1 // Auto-generated file. Do not edit!
2 // Template: src/f32-igemm/sse-shuffle.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/igemm.h>
15
16
xnn_f32_igemm_minmax_ukernel_1x8s4__sse(size_t mr,size_t nc,size_t kc,size_t ks,const float ** restrict a,const float * restrict w,float * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const float * zero,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_igemm_minmax_ukernel_1x8s4__sse(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 size_t ks,
22 const float**restrict a,
23 const float*restrict w,
24 float*restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 size_t a_offset,
28 const float* zero,
29 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31 assert(mr != 0);
32 assert(mr <= 1);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(kc % sizeof(float) == 0);
36 assert(ks != 0);
37 assert(ks % (1 * sizeof(void*)) == 0);
38 assert(a_offset % sizeof(float) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 float* c0 = c;
44
45 do {
46 __m128 vacc0x0123 = _mm_load_ps(w);
47 __m128 vacc0x4567 = _mm_load_ps(w + 4);
48 w += 8;
49
50 size_t p = ks;
51 do {
52 const float* restrict a0 = a[0];
53 assert(a0 != NULL);
54 if XNN_UNPREDICTABLE(a0 != zero) {
55 a0 = (const float*) ((uintptr_t) a0 + a_offset);
56 }
57 a += 1;
58
59 size_t k = kc;
60 while (k >= 4 * sizeof(float)) {
61 __m128 va0 = _mm_loadu_ps(a0);
62 a0 += 4;
63
64
65 const __m128 vb0123c0 = _mm_load_ps(w + 0);
66 const __m128 vb4567c0 = _mm_load_ps(w + 4);
67
68 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0, vb0123c0));
69 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0, vb4567c0));
70
71 va0 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(0, 3, 2, 1));
72
73 const __m128 vb0123c1 = _mm_load_ps(w + 8);
74 const __m128 vb4567c1 = _mm_load_ps(w + 12);
75
76 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0, vb0123c1));
77 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0, vb4567c1));
78
79 va0 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(0, 3, 2, 1));
80
81 const __m128 vb0123c2 = _mm_load_ps(w + 16);
82 const __m128 vb4567c2 = _mm_load_ps(w + 20);
83
84 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0, vb0123c2));
85 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0, vb4567c2));
86
87 va0 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(0, 3, 2, 1));
88
89 const __m128 vb0123c3 = _mm_load_ps(w + 24);
90 const __m128 vb4567c3 = _mm_load_ps(w + 28);
91
92 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0, vb0123c3));
93 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0, vb4567c3));
94
95
96 w += 32;
97 k -= 4 * sizeof(float);
98 }
99 if XNN_UNLIKELY(k != 0) {
100 __m128 va0 = _mm_loadu_ps(a0);
101 a0 = (const float*) ((uintptr_t) a0 + k);
102
103
104 const __m128 vb0123c0 = _mm_load_ps(w + 0);
105 const __m128 vb4567c0 = _mm_load_ps(w + 4);
106
107 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c0), va0), vb0123c0));
108 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c0), va0), vb4567c0));
109
110 va0 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(0, 3, 2, 1));
111
112 const __m128 vb0123c1 = _mm_load_ps(w + 8);
113 const __m128 vb4567c1 = _mm_load_ps(w + 12);
114
115 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c1), va0), vb0123c1));
116 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c1), va0), vb4567c1));
117
118 va0 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(0, 3, 2, 1));
119
120 const __m128 vb0123c2 = _mm_load_ps(w + 16);
121 const __m128 vb4567c2 = _mm_load_ps(w + 20);
122
123 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c2), va0), vb0123c2));
124 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c2), va0), vb4567c2));
125
126 va0 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(0, 3, 2, 1));
127
128 const __m128 vb0123c3 = _mm_load_ps(w + 24);
129 const __m128 vb4567c3 = _mm_load_ps(w + 28);
130
131 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c3), va0), vb0123c3));
132 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c3), va0), vb4567c3));
133
134
135 w += 32;
136 }
137 p -= 1 * sizeof(void*);
138 } while (p != 0);
139
140 const __m128 vmax = _mm_load_ps(params->sse.max);
141 vacc0x0123 = _mm_min_ps(vacc0x0123, vmax);
142 vacc0x4567 = _mm_min_ps(vacc0x4567, vmax);
143
144 const __m128 vmin = _mm_load_ps(params->sse.min);
145 vacc0x0123 = _mm_max_ps(vacc0x0123, vmin);
146 vacc0x4567 = _mm_max_ps(vacc0x4567, vmin);
147
148 if XNN_LIKELY(nc >= 8) {
149 _mm_storeu_ps(c0, vacc0x0123);
150 _mm_storeu_ps(c0 + 4, vacc0x4567);
151 c0 = (float*) ((uintptr_t) c0 + cn_stride);
152
153 a = (const float**restrict) ((uintptr_t) a - ks);
154 nc -= 8;
155 } else {
156 if (nc & 4) {
157 _mm_storeu_ps(c0, vacc0x0123);
158
159 vacc0x0123 = vacc0x4567;
160
161 c0 += 4;
162 }
163 if (nc & 2) {
164 _mm_storel_pi((__m64*) c0, vacc0x0123);
165
166 vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123);
167
168 c0 += 2;
169 }
170 if (nc & 1) {
171 _mm_store_ss(c0, vacc0x0123);
172 }
173
174 nc = 0;
175 }
176 } while (nc != 0);
177 }
178