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 #include <math.h>
12
13 #include <xnnpack/math.h>
14 #include <xnnpack/gemm.h>
15 #include <xnnpack/unaligned.h>
16
17
xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf(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)])18 void xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 const uint8_t* restrict a,
23 size_t a_stride,
24 const void* restrict w,
25 uint8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 const union xnn_qu8_conv_minmax_params 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 uint8_t* a0 = a;
36 uint8_t* c0 = c;
37
38 const int32_t vb_zero_point = params->fp32_scalar_lrintf.kernel_zero_point;
39 do {
40 int32_t vacc0x0 = unaligned_indexed_load_s32(w, 0);
41 int32_t vacc0x1 = unaligned_indexed_load_s32(w, 1);
42 w = (const void*) ((const int32_t*) w + 2);
43
44 size_t k = kc;
45 do {
46 const int32_t va0 = (int32_t) (uint32_t) *a0++;
47
48 const int32_t vb0 = (int32_t) (uint32_t) ((const uint8_t*) w)[0] - vb_zero_point;
49 const int32_t vb1 = (int32_t) (uint32_t) ((const uint8_t*) w)[1] - vb_zero_point;
50 w = (const void*) ((const uint8_t*) w + 2);
51
52 vacc0x0 += va0 * vb0;
53 vacc0x1 += va0 * vb1;
54
55 k -= sizeof(uint8_t);
56 } while (k != 0);
57
58 float vfpacc0x0 = (float) vacc0x0;
59 float vfpacc0x1 = (float) vacc0x1;
60
61 const float vscale = params->fp32_scalar_lrintf.scale;
62 vfpacc0x0 *= vscale;
63 vfpacc0x1 *= vscale;
64
65 const float voutput_min_less_zero_point = params->fp32_scalar_lrintf.output_min_less_zero_point;
66 vfpacc0x0 = math_max_f32(vfpacc0x0, voutput_min_less_zero_point);
67 vfpacc0x1 = math_max_f32(vfpacc0x1, voutput_min_less_zero_point);
68
69 const float voutput_max_less_zero_point = params->fp32_scalar_lrintf.output_max_less_zero_point;
70 vfpacc0x0 = math_min_f32(vfpacc0x0, voutput_max_less_zero_point);
71 vfpacc0x1 = math_min_f32(vfpacc0x1, voutput_max_less_zero_point);
72
73 const int32_t vrndacc0x0 = (int32_t) lrintf(vfpacc0x0);
74 const int32_t vrndacc0x1 = (int32_t) lrintf(vfpacc0x1);
75
76 const int32_t voutput_zero_point = params->fp32_scalar_lrintf.output_zero_point;
77 int32_t vout0x0 = vrndacc0x0 + voutput_zero_point;
78 int32_t vout0x1 = vrndacc0x1 + voutput_zero_point;
79
80 if XNN_LIKELY(nc >= 2) {
81 c0[0] = (uint8_t) vout0x0;
82 c0[1] = (uint8_t) vout0x1;
83
84 a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
85
86 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
87
88 nc -= 2;
89 } else {
90 if (nc & 1) {
91 c0[0] = (uint8_t) vout0x0;
92 }
93
94 nc = 0;
95 }
96 } while (nc != 0);
97 }
98