xref: /aosp_15_r20/external/XNNPACK/src/qs8-gemm/gen/3x4-minmax-fp32-scalar-fmagic.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_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_fmagic(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_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])16 void xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_fmagic(
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_qs8_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 vscale = params->fp32_scalar_fmagic.scale;
105     vfpacc0x0 *= vscale;
106     vfpacc0x1 *= vscale;
107     vfpacc0x2 *= vscale;
108     vfpacc0x3 *= vscale;
109     vfpacc1x0 *= vscale;
110     vfpacc1x1 *= vscale;
111     vfpacc1x2 *= vscale;
112     vfpacc1x3 *= vscale;
113     vfpacc2x0 *= vscale;
114     vfpacc2x1 *= vscale;
115     vfpacc2x2 *= vscale;
116     vfpacc2x3 *= vscale;
117 
118     const float voutput_min_less_zero_point = params->fp32_scalar_fmagic.output_min_less_zero_point;
119     vfpacc0x0 = math_max_f32(vfpacc0x0, voutput_min_less_zero_point);
120     vfpacc0x1 = math_max_f32(vfpacc0x1, voutput_min_less_zero_point);
121     vfpacc0x2 = math_max_f32(vfpacc0x2, voutput_min_less_zero_point);
122     vfpacc0x3 = math_max_f32(vfpacc0x3, voutput_min_less_zero_point);
123     vfpacc1x0 = math_max_f32(vfpacc1x0, voutput_min_less_zero_point);
124     vfpacc1x1 = math_max_f32(vfpacc1x1, voutput_min_less_zero_point);
125     vfpacc1x2 = math_max_f32(vfpacc1x2, voutput_min_less_zero_point);
126     vfpacc1x3 = math_max_f32(vfpacc1x3, voutput_min_less_zero_point);
127     vfpacc2x0 = math_max_f32(vfpacc2x0, voutput_min_less_zero_point);
128     vfpacc2x1 = math_max_f32(vfpacc2x1, voutput_min_less_zero_point);
129     vfpacc2x2 = math_max_f32(vfpacc2x2, voutput_min_less_zero_point);
130     vfpacc2x3 = math_max_f32(vfpacc2x3, voutput_min_less_zero_point);
131 
132     const float voutput_max_less_zero_point = params->fp32_scalar_fmagic.output_max_less_zero_point;
133     vfpacc0x0 = math_min_f32(vfpacc0x0, voutput_max_less_zero_point);
134     vfpacc0x1 = math_min_f32(vfpacc0x1, voutput_max_less_zero_point);
135     vfpacc0x2 = math_min_f32(vfpacc0x2, voutput_max_less_zero_point);
136     vfpacc0x3 = math_min_f32(vfpacc0x3, voutput_max_less_zero_point);
137     vfpacc1x0 = math_min_f32(vfpacc1x0, voutput_max_less_zero_point);
138     vfpacc1x1 = math_min_f32(vfpacc1x1, voutput_max_less_zero_point);
139     vfpacc1x2 = math_min_f32(vfpacc1x2, voutput_max_less_zero_point);
140     vfpacc1x3 = math_min_f32(vfpacc1x3, voutput_max_less_zero_point);
141     vfpacc2x0 = math_min_f32(vfpacc2x0, voutput_max_less_zero_point);
142     vfpacc2x1 = math_min_f32(vfpacc2x1, voutput_max_less_zero_point);
143     vfpacc2x2 = math_min_f32(vfpacc2x2, voutput_max_less_zero_point);
144     vfpacc2x3 = math_min_f32(vfpacc2x3, voutput_max_less_zero_point);
145 
146     const float vmagic_bias = params->fp32_scalar_fmagic.magic_bias;
147     vfpacc0x0 += vmagic_bias;
148     vfpacc0x1 += vmagic_bias;
149     vfpacc0x2 += vmagic_bias;
150     vfpacc0x3 += vmagic_bias;
151     vfpacc1x0 += vmagic_bias;
152     vfpacc1x1 += vmagic_bias;
153     vfpacc1x2 += vmagic_bias;
154     vfpacc1x3 += vmagic_bias;
155     vfpacc2x0 += vmagic_bias;
156     vfpacc2x1 += vmagic_bias;
157     vfpacc2x2 += vmagic_bias;
158     vfpacc2x3 += vmagic_bias;
159 
160     const int32_t vmagic_bias_less_output_zero_point = params->fp32_scalar_fmagic.magic_bias_less_output_zero_point;
161     int32_t vout0x0 = (int32_t) float_as_uint32(vfpacc0x0) - vmagic_bias_less_output_zero_point;
162     int32_t vout0x1 = (int32_t) float_as_uint32(vfpacc0x1) - vmagic_bias_less_output_zero_point;
163     int32_t vout0x2 = (int32_t) float_as_uint32(vfpacc0x2) - vmagic_bias_less_output_zero_point;
164     int32_t vout0x3 = (int32_t) float_as_uint32(vfpacc0x3) - vmagic_bias_less_output_zero_point;
165     int32_t vout1x0 = (int32_t) float_as_uint32(vfpacc1x0) - vmagic_bias_less_output_zero_point;
166     int32_t vout1x1 = (int32_t) float_as_uint32(vfpacc1x1) - vmagic_bias_less_output_zero_point;
167     int32_t vout1x2 = (int32_t) float_as_uint32(vfpacc1x2) - vmagic_bias_less_output_zero_point;
168     int32_t vout1x3 = (int32_t) float_as_uint32(vfpacc1x3) - vmagic_bias_less_output_zero_point;
169     int32_t vout2x0 = (int32_t) float_as_uint32(vfpacc2x0) - vmagic_bias_less_output_zero_point;
170     int32_t vout2x1 = (int32_t) float_as_uint32(vfpacc2x1) - vmagic_bias_less_output_zero_point;
171     int32_t vout2x2 = (int32_t) float_as_uint32(vfpacc2x2) - vmagic_bias_less_output_zero_point;
172     int32_t vout2x3 = (int32_t) float_as_uint32(vfpacc2x3) - vmagic_bias_less_output_zero_point;
173 
174     if XNN_LIKELY(nc >= 4) {
175       c0[0] = (int8_t) vout0x0;
176       c0[1] = (int8_t) vout0x1;
177       c0[2] = (int8_t) vout0x2;
178       c0[3] = (int8_t) vout0x3;
179       c1[0] = (int8_t) vout1x0;
180       c1[1] = (int8_t) vout1x1;
181       c1[2] = (int8_t) vout1x2;
182       c1[3] = (int8_t) vout1x3;
183       c2[0] = (int8_t) vout2x0;
184       c2[1] = (int8_t) vout2x1;
185       c2[2] = (int8_t) vout2x2;
186       c2[3] = (int8_t) vout2x3;
187 
188       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
189       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
190       a2 = (const int8_t*) ((uintptr_t) a2 - kc);
191 
192       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
193       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
194       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
195 
196       nc -= 4;
197     } else {
198       if (nc & 2) {
199         c0[0] = (int8_t) vout0x0;
200         c0[1] = (int8_t) vout0x1;
201         vout0x0 = vout0x2;
202         c0 += 2;
203         c1[0] = (int8_t) vout1x0;
204         c1[1] = (int8_t) vout1x1;
205         vout1x0 = vout1x2;
206         c1 += 2;
207         c2[0] = (int8_t) vout2x0;
208         c2[1] = (int8_t) vout2x1;
209         vout2x0 = vout2x2;
210         c2 += 2;
211       }
212       if (nc & 1) {
213         c0[0] = (int8_t) vout0x0;
214         c1[0] = (int8_t) vout1x0;
215         c2[0] = (int8_t) vout2x0;
216       }
217 
218       nc = 0;
219     }
220   } while (nc != 0);
221 }
222