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 PIXEL_TILE == 1 8$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ" 9#include <assert.h> 10 11#include <xnnpack/ibilinear.h> 12 13 14void xnn_f32_ibilinear_ukernel__scalar_c${CHANNEL_TILE}${"" if PIXEL_TILE == 1 else "x%d" % PIXEL_TILE}( 15 size_t output_pixels, 16 size_t channels, 17 const float**restrict input, 18 size_t input_offset, 19 const float*restrict weights, 20 float*restrict output, 21 size_t output_increment) 22{ 23 assert(output_pixels != 0); 24 assert(channels != 0); 25 assert(channels % sizeof(float) == 0); 26 27 do { 28 const float* i0 = (const float*) ((uintptr_t) input[0] + input_offset); 29 const float* i1 = (const float*) ((uintptr_t) input[1] + input_offset); 30 const float* i2 = (const float*) ((uintptr_t) input[2] + input_offset); 31 const float* i3 = (const float*) ((uintptr_t) input[3] + input_offset); 32 input += 4; 33 34 const float valphah = weights[0]; 35 const float valphav = weights[1]; 36 weights += 2; 37 38 size_t c = channels; 39 $if CHANNEL_TILE > 1: 40 for (; c >= ${CHANNEL_TILE} * sizeof(float); c -= ${CHANNEL_TILE} * sizeof(float)) { 41 $for C in range(CHANNEL_TILE): 42 const float vtl${ABC[C]} = i0[${C}]; 43 const float vtr${ABC[C]} = i1[${C}]; 44 const float vbl${ABC[C]} = i2[${C}]; 45 const float vbr${ABC[C]} = i3[${C}]; 46 i0 += ${CHANNEL_TILE}; 47 i1 += ${CHANNEL_TILE}; 48 i2 += ${CHANNEL_TILE}; 49 i3 += ${CHANNEL_TILE}; 50 51 $for C in range(CHANNEL_TILE): 52 const float vtd${ABC[C]} = vtr${ABC[C]} - vtl${ABC[C]}; 53 const float vbd${ABC[C]} = vbr${ABC[C]} - vbl${ABC[C]}; 54 55 $for C in range(CHANNEL_TILE): 56 const float vt${ABC[C]} = vtl${ABC[C]} + vtd${ABC[C]} * valphah; 57 const float vb${ABC[C]} = vbl${ABC[C]} + vbd${ABC[C]} * valphah; 58 59 $for C in range(CHANNEL_TILE): 60 const float vd${ABC[C]} = vb${ABC[C]} - vt${ABC[C]}; 61 62 $for C in range(CHANNEL_TILE): 63 const float vo${ABC[C]} = vt${ABC[C]} + vd${ABC[C]} * valphav; 64 65 $for C in range(CHANNEL_TILE): 66 output[${C}] = vo${ABC[C]}; 67 output += ${CHANNEL_TILE}; 68 } 69 for (; c >= sizeof(float); c -= sizeof(float)) { 70 const float vtl = *i0++; 71 const float vtr = *i1++; 72 const float vbl = *i2++; 73 const float vbr = *i3++; 74 75 const float vtd = vtr - vtl; 76 const float vbd = vbr - vbl; 77 78 const float vt = vtl + vtd * valphah; 79 const float vb = vbl + vbd * valphah; 80 81 const float vd = vb - vt; 82 83 const float vo = vt + vd * valphav; 84 85 *output++ = vo; 86 } 87 $else: 88 do { 89 const float vtl = *i0++; 90 const float vtr = *i1++; 91 const float vbl = *i2++; 92 const float vbr = *i3++; 93 94 const float vtd = vtr - vtl; 95 const float vbd = vbr - vbl; 96 97 const float vt = vtl + vtd * valphah; 98 const float vb = vbl + vbd * valphah; 99 100 const float vd = vb - vt; 101 102 const float vo = vt + vd * valphav; 103 104 *output++ = vo; 105 106 c -= sizeof(float); 107 } while (c != 0); 108 109 output = (float*) ((uintptr_t) output + output_increment); 110 } while (--output_pixels != 0); 111} 112