Debian Codegraph @

src/qd8-f16-qb4w-gemm/gen/qd8-f16-qb4w-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_f16_qb4w_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 xnn_float16* restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 const struct xnn_f16_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 <= 1);
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 xnn_float16* c0 = c;
41
42
43 kc = round_up_po2(kc, 2);
44 do {
45 const float vksum0 = unaligned_indexed_load_f32(w, 0);
46 const float vksum1 = unaligned_indexed_load_f32(w, 1);
47 const float vinput_zero_point0 = (const float) quantization_params[0].zero_point;
48 float vout0x0 = vksum0 * vinput_zero_point0;
49 float vout0x1 = vksum1 * vinput_zero_point0;
50 w = (const float*) w + 2;
51
52 for (size_t kb=0; kb < kc; kb += bl) {
53 int32_t vacc0x0 = 0;
54 int32_t vacc0x1 = 0;
55 size_t k = bl;
56 for (; k >= 2 * sizeof(uint8_t); k -= 2 * sizeof(uint8_t)) {
57 const int32_t va0c0 = (int32_t) a0[0];
58 const int32_t va0c1 = (int32_t) a0[1];
59 a0 += 2;
60
61 const uint8_t vbi0 = ((const uint8_t*) w)[0];
62 const uint8_t vbi1 = ((const uint8_t*) w)[1];
63 w = (const uint8_t*) w + 2;
64 const int32_t vb0c0 = (int32_t) (int8_t) (vbi0 << 4);
65 const int32_t vb0c1 = (int32_t) (int8_t) (vbi0 & 0xF0);
66 const int32_t vb1c0 = (int32_t) (int8_t) (vbi1 << 4);
67 const int32_t vb1c1 = (int32_t) (int8_t) (vbi1 & 0xF0);
68
69 vacc0x0 += va0c0 * vb0c0;
70 vacc0x1 += va0c0 * vb1c0;
71 vacc0x0 += va0c1 * vb0c1;
72 vacc0x1 += va0c1 * vb1c1;
73 }
74 // accumulate in float
75 float vf0x0 = vacc0x0;
76 const float vfilter_output_scale0 = math_cvt_fp32_bf16(unaligned_indexed_load_u16(w, 0));
77 float vf0x1 = vacc0x1;
78 const float vfilter_output_scale1 = math_cvt_fp32_bf16(unaligned_indexed_load_u16(w, 1));
79
80 vf0x0 *= vfilter_output_scale0;
81 vout0x0 += vf0x0;
82 vf0x1 *= vfilter_output_scale1;
83 vout0x1 += vf0x1;
84 w = (const uint16_t*) w + 2;
85 }
86
87
88 const float vinput_scale0 = quantization_params[0].inv_scale;
89 vout0x0 *= vinput_scale0;
90 vout0x1 *= vinput_scale0;
91
92
93 const float vbias0 = unaligned_indexed_load_f32(w, 0);
94 vout0x0 += vbias0;
95 const float vbias1 = unaligned_indexed_load_f32(w, 1);
96 vout0x1 += vbias1;
97
98 w = (const float*) w + 2;
99
100 const float voutput_min = xnn_float16_to_float(params->scalar.min);
101 vout0x0 = math_max_f32(vout0x0, voutput_min);
102 vout0x1 = math_max_f32(vout0x1, voutput_min);
103
104 const float voutput_max = xnn_float16_to_float(params->scalar.max);
105 vout0x0 = math_min_f32(vout0x0, voutput_max);
106 vout0x1 = math_min_f32(vout0x1, voutput_max);
107
108 if XNN_LIKELY(nc >= 2) {
109 c0[0] = xnn_float16_from_float(vout0x0);
110 c0[1] = xnn_float16_from_float(vout0x1);
111
112 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
113
114 c0 = (xnn_float16*) ((uintptr_t) c0 + cn_stride);
115
116 nc -= 2;
117 } else {
118 if (nc & 1) {
119 c0[0] = xnn_float16_from_float(vout0x0);
120 }
121
122 nc = 0;
123 }
124 } while (nc != 0);
125 }