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_1x4__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_1x4__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 <= 1);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (1 * sizeof(void*)) == 0);
38 assert(a != NULL);
39 assert(w != NULL);
40 assert(c != NULL);
41
42 uint8_t* c0 = c;
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 w = (const void*) ((const int32_t*) w + 4);
51
52 size_t p = ks;
53 do {
54 const uint8_t* restrict a0 = a[0];
55 assert(a0 != NULL);
56 if XNN_UNPREDICTABLE(a0 != zero) {
57 a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
58 }
59 a += 1;
60
61 size_t k = kc;
62 do {
63 const int32_t va0 = (int32_t) (uint32_t) *a0++;
64
65 const int32_t vb0 = (int32_t) (uint32_t) ((const uint8_t*) w)[0] - vb_zero_point;
66 const int32_t vb1 = (int32_t) (uint32_t) ((const uint8_t*) w)[1] - vb_zero_point;
67 const int32_t vb2 = (int32_t) (uint32_t) ((const uint8_t*) w)[2] - vb_zero_point;
68 const int32_t vb3 = (int32_t) (uint32_t) ((const uint8_t*) w)[3] - vb_zero_point;
69 w = (const void*) ((const uint8_t*) w + 4);
70
71 vacc0x0 += va0 * vb0;
72 vacc0x1 += va0 * vb1;
73 vacc0x2 += va0 * vb2;
74 vacc0x3 += va0 * vb3;
75
76 k -= sizeof(uint8_t);
77 } while (k != 0);
78 p -= 1 * sizeof(void*);
79 } while (p != 0);
80
81 float vfpacc0x0 = (float) vacc0x0;
82 float vfpacc0x1 = (float) vacc0x1;
83 float vfpacc0x2 = (float) vacc0x2;
84 float vfpacc0x3 = (float) vacc0x3;
85
86 const float vscale = params->fp32_scalar_fmagic.scale;
87 vfpacc0x0 *= vscale;
88 vfpacc0x1 *= vscale;
89 vfpacc0x2 *= vscale;
90 vfpacc0x3 *= vscale;
91
92 const float voutput_min_less_zero_point = params->fp32_scalar_fmagic.output_min_less_zero_point;
93 vfpacc0x0 = __builtin_wasm_max_f32(vfpacc0x0, voutput_min_less_zero_point);
94 vfpacc0x1 = __builtin_wasm_max_f32(vfpacc0x1, voutput_min_less_zero_point);
95 vfpacc0x2 = __builtin_wasm_max_f32(vfpacc0x2, voutput_min_less_zero_point);
96 vfpacc0x3 = __builtin_wasm_max_f32(vfpacc0x3, voutput_min_less_zero_point);
97
98 const float voutput_max_less_zero_point = params->fp32_scalar_fmagic.output_max_less_zero_point;
99 vfpacc0x0 = __builtin_wasm_min_f32(vfpacc0x0, voutput_max_less_zero_point);
100 vfpacc0x1 = __builtin_wasm_min_f32(vfpacc0x1, voutput_max_less_zero_point);
101 vfpacc0x2 = __builtin_wasm_min_f32(vfpacc0x2, voutput_max_less_zero_point);
102 vfpacc0x3 = __builtin_wasm_min_f32(vfpacc0x3, voutput_max_less_zero_point);
103
104 const float vmagic_bias = params->fp32_scalar_fmagic.magic_bias;
105 vfpacc0x0 += vmagic_bias;
106 vfpacc0x1 += vmagic_bias;
107 vfpacc0x2 += vmagic_bias;
108 vfpacc0x3 += vmagic_bias;
109
110 const int32_t vmagic_bias_less_output_zero_point = params->fp32_scalar_fmagic.magic_bias_less_output_zero_point;
111 int32_t vout0x0 = (int32_t) fp32_to_bits(vfpacc0x0) - vmagic_bias_less_output_zero_point;
112 int32_t vout0x1 = (int32_t) fp32_to_bits(vfpacc0x1) - vmagic_bias_less_output_zero_point;
113 int32_t vout0x2 = (int32_t) fp32_to_bits(vfpacc0x2) - vmagic_bias_less_output_zero_point;
114 int32_t vout0x3 = (int32_t) fp32_to_bits(vfpacc0x3) - vmagic_bias_less_output_zero_point;
115
116 if XNN_LIKELY(nc >= 4) {
117 c0[0] = (uint8_t) vout0x0;
118 c0[1] = (uint8_t) vout0x1;
119 c0[2] = (uint8_t) vout0x2;
120 c0[3] = (uint8_t) vout0x3;
121
122 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
123
124 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
125 nc -= 4;
126 } else {
127 if (nc & 2) {
128 c0[0] = (uint8_t) vout0x0;
129 c0[1] = (uint8_t) vout0x1;
130 vout0x0 = vout0x2;
131 c0 += 2;
132 }
133 if (nc & 1) {
134 c0[0] = (uint8_t) vout0x0;
135 }
136
137 nc = 0;
138 }
139 } while (nc != 0);
140 }
141