1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-igemm/sse-dup.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_1x8__sse_dup(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_1x8__sse_dup(
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)])
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         const __m128 va0 = _mm_loadu_ps(a0);
62         a0 += 4;
63 
64 
65         const __m128 va0c0000 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(0, 0, 0, 0));
66 
67         const __m128 vb0123c0 = _mm_load_ps(w + 0);
68         const __m128 vb4567c0 = _mm_load_ps(w + 4);
69 
70         vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0c0000, vb0123c0));
71         vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0c0000, vb4567c0));
72 
73         const __m128 va0c1111 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(1, 1, 1, 1));
74 
75         const __m128 vb0123c1 = _mm_load_ps(w + 8);
76         const __m128 vb4567c1 = _mm_load_ps(w + 12);
77 
78         vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0c1111, vb0123c1));
79         vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0c1111, vb4567c1));
80 
81         const __m128 va0c2222 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(2, 2, 2, 2));
82 
83         const __m128 vb0123c2 = _mm_load_ps(w + 16);
84         const __m128 vb4567c2 = _mm_load_ps(w + 20);
85 
86         vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0c2222, vb0123c2));
87         vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0c2222, vb4567c2));
88 
89         const __m128 va0c3333 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(3, 3, 3, 3));
90 
91         const __m128 vb0123c3 = _mm_load_ps(w + 24);
92         const __m128 vb4567c3 = _mm_load_ps(w + 28);
93 
94         vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0c3333, vb0123c3));
95         vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0c3333, vb4567c3));
96 
97         w += 32;
98         k -= 4 * sizeof(float);
99       }
100       if XNN_UNLIKELY(k != 0) {
101         do {
102           const __m128 vb0123 = _mm_load_ps(w);
103           const __m128 vb4567 = _mm_load_ps(w + 4);
104           w += 8;
105 
106           const __m128 va0 = _mm_load1_ps(a0);
107           a0 += 1;
108 
109           vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0, vb0123));
110           vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0, vb4567));
111           k -= sizeof(float);
112         } while (k != 0);
113       }
114       p -= 1 * sizeof(void*);
115     } while (p != 0);
116 
117     const __m128 vmax = _mm_load_ps(params->sse.max);
118     vacc0x0123 = _mm_min_ps(vacc0x0123, vmax);
119     vacc0x4567 = _mm_min_ps(vacc0x4567, vmax);
120 
121     const __m128 vmin = _mm_load_ps(params->sse.min);
122     vacc0x0123 = _mm_max_ps(vacc0x0123, vmin);
123     vacc0x4567 = _mm_max_ps(vacc0x4567, vmin);
124 
125     if XNN_LIKELY(nc >= 8) {
126       _mm_storeu_ps(c0, vacc0x0123);
127       _mm_storeu_ps(c0 + 4, vacc0x4567);
128       c0 = (float*) ((uintptr_t) c0 + cn_stride);
129 
130       a = (const float**restrict) ((uintptr_t) a - ks);
131       nc -= 8;
132     } else {
133       if (nc & 4) {
134         _mm_storeu_ps(c0, vacc0x0123);
135 
136         vacc0x0123 = vacc0x4567;
137 
138         c0 += 4;
139       }
140       if (nc & 2) {
141         _mm_storel_pi((__m64*) c0, vacc0x0123);
142 
143         vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123);
144 
145         c0 += 2;
146       }
147       if (nc & 1) {
148         _mm_store_ss(c0, vacc0x0123);
149       }
150 
151       nc = 0;
152     }
153   } while (nc != 0);
154 }
155