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$assert CHANNEL_TILE >= 1 7$assert ROW_TILE >= 1 8$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ" 9#include <assert.h> 10 11#include <xnnpack/math.h> 12#include <xnnpack/vmulcaddc.h> 13 14 15$MIN_F32 = "__builtin_wasm_min_f32" if WASM else "math_min_f32" 16$MAX_F32 = "__builtin_wasm_max_f32" if WASM else "math_max_f32" 17void xnn_f32_vmulcaddc_minmax_ukernel_c${CHANNEL_TILE}__${"wasm" if WASM else "scalar"}_${ROW_TILE}x( 18 size_t rows, 19 size_t channels, 20 const float*restrict input, 21 size_t input_stride, 22 const float*restrict weights, 23 float*restrict output, 24 size_t output_stride, 25 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) 26{ 27 assert(rows != 0); 28 assert(channels != 0); 29 assert(channels % sizeof(float) == 0); 30 31 const size_t input_increment = input_stride * ${ROW_TILE} - channels; 32 const size_t output_increment = output_stride * ${ROW_TILE} - channels; 33 34 const float* i0 = input; 35 float* o0 = output; 36 $for M in range(1, ROW_TILE): 37 const float* i${M} = (const float*) ((uintptr_t) i${M-1} + input_stride); 38 float* o${M} = (float*) ((uintptr_t) o${M-1} + output_stride); 39 40 const float vmin = params->scalar.min; 41 const float vmax = params->scalar.max; 42 do { 43 $for M in range(1, ROW_TILE): 44 $if M % 2 == 0: 45 if XNN_UNPREDICTABLE(rows <= ${M}) { 46 i${M} = i${M-1}; 47 o${M} = o${M-1}; 48 } 49 $else: 50 if XNN_UNPREDICTABLE(rows < ${M+1}) { 51 i${M} = i${M-1}; 52 o${M} = o${M-1}; 53 } 54 55 const float* w = weights; 56 size_t c = channels; 57 $if CHANNEL_TILE > 1: 58 for (; c >= ${CHANNEL_TILE} * sizeof(float); c -= ${CHANNEL_TILE} * sizeof(float)) { 59 $for C in range(CHANNEL_TILE): 60 const float vscale${ABC[C]} = w[${C}]; 61 62 $for M in range(ROW_TILE): 63 $for C in range(CHANNEL_TILE): 64 float vacc${M}x${ABC[C]} = i${M}[${C}]; 65 i${M} += ${CHANNEL_TILE}; 66 67 $for C in range(CHANNEL_TILE): 68 const float vbias${ABC[C]} = w[${C + CHANNEL_TILE}]; 69 70 $for M in range(ROW_TILE): 71 $for C in range(CHANNEL_TILE): 72 vacc${M}x${ABC[C]} = vacc${M}x${ABC[C]} * vscale${ABC[C]} + vbias${ABC[C]}; 73 74 $for M in range(ROW_TILE): 75 $for C in range(CHANNEL_TILE): 76 vacc${M}x${ABC[C]} = ${MAX_F32}(vacc${M}x${ABC[C]}, vmin); 77 78 $for M in range(ROW_TILE): 79 $for C in range(CHANNEL_TILE): 80 vacc${M}x${ABC[C]} = ${MIN_F32}(vacc${M}x${ABC[C]}, vmax); 81 82 $for M in range(ROW_TILE): 83 $for C in range(CHANNEL_TILE): 84 o${M}[${C}] = vacc${M}x${ABC[C]}; 85 o${M} += ${CHANNEL_TILE}; 86 87 w += ${CHANNEL_TILE * 2}; 88 } 89 if XNN_UNLIKELY(c != 0) { 90 do { 91 const float vscale = *w++; 92 93 $for M in range(ROW_TILE): 94 float vacc${M} = *i${M}++; 95 96 const float vbias = w[${CHANNEL_TILE - 1}]; 97 98 $for M in range(ROW_TILE): 99 vacc${M} = vacc${M} * vscale + vbias; 100 101 $for M in range(ROW_TILE): 102 vacc${M} = ${MAX_F32}(vacc${M}, vmin); 103 104 $for M in range(ROW_TILE): 105 vacc${M} = ${MIN_F32}(vacc${M}, vmax); 106 107 $for M in range(ROW_TILE): 108 *o${M}++ = vacc${M}; 109 110 c -= sizeof(float); 111 } while (c != 0); 112 } 113 $else: 114 do { 115 const float vscale = w[0]; 116 117 $for M in range(ROW_TILE): 118 float vacc${M} = *i${M}++; 119 120 const float vbias = w[1]; 121 122 $for M in range(ROW_TILE): 123 vacc${M} = vacc${M} * vscale + vbias; 124 125 $for M in range(ROW_TILE): 126 vacc${M} = ${MAX_F32}(vacc${M}, vmin); 127 128 $for M in range(ROW_TILE): 129 vacc${M} = ${MIN_F32}(vacc${M}, vmax); 130 131 $for M in range(ROW_TILE): 132 *o${M}++ = vacc${M}; 133 134 w += 2; 135 c -= sizeof(float); 136 } while (c != 0); 137 $for M in range(ROW_TILE): 138 i${M} = (const float*) ((uintptr_t) i${M} + input_increment); 139 o${M} = (float*) ((uintptr_t) o${M} + output_increment); 140 rows = doz(rows, ${ROW_TILE}); 141 } while (rows != 0); 142} 143