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 <fp16.h>
13
14 #include <xnnpack/math.h>
15 #include <xnnpack/gemm.h>
16
17
xnn_qu8_gemm_minmax_fp32_ukernel_2x4__wasm_fmagic(size_t mr,size_t nc,size_t kc,const uint8_t * restrict a,size_t a_stride,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qu8_gemm_minmax_fp32_ukernel_2x4__wasm_fmagic(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 const uint8_t* restrict a,
23 size_t a_stride,
24 const void* restrict w,
25 uint8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
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
35 const uint8_t* a0 = a;
36 uint8_t* c0 = c;
37 const uint8_t* a1 = (const uint8_t*) ((uintptr_t) a0 + a_stride);
38 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
39 if XNN_UNPREDICTABLE(mr != 2) {
40 a1 = a0;
41 c1 = c0;
42 }
43
44 const int32_t vb_zero_point = params->fp32_scalar_fmagic.kernel_zero_point;
45 do {
46 int32_t vacc0x0 = ((const int32_t*) w)[0];
47 int32_t vacc0x1 = ((const int32_t*) w)[1];
48 int32_t vacc0x2 = ((const int32_t*) w)[2];
49 int32_t vacc0x3 = ((const int32_t*) w)[3];
50 int32_t vacc1x0 = vacc0x0;
51 int32_t vacc1x1 = vacc0x1;
52 int32_t vacc1x2 = vacc0x2;
53 int32_t vacc1x3 = vacc0x3;
54 w = (const void*) ((const int32_t*) w + 4);
55
56 size_t k = kc;
57 do {
58 const int32_t va0 = (int32_t) (uint32_t) *a0++;
59 const int32_t va1 = (int32_t) (uint32_t) *a1++;
60
61 const int32_t vb0 = (int32_t) (uint32_t) ((const uint8_t*) w)[0] - vb_zero_point;
62 const int32_t vb1 = (int32_t) (uint32_t) ((const uint8_t*) w)[1] - vb_zero_point;
63 const int32_t vb2 = (int32_t) (uint32_t) ((const uint8_t*) w)[2] - vb_zero_point;
64 const int32_t vb3 = (int32_t) (uint32_t) ((const uint8_t*) w)[3] - vb_zero_point;
65 w = (const void*) ((const uint8_t*) w + 4);
66
67 vacc0x0 += va0 * vb0;
68 vacc0x1 += va0 * vb1;
69 vacc0x2 += va0 * vb2;
70 vacc0x3 += va0 * vb3;
71 vacc1x0 += va1 * vb0;
72 vacc1x1 += va1 * vb1;
73 vacc1x2 += va1 * vb2;
74 vacc1x3 += va1 * vb3;
75
76 k -= sizeof(uint8_t);
77 } while (k != 0);
78
79 float vfpacc0x0 = (float) vacc0x0;
80 float vfpacc0x1 = (float) vacc0x1;
81 float vfpacc0x2 = (float) vacc0x2;
82 float vfpacc0x3 = (float) vacc0x3;
83 float vfpacc1x0 = (float) vacc1x0;
84 float vfpacc1x1 = (float) vacc1x1;
85 float vfpacc1x2 = (float) vacc1x2;
86 float vfpacc1x3 = (float) vacc1x3;
87
88 const float vscale = params->fp32_scalar_fmagic.scale;
89 vfpacc0x0 *= vscale;
90 vfpacc0x1 *= vscale;
91 vfpacc0x2 *= vscale;
92 vfpacc0x3 *= vscale;
93 vfpacc1x0 *= vscale;
94 vfpacc1x1 *= vscale;
95 vfpacc1x2 *= vscale;
96 vfpacc1x3 *= vscale;
97
98 const float voutput_min_less_zero_point = params->fp32_scalar_fmagic.output_min_less_zero_point;
99 vfpacc0x0 = __builtin_wasm_max_f32(vfpacc0x0, voutput_min_less_zero_point);
100 vfpacc0x1 = __builtin_wasm_max_f32(vfpacc0x1, voutput_min_less_zero_point);
101 vfpacc0x2 = __builtin_wasm_max_f32(vfpacc0x2, voutput_min_less_zero_point);
102 vfpacc0x3 = __builtin_wasm_max_f32(vfpacc0x3, voutput_min_less_zero_point);
103 vfpacc1x0 = __builtin_wasm_max_f32(vfpacc1x0, voutput_min_less_zero_point);
104 vfpacc1x1 = __builtin_wasm_max_f32(vfpacc1x1, voutput_min_less_zero_point);
105 vfpacc1x2 = __builtin_wasm_max_f32(vfpacc1x2, voutput_min_less_zero_point);
106 vfpacc1x3 = __builtin_wasm_max_f32(vfpacc1x3, voutput_min_less_zero_point);
107
108 const float voutput_max_less_zero_point = params->fp32_scalar_fmagic.output_max_less_zero_point;
109 vfpacc0x0 = __builtin_wasm_min_f32(vfpacc0x0, voutput_max_less_zero_point);
110 vfpacc0x1 = __builtin_wasm_min_f32(vfpacc0x1, voutput_max_less_zero_point);
111 vfpacc0x2 = __builtin_wasm_min_f32(vfpacc0x2, voutput_max_less_zero_point);
112 vfpacc0x3 = __builtin_wasm_min_f32(vfpacc0x3, voutput_max_less_zero_point);
113 vfpacc1x0 = __builtin_wasm_min_f32(vfpacc1x0, voutput_max_less_zero_point);
114 vfpacc1x1 = __builtin_wasm_min_f32(vfpacc1x1, voutput_max_less_zero_point);
115 vfpacc1x2 = __builtin_wasm_min_f32(vfpacc1x2, voutput_max_less_zero_point);
116 vfpacc1x3 = __builtin_wasm_min_f32(vfpacc1x3, voutput_max_less_zero_point);
117
118 const float vmagic_bias = params->fp32_scalar_fmagic.magic_bias;
119 vfpacc0x0 += vmagic_bias;
120 vfpacc0x1 += vmagic_bias;
121 vfpacc0x2 += vmagic_bias;
122 vfpacc0x3 += vmagic_bias;
123 vfpacc1x0 += vmagic_bias;
124 vfpacc1x1 += vmagic_bias;
125 vfpacc1x2 += vmagic_bias;
126 vfpacc1x3 += vmagic_bias;
127
128 const int32_t vmagic_bias_less_output_zero_point = params->fp32_scalar_fmagic.magic_bias_less_output_zero_point;
129 int32_t vout0x0 = (int32_t) fp32_to_bits(vfpacc0x0) - vmagic_bias_less_output_zero_point;
130 int32_t vout0x1 = (int32_t) fp32_to_bits(vfpacc0x1) - vmagic_bias_less_output_zero_point;
131 int32_t vout0x2 = (int32_t) fp32_to_bits(vfpacc0x2) - vmagic_bias_less_output_zero_point;
132 int32_t vout0x3 = (int32_t) fp32_to_bits(vfpacc0x3) - vmagic_bias_less_output_zero_point;
133 int32_t vout1x0 = (int32_t) fp32_to_bits(vfpacc1x0) - vmagic_bias_less_output_zero_point;
134 int32_t vout1x1 = (int32_t) fp32_to_bits(vfpacc1x1) - vmagic_bias_less_output_zero_point;
135 int32_t vout1x2 = (int32_t) fp32_to_bits(vfpacc1x2) - vmagic_bias_less_output_zero_point;
136 int32_t vout1x3 = (int32_t) fp32_to_bits(vfpacc1x3) - vmagic_bias_less_output_zero_point;
137
138 if XNN_LIKELY(nc >= 4) {
139 c0[0] = (uint8_t) vout0x0;
140 c0[1] = (uint8_t) vout0x1;
141 c0[2] = (uint8_t) vout0x2;
142 c0[3] = (uint8_t) vout0x3;
143 c1[0] = (uint8_t) vout1x0;
144 c1[1] = (uint8_t) vout1x1;
145 c1[2] = (uint8_t) vout1x2;
146 c1[3] = (uint8_t) vout1x3;
147
148 a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
149 a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
150
151 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
152 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
153
154 nc -= 4;
155 } else {
156 if (nc & 2) {
157 c0[0] = (uint8_t) vout0x0;
158 c0[1] = (uint8_t) vout0x1;
159 vout0x0 = vout0x2;
160 c0 += 2;
161 c1[0] = (uint8_t) vout1x0;
162 c1[1] = (uint8_t) vout1x1;
163 vout1x0 = vout1x2;
164 c1 += 2;
165 }
166 if (nc & 1) {
167 c0[0] = (uint8_t) vout0x0;
168 c1[0] = (uint8_t) vout1x0;
169 }
170
171 nc = 0;
172 }
173 } while (nc != 0);
174 }
175