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$assert CHANNEL_TILE % 4 == 0 7$assert KERNEL_TILE >= 2 8$assert ACCUMULATORS >= 1 9$ABC = "0123456789ABCDEF" 10#include <assert.h> 11 12#include <psimd.h> 13 14#include <xnnpack/dwconv.h> 15 16 17void xnn_f32_dwconv_ukernel_up${CHANNEL_TILE}x${KERNEL_TILE}__psimd${"" if ACCUMULATORS == 1 else "_acc%d" % ACCUMULATORS}( 18 size_t channels, 19 size_t output_width, 20 const float** input, 21 const float* weights, 22 float* output, 23 size_t input_stride, 24 size_t output_increment, 25 const union xnn_f32_output_params params[restrict static 1]) 26{ 27 assert(channels != 0); 28 assert(output_width != 0); 29 30 const psimd_f32 vmax = psimd_load_splat_f32(¶ms->scalar.max); 31 const psimd_f32 vmin = psimd_load_splat_f32(¶ms->scalar.min); 32 do { 33 $for K in range(KERNEL_TILE): 34 const float* i${K} = input[${K}]; 35 assert(i${K} != NULL); 36 input = (const float**) ((uintptr_t) input + input_stride); 37 38 size_t c = channels; 39 const float* w = weights; 40 for (; c >= ${CHANNEL_TILE}; c -= ${CHANNEL_TILE}) { 41 psimd_f32 vacc${ABC[0:4]}p0 = psimd_load_f32(w); 42 $for C in range(4, CHANNEL_TILE, 4): 43 psimd_f32 vacc${ABC[C:C+4]}p0 = psimd_load_f32(w + ${C}); 44 45 $for K in range(KERNEL_TILE): 46 47 const psimd_f32 vi${K}x${ABC[0:4]} = psimd_load_f32(i${K}); 48 $for C in range(4, CHANNEL_TILE, 4): 49 const psimd_f32 vi${K}x${ABC[C:C+4]} = psimd_load_f32(i${K} + ${C}); 50 i${K} += ${CHANNEL_TILE}; 51 52 $for C in range(0, CHANNEL_TILE, 4): 53 const psimd_f32 vk${K}x${ABC[C:C+4]} = psimd_load_f32(w + ${(K + 1) * CHANNEL_TILE + C}); 54 $for C in range(0, CHANNEL_TILE, 4): 55 $if 1 <= K < ACCUMULATORS: 56 psimd_f32 vacc${ABC[C:C+4]}p${K} = psimd_mul_f32(vi${K}x${ABC[C:C+4]}, vk${K}x${ABC[C:C+4]}); 57 $else: 58 vacc${ABC[C:C+4]}p${K % ACCUMULATORS} = psimd_qfma_f32(vacc${ABC[C:C+4]}p${K % ACCUMULATORS}, vi${K}x${ABC[C:C+4]}, vk${K}x${ABC[C:C+4]}); 59 60 w += ${(KERNEL_TILE + 1) * CHANNEL_TILE}; 61 62 $if ACCUMULATORS > 1: 63 // Add up all accumulators to vacc${ABC[0:CHANNEL_TILE]}p0 64 $ACC_SLICE = 1 65 $while ACC_SLICE < ACCUMULATORS: 66 $for A in range(0, ACCUMULATORS, ACC_SLICE * 2): 67 $if A + ACC_SLICE < ACCUMULATORS: 68 $for C in range(0, CHANNEL_TILE, 4): 69 vacc${ABC[C:C+4]}p${A} = psimd_add_f32(vacc${ABC[C:C+4]}p${A}, vacc${ABC[C:C+4]}p${A + ACC_SLICE}); 70 $ACC_SLICE *= 2 71 72 $for C in range(0, CHANNEL_TILE, 4): 73 psimd_f32 vacc${ABC[C:C+4]} = psimd_max_f32(vacc${ABC[C:C+4]}p0, vmin); 74 $for C in range(0, CHANNEL_TILE, 4): 75 vacc${ABC[C:C+4]} = psimd_min_f32(vacc${ABC[C:C+4]}, vmax); 76 77 psimd_store_f32(output, vacc${ABC[0:4]}); 78 $for C in range(4, CHANNEL_TILE, 4): 79 psimd_store_f32(output + ${C}, vacc${ABC[C:C+4]}); 80 output += ${CHANNEL_TILE}; 81 } 82 $if CHANNEL_TILE > 4: 83 for (; c >= 4; c -= 4) { 84 psimd_f32 vacc0123p0 = psimd_load_f32(w); 85 $for K in range(KERNEL_TILE): 86 87 const psimd_f32 vi${K}x0123 = psimd_load_f32(i${K}); 88 i${K} += 4; 89 90 const psimd_f32 vk${K}x0123 = psimd_load_f32(w + ${(K + 1) * CHANNEL_TILE}); 91 $if 1 <= K < ACCUMULATORS: 92 psimd_f32 vacc0123p${K} = psimd_mul_f32(vi${K}x0123, vk${K}x0123); 93 $else: 94 vacc0123p${K % ACCUMULATORS} = psimd_qfma_f32(vacc0123p${K % ACCUMULATORS}, vi${K}x0123, vk${K}x0123); 95 96 w += 4; 97 98 $if ACCUMULATORS > 1: 99 // Add up all accumulators to vacc${ABC[0:CHANNEL_TILE]}p0 100 $ACC_SLICE = 1 101 $while ACC_SLICE < ACCUMULATORS: 102 $for A in range(0, ACCUMULATORS, ACC_SLICE * 2): 103 $if A + ACC_SLICE < ACCUMULATORS: 104 vacc0123p${A} = psimd_add_f32(vacc0123p${A}, vacc0123p${A + ACC_SLICE}); 105 $ACC_SLICE *= 2 106 107 psimd_f32 vacc0123 = psimd_max_f32(vacc0123p0, vmin); 108 vacc0123 = psimd_min_f32(vacc0123, vmax); 109 110 psimd_store_f32(output, vacc0123); 111 output += 4; 112 } 113 if XNN_UNLIKELY(c != 0) { 114 psimd_f32 vacc0123p0 = psimd_load_f32(w); 115 $for K in range(KERNEL_TILE): 116 117 const psimd_f32 vi${K}x0123 = psimd_load_f32(i${K}); 118 const psimd_f32 vk${K}x0123 = psimd_load_f32(w + ${(K+1) * CHANNEL_TILE}); 119 $if 1 <= K < ACCUMULATORS: 120 psimd_f32 vacc0123p${K} = psimd_mul_f32(vi${K}x0123, vk${K}x0123); 121 $else: 122 vacc0123p${K % ACCUMULATORS} = psimd_qfma_f32(vacc0123p${K % ACCUMULATORS}, vi${K}x0123, vk${K}x0123); 123 124 $if ACCUMULATORS > 1: 125 // Add up all accumulators to vacc${ABC[0:CHANNEL_TILE]}p0 126 $ACC_SLICE = 1 127 $while ACC_SLICE < ACCUMULATORS: 128 $for A in range(0, ACCUMULATORS, ACC_SLICE * 2): 129 $if A + ACC_SLICE < ACCUMULATORS: 130 vacc0123p${A} = psimd_add_f32(vacc0123p${A}, vacc0123p${A + ACC_SLICE}); 131 $ACC_SLICE *= 2 132 133 psimd_f32 vacc0123 = psimd_max_f32(vacc0123p0, vmin); 134 vacc0123 = psimd_min_f32(vacc0123, vmax); 135 136 if (c & 2) { 137 psimd_store2_f32(output, vacc0123); 138 vacc0123 = psimd_concat_hi_f32(vacc0123, vacc0123); 139 output += 2; 140 } 141 if (c & 1) { 142 psimd_store1_f32(output, vacc0123); 143 output += 1; 144 } 145 } 146 147 output = (float*) ((uintptr_t) output + output_increment); 148 } while (--output_width != 0); 149} 150