Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 1 | // Auto-generated file. Do not edit! |
| 2 | // Template: src/f32-vbinary/vop-neon.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 <arm_neon.h> |
| 13 | |
| 14 | #include <xnnpack/common.h> |
| 15 | #include <xnnpack/vbinary.h> |
| 16 | |
| 17 | |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 18 | void xnn_f32_vmul_minmax_ukernel__neon_x8( |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 19 | size_t n, |
| 20 | const float* a, |
| 21 | const float* b, |
| 22 | float* y, |
Marat Dukhan | f196d01 | 2020-04-15 11:50:03 -0700 | [diff] [blame] | 23 | const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 24 | { |
| 25 | assert(n != 0); |
| 26 | assert(n % sizeof(float) == 0); |
| 27 | |
| 28 | const float32x4_t vy_min = vld1q_dup_f32(¶ms->scalar.min); |
| 29 | const float32x4_t vy_max = vld1q_dup_f32(¶ms->scalar.max); |
| 30 | |
| 31 | for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) { |
| 32 | const float32x4_t va0123 = vld1q_f32(a); a += 4; |
| 33 | const float32x4_t vb0123 = vld1q_f32(b); b += 4; |
| 34 | const float32x4_t va4567 = vld1q_f32(a); a += 4; |
| 35 | const float32x4_t vb4567 = vld1q_f32(b); b += 4; |
| 36 | |
| 37 | float32x4_t vy0123 = vmulq_f32(va0123, vb0123); |
| 38 | float32x4_t vy4567 = vmulq_f32(va4567, vb4567); |
| 39 | |
| 40 | vy0123 = vmaxq_f32(vy0123, vy_min); |
| 41 | vy4567 = vmaxq_f32(vy4567, vy_min); |
| 42 | |
| 43 | vy0123 = vminq_f32(vy0123, vy_max); |
| 44 | vy4567 = vminq_f32(vy4567, vy_max); |
| 45 | |
| 46 | vst1q_f32(y, vy0123); y += 4; |
| 47 | vst1q_f32(y, vy4567); y += 4; |
| 48 | } |
| 49 | for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) { |
| 50 | const float32x4_t va0123 = vld1q_f32(a); a += 4; |
| 51 | const float32x4_t vb0123 = vld1q_f32(b); b += 4; |
| 52 | |
| 53 | float32x4_t vy0123 = vmulq_f32(va0123, vb0123); |
| 54 | vy0123 = vmaxq_f32(vy0123, vy_min); |
| 55 | vy0123 = vminq_f32(vy0123, vy_max); |
| 56 | vst1q_f32(y, vy0123); y += 4; |
| 57 | } |
| 58 | if XNN_UNLIKELY(n != 0) { |
| 59 | const float32x4_t va0123 = vld1q_f32(a); |
| 60 | const float32x4_t vb0123 = vld1q_f32(b); |
| 61 | |
| 62 | float32x4_t vy0123 = vmulq_f32(va0123, vb0123); |
| 63 | vy0123 = vmaxq_f32(vy0123, vy_min); |
| 64 | vy0123 = vminq_f32(vy0123, vy_max); |
| 65 | |
| 66 | float32x2_t vy01 = vget_low_f32(vy0123); |
| 67 | if (n & (2 * sizeof(float))) { |
| 68 | vst1_f32(y, vy01); y += 2; |
| 69 | vy01 = vget_high_f32(vy0123); |
| 70 | } |
| 71 | if (n & (1 * sizeof(float))) { |
| 72 | vst1_lane_f32(y, vy01, 0); |
| 73 | } |
| 74 | } |
| 75 | } |