Marat Dukhan | cf5b3c3 | 2020-10-25 19:21:10 -0700 | [diff] [blame^] | 1 | // Auto-generated file. Do not edit! |
| 2 | // Template: src/f32-dwconv2d-chw/3x3s2p1-scalar.c.in |
| 3 | // Generator: tools/xngen |
| 4 | // |
| 5 | // Copyright 2020 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 <xnnpack/dwconv.h> |
| 13 | #include <xnnpack/math.h> |
| 14 | |
| 15 | |
| 16 | void xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_2x1( |
| 17 | size_t input_height, |
| 18 | size_t input_width, |
| 19 | const float* input, |
| 20 | const float* weights, |
| 21 | const float* zero, |
| 22 | float* output, |
| 23 | uint32_t padding_top, |
| 24 | const union xnn_f32_chw_params params[restrict XNN_MIN_ELEMENTS(1)]) |
| 25 | { |
| 26 | assert(input_height != 0); |
| 27 | assert(input_width != 0); |
| 28 | assert(input_width % sizeof(float) == 0); |
| 29 | assert(padding_top >= 0); |
| 30 | assert(padding_top <= 1); |
| 31 | |
| 32 | const float vmin = params->scalar.min; |
| 33 | const float vmax = params->scalar.max; |
| 34 | |
| 35 | const float vbias = weights[0]; |
| 36 | const float vk00 = weights[1]; |
| 37 | const float vk01 = weights[2]; |
| 38 | const float vk02 = weights[3]; |
| 39 | const float vk10 = weights[4]; |
| 40 | const float vk11 = weights[5]; |
| 41 | const float vk12 = weights[6]; |
| 42 | const float vk20 = weights[7]; |
| 43 | const float vk21 = weights[8]; |
| 44 | const float vk22 = weights[9]; |
| 45 | |
| 46 | const size_t output_width = round_down_po2((input_width + (2 /* padding */ - 3 /* kernel size */ + 2 /* subsampling */) * sizeof(float)) / 2, sizeof(float)); |
| 47 | |
| 48 | const float* i0 = (const float*) ((uintptr_t) input - ((-padding_top) & input_width)); |
| 49 | const float* i1 = (const float*) ((uintptr_t) i0 + input_width); |
| 50 | if XNN_UNPREDICTABLE(padding_top != 0) { |
| 51 | i0 = zero; |
| 52 | } |
| 53 | const float* i2 = (const float*) ((uintptr_t) i1 + input_width); |
| 54 | const float* i3 = (const float*) ((uintptr_t) i2 + input_width); |
| 55 | const float* i4 = (const float*) ((uintptr_t) i3 + input_width); |
| 56 | |
| 57 | float* o0 = output; |
| 58 | float* o1 = (float*) ((uintptr_t) o0 + output_width); |
| 59 | |
| 60 | size_t padded_input_height = input_height + padding_top + 1 /* padding bottom */; |
| 61 | size_t output_height = (padded_input_height - 3 /* kernel size */ + 2 /* subsampling */) / 2; |
| 62 | do { |
| 63 | if XNN_UNPREDICTABLE(padded_input_height < 4) { |
| 64 | i2 = zero; |
| 65 | } |
| 66 | if XNN_UNPREDICTABLE(padded_input_height < 5) { |
| 67 | i3 = zero; |
| 68 | o1 = o0; |
| 69 | } |
| 70 | if XNN_UNPREDICTABLE(padded_input_height < 6) { |
| 71 | i4 = zero; |
| 72 | } |
| 73 | |
| 74 | float vi0x0 = 0.0f; |
| 75 | float vi1x0 = 0.0f; |
| 76 | float vi2x0 = 0.0f; |
| 77 | float vi3x0 = 0.0f; |
| 78 | float vi4x0 = 0.0f; |
| 79 | |
| 80 | size_t w = input_width; |
| 81 | for (; w >= 2 * sizeof(float); w -= 2 * sizeof(float)) { |
| 82 | const float vi0x1 = i0[0]; |
| 83 | const float vi1x1 = i1[0]; |
| 84 | const float vi2x1 = i2[0]; |
| 85 | const float vi3x1 = i3[0]; |
| 86 | const float vi4x1 = i4[0]; |
| 87 | |
| 88 | float vo0p0 = vbias + vi0x0 * vk00; |
| 89 | float vo1p0 = vbias + vi2x0 * vk00; |
| 90 | vo0p0 += vi1x0 * vk10; |
| 91 | vo1p0 += vi3x0 * vk10; |
| 92 | vo0p0 += vi2x0 * vk20; |
| 93 | vo1p0 += vi4x0 * vk20; |
| 94 | |
| 95 | const float vi0x2 = i0[1]; |
| 96 | i0 += 2; |
| 97 | const float vi1x2 = i1[1]; |
| 98 | i1 += 2; |
| 99 | const float vi2x2 = i2[1]; |
| 100 | i2 += 2; |
| 101 | const float vi3x2 = i3[1]; |
| 102 | i3 += 2; |
| 103 | const float vi4x2 = i4[1]; |
| 104 | i4 += 2; |
| 105 | |
| 106 | vo0p0 += vi0x1 * vk01; |
| 107 | vo1p0 += vi2x1 * vk01; |
| 108 | vo0p0 += vi1x1 * vk11; |
| 109 | vo1p0 += vi3x1 * vk11; |
| 110 | vo0p0 += vi2x1 * vk21; |
| 111 | vo1p0 += vi4x1 * vk21; |
| 112 | |
| 113 | vi0x0 = vi0x2; |
| 114 | vi1x0 = vi1x2; |
| 115 | vi2x0 = vi2x2; |
| 116 | vi3x0 = vi3x2; |
| 117 | vi4x0 = vi4x2; |
| 118 | |
| 119 | vo0p0 += vi0x2 * vk02; |
| 120 | vo1p0 += vi2x2 * vk02; |
| 121 | vo0p0 += vi1x2 * vk12; |
| 122 | vo1p0 += vi3x2 * vk12; |
| 123 | vo0p0 += vi2x2 * vk22; |
| 124 | vo1p0 += vi4x2 * vk22; |
| 125 | |
| 126 | |
| 127 | float vo0 = math_max_f32(vo0p0, vmin); |
| 128 | float vo1 = math_max_f32(vo1p0, vmin); |
| 129 | |
| 130 | vo0 = math_min_f32(vo0, vmax); |
| 131 | vo1 = math_min_f32(vo1, vmax); |
| 132 | |
| 133 | *o1++ = vo1; |
| 134 | *o0++ = vo0; |
| 135 | } |
| 136 | // Potentially process the last pixel. |
| 137 | assert(w <= 1 * sizeof(float)); |
| 138 | if (w != 0) { |
| 139 | const float vi0x1 = *i0++; |
| 140 | const float vi1x1 = *i1++; |
| 141 | const float vi2x1 = *i2++; |
| 142 | const float vi3x1 = *i3++; |
| 143 | const float vi4x1 = *i4++; |
| 144 | |
| 145 | float vo0p0 = vbias + vi0x0 * vk00; |
| 146 | float vo1p0 = vbias + vi2x0 * vk00; |
| 147 | vo0p0 += vi1x0 * vk10; |
| 148 | vo1p0 += vi3x0 * vk10; |
| 149 | vo0p0 += vi2x0 * vk20; |
| 150 | vo1p0 += vi4x0 * vk20; |
| 151 | |
| 152 | vo0p0 += vi0x1 * vk01; |
| 153 | vo1p0 += vi2x1 * vk01; |
| 154 | vo0p0 += vi1x1 * vk11; |
| 155 | vo1p0 += vi3x1 * vk11; |
| 156 | vo0p0 += vi2x1 * vk21; |
| 157 | vo1p0 += vi4x1 * vk21; |
| 158 | |
| 159 | |
| 160 | float vo0 = math_max_f32(vo0p0, vmin); |
| 161 | float vo1 = math_max_f32(vo1p0, vmin); |
| 162 | |
| 163 | vo0 = math_min_f32(vo0, vmax); |
| 164 | vo1 = math_min_f32(vo1, vmax); |
| 165 | |
| 166 | *o1++ = vo1; |
| 167 | *o0++ = vo0; |
| 168 | } |
| 169 | |
| 170 | i0 = (const float*) ((uintptr_t) i3); |
| 171 | i1 = (const float*) ((uintptr_t) i4); |
| 172 | i2 = (const float*) ((uintptr_t) i1 + input_width); |
| 173 | i3 = (const float*) ((uintptr_t) i2 + input_width); |
| 174 | i4 = (const float*) ((uintptr_t) i3 + input_width); |
| 175 | |
| 176 | o0 = o1; |
| 177 | o1 = (float*) ((uintptr_t) o0 + output_width); |
| 178 | |
| 179 | output_height = doz(output_height, 2); |
| 180 | padded_input_height = doz(padded_input_height, 4); |
| 181 | } while (output_height != 0); |
| 182 | } |