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_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf(size_t mr,size_t nc,size_t kc,const int8_t * restrict a,size_t a_stride,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 const int8_t* restrict a,
23 size_t a_stride,
24 const void* restrict w,
25 int8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 const union xnn_qc8_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 int8_t* a0 = a;
36 int8_t* c0 = c;
37
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) *a0++;
46
47 const int32_t vb0 = (int32_t) ((const int8_t*) w)[0];
48 const int32_t vb1 = (int32_t) ((const int8_t*) w)[1];
49 w = (const void*) ((const int8_t*) w + 2);
50
51 vacc0x0 += va0 * vb0;
52 vacc0x1 += va0 * vb1;
53
54 k -= sizeof(int8_t);
55 } while (k != 0);
56
57 float vfpacc0x0 = (float) vacc0x0;
58 float vfpacc0x1 = (float) vacc0x1;
59
60 const float vscale0 = unaligned_indexed_load_f32(w, 0);
61 vfpacc0x0 *= vscale0;
62 const float vscale1 = unaligned_indexed_load_f32(w, 1);
63 vfpacc0x1 *= vscale1;
64 w = (const void*) ((const float*) w + 2);
65
66 const float voutput_min_less_zero_point = params->fp32_scalar_lrintf.output_min_less_zero_point;
67 vfpacc0x0 = math_max_f32(vfpacc0x0, voutput_min_less_zero_point);
68 vfpacc0x1 = math_max_f32(vfpacc0x1, voutput_min_less_zero_point);
69
70 const float voutput_max_less_zero_point = params->fp32_scalar_lrintf.output_max_less_zero_point;
71 vfpacc0x0 = math_min_f32(vfpacc0x0, voutput_max_less_zero_point);
72 vfpacc0x1 = math_min_f32(vfpacc0x1, voutput_max_less_zero_point);
73
74 const int32_t vrndacc0x0 = (int32_t) lrintf(vfpacc0x0);
75 const int32_t vrndacc0x1 = (int32_t) lrintf(vfpacc0x1);
76
77 const int32_t voutput_zero_point = params->fp32_scalar_lrintf.output_zero_point;
78 int32_t vout0x0 = vrndacc0x0 + voutput_zero_point;
79 int32_t vout0x1 = vrndacc0x1 + voutput_zero_point;
80
81 if XNN_LIKELY(nc >= 2) {
82 c0[0] = (int8_t) vout0x0;
83 c0[1] = (int8_t) vout0x1;
84
85 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
86
87 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
88
89 nc -= 2;
90 } else {
91 if (nc & 1) {
92 c0[0] = (int8_t) vout0x0;
93 }
94
95 nc = 0;
96 }
97 } while (nc != 0);
98 }
99