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_x16(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_x16(
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 uint8x16_t voutput_min = vld1q_dup_u8(¶ms->neon.output_min);
34 const uint8x16_t voutput_max = vld1q_dup_u8(¶ms->neon.output_max);
35 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
36 float32x4_t vx0123 = vld1q_f32(x); x += 4;
37 float32x4_t vx4567 = vld1q_f32(x); x += 4;
38 float32x4_t vx89AB = vld1q_f32(x); x += 4;
39 float32x4_t vxCDEF = vld1q_f32(x); x += 4;
40
41 vx0123 = vmulq_f32(vx0123, vscale);
42 vx4567 = vmulq_f32(vx4567, vscale);
43 vx89AB = vmulq_f32(vx89AB, vscale);
44 vxCDEF = vmulq_f32(vxCDEF, vscale);
45
46 vx0123 = vaddq_f32(vx0123, vmagic_bias);
47 vx4567 = vaddq_f32(vx4567, vmagic_bias);
48 vx89AB = vaddq_f32(vx89AB, vmagic_bias);
49 vxCDEF = vaddq_f32(vxCDEF, vmagic_bias);
50
51 const int32x4_t vacc0123 = vqsubq_s32(vreinterpretq_s32_f32(vx0123), vmagic_bias_less_zero_point);
52 const int32x4_t vacc4567 = vqsubq_s32(vreinterpretq_s32_f32(vx4567), vmagic_bias_less_zero_point);
53 const int32x4_t vacc89AB = vqsubq_s32(vreinterpretq_s32_f32(vx89AB), vmagic_bias_less_zero_point);
54 const int32x4_t vaccCDEF = vqsubq_s32(vreinterpretq_s32_f32(vxCDEF), vmagic_bias_less_zero_point);
55
56 const int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
57 const int16x8_t vacc89ABCDEF = vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF));
58
59 uint8x16_t vy0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc01234567), vqmovun_s16(vacc89ABCDEF));
60
61 vy0123456789ABCDEF = vmaxq_u8(vy0123456789ABCDEF, voutput_min);
62
63 vy0123456789ABCDEF = vminq_u8(vy0123456789ABCDEF, voutput_max);
64
65 vst1q_u8(y, vy0123456789ABCDEF); y += 16;
66 }
67 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
68 float32x4_t vx_lo = vld1q_f32(x); x += 4;
69 float32x4_t vx_hi = vld1q_f32(x); x += 4;
70
71 vx_lo = vmulq_f32(vx_lo, vscale);
72 vx_hi = vmulq_f32(vx_hi, vscale);
73
74 vx_lo = vaddq_f32(vx_lo, vmagic_bias);
75 vx_hi = vaddq_f32(vx_hi, vmagic_bias);
76
77 const int32x4_t vacc_lo = vqsubq_s32(vreinterpretq_s32_f32(vx_lo), vmagic_bias_less_zero_point);
78 const int32x4_t vacc_hi = vqsubq_s32(vreinterpretq_s32_f32(vx_hi), vmagic_bias_less_zero_point);
79
80 const int16x8_t vacc = vcombine_s16(vqmovn_s32(vacc_lo), vqmovn_s32(vacc_hi));
81
82 uint8x8_t vy = vqmovun_s16(vacc);
83 vy = vmax_u8(vy, vget_low_u8(voutput_min));
84 vy = vmin_u8(vy, vget_low_u8(voutput_max));
85 vst1_u8(y, vy); y += 8;
86 }
87 if XNN_UNLIKELY(n != 0) {
88 assert(n >= 1 * sizeof(float));
89 assert(n <= 7 * sizeof(float));
90 float32x4_t vx_lo = vld1q_f32(x);
91 const float* x_hi = (const float*) ((uintptr_t) x + (n & (4 * sizeof(float))));
92 float32x4_t vx_hi = vld1q_f32(x_hi);
93
94 vx_lo = vmulq_f32(vx_lo, vscale);
95 vx_hi = vmulq_f32(vx_hi, vscale);
96
97 vx_lo = vaddq_f32(vx_lo, vmagic_bias);
98 vx_hi = vaddq_f32(vx_hi, vmagic_bias);
99
100 const int32x4_t vacc_lo = vqsubq_s32(vreinterpretq_s32_f32(vx_lo), vmagic_bias_less_zero_point);
101 const int32x4_t vacc_hi = vqsubq_s32(vreinterpretq_s32_f32(vx_hi), vmagic_bias_less_zero_point);
102
103 const int16x8_t vacc = vcombine_s16(vqmovn_s32(vacc_lo), vqmovn_s32(vacc_hi));
104
105 uint8x8_t vy = vqmovun_s16(vacc);
106 vy = vmax_u8(vy, vget_low_u8(voutput_min));
107 vy = vmin_u8(vy, vget_low_u8(voutput_max));
108
109 if (n & (4 * sizeof(float))) {
110 vst1_lane_u32((void*) y, vreinterpret_u32_u8(vy), 0); y += 4;
111 vy = vext_u8(vy, vy, 4);
112 }
113 if (n & (2 * sizeof(float))) {
114 vst1_lane_u16((void*) y, vreinterpret_u16_u8(vy), 0); y += 2;
115 vy = vext_u8(vy, vy, 2);
116 }
117 if (n & (1 * sizeof(float))) {
118 vst1_lane_u8(y, vy, 0);
119 }
120 }
121 }
122