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 CHANNEL_TILE >= 4 8$assert PIXEL_TILE == 1 9$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ" 10$VMULADDQ_F32 = "vfmaq_f32" if FMA else "vmlaq_f32" 11#include <assert.h> 12 13#include <arm_neon.h> 14 15#include <xnnpack/common.h> 16#include <xnnpack/ibilinear.h> 17 18 19void xnn_f32_ibilinear_ukernel__${"neonfma" if FMA else "neon"}_c${CHANNEL_TILE}${"" if PIXEL_TILE == 1 else "x%d" % PIXEL_TILE}( 20 size_t output_pixels, 21 size_t channels, 22 const float**restrict input, 23 size_t input_offset, 24 const float*restrict weights, 25 float*restrict output, 26 size_t output_increment) XNN_DISABLE_TSAN 27{ 28 assert(output_pixels != 0); 29 assert(channels != 0); 30 assert(channels % sizeof(float) == 0); 31 32 do { 33 const float* i0 = (const float*) ((uintptr_t) input[0] + input_offset); 34 const float* i1 = (const float*) ((uintptr_t) input[1] + input_offset); 35 const float* i2 = (const float*) ((uintptr_t) input[2] + input_offset); 36 const float* i3 = (const float*) ((uintptr_t) input[3] + input_offset); 37 input += 4; 38 39 const float32x2_t valphahv = vld1_f32(weights); weights += 2; 40 $if FMA: 41 #if XNN_ARCH_ARM 42 const float32x4_t valphah = vdupq_lane_f32(valphahv, 0); 43 const float32x4_t valphav = vdupq_lane_f32(valphahv, 1); 44 #endif 45 46 size_t c = channels; 47 for (; c >= ${CHANNEL_TILE} * sizeof(float); c -= ${CHANNEL_TILE} * sizeof(float)) { 48 $for C in range(0, CHANNEL_TILE, 4): 49 const float32x4_t vtl${ABC[C:C+4]} = vld1q_f32(i0); i0 += 4; 50 const float32x4_t vtr${ABC[C:C+4]} = vld1q_f32(i1); i1 += 4; 51 const float32x4_t vbl${ABC[C:C+4]} = vld1q_f32(i2); i2 += 4; 52 const float32x4_t vbr${ABC[C:C+4]} = vld1q_f32(i3); i3 += 4; 53 54 $for C in range(0, CHANNEL_TILE, 4): 55 const float32x4_t vtd${ABC[C:C+4]} = vsubq_f32(vtr${ABC[C:C+4]}, vtl${ABC[C:C+4]}); 56 const float32x4_t vbd${ABC[C:C+4]} = vsubq_f32(vbr${ABC[C:C+4]}, vbl${ABC[C:C+4]}); 57 58 $if FMA: 59 #if XNN_ARCH_ARM 60 $for C in range(0, CHANNEL_TILE, 4): 61 const float32x4_t vt${ABC[C:C+4]} = vfmaq_f32(vtl${ABC[C:C+4]}, vtd${ABC[C:C+4]}, valphah); 62 const float32x4_t vb${ABC[C:C+4]} = vfmaq_f32(vbl${ABC[C:C+4]}, vbd${ABC[C:C+4]}, valphah); 63 #else 64 $for C in range(0, CHANNEL_TILE, 4): 65 const float32x4_t vt${ABC[C:C+4]} = vfmaq_lane_f32(vtl${ABC[C:C+4]}, vtd${ABC[C:C+4]}, valphahv, 0); 66 const float32x4_t vb${ABC[C:C+4]} = vfmaq_lane_f32(vbl${ABC[C:C+4]}, vbd${ABC[C:C+4]}, valphahv, 0); 67 #endif 68 $else: 69 $for C in range(0, CHANNEL_TILE, 4): 70 const float32x4_t vt${ABC[C:C+4]} = vmlaq_lane_f32(vtl${ABC[C:C+4]}, vtd${ABC[C:C+4]}, valphahv, 0); 71 const float32x4_t vb${ABC[C:C+4]} = vmlaq_lane_f32(vbl${ABC[C:C+4]}, vbd${ABC[C:C+4]}, valphahv, 0); 72 73 $for C in range(0, CHANNEL_TILE, 4): 74 const float32x4_t vd${ABC[C:C+4]} = vsubq_f32(vb${ABC[C:C+4]}, vt${ABC[C:C+4]}); 75 76 $if FMA: 77 #if XNN_ARCH_ARM 78 $for C in range(0, CHANNEL_TILE, 4): 79 const float32x4_t vo${ABC[C:C+4]} = vfmaq_f32(vt${ABC[C:C+4]}, vd${ABC[C:C+4]}, valphav); 80 #else 81 $for C in range(0, CHANNEL_TILE, 4): 82 const float32x4_t vo${ABC[C:C+4]} = vfmaq_lane_f32(vt${ABC[C:C+4]}, vd${ABC[C:C+4]}, valphahv, 1); 83 #endif 84 $else: 85 $for C in range(0, CHANNEL_TILE, 4): 86 const float32x4_t vo${ABC[C:C+4]} = vmlaq_lane_f32(vt${ABC[C:C+4]}, vd${ABC[C:C+4]}, valphahv, 1); 87 88 $for C in range(0, CHANNEL_TILE, 4): 89 vst1q_f32(output, vo${ABC[C:C+4]}); output += 4; 90 } 91 $if CHANNEL_TILE > 4: 92 for (; c >= 4 * sizeof(float); c -= 4 * sizeof(float)) { 93 const float32x4_t vtl0123 = vld1q_f32(i0); i0 += 4; 94 const float32x4_t vtr0123 = vld1q_f32(i1); i1 += 4; 95 const float32x4_t vbl0123 = vld1q_f32(i2); i2 += 4; 96 const float32x4_t vbr0123 = vld1q_f32(i3); i3 += 4; 97 98 const float32x4_t vtd0123 = vsubq_f32(vtr0123, vtl0123); 99 const float32x4_t vbd0123 = vsubq_f32(vbr0123, vbl0123); 100 101 $if FMA: 102 #if XNN_ARCH_ARM 103 const float32x4_t vt0123 = vfmaq_f32(vtl0123, vtd0123, valphah); 104 const float32x4_t vb0123 = vfmaq_f32(vbl0123, vbd0123, valphah); 105 #else 106 const float32x4_t vt0123 = vfmaq_lane_f32(vtl0123, vtd0123, valphahv, 0); 107 const float32x4_t vb0123 = vfmaq_lane_f32(vbl0123, vbd0123, valphahv, 0); 108 #endif 109 $else: 110 const float32x4_t vt0123 = vmlaq_lane_f32(vtl0123, vtd0123, valphahv, 0); 111 const float32x4_t vb0123 = vmlaq_lane_f32(vbl0123, vbd0123, valphahv, 0); 112 113 const float32x4_t vd0123 = vsubq_f32(vb0123, vt0123); 114 115 $if FMA: 116 #if XNN_ARCH_ARM 117 const float32x4_t vo0123 = vfmaq_f32(vt0123, vd0123, valphav); 118 #else 119 const float32x4_t vo0123 = vfmaq_lane_f32(vt0123, vd0123, valphahv, 1); 120 #endif 121 $else: 122 const float32x4_t vo0123 = vmlaq_lane_f32(vt0123, vd0123, valphahv, 1); 123 124 vst1q_f32(output, vo0123); 125 output += 4; 126 } 127 if XNN_UNLIKELY(c != 0) { 128 const float32x4_t vtl0123 = vld1q_f32(i0); 129 const float32x4_t vtr0123 = vld1q_f32(i1); 130 const float32x4_t vbl0123 = vld1q_f32(i2); 131 const float32x4_t vbr0123 = vld1q_f32(i3); 132 133 const float32x4_t vtd0123 = vsubq_f32(vtr0123, vtl0123); 134 const float32x4_t vbd0123 = vsubq_f32(vbr0123, vbl0123); 135 136 $if FMA: 137 #if XNN_ARCH_ARM 138 const float32x4_t vt0123 = vfmaq_f32(vtl0123, vtd0123, valphah); 139 const float32x4_t vb0123 = vfmaq_f32(vbl0123, vbd0123, valphah); 140 #else 141 const float32x4_t vt0123 = vfmaq_lane_f32(vtl0123, vtd0123, valphahv, 0); 142 const float32x4_t vb0123 = vfmaq_lane_f32(vbl0123, vbd0123, valphahv, 0); 143 #endif 144 $else: 145 const float32x4_t vt0123 = vmlaq_lane_f32(vtl0123, vtd0123, valphahv, 0); 146 const float32x4_t vb0123 = vmlaq_lane_f32(vbl0123, vbd0123, valphahv, 0); 147 148 const float32x4_t vd0123 = vsubq_f32(vb0123, vt0123); 149 150 $if FMA: 151 #if XNN_ARCH_ARM 152 float32x4_t vo0123 = vfmaq_f32(vt0123, vd0123, valphav); 153 #else 154 float32x4_t vo0123 = vfmaq_lane_f32(vt0123, vd0123, valphahv, 1); 155 #endif 156 $else: 157 const float32x4_t vo0123 = vmlaq_lane_f32(vt0123, vd0123, valphahv, 1); 158 159 float32x2_t vo01 = vget_low_f32(vo0123); 160 if (c & (2 * sizeof(float))) { 161 vst1_f32(output, vo01); output += 2; 162 vo01 = vget_high_f32(vo0123); 163 } 164 if (c & (1 * sizeof(float))) { 165 vst1_lane_f32(output, vo01, 0); output += 1; 166 } 167 } 168 169 output = (float*) ((uintptr_t) output + output_increment); 170 } while (--output_pixels != 0); 171} 172