Marat Dukhan | 5098c3e | 2019-11-07 12:01:19 -0800 | [diff] [blame] | 1 | // Auto-generated file. Do not edit! |
| 2 | // Template: src/f32-dwconv/up-sse.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 <xmmintrin.h> |
| 13 | |
| 14 | #include <xnnpack/dwconv.h> |
| 15 | |
| 16 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 17 | void xnn_f32_dwconv_minmax_ukernel_up8x4__sse( |
Marat Dukhan | 5098c3e | 2019-11-07 12:01:19 -0800 | [diff] [blame] | 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 | f196d01 | 2020-04-15 11:50:03 -0700 | [diff] [blame] | 25 | const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) |
Marat Dukhan | 5098c3e | 2019-11-07 12:01:19 -0800 | [diff] [blame] | 26 | { |
| 27 | assert(channels != 0); |
| 28 | assert(output_width != 0); |
| 29 | |
| 30 | const __m128 vmax = _mm_load_ps(params->sse.max); |
| 31 | const __m128 vmin = _mm_load_ps(params->sse.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); |
Marat Dukhan | 5098c3e | 2019-11-07 12:01:19 -0800 | [diff] [blame] | 35 | const float* i1 = input[1]; |
Marat Dukhan | 6866099 | 2020-02-03 13:31:12 -0800 | [diff] [blame] | 36 | assert(i1 != NULL); |
Marat Dukhan | 5098c3e | 2019-11-07 12:01:19 -0800 | [diff] [blame] | 37 | const float* i2 = input[2]; |
Marat Dukhan | 6866099 | 2020-02-03 13:31:12 -0800 | [diff] [blame] | 38 | assert(i2 != NULL); |
Marat Dukhan | 5098c3e | 2019-11-07 12:01:19 -0800 | [diff] [blame] | 39 | const float* i3 = input[3]; |
Marat Dukhan | 6866099 | 2020-02-03 13:31:12 -0800 | [diff] [blame] | 40 | assert(i3 != NULL); |
Marat Dukhan | 5098c3e | 2019-11-07 12:01:19 -0800 | [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 >= 8; c -= 8) { |
| 46 | __m128 vacc0123p0 = _mm_load_ps(w); |
| 47 | __m128 vacc4567p0 = _mm_load_ps(w + 4); |
| 48 | |
| 49 | |
| 50 | const __m128 vi0x0123 = _mm_loadu_ps(i0); |
| 51 | const __m128 vi0x4567 = _mm_loadu_ps(i0 + 4); |
| 52 | i0 += 8; |
| 53 | |
| 54 | const __m128 vk0x0123 = _mm_load_ps(w + 8); |
| 55 | const __m128 vk0x4567 = _mm_load_ps(w + 12); |
| 56 | vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi0x0123, vk0x0123)); |
| 57 | vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi0x4567, vk0x4567)); |
| 58 | |
| 59 | const __m128 vi1x0123 = _mm_loadu_ps(i1); |
| 60 | const __m128 vi1x4567 = _mm_loadu_ps(i1 + 4); |
| 61 | i1 += 8; |
| 62 | |
| 63 | const __m128 vk1x0123 = _mm_load_ps(w + 16); |
| 64 | const __m128 vk1x4567 = _mm_load_ps(w + 20); |
| 65 | vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi1x0123, vk1x0123)); |
| 66 | vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi1x4567, vk1x4567)); |
| 67 | |
| 68 | const __m128 vi2x0123 = _mm_loadu_ps(i2); |
| 69 | const __m128 vi2x4567 = _mm_loadu_ps(i2 + 4); |
| 70 | i2 += 8; |
| 71 | |
| 72 | const __m128 vk2x0123 = _mm_load_ps(w + 24); |
| 73 | const __m128 vk2x4567 = _mm_load_ps(w + 28); |
| 74 | vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi2x0123, vk2x0123)); |
| 75 | vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi2x4567, vk2x4567)); |
| 76 | |
| 77 | const __m128 vi3x0123 = _mm_loadu_ps(i3); |
| 78 | const __m128 vi3x4567 = _mm_loadu_ps(i3 + 4); |
| 79 | i3 += 8; |
| 80 | |
| 81 | const __m128 vk3x0123 = _mm_load_ps(w + 32); |
| 82 | const __m128 vk3x4567 = _mm_load_ps(w + 36); |
| 83 | vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi3x0123, vk3x0123)); |
| 84 | vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi3x4567, vk3x4567)); |
| 85 | |
| 86 | w += 40; |
| 87 | |
| 88 | |
| 89 | __m128 vacc0123 = _mm_max_ps(vacc0123p0, vmin); |
| 90 | __m128 vacc4567 = _mm_max_ps(vacc4567p0, vmin); |
| 91 | vacc0123 = _mm_min_ps(vacc0123, vmax); |
| 92 | vacc4567 = _mm_min_ps(vacc4567, vmax); |
| 93 | |
| 94 | _mm_storeu_ps(output, vacc0123); |
| 95 | _mm_storeu_ps(output + 4, vacc4567); |
| 96 | output += 8; |
| 97 | } |
| 98 | for (; c >= 4; c -= 4) { |
| 99 | __m128 vacc0123p0 = _mm_load_ps(w); |
| 100 | |
| 101 | const __m128 vi0x0123 = _mm_loadu_ps(i0); |
| 102 | i0 += 4; |
| 103 | |
| 104 | const __m128 vk0x0123 = _mm_load_ps(w + 8); |
| 105 | vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi0x0123, vk0x0123)); |
| 106 | |
| 107 | const __m128 vi1x0123 = _mm_loadu_ps(i1); |
| 108 | i1 += 4; |
| 109 | |
| 110 | const __m128 vk1x0123 = _mm_load_ps(w + 16); |
| 111 | vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi1x0123, vk1x0123)); |
| 112 | |
| 113 | const __m128 vi2x0123 = _mm_loadu_ps(i2); |
| 114 | i2 += 4; |
| 115 | |
| 116 | const __m128 vk2x0123 = _mm_load_ps(w + 24); |
| 117 | vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi2x0123, vk2x0123)); |
| 118 | |
| 119 | const __m128 vi3x0123 = _mm_loadu_ps(i3); |
| 120 | i3 += 4; |
| 121 | |
| 122 | const __m128 vk3x0123 = _mm_load_ps(w + 32); |
| 123 | vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi3x0123, vk3x0123)); |
| 124 | |
| 125 | w += 4; |
| 126 | |
| 127 | |
| 128 | __m128 vacc0123 = _mm_max_ps(vacc0123p0, vmin); |
| 129 | vacc0123 = _mm_min_ps(vacc0123, vmax); |
| 130 | |
| 131 | _mm_storeu_ps(output, vacc0123); |
| 132 | output += 4; |
| 133 | } |
| 134 | if XNN_UNLIKELY(c != 0) { |
| 135 | __m128 vacc0123p0 = _mm_load_ps(w); |
| 136 | |
| 137 | const __m128 vi0x0123 = _mm_loadu_ps(i0); |
| 138 | const __m128 vk0x0123 = _mm_load_ps(w + 8); |
| 139 | vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi0x0123, vk0x0123)); |
| 140 | |
| 141 | const __m128 vi1x0123 = _mm_loadu_ps(i1); |
| 142 | const __m128 vk1x0123 = _mm_load_ps(w + 16); |
| 143 | vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi1x0123, vk1x0123)); |
| 144 | |
| 145 | const __m128 vi2x0123 = _mm_loadu_ps(i2); |
| 146 | const __m128 vk2x0123 = _mm_load_ps(w + 24); |
| 147 | vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi2x0123, vk2x0123)); |
| 148 | |
| 149 | const __m128 vi3x0123 = _mm_loadu_ps(i3); |
| 150 | const __m128 vk3x0123 = _mm_load_ps(w + 32); |
| 151 | vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi3x0123, vk3x0123)); |
| 152 | |
| 153 | |
| 154 | __m128 vacc0123 = _mm_max_ps(vacc0123p0, vmin); |
| 155 | vacc0123 = _mm_min_ps(vacc0123, vmax); |
| 156 | |
| 157 | if (c & 2) { |
| 158 | _mm_storel_pi((__m64*) output, vacc0123); |
| 159 | vacc0123 = _mm_movehl_ps(vacc0123, vacc0123); |
| 160 | output += 2; |
| 161 | } |
| 162 | if (c & 1) { |
| 163 | _mm_store_ss(output, vacc0123); |
| 164 | output += 1; |
| 165 | } |
| 166 | } |
| 167 | |
| 168 | output = (float*) ((uintptr_t) output + output_increment); |
| 169 | } while (--output_width != 0); |
| 170 | } |