xref: /aosp_15_r20/external/XNNPACK/src/qc8-gemm/gen/4x4-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_4x4__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_4x4__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 <= 4);
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   const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
48   int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
49   if XNN_UNPREDICTABLE(mr != 4) {
50     a3 = a2;
51     c3 = c2;
52   }
53 
54   do {
55     int32_t vacc0x0 = ((const int32_t*) w)[0];
56     int32_t vacc0x1 = ((const int32_t*) w)[1];
57     int32_t vacc0x2 = ((const int32_t*) w)[2];
58     int32_t vacc0x3 = ((const int32_t*) w)[3];
59     int32_t vacc1x0 = vacc0x0;
60     int32_t vacc1x1 = vacc0x1;
61     int32_t vacc1x2 = vacc0x2;
62     int32_t vacc1x3 = vacc0x3;
63     int32_t vacc2x0 = vacc0x0;
64     int32_t vacc2x1 = vacc0x1;
65     int32_t vacc2x2 = vacc0x2;
66     int32_t vacc2x3 = vacc0x3;
67     int32_t vacc3x0 = vacc0x0;
68     int32_t vacc3x1 = vacc0x1;
69     int32_t vacc3x2 = vacc0x2;
70     int32_t vacc3x3 = vacc0x3;
71     w = (const void*) ((const int32_t*) w + 4);
72 
73     size_t k = kc;
74     do {
75       const int32_t va0 = (int32_t) *a0++;
76       const int32_t va1 = (int32_t) *a1++;
77       const int32_t va2 = (int32_t) *a2++;
78       const int32_t va3 = (int32_t) *a3++;
79 
80       const int32_t vb0 = (int32_t) ((const int8_t*) w)[0];
81       const int32_t vb1 = (int32_t) ((const int8_t*) w)[1];
82       const int32_t vb2 = (int32_t) ((const int8_t*) w)[2];
83       const int32_t vb3 = (int32_t) ((const int8_t*) w)[3];
84       w = (const void*) ((const int8_t*) w + 4);
85 
86       vacc0x0 += va0 * vb0;
87       vacc0x1 += va0 * vb1;
88       vacc0x2 += va0 * vb2;
89       vacc0x3 += va0 * vb3;
90       vacc1x0 += va1 * vb0;
91       vacc1x1 += va1 * vb1;
92       vacc1x2 += va1 * vb2;
93       vacc1x3 += va1 * vb3;
94       vacc2x0 += va2 * vb0;
95       vacc2x1 += va2 * vb1;
96       vacc2x2 += va2 * vb2;
97       vacc2x3 += va2 * vb3;
98       vacc3x0 += va3 * vb0;
99       vacc3x1 += va3 * vb1;
100       vacc3x2 += va3 * vb2;
101       vacc3x3 += va3 * vb3;
102 
103       k -= sizeof(int8_t);
104     } while (k != 0);
105 
106     float vfpacc0x0 = (float) vacc0x0;
107     float vfpacc0x1 = (float) vacc0x1;
108     float vfpacc0x2 = (float) vacc0x2;
109     float vfpacc0x3 = (float) vacc0x3;
110     float vfpacc1x0 = (float) vacc1x0;
111     float vfpacc1x1 = (float) vacc1x1;
112     float vfpacc1x2 = (float) vacc1x2;
113     float vfpacc1x3 = (float) vacc1x3;
114     float vfpacc2x0 = (float) vacc2x0;
115     float vfpacc2x1 = (float) vacc2x1;
116     float vfpacc2x2 = (float) vacc2x2;
117     float vfpacc2x3 = (float) vacc2x3;
118     float vfpacc3x0 = (float) vacc3x0;
119     float vfpacc3x1 = (float) vacc3x1;
120     float vfpacc3x2 = (float) vacc3x2;
121     float vfpacc3x3 = (float) vacc3x3;
122 
123     const float vscale0 = ((const float*) w)[0];
124     vfpacc0x0 *= vscale0;
125     vfpacc1x0 *= vscale0;
126     vfpacc2x0 *= vscale0;
127     vfpacc3x0 *= vscale0;
128     const float vscale1 = ((const float*) w)[1];
129     vfpacc0x1 *= vscale1;
130     vfpacc1x1 *= vscale1;
131     vfpacc2x1 *= vscale1;
132     vfpacc3x1 *= vscale1;
133     const float vscale2 = ((const float*) w)[2];
134     vfpacc0x2 *= vscale2;
135     vfpacc1x2 *= vscale2;
136     vfpacc2x2 *= vscale2;
137     vfpacc3x2 *= vscale2;
138     const float vscale3 = ((const float*) w)[3];
139     vfpacc0x3 *= vscale3;
140     vfpacc1x3 *= vscale3;
141     vfpacc2x3 *= vscale3;
142     vfpacc3x3 *= vscale3;
143     w = (const void*) ((const float*) w + 4);
144 
145     const float vmagic_bias = params->fp32_scalar_imagic.magic_bias;
146     vfpacc0x0 += vmagic_bias;
147     vfpacc0x1 += vmagic_bias;
148     vfpacc0x2 += vmagic_bias;
149     vfpacc0x3 += vmagic_bias;
150     vfpacc1x0 += vmagic_bias;
151     vfpacc1x1 += vmagic_bias;
152     vfpacc1x2 += vmagic_bias;
153     vfpacc1x3 += vmagic_bias;
154     vfpacc2x0 += vmagic_bias;
155     vfpacc2x1 += vmagic_bias;
156     vfpacc2x2 += vmagic_bias;
157     vfpacc2x3 += vmagic_bias;
158     vfpacc3x0 += vmagic_bias;
159     vfpacc3x1 += vmagic_bias;
160     vfpacc3x2 += vmagic_bias;
161     vfpacc3x3 += vmagic_bias;
162 
163     int32_t vout0x0 = (int32_t) float_as_uint32(vfpacc0x0);
164     int32_t vout0x1 = (int32_t) float_as_uint32(vfpacc0x1);
165     int32_t vout0x2 = (int32_t) float_as_uint32(vfpacc0x2);
166     int32_t vout0x3 = (int32_t) float_as_uint32(vfpacc0x3);
167     int32_t vout1x0 = (int32_t) float_as_uint32(vfpacc1x0);
168     int32_t vout1x1 = (int32_t) float_as_uint32(vfpacc1x1);
169     int32_t vout1x2 = (int32_t) float_as_uint32(vfpacc1x2);
170     int32_t vout1x3 = (int32_t) float_as_uint32(vfpacc1x3);
171     int32_t vout2x0 = (int32_t) float_as_uint32(vfpacc2x0);
172     int32_t vout2x1 = (int32_t) float_as_uint32(vfpacc2x1);
173     int32_t vout2x2 = (int32_t) float_as_uint32(vfpacc2x2);
174     int32_t vout2x3 = (int32_t) float_as_uint32(vfpacc2x3);
175     int32_t vout3x0 = (int32_t) float_as_uint32(vfpacc3x0);
176     int32_t vout3x1 = (int32_t) float_as_uint32(vfpacc3x1);
177     int32_t vout3x2 = (int32_t) float_as_uint32(vfpacc3x2);
178     int32_t vout3x3 = (int32_t) float_as_uint32(vfpacc3x3);
179 
180     const int32_t vmagic_min = params->fp32_scalar_imagic.magic_min;
181     vout0x0 = math_max_s32(vout0x0, vmagic_min);
182     vout0x1 = math_max_s32(vout0x1, vmagic_min);
183     vout0x2 = math_max_s32(vout0x2, vmagic_min);
184     vout0x3 = math_max_s32(vout0x3, vmagic_min);
185     vout1x0 = math_max_s32(vout1x0, vmagic_min);
186     vout1x1 = math_max_s32(vout1x1, vmagic_min);
187     vout1x2 = math_max_s32(vout1x2, vmagic_min);
188     vout1x3 = math_max_s32(vout1x3, vmagic_min);
189     vout2x0 = math_max_s32(vout2x0, vmagic_min);
190     vout2x1 = math_max_s32(vout2x1, vmagic_min);
191     vout2x2 = math_max_s32(vout2x2, vmagic_min);
192     vout2x3 = math_max_s32(vout2x3, vmagic_min);
193     vout3x0 = math_max_s32(vout3x0, vmagic_min);
194     vout3x1 = math_max_s32(vout3x1, vmagic_min);
195     vout3x2 = math_max_s32(vout3x2, vmagic_min);
196     vout3x3 = math_max_s32(vout3x3, vmagic_min);
197 
198     const int32_t vmagic_max = params->fp32_scalar_imagic.magic_max;
199     vout0x0 = math_min_s32(vout0x0, vmagic_max);
200     vout0x1 = math_min_s32(vout0x1, vmagic_max);
201     vout0x2 = math_min_s32(vout0x2, vmagic_max);
202     vout0x3 = math_min_s32(vout0x3, vmagic_max);
203     vout1x0 = math_min_s32(vout1x0, vmagic_max);
204     vout1x1 = math_min_s32(vout1x1, vmagic_max);
205     vout1x2 = math_min_s32(vout1x2, vmagic_max);
206     vout1x3 = math_min_s32(vout1x3, vmagic_max);
207     vout2x0 = math_min_s32(vout2x0, vmagic_max);
208     vout2x1 = math_min_s32(vout2x1, vmagic_max);
209     vout2x2 = math_min_s32(vout2x2, vmagic_max);
210     vout2x3 = math_min_s32(vout2x3, vmagic_max);
211     vout3x0 = math_min_s32(vout3x0, vmagic_max);
212     vout3x1 = math_min_s32(vout3x1, vmagic_max);
213     vout3x2 = math_min_s32(vout3x2, vmagic_max);
214     vout3x3 = math_min_s32(vout3x3, vmagic_max);
215 
216     const int32_t vmagic_bias_less_zero_point = params->fp32_scalar_imagic.magic_bias_less_zero_point;
217     vout0x0 -= vmagic_bias_less_zero_point;
218     vout0x1 -= vmagic_bias_less_zero_point;
219     vout0x2 -= vmagic_bias_less_zero_point;
220     vout0x3 -= vmagic_bias_less_zero_point;
221     vout1x0 -= vmagic_bias_less_zero_point;
222     vout1x1 -= vmagic_bias_less_zero_point;
223     vout1x2 -= vmagic_bias_less_zero_point;
224     vout1x3 -= vmagic_bias_less_zero_point;
225     vout2x0 -= vmagic_bias_less_zero_point;
226     vout2x1 -= vmagic_bias_less_zero_point;
227     vout2x2 -= vmagic_bias_less_zero_point;
228     vout2x3 -= vmagic_bias_less_zero_point;
229     vout3x0 -= vmagic_bias_less_zero_point;
230     vout3x1 -= vmagic_bias_less_zero_point;
231     vout3x2 -= vmagic_bias_less_zero_point;
232     vout3x3 -= vmagic_bias_less_zero_point;
233 
234     if XNN_LIKELY(nc >= 4) {
235       c0[0] = (int8_t) vout0x0;
236       c0[1] = (int8_t) vout0x1;
237       c0[2] = (int8_t) vout0x2;
238       c0[3] = (int8_t) vout0x3;
239       c1[0] = (int8_t) vout1x0;
240       c1[1] = (int8_t) vout1x1;
241       c1[2] = (int8_t) vout1x2;
242       c1[3] = (int8_t) vout1x3;
243       c2[0] = (int8_t) vout2x0;
244       c2[1] = (int8_t) vout2x1;
245       c2[2] = (int8_t) vout2x2;
246       c2[3] = (int8_t) vout2x3;
247       c3[0] = (int8_t) vout3x0;
248       c3[1] = (int8_t) vout3x1;
249       c3[2] = (int8_t) vout3x2;
250       c3[3] = (int8_t) vout3x3;
251 
252       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
253       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
254       a2 = (const int8_t*) ((uintptr_t) a2 - kc);
255       a3 = (const int8_t*) ((uintptr_t) a3 - kc);
256 
257       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
258       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
259       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
260       c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
261 
262       nc -= 4;
263     } else {
264       if (nc & 2) {
265         c0[0] = (int8_t) vout0x0;
266         c0[1] = (int8_t) vout0x1;
267         vout0x0 = vout0x2;
268         c0 += 2;
269         c1[0] = (int8_t) vout1x0;
270         c1[1] = (int8_t) vout1x1;
271         vout1x0 = vout1x2;
272         c1 += 2;
273         c2[0] = (int8_t) vout2x0;
274         c2[1] = (int8_t) vout2x1;
275         vout2x0 = vout2x2;
276         c2 += 2;
277         c3[0] = (int8_t) vout3x0;
278         c3[1] = (int8_t) vout3x1;
279         vout3x0 = vout3x2;
280         c3 += 2;
281       }
282       if (nc & 1) {
283         c0[0] = (int8_t) vout0x0;
284         c1[0] = (int8_t) vout1x0;
285         c2[0] = (int8_t) vout2x0;
286         c3[0] = (int8_t) vout3x0;
287       }
288 
289       nc = 0;
290     }
291   } while (nc != 0);
292 }
293