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_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_fmagic(size_t mr,size_t nc,size_t kc,const int8_t * restrict a,size_t a_stride,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_fmagic(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 const int8_t* restrict a,
22 size_t a_stride,
23 const void* restrict w,
24 int8_t* restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
28 {
29 assert(mr != 0);
30 assert(mr <= 3);
31 assert(nc != 0);
32 assert(kc != 0);
33
34 const int8_t* a0 = a;
35 int8_t* c0 = c;
36 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
37 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
38 if XNN_UNPREDICTABLE(mr < 2) {
39 a1 = a0;
40 c1 = c0;
41 }
42 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
43 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
44 if XNN_UNPREDICTABLE(mr <= 2) {
45 a2 = a1;
46 c2 = c1;
47 }
48
49 do {
50 int32_t vacc0x0 = unaligned_indexed_load_s32(w, 0);
51 int32_t vacc0x1 = unaligned_indexed_load_s32(w, 1);
52 int32_t vacc1x0 = vacc0x0;
53 int32_t vacc1x1 = vacc0x1;
54 int32_t vacc2x0 = vacc0x0;
55 int32_t vacc2x1 = vacc0x1;
56 w = (const void*) ((const int32_t*) w + 2);
57
58 size_t k = kc;
59 do {
60 const int32_t va0 = (int32_t) *a0++;
61 const int32_t va1 = (int32_t) *a1++;
62 const int32_t va2 = (int32_t) *a2++;
63
64 const int32_t vb0 = (int32_t) ((const int8_t*) w)[0];
65 const int32_t vb1 = (int32_t) ((const int8_t*) w)[1];
66 w = (const void*) ((const int8_t*) w + 2);
67
68 vacc0x0 += va0 * vb0;
69 vacc0x1 += va0 * vb1;
70 vacc1x0 += va1 * vb0;
71 vacc1x1 += va1 * vb1;
72 vacc2x0 += va2 * vb0;
73 vacc2x1 += va2 * vb1;
74
75 k -= sizeof(int8_t);
76 } while (k != 0);
77
78 float vfpacc0x0 = (float) vacc0x0;
79 float vfpacc0x1 = (float) vacc0x1;
80 float vfpacc1x0 = (float) vacc1x0;
81 float vfpacc1x1 = (float) vacc1x1;
82 float vfpacc2x0 = (float) vacc2x0;
83 float vfpacc2x1 = (float) vacc2x1;
84
85 const float vscale0 = unaligned_indexed_load_f32(w, 0);
86 vfpacc0x0 *= vscale0;
87 vfpacc1x0 *= vscale0;
88 vfpacc2x0 *= vscale0;
89 const float vscale1 = unaligned_indexed_load_f32(w, 1);
90 vfpacc0x1 *= vscale1;
91 vfpacc1x1 *= vscale1;
92 vfpacc2x1 *= vscale1;
93 w = (const void*) ((const float*) w + 2);
94
95 const float voutput_min_less_zero_point = params->fp32_scalar_fmagic.output_min_less_zero_point;
96 vfpacc0x0 = math_max_f32(vfpacc0x0, voutput_min_less_zero_point);
97 vfpacc0x1 = math_max_f32(vfpacc0x1, voutput_min_less_zero_point);
98 vfpacc1x0 = math_max_f32(vfpacc1x0, voutput_min_less_zero_point);
99 vfpacc1x1 = math_max_f32(vfpacc1x1, voutput_min_less_zero_point);
100 vfpacc2x0 = math_max_f32(vfpacc2x0, voutput_min_less_zero_point);
101 vfpacc2x1 = math_max_f32(vfpacc2x1, voutput_min_less_zero_point);
102
103 const float voutput_max_less_zero_point = params->fp32_scalar_fmagic.output_max_less_zero_point;
104 vfpacc0x0 = math_min_f32(vfpacc0x0, voutput_max_less_zero_point);
105 vfpacc0x1 = math_min_f32(vfpacc0x1, voutput_max_less_zero_point);
106 vfpacc1x0 = math_min_f32(vfpacc1x0, voutput_max_less_zero_point);
107 vfpacc1x1 = math_min_f32(vfpacc1x1, voutput_max_less_zero_point);
108 vfpacc2x0 = math_min_f32(vfpacc2x0, voutput_max_less_zero_point);
109 vfpacc2x1 = math_min_f32(vfpacc2x1, 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 vfpacc2x0 += vmagic_bias;
117 vfpacc2x1 += vmagic_bias;
118
119 const int32_t vmagic_bias_less_output_zero_point = params->fp32_scalar_fmagic.magic_bias_less_output_zero_point;
120 int32_t vout0x0 = (int32_t) float_as_uint32(vfpacc0x0) - vmagic_bias_less_output_zero_point;
121 int32_t vout0x1 = (int32_t) float_as_uint32(vfpacc0x1) - vmagic_bias_less_output_zero_point;
122 int32_t vout1x0 = (int32_t) float_as_uint32(vfpacc1x0) - vmagic_bias_less_output_zero_point;
123 int32_t vout1x1 = (int32_t) float_as_uint32(vfpacc1x1) - vmagic_bias_less_output_zero_point;
124 int32_t vout2x0 = (int32_t) float_as_uint32(vfpacc2x0) - vmagic_bias_less_output_zero_point;
125 int32_t vout2x1 = (int32_t) float_as_uint32(vfpacc2x1) - vmagic_bias_less_output_zero_point;
126
127 if XNN_LIKELY(nc >= 2) {
128 c0[0] = (int8_t) vout0x0;
129 c0[1] = (int8_t) vout0x1;
130 c1[0] = (int8_t) vout1x0;
131 c1[1] = (int8_t) vout1x1;
132 c2[0] = (int8_t) vout2x0;
133 c2[1] = (int8_t) vout2x1;
134
135 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
136 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
137 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
138
139 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
140 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
141 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
142
143 nc -= 2;
144 } else {
145 if (nc & 1) {
146 c0[0] = (int8_t) vout0x0;
147 c1[0] = (int8_t) vout1x0;
148 c2[0] = (int8_t) vout2x0;
149 }
150
151 nc = 0;
152 }
153 } while (nc != 0);
154 }
155