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/math.h>
13 #include <xnnpack/gemm.h>
14 #include <xnnpack/unaligned.h>
15
16
xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic(size_t mr,size_t nc,size_t kc,const uint8_t * restrict a,size_t a_stride,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 const uint8_t* restrict a,
22 size_t a_stride,
23 const void* restrict w,
24 uint8_t* restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
28 {
29 assert(mr != 0);
30 assert(mr <= 1);
31 assert(nc != 0);
32 assert(kc != 0);
33
34 const uint8_t* a0 = a;
35 uint8_t* c0 = c;
36
37 const int32_t vb_zero_point = params->fp32_scalar_imagic.kernel_zero_point;
38 do {
39 int32_t vacc0x0 = unaligned_indexed_load_s32(w, 0);
40 int32_t vacc0x1 = unaligned_indexed_load_s32(w, 1);
41 w = (const void*) ((const int32_t*) w + 2);
42
43 size_t k = kc;
44 do {
45 const int32_t va0 = (int32_t) (uint32_t) *a0++;
46
47 const int32_t vb0 = (int32_t) (uint32_t) ((const uint8_t*) w)[0] - vb_zero_point;
48 const int32_t vb1 = (int32_t) (uint32_t) ((const uint8_t*) w)[1] - vb_zero_point;
49 w = (const void*) ((const uint8_t*) w + 2);
50
51 vacc0x0 += va0 * vb0;
52 vacc0x1 += va0 * vb1;
53
54 k -= sizeof(uint8_t);
55 } while (k != 0);
56
57 float vfpacc0x0 = (float) vacc0x0;
58 float vfpacc0x1 = (float) vacc0x1;
59
60 const float vscale = params->fp32_scalar_imagic.scale;
61 vfpacc0x0 *= vscale;
62 vfpacc0x1 *= vscale;
63
64 const float vmagic_bias = params->fp32_scalar_imagic.magic_bias;
65 vfpacc0x0 += vmagic_bias;
66 vfpacc0x1 += vmagic_bias;
67
68 int32_t vout0x0 = (int32_t) float_as_uint32(vfpacc0x0);
69 int32_t vout0x1 = (int32_t) float_as_uint32(vfpacc0x1);
70
71 const int32_t vmagic_min = params->fp32_scalar_imagic.magic_min;
72 vout0x0 = math_max_s32(vout0x0, vmagic_min);
73 vout0x1 = math_max_s32(vout0x1, vmagic_min);
74
75 const int32_t vmagic_max = params->fp32_scalar_imagic.magic_max;
76 vout0x0 = math_min_s32(vout0x0, vmagic_max);
77 vout0x1 = math_min_s32(vout0x1, vmagic_max);
78
79 const int32_t vmagic_bias_less_zero_point = params->fp32_scalar_imagic.magic_bias_less_zero_point;
80 vout0x0 -= vmagic_bias_less_zero_point;
81 vout0x1 -= vmagic_bias_less_zero_point;
82
83 if XNN_LIKELY(nc >= 2) {
84 c0[0] = (uint8_t) vout0x0;
85 c0[1] = (uint8_t) vout0x1;
86
87 a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
88
89 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
90
91 nc -= 2;
92 } else {
93 if (nc & 1) {
94 c0[0] = (uint8_t) vout0x0;
95 }
96
97 nc = 0;
98 }
99 } while (nc != 0);
100 }
101