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 = "0123456789ABCDEFGHIJKLMN" 10#include <assert.h> 11 12#include <psimd.h> 13 14#include <xnnpack/bilinear.h> 15 16 17void xnn_f32_bilinear_ukernel__psimd_c${CHANNEL_TILE}${"" if PIXEL_TILE == 1 else "x%d" % PIXEL_TILE}( 18 size_t output_pixels, 19 size_t channels, 20 const float**restrict input, 21 size_t input_offset, 22 const float*restrict weights, 23 float*restrict output, 24 size_t output_increment) 25{ 26 assert(output_pixels != 0); 27 assert(channels != 0); 28 assert(channels % sizeof(float) == 0); 29 30 do { 31 const float* i0 = (const float*) ((uintptr_t) input[0] + input_offset); 32 const float* i1 = (const float*) ((uintptr_t) input[1] + input_offset); 33 const float* i2 = (const float*) ((uintptr_t) input[2] + input_offset); 34 const float* i3 = (const float*) ((uintptr_t) input[3] + input_offset); 35 input += 4; 36 37 const psimd_f32 valphah = psimd_load_splat_f32(weights); 38 const psimd_f32 valphav = psimd_load_splat_f32(weights + 1); 39 weights += 2; 40 41 size_t c = channels; 42 for (; c >= ${CHANNEL_TILE} * sizeof(float); c -= ${CHANNEL_TILE} * sizeof(float)) { 43 const psimd_f32 vtl${ABC[0:4]} = psimd_load_f32(i0); 44 const psimd_f32 vtr${ABC[0:4]} = psimd_load_f32(i1); 45 const psimd_f32 vbl${ABC[0:4]} = psimd_load_f32(i2); 46 const psimd_f32 vbr${ABC[0:4]} = psimd_load_f32(i3); 47 $for C in range(4, CHANNEL_TILE, 4): 48 const psimd_f32 vtl${ABC[C:C+4]} = psimd_load_f32(i0 + ${C}); 49 const psimd_f32 vtr${ABC[C:C+4]} = psimd_load_f32(i1 + ${C}); 50 const psimd_f32 vbl${ABC[C:C+4]} = psimd_load_f32(i2 + ${C}); 51 const psimd_f32 vbr${ABC[C:C+4]} = psimd_load_f32(i3 + ${C}); 52 i0 += ${CHANNEL_TILE}; 53 i1 += ${CHANNEL_TILE}; 54 i2 += ${CHANNEL_TILE}; 55 i3 += ${CHANNEL_TILE}; 56 57 $for C in range(0, CHANNEL_TILE, 4): 58 const psimd_f32 vtd${ABC[C:C+4]} = psimd_sub_f32(vtr${ABC[C:C+4]}, vtl${ABC[C:C+4]}); 59 const psimd_f32 vbd${ABC[C:C+4]} = psimd_sub_f32(vbr${ABC[C:C+4]}, vbl${ABC[C:C+4]}); 60 61 $for C in range(0, CHANNEL_TILE, 4): 62 const psimd_f32 vt${ABC[C:C+4]} = psimd_qfma_f32(vtl${ABC[C:C+4]}, vtd${ABC[C:C+4]}, valphah); 63 const psimd_f32 vb${ABC[C:C+4]} = psimd_qfma_f32(vbl${ABC[C:C+4]}, vbd${ABC[C:C+4]}, valphah); 64 65 $for C in range(0, CHANNEL_TILE, 4): 66 const psimd_f32 vd${ABC[C:C+4]} = psimd_sub_f32(vb${ABC[C:C+4]}, vt${ABC[C:C+4]}); 67 68 $for C in range(0, CHANNEL_TILE, 4): 69 const psimd_f32 vo${ABC[C:C+4]} = psimd_qfma_f32(vt${ABC[C:C+4]}, vd${ABC[C:C+4]}, valphav); 70 71 psimd_store_f32(output, vo${ABC[0:4]}); 72 $for C in range(4, CHANNEL_TILE, 4): 73 psimd_store_f32(output + ${C}, vo${ABC[C:C+4]}); 74 output += ${CHANNEL_TILE}; 75 } 76 $if CHANNEL_TILE > 4: 77 for (; c >= 4 * sizeof(float); c -= 4 * sizeof(float)) { 78 const psimd_f32 vtl0123 = psimd_load_f32(i0); 79 const psimd_f32 vtr0123 = psimd_load_f32(i1); 80 const psimd_f32 vbl0123 = psimd_load_f32(i2); 81 const psimd_f32 vbr0123 = psimd_load_f32(i3); 82 i0 += 4; 83 i1 += 4; 84 i2 += 4; 85 i3 += 4; 86 87 const psimd_f32 vtd0123 = psimd_sub_f32(vtr0123, vtl0123); 88 const psimd_f32 vbd0123 = psimd_sub_f32(vbr0123, vbl0123); 89 90 const psimd_f32 vt0123 = psimd_qfma_f32(vtl0123, vtd0123, valphah); 91 const psimd_f32 vb0123 = psimd_qfma_f32(vbl0123, vbd0123, valphah); 92 93 const psimd_f32 vd0123 = psimd_sub_f32(vb0123, vt0123); 94 95 const psimd_f32 vo0123 = psimd_qfma_f32(vt0123, vd0123, valphav); 96 97 psimd_store_f32(output, vo0123); 98 output += 4; 99 } 100 if XNN_UNLIKELY(c != 0) { 101 const psimd_f32 vtl0123 = psimd_load_f32(i0); 102 const psimd_f32 vtr0123 = psimd_load_f32(i1); 103 const psimd_f32 vbl0123 = psimd_load_f32(i2); 104 const psimd_f32 vbr0123 = psimd_load_f32(i3); 105 106 const psimd_f32 vtd0123 = psimd_sub_f32(vtr0123, vtl0123); 107 const psimd_f32 vbd0123 = psimd_sub_f32(vbr0123, vbl0123); 108 109 const psimd_f32 vt0123 = psimd_qfma_f32(vtl0123, vtd0123, valphah); 110 const psimd_f32 vb0123 = psimd_qfma_f32(vbl0123, vbd0123, valphah); 111 112 const psimd_f32 vd0123 = psimd_sub_f32(vb0123, vt0123); 113 114 psimd_f32 vo0123 = psimd_qfma_f32(vt0123, vd0123, valphav); 115 116 if (c & (2 * sizeof(float))) { 117 psimd_store2_f32(output, vo0123); 118 vo0123 = psimd_concat_hi_f32(vo0123, vo0123); 119 output += 2; 120 } 121 if (c & (1 * sizeof(float))) { 122 psimd_store1_f32(output, vo0123); 123 output += 1; 124 } 125 } 126 127 output = (float*) ((uintptr_t) output + output_increment); 128 } while (--output_pixels != 0); 129} 130