xref: /aosp_15_r20/external/XNNPACK/src/qs8-gemm/gen/4x2-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 #include <xnnpack/unaligned.h>
15 
16 
xnn_qs8_gemm_minmax_fp32_ukernel_4x2__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)])17 void xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_fmagic(
18     size_t mr,
19     size_t nc,
20     size_t kc,
21     const int8_t* restrict a,
22     size_t a_stride,
23     const void* restrict w,
24     int8_t* restrict c,
25     size_t cm_stride,
26     size_t cn_stride,
27     const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
28 {
29   assert(mr != 0);
30   assert(mr <= 4);
31   assert(nc != 0);
32   assert(kc != 0);
33 
34   const int8_t* a0 = a;
35   int8_t* c0 = c;
36   const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
37   int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
38   if XNN_UNPREDICTABLE(mr < 2) {
39     a1 = a0;
40     c1 = c0;
41   }
42   const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
43   int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
44   if XNN_UNPREDICTABLE(mr <= 2) {
45     a2 = a1;
46     c2 = c1;
47   }
48   const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
49   int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
50   if XNN_UNPREDICTABLE(mr != 4) {
51     a3 = a2;
52     c3 = c2;
53   }
54 
55   do {
56     int32_t vacc0x0 = unaligned_indexed_load_s32(w, 0);
57     int32_t vacc0x1 = unaligned_indexed_load_s32(w, 1);
58     int32_t vacc1x0 = vacc0x0;
59     int32_t vacc1x1 = vacc0x1;
60     int32_t vacc2x0 = vacc0x0;
61     int32_t vacc2x1 = vacc0x1;
62     int32_t vacc3x0 = vacc0x0;
63     int32_t vacc3x1 = vacc0x1;
64     w = (const void*) ((const int32_t*) w + 2);
65 
66     size_t k = kc;
67     do {
68       const int32_t va0 = (int32_t) *a0++;
69       const int32_t va1 = (int32_t) *a1++;
70       const int32_t va2 = (int32_t) *a2++;
71       const int32_t va3 = (int32_t) *a3++;
72 
73       const int32_t vb0 = (int32_t) ((const int8_t*) w)[0];
74       const int32_t vb1 = (int32_t) ((const int8_t*) w)[1];
75       w = (const void*) ((const int8_t*) w + 2);
76 
77       vacc0x0 += va0 * vb0;
78       vacc0x1 += va0 * vb1;
79       vacc1x0 += va1 * vb0;
80       vacc1x1 += va1 * vb1;
81       vacc2x0 += va2 * vb0;
82       vacc2x1 += va2 * vb1;
83       vacc3x0 += va3 * vb0;
84       vacc3x1 += va3 * vb1;
85 
86       k -= sizeof(int8_t);
87     } while (k != 0);
88 
89     float vfpacc0x0 = (float) vacc0x0;
90     float vfpacc0x1 = (float) vacc0x1;
91     float vfpacc1x0 = (float) vacc1x0;
92     float vfpacc1x1 = (float) vacc1x1;
93     float vfpacc2x0 = (float) vacc2x0;
94     float vfpacc2x1 = (float) vacc2x1;
95     float vfpacc3x0 = (float) vacc3x0;
96     float vfpacc3x1 = (float) vacc3x1;
97 
98     const float vscale = params->fp32_scalar_fmagic.scale;
99     vfpacc0x0 *= vscale;
100     vfpacc0x1 *= vscale;
101     vfpacc1x0 *= vscale;
102     vfpacc1x1 *= vscale;
103     vfpacc2x0 *= vscale;
104     vfpacc2x1 *= vscale;
105     vfpacc3x0 *= vscale;
106     vfpacc3x1 *= vscale;
107 
108     const float voutput_min_less_zero_point = params->fp32_scalar_fmagic.output_min_less_zero_point;
109     vfpacc0x0 = math_max_f32(vfpacc0x0, voutput_min_less_zero_point);
110     vfpacc0x1 = math_max_f32(vfpacc0x1, voutput_min_less_zero_point);
111     vfpacc1x0 = math_max_f32(vfpacc1x0, voutput_min_less_zero_point);
112     vfpacc1x1 = math_max_f32(vfpacc1x1, voutput_min_less_zero_point);
113     vfpacc2x0 = math_max_f32(vfpacc2x0, voutput_min_less_zero_point);
114     vfpacc2x1 = math_max_f32(vfpacc2x1, voutput_min_less_zero_point);
115     vfpacc3x0 = math_max_f32(vfpacc3x0, voutput_min_less_zero_point);
116     vfpacc3x1 = math_max_f32(vfpacc3x1, voutput_min_less_zero_point);
117 
118     const float voutput_max_less_zero_point = params->fp32_scalar_fmagic.output_max_less_zero_point;
119     vfpacc0x0 = math_min_f32(vfpacc0x0, voutput_max_less_zero_point);
120     vfpacc0x1 = math_min_f32(vfpacc0x1, voutput_max_less_zero_point);
121     vfpacc1x0 = math_min_f32(vfpacc1x0, voutput_max_less_zero_point);
122     vfpacc1x1 = math_min_f32(vfpacc1x1, voutput_max_less_zero_point);
123     vfpacc2x0 = math_min_f32(vfpacc2x0, voutput_max_less_zero_point);
124     vfpacc2x1 = math_min_f32(vfpacc2x1, voutput_max_less_zero_point);
125     vfpacc3x0 = math_min_f32(vfpacc3x0, voutput_max_less_zero_point);
126     vfpacc3x1 = math_min_f32(vfpacc3x1, voutput_max_less_zero_point);
127 
128     const float vmagic_bias = params->fp32_scalar_fmagic.magic_bias;
129     vfpacc0x0 += vmagic_bias;
130     vfpacc0x1 += vmagic_bias;
131     vfpacc1x0 += vmagic_bias;
132     vfpacc1x1 += vmagic_bias;
133     vfpacc2x0 += vmagic_bias;
134     vfpacc2x1 += vmagic_bias;
135     vfpacc3x0 += vmagic_bias;
136     vfpacc3x1 += vmagic_bias;
137 
138     const int32_t vmagic_bias_less_output_zero_point = params->fp32_scalar_fmagic.magic_bias_less_output_zero_point;
139     int32_t vout0x0 = (int32_t) float_as_uint32(vfpacc0x0) - vmagic_bias_less_output_zero_point;
140     int32_t vout0x1 = (int32_t) float_as_uint32(vfpacc0x1) - vmagic_bias_less_output_zero_point;
141     int32_t vout1x0 = (int32_t) float_as_uint32(vfpacc1x0) - vmagic_bias_less_output_zero_point;
142     int32_t vout1x1 = (int32_t) float_as_uint32(vfpacc1x1) - vmagic_bias_less_output_zero_point;
143     int32_t vout2x0 = (int32_t) float_as_uint32(vfpacc2x0) - vmagic_bias_less_output_zero_point;
144     int32_t vout2x1 = (int32_t) float_as_uint32(vfpacc2x1) - vmagic_bias_less_output_zero_point;
145     int32_t vout3x0 = (int32_t) float_as_uint32(vfpacc3x0) - vmagic_bias_less_output_zero_point;
146     int32_t vout3x1 = (int32_t) float_as_uint32(vfpacc3x1) - vmagic_bias_less_output_zero_point;
147 
148     if XNN_LIKELY(nc >= 2) {
149       c0[0] = (int8_t) vout0x0;
150       c0[1] = (int8_t) vout0x1;
151       c1[0] = (int8_t) vout1x0;
152       c1[1] = (int8_t) vout1x1;
153       c2[0] = (int8_t) vout2x0;
154       c2[1] = (int8_t) vout2x1;
155       c3[0] = (int8_t) vout3x0;
156       c3[1] = (int8_t) vout3x1;
157 
158       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
159       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
160       a2 = (const int8_t*) ((uintptr_t) a2 - kc);
161       a3 = (const int8_t*) ((uintptr_t) a3 - kc);
162 
163       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
164       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
165       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
166       c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
167 
168       nc -= 2;
169     } else {
170       if (nc & 1) {
171         c0[0] = (int8_t) vout0x0;
172         c1[0] = (int8_t) vout1x0;
173         c2[0] = (int8_t) vout2x0;
174         c3[0] = (int8_t) vout3x0;
175       }
176 
177       nc = 0;
178     }
179   } while (nc != 0);
180 }
181