1# Copyright 2019 Google LLC 2# 3# This source code is licensed under the BSD-style license found in the 4# LICENSE file in the root directory of this source tree. 5 6# AArch64 assembly 7- name: xnn_f16_gemm_minmax_ukernel_1x8__aarch64_neonfp16arith_ld64 8 init: xnn_init_f16_minmax_neon_params 9 k-block: 4 10 assembly: true 11- name: xnn_f16_gemm_minmax_ukernel_1x16__aarch64_neonfp16arith_ld32 12 init: xnn_init_f16_minmax_neon_params 13 k-block: 2 14 assembly: true 15- name: xnn_f16_gemm_minmax_ukernel_1x16__aarch64_neonfp16arith_ld64 16 init: xnn_init_f16_minmax_neon_params 17 k-block: 4 18 assembly: true 19- name: xnn_f16_gemm_minmax_ukernel_4x8__aarch64_neonfp16arith_ld64 20 init: xnn_init_f16_minmax_neon_params 21 k-block: 4 22 assembly: true 23- name: xnn_f16_gemm_minmax_ukernel_4x16__aarch64_neonfp16arith_ld32 24 init: xnn_init_f16_minmax_neon_params 25 k-block: 2 26 assembly: true 27- name: xnn_f16_gemm_minmax_ukernel_4x16__aarch64_neonfp16arith_ld64 28 init: xnn_init_f16_minmax_neon_params 29 k-block: 4 30 assembly: true 31- name: xnn_f16_gemm_minmax_ukernel_6x8__aarch64_neonfp16arith_ld64 32 init: xnn_init_f16_minmax_neon_params 33 k-block: 4 34 assembly: true 35- name: xnn_f16_gemm_minmax_ukernel_6x16__aarch64_neonfp16arith_cortex_a55 36 init: xnn_init_f16_minmax_neon_params 37 k-block: 4 38 assembly: true 39- name: xnn_f16_gemm_minmax_ukernel_6x16__aarch64_neonfp16arith_cortex_a55r0 40 init: xnn_init_f16_minmax_neon_params 41 k-block: 4 42 assembly: true 43- name: xnn_f16_gemm_minmax_ukernel_6x16__aarch64_neonfp16arith_cortex_a75 44 init: xnn_init_f16_minmax_neon_params 45 k-block: 4 46 assembly: true 47- name: xnn_f16_gemm_minmax_ukernel_6x16__aarch64_neonfp16arith_ld32 48 init: xnn_init_f16_minmax_neon_params 49 k-block: 2 50 assembly: true 51- name: xnn_f16_gemm_minmax_ukernel_6x16__aarch64_neonfp16arith_ld64 52 init: xnn_init_f16_minmax_neon_params 53 k-block: 4 54 assembly: true 55- name: xnn_f16_gemm_minmax_ukernel_8x8__aarch64_neonfp16arith_ld64 56 init: xnn_init_f16_minmax_neon_params 57 k-block: 4 58 assembly: true 59 60# ARM NEON+FP16ARITH 61- name: xnn_f16_gemm_minmax_ukernel_1x8__neonfp16arith_ld64 62 init: xnn_init_f16_minmax_neon_params 63 k-block: 4 64- name: xnn_f16_gemm_minmax_ukernel_1x16__neonfp16arith_ld64 65 init: xnn_init_f16_minmax_neon_params 66 k-block: 4 67- name: xnn_f16_gemm_minmax_ukernel_4x8__neonfp16arith_ld64 68 init: xnn_init_f16_minmax_neon_params 69 k-block: 4 70- name: xnn_f16_gemm_minmax_ukernel_4x16__neonfp16arith_ld64 71 init: xnn_init_f16_minmax_neon_params 72 k-block: 4 73- name: xnn_f16_gemm_minmax_ukernel_6x8__neonfp16arith_ld64 74 init: xnn_init_f16_minmax_neon_params 75 k-block: 4 76- name: xnn_f16_gemm_minmax_ukernel_6x16__neonfp16arith_ld64 77 init: xnn_init_f16_minmax_neon_params 78 k-block: 4 79- name: xnn_f16_gemm_minmax_ukernel_8x8__neonfp16arith_ld64 80 init: xnn_init_f16_minmax_neon_params 81 k-block: 4 82- name: xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64 83 init: xnn_init_f16_minmax_neon_params 84 k-block: 4 85 86# x86 AVX2 87- name: xnn_f16_gemm_minmax_ukernel_1x8__avx2_broadcast 88 init: xnn_init_f16_minmax_avx_params 89 k-block: 1 90- name: xnn_f16_gemm_minmax_ukernel_1x16__avx2_broadcast 91 init: xnn_init_f16_minmax_avx_params 92 k-block: 1 93- name: xnn_f16_gemm_minmax_ukernel_3x16__avx2_broadcast 94 init: xnn_init_f16_minmax_avx_params 95 k-block: 1 96- name: xnn_f16_gemm_minmax_ukernel_4x8__avx2_broadcast 97 init: xnn_init_f16_minmax_avx_params 98 k-block: 1 99- name: xnn_f16_gemm_minmax_ukernel_4x16__avx2_broadcast 100 init: xnn_init_f16_minmax_avx_params 101 k-block: 1 102- name: xnn_f16_gemm_minmax_ukernel_5x8__avx2_broadcast 103 init: xnn_init_f16_minmax_avx_params 104 k-block: 1 105- name: xnn_f16_gemm_minmax_ukernel_5x16__avx2_broadcast 106 init: xnn_init_f16_minmax_avx_params 107 k-block: 1 108- name: xnn_f16_gemm_minmax_ukernel_6x8__avx2_broadcast 109 init: xnn_init_f16_minmax_avx_params 110 k-block: 1 111- name: xnn_f16_gemm_minmax_ukernel_7x8__avx2_broadcast 112 init: xnn_init_f16_minmax_avx_params 113 k-block: 1 114