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_1x2__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_1x2__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 <= 1);
31 assert(nc != 0);
32 assert(kc != 0);
33
34 const int8_t* a0 = a;
35 int8_t* c0 = c;
36
37 do {
38 int32_t vacc0x0 = unaligned_indexed_load_s32(w, 0);
39 int32_t vacc0x1 = unaligned_indexed_load_s32(w, 1);
40 w = (const void*) ((const int32_t*) w + 2);
41
42 size_t k = kc;
43 do {
44 const int32_t va0 = (int32_t) *a0++;
45
46 const int32_t vb0 = (int32_t) ((const int8_t*) w)[0];
47 const int32_t vb1 = (int32_t) ((const int8_t*) w)[1];
48 w = (const void*) ((const int8_t*) w + 2);
49
50 vacc0x0 += va0 * vb0;
51 vacc0x1 += va0 * vb1;
52
53 k -= sizeof(int8_t);
54 } while (k != 0);
55
56 float vfpacc0x0 = (float) vacc0x0;
57 float vfpacc0x1 = (float) vacc0x1;
58
59 const float vscale0 = unaligned_indexed_load_f32(w, 0);
60 vfpacc0x0 *= vscale0;
61 const float vscale1 = unaligned_indexed_load_f32(w, 1);
62 vfpacc0x1 *= vscale1;
63 w = (const void*) ((const float*) w + 2);
64
65 const float voutput_min_less_zero_point = params->fp32_scalar_fmagic.output_min_less_zero_point;
66 vfpacc0x0 = math_max_f32(vfpacc0x0, voutput_min_less_zero_point);
67 vfpacc0x1 = math_max_f32(vfpacc0x1, voutput_min_less_zero_point);
68
69 const float voutput_max_less_zero_point = params->fp32_scalar_fmagic.output_max_less_zero_point;
70 vfpacc0x0 = math_min_f32(vfpacc0x0, voutput_max_less_zero_point);
71 vfpacc0x1 = math_min_f32(vfpacc0x1, voutput_max_less_zero_point);
72
73 const float vmagic_bias = params->fp32_scalar_fmagic.magic_bias;
74 vfpacc0x0 += vmagic_bias;
75 vfpacc0x1 += vmagic_bias;
76
77 const int32_t vmagic_bias_less_output_zero_point = params->fp32_scalar_fmagic.magic_bias_less_output_zero_point;
78 int32_t vout0x0 = (int32_t) float_as_uint32(vfpacc0x0) - vmagic_bias_less_output_zero_point;
79 int32_t vout0x1 = (int32_t) float_as_uint32(vfpacc0x1) - vmagic_bias_less_output_zero_point;
80
81 if XNN_LIKELY(nc >= 2) {
82 c0[0] = (int8_t) vout0x0;
83 c0[1] = (int8_t) vout0x1;
84
85 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
86
87 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
88
89 nc -= 2;
90 } else {
91 if (nc & 1) {
92 c0[0] = (int8_t) vout0x0;
93 }
94
95 nc = 0;
96 }
97 } while (nc != 0);
98 }
99