1 // Auto-generated file. Do not edit!
2 // Template: src/f32-qs8-vcvt/neon.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__neon_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__neon_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->neon.scale);
31 const float32x4_t vmagic_bias = vld1q_dup_f32(¶ms->neon.magic_bias);
32 const int32x4_t vmagic_bias_less_zero_point = vld1q_dup_s32(¶ms->neon.magic_bias_less_zero_point);
33 const uint8x8_t voutput_min = vld1_dup_u8(¶ms->neon.output_min);
34 const uint8x8_t voutput_max = vld1_dup_u8(¶ms->neon.output_max);
35 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
36 float32x4_t vx_lo = vld1q_f32(x); x += 4;
37 float32x4_t vx_hi = vld1q_f32(x); x += 4;
38
39 vx_lo = vmulq_f32(vx_lo, vscale);
40 vx_hi = vmulq_f32(vx_hi, vscale);
41
42 vx_lo = vaddq_f32(vx_lo, vmagic_bias);
43 vx_hi = vaddq_f32(vx_hi, vmagic_bias);
44
45 const int32x4_t vacc_lo = vqsubq_s32(vreinterpretq_s32_f32(vx_lo), vmagic_bias_less_zero_point);
46 const int32x4_t vacc_hi = vqsubq_s32(vreinterpretq_s32_f32(vx_hi), vmagic_bias_less_zero_point);
47
48 const int16x8_t vacc = vcombine_s16(vqmovn_s32(vacc_lo), vqmovn_s32(vacc_hi));
49
50 uint8x8_t vy = vqmovun_s16(vacc);
51 vy = vmax_u8(vy, voutput_min);
52 vy = vmin_u8(vy, voutput_max);
53 vst1_u8(y, vy); y += 8;
54 }
55 if XNN_UNLIKELY(n != 0) {
56 assert(n >= 1 * sizeof(float));
57 assert(n <= 7 * sizeof(float));
58 float32x4_t vx_lo = vld1q_f32(x);
59 const float* x_hi = (const float*) ((uintptr_t) x + (n & (4 * sizeof(float))));
60 float32x4_t vx_hi = vld1q_f32(x_hi);
61
62 vx_lo = vmulq_f32(vx_lo, vscale);
63 vx_hi = vmulq_f32(vx_hi, vscale);
64
65 vx_lo = vaddq_f32(vx_lo, vmagic_bias);
66 vx_hi = vaddq_f32(vx_hi, vmagic_bias);
67
68 const int32x4_t vacc_lo = vqsubq_s32(vreinterpretq_s32_f32(vx_lo), vmagic_bias_less_zero_point);
69 const int32x4_t vacc_hi = vqsubq_s32(vreinterpretq_s32_f32(vx_hi), vmagic_bias_less_zero_point);
70
71 const int16x8_t vacc = vcombine_s16(vqmovn_s32(vacc_lo), vqmovn_s32(vacc_hi));
72
73 uint8x8_t vy = vqmovun_s16(vacc);
74 vy = vmax_u8(vy, voutput_min);
75 vy = vmin_u8(vy, voutput_max);
76
77 if (n & (4 * sizeof(float))) {
78 vst1_lane_u32((void*) y, vreinterpret_u32_u8(vy), 0); y += 4;
79 vy = vext_u8(vy, vy, 4);
80 }
81 if (n & (2 * sizeof(float))) {
82 vst1_lane_u16((void*) y, vreinterpret_u16_u8(vy), 0); y += 2;
83 vy = vext_u8(vy, vy, 2);
84 }
85 if (n & (1 * sizeof(float))) {
86 vst1_lane_u8(y, vy, 0);
87 }
88 }
89 }
90