Frank Barchard | 0a3093c | 2021-08-31 09:58:11 -0700 | [diff] [blame] | 1 | // Auto-generated file. Do not edit! |
| 2 | // Template: src/qs8-vaddc/neon.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 <arm_neon.h> |
| 13 | |
Marat Dukhan | 6428725 | 2021-09-07 16:20:03 -0700 | [diff] [blame] | 14 | #include <xnnpack/vaddsub.h> |
Frank Barchard | 0a3093c | 2021-08-31 09:58:11 -0700 | [diff] [blame] | 15 | |
| 16 | |
| 17 | void xnn_qu8_vaddc_minmax_ukernel__neon_ld64_x32( |
| 18 | size_t n, |
| 19 | const uint8_t* input_a, |
| 20 | const uint8_t* input_b, |
| 21 | uint8_t* output, |
Marat Dukhan | 7be427a | 2021-12-13 23:38:20 -0800 | [diff] [blame] | 22 | const union xnn_qu8_addsub_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS |
Frank Barchard | 0a3093c | 2021-08-31 09:58:11 -0700 | [diff] [blame] | 23 | { |
| 24 | const uint8x8_t va_zero_point = vld1_dup_u8(¶ms->neon.a_zero_point); |
| 25 | const int32x4_t va_multiplier = vld1q_dup_s32(¶ms->neon.a_multiplier); |
| 26 | const int32x4_t vright_shift = vld1q_dup_s32(¶ms->neon.right_shift); |
| 27 | const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->neon.output_zero_point); |
| 28 | const uint8x16_t voutput_min = vld1q_dup_u8(¶ms->neon.output_min); |
| 29 | const uint8x16_t voutput_max = vld1q_dup_u8(¶ms->neon.output_max); |
| 30 | |
| 31 | const int32_t vxb = (int32_t) *input_b - (int32_t) params->neon.b_zero_point; |
| 32 | const int32_t vb = params->neon.b_multiplier; |
| 33 | const int32x4_t vbias = vdupq_n_s32(vxb * vb); |
| 34 | |
| 35 | for (; n >= 32 * sizeof(uint8_t); n -= 32 * sizeof(uint8_t)) { |
| 36 | const uint8x8_t va01234567 = vld1_u8(input_a); input_a += 8; |
| 37 | const uint8x8_t va89ABCDEF = vld1_u8(input_a); input_a += 8; |
| 38 | const uint8x8_t vaGHIJKLMN = vld1_u8(input_a); input_a += 8; |
| 39 | const uint8x8_t vaOPQRSTUV = vld1_u8(input_a); input_a += 8; |
| 40 | |
| 41 | const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(va01234567, va_zero_point)); |
| 42 | const int16x8_t vxa89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(va89ABCDEF, va_zero_point)); |
| 43 | const int16x8_t vxaGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vaGHIJKLMN, va_zero_point)); |
| 44 | const int16x8_t vxaOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vaOPQRSTUV, va_zero_point)); |
| 45 | |
| 46 | int32x4_t vacc0123 = vmlaq_s32(vbias, vmovl_s16(vget_low_s16(vxa01234567)), va_multiplier); |
| 47 | int32x4_t vacc4567 = vmlaq_s32(vbias, vmovl_s16(vget_high_s16(vxa01234567)), va_multiplier); |
| 48 | int32x4_t vacc89AB = vmlaq_s32(vbias, vmovl_s16(vget_low_s16(vxa89ABCDEF)), va_multiplier); |
| 49 | int32x4_t vaccCDEF = vmlaq_s32(vbias, vmovl_s16(vget_high_s16(vxa89ABCDEF)), va_multiplier); |
| 50 | int32x4_t vaccGHIJ = vmlaq_s32(vbias, vmovl_s16(vget_low_s16(vxaGHIJKLMN)), va_multiplier); |
| 51 | int32x4_t vaccKLMN = vmlaq_s32(vbias, vmovl_s16(vget_high_s16(vxaGHIJKLMN)), va_multiplier); |
| 52 | int32x4_t vaccOPQR = vmlaq_s32(vbias, vmovl_s16(vget_low_s16(vxaOPQRSTUV)), va_multiplier); |
| 53 | int32x4_t vaccSTUV = vmlaq_s32(vbias, vmovl_s16(vget_high_s16(vxaOPQRSTUV)), va_multiplier); |
| 54 | |
| 55 | vacc0123 = vrshlq_s32(vacc0123, vright_shift); |
| 56 | vacc4567 = vrshlq_s32(vacc4567, vright_shift); |
| 57 | vacc89AB = vrshlq_s32(vacc89AB, vright_shift); |
| 58 | vaccCDEF = vrshlq_s32(vaccCDEF, vright_shift); |
| 59 | vaccGHIJ = vrshlq_s32(vaccGHIJ, vright_shift); |
| 60 | vaccKLMN = vrshlq_s32(vaccKLMN, vright_shift); |
| 61 | vaccOPQR = vrshlq_s32(vaccOPQR, vright_shift); |
| 62 | vaccSTUV = vrshlq_s32(vaccSTUV, vright_shift); |
| 63 | |
| 64 | const int16x8_t vacc01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)), voutput_zero_point); |
| 65 | const int16x8_t vacc89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF)), voutput_zero_point); |
| 66 | const int16x8_t vaccGHIJKLMN = vqaddq_s16(vcombine_s16(vqmovn_s32(vaccGHIJ), vqmovn_s32(vaccKLMN)), voutput_zero_point); |
| 67 | const int16x8_t vaccOPQRSTUV = vqaddq_s16(vcombine_s16(vqmovn_s32(vaccOPQR), vqmovn_s32(vaccSTUV)), voutput_zero_point); |
| 68 | |
| 69 | uint8x16_t vout0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc01234567), vqmovun_s16(vacc89ABCDEF)); |
| 70 | uint8x16_t voutGHIJKLMNOPQRSTUV = vcombine_u8(vqmovun_s16(vaccGHIJKLMN), vqmovun_s16(vaccOPQRSTUV)); |
| 71 | |
| 72 | vout0123456789ABCDEF = vmaxq_u8(vout0123456789ABCDEF, voutput_min); |
| 73 | voutGHIJKLMNOPQRSTUV = vmaxq_u8(voutGHIJKLMNOPQRSTUV, voutput_min); |
| 74 | |
| 75 | vout0123456789ABCDEF = vminq_u8(vout0123456789ABCDEF, voutput_max); |
| 76 | voutGHIJKLMNOPQRSTUV = vminq_u8(voutGHIJKLMNOPQRSTUV, voutput_max); |
| 77 | |
| 78 | vst1q_u8(output, vout0123456789ABCDEF); output += 16; |
| 79 | vst1q_u8(output, voutGHIJKLMNOPQRSTUV); output += 16; |
| 80 | } |
| 81 | if XNN_UNLIKELY(n != 0) { |
| 82 | do { |
| 83 | const uint8x8_t va01234567 = vld1_u8(input_a); input_a += 8; |
| 84 | |
| 85 | const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(va01234567, va_zero_point)); |
| 86 | |
| 87 | int32x4_t vacc0123 = vmlaq_s32(vbias, vmovl_s16(vget_low_s16(vxa01234567)), va_multiplier); |
| 88 | int32x4_t vacc4567 = vmlaq_s32(vbias, vmovl_s16(vget_high_s16(vxa01234567)), va_multiplier); |
| 89 | |
| 90 | vacc0123 = vrshlq_s32(vacc0123, vright_shift); |
| 91 | vacc4567 = vrshlq_s32(vacc4567, vright_shift); |
| 92 | |
| 93 | const int16x8_t vacc01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)), voutput_zero_point); |
| 94 | |
| 95 | uint8x8_t vout01234567 = vqmovun_s16(vacc01234567); |
| 96 | vout01234567 = vmax_u8(vout01234567, vget_low_u8(voutput_min)); |
| 97 | vout01234567 = vmin_u8(vout01234567, vget_low_u8(voutput_max)); |
| 98 | |
| 99 | if XNN_LIKELY(n >= (8 * sizeof(uint8_t))) { |
| 100 | vst1_u8(output, vout01234567); output += 8; |
| 101 | n -= 8 * sizeof(uint8_t); |
| 102 | } else { |
| 103 | if (n & (4 * sizeof(uint8_t))) { |
Marat Dukhan | 5f7cf55 | 2021-11-25 17:37:03 -0800 | [diff] [blame] | 104 | vst1_lane_u32((void*) output, vreinterpret_u32_u8(vout01234567), 0); output += 4; |
Frank Barchard | 0a3093c | 2021-08-31 09:58:11 -0700 | [diff] [blame] | 105 | vout01234567 = vext_u8(vout01234567, vout01234567, 4); |
| 106 | } |
| 107 | if (n & (2 * sizeof(uint8_t))) { |
Marat Dukhan | 5f7cf55 | 2021-11-25 17:37:03 -0800 | [diff] [blame] | 108 | vst1_lane_u16((void*) output, vreinterpret_u16_u8(vout01234567), 0); output += 2; |
Frank Barchard | 0a3093c | 2021-08-31 09:58:11 -0700 | [diff] [blame] | 109 | vout01234567 = vext_u8(vout01234567, vout01234567, 2); |
| 110 | } |
| 111 | if (n & (1 * sizeof(uint8_t))) { |
| 112 | vst1_lane_u8(output, vout01234567, 0); |
| 113 | } |
| 114 | n = 0; |
| 115 | } |
| 116 | } while (n != 0); |
| 117 | } |
| 118 | } |