xref: /aosp_15_r20/external/XNNPACK/src/qu8-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_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic(size_t mr,size_t nc,size_t kc,const uint8_t * restrict a,size_t a_stride,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])16 void xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic(
17     size_t mr,
18     size_t nc,
19     size_t kc,
20     const uint8_t* restrict a,
21     size_t a_stride,
22     const void* restrict w,
23     uint8_t* restrict c,
24     size_t cm_stride,
25     size_t cn_stride,
26     const union xnn_qu8_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 uint8_t* a0 = a;
34   uint8_t* c0 = c;
35   const uint8_t* a1 = (const uint8_t*) ((uintptr_t) a0 + a_stride);
36   uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
37   if XNN_UNPREDICTABLE(mr < 2) {
38     a1 = a0;
39     c1 = c0;
40   }
41   const uint8_t* a2 = (const uint8_t*) ((uintptr_t) a1 + a_stride);
42   uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
43   if XNN_UNPREDICTABLE(mr <= 2) {
44     a2 = a1;
45     c2 = c1;
46   }
47 
48   const int32_t vb_zero_point = params->fp32_scalar_imagic.kernel_zero_point;
49   do {
50     int32_t vacc0x0 = ((const int32_t*) w)[0];
51     int32_t vacc0x1 = ((const int32_t*) w)[1];
52     int32_t vacc0x2 = ((const int32_t*) w)[2];
53     int32_t vacc0x3 = ((const int32_t*) w)[3];
54     int32_t vacc1x0 = vacc0x0;
55     int32_t vacc1x1 = vacc0x1;
56     int32_t vacc1x2 = vacc0x2;
57     int32_t vacc1x3 = vacc0x3;
58     int32_t vacc2x0 = vacc0x0;
59     int32_t vacc2x1 = vacc0x1;
60     int32_t vacc2x2 = vacc0x2;
61     int32_t vacc2x3 = vacc0x3;
62     w = (const void*) ((const int32_t*) w + 4);
63 
64     size_t k = kc;
65     do {
66       const int32_t va0 = (int32_t) (uint32_t) *a0++;
67       const int32_t va1 = (int32_t) (uint32_t) *a1++;
68       const int32_t va2 = (int32_t) (uint32_t) *a2++;
69 
70       const int32_t vb0 = (int32_t) (uint32_t) ((const uint8_t*) w)[0] - vb_zero_point;
71       const int32_t vb1 = (int32_t) (uint32_t) ((const uint8_t*) w)[1] - vb_zero_point;
72       const int32_t vb2 = (int32_t) (uint32_t) ((const uint8_t*) w)[2] - vb_zero_point;
73       const int32_t vb3 = (int32_t) (uint32_t) ((const uint8_t*) w)[3] - vb_zero_point;
74       w = (const void*) ((const uint8_t*) w + 4);
75 
76       vacc0x0 += va0 * vb0;
77       vacc0x1 += va0 * vb1;
78       vacc0x2 += va0 * vb2;
79       vacc0x3 += va0 * vb3;
80       vacc1x0 += va1 * vb0;
81       vacc1x1 += va1 * vb1;
82       vacc1x2 += va1 * vb2;
83       vacc1x3 += va1 * vb3;
84       vacc2x0 += va2 * vb0;
85       vacc2x1 += va2 * vb1;
86       vacc2x2 += va2 * vb2;
87       vacc2x3 += va2 * vb3;
88 
89       k -= sizeof(uint8_t);
90     } while (k != 0);
91 
92     float vfpacc0x0 = (float) vacc0x0;
93     float vfpacc0x1 = (float) vacc0x1;
94     float vfpacc0x2 = (float) vacc0x2;
95     float vfpacc0x3 = (float) vacc0x3;
96     float vfpacc1x0 = (float) vacc1x0;
97     float vfpacc1x1 = (float) vacc1x1;
98     float vfpacc1x2 = (float) vacc1x2;
99     float vfpacc1x3 = (float) vacc1x3;
100     float vfpacc2x0 = (float) vacc2x0;
101     float vfpacc2x1 = (float) vacc2x1;
102     float vfpacc2x2 = (float) vacc2x2;
103     float vfpacc2x3 = (float) vacc2x3;
104 
105     const float vscale = params->fp32_scalar_imagic.scale;
106     vfpacc0x0 *= vscale;
107     vfpacc0x1 *= vscale;
108     vfpacc0x2 *= vscale;
109     vfpacc0x3 *= vscale;
110     vfpacc1x0 *= vscale;
111     vfpacc1x1 *= vscale;
112     vfpacc1x2 *= vscale;
113     vfpacc1x3 *= vscale;
114     vfpacc2x0 *= vscale;
115     vfpacc2x1 *= vscale;
116     vfpacc2x2 *= vscale;
117     vfpacc2x3 *= vscale;
118 
119     const float vmagic_bias = params->fp32_scalar_imagic.magic_bias;
120     vfpacc0x0 += vmagic_bias;
121     vfpacc0x1 += vmagic_bias;
122     vfpacc0x2 += vmagic_bias;
123     vfpacc0x3 += vmagic_bias;
124     vfpacc1x0 += vmagic_bias;
125     vfpacc1x1 += vmagic_bias;
126     vfpacc1x2 += vmagic_bias;
127     vfpacc1x3 += vmagic_bias;
128     vfpacc2x0 += vmagic_bias;
129     vfpacc2x1 += vmagic_bias;
130     vfpacc2x2 += vmagic_bias;
131     vfpacc2x3 += vmagic_bias;
132 
133     int32_t vout0x0 = (int32_t) float_as_uint32(vfpacc0x0);
134     int32_t vout0x1 = (int32_t) float_as_uint32(vfpacc0x1);
135     int32_t vout0x2 = (int32_t) float_as_uint32(vfpacc0x2);
136     int32_t vout0x3 = (int32_t) float_as_uint32(vfpacc0x3);
137     int32_t vout1x0 = (int32_t) float_as_uint32(vfpacc1x0);
138     int32_t vout1x1 = (int32_t) float_as_uint32(vfpacc1x1);
139     int32_t vout1x2 = (int32_t) float_as_uint32(vfpacc1x2);
140     int32_t vout1x3 = (int32_t) float_as_uint32(vfpacc1x3);
141     int32_t vout2x0 = (int32_t) float_as_uint32(vfpacc2x0);
142     int32_t vout2x1 = (int32_t) float_as_uint32(vfpacc2x1);
143     int32_t vout2x2 = (int32_t) float_as_uint32(vfpacc2x2);
144     int32_t vout2x3 = (int32_t) float_as_uint32(vfpacc2x3);
145 
146     const int32_t vmagic_min = params->fp32_scalar_imagic.magic_min;
147     vout0x0 = math_max_s32(vout0x0, vmagic_min);
148     vout0x1 = math_max_s32(vout0x1, vmagic_min);
149     vout0x2 = math_max_s32(vout0x2, vmagic_min);
150     vout0x3 = math_max_s32(vout0x3, vmagic_min);
151     vout1x0 = math_max_s32(vout1x0, vmagic_min);
152     vout1x1 = math_max_s32(vout1x1, vmagic_min);
153     vout1x2 = math_max_s32(vout1x2, vmagic_min);
154     vout1x3 = math_max_s32(vout1x3, vmagic_min);
155     vout2x0 = math_max_s32(vout2x0, vmagic_min);
156     vout2x1 = math_max_s32(vout2x1, vmagic_min);
157     vout2x2 = math_max_s32(vout2x2, vmagic_min);
158     vout2x3 = math_max_s32(vout2x3, vmagic_min);
159 
160     const int32_t vmagic_max = params->fp32_scalar_imagic.magic_max;
161     vout0x0 = math_min_s32(vout0x0, vmagic_max);
162     vout0x1 = math_min_s32(vout0x1, vmagic_max);
163     vout0x2 = math_min_s32(vout0x2, vmagic_max);
164     vout0x3 = math_min_s32(vout0x3, vmagic_max);
165     vout1x0 = math_min_s32(vout1x0, vmagic_max);
166     vout1x1 = math_min_s32(vout1x1, vmagic_max);
167     vout1x2 = math_min_s32(vout1x2, vmagic_max);
168     vout1x3 = math_min_s32(vout1x3, vmagic_max);
169     vout2x0 = math_min_s32(vout2x0, vmagic_max);
170     vout2x1 = math_min_s32(vout2x1, vmagic_max);
171     vout2x2 = math_min_s32(vout2x2, vmagic_max);
172     vout2x3 = math_min_s32(vout2x3, vmagic_max);
173 
174     const int32_t vmagic_bias_less_zero_point = params->fp32_scalar_imagic.magic_bias_less_zero_point;
175     vout0x0 -= vmagic_bias_less_zero_point;
176     vout0x1 -= vmagic_bias_less_zero_point;
177     vout0x2 -= vmagic_bias_less_zero_point;
178     vout0x3 -= vmagic_bias_less_zero_point;
179     vout1x0 -= vmagic_bias_less_zero_point;
180     vout1x1 -= vmagic_bias_less_zero_point;
181     vout1x2 -= vmagic_bias_less_zero_point;
182     vout1x3 -= vmagic_bias_less_zero_point;
183     vout2x0 -= vmagic_bias_less_zero_point;
184     vout2x1 -= vmagic_bias_less_zero_point;
185     vout2x2 -= vmagic_bias_less_zero_point;
186     vout2x3 -= vmagic_bias_less_zero_point;
187 
188     if XNN_LIKELY(nc >= 4) {
189       c0[0] = (uint8_t) vout0x0;
190       c0[1] = (uint8_t) vout0x1;
191       c0[2] = (uint8_t) vout0x2;
192       c0[3] = (uint8_t) vout0x3;
193       c1[0] = (uint8_t) vout1x0;
194       c1[1] = (uint8_t) vout1x1;
195       c1[2] = (uint8_t) vout1x2;
196       c1[3] = (uint8_t) vout1x3;
197       c2[0] = (uint8_t) vout2x0;
198       c2[1] = (uint8_t) vout2x1;
199       c2[2] = (uint8_t) vout2x2;
200       c2[3] = (uint8_t) vout2x3;
201 
202       a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
203       a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
204       a2 = (const uint8_t*) ((uintptr_t) a2 - kc);
205 
206       c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
207       c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
208       c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
209 
210       nc -= 4;
211     } else {
212       if (nc & 2) {
213         c0[0] = (uint8_t) vout0x0;
214         c0[1] = (uint8_t) vout0x1;
215         vout0x0 = vout0x2;
216         c0 += 2;
217         c1[0] = (uint8_t) vout1x0;
218         c1[1] = (uint8_t) vout1x1;
219         vout1x0 = vout1x2;
220         c1 += 2;
221         c2[0] = (uint8_t) vout2x0;
222         c2[1] = (uint8_t) vout2x1;
223         vout2x0 = vout2x2;
224         c2 += 2;
225       }
226       if (nc & 1) {
227         c0[0] = (uint8_t) vout0x0;
228         c1[0] = (uint8_t) vout1x0;
229         c2[0] = (uint8_t) vout2x0;
230       }
231 
232       nc = 0;
233     }
234   } while (nc != 0);
235 }
236