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 #include <xnnpack/unaligned.h>
15
16
xnn_qu8_igemm_minmax_fp32_ukernel_2x2__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)])17 void xnn_qu8_igemm_minmax_fp32_ukernel_2x2__wasm_fmagic(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 size_t ks,
22 const uint8_t**restrict a,
23 const void*restrict w,
24 uint8_t*restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 size_t a_offset,
28 const uint8_t* zero,
29 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
30 {
31 assert(mr != 0);
32 assert(mr <= 2);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(ks != 0);
36 assert(ks % (2 * sizeof(void*)) == 0);
37 assert(a != NULL);
38 assert(w != NULL);
39 assert(c != NULL);
40
41 uint8_t* c0 = c;
42 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
43 if XNN_UNPREDICTABLE(mr != 2) {
44 c1 = c0;
45 }
46
47 const int32_t vb_zero_point = params->fp32_scalar_fmagic.kernel_zero_point;
48 do {
49 int32_t vacc0x0 = unaligned_indexed_load_s32(w, 0);
50 int32_t vacc0x1 = unaligned_indexed_load_s32(w, 1);
51 int32_t vacc1x0 = vacc0x0;
52 int32_t vacc1x1 = vacc0x1;
53 w = (const void*) ((const int32_t*) w + 2);
54
55 size_t p = ks;
56 do {
57 const uint8_t* restrict a0 = a[0];
58 assert(a0 != NULL);
59 if XNN_UNPREDICTABLE(a0 != zero) {
60 a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
61 }
62 const uint8_t* restrict a1 = a[1];
63 assert(a1 != NULL);
64 if XNN_UNPREDICTABLE(a1 != zero) {
65 a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
66 }
67 a += 2;
68
69 size_t k = kc;
70 do {
71 const int32_t va0 = (int32_t) (uint32_t) *a0++;
72 const int32_t va1 = (int32_t) (uint32_t) *a1++;
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
83 k -= sizeof(uint8_t);
84 } while (k != 0);
85 p -= 2 * sizeof(void*);
86 } while (p != 0);
87
88 float vfpacc0x0 = (float) vacc0x0;
89 float vfpacc0x1 = (float) vacc0x1;
90 float vfpacc1x0 = (float) vacc1x0;
91 float vfpacc1x1 = (float) vacc1x1;
92
93 const float vscale = params->fp32_scalar_fmagic.scale;
94 vfpacc0x0 *= vscale;
95 vfpacc0x1 *= vscale;
96 vfpacc1x0 *= vscale;
97 vfpacc1x1 *= vscale;
98
99 const float voutput_min_less_zero_point = params->fp32_scalar_fmagic.output_min_less_zero_point;
100 vfpacc0x0 = __builtin_wasm_max_f32(vfpacc0x0, voutput_min_less_zero_point);
101 vfpacc0x1 = __builtin_wasm_max_f32(vfpacc0x1, voutput_min_less_zero_point);
102 vfpacc1x0 = __builtin_wasm_max_f32(vfpacc1x0, voutput_min_less_zero_point);
103 vfpacc1x1 = __builtin_wasm_max_f32(vfpacc1x1, voutput_min_less_zero_point);
104
105 const float voutput_max_less_zero_point = params->fp32_scalar_fmagic.output_max_less_zero_point;
106 vfpacc0x0 = __builtin_wasm_min_f32(vfpacc0x0, voutput_max_less_zero_point);
107 vfpacc0x1 = __builtin_wasm_min_f32(vfpacc0x1, voutput_max_less_zero_point);
108 vfpacc1x0 = __builtin_wasm_min_f32(vfpacc1x0, voutput_max_less_zero_point);
109 vfpacc1x1 = __builtin_wasm_min_f32(vfpacc1x1, voutput_max_less_zero_point);
110
111 const float vmagic_bias = params->fp32_scalar_fmagic.magic_bias;
112 vfpacc0x0 += vmagic_bias;
113 vfpacc0x1 += vmagic_bias;
114 vfpacc1x0 += vmagic_bias;
115 vfpacc1x1 += vmagic_bias;
116
117 const int32_t vmagic_bias_less_output_zero_point = params->fp32_scalar_fmagic.magic_bias_less_output_zero_point;
118 int32_t vout0x0 = (int32_t) float_as_uint32(vfpacc0x0) - vmagic_bias_less_output_zero_point;
119 int32_t vout0x1 = (int32_t) float_as_uint32(vfpacc0x1) - vmagic_bias_less_output_zero_point;
120 int32_t vout1x0 = (int32_t) float_as_uint32(vfpacc1x0) - vmagic_bias_less_output_zero_point;
121 int32_t vout1x1 = (int32_t) float_as_uint32(vfpacc1x1) - vmagic_bias_less_output_zero_point;
122
123 if XNN_LIKELY(nc >= 2) {
124 c1[0] = (uint8_t) vout1x0;
125 c1[1] = (uint8_t) vout1x1;
126 c0[0] = (uint8_t) vout0x0;
127 c0[1] = (uint8_t) vout0x1;
128
129 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
130 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
131
132 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
133 nc -= 2;
134 } else {
135 if (nc & 1) {
136 c1[0] = (uint8_t) vout1x0;
137 c0[0] = (uint8_t) vout0x0;
138 }
139
140 nc = 0;
141 }
142 } while (nc != 0);
143 }
144