1 // Auto-generated file. Do not edit!
2 // Template: src/f32-gemm/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/gemm.h>
15
16
xnn_f32_gemm_minmax_ukernel_1x8s4__sse(size_t mr,size_t nc,size_t kc,const float * restrict a,size_t a_stride,const float * restrict w,float * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_gemm_minmax_ukernel_1x8s4__sse(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 const float*restrict a,
22 size_t a_stride,
23 const float*restrict w,
24 float*restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28 {
29 assert(mr != 0);
30 assert(mr <= 1);
31 assert(nc != 0);
32 assert(kc != 0);
33 assert(kc % sizeof(float) == 0);
34 assert(a != NULL);
35 assert(w != NULL);
36 assert(c != NULL);
37
38 const float* a0 = a;
39 float* c0 = c;
40
41 do {
42 __m128 vacc0x0123 = _mm_load_ps(w + 0);
43 __m128 vacc0x4567 = _mm_load_ps(w + 4);
44 w += 8;
45
46 size_t k = kc;
47 while (k >= 4 * sizeof(float)) {
48 __m128 va0 = _mm_loadu_ps(a0);
49 a0 += 4;
50
51
52 const __m128 vb0123c0 = _mm_load_ps(w + 0);
53 const __m128 vb4567c0 = _mm_load_ps(w + 4);
54
55 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0, vb0123c0));
56 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0, vb4567c0));
57
58 va0 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(0, 3, 2, 1));
59
60 const __m128 vb0123c1 = _mm_load_ps(w + 8);
61 const __m128 vb4567c1 = _mm_load_ps(w + 12);
62
63 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0, vb0123c1));
64 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0, vb4567c1));
65
66 va0 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(0, 3, 2, 1));
67
68 const __m128 vb0123c2 = _mm_load_ps(w + 16);
69 const __m128 vb4567c2 = _mm_load_ps(w + 20);
70
71 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0, vb0123c2));
72 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0, vb4567c2));
73
74 va0 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(0, 3, 2, 1));
75
76 const __m128 vb0123c3 = _mm_load_ps(w + 24);
77 const __m128 vb4567c3 = _mm_load_ps(w + 28);
78
79 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0, vb0123c3));
80 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0, vb4567c3));
81
82
83 w += 32;
84 k -= 4 * sizeof(float);
85 }
86 if XNN_UNLIKELY(k != 0) {
87 __m128 va0 = _mm_loadu_ps(a0);
88 a0 = (const float*) ((uintptr_t) a0 + k);
89
90
91 const __m128 vb0123c0 = _mm_load_ps(w + 0);
92 const __m128 vb4567c0 = _mm_load_ps(w + 4);
93
94 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c0), va0), vb0123c0));
95 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c0), va0), vb4567c0));
96
97 va0 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(0, 3, 2, 1));
98
99 const __m128 vb0123c1 = _mm_load_ps(w + 8);
100 const __m128 vb4567c1 = _mm_load_ps(w + 12);
101
102 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c1), va0), vb0123c1));
103 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c1), va0), vb4567c1));
104
105 va0 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(0, 3, 2, 1));
106
107 const __m128 vb0123c2 = _mm_load_ps(w + 16);
108 const __m128 vb4567c2 = _mm_load_ps(w + 20);
109
110 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c2), va0), vb0123c2));
111 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c2), va0), vb4567c2));
112
113 va0 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(0, 3, 2, 1));
114
115 const __m128 vb0123c3 = _mm_load_ps(w + 24);
116 const __m128 vb4567c3 = _mm_load_ps(w + 28);
117
118 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c3), va0), vb0123c3));
119 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c3), va0), vb4567c3));
120
121
122 w += 32;
123 }
124
125 const __m128 vmax = _mm_load_ps(params->sse.max);
126 vacc0x0123 = _mm_min_ps(vacc0x0123, vmax);
127 vacc0x4567 = _mm_min_ps(vacc0x4567, vmax);
128
129 const __m128 vmin = _mm_load_ps(params->sse.min);
130 vacc0x0123 = _mm_max_ps(vacc0x0123, vmin);
131 vacc0x4567 = _mm_max_ps(vacc0x4567, vmin);
132
133 if XNN_LIKELY(nc >= 8) {
134 _mm_storeu_ps(c0, vacc0x0123);
135 _mm_storeu_ps(c0 + 4, vacc0x4567);
136 c0 = (float*) ((uintptr_t) c0 + cn_stride);
137
138 a0 = (const float*) ((uintptr_t) a0 - kc);
139
140 nc -= 8;
141 } else {
142 if (nc & 4) {
143 _mm_storeu_ps(c0, vacc0x0123);
144
145 vacc0x0123 = vacc0x4567;
146
147 c0 += 4;
148 }
149 if (nc & 2) {
150 _mm_storel_pi((__m64*) c0, vacc0x0123);
151
152 vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123);
153
154 c0 += 2;
155 }
156 if (nc & 1) {
157 _mm_store_ss(c0, vacc0x0123);
158 }
159
160 nc = 0;
161 }
162 } while (nc != 0);
163 }
164