xref: /aosp_15_r20/external/XNNPACK/src/qu8-f32-vcvt/gen/vcvt-neon-x24.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-f32-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_qu8_f32_vcvt_ukernel__neon_x24(size_t n,const uint8_t * x,float * y,const union xnn_qu8_f32_cvt_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qu8_f32_vcvt_ukernel__neon_x24(
20     size_t n,
21     const uint8_t* x,
22     float* y,
23     const union xnn_qu8_f32_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
24 {
25   assert(n != 0);
26   assert(n % sizeof(uint8_t) == 0);
27   assert(x != NULL);
28   assert(y != NULL);
29 
30   const int16x8_t vminus_zero_point = vreinterpretq_s16_u32(vld1q_dup_u32((const void*) params->neon.minus_zero_point));
31   const float32x4_t vscale = vld1q_dup_f32(&params->neon.scale);
32   for (; n >= 24 * sizeof(uint8_t); n -= 24 * sizeof(uint8_t)) {
33     const uint8x8_t vx01234567 = vld1_u8(x); x += 8;
34     const uint8x8_t vx89ABCDEF = vld1_u8(x); x += 8;
35     const uint8x8_t vxGHIJKLMN = vld1_u8(x); x += 8;
36 
37     const int16x8_t vhx01234567 = vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(vminus_zero_point), vx01234567));
38     const int16x8_t vhx89ABCDEF = vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(vminus_zero_point), vx89ABCDEF));
39     const int16x8_t vhxGHIJKLMN = vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(vminus_zero_point), vxGHIJKLMN));
40 
41     const int32x4_t vwx0123 = vmovl_s16(vget_low_s16(vhx01234567));
42     const int32x4_t vwx4567 = vmovl_s16(vget_high_s16(vhx01234567));
43     const int32x4_t vwx89AB = vmovl_s16(vget_low_s16(vhx89ABCDEF));
44     const int32x4_t vwxCDEF = vmovl_s16(vget_high_s16(vhx89ABCDEF));
45     const int32x4_t vwxGHIJ = vmovl_s16(vget_low_s16(vhxGHIJKLMN));
46     const int32x4_t vwxKLMN = vmovl_s16(vget_high_s16(vhxGHIJKLMN));
47 
48     float32x4_t vy0123 = vcvtq_f32_s32(vwx0123);
49     float32x4_t vy4567 = vcvtq_f32_s32(vwx4567);
50     float32x4_t vy89AB = vcvtq_f32_s32(vwx89AB);
51     float32x4_t vyCDEF = vcvtq_f32_s32(vwxCDEF);
52     float32x4_t vyGHIJ = vcvtq_f32_s32(vwxGHIJ);
53     float32x4_t vyKLMN = vcvtq_f32_s32(vwxKLMN);
54 
55     vy0123 = vmulq_f32(vy0123, vscale);
56     vy4567 = vmulq_f32(vy4567, vscale);
57     vy89AB = vmulq_f32(vy89AB, vscale);
58     vyCDEF = vmulq_f32(vyCDEF, vscale);
59     vyGHIJ = vmulq_f32(vyGHIJ, vscale);
60     vyKLMN = vmulq_f32(vyKLMN, vscale);
61 
62     vst1q_f32(y, vy0123); y += 4;
63     vst1q_f32(y, vy4567); y += 4;
64     vst1q_f32(y, vy89AB); y += 4;
65     vst1q_f32(y, vyCDEF); y += 4;
66     vst1q_f32(y, vyGHIJ); y += 4;
67     vst1q_f32(y, vyKLMN); y += 4;
68   }
69   for (; n >= 8 * sizeof(uint8_t); n -= 8 * sizeof(uint8_t)) {
70     const uint8x8_t vx = vld1_u8(x); x += 8;
71 
72     const int16x8_t vhx = vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(vminus_zero_point), vx));
73 
74     const int32x4_t vwx_lo = vmovl_s16(vget_low_s16(vhx));
75     const int32x4_t vwx_hi = vmovl_s16(vget_high_s16(vhx));
76 
77     float32x4_t vy_lo = vcvtq_f32_s32(vwx_lo);
78     float32x4_t vy_hi = vcvtq_f32_s32(vwx_hi);
79 
80     vy_lo = vmulq_f32(vy_lo, vscale);
81     vy_hi = vmulq_f32(vy_hi, vscale);
82 
83     vst1q_f32(y, vy_lo); y += 4;
84     vst1q_f32(y, vy_hi); y += 4;
85   }
86   if XNN_UNLIKELY(n != 0) {
87     assert(n >= 1 * sizeof(uint8_t));
88     assert(n <= 7 * sizeof(uint8_t));
89 
90     const uint8x8_t vx = vld1_u8(x);
91 
92     const int16x8_t vhx = vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(vminus_zero_point), vx));
93 
94     const int32x4_t vwx_lo = vmovl_s16(vget_low_s16(vhx));
95     const int32x4_t vwx_hi = vmovl_s16(vget_high_s16(vhx));
96 
97     float32x4_t vy = vcvtq_f32_s32(vwx_lo);
98     vy = vmulq_f32(vy, vscale);
99 
100     if (n & (4 * sizeof(uint8_t))) {
101       vst1q_f32(y, vy); y += 4;
102       vy = vcvtq_f32_s32(vwx_hi);
103       vy = vmulq_f32(vy, vscale);
104     }
105     float32x2_t vy_lo = vget_low_f32(vy);
106     if (n & (2 * sizeof(uint8_t))) {
107       vst1_f32(y, vy_lo); y += 2;
108       vy_lo = vget_high_f32(vy);
109     }
110     if (n & (1 * sizeof(uint8_t))) {
111       vst1_lane_f32(y, vy_lo, 0);
112     }
113   }
114 }
115