| 1 |
// Auto-generated file. Do not edit! |
| 2 |
// Template: src/qs8-vcvt/neon.c.in |
| 3 |
// Generator: tools/xngen |
| 4 |
// |
| 5 |
// Copyright 2022 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/vcvt.h" |
| 16 |
|
| 17 |
|
| 18 |
void xnn_qu8_vcvt_ukernel__neon_u16( |
| 19 |
size_t batch, |
| 20 |
const uint8_t* input, |
| 21 |
uint8_t* output, |
| 22 |
const struct xnn_qu8_cvt_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 != NULL); |
| 27 |
assert(output != NULL); |
| 28 |
|
| 29 |
const uint16x8_t vinput_zero_point = vld1q_dup_u16(¶ms->scalar.input_zero_point); |
| 30 |
const int16x8_t vmultiplier = vdupq_n_s16(-params->scalar.multiplier); |
| 31 |
const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->scalar.output_zero_point); |
| 32 |
for (; batch >= 16 * sizeof(uint8_t); batch -= 16 * sizeof(uint8_t)) { |
| 33 |
const uint8x16_t vx0 = vld1q_u8(input); input += 16; |
| 34 |
|
| 35 |
int16x8_t vacc0 = vreinterpretq_s16_u16(vsubw_u8(vinput_zero_point, vget_low_u8(vx0))); |
| 36 |
int16x8_t vacc1 = vreinterpretq_s16_u16(vsubw_u8(vinput_zero_point, vget_high_u8(vx0))); |
| 37 |
|
| 38 |
vacc0 = vshlq_n_s16(vacc0, 7); |
| 39 |
vacc1 = vshlq_n_s16(vacc1, 7); |
| 40 |
|
| 41 |
vacc0 = vqrdmulhq_s16(vacc0, vmultiplier); |
| 42 |
vacc1 = vqrdmulhq_s16(vacc1, vmultiplier); |
| 43 |
|
| 44 |
vacc0 = vqaddq_s16(vacc0, voutput_zero_point); |
| 45 |
vacc1 = vqaddq_s16(vacc1, voutput_zero_point); |
| 46 |
|
| 47 |
const uint8x16_t vy0 = vcombine_u8(vqmovun_s16(vacc0), vqmovun_s16(vacc1)); |
| 48 |
|
| 49 |
vst1q_u8(output, vy0); output += 16; |
| 50 |
} |
| 51 |
for (; batch >= 8 * sizeof(uint8_t); batch -= 8 * sizeof(uint8_t)) { |
| 52 |
const uint8x8_t vx = vld1_u8(input); input += 8; |
| 53 |
int16x8_t vacc = vreinterpretq_s16_u16(vsubw_u8(vinput_zero_point, vx)); |
| 54 |
vacc = vshlq_n_s16(vacc, 7); |
| 55 |
vacc = vqrdmulhq_s16(vacc, vmultiplier); |
| 56 |
vacc = vqaddq_s16(vacc, voutput_zero_point); |
| 57 |
const uint8x8_t vy = vqmovun_s16(vacc); |
| 58 |
vst1_u8(output, vy); output += 8; |
| 59 |
} |
| 60 |
if XNN_UNLIKELY(batch != 0) { |
| 61 |
assert(batch >= 1 * sizeof(uint8_t)); |
| 62 |
assert(batch <= 7 * sizeof(uint8_t)); |
| 63 |
|
| 64 |
const uint8x8_t vx = vld1_u8(input); |
| 65 |
int16x8_t vacc = vreinterpretq_s16_u16(vsubw_u8(vinput_zero_point, vx)); |
| 66 |
vacc = vshlq_n_s16(vacc, 7); |
| 67 |
vacc = vqrdmulhq_s16(vacc, vmultiplier); |
| 68 |
vacc = vqaddq_s16(vacc, voutput_zero_point); |
| 69 |
uint8x8_t vy = vqmovun_s16(vacc); |
| 70 |
|
| 71 |
if (batch & (4 * sizeof(uint8_t))) { |
| 72 |
vst1_lane_u32((void*) output, vreinterpret_u32_u8(vy), 0); output += 4; |
| 73 |
vy = vext_u8(vy, vy, 4); |
| 74 |
} |
| 75 |
if (batch & (2 * sizeof(uint8_t))) { |
| 76 |
vst1_lane_u16((void*) output, vreinterpret_u16_u8(vy), 0); output += 2; |
| 77 |
vy = vext_u8(vy, vy, 2); |
| 78 |
} |
| 79 |
if (batch & (1 * sizeof(uint8_t))) { |
| 80 |
vst1_lane_u8(output, vy, 0); |
| 81 |
} |
| 82 |
} |
| 83 |
} |