xref: /aosp_15_r20/external/XNNPACK/src/qc8-gemm/gen/3x4-minmax-fp32-scalar-imagic.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
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 
15 
xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic(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)])16 void xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic(
17     size_t mr,
18     size_t nc,
19     size_t kc,
20     const int8_t* restrict a,
21     size_t a_stride,
22     const void* restrict w,
23     int8_t* restrict c,
24     size_t cm_stride,
25     size_t cn_stride,
26     const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
27 {
28   assert(mr != 0);
29   assert(mr <= 3);
30   assert(nc != 0);
31   assert(kc != 0);
32 
33   const int8_t* a0 = a;
34   int8_t* c0 = c;
35   const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
36   int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
37   if XNN_UNPREDICTABLE(mr < 2) {
38     a1 = a0;
39     c1 = c0;
40   }
41   const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
42   int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
43   if XNN_UNPREDICTABLE(mr <= 2) {
44     a2 = a1;
45     c2 = c1;
46   }
47 
48   do {
49     int32_t vacc0x0 = ((const int32_t*) w)[0];
50     int32_t vacc0x1 = ((const int32_t*) w)[1];
51     int32_t vacc0x2 = ((const int32_t*) w)[2];
52     int32_t vacc0x3 = ((const int32_t*) w)[3];
53     int32_t vacc1x0 = vacc0x0;
54     int32_t vacc1x1 = vacc0x1;
55     int32_t vacc1x2 = vacc0x2;
56     int32_t vacc1x3 = vacc0x3;
57     int32_t vacc2x0 = vacc0x0;
58     int32_t vacc2x1 = vacc0x1;
59     int32_t vacc2x2 = vacc0x2;
60     int32_t vacc2x3 = vacc0x3;
61     w = (const void*) ((const int32_t*) w + 4);
62 
63     size_t k = kc;
64     do {
65       const int32_t va0 = (int32_t) *a0++;
66       const int32_t va1 = (int32_t) *a1++;
67       const int32_t va2 = (int32_t) *a2++;
68 
69       const int32_t vb0 = (int32_t) ((const int8_t*) w)[0];
70       const int32_t vb1 = (int32_t) ((const int8_t*) w)[1];
71       const int32_t vb2 = (int32_t) ((const int8_t*) w)[2];
72       const int32_t vb3 = (int32_t) ((const int8_t*) w)[3];
73       w = (const void*) ((const int8_t*) w + 4);
74 
75       vacc0x0 += va0 * vb0;
76       vacc0x1 += va0 * vb1;
77       vacc0x2 += va0 * vb2;
78       vacc0x3 += va0 * vb3;
79       vacc1x0 += va1 * vb0;
80       vacc1x1 += va1 * vb1;
81       vacc1x2 += va1 * vb2;
82       vacc1x3 += va1 * vb3;
83       vacc2x0 += va2 * vb0;
84       vacc2x1 += va2 * vb1;
85       vacc2x2 += va2 * vb2;
86       vacc2x3 += va2 * vb3;
87 
88       k -= sizeof(int8_t);
89     } while (k != 0);
90 
91     float vfpacc0x0 = (float) vacc0x0;
92     float vfpacc0x1 = (float) vacc0x1;
93     float vfpacc0x2 = (float) vacc0x2;
94     float vfpacc0x3 = (float) vacc0x3;
95     float vfpacc1x0 = (float) vacc1x0;
96     float vfpacc1x1 = (float) vacc1x1;
97     float vfpacc1x2 = (float) vacc1x2;
98     float vfpacc1x3 = (float) vacc1x3;
99     float vfpacc2x0 = (float) vacc2x0;
100     float vfpacc2x1 = (float) vacc2x1;
101     float vfpacc2x2 = (float) vacc2x2;
102     float vfpacc2x3 = (float) vacc2x3;
103 
104     const float vscale0 = ((const float*) w)[0];
105     vfpacc0x0 *= vscale0;
106     vfpacc1x0 *= vscale0;
107     vfpacc2x0 *= vscale0;
108     const float vscale1 = ((const float*) w)[1];
109     vfpacc0x1 *= vscale1;
110     vfpacc1x1 *= vscale1;
111     vfpacc2x1 *= vscale1;
112     const float vscale2 = ((const float*) w)[2];
113     vfpacc0x2 *= vscale2;
114     vfpacc1x2 *= vscale2;
115     vfpacc2x2 *= vscale2;
116     const float vscale3 = ((const float*) w)[3];
117     vfpacc0x3 *= vscale3;
118     vfpacc1x3 *= vscale3;
119     vfpacc2x3 *= vscale3;
120     w = (const void*) ((const float*) w + 4);
121 
122     const float vmagic_bias = params->fp32_scalar_imagic.magic_bias;
123     vfpacc0x0 += vmagic_bias;
124     vfpacc0x1 += vmagic_bias;
125     vfpacc0x2 += vmagic_bias;
126     vfpacc0x3 += vmagic_bias;
127     vfpacc1x0 += vmagic_bias;
128     vfpacc1x1 += vmagic_bias;
129     vfpacc1x2 += vmagic_bias;
130     vfpacc1x3 += vmagic_bias;
131     vfpacc2x0 += vmagic_bias;
132     vfpacc2x1 += vmagic_bias;
133     vfpacc2x2 += vmagic_bias;
134     vfpacc2x3 += vmagic_bias;
135 
136     int32_t vout0x0 = (int32_t) float_as_uint32(vfpacc0x0);
137     int32_t vout0x1 = (int32_t) float_as_uint32(vfpacc0x1);
138     int32_t vout0x2 = (int32_t) float_as_uint32(vfpacc0x2);
139     int32_t vout0x3 = (int32_t) float_as_uint32(vfpacc0x3);
140     int32_t vout1x0 = (int32_t) float_as_uint32(vfpacc1x0);
141     int32_t vout1x1 = (int32_t) float_as_uint32(vfpacc1x1);
142     int32_t vout1x2 = (int32_t) float_as_uint32(vfpacc1x2);
143     int32_t vout1x3 = (int32_t) float_as_uint32(vfpacc1x3);
144     int32_t vout2x0 = (int32_t) float_as_uint32(vfpacc2x0);
145     int32_t vout2x1 = (int32_t) float_as_uint32(vfpacc2x1);
146     int32_t vout2x2 = (int32_t) float_as_uint32(vfpacc2x2);
147     int32_t vout2x3 = (int32_t) float_as_uint32(vfpacc2x3);
148 
149     const int32_t vmagic_min = params->fp32_scalar_imagic.magic_min;
150     vout0x0 = math_max_s32(vout0x0, vmagic_min);
151     vout0x1 = math_max_s32(vout0x1, vmagic_min);
152     vout0x2 = math_max_s32(vout0x2, vmagic_min);
153     vout0x3 = math_max_s32(vout0x3, vmagic_min);
154     vout1x0 = math_max_s32(vout1x0, vmagic_min);
155     vout1x1 = math_max_s32(vout1x1, vmagic_min);
156     vout1x2 = math_max_s32(vout1x2, vmagic_min);
157     vout1x3 = math_max_s32(vout1x3, vmagic_min);
158     vout2x0 = math_max_s32(vout2x0, vmagic_min);
159     vout2x1 = math_max_s32(vout2x1, vmagic_min);
160     vout2x2 = math_max_s32(vout2x2, vmagic_min);
161     vout2x3 = math_max_s32(vout2x3, vmagic_min);
162 
163     const int32_t vmagic_max = params->fp32_scalar_imagic.magic_max;
164     vout0x0 = math_min_s32(vout0x0, vmagic_max);
165     vout0x1 = math_min_s32(vout0x1, vmagic_max);
166     vout0x2 = math_min_s32(vout0x2, vmagic_max);
167     vout0x3 = math_min_s32(vout0x3, vmagic_max);
168     vout1x0 = math_min_s32(vout1x0, vmagic_max);
169     vout1x1 = math_min_s32(vout1x1, vmagic_max);
170     vout1x2 = math_min_s32(vout1x2, vmagic_max);
171     vout1x3 = math_min_s32(vout1x3, vmagic_max);
172     vout2x0 = math_min_s32(vout2x0, vmagic_max);
173     vout2x1 = math_min_s32(vout2x1, vmagic_max);
174     vout2x2 = math_min_s32(vout2x2, vmagic_max);
175     vout2x3 = math_min_s32(vout2x3, vmagic_max);
176 
177     const int32_t vmagic_bias_less_zero_point = params->fp32_scalar_imagic.magic_bias_less_zero_point;
178     vout0x0 -= vmagic_bias_less_zero_point;
179     vout0x1 -= vmagic_bias_less_zero_point;
180     vout0x2 -= vmagic_bias_less_zero_point;
181     vout0x3 -= vmagic_bias_less_zero_point;
182     vout1x0 -= vmagic_bias_less_zero_point;
183     vout1x1 -= vmagic_bias_less_zero_point;
184     vout1x2 -= vmagic_bias_less_zero_point;
185     vout1x3 -= vmagic_bias_less_zero_point;
186     vout2x0 -= vmagic_bias_less_zero_point;
187     vout2x1 -= vmagic_bias_less_zero_point;
188     vout2x2 -= vmagic_bias_less_zero_point;
189     vout2x3 -= vmagic_bias_less_zero_point;
190 
191     if XNN_LIKELY(nc >= 4) {
192       c0[0] = (int8_t) vout0x0;
193       c0[1] = (int8_t) vout0x1;
194       c0[2] = (int8_t) vout0x2;
195       c0[3] = (int8_t) vout0x3;
196       c1[0] = (int8_t) vout1x0;
197       c1[1] = (int8_t) vout1x1;
198       c1[2] = (int8_t) vout1x2;
199       c1[3] = (int8_t) vout1x3;
200       c2[0] = (int8_t) vout2x0;
201       c2[1] = (int8_t) vout2x1;
202       c2[2] = (int8_t) vout2x2;
203       c2[3] = (int8_t) vout2x3;
204 
205       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
206       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
207       a2 = (const int8_t*) ((uintptr_t) a2 - kc);
208 
209       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
210       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
211       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
212 
213       nc -= 4;
214     } else {
215       if (nc & 2) {
216         c0[0] = (int8_t) vout0x0;
217         c0[1] = (int8_t) vout0x1;
218         vout0x0 = vout0x2;
219         c0 += 2;
220         c1[0] = (int8_t) vout1x0;
221         c1[1] = (int8_t) vout1x1;
222         vout1x0 = vout1x2;
223         c1 += 2;
224         c2[0] = (int8_t) vout2x0;
225         c2[1] = (int8_t) vout2x1;
226         vout2x0 = vout2x2;
227         c2 += 2;
228       }
229       if (nc & 1) {
230         c0[0] = (int8_t) vout0x0;
231         c1[0] = (int8_t) vout1x0;
232         c2[0] = (int8_t) vout2x0;
233       }
234 
235       nc = 0;
236     }
237   } while (nc != 0);
238 }
239