1// Copyright 2022 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$assert CHANNEL_TILE % 8 == 0 7$assert CHANNEL_TILE >= 8 8$assert PIXEL_TILE == 1 9$ABC = "456789ABCDEFGHIJKLMNOPQRSTUVWXYZ" 10#include <assert.h> 11 12#include <arm_neon.h> 13 14#include <xnnpack/common.h> 15#include <xnnpack/ibilinear.h> 16 17 18void xnn_f16_ibilinear_ukernel__neonfp16arith_c${CHANNEL_TILE}${"" if PIXEL_TILE == 1 else "x%d" % PIXEL_TILE}( 19 size_t output_pixels, 20 size_t channels, 21 const void**restrict input, 22 size_t input_offset, 23 const void*restrict weights, 24 void*restrict output, 25 size_t output_increment) XNN_OOB_READS 26{ 27 assert(output_pixels != 0); 28 assert(channels != 0); 29 assert(channels % sizeof(__fp16) == 0); 30 31 __fp16* o = (__fp16*) output; 32 do { 33 const __fp16* i0 = (const __fp16*) ((uintptr_t) input[0] + input_offset); 34 const __fp16* i1 = (const __fp16*) ((uintptr_t) input[1] + input_offset); 35 const __fp16* i2 = (const __fp16*) ((uintptr_t) input[2] + input_offset); 36 const __fp16* i3 = (const __fp16*) ((uintptr_t) input[3] + input_offset); 37 input += 4; 38 39 const float16x8_t valphah = vld1q_dup_f16(weights); weights = (const __fp16*) weights + 1; 40 const float16x8_t valphav = vld1q_dup_f16(weights); weights = (const __fp16*) weights + 1; 41 42 size_t c = channels; 43 $if CHANNEL_TILE > 8: 44 for (; c >= ${CHANNEL_TILE} * sizeof(__fp16); c -= ${CHANNEL_TILE} * sizeof(__fp16)) { 45 $for C in range(0, CHANNEL_TILE, 8): 46 const float16x8_t vtl${ABC[C:C+8]} = vld1q_f16(i0); i0 += 8; 47 const float16x8_t vtr${ABC[C:C+8]} = vld1q_f16(i1); i1 += 8; 48 const float16x8_t vbl${ABC[C:C+8]} = vld1q_f16(i2); i2 += 8; 49 const float16x8_t vbr${ABC[C:C+8]} = vld1q_f16(i3); i3 += 8; 50 51 $for C in range(0, CHANNEL_TILE, 8): 52 const float16x8_t vtd${ABC[C:C+8]} = vsubq_f16(vtr${ABC[C:C+8]}, vtl${ABC[C:C+8]}); 53 const float16x8_t vbd${ABC[C:C+8]} = vsubq_f16(vbr${ABC[C:C+8]}, vbl${ABC[C:C+8]}); 54 55 $for C in range(0, CHANNEL_TILE, 8): 56 const float16x8_t vt${ABC[C:C+8]} = vfmaq_f16(vtl${ABC[C:C+8]}, vtd${ABC[C:C+8]}, valphah); 57 const float16x8_t vb${ABC[C:C+8]} = vfmaq_f16(vbl${ABC[C:C+8]}, vbd${ABC[C:C+8]}, valphah); 58 59 $for C in range(0, CHANNEL_TILE, 8): 60 const float16x8_t vd${ABC[C:C+8]} = vsubq_f16(vb${ABC[C:C+8]}, vt${ABC[C:C+8]}); 61 62 $for C in range(0, CHANNEL_TILE, 8): 63 const float16x8_t vo${ABC[C:C+8]} = vfmaq_f16(vt${ABC[C:C+8]}, vd${ABC[C:C+8]}, valphav); 64 65 $for C in range(0, CHANNEL_TILE, 8): 66 vst1q_f16(o, vo${ABC[C:C+8]}); o += 8; 67 } 68 for (; c >= 8 * sizeof(__fp16); c -= 8 * sizeof(__fp16)) { 69 const float16x8_t vtl = vld1q_f16(i0); i0 += 8; 70 const float16x8_t vtr = vld1q_f16(i1); i1 += 8; 71 const float16x8_t vbl = vld1q_f16(i2); i2 += 8; 72 const float16x8_t vbr = vld1q_f16(i3); i3 += 8; 73 74 const float16x8_t vtd = vsubq_f16(vtr, vtl); 75 const float16x8_t vbd = vsubq_f16(vbr, vbl); 76 77 const float16x8_t vt = vfmaq_f16(vtl, vtd, valphah); 78 const float16x8_t vb = vfmaq_f16(vbl, vbd, valphah); 79 80 const float16x8_t vd = vsubq_f16(vb, vt); 81 82 const float16x8_t vo = vfmaq_f16(vt, vd, valphav); 83 84 vst1q_f16(o, vo); o += 8; 85 } 86 if XNN_UNLIKELY(c != 0) { 87 const float16x8_t vtl = vld1q_f16(i0); 88 const float16x8_t vtr = vld1q_f16(i1); 89 const float16x8_t vbl = vld1q_f16(i2); 90 const float16x8_t vbr = vld1q_f16(i3); 91 92 const float16x8_t vtd = vsubq_f16(vtr, vtl); 93 const float16x8_t vbd = vsubq_f16(vbr, vbl); 94 95 const float16x8_t vt = vfmaq_f16(vtl, vtd, valphah); 96 const float16x8_t vb = vfmaq_f16(vbl, vbd, valphah); 97 98 const float16x8_t vd = vsubq_f16(vb, vt); 99 100 float16x8_t vo = vfmaq_f16(vt, vd, valphav); 101 102 float16x4_t vo_lo = vget_low_f16(vo); 103 if (c & (4 * sizeof(__fp16))) { 104 vst1_f16(o, vo_lo); o += 4; 105 vo_lo = vget_high_f16(vo); 106 } 107 if (c & (2 * sizeof(__fp16))) { 108 vst1_lane_u32((void*) o, vreinterpret_u32_f16(vo_lo), 0); o += 2; 109 vo_lo = vext_f16(vo_lo, vo_lo, 2); 110 } 111 if (c & (1 * sizeof(__fp16))) { 112 vst1_lane_f16(o, vo_lo, 0); o += 1; 113 } 114 } 115 116 o = (__fp16*) ((uintptr_t) o + output_increment); 117 } while (--output_pixels != 0); 118} 119