Debian Codegraph @

src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-gemm-1x1-minmax-scalar.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/scalar.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 "xnnpack/gemm.h"
13 #include "xnnpack/math.h"
14 #include "xnnpack/unaligned.h"
15
16
17 void xnn_qd8_f32_qc4w_gemm_minmax_ukernel_1x1__scalar(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 const int8_t* restrict a,
22 size_t a_stride,
23 const void* restrict w,
24 float* restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 const struct xnn_f32_qc4w_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
28 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)])
29 {
30 assert(mr != 0);
31 assert(mr <= 1);
32 assert(nc != 0);
33 assert(kc != 0);
34
35 const int8_t* a0 = a;
36 float* c0 = c;
37
38
39 kc = round_up_po2(kc, 2);
40 do {
41 const int32_t vksum0 = unaligned_indexed_load_s32(w, 0);
42 const int32_t vinput_zero_point0 = quantization_params[0].zero_point;
43 int32_t vacc0x0 = vksum0 * vinput_zero_point0;
44 w = (const int32_t*) w + 1;
45
46 size_t k = kc;
47 for (; k >= 2 * sizeof(uint8_t); k -= 2 * sizeof(uint8_t)) {
48 const int32_t va0c0 = (int32_t) a0[0];
49 const int32_t va0c1 = (int32_t) a0[1];
50 a0 += 2;
51
52 const uint8_t vbi0 = ((const uint8_t*) w)[0];
53 w = (const uint8_t*) w + 1;
54 const int32_t vb0c0 = (int32_t) (int8_t) (vbi0 << 4);
55 const int32_t vb0c1 = (int32_t) (int8_t) (vbi0 & 0xF0);
56
57 vacc0x0 += va0c0 * vb0c0;
58 vacc0x0 += va0c1 * vb0c1;
59 }
60
61 float vout0x0 = (float) math_asr_s32(vacc0x0, 4);
62
63 const float vinput_scale0 = quantization_params[0].inv_scale;
64 vout0x0 *= vinput_scale0;
65
66 const float vfilter_output_scale0 = unaligned_indexed_load_f32(w, 0);
67 vout0x0 *= vfilter_output_scale0;
68
69 const float vbias0 = unaligned_indexed_load_f32(w, 1);
70 vout0x0 += vbias0;
71
72 w = (const float*) w + 2;
73
74 const float voutput_min = params->scalar.min;
75 vout0x0 = math_max_f32(vout0x0, voutput_min);
76
77 const float voutput_max = params->scalar.max;
78 vout0x0 = math_min_f32(vout0x0, voutput_max);
79
80 if XNN_LIKELY(nc >= 1) {
81 c0[0] = vout0x0;
82
83 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
84
85 c0 = (float*) ((uintptr_t) c0 + cn_stride);
86
87 nc -= 1;
88 } else {
89
90 nc = 0;
91 }
92 } while (nc != 0);
93 }