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_3x2__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_3x2__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 <= 3);
32 assert(nc != 0);
33 assert(kc != 0);
34
35 const int8_t* a0 = a;
36 int8_t* c0 = c;
37 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
38 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
39 if XNN_UNPREDICTABLE(mr < 2) {
40 a1 = a0;
41 c1 = c0;
42 }
43 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
44 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
45 if XNN_UNPREDICTABLE(mr <= 2) {
46 a2 = a1;
47 c2 = c1;
48 }
49
50 do {
51 int32_t vacc0x0 = unaligned_indexed_load_s32(w, 0);
52 int32_t vacc0x1 = unaligned_indexed_load_s32(w, 1);
53 int32_t vacc1x0 = vacc0x0;
54 int32_t vacc1x1 = vacc0x1;
55 int32_t vacc2x0 = vacc0x0;
56 int32_t vacc2x1 = vacc0x1;
57 w = (const void*) ((const int32_t*) w + 2);
58
59 size_t k = kc;
60 do {
61 const int32_t va0 = (int32_t) *a0++;
62 const int32_t va1 = (int32_t) *a1++;
63 const int32_t va2 = (int32_t) *a2++;
64
65 const int32_t vb0 = (int32_t) ((const int8_t*) w)[0];
66 const int32_t vb1 = (int32_t) ((const int8_t*) w)[1];
67 w = (const void*) ((const int8_t*) w + 2);
68
69 vacc0x0 += va0 * vb0;
70 vacc0x1 += va0 * vb1;
71 vacc1x0 += va1 * vb0;
72 vacc1x1 += va1 * vb1;
73 vacc2x0 += va2 * vb0;
74 vacc2x1 += va2 * vb1;
75
76 k -= sizeof(int8_t);
77 } while (k != 0);
78
79 float vfpacc0x0 = (float) vacc0x0;
80 float vfpacc0x1 = (float) vacc0x1;
81 float vfpacc1x0 = (float) vacc1x0;
82 float vfpacc1x1 = (float) vacc1x1;
83 float vfpacc2x0 = (float) vacc2x0;
84 float vfpacc2x1 = (float) vacc2x1;
85
86 const float vscale0 = unaligned_indexed_load_f32(w, 0);
87 vfpacc0x0 *= vscale0;
88 vfpacc1x0 *= vscale0;
89 vfpacc2x0 *= vscale0;
90 const float vscale1 = unaligned_indexed_load_f32(w, 1);
91 vfpacc0x1 *= vscale1;
92 vfpacc1x1 *= vscale1;
93 vfpacc2x1 *= vscale1;
94 w = (const void*) ((const float*) w + 2);
95
96 const float voutput_min_less_zero_point = params->fp32_scalar_lrintf.output_min_less_zero_point;
97 vfpacc0x0 = math_max_f32(vfpacc0x0, voutput_min_less_zero_point);
98 vfpacc0x1 = math_max_f32(vfpacc0x1, voutput_min_less_zero_point);
99 vfpacc1x0 = math_max_f32(vfpacc1x0, voutput_min_less_zero_point);
100 vfpacc1x1 = math_max_f32(vfpacc1x1, voutput_min_less_zero_point);
101 vfpacc2x0 = math_max_f32(vfpacc2x0, voutput_min_less_zero_point);
102 vfpacc2x1 = math_max_f32(vfpacc2x1, voutput_min_less_zero_point);
103
104 const float voutput_max_less_zero_point = params->fp32_scalar_lrintf.output_max_less_zero_point;
105 vfpacc0x0 = math_min_f32(vfpacc0x0, voutput_max_less_zero_point);
106 vfpacc0x1 = math_min_f32(vfpacc0x1, voutput_max_less_zero_point);
107 vfpacc1x0 = math_min_f32(vfpacc1x0, voutput_max_less_zero_point);
108 vfpacc1x1 = math_min_f32(vfpacc1x1, voutput_max_less_zero_point);
109 vfpacc2x0 = math_min_f32(vfpacc2x0, voutput_max_less_zero_point);
110 vfpacc2x1 = math_min_f32(vfpacc2x1, voutput_max_less_zero_point);
111
112 const int32_t vrndacc0x0 = (int32_t) lrintf(vfpacc0x0);
113 const int32_t vrndacc0x1 = (int32_t) lrintf(vfpacc0x1);
114 const int32_t vrndacc1x0 = (int32_t) lrintf(vfpacc1x0);
115 const int32_t vrndacc1x1 = (int32_t) lrintf(vfpacc1x1);
116 const int32_t vrndacc2x0 = (int32_t) lrintf(vfpacc2x0);
117 const int32_t vrndacc2x1 = (int32_t) lrintf(vfpacc2x1);
118
119 const int32_t voutput_zero_point = params->fp32_scalar_lrintf.output_zero_point;
120 int32_t vout0x0 = vrndacc0x0 + voutput_zero_point;
121 int32_t vout0x1 = vrndacc0x1 + voutput_zero_point;
122 int32_t vout1x0 = vrndacc1x0 + voutput_zero_point;
123 int32_t vout1x1 = vrndacc1x1 + voutput_zero_point;
124 int32_t vout2x0 = vrndacc2x0 + voutput_zero_point;
125 int32_t vout2x1 = vrndacc2x1 + voutput_zero_point;
126
127 if XNN_LIKELY(nc >= 2) {
128 c0[0] = (int8_t) vout0x0;
129 c0[1] = (int8_t) vout0x1;
130 c1[0] = (int8_t) vout1x0;
131 c1[1] = (int8_t) vout1x1;
132 c2[0] = (int8_t) vout2x0;
133 c2[1] = (int8_t) vout2x1;
134
135 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
136 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
137 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
138
139 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
140 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
141 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
142
143 nc -= 2;
144 } else {
145 if (nc & 1) {
146 c0[0] = (int8_t) vout0x0;
147 c1[0] = (int8_t) vout1x0;
148 c2[0] = (int8_t) vout2x0;
149 }
150
151 nc = 0;
152 }
153 } while (nc != 0);
154 }
155