1 // Auto-generated file. Do not edit!
2 // Template: src/f32-qs8-vcvt/neonv8.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2021 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/intrinsics-polyfill.h>
16 #include <xnnpack/vcvt.h>
17
18
xnn_f32_qu8_vcvt_ukernel__neonv8_x8(size_t n,const float * x,uint8_t * y,const union xnn_f32_qu8_cvt_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_f32_qu8_vcvt_ukernel__neonv8_x8(
20 size_t n,
21 const float* x,
22 uint8_t* y,
23 const union xnn_f32_qu8_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
24 {
25 assert(n != 0);
26 assert(n % sizeof(float) == 0);
27 assert(x != NULL);
28 assert(y != NULL);
29
30 const float32x4_t vscale = vld1q_dup_f32(¶ms->neonv8.scale);
31 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->neonv8.output_zero_point);
32 const uint8x8_t voutput_min = vld1_dup_u8(¶ms->neonv8.output_min);
33 const uint8x8_t voutput_max = vld1_dup_u8(¶ms->neonv8.output_max);
34 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
35 float32x4_t vx_lo = vld1q_f32(x); x += 4;
36 float32x4_t vx_hi = vld1q_f32(x); x += 4;
37
38 vx_lo = vmulq_f32(vx_lo, vscale);
39 vx_hi = vmulq_f32(vx_hi, vscale);
40
41 const int32x4_t vacc_lo = vcvtnq_s32_f32(vx_lo);
42 const int32x4_t vacc_hi = vcvtnq_s32_f32(vx_hi);
43
44 int16x8_t vacc = vcombine_s16(vqmovn_s32(vacc_lo), vqmovn_s32(vacc_hi));
45 vacc = vqaddq_s16(vacc, voutput_zero_point);
46
47 uint8x8_t vy = vqmovun_s16(vacc);
48 vy = vmax_u8(vy, voutput_min);
49 vy = vmin_u8(vy, voutput_max);
50 vst1_u8(y, vy); y += 8;
51 }
52 if XNN_UNLIKELY(n != 0) {
53 assert(n >= 1 * sizeof(float));
54 assert(n <= 7 * sizeof(float));
55 float32x4_t vx_lo = vld1q_f32(x);
56 const float* x_hi = (const float*) ((uintptr_t) x + (n & (4 * sizeof(float))));
57 float32x4_t vx_hi = vld1q_f32(x_hi);
58
59 vx_lo = vmulq_f32(vx_lo, vscale);
60 vx_hi = vmulq_f32(vx_hi, vscale);
61
62 const int32x4_t vacc_lo = vcvtnq_s32_f32(vx_lo);
63 const int32x4_t vacc_hi = vcvtnq_s32_f32(vx_hi);
64
65 int16x8_t vacc = vcombine_s16(vqmovn_s32(vacc_lo), vqmovn_s32(vacc_hi));
66 vacc = vqaddq_s16(vacc, voutput_zero_point);
67
68 uint8x8_t vy = vqmovun_s16(vacc);
69 vy = vmax_u8(vy, voutput_min);
70 vy = vmin_u8(vy, voutput_max);
71
72 if (n & (4 * sizeof(float))) {
73 vst1_lane_u32((void*) y, vreinterpret_u32_u8(vy), 0); y += 4;
74 vy = vext_u8(vy, vy, 4);
75 }
76 if (n & (2 * sizeof(float))) {
77 vst1_lane_u16((void*) y, vreinterpret_u16_u8(vy), 0); y += 2;
78 vy = vext_u8(vy, vy, 2);
79 }
80 if (n & (1 * sizeof(float))) {
81 vst1_lane_u8(y, vy, 0);
82 }
83 }
84 }
85