xref: /aosp_15_r20/external/XNNPACK/src/math/cvt-f32-qu8-neonv8.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Copyright 2021 Google LLC
2 //
3 // This source code is licensed under the BSD-style license found in the
4 // LICENSE file in the root directory of this source tree.
5 
6 #include <assert.h>
7 #include <stddef.h>
8 #include <stdint.h>
9 
10 #include <arm_neon.h>
11 
12 #include <xnnpack/intrinsics-polyfill.h>
13 #include <xnnpack/math-stubs.h>
14 
15 
xnn_math_f32_qu8_cvt__neonv8(size_t n,const float * input,uint8_t * output,uint8_t output_zero_point)16 void xnn_math_f32_qu8_cvt__neonv8(
17     size_t n,
18     const float* input,
19     uint8_t* output,
20     uint8_t output_zero_point)
21 {
22   assert(n % (8 * sizeof(int8_t)) == 0);
23 
24   const int16x8_t voutput_zero_point = vdupq_n_s16((int16_t) (uint16_t) output_zero_point);
25   for (; n != 0; n -= 8 * sizeof(int8_t)) {
26     const float32x4_t vx_lo = vld1q_f32(input); input += 4;
27     const float32x4_t vx_hi = vld1q_f32(input); input += 4;
28 
29     const int32x4_t vy_lo = vcvtnq_s32_f32(vx_lo);
30     const int32x4_t vy_hi = vcvtnq_s32_f32(vx_hi);
31 
32     const int16x8_t vy = vqaddq_s16(vcombine_s16(vqmovn_s32(vy_lo), vqmovn_s32(vy_hi)), voutput_zero_point);
33 
34     const uint8x8_t vout = vqmovun_s16(vy);
35     vst1_u8(output, vout); output += 8;
36   }
37 }
38