xref: /aosp_15_r20/external/XNNPACK/src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-vmulcaddc/sse.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2019 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 <xmmintrin.h>
13 
14 #include <xnnpack/math.h>
15 #include <xnnpack/vmulcaddc.h>
16 
17 
xnn_f32_vmulcaddc_minmax_ukernel_c4__sse_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)])18 void xnn_f32_vmulcaddc_minmax_ukernel_c4__sse_2x(
19     size_t rows,
20     size_t channels,
21     const float*restrict input,
22     size_t input_stride,
23     const float*restrict weights,
24     float*restrict output,
25     size_t output_stride,
26     const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
27 {
28   assert(rows != 0);
29   assert(channels != 0);
30   assert(channels % sizeof(float) == 0);
31 
32   const float* i0 = input;
33   float* o0 = output;
34   const float* i1 = (const float*) ((uintptr_t) i0 + input_stride);
35   float* o1 = (float*) ((uintptr_t) o0 + output_stride);
36 
37   const size_t input_increment = input_stride * 2 - channels;
38   const size_t output_increment = output_stride * 2 - channels;
39 
40   const __m128 vmin = _mm_load_ps(params->sse.min);
41   const __m128 vmax = _mm_load_ps(params->sse.max);
42   do {
43     if XNN_UNPREDICTABLE(rows < 2) {
44       i1 = i0;
45       o1 = o0;
46     }
47 
48     const float* w = weights;
49     size_t c = channels;
50     for (; c >= 4 * sizeof(float); c -= 4 * sizeof(float)) {
51       const __m128 vscale0123 = _mm_load_ps(w);
52 
53       __m128 vacc0x0123 = _mm_loadu_ps(i0);
54       i0 += 4;
55       __m128 vacc1x0123 = _mm_loadu_ps(i1);
56       i1 += 4;
57 
58       vacc0x0123 = _mm_mul_ps(vacc0x0123, vscale0123);
59       vacc1x0123 = _mm_mul_ps(vacc1x0123, vscale0123);
60 
61       const __m128 vbias0123 = _mm_load_ps(w + 4);
62 
63       vacc0x0123 = _mm_add_ps(vacc0x0123, vbias0123);
64       vacc1x0123 = _mm_add_ps(vacc1x0123, vbias0123);
65 
66       vacc0x0123 = _mm_max_ps(vacc0x0123, vmin);
67       vacc1x0123 = _mm_max_ps(vacc1x0123, vmin);
68 
69       vacc0x0123 = _mm_min_ps(vacc0x0123, vmax);
70       vacc1x0123 = _mm_min_ps(vacc1x0123, vmax);
71 
72       _mm_storeu_ps(o0, vacc0x0123);
73       o0 += 4;
74       _mm_storeu_ps(o1, vacc1x0123);
75       o1 += 4;
76 
77       w += 8;
78     }
79     if XNN_UNLIKELY(c != 0) {
80       const __m128 vscale0123 = _mm_load_ps(w);
81 
82       __m128 vacc0x0123 = _mm_loadu_ps(i0);
83       i0 = (const float*) ((uintptr_t) i0 + c);
84       __m128 vacc1x0123 = _mm_loadu_ps(i1);
85       i1 = (const float*) ((uintptr_t) i1 + c);
86 
87       vacc0x0123 = _mm_mul_ps(vacc0x0123, vscale0123);
88       vacc1x0123 = _mm_mul_ps(vacc1x0123, vscale0123);
89 
90       const __m128 vbias0123 = _mm_load_ps(w + 4);
91 
92       vacc0x0123 = _mm_add_ps(vacc0x0123, vbias0123);
93       vacc1x0123 = _mm_add_ps(vacc1x0123, vbias0123);
94 
95       vacc0x0123 = _mm_max_ps(vacc0x0123, vmin);
96       vacc1x0123 = _mm_max_ps(vacc1x0123, vmin);
97 
98       vacc0x0123 = _mm_min_ps(vacc0x0123, vmax);
99       vacc1x0123 = _mm_min_ps(vacc1x0123, vmax);
100 
101       if (c & (2 * sizeof(float))) {
102         _mm_storel_pi((__m64*) o0, vacc0x0123);
103         _mm_storel_pi((__m64*) o1, vacc1x0123);
104 
105         vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123);
106         vacc1x0123 = _mm_movehl_ps(vacc1x0123, vacc1x0123);
107 
108         o0 += 2;
109         o1 += 2;
110       }
111       if (c & (1 * sizeof(float))) {
112         _mm_store_ss(o0, vacc0x0123);
113         _mm_store_ss(o1, vacc1x0123);
114 
115         o0 += 1;
116         o1 += 1;
117       }
118     }
119     i0 = (const float*) ((uintptr_t) i0 + input_increment);
120     o0 = (float*) ((uintptr_t) o0 + output_increment);
121     i1 = (const float*) ((uintptr_t) i1 + input_increment);
122     o1 = (float*) ((uintptr_t) o1 + output_increment);
123     rows = doz(rows, 2);
124   } while (rows != 0);
125 }
126