XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1 | // Auto-generated file. Do not edit! |
| 2 | // Template: src/f32-dwconv/up-psimd.c.in |
| 3 | // Generator: tools/xngen |
| 4 | // |
| 5 | // Copyright 2019 Google LLC |
| 6 | // |
| 7 | // This source code is licensed under the BSD-style license found in the |
| 8 | // LICENSE file in the root directory of this source tree. |
| 9 | |
| 10 | #include <assert.h> |
| 11 | |
| 12 | #include <psimd.h> |
| 13 | |
| 14 | #include <xnnpack/dwconv.h> |
| 15 | |
| 16 | |
| 17 | void xnn_f32_dwconv_ukernel_up4x4__psimd( |
| 18 | size_t channels, |
| 19 | size_t output_width, |
| 20 | const float** input, |
| 21 | const float* weights, |
| 22 | float* output, |
| 23 | size_t input_stride, |
| 24 | size_t output_increment, |
Marat Dukhan | eb09a6b | 2020-04-08 17:34:32 -0700 | [diff] [blame] | 25 | const union xnn_f32_minmax_params params[restrict static 1]) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 26 | { |
| 27 | assert(channels != 0); |
| 28 | assert(output_width != 0); |
| 29 | |
| 30 | const psimd_f32 vmax = psimd_load_splat_f32(¶ms->scalar.max); |
| 31 | const psimd_f32 vmin = psimd_load_splat_f32(¶ms->scalar.min); |
| 32 | do { |
| 33 | const float* i0 = input[0]; |
Marat Dukhan | 6866099 | 2020-02-03 13:31:12 -0800 | [diff] [blame] | 34 | assert(i0 != NULL); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 35 | const float* i1 = input[1]; |
Marat Dukhan | 6866099 | 2020-02-03 13:31:12 -0800 | [diff] [blame] | 36 | assert(i1 != NULL); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 37 | const float* i2 = input[2]; |
Marat Dukhan | 6866099 | 2020-02-03 13:31:12 -0800 | [diff] [blame] | 38 | assert(i2 != NULL); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 39 | const float* i3 = input[3]; |
Marat Dukhan | 6866099 | 2020-02-03 13:31:12 -0800 | [diff] [blame] | 40 | assert(i3 != NULL); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 41 | input = (const float**) ((uintptr_t) input + input_stride); |
| 42 | |
| 43 | size_t c = channels; |
| 44 | const float* w = weights; |
| 45 | for (; c >= 4; c -= 4) { |
Marat Dukhan | 5098c3e | 2019-11-07 12:01:19 -0800 | [diff] [blame] | 46 | psimd_f32 vacc0123p0 = psimd_load_f32(w); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 47 | |
Marat Dukhan | 5098c3e | 2019-11-07 12:01:19 -0800 | [diff] [blame] | 48 | |
| 49 | const psimd_f32 vi0x0123 = psimd_load_f32(i0); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 50 | i0 += 4; |
| 51 | |
Marat Dukhan | 5098c3e | 2019-11-07 12:01:19 -0800 | [diff] [blame] | 52 | const psimd_f32 vk0x0123 = psimd_load_f32(w + 4); |
| 53 | vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi0x0123, vk0x0123); |
| 54 | |
| 55 | const psimd_f32 vi1x0123 = psimd_load_f32(i1); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 56 | i1 += 4; |
| 57 | |
Marat Dukhan | 5098c3e | 2019-11-07 12:01:19 -0800 | [diff] [blame] | 58 | const psimd_f32 vk1x0123 = psimd_load_f32(w + 8); |
| 59 | vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi1x0123, vk1x0123); |
| 60 | |
| 61 | const psimd_f32 vi2x0123 = psimd_load_f32(i2); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 62 | i2 += 4; |
| 63 | |
Marat Dukhan | 5098c3e | 2019-11-07 12:01:19 -0800 | [diff] [blame] | 64 | const psimd_f32 vk2x0123 = psimd_load_f32(w + 12); |
| 65 | vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi2x0123, vk2x0123); |
| 66 | |
| 67 | const psimd_f32 vi3x0123 = psimd_load_f32(i3); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 68 | i3 += 4; |
| 69 | |
Marat Dukhan | 5098c3e | 2019-11-07 12:01:19 -0800 | [diff] [blame] | 70 | const psimd_f32 vk3x0123 = psimd_load_f32(w + 16); |
| 71 | vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi3x0123, vk3x0123); |
| 72 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 73 | w += 20; |
| 74 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 75 | |
Marat Dukhan | 5098c3e | 2019-11-07 12:01:19 -0800 | [diff] [blame] | 76 | psimd_f32 vacc0123 = psimd_max_f32(vacc0123p0, vmin); |
| 77 | vacc0123 = psimd_min_f32(vacc0123, vmax); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 78 | |
Marat Dukhan | 5098c3e | 2019-11-07 12:01:19 -0800 | [diff] [blame] | 79 | psimd_store_f32(output, vacc0123); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 80 | output += 4; |
| 81 | } |
| 82 | if XNN_UNLIKELY(c != 0) { |
Marat Dukhan | 5098c3e | 2019-11-07 12:01:19 -0800 | [diff] [blame] | 83 | psimd_f32 vacc0123p0 = psimd_load_f32(w); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 84 | |
Marat Dukhan | 5098c3e | 2019-11-07 12:01:19 -0800 | [diff] [blame] | 85 | const psimd_f32 vi0x0123 = psimd_load_f32(i0); |
| 86 | const psimd_f32 vk0x0123 = psimd_load_f32(w + 4); |
| 87 | vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi0x0123, vk0x0123); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 88 | |
Marat Dukhan | 5098c3e | 2019-11-07 12:01:19 -0800 | [diff] [blame] | 89 | const psimd_f32 vi1x0123 = psimd_load_f32(i1); |
| 90 | const psimd_f32 vk1x0123 = psimd_load_f32(w + 8); |
| 91 | vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi1x0123, vk1x0123); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 92 | |
Marat Dukhan | 5098c3e | 2019-11-07 12:01:19 -0800 | [diff] [blame] | 93 | const psimd_f32 vi2x0123 = psimd_load_f32(i2); |
| 94 | const psimd_f32 vk2x0123 = psimd_load_f32(w + 12); |
| 95 | vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi2x0123, vk2x0123); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 96 | |
Marat Dukhan | 5098c3e | 2019-11-07 12:01:19 -0800 | [diff] [blame] | 97 | const psimd_f32 vi3x0123 = psimd_load_f32(i3); |
| 98 | const psimd_f32 vk3x0123 = psimd_load_f32(w + 16); |
| 99 | vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi3x0123, vk3x0123); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 100 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 101 | |
Marat Dukhan | 5098c3e | 2019-11-07 12:01:19 -0800 | [diff] [blame] | 102 | psimd_f32 vacc0123 = psimd_max_f32(vacc0123p0, vmin); |
| 103 | vacc0123 = psimd_min_f32(vacc0123, vmax); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 104 | |
| 105 | if (c & 2) { |
Marat Dukhan | 5098c3e | 2019-11-07 12:01:19 -0800 | [diff] [blame] | 106 | psimd_store2_f32(output, vacc0123); |
| 107 | vacc0123 = psimd_concat_hi_f32(vacc0123, vacc0123); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 108 | output += 2; |
| 109 | } |
| 110 | if (c & 1) { |
Marat Dukhan | 5098c3e | 2019-11-07 12:01:19 -0800 | [diff] [blame] | 111 | psimd_store1_f32(output, vacc0123); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 112 | output += 1; |
| 113 | } |
| 114 | } |
| 115 | |
| 116 | output = (float*) ((uintptr_t) output + output_increment); |
| 117 | } while (--output_width != 0); |
| 118 | } |