| # Copyright 2021 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_fp32_ukernel_1x8c2__neon_mlal_ld1r |
| init: xnn_init_qs8_conv_minmax_fp32_neon_params |
| k-block: 16 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_2x8c2__neon_mlal_ld1r |
| init: xnn_init_qs8_conv_minmax_fp32_neon_params |
| k-block: 16 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld1r |
| init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| k-block: 16 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_2x8c2__neonv8_mlal_ld1r |
| init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| k-block: 16 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld2r |
| init: xnn_init_qs8_conv_minmax_fp32_neon_params |
| k-block: 16 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_2x8c2__neon_mlal_ld2r |
| init: xnn_init_qs8_conv_minmax_fp32_neon_params |
| k-block: 16 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld2r |
| init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| k-block: 16 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_2x8c2__neonv8_mlal_ld2r |
| init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| k-block: 16 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld4r |
| init: xnn_init_qs8_conv_minmax_fp32_neon_params |
| k-block: 16 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_2x8c2__neon_mlal_ld4r |
| init: xnn_init_qs8_conv_minmax_fp32_neon_params |
| k-block: 16 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld4r |
| init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| k-block: 16 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_2x8c2__neonv8_mlal_ld4r |
| init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| k-block: 16 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_1x8c4s2__neon_mlal |
| init: xnn_init_qs8_conv_minmax_fp32_neon_params |
| k-block: 16 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4s2__neon_mlal |
| init: xnn_init_qs8_conv_minmax_fp32_neon_params |
| k-block: 16 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_1x8c4s2__neonv8_mlal |
| init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| k-block: 16 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4s2__neonv8_mlal |
| init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| k-block: 16 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2s4__neon_mlal |
| init: xnn_init_qs8_conv_minmax_fp32_neon_params |
| k-block: 16 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_2x8c2s4__neon_mlal |
| init: xnn_init_qs8_conv_minmax_fp32_neon_params |
| k-block: 16 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2s4__neonv8_mlal |
| init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| k-block: 16 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_2x8c2s4__neonv8_mlal |
| init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| k-block: 16 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_1x8c4__neon_mlal_dup |
| init: xnn_init_qs8_conv_minmax_fp32_neon_params |
| k-block: 16 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_dup |
| init: xnn_init_qs8_conv_minmax_fp32_neon_params |
| k-block: 16 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_1x8c4__neonv8_mlal_dup |
| init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| k-block: 16 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup |
| init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| k-block: 16 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_1x8c4__neon_mlal_ld1r |
| init: xnn_init_qs8_conv_minmax_fp32_neon_params |
| k-block: 16 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld1r |
| init: xnn_init_qs8_conv_minmax_fp32_neon_params |
| k-block: 16 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_1x8c4__neonv8_mlal_ld1r |
| init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| k-block: 16 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld1r |
| init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| k-block: 16 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_1x8c4__neon_mlal_ld2r |
| init: xnn_init_qs8_conv_minmax_fp32_neon_params |
| k-block: 16 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld2r |
| init: xnn_init_qs8_conv_minmax_fp32_neon_params |
| k-block: 16 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_1x8c4__neonv8_mlal_ld2r |
| init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| k-block: 16 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld2r |
| init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| k-block: 16 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_dup |
| init: xnn_init_qs8_conv_minmax_fp32_neon_params |
| k-block: 16 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_2x8c2__neon_mlal_dup |
| init: xnn_init_qs8_conv_minmax_fp32_neon_params |
| k-block: 16 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_dup |
| init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| k-block: 16 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_2x8c2__neonv8_mlal_dup |
| init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| k-block: 16 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__aarch64_neon_mlal |
| init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| k-block: 16 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__aarch64_neon_mlal_prfm |
| init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| k-block: 16 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__aarch64_neon_mull |
| init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__aarch64_neon_mlal_cortex_a53 |
| init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| k-block: 16 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__aarch64_neon_mlal_prfm_cortex_a53 |
| init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| k-block: 16 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal |
| init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| k-block: 16 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__aarch64_neon_mlal |
| init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| k-block: 16 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__aarch64_neon_mlal_prfm |
| init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| k-block: 16 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__aarch64_neon_mlal_cortex_a53 |
| init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| k-block: 16 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__aarch64_neon_mlal_prfm_cortex_a53 |
| init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| k-block: 16 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53 |
| init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a53 |
| init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_4x16__aarch64_neon_mlal_lane_prfm_ld64 |
| init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_4x16__aarch64_neon_mlal_lane_ld64 |
| init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_1x16c4__aarch64_neondot_ld32 |
| init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| k-block: 4 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_1x16c4__aarch64_neondot_ld64 |
| init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_4x16c4__aarch64_neondot_ld32 |
| init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| k-block: 4 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_4x16c4__aarch64_neondot_ld64 |
| init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_4x16c4__aarch64_neondot_ld128 |
| init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| k-block: 16 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_4x16c4__aarch64_neondot_cortex_a55 |
| init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| k-block: 16 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_1x16__neon_mlal_lane |
| init: xnn_init_qs8_conv_minmax_fp32_neon_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_4x16__neon_mlal_lane |
| init: xnn_init_qs8_conv_minmax_fp32_neon_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_1x16__neonv8_mlal_lane |
| init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_4x16__neonv8_mlal_lane |
| init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__neon_mlal |
| init: xnn_init_qs8_conv_minmax_fp32_neon_params |
| k-block: 16 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal |
| init: xnn_init_qs8_conv_minmax_fp32_neon_params |
| k-block: 16 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal |
| init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| k-block: 16 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal |
| init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| k-block: 16 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_1x8c4__neondot |
| init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_1x4c2__sse2_ld64 |
| init: xnn_init_qs8_conv_minmax_fp32_sse2_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_2x4c2__sse2_ld64 |
| init: xnn_init_qs8_conv_minmax_fp32_sse2_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2__sse2_ld64 |
| init: xnn_init_qs8_conv_minmax_fp32_sse2_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_4x4c2__sse2_ld64 |
| init: xnn_init_qs8_conv_minmax_fp32_sse2_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_1x4c2__sse41_ld64 |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_2x4c2__sse41_ld64 |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2__sse41_ld64 |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_4x4c2__sse41_ld64 |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_1x4c2__avx_ld64 |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_2x4c2__avx_ld64 |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2__avx_ld64 |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_4x4c2__avx_ld64 |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_1x4c2__xop_ld64 |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_2x4c2__xop_ld64 |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2__xop_ld64 |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_4x4c2__xop_ld64 |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_1x4c2__sse2_ld128 |
| init: xnn_init_qs8_conv_minmax_fp32_sse2_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_2x4c2__sse2_ld128 |
| init: xnn_init_qs8_conv_minmax_fp32_sse2_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2__sse2_ld128 |
| init: xnn_init_qs8_conv_minmax_fp32_sse2_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_4x4c2__sse2_ld128 |
| init: xnn_init_qs8_conv_minmax_fp32_sse2_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_1x4c2__sse41_ld128 |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_2x4c2__sse41_ld128 |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2__sse41_ld128 |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_4x4c2__sse41_ld128 |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_1x4c2__avx_ld128 |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_2x4c2__avx_ld128 |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2__avx_ld128 |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_4x4c2__avx_ld128 |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_1x4c2__xop_ld128 |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_2x4c2__xop_ld128 |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2__xop_ld128 |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_4x4c2__xop_ld128 |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_1x4c8__sse2_ld64 |
| init: xnn_init_qs8_conv_minmax_fp32_sse2_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_2x4c8__sse2_ld64 |
| init: xnn_init_qs8_conv_minmax_fp32_sse2_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_3x4c8__sse2_ld64 |
| init: xnn_init_qs8_conv_minmax_fp32_sse2_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_1x4c8__ssse3_ld64 |
| init: xnn_init_qs8_conv_minmax_fp32_sse2_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_2x4c8__ssse3_ld64 |
| init: xnn_init_qs8_conv_minmax_fp32_sse2_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_3x4c8__ssse3_ld64 |
| init: xnn_init_qs8_conv_minmax_fp32_sse2_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_1x4c8__sse41_ld64 |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_2x4c8__sse41_ld64 |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_3x4c8__sse41_ld64 |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_1x4c8__avx_ld64 |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_2x4c8__avx_ld64 |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_3x4c8__avx_ld64 |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_1x4c8__xop_ld64 |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_2x4c8__xop_ld64 |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_3x4c8__xop_ld64 |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_1x4c8__sse2_ld128 |
| init: xnn_init_qs8_conv_minmax_fp32_sse2_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_2x4c8__sse2_ld128 |
| init: xnn_init_qs8_conv_minmax_fp32_sse2_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_3x4c8__sse2_ld128 |
| init: xnn_init_qs8_conv_minmax_fp32_sse2_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_1x4c8__ssse3_ld128 |
| init: xnn_init_qs8_conv_minmax_fp32_sse2_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_2x4c8__ssse3_ld128 |
| init: xnn_init_qs8_conv_minmax_fp32_sse2_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_3x4c8__ssse3_ld128 |
| init: xnn_init_qs8_conv_minmax_fp32_sse2_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_1x4c8__sse41_ld128 |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_2x4c8__sse41_ld128 |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_3x4c8__sse41_ld128 |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_1x4c8__avx_ld128 |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_2x4c8__avx_ld128 |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_3x4c8__avx_ld128 |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_1x4c8__xop_ld128 |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_2x4c8__xop_ld128 |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_3x4c8__xop_ld128 |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_xw_minmax_fp32_ukernel_1x4c2__sse2 |
| init: xnn_init_qs8_conv_minmax_fp32_sse2_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_xw_minmax_fp32_ukernel_2x4c2__sse2 |
| init: xnn_init_qs8_conv_minmax_fp32_sse2_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c2__sse2 |
| init: xnn_init_qs8_conv_minmax_fp32_sse2_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_xw_minmax_fp32_ukernel_4x4c2__sse2 |
| init: xnn_init_qs8_conv_minmax_fp32_sse2_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_xw_minmax_fp32_ukernel_1x4c2__sse41 |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_xw_minmax_fp32_ukernel_2x4c2__sse41 |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c2__sse41 |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_xw_minmax_fp32_ukernel_4x4c2__sse41 |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_xw_minmax_fp32_ukernel_1x4c2__avx |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_xw_minmax_fp32_ukernel_2x4c2__avx |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c2__avx |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_xw_minmax_fp32_ukernel_4x4c2__avx |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_xw_minmax_fp32_ukernel_1x4c2__xop |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_xw_minmax_fp32_ukernel_2x4c2__xop |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c2__xop |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_xw_minmax_fp32_ukernel_4x4c2__xop |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_xw_minmax_fp32_ukernel_1x4c8__sse2 |
| init: xnn_init_qs8_conv_minmax_fp32_sse2_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_xw_minmax_fp32_ukernel_2x4c8__sse2 |
| init: xnn_init_qs8_conv_minmax_fp32_sse2_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c8__sse2 |
| init: xnn_init_qs8_conv_minmax_fp32_sse2_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_xw_minmax_fp32_ukernel_1x4c8__ssse3 |
| init: xnn_init_qs8_conv_minmax_fp32_sse2_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_xw_minmax_fp32_ukernel_2x4c8__ssse3 |
| init: xnn_init_qs8_conv_minmax_fp32_sse2_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c8__ssse3 |
| init: xnn_init_qs8_conv_minmax_fp32_sse2_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_xw_minmax_fp32_ukernel_1x4c8__sse41 |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_xw_minmax_fp32_ukernel_2x4c8__sse41 |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c8__sse41 |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_xw_minmax_fp32_ukernel_1x4c8__avx |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_xw_minmax_fp32_ukernel_2x4c8__avx |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c8__avx |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_xw_minmax_fp32_ukernel_1x4c8__xop |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_xw_minmax_fp32_ukernel_2x4c8__xop |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c8__xop |
| init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__avx2 |
| init: xnn_init_qs8_conv_minmax_fp32_avx2_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__avx2 |
| init: xnn_init_qs8_conv_minmax_fp32_avx2_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_3x8c8__avx2 |
| init: xnn_init_qs8_conv_minmax_fp32_avx2_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_xw_minmax_fp32_ukernel_1x8c8__avx2 |
| init: xnn_init_qs8_conv_minmax_fp32_avx2_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_xw_minmax_fp32_ukernel_2x8c8__avx2 |
| init: xnn_init_qs8_conv_minmax_fp32_avx2_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x8c8__avx2 |
| init: xnn_init_qs8_conv_minmax_fp32_avx2_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_1x16c8__avx512skx |
| init: xnn_init_qs8_conv_minmax_fp32_avx512_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_2x16c8__avx512skx |
| init: xnn_init_qs8_conv_minmax_fp32_avx512_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_3x16c8__avx512skx |
| init: xnn_init_qs8_conv_minmax_fp32_avx512_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx |
| init: xnn_init_qs8_conv_minmax_fp32_avx512_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_1x4c2__wasmsimd_dot16x2_ld64 |
| init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_2x4c2__wasmsimd_dot16x2_ld64 |
| init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2__wasmsimd_dot16x2_ld64 |
| init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_4x4c2__wasmsimd_dot16x2_ld64 |
| init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_1x4c2__wasmsimd_dot16x2_ld128 |
| init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_2x4c2__wasmsimd_dot16x2_ld128 |
| init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2__wasmsimd_dot16x2_ld128 |
| init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_4x4c2__wasmsimd_dot16x2_ld128 |
| init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_xw_minmax_fp32_ukernel_1x4c2__wasmsimd_dot16x2 |
| init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_xw_minmax_fp32_ukernel_2x4c2__wasmsimd_dot16x2 |
| init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c2__wasmsimd_dot16x2 |
| init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_xw_minmax_fp32_ukernel_4x4c2__wasmsimd_dot16x2 |
| init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_1x4c8__wasmsimd_dot16x2_ld64 |
| init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_2x4c8__wasmsimd_dot16x2_ld64 |
| init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_3x4c8__wasmsimd_dot16x2_ld64 |
| init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_4x4c8__wasmsimd_dot16x2_ld64 |
| init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_1x4c8__wasmsimd_dot16x2_ld128 |
| init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_2x4c8__wasmsimd_dot16x2_ld128 |
| init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_3x4c8__wasmsimd_dot16x2_ld128 |
| init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_4x4c8__wasmsimd_dot16x2_ld128 |
| init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_xw_minmax_fp32_ukernel_1x4c8__wasmsimd_dot16x2 |
| init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_xw_minmax_fp32_ukernel_2x4c8__wasmsimd_dot16x2 |
| init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c8__wasmsimd_dot16x2 |
| init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_xw_minmax_fp32_ukernel_4x4c8__wasmsimd_dot16x2 |
| init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_1x4c8__wasmsimd_mul16_ld64 |
| init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_2x4c8__wasmsimd_mul16_ld64 |
| init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_3x4c8__wasmsimd_mul16_ld64 |
| init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_1x4c8__wasmsimd_mul16_ld128 |
| init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_2x4c8__wasmsimd_mul16_ld128 |
| init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_3x4c8__wasmsimd_mul16_ld128 |
| init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_xw_minmax_fp32_ukernel_1x4c8__wasmsimd_mul16 |
| init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_xw_minmax_fp32_ukernel_2x4c8__wasmsimd_mul16 |
| init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c8__wasmsimd_mul16 |
| init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params |
| k-block: 8 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_lrint |
| init: xnn_init_qs8_conv_minmax_fp32_scalar_lrint_params |
| k-block: 1 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_lrint |
| init: xnn_init_qs8_conv_minmax_fp32_scalar_lrint_params |
| k-block: 1 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_lrint |
| init: xnn_init_qs8_conv_minmax_fp32_scalar_lrint_params |
| k-block: 1 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_lrint |
| init: xnn_init_qs8_conv_minmax_fp32_scalar_lrint_params |
| k-block: 1 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_lrint |
| init: xnn_init_qs8_conv_minmax_fp32_scalar_lrint_params |
| k-block: 1 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_lrint |
| init: xnn_init_qs8_conv_minmax_fp32_scalar_lrint_params |
| k-block: 1 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_lrint |
| init: xnn_init_qs8_conv_minmax_fp32_scalar_lrint_params |
| k-block: 1 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_lrint |
| init: xnn_init_qs8_conv_minmax_fp32_scalar_lrint_params |
| k-block: 1 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_magic |
| init: xnn_init_qs8_conv_minmax_fp32_scalar_magic_params |
| k-block: 1 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_magic |
| init: xnn_init_qs8_conv_minmax_fp32_scalar_magic_params |
| k-block: 1 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_magic |
| init: xnn_init_qs8_conv_minmax_fp32_scalar_magic_params |
| k-block: 1 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_magic |
| init: xnn_init_qs8_conv_minmax_fp32_scalar_magic_params |
| k-block: 1 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_magic |
| init: xnn_init_qs8_conv_minmax_fp32_scalar_magic_params |
| k-block: 1 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_magic |
| init: xnn_init_qs8_conv_minmax_fp32_scalar_magic_params |
| k-block: 1 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_magic |
| init: xnn_init_qs8_conv_minmax_fp32_scalar_magic_params |
| k-block: 1 |
| - name: xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_magic |
| init: xnn_init_qs8_conv_minmax_fp32_scalar_magic_params |
| k-block: 1 |