Debian Codegraph @

src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-2x2-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_qb4w_gemm_minmax_ukernel_2x2__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_qb4w_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 <= 2);
32 assert(nc != 0);
33 assert(kc != 0);
34 size_t bl = params->scalar.blocksize;
35 assert(bl <= round_up_po2(kc, 2));
36 assert(bl != 0);
37 assert(bl % 32 == 0);
38
39 const int8_t* a0 = a;
40 float* c0 = c;
41 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
42 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
43 if XNN_UNPREDICTABLE(mr != 2) {
44 a1 = a0;
45 c1 = c0;
46 }
47
48
49 kc = round_up_po2(kc, 2);
50 do {
51 const float vksum0 = unaligned_indexed_load_f32(w, 0);
52 const float vksum1 = unaligned_indexed_load_f32(w, 1);
53 const float vinput_zero_point0 = (const float) quantization_params[0].zero_point;
54 float vout0x0 = vksum0 * vinput_zero_point0;
55 float vout0x1 = vksum1 * vinput_zero_point0;
56 const float vinput_zero_point1 = (const float) quantization_params[1].zero_point;
57 float vout1x0 = vksum0 * vinput_zero_point1;
58 float vout1x1 = vksum1 * vinput_zero_point1;
59 w = (const float*) w + 2;
60
61 for (size_t kb=0; kb < kc; kb += bl) {
62 int32_t vacc0x0 = 0;
63 int32_t vacc0x1 = 0;
64 int32_t vacc1x0 = 0;
65 int32_t vacc1x1 = 0;
66 size_t k = bl;
67 for (; k >= 2 * sizeof(uint8_t); k -= 2 * sizeof(uint8_t)) {
68 const int32_t va0c0 = (int32_t) a0[0];
69 const int32_t va0c1 = (int32_t) a0[1];
70 a0 += 2;
71 const int32_t va1c0 = (int32_t) a1[0];
72 const int32_t va1c1 = (int32_t) a1[1];
73 a1 += 2;
74
75 const uint8_t vbi0 = ((const uint8_t*) w)[0];
76 const uint8_t vbi1 = ((const uint8_t*) w)[1];
77 w = (const uint8_t*) w + 2;
78 const int32_t vb0c0 = (int32_t) (int8_t) (vbi0 << 4);
79 const int32_t vb0c1 = (int32_t) (int8_t) (vbi0 & 0xF0);
80 const int32_t vb1c0 = (int32_t) (int8_t) (vbi1 << 4);
81 const int32_t vb1c1 = (int32_t) (int8_t) (vbi1 & 0xF0);
82
83 vacc0x0 += va0c0 * vb0c0;
84 vacc0x1 += va0c0 * vb1c0;
85 vacc1x0 += va1c0 * vb0c0;
86 vacc1x1 += va1c0 * vb1c0;
87 vacc0x0 += va0c1 * vb0c1;
88 vacc0x1 += va0c1 * vb1c1;
89 vacc1x0 += va1c1 * vb0c1;
90 vacc1x1 += va1c1 * vb1c1;
91 }
92 // accumulate in float
93 float vf0x0 = vacc0x0;
94 const float vfilter_output_scale0 = math_cvt_fp32_bf16(unaligned_indexed_load_u16(w, 0));
95 float vf0x1 = vacc0x1;
96 const float vfilter_output_scale1 = math_cvt_fp32_bf16(unaligned_indexed_load_u16(w, 1));
97 float vf1x0 = vacc1x0;
98 float vf1x1 = vacc1x1;
99
100 vf0x0 *= vfilter_output_scale0;
101 vout0x0 += vf0x0;
102 vf0x1 *= vfilter_output_scale1;
103 vout0x1 += vf0x1;
104 vf1x0 *= vfilter_output_scale0;
105 vout1x0 += vf1x0;
106 vf1x1 *= vfilter_output_scale1;
107 vout1x1 += vf1x1;
108 w = (const uint16_t*) w + 2;
109 }
110
111
112 const float vinput_scale0 = quantization_params[0].inv_scale;
113 vout0x0 *= vinput_scale0;
114 vout0x1 *= vinput_scale0;
115 const float vinput_scale1 = quantization_params[1].inv_scale;
116 vout1x0 *= vinput_scale1;
117 vout1x1 *= vinput_scale1;
118
119
120 const float vbias0 = unaligned_indexed_load_f32(w, 0);
121 vout0x0 += vbias0;
122 vout1x0 += vbias0;
123 const float vbias1 = unaligned_indexed_load_f32(w, 1);
124 vout0x1 += vbias1;
125 vout1x1 += vbias1;
126
127 w = (const float*) w + 2;
128
129 const float voutput_min = params->scalar.min;
130 vout0x0 = math_max_f32(vout0x0, voutput_min);
131 vout1x0 = math_max_f32(vout1x0, voutput_min);
132 vout0x1 = math_max_f32(vout0x1, voutput_min);
133 vout1x1 = math_max_f32(vout1x1, voutput_min);
134
135 const float voutput_max = params->scalar.max;
136 vout0x0 = math_min_f32(vout0x0, voutput_max);
137 vout1x0 = math_min_f32(vout1x0, voutput_max);
138 vout0x1 = math_min_f32(vout0x1, voutput_max);
139 vout1x1 = math_min_f32(vout1x1, voutput_max);
140
141 if XNN_LIKELY(nc >= 2) {
142 c0[0] = vout0x0;
143 c0[1] = vout0x1;
144 c1[0] = vout1x0;
145 c1[1] = vout1x1;
146
147 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
148 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
149
150 c0 = (float*) ((uintptr_t) c0 + cn_stride);
151 c1 = (float*) ((uintptr_t) c1 + cn_stride);
152
153 nc -= 2;
154 } else {
155 if (nc & 1) {
156 c0[0] = vout0x0;
157 c1[0] = vout1x0;
158 }
159
160 nc = 0;
161 }
162 } while (nc != 0);
163 }