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_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)])18 void xnn_qu8_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 uint8_t**restrict a,
24 const void*restrict w,
25 uint8_t*restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const uint8_t* zero,
30 const union xnn_qu8_conv_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 uint8_t* c0 = c;
43 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
44 if XNN_UNPREDICTABLE(mr != 2) {
45 c1 = c0;
46 }
47
48 const int32_t vb_zero_point = params->fp32_scalar_fmagic.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 w = (const void*) ((const int32_t*) w + 4);
59
60 size_t p = ks;
61 do {
62 const uint8_t* restrict a0 = a[0];
63 assert(a0 != NULL);
64 if XNN_UNPREDICTABLE(a0 != zero) {
65 a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
66 }
67 const uint8_t* restrict a1 = a[1];
68 assert(a1 != NULL);
69 if XNN_UNPREDICTABLE(a1 != zero) {
70 a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
71 }
72 a += 2;
73
74 size_t k = kc;
75 do {
76 const int32_t va0 = (int32_t) (uint32_t) *a0++;
77 const int32_t va1 = (int32_t) (uint32_t) *a1++;
78
79 const int32_t vb0 = (int32_t) (uint32_t) ((const uint8_t*) w)[0] - vb_zero_point;
80 const int32_t vb1 = (int32_t) (uint32_t) ((const uint8_t*) w)[1] - vb_zero_point;
81 const int32_t vb2 = (int32_t) (uint32_t) ((const uint8_t*) w)[2] - vb_zero_point;
82 const int32_t vb3 = (int32_t) (uint32_t) ((const uint8_t*) w)[3] - vb_zero_point;
83 w = (const void*) ((const uint8_t*) w + 4);
84
85 vacc0x0 += va0 * vb0;
86 vacc0x1 += va0 * vb1;
87 vacc0x2 += va0 * vb2;
88 vacc0x3 += va0 * vb3;
89 vacc1x0 += va1 * vb0;
90 vacc1x1 += va1 * vb1;
91 vacc1x2 += va1 * vb2;
92 vacc1x3 += va1 * vb3;
93
94 k -= sizeof(uint8_t);
95 } while (k != 0);
96 p -= 2 * sizeof(void*);
97 } while (p != 0);
98
99 float vfpacc0x0 = (float) vacc0x0;
100 float vfpacc0x1 = (float) vacc0x1;
101 float vfpacc0x2 = (float) vacc0x2;
102 float vfpacc0x3 = (float) vacc0x3;
103 float vfpacc1x0 = (float) vacc1x0;
104 float vfpacc1x1 = (float) vacc1x1;
105 float vfpacc1x2 = (float) vacc1x2;
106 float vfpacc1x3 = (float) vacc1x3;
107
108 const float vscale = params->fp32_scalar_fmagic.scale;
109 vfpacc0x0 *= vscale;
110 vfpacc0x1 *= vscale;
111 vfpacc0x2 *= vscale;
112 vfpacc0x3 *= vscale;
113 vfpacc1x0 *= vscale;
114 vfpacc1x1 *= vscale;
115 vfpacc1x2 *= vscale;
116 vfpacc1x3 *= vscale;
117
118 const float voutput_min_less_zero_point = params->fp32_scalar_fmagic.output_min_less_zero_point;
119 vfpacc0x0 = __builtin_wasm_max_f32(vfpacc0x0, voutput_min_less_zero_point);
120 vfpacc0x1 = __builtin_wasm_max_f32(vfpacc0x1, voutput_min_less_zero_point);
121 vfpacc0x2 = __builtin_wasm_max_f32(vfpacc0x2, voutput_min_less_zero_point);
122 vfpacc0x3 = __builtin_wasm_max_f32(vfpacc0x3, voutput_min_less_zero_point);
123 vfpacc1x0 = __builtin_wasm_max_f32(vfpacc1x0, voutput_min_less_zero_point);
124 vfpacc1x1 = __builtin_wasm_max_f32(vfpacc1x1, voutput_min_less_zero_point);
125 vfpacc1x2 = __builtin_wasm_max_f32(vfpacc1x2, voutput_min_less_zero_point);
126 vfpacc1x3 = __builtin_wasm_max_f32(vfpacc1x3, voutput_min_less_zero_point);
127
128 const float voutput_max_less_zero_point = params->fp32_scalar_fmagic.output_max_less_zero_point;
129 vfpacc0x0 = __builtin_wasm_min_f32(vfpacc0x0, voutput_max_less_zero_point);
130 vfpacc0x1 = __builtin_wasm_min_f32(vfpacc0x1, voutput_max_less_zero_point);
131 vfpacc0x2 = __builtin_wasm_min_f32(vfpacc0x2, voutput_max_less_zero_point);
132 vfpacc0x3 = __builtin_wasm_min_f32(vfpacc0x3, voutput_max_less_zero_point);
133 vfpacc1x0 = __builtin_wasm_min_f32(vfpacc1x0, voutput_max_less_zero_point);
134 vfpacc1x1 = __builtin_wasm_min_f32(vfpacc1x1, voutput_max_less_zero_point);
135 vfpacc1x2 = __builtin_wasm_min_f32(vfpacc1x2, voutput_max_less_zero_point);
136 vfpacc1x3 = __builtin_wasm_min_f32(vfpacc1x3, voutput_max_less_zero_point);
137
138 const float vmagic_bias = params->fp32_scalar_fmagic.magic_bias;
139 vfpacc0x0 += vmagic_bias;
140 vfpacc0x1 += vmagic_bias;
141 vfpacc0x2 += vmagic_bias;
142 vfpacc0x3 += vmagic_bias;
143 vfpacc1x0 += vmagic_bias;
144 vfpacc1x1 += vmagic_bias;
145 vfpacc1x2 += vmagic_bias;
146 vfpacc1x3 += vmagic_bias;
147
148 const int32_t vmagic_bias_less_output_zero_point = params->fp32_scalar_fmagic.magic_bias_less_output_zero_point;
149 int32_t vout0x0 = (int32_t) fp32_to_bits(vfpacc0x0) - vmagic_bias_less_output_zero_point;
150 int32_t vout0x1 = (int32_t) fp32_to_bits(vfpacc0x1) - vmagic_bias_less_output_zero_point;
151 int32_t vout0x2 = (int32_t) fp32_to_bits(vfpacc0x2) - vmagic_bias_less_output_zero_point;
152 int32_t vout0x3 = (int32_t) fp32_to_bits(vfpacc0x3) - vmagic_bias_less_output_zero_point;
153 int32_t vout1x0 = (int32_t) fp32_to_bits(vfpacc1x0) - vmagic_bias_less_output_zero_point;
154 int32_t vout1x1 = (int32_t) fp32_to_bits(vfpacc1x1) - vmagic_bias_less_output_zero_point;
155 int32_t vout1x2 = (int32_t) fp32_to_bits(vfpacc1x2) - vmagic_bias_less_output_zero_point;
156 int32_t vout1x3 = (int32_t) fp32_to_bits(vfpacc1x3) - vmagic_bias_less_output_zero_point;
157
158 if XNN_LIKELY(nc >= 4) {
159 c1[0] = (uint8_t) vout1x0;
160 c1[1] = (uint8_t) vout1x1;
161 c1[2] = (uint8_t) vout1x2;
162 c1[3] = (uint8_t) vout1x3;
163 c0[0] = (uint8_t) vout0x0;
164 c0[1] = (uint8_t) vout0x1;
165 c0[2] = (uint8_t) vout0x2;
166 c0[3] = (uint8_t) vout0x3;
167
168 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
169 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
170
171 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
172 nc -= 4;
173 } else {
174 if (nc & 2) {
175 c1[0] = (uint8_t) vout1x0;
176 c1[1] = (uint8_t) vout1x1;
177 vout1x0 = vout1x2;
178 c1 += 2;
179 c0[0] = (uint8_t) vout0x0;
180 c0[1] = (uint8_t) vout0x1;
181 vout0x0 = vout0x2;
182 c0 += 2;
183 }
184 if (nc & 1) {
185 c1[0] = (uint8_t) vout1x0;
186 c0[0] = (uint8_t) vout0x0;
187 }
188
189 nc = 0;
190 }
191 } while (nc != 0);
192 }
193