Debian Codegraph @

src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-gemm-1x2-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_1x2__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 vksum1 = unaligned_indexed_load_s32(w, 1);
43 const int32_t vinput_zero_point0 = quantization_params[0].zero_point;
44 int32_t vacc0x0 = vksum0 * vinput_zero_point0;
45 int32_t vacc0x1 = vksum1 * vinput_zero_point0;
46 w = (const int32_t*) w + 2;
47
48 size_t k = kc;
49 for (; k >= 2 * sizeof(uint8_t); k -= 2 * sizeof(uint8_t)) {
50 const int32_t va0c0 = (int32_t) a0[0];
51 const int32_t va0c1 = (int32_t) a0[1];
52 a0 += 2;
53
54 const uint8_t vbi0 = ((const uint8_t*) w)[0];
55 const uint8_t vbi1 = ((const uint8_t*) w)[1];
56 w = (const uint8_t*) w + 2;
57 const int32_t vb0c0 = (int32_t) (int8_t) (vbi0 << 4);
58 const int32_t vb0c1 = (int32_t) (int8_t) (vbi0 & 0xF0);
59 const int32_t vb1c0 = (int32_t) (int8_t) (vbi1 << 4);
60 const int32_t vb1c1 = (int32_t) (int8_t) (vbi1 & 0xF0);
61
62 vacc0x0 += va0c0 * vb0c0;
63 vacc0x1 += va0c0 * vb1c0;
64 vacc0x0 += va0c1 * vb0c1;
65 vacc0x1 += va0c1 * vb1c1;
66 }
67
68 float vout0x0 = (float) math_asr_s32(vacc0x0, 4);
69 float vout0x1 = (float) math_asr_s32(vacc0x1, 4);
70
71 const float vinput_scale0 = quantization_params[0].inv_scale;
72 vout0x0 *= vinput_scale0;
73 vout0x1 *= vinput_scale0;
74
75 const float vfilter_output_scale0 = unaligned_indexed_load_f32(w, 0);
76 vout0x0 *= vfilter_output_scale0;
77 const float vfilter_output_scale1 = unaligned_indexed_load_f32(w, 1);
78 vout0x1 *= vfilter_output_scale1;
79
80 const float vbias0 = unaligned_indexed_load_f32(w, 2);
81 vout0x0 += vbias0;
82 const float vbias1 = unaligned_indexed_load_f32(w, 3);
83 vout0x1 += vbias1;
84
85 w = (const float*) w + 4;
86
87 const float voutput_min = params->scalar.min;
88 vout0x0 = math_max_f32(vout0x0, voutput_min);
89 vout0x1 = math_max_f32(vout0x1, voutput_min);
90
91 const float voutput_max = params->scalar.max;
92 vout0x0 = math_min_f32(vout0x0, voutput_max);
93 vout0x1 = math_min_f32(vout0x1, voutput_max);
94
95 if XNN_LIKELY(nc >= 2) {
96 c0[0] = vout0x0;
97 c0[1] = vout0x1;
98
99 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
100
101 c0 = (float*) ((uintptr_t) c0 + cn_stride);
102
103 nc -= 2;
104 } else {
105 if (nc & 1) {
106 c0[0] = vout0x0;
107 }
108
109 nc = 0;
110 }
111 } while (nc != 0);
112 }