1*4bdc9457SAndroid Build Coastguard Worker // Auto-generated file. Do not edit!
2*4bdc9457SAndroid Build Coastguard Worker // Template: src/qs8-gemm/scalar.c.in
3*4bdc9457SAndroid Build Coastguard Worker // Generator: tools/xngen
4*4bdc9457SAndroid Build Coastguard Worker //
5*4bdc9457SAndroid Build Coastguard Worker // Copyright 2021 Google LLC
6*4bdc9457SAndroid Build Coastguard Worker //
7*4bdc9457SAndroid Build Coastguard Worker // This source code is licensed under the BSD-style license found in the
8*4bdc9457SAndroid Build Coastguard Worker // LICENSE file in the root directory of this source tree.
9*4bdc9457SAndroid Build Coastguard Worker
10*4bdc9457SAndroid Build Coastguard Worker #include <assert.h>
11*4bdc9457SAndroid Build Coastguard Worker #include <math.h>
12*4bdc9457SAndroid Build Coastguard Worker
13*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/math.h>
14*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/gemm.h>
15*4bdc9457SAndroid Build Coastguard Worker
16*4bdc9457SAndroid Build Coastguard Worker
xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf(size_t mr,size_t nc,size_t kc,const uint8_t * restrict a,size_t a_stride,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17*4bdc9457SAndroid Build Coastguard Worker void xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf(
18*4bdc9457SAndroid Build Coastguard Worker size_t mr,
19*4bdc9457SAndroid Build Coastguard Worker size_t nc,
20*4bdc9457SAndroid Build Coastguard Worker size_t kc,
21*4bdc9457SAndroid Build Coastguard Worker const uint8_t* restrict a,
22*4bdc9457SAndroid Build Coastguard Worker size_t a_stride,
23*4bdc9457SAndroid Build Coastguard Worker const void* restrict w,
24*4bdc9457SAndroid Build Coastguard Worker uint8_t* restrict c,
25*4bdc9457SAndroid Build Coastguard Worker size_t cm_stride,
26*4bdc9457SAndroid Build Coastguard Worker size_t cn_stride,
27*4bdc9457SAndroid Build Coastguard Worker const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
28*4bdc9457SAndroid Build Coastguard Worker {
29*4bdc9457SAndroid Build Coastguard Worker assert(mr != 0);
30*4bdc9457SAndroid Build Coastguard Worker assert(mr <= 1);
31*4bdc9457SAndroid Build Coastguard Worker assert(nc != 0);
32*4bdc9457SAndroid Build Coastguard Worker assert(kc != 0);
33*4bdc9457SAndroid Build Coastguard Worker
34*4bdc9457SAndroid Build Coastguard Worker const uint8_t* a0 = a;
35*4bdc9457SAndroid Build Coastguard Worker uint8_t* c0 = c;
36*4bdc9457SAndroid Build Coastguard Worker
37*4bdc9457SAndroid Build Coastguard Worker const int32_t vb_zero_point = params->fp32_scalar_lrintf.kernel_zero_point;
38*4bdc9457SAndroid Build Coastguard Worker do {
39*4bdc9457SAndroid Build Coastguard Worker int32_t vacc0x0 = ((const int32_t*) w)[0];
40*4bdc9457SAndroid Build Coastguard Worker int32_t vacc0x1 = ((const int32_t*) w)[1];
41*4bdc9457SAndroid Build Coastguard Worker int32_t vacc0x2 = ((const int32_t*) w)[2];
42*4bdc9457SAndroid Build Coastguard Worker int32_t vacc0x3 = ((const int32_t*) w)[3];
43*4bdc9457SAndroid Build Coastguard Worker w = (const void*) ((const int32_t*) w + 4);
44*4bdc9457SAndroid Build Coastguard Worker
45*4bdc9457SAndroid Build Coastguard Worker size_t k = kc;
46*4bdc9457SAndroid Build Coastguard Worker do {
47*4bdc9457SAndroid Build Coastguard Worker const int32_t va0 = (int32_t) (uint32_t) *a0++;
48*4bdc9457SAndroid Build Coastguard Worker
49*4bdc9457SAndroid Build Coastguard Worker const int32_t vb0 = (int32_t) (uint32_t) ((const uint8_t*) w)[0] - vb_zero_point;
50*4bdc9457SAndroid Build Coastguard Worker const int32_t vb1 = (int32_t) (uint32_t) ((const uint8_t*) w)[1] - vb_zero_point;
51*4bdc9457SAndroid Build Coastguard Worker const int32_t vb2 = (int32_t) (uint32_t) ((const uint8_t*) w)[2] - vb_zero_point;
52*4bdc9457SAndroid Build Coastguard Worker const int32_t vb3 = (int32_t) (uint32_t) ((const uint8_t*) w)[3] - vb_zero_point;
53*4bdc9457SAndroid Build Coastguard Worker w = (const void*) ((const uint8_t*) w + 4);
54*4bdc9457SAndroid Build Coastguard Worker
55*4bdc9457SAndroid Build Coastguard Worker vacc0x0 += va0 * vb0;
56*4bdc9457SAndroid Build Coastguard Worker vacc0x1 += va0 * vb1;
57*4bdc9457SAndroid Build Coastguard Worker vacc0x2 += va0 * vb2;
58*4bdc9457SAndroid Build Coastguard Worker vacc0x3 += va0 * vb3;
59*4bdc9457SAndroid Build Coastguard Worker
60*4bdc9457SAndroid Build Coastguard Worker k -= sizeof(uint8_t);
61*4bdc9457SAndroid Build Coastguard Worker } while (k != 0);
62*4bdc9457SAndroid Build Coastguard Worker
63*4bdc9457SAndroid Build Coastguard Worker float vfpacc0x0 = (float) vacc0x0;
64*4bdc9457SAndroid Build Coastguard Worker float vfpacc0x1 = (float) vacc0x1;
65*4bdc9457SAndroid Build Coastguard Worker float vfpacc0x2 = (float) vacc0x2;
66*4bdc9457SAndroid Build Coastguard Worker float vfpacc0x3 = (float) vacc0x3;
67*4bdc9457SAndroid Build Coastguard Worker
68*4bdc9457SAndroid Build Coastguard Worker const float vscale = params->fp32_scalar_lrintf.scale;
69*4bdc9457SAndroid Build Coastguard Worker vfpacc0x0 *= vscale;
70*4bdc9457SAndroid Build Coastguard Worker vfpacc0x1 *= vscale;
71*4bdc9457SAndroid Build Coastguard Worker vfpacc0x2 *= vscale;
72*4bdc9457SAndroid Build Coastguard Worker vfpacc0x3 *= vscale;
73*4bdc9457SAndroid Build Coastguard Worker
74*4bdc9457SAndroid Build Coastguard Worker const float voutput_min_less_zero_point = params->fp32_scalar_lrintf.output_min_less_zero_point;
75*4bdc9457SAndroid Build Coastguard Worker vfpacc0x0 = math_max_f32(vfpacc0x0, voutput_min_less_zero_point);
76*4bdc9457SAndroid Build Coastguard Worker vfpacc0x1 = math_max_f32(vfpacc0x1, voutput_min_less_zero_point);
77*4bdc9457SAndroid Build Coastguard Worker vfpacc0x2 = math_max_f32(vfpacc0x2, voutput_min_less_zero_point);
78*4bdc9457SAndroid Build Coastguard Worker vfpacc0x3 = math_max_f32(vfpacc0x3, voutput_min_less_zero_point);
79*4bdc9457SAndroid Build Coastguard Worker
80*4bdc9457SAndroid Build Coastguard Worker const float voutput_max_less_zero_point = params->fp32_scalar_lrintf.output_max_less_zero_point;
81*4bdc9457SAndroid Build Coastguard Worker vfpacc0x0 = math_min_f32(vfpacc0x0, voutput_max_less_zero_point);
82*4bdc9457SAndroid Build Coastguard Worker vfpacc0x1 = math_min_f32(vfpacc0x1, voutput_max_less_zero_point);
83*4bdc9457SAndroid Build Coastguard Worker vfpacc0x2 = math_min_f32(vfpacc0x2, voutput_max_less_zero_point);
84*4bdc9457SAndroid Build Coastguard Worker vfpacc0x3 = math_min_f32(vfpacc0x3, voutput_max_less_zero_point);
85*4bdc9457SAndroid Build Coastguard Worker
86*4bdc9457SAndroid Build Coastguard Worker const int32_t vrndacc0x0 = (int32_t) lrintf(vfpacc0x0);
87*4bdc9457SAndroid Build Coastguard Worker const int32_t vrndacc0x1 = (int32_t) lrintf(vfpacc0x1);
88*4bdc9457SAndroid Build Coastguard Worker const int32_t vrndacc0x2 = (int32_t) lrintf(vfpacc0x2);
89*4bdc9457SAndroid Build Coastguard Worker const int32_t vrndacc0x3 = (int32_t) lrintf(vfpacc0x3);
90*4bdc9457SAndroid Build Coastguard Worker
91*4bdc9457SAndroid Build Coastguard Worker const int32_t voutput_zero_point = params->fp32_scalar_lrintf.output_zero_point;
92*4bdc9457SAndroid Build Coastguard Worker int32_t vout0x0 = vrndacc0x0 + voutput_zero_point;
93*4bdc9457SAndroid Build Coastguard Worker int32_t vout0x1 = vrndacc0x1 + voutput_zero_point;
94*4bdc9457SAndroid Build Coastguard Worker int32_t vout0x2 = vrndacc0x2 + voutput_zero_point;
95*4bdc9457SAndroid Build Coastguard Worker int32_t vout0x3 = vrndacc0x3 + voutput_zero_point;
96*4bdc9457SAndroid Build Coastguard Worker
97*4bdc9457SAndroid Build Coastguard Worker if XNN_LIKELY(nc >= 4) {
98*4bdc9457SAndroid Build Coastguard Worker c0[0] = (uint8_t) vout0x0;
99*4bdc9457SAndroid Build Coastguard Worker c0[1] = (uint8_t) vout0x1;
100*4bdc9457SAndroid Build Coastguard Worker c0[2] = (uint8_t) vout0x2;
101*4bdc9457SAndroid Build Coastguard Worker c0[3] = (uint8_t) vout0x3;
102*4bdc9457SAndroid Build Coastguard Worker
103*4bdc9457SAndroid Build Coastguard Worker a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
104*4bdc9457SAndroid Build Coastguard Worker
105*4bdc9457SAndroid Build Coastguard Worker c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
106*4bdc9457SAndroid Build Coastguard Worker
107*4bdc9457SAndroid Build Coastguard Worker nc -= 4;
108*4bdc9457SAndroid Build Coastguard Worker } else {
109*4bdc9457SAndroid Build Coastguard Worker if (nc & 2) {
110*4bdc9457SAndroid Build Coastguard Worker c0[0] = (uint8_t) vout0x0;
111*4bdc9457SAndroid Build Coastguard Worker c0[1] = (uint8_t) vout0x1;
112*4bdc9457SAndroid Build Coastguard Worker vout0x0 = vout0x2;
113*4bdc9457SAndroid Build Coastguard Worker c0 += 2;
114*4bdc9457SAndroid Build Coastguard Worker }
115*4bdc9457SAndroid Build Coastguard Worker if (nc & 1) {
116*4bdc9457SAndroid Build Coastguard Worker c0[0] = (uint8_t) vout0x0;
117*4bdc9457SAndroid Build Coastguard Worker }
118*4bdc9457SAndroid Build Coastguard Worker
119*4bdc9457SAndroid Build Coastguard Worker nc = 0;
120*4bdc9457SAndroid Build Coastguard Worker }
121*4bdc9457SAndroid Build Coastguard Worker } while (nc != 0);
122*4bdc9457SAndroid Build Coastguard Worker }
123