1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/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 <fp16.h>
13
14 #include <xnnpack/math.h>
15 #include <xnnpack/gemm.h>
16
17
xnn_qc8_igemm_minmax_fp32_ukernel_2x4__wasm_fmagic(size_t mr,size_t nc,size_t kc,size_t ks,const int8_t ** restrict a,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const int8_t * zero,const union xnn_qs8_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qc8_igemm_minmax_fp32_ukernel_2x4__wasm_fmagic(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const int8_t**restrict a,
24 const void*restrict w,
25 int8_t*restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const int8_t* zero,
30 const union xnn_qs8_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
31 {
32 assert(mr != 0);
33 assert(mr <= 2);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (2 * sizeof(void*)) == 0);
38 assert(a != NULL);
39 assert(w != NULL);
40 assert(c != NULL);
41
42 int8_t* c0 = c;
43 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
44 if XNN_UNPREDICTABLE(mr != 2) {
45 c1 = c0;
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 w = (const void*) ((const int32_t*) w + 4);
58
59 size_t p = ks;
60 do {
61 const int8_t* restrict a0 = a[0];
62 assert(a0 != NULL);
63 if XNN_UNPREDICTABLE(a0 != zero) {
64 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
65 }
66 const int8_t* restrict a1 = a[1];
67 assert(a1 != NULL);
68 if XNN_UNPREDICTABLE(a1 != zero) {
69 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
70 }
71 a += 2;
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
78 const int32_t vb0 = (int32_t) ((const int8_t*) w)[0];
79 const int32_t vb1 = (int32_t) ((const int8_t*) w)[1];
80 const int32_t vb2 = (int32_t) ((const int8_t*) w)[2];
81 const int32_t vb3 = (int32_t) ((const int8_t*) w)[3];
82 w = (const void*) ((const int8_t*) w + 4);
83
84 vacc0x0 += va0 * vb0;
85 vacc0x1 += va0 * vb1;
86 vacc0x2 += va0 * vb2;
87 vacc0x3 += va0 * vb3;
88 vacc1x0 += va1 * vb0;
89 vacc1x1 += va1 * vb1;
90 vacc1x2 += va1 * vb2;
91 vacc1x3 += va1 * vb3;
92
93 k -= sizeof(int8_t);
94 } while (k != 0);
95 p -= 2 * sizeof(void*);
96 } while (p != 0);
97
98 float vfpacc0x0 = (float) vacc0x0;
99 float vfpacc0x1 = (float) vacc0x1;
100 float vfpacc0x2 = (float) vacc0x2;
101 float vfpacc0x3 = (float) vacc0x3;
102 float vfpacc1x0 = (float) vacc1x0;
103 float vfpacc1x1 = (float) vacc1x1;
104 float vfpacc1x2 = (float) vacc1x2;
105 float vfpacc1x3 = (float) vacc1x3;
106
107 const float vscale0 = ((const float*) w)[0];
108 vfpacc0x0 *= vscale0;
109 vfpacc1x0 *= vscale0;
110 const float vscale1 = ((const float*) w)[1];
111 vfpacc0x1 *= vscale1;
112 vfpacc1x1 *= vscale1;
113 const float vscale2 = ((const float*) w)[2];
114 vfpacc0x2 *= vscale2;
115 vfpacc1x2 *= vscale2;
116 const float vscale3 = ((const float*) w)[3];
117 vfpacc0x3 *= vscale3;
118 vfpacc1x3 *= vscale3;
119 w = (const void*) ((const float*) w + 4);
120
121 const float voutput_min_less_zero_point = params->scalar_fmagic.output_min_less_zero_point;
122 vfpacc0x0 = __builtin_wasm_max_f32(vfpacc0x0, voutput_min_less_zero_point);
123 vfpacc0x1 = __builtin_wasm_max_f32(vfpacc0x1, voutput_min_less_zero_point);
124 vfpacc0x2 = __builtin_wasm_max_f32(vfpacc0x2, voutput_min_less_zero_point);
125 vfpacc0x3 = __builtin_wasm_max_f32(vfpacc0x3, voutput_min_less_zero_point);
126 vfpacc1x0 = __builtin_wasm_max_f32(vfpacc1x0, voutput_min_less_zero_point);
127 vfpacc1x1 = __builtin_wasm_max_f32(vfpacc1x1, voutput_min_less_zero_point);
128 vfpacc1x2 = __builtin_wasm_max_f32(vfpacc1x2, voutput_min_less_zero_point);
129 vfpacc1x3 = __builtin_wasm_max_f32(vfpacc1x3, voutput_min_less_zero_point);
130
131 const float voutput_max_less_zero_point = params->scalar_fmagic.output_max_less_zero_point;
132 vfpacc0x0 = __builtin_wasm_min_f32(vfpacc0x0, voutput_max_less_zero_point);
133 vfpacc0x1 = __builtin_wasm_min_f32(vfpacc0x1, voutput_max_less_zero_point);
134 vfpacc0x2 = __builtin_wasm_min_f32(vfpacc0x2, voutput_max_less_zero_point);
135 vfpacc0x3 = __builtin_wasm_min_f32(vfpacc0x3, voutput_max_less_zero_point);
136 vfpacc1x0 = __builtin_wasm_min_f32(vfpacc1x0, voutput_max_less_zero_point);
137 vfpacc1x1 = __builtin_wasm_min_f32(vfpacc1x1, voutput_max_less_zero_point);
138 vfpacc1x2 = __builtin_wasm_min_f32(vfpacc1x2, voutput_max_less_zero_point);
139 vfpacc1x3 = __builtin_wasm_min_f32(vfpacc1x3, voutput_max_less_zero_point);
140
141 const float vmagic_bias = params->scalar_fmagic.magic_bias;
142 vfpacc0x0 += vmagic_bias;
143 vfpacc0x1 += vmagic_bias;
144 vfpacc0x2 += vmagic_bias;
145 vfpacc0x3 += vmagic_bias;
146 vfpacc1x0 += vmagic_bias;
147 vfpacc1x1 += vmagic_bias;
148 vfpacc1x2 += vmagic_bias;
149 vfpacc1x3 += vmagic_bias;
150
151 const int32_t vmagic_bias_less_output_zero_point = params->scalar_fmagic.magic_bias_less_output_zero_point;
152 int32_t vout0x0 = (int32_t) fp32_to_bits(vfpacc0x0) - vmagic_bias_less_output_zero_point;
153 int32_t vout0x1 = (int32_t) fp32_to_bits(vfpacc0x1) - vmagic_bias_less_output_zero_point;
154 int32_t vout0x2 = (int32_t) fp32_to_bits(vfpacc0x2) - vmagic_bias_less_output_zero_point;
155 int32_t vout0x3 = (int32_t) fp32_to_bits(vfpacc0x3) - vmagic_bias_less_output_zero_point;
156 int32_t vout1x0 = (int32_t) fp32_to_bits(vfpacc1x0) - vmagic_bias_less_output_zero_point;
157 int32_t vout1x1 = (int32_t) fp32_to_bits(vfpacc1x1) - vmagic_bias_less_output_zero_point;
158 int32_t vout1x2 = (int32_t) fp32_to_bits(vfpacc1x2) - vmagic_bias_less_output_zero_point;
159 int32_t vout1x3 = (int32_t) fp32_to_bits(vfpacc1x3) - vmagic_bias_less_output_zero_point;
160
161 if XNN_LIKELY(nc >= 4) {
162 c1[0] = (int8_t) vout1x0;
163 c1[1] = (int8_t) vout1x1;
164 c1[2] = (int8_t) vout1x2;
165 c1[3] = (int8_t) vout1x3;
166 c0[0] = (int8_t) vout0x0;
167 c0[1] = (int8_t) vout0x1;
168 c0[2] = (int8_t) vout0x2;
169 c0[3] = (int8_t) vout0x3;
170
171 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
172 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
173
174 a = (const int8_t**restrict) ((uintptr_t) a - ks);
175 nc -= 4;
176 } else {
177 if (nc & 2) {
178 c1[0] = (int8_t) vout1x0;
179 c1[1] = (int8_t) vout1x1;
180 vout1x0 = vout1x2;
181 c1 += 2;
182 c0[0] = (int8_t) vout0x0;
183 c0[1] = (int8_t) vout0x1;
184 vout0x0 = vout0x2;
185 c0 += 2;
186 }
187 if (nc & 1) {
188 c1[0] = (int8_t) vout1x0;
189 c0[0] = (int8_t) vout0x0;
190 }
191
192 nc = 0;
193 }
194 } while (nc != 0);
195 }
196