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_c8__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_c8__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 >= 8 * sizeof(float); c -= 8 * sizeof(float)) {
51 const __m128 vscale0123 = _mm_load_ps(w);
52 const __m128 vscale4567 = _mm_load_ps(w + 4);
53
54 __m128 vacc0x0123 = _mm_loadu_ps(i0);
55 __m128 vacc0x4567 = _mm_loadu_ps(i0 + 4);
56 i0 += 8;
57 __m128 vacc1x0123 = _mm_loadu_ps(i1);
58 __m128 vacc1x4567 = _mm_loadu_ps(i1 + 4);
59 i1 += 8;
60
61 vacc0x0123 = _mm_mul_ps(vacc0x0123, vscale0123);
62 vacc0x4567 = _mm_mul_ps(vacc0x4567, vscale4567);
63 vacc1x0123 = _mm_mul_ps(vacc1x0123, vscale0123);
64 vacc1x4567 = _mm_mul_ps(vacc1x4567, vscale4567);
65
66 const __m128 vbias0123 = _mm_load_ps(w + 8);
67 const __m128 vbias4567 = _mm_load_ps(w + 12);
68
69 vacc0x0123 = _mm_add_ps(vacc0x0123, vbias0123);
70 vacc0x4567 = _mm_add_ps(vacc0x4567, vbias4567);
71 vacc1x0123 = _mm_add_ps(vacc1x0123, vbias0123);
72 vacc1x4567 = _mm_add_ps(vacc1x4567, vbias4567);
73
74 vacc0x0123 = _mm_max_ps(vacc0x0123, vmin);
75 vacc0x4567 = _mm_max_ps(vacc0x4567, vmin);
76 vacc1x0123 = _mm_max_ps(vacc1x0123, vmin);
77 vacc1x4567 = _mm_max_ps(vacc1x4567, vmin);
78
79 vacc0x0123 = _mm_min_ps(vacc0x0123, vmax);
80 vacc0x4567 = _mm_min_ps(vacc0x4567, vmax);
81 vacc1x0123 = _mm_min_ps(vacc1x0123, vmax);
82 vacc1x4567 = _mm_min_ps(vacc1x4567, vmax);
83
84 _mm_storeu_ps(o0, vacc0x0123);
85 _mm_storeu_ps(o0 + 4, vacc0x4567);
86 o0 += 8;
87 _mm_storeu_ps(o1, vacc1x0123);
88 _mm_storeu_ps(o1 + 4, vacc1x4567);
89 o1 += 8;
90
91 w += 16;
92 }
93 for (; c >= 4 * sizeof(float); c -= 4 * sizeof(float)) {
94 const __m128 vscale0123 = _mm_load_ps(w);
95
96 __m128 vacc0x0123 = _mm_loadu_ps(i0);
97 i0 += 4;
98 __m128 vacc1x0123 = _mm_loadu_ps(i1);
99 i1 += 4;
100
101 vacc0x0123 = _mm_mul_ps(vacc0x0123, vscale0123);
102 vacc1x0123 = _mm_mul_ps(vacc1x0123, vscale0123);
103
104 const __m128 vbias0123 = _mm_load_ps(w + 8);
105
106 vacc0x0123 = _mm_add_ps(vacc0x0123, vbias0123);
107 vacc1x0123 = _mm_add_ps(vacc1x0123, vbias0123);
108
109 vacc0x0123 = _mm_max_ps(vacc0x0123, vmin);
110 vacc1x0123 = _mm_max_ps(vacc1x0123, vmin);
111
112 vacc0x0123 = _mm_min_ps(vacc0x0123, vmax);
113 vacc1x0123 = _mm_min_ps(vacc1x0123, vmax);
114
115 _mm_storeu_ps(o0, vacc0x0123);
116 o0 += 4;
117 _mm_storeu_ps(o1, vacc1x0123);
118 o1 += 4;
119
120 w += 4;
121 }
122 if XNN_UNLIKELY(c != 0) {
123 const __m128 vscale0123 = _mm_load_ps(w);
124
125 __m128 vacc0x0123 = _mm_loadu_ps(i0);
126 i0 = (const float*) ((uintptr_t) i0 + c);
127 __m128 vacc1x0123 = _mm_loadu_ps(i1);
128 i1 = (const float*) ((uintptr_t) i1 + c);
129
130 vacc0x0123 = _mm_mul_ps(vacc0x0123, vscale0123);
131 vacc1x0123 = _mm_mul_ps(vacc1x0123, vscale0123);
132
133 const __m128 vbias0123 = _mm_load_ps(w + 8);
134
135 vacc0x0123 = _mm_add_ps(vacc0x0123, vbias0123);
136 vacc1x0123 = _mm_add_ps(vacc1x0123, vbias0123);
137
138 vacc0x0123 = _mm_max_ps(vacc0x0123, vmin);
139 vacc1x0123 = _mm_max_ps(vacc1x0123, vmin);
140
141 vacc0x0123 = _mm_min_ps(vacc0x0123, vmax);
142 vacc1x0123 = _mm_min_ps(vacc1x0123, vmax);
143
144 if (c & (2 * sizeof(float))) {
145 _mm_storel_pi((__m64*) o0, vacc0x0123);
146 _mm_storel_pi((__m64*) o1, vacc1x0123);
147
148 vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123);
149 vacc1x0123 = _mm_movehl_ps(vacc1x0123, vacc1x0123);
150
151 o0 += 2;
152 o1 += 2;
153 }
154 if (c & (1 * sizeof(float))) {
155 _mm_store_ss(o0, vacc0x0123);
156 _mm_store_ss(o1, vacc1x0123);
157
158 o0 += 1;
159 o1 += 1;
160 }
161 }
162 i0 = (const float*) ((uintptr_t) i0 + input_increment);
163 o0 = (float*) ((uintptr_t) o0 + output_increment);
164 i1 = (const float*) ((uintptr_t) i1 + input_increment);
165 o1 = (float*) ((uintptr_t) o1 + output_increment);
166 rows = doz(rows, 2);
167 } while (rows != 0);
168 }
169