# Copyright 2020 Google LLC # # This source code is licensed under the BSD-style license found in the # LICENSE file in the root directory of this source tree. - name: xnn_qs8_gemm_minmax_ukernel_2x8c8__aarch64_neon_mlal_padal k-block: 16 - name: xnn_qs8_gemm_minmax_ukernel_2x8c8__aarch64_neon_mull_padal k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_1x8__neon_mlal_lane k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_4x8__neon_mlal_lane k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_1x16__neon_mlal_lane k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_2x16__neon_mlal_lane k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_3x16__neon_mlal_lane k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_1x8__neon_mull_addw_dup k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_2x8__neon_mull_addw_dup k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_3x8__neon_mull_addw_dup k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_4x8__neon_mull_addw_dup k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_1x16__neon_mull_addw_dup k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_2x16__neon_mull_addw_dup k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_3x16__neon_mull_addw_dup k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_1x8c2__neon_mull_padal_dup k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mull_padal_dup k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mull_padal_dup k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mull_padal_dup k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mull_padal_dup k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mull_padal_dup k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_3x16c2__neon_mull_padal_dup k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_4x16c2__neon_mull_padal_dup k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_1x8c2__neon_mlal_padal_dup k-block: 16 - name: xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup k-block: 16 - name: xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup k-block: 16 - name: xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup k-block: 16 - name: xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup k-block: 16 - name: xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mlal_padal_dup k-block: 16 - name: xnn_qs8_gemm_minmax_ukernel_3x16c2__neon_mlal_padal_dup k-block: 16 - name: xnn_qs8_gemm_minmax_ukernel_4x16c2__neon_mlal_padal_dup k-block: 16 - name: xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mull_padal k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mull_padal k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mlal_padal k-block: 16 - name: xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mlal_padal k-block: 16 - name: xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal k-block: 16 - name: xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal k-block: 16 - name: xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mlal_padal k-block: 16 - name: xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal k-block: 16 - name: xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal k-block: 16 - name: xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal k-block: 16 - name: xnn_qs8_gemm_minmax_ukernel_1x8c16__neon_mlal_padal k-block: 16 - name: xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal k-block: 16 - name: xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal k-block: 16 - name: xnn_qs8_gemm_minmax_ukernel_4x8c16__neon_mlal_padal k-block: 16 - name: xnn_qs8_gemm_minmax_ukernel_1x16c16__neon_mlal_padal k-block: 16 - name: xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal k-block: 16 - name: xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal k-block: 16 - name: xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal k-block: 16 - name: xnn_qs8_gemm_minmax_ukernel_1x8c4__neondot k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_4x8c4__neondot k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_6x8c4__neondot k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_8x8c4__neondot k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_1x16c4__neondot k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_2x8c16__aarch64_neon_mlal_padal k-block: 16 - name: xnn_qs8_gemm_minmax_ukernel_1x16c4__aarch64_neondot_ld32 k-block: 4 - name: xnn_qs8_gemm_minmax_ukernel_1x16c4__aarch64_neondot_ld64 k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_4x16c4__aarch64_neondot_ld32 k-block: 4 - name: xnn_qs8_gemm_minmax_ukernel_4x16c4__aarch64_neondot_ld64 k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_4x16c4__aarch64_neondot_cortex_a55 k-block: 16 - name: xnn_qs8_gemm_minmax_ukernel_1x4c2__sse2_ld64 k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_4x4c2__sse2_ld64 k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_1x4c2__ssse3_ld64 k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_4x4c2__ssse3_ld64 k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_1x4c2__sse41_ld64 k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_4x4c2__sse41_ld64 k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_1x4c2__xop_ld64 k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_4x4c2__xop_ld64 k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_1x4c2__sse2_ld128 k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_4x4c2__sse2_ld128 k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_1x4c2__ssse3_ld128 k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_4x4c2__ssse3_ld128 k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_1x4c2__sse41_ld128 k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_4x4c2__sse41_ld128 k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_1x4c2__xop_ld128 k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_4x4c2__xop_ld128 k-block: 8 - name: xnn_qs8_gemm_xw_minmax_ukernel_1x4c2__sse2 k-block: 8 - name: xnn_qs8_gemm_xw_minmax_ukernel_4x4c2__sse2 k-block: 8 - name: xnn_qs8_gemm_xw_minmax_ukernel_1x4c2__ssse3 k-block: 8 - name: xnn_qs8_gemm_xw_minmax_ukernel_4x4c2__ssse3 k-block: 8 - name: xnn_qs8_gemm_xw_minmax_ukernel_1x4c2__sse41 k-block: 8 - name: xnn_qs8_gemm_xw_minmax_ukernel_4x4c2__sse41 k-block: 8 - name: xnn_qs8_gemm_xw_minmax_ukernel_1x4c2__xop k-block: 8 - name: xnn_qs8_gemm_xw_minmax_ukernel_4x4c2__xop k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_1x4c8__sse2_ld64 k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_2x4c8__sse2_ld64 k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_3x4c8__sse2_ld64 k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_1x4c8__ssse3_ld64 k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_2x4c8__ssse3_ld64 k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_3x4c8__ssse3_ld64 k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_1x4c8__sse41_ld64 k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_2x4c8__sse41_ld64 k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_3x4c8__sse41_ld64 k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_1x4c8__xop_ld64 k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_2x4c8__xop_ld64 k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_3x4c8__xop_ld64 k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_1x4c8__sse2_ld128 k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_2x4c8__sse2_ld128 k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_3x4c8__sse2_ld128 k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_1x4c8__ssse3_ld128 k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_2x4c8__ssse3_ld128 k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_3x4c8__ssse3_ld128 k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_1x4c8__sse41_ld128 k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_2x4c8__sse41_ld128 k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_3x4c8__sse41_ld128 k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_1x4c8__xop_ld128 k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_2x4c8__xop_ld128 k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_3x4c8__xop_ld128 k-block: 8 - name: xnn_qs8_gemm_xw_minmax_ukernel_1x4c8__sse2 k-block: 8 - name: xnn_qs8_gemm_xw_minmax_ukernel_2x4c8__sse2 k-block: 8 - name: xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__sse2 k-block: 8 - name: xnn_qs8_gemm_xw_minmax_ukernel_1x4c8__ssse3 k-block: 8 - name: xnn_qs8_gemm_xw_minmax_ukernel_2x4c8__ssse3 k-block: 8 - name: xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__ssse3 k-block: 8 - name: xnn_qs8_gemm_xw_minmax_ukernel_1x4c8__sse41 k-block: 8 - name: xnn_qs8_gemm_xw_minmax_ukernel_2x4c8__sse41 k-block: 8 - name: xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__sse41 k-block: 8 - name: xnn_qs8_gemm_xw_minmax_ukernel_1x4c8__xop k-block: 8 - name: xnn_qs8_gemm_xw_minmax_ukernel_2x4c8__xop k-block: 8 - name: xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__xop k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_1x8c8__avx2 k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_2x8c8__avx2 k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_3x8c8__avx2 k-block: 8 - name: xnn_qs8_gemm_xw_minmax_ukernel_1x8c8__avx2 k-block: 8 - name: xnn_qs8_gemm_xw_minmax_ukernel_2x8c8__avx2 k-block: 8 - name: xnn_qs8_gemm_xw_minmax_ukernel_3x8c8__avx2 k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_1x16c8__avx512skx k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_2x16c8__avx512skx k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_3x16c8__avx512skx k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_4x16c8__avx512skx k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_1x4c8__wasmsimd_ld64 k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_2x4c8__wasmsimd_ld64 k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_3x4c8__wasmsimd_ld64 k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_1x4c8__wasmsimd_ld128 k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_2x4c8__wasmsimd_ld128 k-block: 8 - name: xnn_qs8_gemm_minmax_ukernel_3x4c8__wasmsimd_ld128 k-block: 8 - name: xnn_qs8_gemm_xw_minmax_ukernel_1x4c8__wasmsimd k-block: 8 - name: xnn_qs8_gemm_xw_minmax_ukernel_2x4c8__wasmsimd k-block: 8 - name: xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__wasmsimd k-block: 8