| 1 |
// Auto-generated file. Do not edit! |
| 2 |
// Template: src/qs8-vadd/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 |
|
| 14 |
#include "xnnpack/vbinary.h" |
| 15 |
|
| 16 |
|
| 17 |
void xnn_qu8_vadd_minmax_ukernel__neon_ld128_u16( |
| 18 |
size_t batch, |
| 19 |
const uint8_t* input_a, |
| 20 |
const uint8_t* input_b, |
| 21 |
uint8_t* output, |
| 22 |
const struct xnn_qu8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS |
| 23 |
{ |
| 24 |
assert(batch != 0); |
| 25 |
assert(batch % sizeof(uint8_t) == 0); |
| 26 |
assert(input_a != NULL); |
| 27 |
assert(input_b != NULL); |
| 28 |
assert(output != NULL); |
| 29 |
|
| 30 |
#if XNN_ARCH_ARM64 |
| 31 |
const uint8x16_t va_zero_point = vld1q_dup_u8(¶ms->scalar.a_zero_point); |
| 32 |
const uint8x16_t vb_zero_point = vld1q_dup_u8(¶ms->scalar.b_zero_point); |
| 33 |
#else |
| 34 |
const uint8x8_t va_zero_point = vld1_dup_u8(¶ms->scalar.a_zero_point); |
| 35 |
const uint8x8_t vb_zero_point = vld1_dup_u8(¶ms->scalar.b_zero_point); |
| 36 |
#endif |
| 37 |
const int32x4_t va_multiplier = vld1q_dup_s32(¶ms->scalar.a_multiplier); |
| 38 |
const int32x4_t vb_multiplier = vld1q_dup_s32(¶ms->scalar.b_multiplier); |
| 39 |
const int32x4_t vright_shift = vdupq_n_s32(-params->scalar.shift); |
| 40 |
const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->scalar.output_zero_point); |
| 41 |
const uint8x16_t voutput_min = vld1q_dup_u8(¶ms->scalar.output_min); |
| 42 |
const uint8x16_t voutput_max = vld1q_dup_u8(¶ms->scalar.output_max); |
| 43 |
|
| 44 |
for (; batch >= 16 * sizeof(uint8_t); batch -= 16 * sizeof(uint8_t)) { |
| 45 |
const uint8x16_t va0123456789ABCDEF = vld1q_u8(input_a); input_a += 16; |
| 46 |
const uint8x16_t vb0123456789ABCDEF = vld1q_u8(input_b); input_b += 16; |
| 47 |
|
| 48 |
#if XNN_ARCH_ARM64 |
| 49 |
const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(vget_low_u8(va0123456789ABCDEF), vget_low_u8(va_zero_point))); |
| 50 |
const int16x8_t vxa89ABCDEF = vreinterpretq_s16_u16(vsubl_high_u8(va0123456789ABCDEF, va_zero_point)); |
| 51 |
const int16x8_t vxb01234567 = vreinterpretq_s16_u16(vsubl_u8(vget_low_u8(vb0123456789ABCDEF), vget_low_u8(vb_zero_point))); |
| 52 |
const int16x8_t vxb89ABCDEF = vreinterpretq_s16_u16(vsubl_high_u8(vb0123456789ABCDEF, vb_zero_point)); |
| 53 |
#else // !XNN_ARCH_ARM64 |
| 54 |
const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(vget_low_u8(va0123456789ABCDEF), va_zero_point)); |
| 55 |
const int16x8_t vxa89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vget_high_u8(va0123456789ABCDEF), va_zero_point)); |
| 56 |
const int16x8_t vxb01234567 = vreinterpretq_s16_u16(vsubl_u8(vget_low_u8(vb0123456789ABCDEF), vb_zero_point)); |
| 57 |
const int16x8_t vxb89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vget_high_u8(vb0123456789ABCDEF), vb_zero_point)); |
| 58 |
#endif // XNN_ARCH_ARM64 |
| 59 |
|
| 60 |
int32x4_t vacc0123 = vmulq_s32(vmovl_s16(vget_low_s16(vxa01234567)), va_multiplier); |
| 61 |
int32x4_t vacc4567 = vmulq_s32(vmovl_s16(vget_high_s16(vxa01234567)), va_multiplier); |
| 62 |
int32x4_t vacc89AB = vmulq_s32(vmovl_s16(vget_low_s16(vxa89ABCDEF)), va_multiplier); |
| 63 |
int32x4_t vaccCDEF = vmulq_s32(vmovl_s16(vget_high_s16(vxa89ABCDEF)), va_multiplier); |
| 64 |
|
| 65 |
vacc0123 = vmlaq_s32(vacc0123, vmovl_s16(vget_low_s16(vxb01234567)), vb_multiplier); |
| 66 |
vacc4567 = vmlaq_s32(vacc4567, vmovl_s16(vget_high_s16(vxb01234567)), vb_multiplier); |
| 67 |
vacc89AB = vmlaq_s32(vacc89AB, vmovl_s16(vget_low_s16(vxb89ABCDEF)), vb_multiplier); |
| 68 |
vaccCDEF = vmlaq_s32(vaccCDEF, vmovl_s16(vget_high_s16(vxb89ABCDEF)), vb_multiplier); |
| 69 |
|
| 70 |
vacc0123 = vrshlq_s32(vacc0123, vright_shift); |
| 71 |
vacc4567 = vrshlq_s32(vacc4567, vright_shift); |
| 72 |
vacc89AB = vrshlq_s32(vacc89AB, vright_shift); |
| 73 |
vaccCDEF = vrshlq_s32(vaccCDEF, vright_shift); |
| 74 |
|
| 75 |
const int16x8_t vacc01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)), voutput_zero_point); |
| 76 |
const int16x8_t vacc89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF)), voutput_zero_point); |
| 77 |
|
| 78 |
uint8x16_t vout0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc01234567), vqmovun_s16(vacc89ABCDEF)); |
| 79 |
|
| 80 |
vout0123456789ABCDEF = vmaxq_u8(vout0123456789ABCDEF, voutput_min); |
| 81 |
|
| 82 |
vout0123456789ABCDEF = vminq_u8(vout0123456789ABCDEF, voutput_max); |
| 83 |
|
| 84 |
vst1q_u8(output, vout0123456789ABCDEF); output += 16; |
| 85 |
} |
| 86 |
if XNN_UNLIKELY(batch != 0) { |
| 87 |
do { |
| 88 |
const uint8x8_t va01234567 = vld1_u8(input_a); input_a += 8; |
| 89 |
const uint8x8_t vb01234567 = vld1_u8(input_b); input_b += 8; |
| 90 |
|
| 91 |
#if XNN_ARCH_ARM64 |
| 92 |
const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(va01234567, vget_low_u8(va_zero_point))); |
| 93 |
const int16x8_t vxb01234567 = vreinterpretq_s16_u16(vsubl_u8(vb01234567, vget_low_u8(vb_zero_point))); |
| 94 |
#else // !XNN_ARCH_ARM64 |
| 95 |
const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(va01234567, va_zero_point)); |
| 96 |
const int16x8_t vxb01234567 = vreinterpretq_s16_u16(vsubl_u8(vb01234567, vb_zero_point)); |
| 97 |
#endif // XNN_ARCH_ARM64 |
| 98 |
|
| 99 |
int32x4_t vacc0123 = vmulq_s32(vmovl_s16(vget_low_s16(vxa01234567)), va_multiplier); |
| 100 |
int32x4_t vacc4567 = vmulq_s32(vmovl_s16(vget_high_s16(vxa01234567)), va_multiplier); |
| 101 |
|
| 102 |
vacc0123 = vmlaq_s32(vacc0123, vmovl_s16(vget_low_s16(vxb01234567)), vb_multiplier); |
| 103 |
vacc4567 = vmlaq_s32(vacc4567, vmovl_s16(vget_high_s16(vxb01234567)), vb_multiplier); |
| 104 |
|
| 105 |
vacc0123 = vrshlq_s32(vacc0123, vright_shift); |
| 106 |
vacc4567 = vrshlq_s32(vacc4567, vright_shift); |
| 107 |
|
| 108 |
const int16x8_t vacc01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)), voutput_zero_point); |
| 109 |
|
| 110 |
uint8x8_t vout01234567 = vqmovun_s16(vacc01234567); |
| 111 |
vout01234567 = vmax_u8(vout01234567, vget_low_u8(voutput_min)); |
| 112 |
vout01234567 = vmin_u8(vout01234567, vget_low_u8(voutput_max)); |
| 113 |
|
| 114 |
if XNN_LIKELY(batch >= (8 * sizeof(uint8_t))) { |
| 115 |
vst1_u8(output, vout01234567); output += 8; |
| 116 |
batch -= 8 * sizeof(uint8_t); |
| 117 |
} else { |
| 118 |
if (batch & (4 * sizeof(uint8_t))) { |
| 119 |
vst1_lane_u32((void*) output, vreinterpret_u32_u8(vout01234567), 0); output += 4; |
| 120 |
vout01234567 = vext_u8(vout01234567, vout01234567, 4); |
| 121 |
} |
| 122 |
if (batch & (2 * sizeof(uint8_t))) { |
| 123 |
vst1_lane_u16((void*) output, vreinterpret_u16_u8(vout01234567), 0); output += 2; |
| 124 |
vout01234567 = vext_u8(vout01234567, vout01234567, 2); |
| 125 |
} |
| 126 |
if (batch & (1 * sizeof(uint8_t))) { |
| 127 |
vst1_lane_u8(output, vout01234567, 0); |
| 128 |
} |
| 129 |
batch = 0; |
| 130 |
} |
| 131 |
} while (batch != 0); |
| 132 |
} |
| 133 |
} |