• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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_1x2__scalar_imagic(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_1x2__scalar_imagic(
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_imagic.kernel_zero_point;
45   do {
46     int32_t vacc0x0 = ((const int32_t*) w)[0];
47     int32_t vacc0x1 = ((const int32_t*) w)[1];
48     w = (const void*) ((const int32_t*) w + 2);
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         w = (const void*) ((const uint8_t*) w + 2);
66 
67         vacc0x0 += va0 * vb0;
68         vacc0x1 += va0 * vb1;
69 
70         k -= sizeof(uint8_t);
71       } while (k != 0);
72       p -= 1 * sizeof(void*);
73     } while (p != 0);
74 
75     float vfpacc0x0 = (float) vacc0x0;
76     float vfpacc0x1 = (float) vacc0x1;
77 
78     const float vscale = params->fp32_scalar_imagic.scale;
79     vfpacc0x0 *= vscale;
80     vfpacc0x1 *= vscale;
81 
82     const float vmagic_bias = params->fp32_scalar_imagic.magic_bias;
83     vfpacc0x0 += vmagic_bias;
84     vfpacc0x1 += vmagic_bias;
85 
86     int32_t vout0x0 = (int32_t) fp32_to_bits(vfpacc0x0);
87     int32_t vout0x1 = (int32_t) fp32_to_bits(vfpacc0x1);
88 
89     const int32_t vmagic_min = params->fp32_scalar_imagic.magic_min;
90     vout0x0 = math_max_s32(vout0x0, vmagic_min);
91     vout0x1 = math_max_s32(vout0x1, vmagic_min);
92 
93     const int32_t vmagic_max = params->fp32_scalar_imagic.magic_max;
94     vout0x0 = math_min_s32(vout0x0, vmagic_max);
95     vout0x1 = math_min_s32(vout0x1, vmagic_max);
96 
97     const int32_t vmagic_bias_less_zero_point = params->fp32_scalar_imagic.magic_bias_less_zero_point;
98     vout0x0 -= vmagic_bias_less_zero_point;
99     vout0x1 -= vmagic_bias_less_zero_point;
100 
101     if XNN_LIKELY(nc >= 2) {
102       c0[0] = (uint8_t) vout0x0;
103       c0[1] = (uint8_t) vout0x1;
104 
105       c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
106 
107       a = (const uint8_t**restrict) ((uintptr_t) a - ks);
108       nc -= 2;
109     } else {
110       if (nc & 1) {
111         c0[0] = (uint8_t) vout0x0;
112       }
113 
114       nc = 0;
115     }
116   } while (nc != 0);
117 }
118