1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/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_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf(size_t mr,size_t nc,size_t kc,size_t ks,const int8_t ** restrict a,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const int8_t * zero,const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const int8_t**restrict a,
24 const void*restrict w,
25 int8_t*restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const int8_t* zero,
30 const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
31 {
32 assert(mr != 0);
33 assert(mr <= 4);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (4 * sizeof(void*)) == 0);
38 assert(a != NULL);
39 assert(w != NULL);
40 assert(c != NULL);
41
42 int8_t* c0 = c;
43 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
44 if XNN_UNPREDICTABLE(mr < 2) {
45 c1 = c0;
46 }
47 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
48 if XNN_UNPREDICTABLE(mr <= 2) {
49 c2 = c1;
50 }
51 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
52 if XNN_UNPREDICTABLE(mr != 4) {
53 c3 = c2;
54 }
55
56 do {
57 int32_t vacc0x0 = unaligned_indexed_load_s32(w, 0);
58 int32_t vacc0x1 = unaligned_indexed_load_s32(w, 1);
59 int32_t vacc1x0 = vacc0x0;
60 int32_t vacc1x1 = vacc0x1;
61 int32_t vacc2x0 = vacc0x0;
62 int32_t vacc2x1 = vacc0x1;
63 int32_t vacc3x0 = vacc0x0;
64 int32_t vacc3x1 = vacc0x1;
65 w = (const void*) ((const int32_t*) w + 2);
66
67 size_t p = ks;
68 do {
69 const int8_t* restrict a0 = a[0];
70 assert(a0 != NULL);
71 if XNN_UNPREDICTABLE(a0 != zero) {
72 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
73 }
74 const int8_t* restrict a1 = a[1];
75 assert(a1 != NULL);
76 if XNN_UNPREDICTABLE(a1 != zero) {
77 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
78 }
79 const int8_t* restrict a2 = a[2];
80 assert(a2 != NULL);
81 if XNN_UNPREDICTABLE(a2 != zero) {
82 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
83 }
84 const int8_t* restrict a3 = a[3];
85 assert(a3 != NULL);
86 if XNN_UNPREDICTABLE(a3 != zero) {
87 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
88 }
89 a += 4;
90
91 size_t k = kc;
92 do {
93 const int32_t va0 = (int32_t) *a0++;
94 const int32_t va1 = (int32_t) *a1++;
95 const int32_t va2 = (int32_t) *a2++;
96 const int32_t va3 = (int32_t) *a3++;
97
98 const int32_t vb0 = (int32_t) ((const int8_t*) w)[0];
99 const int32_t vb1 = (int32_t) ((const int8_t*) w)[1];
100 w = (const void*) ((const int8_t*) w + 2);
101
102 vacc0x0 += va0 * vb0;
103 vacc0x1 += va0 * vb1;
104 vacc1x0 += va1 * vb0;
105 vacc1x1 += va1 * vb1;
106 vacc2x0 += va2 * vb0;
107 vacc2x1 += va2 * vb1;
108 vacc3x0 += va3 * vb0;
109 vacc3x1 += va3 * vb1;
110
111 k -= sizeof(int8_t);
112 } while (k != 0);
113 p -= 4 * sizeof(void*);
114 } while (p != 0);
115
116 float vfpacc0x0 = (float) vacc0x0;
117 float vfpacc0x1 = (float) vacc0x1;
118 float vfpacc1x0 = (float) vacc1x0;
119 float vfpacc1x1 = (float) vacc1x1;
120 float vfpacc2x0 = (float) vacc2x0;
121 float vfpacc2x1 = (float) vacc2x1;
122 float vfpacc3x0 = (float) vacc3x0;
123 float vfpacc3x1 = (float) vacc3x1;
124
125 const float vscale0 = unaligned_indexed_load_f32(w, 0);
126 vfpacc0x0 *= vscale0;
127 vfpacc1x0 *= vscale0;
128 vfpacc2x0 *= vscale0;
129 vfpacc3x0 *= vscale0;
130 const float vscale1 = unaligned_indexed_load_f32(w, 1);
131 vfpacc0x1 *= vscale1;
132 vfpacc1x1 *= vscale1;
133 vfpacc2x1 *= vscale1;
134 vfpacc3x1 *= vscale1;
135 w = (const void*) ((const float*) w + 2);
136
137 const float voutput_min_less_zero_point = params->fp32_scalar_lrintf.output_min_less_zero_point;
138 vfpacc0x0 = math_max_f32(vfpacc0x0, voutput_min_less_zero_point);
139 vfpacc0x1 = math_max_f32(vfpacc0x1, voutput_min_less_zero_point);
140 vfpacc1x0 = math_max_f32(vfpacc1x0, voutput_min_less_zero_point);
141 vfpacc1x1 = math_max_f32(vfpacc1x1, voutput_min_less_zero_point);
142 vfpacc2x0 = math_max_f32(vfpacc2x0, voutput_min_less_zero_point);
143 vfpacc2x1 = math_max_f32(vfpacc2x1, voutput_min_less_zero_point);
144 vfpacc3x0 = math_max_f32(vfpacc3x0, voutput_min_less_zero_point);
145 vfpacc3x1 = math_max_f32(vfpacc3x1, voutput_min_less_zero_point);
146
147 const float voutput_max_less_zero_point = params->fp32_scalar_lrintf.output_max_less_zero_point;
148 vfpacc0x0 = math_min_f32(vfpacc0x0, voutput_max_less_zero_point);
149 vfpacc0x1 = math_min_f32(vfpacc0x1, voutput_max_less_zero_point);
150 vfpacc1x0 = math_min_f32(vfpacc1x0, voutput_max_less_zero_point);
151 vfpacc1x1 = math_min_f32(vfpacc1x1, voutput_max_less_zero_point);
152 vfpacc2x0 = math_min_f32(vfpacc2x0, voutput_max_less_zero_point);
153 vfpacc2x1 = math_min_f32(vfpacc2x1, voutput_max_less_zero_point);
154 vfpacc3x0 = math_min_f32(vfpacc3x0, voutput_max_less_zero_point);
155 vfpacc3x1 = math_min_f32(vfpacc3x1, voutput_max_less_zero_point);
156
157 const int32_t vrndacc0x0 = (int32_t) lrintf(vfpacc0x0);
158 const int32_t vrndacc0x1 = (int32_t) lrintf(vfpacc0x1);
159 const int32_t vrndacc1x0 = (int32_t) lrintf(vfpacc1x0);
160 const int32_t vrndacc1x1 = (int32_t) lrintf(vfpacc1x1);
161 const int32_t vrndacc2x0 = (int32_t) lrintf(vfpacc2x0);
162 const int32_t vrndacc2x1 = (int32_t) lrintf(vfpacc2x1);
163 const int32_t vrndacc3x0 = (int32_t) lrintf(vfpacc3x0);
164 const int32_t vrndacc3x1 = (int32_t) lrintf(vfpacc3x1);
165
166 const int32_t voutput_zero_point = params->fp32_scalar_lrintf.output_zero_point;
167 int32_t vout0x0 = vrndacc0x0 + voutput_zero_point;
168 int32_t vout0x1 = vrndacc0x1 + voutput_zero_point;
169 int32_t vout1x0 = vrndacc1x0 + voutput_zero_point;
170 int32_t vout1x1 = vrndacc1x1 + voutput_zero_point;
171 int32_t vout2x0 = vrndacc2x0 + voutput_zero_point;
172 int32_t vout2x1 = vrndacc2x1 + voutput_zero_point;
173 int32_t vout3x0 = vrndacc3x0 + voutput_zero_point;
174 int32_t vout3x1 = vrndacc3x1 + voutput_zero_point;
175
176 if XNN_LIKELY(nc >= 2) {
177 c3[0] = (int8_t) vout3x0;
178 c3[1] = (int8_t) vout3x1;
179 c2[0] = (int8_t) vout2x0;
180 c2[1] = (int8_t) vout2x1;
181 c1[0] = (int8_t) vout1x0;
182 c1[1] = (int8_t) vout1x1;
183 c0[0] = (int8_t) vout0x0;
184 c0[1] = (int8_t) vout0x1;
185
186 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
187 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
188 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
189 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
190
191 a = (const int8_t**restrict) ((uintptr_t) a - ks);
192 nc -= 2;
193 } else {
194 if (nc & 1) {
195 c3[0] = (int8_t) vout3x0;
196 c2[0] = (int8_t) vout2x0;
197 c1[0] = (int8_t) vout1x0;
198 c0[0] = (int8_t) vout0x0;
199 }
200
201 nc = 0;
202 }
203 } while (nc != 0);
204 }
205