xref: /aosp_15_r20/external/XNNPACK/src/f32-qu8-vcvt/gen/vcvt-neonv8-x16.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
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_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__neonv8_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(&params->neonv8.scale);
31   const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->neonv8.output_zero_point);
32   const uint8x16_t voutput_min = vld1q_dup_u8(&params->neonv8.output_min);
33   const uint8x16_t voutput_max = vld1q_dup_u8(&params->neonv8.output_max);
34   for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
35     float32x4_t vx0123 = vld1q_f32(x); x += 4;
36     float32x4_t vx4567 = vld1q_f32(x); x += 4;
37     float32x4_t vx89AB = vld1q_f32(x); x += 4;
38     float32x4_t vxCDEF = vld1q_f32(x); x += 4;
39 
40     vx0123 = vmulq_f32(vx0123, vscale);
41     vx4567 = vmulq_f32(vx4567, vscale);
42     vx89AB = vmulq_f32(vx89AB, vscale);
43     vxCDEF = vmulq_f32(vxCDEF, vscale);
44 
45     const int32x4_t vacc0123 = vcvtnq_s32_f32(vx0123);
46     const int32x4_t vacc4567 = vcvtnq_s32_f32(vx4567);
47     const int32x4_t vacc89AB = vcvtnq_s32_f32(vx89AB);
48     const int32x4_t vaccCDEF = vcvtnq_s32_f32(vxCDEF);
49 
50     int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
51     int16x8_t vacc89ABCDEF = vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF));
52 
53     vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
54     vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
55 
56     uint8x16_t vy0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc01234567), vqmovun_s16(vacc89ABCDEF));
57 
58     vy0123456789ABCDEF = vmaxq_u8(vy0123456789ABCDEF, voutput_min);
59 
60     vy0123456789ABCDEF = vminq_u8(vy0123456789ABCDEF, voutput_max);
61 
62     vst1q_u8(y, vy0123456789ABCDEF); y += 16;
63   }
64   for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
65     float32x4_t vx_lo = vld1q_f32(x); x += 4;
66     float32x4_t vx_hi = vld1q_f32(x); x += 4;
67 
68     vx_lo = vmulq_f32(vx_lo, vscale);
69     vx_hi = vmulq_f32(vx_hi, vscale);
70 
71     const int32x4_t vacc_lo = vcvtnq_s32_f32(vx_lo);
72     const int32x4_t vacc_hi = vcvtnq_s32_f32(vx_hi);
73 
74     int16x8_t vacc = vcombine_s16(vqmovn_s32(vacc_lo), vqmovn_s32(vacc_hi));
75     vacc = vqaddq_s16(vacc, voutput_zero_point);
76 
77     uint8x8_t vy = vqmovun_s16(vacc);
78     vy = vmax_u8(vy, vget_low_u8(voutput_min));
79     vy = vmin_u8(vy, vget_low_u8(voutput_max));
80     vst1_u8(y, vy); y += 8;
81   }
82   if XNN_UNLIKELY(n != 0) {
83     assert(n >= 1 * sizeof(float));
84     assert(n <= 7 * sizeof(float));
85     float32x4_t vx_lo = vld1q_f32(x);
86     const float* x_hi = (const float*) ((uintptr_t) x + (n & (4 * sizeof(float))));
87     float32x4_t vx_hi = vld1q_f32(x_hi);
88 
89     vx_lo = vmulq_f32(vx_lo, vscale);
90     vx_hi = vmulq_f32(vx_hi, vscale);
91 
92     const int32x4_t vacc_lo = vcvtnq_s32_f32(vx_lo);
93     const int32x4_t vacc_hi = vcvtnq_s32_f32(vx_hi);
94 
95     int16x8_t vacc = vcombine_s16(vqmovn_s32(vacc_lo), vqmovn_s32(vacc_hi));
96     vacc = vqaddq_s16(vacc, voutput_zero_point);
97 
98     uint8x8_t vy = vqmovun_s16(vacc);
99     vy = vmax_u8(vy, vget_low_u8(voutput_min));
100     vy = vmin_u8(vy, vget_low_u8(voutput_max));
101 
102     if (n & (4 * sizeof(float))) {
103       vst1_lane_u32((void*) y, vreinterpret_u32_u8(vy), 0); y += 4;
104       vy = vext_u8(vy, vy, 4);
105     }
106     if (n & (2 * sizeof(float))) {
107       vst1_lane_u16((void*) y, vreinterpret_u16_u8(vy), 0); y += 2;
108       vy = vext_u8(vy, vy, 2);
109     }
110     if (n & (1 * sizeof(float))) {
111       vst1_lane_u8(y, vy, 0);
112     }
113   }
114 }
115