xref: /aosp_15_r20/external/XNNPACK/src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1*4bdc9457SAndroid Build Coastguard Worker // Auto-generated file. Do not edit!
2*4bdc9457SAndroid Build Coastguard Worker //   Template: src/f32-vmulcaddc/scalar.c.in
3*4bdc9457SAndroid Build Coastguard Worker //   Generator: tools/xngen
4*4bdc9457SAndroid Build Coastguard Worker //
5*4bdc9457SAndroid Build Coastguard Worker // Copyright 2019 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 
12*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/math.h>
13*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/vmulcaddc.h>
14*4bdc9457SAndroid Build Coastguard Worker 
15*4bdc9457SAndroid Build Coastguard Worker 
xnn_f32_vmulcaddc_minmax_ukernel_c4__scalar_2x(size_t rows,size_t channels,const float * restrict input,size_t input_stride,const float * restrict weights,float * restrict output,size_t output_stride,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])16*4bdc9457SAndroid Build Coastguard Worker void xnn_f32_vmulcaddc_minmax_ukernel_c4__scalar_2x(
17*4bdc9457SAndroid Build Coastguard Worker     size_t rows,
18*4bdc9457SAndroid Build Coastguard Worker     size_t channels,
19*4bdc9457SAndroid Build Coastguard Worker     const float*restrict input,
20*4bdc9457SAndroid Build Coastguard Worker     size_t input_stride,
21*4bdc9457SAndroid Build Coastguard Worker     const float*restrict weights,
22*4bdc9457SAndroid Build Coastguard Worker     float*restrict output,
23*4bdc9457SAndroid Build Coastguard Worker     size_t output_stride,
24*4bdc9457SAndroid Build Coastguard Worker     const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
25*4bdc9457SAndroid Build Coastguard Worker {
26*4bdc9457SAndroid Build Coastguard Worker   assert(rows != 0);
27*4bdc9457SAndroid Build Coastguard Worker   assert(channels != 0);
28*4bdc9457SAndroid Build Coastguard Worker   assert(channels % sizeof(float) == 0);
29*4bdc9457SAndroid Build Coastguard Worker 
30*4bdc9457SAndroid Build Coastguard Worker   const size_t input_increment = input_stride * 2 - channels;
31*4bdc9457SAndroid Build Coastguard Worker   const size_t output_increment = output_stride * 2 - channels;
32*4bdc9457SAndroid Build Coastguard Worker 
33*4bdc9457SAndroid Build Coastguard Worker   const float* i0 = input;
34*4bdc9457SAndroid Build Coastguard Worker   float* o0 = output;
35*4bdc9457SAndroid Build Coastguard Worker   const float* i1 = (const float*) ((uintptr_t) i0 + input_stride);
36*4bdc9457SAndroid Build Coastguard Worker   float* o1 = (float*) ((uintptr_t) o0 + output_stride);
37*4bdc9457SAndroid Build Coastguard Worker 
38*4bdc9457SAndroid Build Coastguard Worker   const float vmin = params->scalar.min;
39*4bdc9457SAndroid Build Coastguard Worker   const float vmax = params->scalar.max;
40*4bdc9457SAndroid Build Coastguard Worker   do {
41*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(rows < 2) {
42*4bdc9457SAndroid Build Coastguard Worker       i1 = i0;
43*4bdc9457SAndroid Build Coastguard Worker       o1 = o0;
44*4bdc9457SAndroid Build Coastguard Worker     }
45*4bdc9457SAndroid Build Coastguard Worker 
46*4bdc9457SAndroid Build Coastguard Worker     const float* w = weights;
47*4bdc9457SAndroid Build Coastguard Worker     size_t c = channels;
48*4bdc9457SAndroid Build Coastguard Worker     for (; c >= 4 * sizeof(float); c -= 4 * sizeof(float)) {
49*4bdc9457SAndroid Build Coastguard Worker       const float vscale0 = w[0];
50*4bdc9457SAndroid Build Coastguard Worker       const float vscale1 = w[1];
51*4bdc9457SAndroid Build Coastguard Worker       const float vscale2 = w[2];
52*4bdc9457SAndroid Build Coastguard Worker       const float vscale3 = w[3];
53*4bdc9457SAndroid Build Coastguard Worker 
54*4bdc9457SAndroid Build Coastguard Worker       float vacc0x0 = i0[0];
55*4bdc9457SAndroid Build Coastguard Worker       float vacc0x1 = i0[1];
56*4bdc9457SAndroid Build Coastguard Worker       float vacc0x2 = i0[2];
57*4bdc9457SAndroid Build Coastguard Worker       float vacc0x3 = i0[3];
58*4bdc9457SAndroid Build Coastguard Worker       i0 += 4;
59*4bdc9457SAndroid Build Coastguard Worker       float vacc1x0 = i1[0];
60*4bdc9457SAndroid Build Coastguard Worker       float vacc1x1 = i1[1];
61*4bdc9457SAndroid Build Coastguard Worker       float vacc1x2 = i1[2];
62*4bdc9457SAndroid Build Coastguard Worker       float vacc1x3 = i1[3];
63*4bdc9457SAndroid Build Coastguard Worker       i1 += 4;
64*4bdc9457SAndroid Build Coastguard Worker 
65*4bdc9457SAndroid Build Coastguard Worker       const float vbias0 = w[4];
66*4bdc9457SAndroid Build Coastguard Worker       const float vbias1 = w[5];
67*4bdc9457SAndroid Build Coastguard Worker       const float vbias2 = w[6];
68*4bdc9457SAndroid Build Coastguard Worker       const float vbias3 = w[7];
69*4bdc9457SAndroid Build Coastguard Worker 
70*4bdc9457SAndroid Build Coastguard Worker       vacc0x0 = vacc0x0 * vscale0 + vbias0;
71*4bdc9457SAndroid Build Coastguard Worker       vacc0x1 = vacc0x1 * vscale1 + vbias1;
72*4bdc9457SAndroid Build Coastguard Worker       vacc0x2 = vacc0x2 * vscale2 + vbias2;
73*4bdc9457SAndroid Build Coastguard Worker       vacc0x3 = vacc0x3 * vscale3 + vbias3;
74*4bdc9457SAndroid Build Coastguard Worker       vacc1x0 = vacc1x0 * vscale0 + vbias0;
75*4bdc9457SAndroid Build Coastguard Worker       vacc1x1 = vacc1x1 * vscale1 + vbias1;
76*4bdc9457SAndroid Build Coastguard Worker       vacc1x2 = vacc1x2 * vscale2 + vbias2;
77*4bdc9457SAndroid Build Coastguard Worker       vacc1x3 = vacc1x3 * vscale3 + vbias3;
78*4bdc9457SAndroid Build Coastguard Worker 
79*4bdc9457SAndroid Build Coastguard Worker       vacc0x0 = math_max_f32(vacc0x0, vmin);
80*4bdc9457SAndroid Build Coastguard Worker       vacc0x1 = math_max_f32(vacc0x1, vmin);
81*4bdc9457SAndroid Build Coastguard Worker       vacc0x2 = math_max_f32(vacc0x2, vmin);
82*4bdc9457SAndroid Build Coastguard Worker       vacc0x3 = math_max_f32(vacc0x3, vmin);
83*4bdc9457SAndroid Build Coastguard Worker       vacc1x0 = math_max_f32(vacc1x0, vmin);
84*4bdc9457SAndroid Build Coastguard Worker       vacc1x1 = math_max_f32(vacc1x1, vmin);
85*4bdc9457SAndroid Build Coastguard Worker       vacc1x2 = math_max_f32(vacc1x2, vmin);
86*4bdc9457SAndroid Build Coastguard Worker       vacc1x3 = math_max_f32(vacc1x3, vmin);
87*4bdc9457SAndroid Build Coastguard Worker 
88*4bdc9457SAndroid Build Coastguard Worker       vacc0x0 = math_min_f32(vacc0x0, vmax);
89*4bdc9457SAndroid Build Coastguard Worker       vacc0x1 = math_min_f32(vacc0x1, vmax);
90*4bdc9457SAndroid Build Coastguard Worker       vacc0x2 = math_min_f32(vacc0x2, vmax);
91*4bdc9457SAndroid Build Coastguard Worker       vacc0x3 = math_min_f32(vacc0x3, vmax);
92*4bdc9457SAndroid Build Coastguard Worker       vacc1x0 = math_min_f32(vacc1x0, vmax);
93*4bdc9457SAndroid Build Coastguard Worker       vacc1x1 = math_min_f32(vacc1x1, vmax);
94*4bdc9457SAndroid Build Coastguard Worker       vacc1x2 = math_min_f32(vacc1x2, vmax);
95*4bdc9457SAndroid Build Coastguard Worker       vacc1x3 = math_min_f32(vacc1x3, vmax);
96*4bdc9457SAndroid Build Coastguard Worker 
97*4bdc9457SAndroid Build Coastguard Worker       o0[0] = vacc0x0;
98*4bdc9457SAndroid Build Coastguard Worker       o0[1] = vacc0x1;
99*4bdc9457SAndroid Build Coastguard Worker       o0[2] = vacc0x2;
100*4bdc9457SAndroid Build Coastguard Worker       o0[3] = vacc0x3;
101*4bdc9457SAndroid Build Coastguard Worker       o0 += 4;
102*4bdc9457SAndroid Build Coastguard Worker       o1[0] = vacc1x0;
103*4bdc9457SAndroid Build Coastguard Worker       o1[1] = vacc1x1;
104*4bdc9457SAndroid Build Coastguard Worker       o1[2] = vacc1x2;
105*4bdc9457SAndroid Build Coastguard Worker       o1[3] = vacc1x3;
106*4bdc9457SAndroid Build Coastguard Worker       o1 += 4;
107*4bdc9457SAndroid Build Coastguard Worker 
108*4bdc9457SAndroid Build Coastguard Worker       w += 8;
109*4bdc9457SAndroid Build Coastguard Worker     }
110*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNLIKELY(c != 0) {
111*4bdc9457SAndroid Build Coastguard Worker       do {
112*4bdc9457SAndroid Build Coastguard Worker         const float vscale = *w++;
113*4bdc9457SAndroid Build Coastguard Worker 
114*4bdc9457SAndroid Build Coastguard Worker         float vacc0 = *i0++;
115*4bdc9457SAndroid Build Coastguard Worker         float vacc1 = *i1++;
116*4bdc9457SAndroid Build Coastguard Worker 
117*4bdc9457SAndroid Build Coastguard Worker         const float vbias = w[3];
118*4bdc9457SAndroid Build Coastguard Worker 
119*4bdc9457SAndroid Build Coastguard Worker         vacc0 = vacc0 * vscale + vbias;
120*4bdc9457SAndroid Build Coastguard Worker         vacc1 = vacc1 * vscale + vbias;
121*4bdc9457SAndroid Build Coastguard Worker 
122*4bdc9457SAndroid Build Coastguard Worker         vacc0 = math_max_f32(vacc0, vmin);
123*4bdc9457SAndroid Build Coastguard Worker         vacc1 = math_max_f32(vacc1, vmin);
124*4bdc9457SAndroid Build Coastguard Worker 
125*4bdc9457SAndroid Build Coastguard Worker         vacc0 = math_min_f32(vacc0, vmax);
126*4bdc9457SAndroid Build Coastguard Worker         vacc1 = math_min_f32(vacc1, vmax);
127*4bdc9457SAndroid Build Coastguard Worker 
128*4bdc9457SAndroid Build Coastguard Worker         *o0++ = vacc0;
129*4bdc9457SAndroid Build Coastguard Worker         *o1++ = vacc1;
130*4bdc9457SAndroid Build Coastguard Worker 
131*4bdc9457SAndroid Build Coastguard Worker         c -= sizeof(float);
132*4bdc9457SAndroid Build Coastguard Worker       } while (c != 0);
133*4bdc9457SAndroid Build Coastguard Worker     }
134*4bdc9457SAndroid Build Coastguard Worker     i0 = (const float*) ((uintptr_t) i0 + input_increment);
135*4bdc9457SAndroid Build Coastguard Worker     o0 = (float*) ((uintptr_t) o0 + output_increment);
136*4bdc9457SAndroid Build Coastguard Worker     i1 = (const float*) ((uintptr_t) i1 + input_increment);
137*4bdc9457SAndroid Build Coastguard Worker     o1 = (float*) ((uintptr_t) o1 + output_increment);
138*4bdc9457SAndroid Build Coastguard Worker     rows = doz(rows, 2);
139*4bdc9457SAndroid Build Coastguard Worker   } while (rows != 0);
140*4bdc9457SAndroid Build Coastguard Worker }
141