• 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 <xnnpack/math.h>
13 #include <xnnpack/gemm.h>
14 #include <xnnpack/unaligned.h>
15 
16 
xnn_qu8_igemm_minmax_fp32_ukernel_3x2__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)])17 void xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic(
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 <= 3);
33   assert(nc != 0);
34   assert(kc != 0);
35   assert(ks != 0);
36   assert(ks % (3 * 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   uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
47   if XNN_UNPREDICTABLE(mr <= 2) {
48     c2 = c1;
49   }
50 
51   const int32_t vb_zero_point = params->fp32_scalar_imagic.kernel_zero_point;
52   do {
53     int32_t vacc0x0 = unaligned_indexed_load_s32(w, 0);
54     int32_t vacc0x1 = unaligned_indexed_load_s32(w, 1);
55     int32_t vacc1x0 = vacc0x0;
56     int32_t vacc1x1 = vacc0x1;
57     int32_t vacc2x0 = vacc0x0;
58     int32_t vacc2x1 = vacc0x1;
59     w = (const void*) ((const int32_t*) w + 2);
60 
61     size_t p = ks;
62     do {
63       const uint8_t* restrict a0 = a[0];
64       assert(a0 != NULL);
65       if XNN_UNPREDICTABLE(a0 != zero) {
66         a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
67       }
68       const uint8_t* restrict a1 = a[1];
69       assert(a1 != NULL);
70       if XNN_UNPREDICTABLE(a1 != zero) {
71         a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
72       }
73       const uint8_t* restrict a2 = a[2];
74       assert(a2 != NULL);
75       if XNN_UNPREDICTABLE(a2 != zero) {
76         a2 = (const uint8_t*) ((uintptr_t) a2 + a_offset);
77       }
78       a += 3;
79 
80       size_t k = kc;
81       do {
82         const int32_t va0 = (int32_t) (uint32_t) *a0++;
83         const int32_t va1 = (int32_t) (uint32_t) *a1++;
84         const int32_t va2 = (int32_t) (uint32_t) *a2++;
85 
86         const int32_t vb0 = (int32_t) (uint32_t) ((const uint8_t*) w)[0] - vb_zero_point;
87         const int32_t vb1 = (int32_t) (uint32_t) ((const uint8_t*) w)[1] - vb_zero_point;
88         w = (const void*) ((const uint8_t*) w + 2);
89 
90         vacc0x0 += va0 * vb0;
91         vacc0x1 += va0 * vb1;
92         vacc1x0 += va1 * vb0;
93         vacc1x1 += va1 * vb1;
94         vacc2x0 += va2 * vb0;
95         vacc2x1 += va2 * vb1;
96 
97         k -= sizeof(uint8_t);
98       } while (k != 0);
99       p -= 3 * sizeof(void*);
100     } while (p != 0);
101 
102     float vfpacc0x0 = (float) vacc0x0;
103     float vfpacc0x1 = (float) vacc0x1;
104     float vfpacc1x0 = (float) vacc1x0;
105     float vfpacc1x1 = (float) vacc1x1;
106     float vfpacc2x0 = (float) vacc2x0;
107     float vfpacc2x1 = (float) vacc2x1;
108 
109     const float vscale = params->fp32_scalar_imagic.scale;
110     vfpacc0x0 *= vscale;
111     vfpacc0x1 *= vscale;
112     vfpacc1x0 *= vscale;
113     vfpacc1x1 *= vscale;
114     vfpacc2x0 *= vscale;
115     vfpacc2x1 *= vscale;
116 
117     const float vmagic_bias = params->fp32_scalar_imagic.magic_bias;
118     vfpacc0x0 += vmagic_bias;
119     vfpacc0x1 += vmagic_bias;
120     vfpacc1x0 += vmagic_bias;
121     vfpacc1x1 += vmagic_bias;
122     vfpacc2x0 += vmagic_bias;
123     vfpacc2x1 += vmagic_bias;
124 
125     int32_t vout0x0 = (int32_t) float_as_uint32(vfpacc0x0);
126     int32_t vout0x1 = (int32_t) float_as_uint32(vfpacc0x1);
127     int32_t vout1x0 = (int32_t) float_as_uint32(vfpacc1x0);
128     int32_t vout1x1 = (int32_t) float_as_uint32(vfpacc1x1);
129     int32_t vout2x0 = (int32_t) float_as_uint32(vfpacc2x0);
130     int32_t vout2x1 = (int32_t) float_as_uint32(vfpacc2x1);
131 
132     const int32_t vmagic_min = params->fp32_scalar_imagic.magic_min;
133     vout0x0 = math_max_s32(vout0x0, vmagic_min);
134     vout0x1 = math_max_s32(vout0x1, vmagic_min);
135     vout1x0 = math_max_s32(vout1x0, vmagic_min);
136     vout1x1 = math_max_s32(vout1x1, vmagic_min);
137     vout2x0 = math_max_s32(vout2x0, vmagic_min);
138     vout2x1 = math_max_s32(vout2x1, vmagic_min);
139 
140     const int32_t vmagic_max = params->fp32_scalar_imagic.magic_max;
141     vout0x0 = math_min_s32(vout0x0, vmagic_max);
142     vout0x1 = math_min_s32(vout0x1, vmagic_max);
143     vout1x0 = math_min_s32(vout1x0, vmagic_max);
144     vout1x1 = math_min_s32(vout1x1, vmagic_max);
145     vout2x0 = math_min_s32(vout2x0, vmagic_max);
146     vout2x1 = math_min_s32(vout2x1, vmagic_max);
147 
148     const int32_t vmagic_bias_less_zero_point = params->fp32_scalar_imagic.magic_bias_less_zero_point;
149     vout0x0 -= vmagic_bias_less_zero_point;
150     vout0x1 -= vmagic_bias_less_zero_point;
151     vout1x0 -= vmagic_bias_less_zero_point;
152     vout1x1 -= vmagic_bias_less_zero_point;
153     vout2x0 -= vmagic_bias_less_zero_point;
154     vout2x1 -= vmagic_bias_less_zero_point;
155 
156     if XNN_LIKELY(nc >= 2) {
157       c2[0] = (uint8_t) vout2x0;
158       c2[1] = (uint8_t) vout2x1;
159       c1[0] = (uint8_t) vout1x0;
160       c1[1] = (uint8_t) vout1x1;
161       c0[0] = (uint8_t) vout0x0;
162       c0[1] = (uint8_t) vout0x1;
163 
164       c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
165       c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
166       c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
167 
168       a = (const uint8_t**restrict) ((uintptr_t) a - ks);
169       nc -= 2;
170     } else {
171       if (nc & 1) {
172         c2[0] = (uint8_t) vout2x0;
173         c1[0] = (uint8_t) vout1x0;
174         c0[0] = (uint8_t) vout0x0;
175       }
176 
177       nc = 0;
178     }
179   } while (nc != 0);
180 }
181