Debian Codegraph @

src/qu8-vcvt/gen/qu8-vcvt-armsimd32-u4.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-vcvt/armsimd32.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2022 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_acle.h>
13
14 #include "xnnpack/intrinsics-polyfill.h"
15 #include "xnnpack/math.h"
16 #include "xnnpack/unaligned.h"
17 #include "xnnpack/vcvt.h"
18
19
20 void xnn_qu8_vcvt_ukernel__armsimd32_u4(
21 size_t batch,
22 const uint8_t* input,
23 uint8_t* output,
24 const struct xnn_qu8_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
25 {
26 assert(batch != 0);
27 assert(batch % sizeof(uint8_t) == 0);
28 assert(input != NULL);
29 assert(output != NULL);
30
31 const uint16x2_t vminus_input_zero_point = (uint16x2_t) broadcast2x_uint16(-params->scalar.input_zero_point);
32 const int32_t vbias = (int32_t) ((uint32_t) (int32_t) params->scalar.output_zero_point << 1) + INT32_C(1);
33 const int32_t vmultiplier = (int32_t) params->scalar.multiplier << 9;
34 for (; batch >= 4 * sizeof(uint8_t); batch -= 4 * sizeof(uint8_t)) {
35 const uint8x4_t vx0123 = (uint8x4_t) unaligned_load_u32(input);
36 input += 4;
37
38 const uint16x2_t vx02 = __uxtab16(vminus_input_zero_point, vx0123);
39 const uint16x2_t vx13 = __uxtab16(vminus_input_zero_point, __ror(vx0123, 8));
40
41 int32_t vacc0 = __smlawb(vmultiplier, vx02, vbias);
42 int32_t vacc1 = __smlawb(vmultiplier, vx13, vbias);
43 int32_t vacc2 = __smlawt(vmultiplier, vx02, vbias);
44 int32_t vacc3 = __smlawt(vmultiplier, vx13, vbias);
45
46 vacc0 = __usat(math_asr_s32(vacc0, 1), 8);
47 vacc1 = __usat(math_asr_s32(vacc1, 1), 8);
48 vacc2 = __usat(math_asr_s32(vacc2, 1), 8);
49 vacc3 = __usat(math_asr_s32(vacc3, 1), 8);
50
51 output[0] = (uint8_t) vacc0;
52 output[1] = (uint8_t) vacc1;
53 output[2] = (uint8_t) vacc2;
54 output[3] = (uint8_t) vacc3;
55 output += 4;
56 }
57 if XNN_UNLIKELY(batch != 0) {
58 const uint8x4_t vx0123 = (uint8x4_t) unaligned_load_u32(input);
59
60 const uint16x2_t vx02 = __uxtab16(vminus_input_zero_point, vx0123);
61 const uint16x2_t vx13 = __uxtab16(vminus_input_zero_point, __ror(vx0123, 8));
62
63 int32_t vacc0 = __smlawb(vmultiplier, vx02, vbias);
64 int32_t vacc1 = __smlawb(vmultiplier, vx13, vbias);
65 const int32_t vacc2 = __smlawt(vmultiplier, vx02, vbias);
66
67 vacc0 = __usat(math_asr_s32(vacc0, 1), 8);
68 vacc1 = __usat(math_asr_s32(vacc1, 1), 8);
69
70 if (batch & (2 * sizeof(uint8_t))) {
71 output[0] = (uint8_t) vacc0;
72 output[1] = (uint8_t) vacc1;
73 vacc0 = __usat(math_asr_s32(vacc2, 1), 8);
74 output += 2;
75 }
76 if (batch & (1 * sizeof(uint8_t))) {
77 output[0] = (uint8_t) vacc0;
78 }
79 }
80 }