Debian Codegraph @

src/qu8-vcvt/gen/qu8-vcvt-neon-u8.c

raw
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_u8(
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(&params->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(&params->scalar.output_zero_point);
32 for (; batch >= 8 * sizeof(uint8_t); batch -= 8 * sizeof(uint8_t)) {
33 const uint8x8_t vx = vld1_u8(input); input += 8;
34 int16x8_t vacc = vreinterpretq_s16_u16(vsubw_u8(vinput_zero_point, vx));
35 vacc = vshlq_n_s16(vacc, 7);
36 vacc = vqrdmulhq_s16(vacc, vmultiplier);
37 vacc = vqaddq_s16(vacc, voutput_zero_point);
38 const uint8x8_t vy = vqmovun_s16(vacc);
39 vst1_u8(output, vy); output += 8;
40 }
41 if XNN_UNLIKELY(batch != 0) {
42 assert(batch >= 1 * sizeof(uint8_t));
43 assert(batch <= 7 * sizeof(uint8_t));
44
45 const uint8x8_t vx = vld1_u8(input);
46 int16x8_t vacc = vreinterpretq_s16_u16(vsubw_u8(vinput_zero_point, vx));
47 vacc = vshlq_n_s16(vacc, 7);
48 vacc = vqrdmulhq_s16(vacc, vmultiplier);
49 vacc = vqaddq_s16(vacc, voutput_zero_point);
50 uint8x8_t vy = vqmovun_s16(vacc);
51
52 if (batch & (4 * sizeof(uint8_t))) {
53 vst1_lane_u32((void*) output, vreinterpret_u32_u8(vy), 0); output += 4;
54 vy = vext_u8(vy, vy, 4);
55 }
56 if (batch & (2 * sizeof(uint8_t))) {
57 vst1_lane_u16((void*) output, vreinterpret_u16_u8(vy), 0); output += 2;
58 vy = vext_u8(vy, vy, 2);
59 }
60 if (batch & (1 * sizeof(uint8_t))) {
61 vst1_lane_u8(output, vy, 0);
62 }
63 }
64 }