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-scalar.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 <xnnpack/dwconv.h> |
| 13 | #include <xnnpack/math.h> |
| 14 | |
| 15 | |
| 16 | void xnn_f32_dwconv_ukernel_up2x4__scalar( |
| 17 | size_t channels, |
| 18 | size_t output_width, |
| 19 | const float** input, |
| 20 | const float* weights, |
| 21 | float* output, |
| 22 | size_t input_stride, |
| 23 | size_t output_increment, |
| 24 | const union xnn_f32_output_params params[restrict static 1]) |
| 25 | { |
| 26 | assert(channels != 0); |
| 27 | assert(output_width != 0); |
| 28 | |
| 29 | const float vmin = params->scalar.min; |
| 30 | const float vmax = params->scalar.max; |
| 31 | do { |
| 32 | const float* i0 = input[0]; |
| 33 | const float* i1 = input[1]; |
| 34 | const float* i2 = input[2]; |
| 35 | const float* i3 = input[3]; |
| 36 | input = (const float**) ((uintptr_t) input + input_stride); |
| 37 | |
| 38 | size_t c = channels; |
| 39 | const float* w = weights; |
| 40 | for (; c >= 2; c -= 2) { |
| 41 | float vacc0p0 = w[0]; |
| 42 | float vacc1p0 = w[1]; |
| 43 | |
| 44 | |
| 45 | const float vi0x0 = i0[0]; |
| 46 | const float vi0x1 = i0[1]; |
| 47 | i0 += 2; |
| 48 | |
| 49 | const float vk0x0 = w[2]; |
| 50 | vacc0p0 += vi0x0 * vk0x0; |
| 51 | const float vk0x1 = w[3]; |
| 52 | vacc1p0 += vi0x1 * vk0x1; |
| 53 | |
| 54 | const float vi1x0 = i1[0]; |
| 55 | const float vi1x1 = i1[1]; |
| 56 | i1 += 2; |
| 57 | |
| 58 | const float vk1x0 = w[4]; |
| 59 | vacc0p0 += vi1x0 * vk1x0; |
| 60 | const float vk1x1 = w[5]; |
| 61 | vacc1p0 += vi1x1 * vk1x1; |
| 62 | |
| 63 | const float vi2x0 = i2[0]; |
| 64 | const float vi2x1 = i2[1]; |
| 65 | i2 += 2; |
| 66 | |
| 67 | const float vk2x0 = w[6]; |
| 68 | vacc0p0 += vi2x0 * vk2x0; |
| 69 | const float vk2x1 = w[7]; |
| 70 | vacc1p0 += vi2x1 * vk2x1; |
| 71 | |
| 72 | const float vi3x0 = i3[0]; |
| 73 | const float vi3x1 = i3[1]; |
| 74 | i3 += 2; |
| 75 | |
| 76 | const float vk3x0 = w[8]; |
| 77 | vacc0p0 += vi3x0 * vk3x0; |
| 78 | const float vk3x1 = w[9]; |
| 79 | vacc1p0 += vi3x1 * vk3x1; |
| 80 | |
| 81 | w += 10; |
| 82 | |
| 83 | |
| 84 | float vacc0 = math_max_f32(vacc0p0, vmin); |
| 85 | float vacc1 = math_max_f32(vacc1p0, vmin); |
| 86 | |
| 87 | vacc0 = math_min_f32(vacc0, vmax); |
| 88 | vacc1 = math_min_f32(vacc1, vmax); |
| 89 | |
| 90 | output[0] = vacc0; |
| 91 | output[1] = vacc1; |
| 92 | output += 2; |
| 93 | } |
| 94 | for (; c >= 1; c -= 1) { |
| 95 | float vacc0p0 = *w++; |
| 96 | |
| 97 | const float vi0 = *i0++; |
| 98 | const float vk0 = w[1]; |
| 99 | vacc0p0 += vi0 * vk0; |
| 100 | const float vi1 = *i1++; |
| 101 | const float vk1 = w[3]; |
| 102 | vacc0p0 += vi1 * vk1; |
| 103 | const float vi2 = *i2++; |
| 104 | const float vk2 = w[5]; |
| 105 | vacc0p0 += vi2 * vk2; |
| 106 | const float vi3 = *i3++; |
| 107 | const float vk3 = w[7]; |
| 108 | vacc0p0 += vi3 * vk3; |
| 109 | |
| 110 | |
| 111 | float vacc0 = math_max_f32(vacc0p0, vmin); |
| 112 | vacc0 = math_min_f32(vacc0, vmax); |
| 113 | *output++ = vacc0; |
| 114 | } |
| 115 | |
| 116 | output = (float*) ((uintptr_t) output + output_increment); |
| 117 | } while (--output_width != 0); |
| 118 | } |