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 <xnnpack/math.h>
13 #include <xnnpack/gemm.h>
14
15
xnn_qu8_igemm_minmax_fp32_ukernel_2x4__wasm_fmagic(size_t mr,size_t nc,size_t kc,size_t ks,const uint8_t ** restrict a,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const uint8_t * zero,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])16 void xnn_qu8_igemm_minmax_fp32_ukernel_2x4__wasm_fmagic(
17 size_t mr,
18 size_t nc,
19 size_t kc,
20 size_t ks,
21 const uint8_t**restrict a,
22 const void*restrict w,
23 uint8_t*restrict c,
24 size_t cm_stride,
25 size_t cn_stride,
26 size_t a_offset,
27 const uint8_t* zero,
28 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
29 {
30 assert(mr != 0);
31 assert(mr <= 2);
32 assert(nc != 0);
33 assert(kc != 0);
34 assert(ks != 0);
35 assert(ks % (2 * sizeof(void*)) == 0);
36 assert(a != NULL);
37 assert(w != NULL);
38 assert(c != NULL);
39
40 uint8_t* c0 = c;
41 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
42 if XNN_UNPREDICTABLE(mr != 2) {
43 c1 = c0;
44 }
45
46 const int32_t vb_zero_point = params->fp32_scalar_fmagic.kernel_zero_point;
47 do {
48 int32_t vacc0x0 = ((const int32_t*) w)[0];
49 int32_t vacc0x1 = ((const int32_t*) w)[1];
50 int32_t vacc0x2 = ((const int32_t*) w)[2];
51 int32_t vacc0x3 = ((const int32_t*) w)[3];
52 int32_t vacc1x0 = vacc0x0;
53 int32_t vacc1x1 = vacc0x1;
54 int32_t vacc1x2 = vacc0x2;
55 int32_t vacc1x3 = vacc0x3;
56 w = (const void*) ((const int32_t*) w + 4);
57
58 size_t p = ks;
59 do {
60 const uint8_t* restrict a0 = a[0];
61 assert(a0 != NULL);
62 if XNN_UNPREDICTABLE(a0 != zero) {
63 a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
64 }
65 const uint8_t* restrict a1 = a[1];
66 assert(a1 != NULL);
67 if XNN_UNPREDICTABLE(a1 != zero) {
68 a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
69 }
70 a += 2;
71
72 size_t k = kc;
73 do {
74 const int32_t va0 = (int32_t) (uint32_t) *a0++;
75 const int32_t va1 = (int32_t) (uint32_t) *a1++;
76
77 const int32_t vb0 = (int32_t) (uint32_t) ((const uint8_t*) w)[0] - vb_zero_point;
78 const int32_t vb1 = (int32_t) (uint32_t) ((const uint8_t*) w)[1] - vb_zero_point;
79 const int32_t vb2 = (int32_t) (uint32_t) ((const uint8_t*) w)[2] - vb_zero_point;
80 const int32_t vb3 = (int32_t) (uint32_t) ((const uint8_t*) w)[3] - vb_zero_point;
81 w = (const void*) ((const uint8_t*) w + 4);
82
83 vacc0x0 += va0 * vb0;
84 vacc0x1 += va0 * vb1;
85 vacc0x2 += va0 * vb2;
86 vacc0x3 += va0 * vb3;
87 vacc1x0 += va1 * vb0;
88 vacc1x1 += va1 * vb1;
89 vacc1x2 += va1 * vb2;
90 vacc1x3 += va1 * vb3;
91
92 k -= sizeof(uint8_t);
93 } while (k != 0);
94 p -= 2 * sizeof(void*);
95 } while (p != 0);
96
97 float vfpacc0x0 = (float) vacc0x0;
98 float vfpacc0x1 = (float) vacc0x1;
99 float vfpacc0x2 = (float) vacc0x2;
100 float vfpacc0x3 = (float) vacc0x3;
101 float vfpacc1x0 = (float) vacc1x0;
102 float vfpacc1x1 = (float) vacc1x1;
103 float vfpacc1x2 = (float) vacc1x2;
104 float vfpacc1x3 = (float) vacc1x3;
105
106 const float vscale = params->fp32_scalar_fmagic.scale;
107 vfpacc0x0 *= vscale;
108 vfpacc0x1 *= vscale;
109 vfpacc0x2 *= vscale;
110 vfpacc0x3 *= vscale;
111 vfpacc1x0 *= vscale;
112 vfpacc1x1 *= vscale;
113 vfpacc1x2 *= vscale;
114 vfpacc1x3 *= vscale;
115
116 const float voutput_min_less_zero_point = params->fp32_scalar_fmagic.output_min_less_zero_point;
117 vfpacc0x0 = __builtin_wasm_max_f32(vfpacc0x0, voutput_min_less_zero_point);
118 vfpacc0x1 = __builtin_wasm_max_f32(vfpacc0x1, voutput_min_less_zero_point);
119 vfpacc0x2 = __builtin_wasm_max_f32(vfpacc0x2, voutput_min_less_zero_point);
120 vfpacc0x3 = __builtin_wasm_max_f32(vfpacc0x3, voutput_min_less_zero_point);
121 vfpacc1x0 = __builtin_wasm_max_f32(vfpacc1x0, voutput_min_less_zero_point);
122 vfpacc1x1 = __builtin_wasm_max_f32(vfpacc1x1, voutput_min_less_zero_point);
123 vfpacc1x2 = __builtin_wasm_max_f32(vfpacc1x2, voutput_min_less_zero_point);
124 vfpacc1x3 = __builtin_wasm_max_f32(vfpacc1x3, voutput_min_less_zero_point);
125
126 const float voutput_max_less_zero_point = params->fp32_scalar_fmagic.output_max_less_zero_point;
127 vfpacc0x0 = __builtin_wasm_min_f32(vfpacc0x0, voutput_max_less_zero_point);
128 vfpacc0x1 = __builtin_wasm_min_f32(vfpacc0x1, voutput_max_less_zero_point);
129 vfpacc0x2 = __builtin_wasm_min_f32(vfpacc0x2, voutput_max_less_zero_point);
130 vfpacc0x3 = __builtin_wasm_min_f32(vfpacc0x3, voutput_max_less_zero_point);
131 vfpacc1x0 = __builtin_wasm_min_f32(vfpacc1x0, voutput_max_less_zero_point);
132 vfpacc1x1 = __builtin_wasm_min_f32(vfpacc1x1, voutput_max_less_zero_point);
133 vfpacc1x2 = __builtin_wasm_min_f32(vfpacc1x2, voutput_max_less_zero_point);
134 vfpacc1x3 = __builtin_wasm_min_f32(vfpacc1x3, voutput_max_less_zero_point);
135
136 const float vmagic_bias = params->fp32_scalar_fmagic.magic_bias;
137 vfpacc0x0 += vmagic_bias;
138 vfpacc0x1 += vmagic_bias;
139 vfpacc0x2 += vmagic_bias;
140 vfpacc0x3 += vmagic_bias;
141 vfpacc1x0 += vmagic_bias;
142 vfpacc1x1 += vmagic_bias;
143 vfpacc1x2 += vmagic_bias;
144 vfpacc1x3 += vmagic_bias;
145
146 const int32_t vmagic_bias_less_output_zero_point = params->fp32_scalar_fmagic.magic_bias_less_output_zero_point;
147 int32_t vout0x0 = (int32_t) float_as_uint32(vfpacc0x0) - vmagic_bias_less_output_zero_point;
148 int32_t vout0x1 = (int32_t) float_as_uint32(vfpacc0x1) - vmagic_bias_less_output_zero_point;
149 int32_t vout0x2 = (int32_t) float_as_uint32(vfpacc0x2) - vmagic_bias_less_output_zero_point;
150 int32_t vout0x3 = (int32_t) float_as_uint32(vfpacc0x3) - vmagic_bias_less_output_zero_point;
151 int32_t vout1x0 = (int32_t) float_as_uint32(vfpacc1x0) - vmagic_bias_less_output_zero_point;
152 int32_t vout1x1 = (int32_t) float_as_uint32(vfpacc1x1) - vmagic_bias_less_output_zero_point;
153 int32_t vout1x2 = (int32_t) float_as_uint32(vfpacc1x2) - vmagic_bias_less_output_zero_point;
154 int32_t vout1x3 = (int32_t) float_as_uint32(vfpacc1x3) - vmagic_bias_less_output_zero_point;
155
156 if XNN_LIKELY(nc >= 4) {
157 c1[0] = (uint8_t) vout1x0;
158 c1[1] = (uint8_t) vout1x1;
159 c1[2] = (uint8_t) vout1x2;
160 c1[3] = (uint8_t) vout1x3;
161 c0[0] = (uint8_t) vout0x0;
162 c0[1] = (uint8_t) vout0x1;
163 c0[2] = (uint8_t) vout0x2;
164 c0[3] = (uint8_t) vout0x3;
165
166 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
167 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
168
169 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
170 nc -= 4;
171 } else {
172 if (nc & 2) {
173 c1[0] = (uint8_t) vout1x0;
174 c1[1] = (uint8_t) vout1x1;
175 vout1x0 = vout1x2;
176 c1 += 2;
177 c0[0] = (uint8_t) vout0x0;
178 c0[1] = (uint8_t) vout0x1;
179 vout0x0 = vout0x2;
180 c0 += 2;
181 }
182 if (nc & 1) {
183 c1[0] = (uint8_t) vout1x0;
184 c0[0] = (uint8_t) vout0x0;
185 }
186
187 nc = 0;
188 }
189 } while (nc != 0);
190 }
191