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_qu8_gemm_minmax_fp32_ukernel_4x2__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)])17 void xnn_qu8_gemm_minmax_fp32_ukernel_4x2__wasm_fmagic(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 const uint8_t* restrict a,
22 size_t a_stride,
23 const void* restrict w,
24 uint8_t* restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 const union xnn_qu8_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 uint8_t* a0 = a;
35 uint8_t* c0 = c;
36 const uint8_t* a1 = (const uint8_t*) ((uintptr_t) a0 + a_stride);
37 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
38 if XNN_UNPREDICTABLE(mr < 2) {
39 a1 = a0;
40 c1 = c0;
41 }
42 const uint8_t* a2 = (const uint8_t*) ((uintptr_t) a1 + a_stride);
43 uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
44 if XNN_UNPREDICTABLE(mr <= 2) {
45 a2 = a1;
46 c2 = c1;
47 }
48 const uint8_t* a3 = (const uint8_t*) ((uintptr_t) a2 + a_stride);
49 uint8_t* c3 = (uint8_t*) ((uintptr_t) c2 + cm_stride);
50 if XNN_UNPREDICTABLE(mr != 4) {
51 a3 = a2;
52 c3 = c2;
53 }
54
55 const int32_t vb_zero_point = params->fp32_scalar_fmagic.kernel_zero_point;
56 do {
57 int32_t vacc0x0 = unaligned_indexed_load_s32(w, 0);
58 int32_t vacc0x1 = unaligned_indexed_load_s32(w, 1);
59 int32_t vacc1x0 = vacc0x0;
60 int32_t vacc1x1 = vacc0x1;
61 int32_t vacc2x0 = vacc0x0;
62 int32_t vacc2x1 = vacc0x1;
63 int32_t vacc3x0 = vacc0x0;
64 int32_t vacc3x1 = vacc0x1;
65 w = (const void*) ((const int32_t*) w + 2);
66
67 size_t k = kc;
68 do {
69 const int32_t va0 = (int32_t) (uint32_t) *a0++;
70 const int32_t va1 = (int32_t) (uint32_t) *a1++;
71 const int32_t va2 = (int32_t) (uint32_t) *a2++;
72 const int32_t va3 = (int32_t) (uint32_t) *a3++;
73
74 const int32_t vb0 = (int32_t) (uint32_t) ((const uint8_t*) w)[0] - vb_zero_point;
75 const int32_t vb1 = (int32_t) (uint32_t) ((const uint8_t*) w)[1] - vb_zero_point;
76 w = (const void*) ((const uint8_t*) w + 2);
77
78 vacc0x0 += va0 * vb0;
79 vacc0x1 += va0 * vb1;
80 vacc1x0 += va1 * vb0;
81 vacc1x1 += va1 * vb1;
82 vacc2x0 += va2 * vb0;
83 vacc2x1 += va2 * vb1;
84 vacc3x0 += va3 * vb0;
85 vacc3x1 += va3 * vb1;
86
87 k -= sizeof(uint8_t);
88 } while (k != 0);
89
90 float vfpacc0x0 = (float) vacc0x0;
91 float vfpacc0x1 = (float) vacc0x1;
92 float vfpacc1x0 = (float) vacc1x0;
93 float vfpacc1x1 = (float) vacc1x1;
94 float vfpacc2x0 = (float) vacc2x0;
95 float vfpacc2x1 = (float) vacc2x1;
96 float vfpacc3x0 = (float) vacc3x0;
97 float vfpacc3x1 = (float) vacc3x1;
98
99 const float vscale = params->fp32_scalar_fmagic.scale;
100 vfpacc0x0 *= vscale;
101 vfpacc0x1 *= vscale;
102 vfpacc1x0 *= vscale;
103 vfpacc1x1 *= vscale;
104 vfpacc2x0 *= vscale;
105 vfpacc2x1 *= vscale;
106 vfpacc3x0 *= vscale;
107 vfpacc3x1 *= vscale;
108
109 const float voutput_min_less_zero_point = params->fp32_scalar_fmagic.output_min_less_zero_point;
110 vfpacc0x0 = __builtin_wasm_max_f32(vfpacc0x0, voutput_min_less_zero_point);
111 vfpacc0x1 = __builtin_wasm_max_f32(vfpacc0x1, voutput_min_less_zero_point);
112 vfpacc1x0 = __builtin_wasm_max_f32(vfpacc1x0, voutput_min_less_zero_point);
113 vfpacc1x1 = __builtin_wasm_max_f32(vfpacc1x1, voutput_min_less_zero_point);
114 vfpacc2x0 = __builtin_wasm_max_f32(vfpacc2x0, voutput_min_less_zero_point);
115 vfpacc2x1 = __builtin_wasm_max_f32(vfpacc2x1, voutput_min_less_zero_point);
116 vfpacc3x0 = __builtin_wasm_max_f32(vfpacc3x0, voutput_min_less_zero_point);
117 vfpacc3x1 = __builtin_wasm_max_f32(vfpacc3x1, voutput_min_less_zero_point);
118
119 const float voutput_max_less_zero_point = params->fp32_scalar_fmagic.output_max_less_zero_point;
120 vfpacc0x0 = __builtin_wasm_min_f32(vfpacc0x0, voutput_max_less_zero_point);
121 vfpacc0x1 = __builtin_wasm_min_f32(vfpacc0x1, voutput_max_less_zero_point);
122 vfpacc1x0 = __builtin_wasm_min_f32(vfpacc1x0, voutput_max_less_zero_point);
123 vfpacc1x1 = __builtin_wasm_min_f32(vfpacc1x1, voutput_max_less_zero_point);
124 vfpacc2x0 = __builtin_wasm_min_f32(vfpacc2x0, voutput_max_less_zero_point);
125 vfpacc2x1 = __builtin_wasm_min_f32(vfpacc2x1, voutput_max_less_zero_point);
126 vfpacc3x0 = __builtin_wasm_min_f32(vfpacc3x0, voutput_max_less_zero_point);
127 vfpacc3x1 = __builtin_wasm_min_f32(vfpacc3x1, voutput_max_less_zero_point);
128
129 const float vmagic_bias = params->fp32_scalar_fmagic.magic_bias;
130 vfpacc0x0 += vmagic_bias;
131 vfpacc0x1 += vmagic_bias;
132 vfpacc1x0 += vmagic_bias;
133 vfpacc1x1 += vmagic_bias;
134 vfpacc2x0 += vmagic_bias;
135 vfpacc2x1 += vmagic_bias;
136 vfpacc3x0 += vmagic_bias;
137 vfpacc3x1 += vmagic_bias;
138
139 const int32_t vmagic_bias_less_output_zero_point = params->fp32_scalar_fmagic.magic_bias_less_output_zero_point;
140 int32_t vout0x0 = (int32_t) float_as_uint32(vfpacc0x0) - vmagic_bias_less_output_zero_point;
141 int32_t vout0x1 = (int32_t) float_as_uint32(vfpacc0x1) - vmagic_bias_less_output_zero_point;
142 int32_t vout1x0 = (int32_t) float_as_uint32(vfpacc1x0) - vmagic_bias_less_output_zero_point;
143 int32_t vout1x1 = (int32_t) float_as_uint32(vfpacc1x1) - vmagic_bias_less_output_zero_point;
144 int32_t vout2x0 = (int32_t) float_as_uint32(vfpacc2x0) - vmagic_bias_less_output_zero_point;
145 int32_t vout2x1 = (int32_t) float_as_uint32(vfpacc2x1) - vmagic_bias_less_output_zero_point;
146 int32_t vout3x0 = (int32_t) float_as_uint32(vfpacc3x0) - vmagic_bias_less_output_zero_point;
147 int32_t vout3x1 = (int32_t) float_as_uint32(vfpacc3x1) - vmagic_bias_less_output_zero_point;
148
149 if XNN_LIKELY(nc >= 2) {
150 c0[0] = (uint8_t) vout0x0;
151 c0[1] = (uint8_t) vout0x1;
152 c1[0] = (uint8_t) vout1x0;
153 c1[1] = (uint8_t) vout1x1;
154 c2[0] = (uint8_t) vout2x0;
155 c2[1] = (uint8_t) vout2x1;
156 c3[0] = (uint8_t) vout3x0;
157 c3[1] = (uint8_t) vout3x1;
158
159 a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
160 a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
161 a2 = (const uint8_t*) ((uintptr_t) a2 - kc);
162 a3 = (const uint8_t*) ((uintptr_t) a3 - kc);
163
164 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
165 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
166 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
167 c3 = (uint8_t*) ((uintptr_t) c3 + cn_stride);
168
169 nc -= 2;
170 } else {
171 if (nc & 1) {
172 c0[0] = (uint8_t) vout0x0;
173 c1[0] = (uint8_t) vout1x0;
174 c2[0] = (uint8_t) vout2x0;
175 c3[0] = (uint8_t) vout3x0;
176 }
177
178 nc = 0;
179 }
180 } while (nc != 0);
181 }
182