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
15
xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_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)])16 void xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_fmagic(
17 size_t mr,
18 size_t nc,
19 size_t kc,
20 size_t ks,
21 const uint8_t**restrict a,
22 const void*restrict w,
23 uint8_t*restrict c,
24 size_t cm_stride,
25 size_t cn_stride,
26 size_t a_offset,
27 const uint8_t* zero,
28 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
29 {
30 assert(mr != 0);
31 assert(mr <= 1);
32 assert(nc != 0);
33 assert(kc != 0);
34 assert(ks != 0);
35 assert(ks % (1 * sizeof(void*)) == 0);
36 assert(a != NULL);
37 assert(w != NULL);
38 assert(c != NULL);
39
40 uint8_t* c0 = c;
41
42 const int32_t vb_zero_point = params->fp32_scalar_fmagic.kernel_zero_point;
43 do {
44 int32_t vacc0x0 = ((const int32_t*) w)[0];
45 int32_t vacc0x1 = ((const int32_t*) w)[1];
46 int32_t vacc0x2 = ((const int32_t*) w)[2];
47 int32_t vacc0x3 = ((const int32_t*) w)[3];
48 w = (const void*) ((const int32_t*) w + 4);
49
50 size_t p = ks;
51 do {
52 const uint8_t* restrict a0 = a[0];
53 assert(a0 != NULL);
54 if XNN_UNPREDICTABLE(a0 != zero) {
55 a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
56 }
57 a += 1;
58
59 size_t k = kc;
60 do {
61 const int32_t va0 = (int32_t) (uint32_t) *a0++;
62
63 const int32_t vb0 = (int32_t) (uint32_t) ((const uint8_t*) w)[0] - vb_zero_point;
64 const int32_t vb1 = (int32_t) (uint32_t) ((const uint8_t*) w)[1] - vb_zero_point;
65 const int32_t vb2 = (int32_t) (uint32_t) ((const uint8_t*) w)[2] - vb_zero_point;
66 const int32_t vb3 = (int32_t) (uint32_t) ((const uint8_t*) w)[3] - vb_zero_point;
67 w = (const void*) ((const uint8_t*) w + 4);
68
69 vacc0x0 += va0 * vb0;
70 vacc0x1 += va0 * vb1;
71 vacc0x2 += va0 * vb2;
72 vacc0x3 += va0 * vb3;
73
74 k -= sizeof(uint8_t);
75 } while (k != 0);
76 p -= 1 * sizeof(void*);
77 } while (p != 0);
78
79 float vfpacc0x0 = (float) vacc0x0;
80 float vfpacc0x1 = (float) vacc0x1;
81 float vfpacc0x2 = (float) vacc0x2;
82 float vfpacc0x3 = (float) vacc0x3;
83
84 const float vscale = params->fp32_scalar_fmagic.scale;
85 vfpacc0x0 *= vscale;
86 vfpacc0x1 *= vscale;
87 vfpacc0x2 *= vscale;
88 vfpacc0x3 *= vscale;
89
90 const float voutput_min_less_zero_point = params->fp32_scalar_fmagic.output_min_less_zero_point;
91 vfpacc0x0 = math_max_f32(vfpacc0x0, voutput_min_less_zero_point);
92 vfpacc0x1 = math_max_f32(vfpacc0x1, voutput_min_less_zero_point);
93 vfpacc0x2 = math_max_f32(vfpacc0x2, voutput_min_less_zero_point);
94 vfpacc0x3 = math_max_f32(vfpacc0x3, voutput_min_less_zero_point);
95
96 const float voutput_max_less_zero_point = params->fp32_scalar_fmagic.output_max_less_zero_point;
97 vfpacc0x0 = math_min_f32(vfpacc0x0, voutput_max_less_zero_point);
98 vfpacc0x1 = math_min_f32(vfpacc0x1, voutput_max_less_zero_point);
99 vfpacc0x2 = math_min_f32(vfpacc0x2, voutput_max_less_zero_point);
100 vfpacc0x3 = math_min_f32(vfpacc0x3, voutput_max_less_zero_point);
101
102 const float vmagic_bias = params->fp32_scalar_fmagic.magic_bias;
103 vfpacc0x0 += vmagic_bias;
104 vfpacc0x1 += vmagic_bias;
105 vfpacc0x2 += vmagic_bias;
106 vfpacc0x3 += vmagic_bias;
107
108 const int32_t vmagic_bias_less_output_zero_point = params->fp32_scalar_fmagic.magic_bias_less_output_zero_point;
109 int32_t vout0x0 = (int32_t) float_as_uint32(vfpacc0x0) - vmagic_bias_less_output_zero_point;
110 int32_t vout0x1 = (int32_t) float_as_uint32(vfpacc0x1) - vmagic_bias_less_output_zero_point;
111 int32_t vout0x2 = (int32_t) float_as_uint32(vfpacc0x2) - vmagic_bias_less_output_zero_point;
112 int32_t vout0x3 = (int32_t) float_as_uint32(vfpacc0x3) - vmagic_bias_less_output_zero_point;
113
114 if XNN_LIKELY(nc >= 4) {
115 c0[0] = (uint8_t) vout0x0;
116 c0[1] = (uint8_t) vout0x1;
117 c0[2] = (uint8_t) vout0x2;
118 c0[3] = (uint8_t) vout0x3;
119
120 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
121
122 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
123 nc -= 4;
124 } else {
125 if (nc & 2) {
126 c0[0] = (uint8_t) vout0x0;
127 c0[1] = (uint8_t) vout0x1;
128 vout0x0 = vout0x2;
129 c0 += 2;
130 }
131 if (nc & 1) {
132 c0[0] = (uint8_t) vout0x0;
133 }
134
135 nc = 0;
136 }
137 } while (nc != 0);
138 }
139