 131072, 16777216}, {16, 8, 32}, {2, 4, 4}, 1, (WGType) 1, 261, 0, 4096, {16, 16, 4}, {true, true, true}}, {'E', 17, {262968, 34374.8, 2055.74, 433.259, 0, 0, 2.05239, 1.96914, 5.76436, 13.1671, 0.0309952, 0.0309952, 0, 1, 1.16811, 1.00132, 7.47888e-13}}},
 {{'G', "gemm", {"H", "H", "S"}, {"T", "N", "N"}}, {-1, -1, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {-1, 128, -1}, {16, 16, 1}, "ABI"}, "am32+C32@64 at32 aS wg 1x1x16 ikr af vav sr sb256 bk0 bm0 sys rr", {16, (LoopType) 255, 128, {(LoopType) 0, (LoopType) 1, (LoopType) 2}, {16777216, 8192, 16777216}, {8192, 8192, 16777216}, {4, 16, 32}, {1, 1, 16}, 1, (WGType) 1, 4357, 0, 256, {16, 16, 4}, {true, true, true}}, {'E', 17, {-164022, 454944, 5845.92, 1045.76, 0, 0, 0.771733, 1.70851, 273.177, 272.825, 0.111732, 0.0896612, 0.00579836, 0.491637, 1.14958, 0.953248, 1.54584e-12}}},
 {{'G', "gemm", {"H", "H", "S"}, {"T", "N", "N"}}, {-1, -1, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {-1, 128, -1}, {16, 16, 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, {16, 16, 4}, {true, true, true}}, {'E', 17, {241719, 28050.5, 9262.5, 2137.5, 552672, 876888, 1.91538, 1.82334, 3.84174, 8.3282, 0.015781, 0.00988075, 0.00545694, 0.959871, 1.33252, 0.965141, 5.50087e-12}}},
+{{'G', "gemm", {"H", "H", "S"}, {"T", "N", "N"}}, {-1, -1, {-1, 1, 3968}, {-1, 1, 4096}, {-1, 1, 3968}, {-1, 1, 4096}, {16, 16, 1}, "ABIH"}, "am32+C32@64 at32 aS wg 1x1x16 ikr af vav sr sb256 bk0 bm0 sys rr", {16, (LoopType) 255, 128, {(LoopType) 0, (LoopType) 1, (LoopType) 2}, {16777216, 8192, 16777216}, {8192, 8192, 16777216}, {4, 16, 32}, {1, 1, 16}, 1, (WGType) 1, 268439813, 0, 256, {16, 16, 4}, {true, true, true}}, {'W', 1, {64}}},
 {{'G', "gemm", {"N", "H", "S"}, {"N", "N", "N"}}, {-1, -1, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {16, 16, 1}, "ABI"}, "am32+m64@8 am32+m16@64 aB wg 4x8 ca3 ks32 nb 4x0 sys xaf rr vav bo ar br sn dm grf256", {16, (LoopType) 255, 256, {(LoopType) 128, (LoopType) 255, (LoopType) 255}, {1048576, 524288, 16777216}, {1048576, 524288, 16777216}, {64, 32, 32}, {4, 8, 1}, 1, (WGType) 1, 257, 49152, 0, {16, 16, 4}, {true, true, true}}, {'W', 1, {2048}}},
 {{'G', "gemm", {"N", "H", "S"}, {"T", "N", "N"}}, {-1, -1, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {-1, -1, -1}, {16, 16, 1}, "ABI"}, "at16x2+m32@40 am16+m16@128 aB wg 32x1 sys xaf rr fx vav bo ar br sm sn dm grf256", {16, (LoopType) 255, 256, {(LoopType) 128, (LoopType) 255, (LoopType) 255}, {262144, 2097152, 16777216}, {262144, 2097152, 16777216}, {16, 128, 16}, {32, 1, 1}, 1, (WGType) 1, 257, 0, 0, {16, 16, 4}, {true, true, true}}, {'W', 1, {2048}}},
 {{'G', "gemm", {"N", "H", "S"}, {"T", "N", "N"}}, {-1, -1, {-1, -1, -1}, {-1, 8, -1}, {-1, -1, -1}, {-1, 8, -1}, {16, 16, 1}, "IAB"}, "at32 am128 aB wg 2x1x8 ikr xaf st vav hi pt sr br sb128 bk0 bm0 nmk sys", {16, (LoopType) 255, 128, {(LoopType) 209, (LoopType) 255, (LoopType) 2}, {16777216, 131072, 16777216}, {262144, 131072, 16777216}, {16, 8, 128}, {2, 1, 8}, 1, (WGType) 0, 4357, 0, 1024, {16, 16, 4}, {true, true, true}}, {'W', 1, {128}}},