Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 1 | // Auto-generated file. Do not edit! |
| 2 | // Template: src/f32-vbinary/vopc-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/common.h> |
| 13 | #include <xnnpack/math.h> |
| 14 | #include <xnnpack/vbinary.h> |
| 15 | |
| 16 | |
| 17 | void xnn_f32_vmaxc_ukernel__scalar_x4( |
| 18 | size_t n, |
| 19 | const float* a, |
| 20 | const float* b, |
| 21 | float* y, |
Marat Dukhan | f196d01 | 2020-04-15 11:50:03 -0700 | [diff] [blame] | 22 | const union xnn_f32_default_params params[restrict XNN_MIN_ELEMENTS(1)]) |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 23 | { |
| 24 | assert(n != 0); |
| 25 | assert(n % sizeof(float) == 0); |
Frank Barchard | 0822dde | 2020-07-04 12:47:24 -0700 | [diff] [blame] | 26 | assert(a != NULL); |
| 27 | assert(b != NULL); |
| 28 | assert(y != NULL); |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 29 | |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 30 | |
| 31 | const float vb = *b; |
| 32 | for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) { |
| 33 | const float va0 = a[0]; |
| 34 | const float va1 = a[1]; |
| 35 | const float va2 = a[2]; |
| 36 | const float va3 = a[3]; |
| 37 | a += 4; |
| 38 | |
| 39 | float vy0 = math_max_f32(va0, vb); |
| 40 | float vy1 = math_max_f32(va1, vb); |
| 41 | float vy2 = math_max_f32(va2, vb); |
| 42 | float vy3 = math_max_f32(va3, vb); |
| 43 | |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 44 | |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 45 | |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 46 | y[0] = vy0; |
| 47 | y[1] = vy1; |
| 48 | y[2] = vy2; |
| 49 | y[3] = vy3; |
| 50 | y += 4; |
| 51 | } |
| 52 | if XNN_UNLIKELY(n != 0) { |
| 53 | do { |
| 54 | const float va = *a++; |
| 55 | float vy = math_max_f32(va, vb); |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 56 | *y++ = vy; |
| 57 | n -= sizeof(float); |
| 58 | } while (n != 0); |
| 59 | } |
| 60 | } |