1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-gemm/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 <emmintrin.h>
13 
14 #include <xnnpack/gemm.h>
15 
16 
xnn_f32_gemminc_minmax_ukernel_1x8__sse2_dup(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 float * restrict acc,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_gemminc_minmax_ukernel_1x8__sse2_dup(
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 float*restrict acc,
28     const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
29 {
30   assert(mr != 0);
31   assert(mr <= 1);
32   assert(nc != 0);
33   assert(kc != 0);
34   assert(kc % sizeof(float) == 0);
35   assert(a != NULL);
36   assert(w != NULL);
37   assert(c != NULL);
38   assert(acc != NULL);
39 
40   const float* a0 = a;
41   float* c0 = c;
42 
43   do {
44     __m128 vacc0x0123 = _mm_load_ps(acc + 0);
45     __m128 vacc0x4567 = _mm_load_ps(acc + 4);
46     acc += 8;
47 
48     size_t k = kc;
49     while (k >= 4 * sizeof(float)) {
50       const __m128 va0 = _mm_loadu_ps(a0);
51       a0 += 4;
52 
53 
54       const __m128 va0c0000 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(va0), _MM_SHUFFLE(0, 0, 0, 0)));
55 
56       const __m128 vb0123c0 = _mm_load_ps(w + 0);
57       const __m128 vb4567c0 = _mm_load_ps(w + 4);
58 
59       vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0c0000, vb0123c0));
60       vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0c0000, vb4567c0));
61 
62       const __m128 va0c1111 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(va0), _MM_SHUFFLE(1, 1, 1, 1)));
63 
64       const __m128 vb0123c1 = _mm_load_ps(w + 8);
65       const __m128 vb4567c1 = _mm_load_ps(w + 12);
66 
67       vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0c1111, vb0123c1));
68       vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0c1111, vb4567c1));
69 
70       const __m128 va0c2222 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(va0), _MM_SHUFFLE(2, 2, 2, 2)));
71 
72       const __m128 vb0123c2 = _mm_load_ps(w + 16);
73       const __m128 vb4567c2 = _mm_load_ps(w + 20);
74 
75       vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0c2222, vb0123c2));
76       vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0c2222, vb4567c2));
77 
78       const __m128 va0c3333 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(3, 3, 3, 3));
79 
80       const __m128 vb0123c3 = _mm_load_ps(w + 24);
81       const __m128 vb4567c3 = _mm_load_ps(w + 28);
82 
83       vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0c3333, vb0123c3));
84       vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0c3333, vb4567c3));
85 
86       w += 32;
87       k -= 4 * sizeof(float);
88     }
89     if XNN_UNLIKELY(k != 0) {
90       do {
91         const __m128 va0 = _mm_load1_ps(a0);
92         a0 += 1;
93 
94         const __m128 vb0123 = _mm_load_ps(w);
95         const __m128 vb4567 = _mm_load_ps(w + 4);
96         w += 8;
97 
98         vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0, vb0123));
99         vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0, vb4567));
100 
101         k -= sizeof(float);
102       } while (k != 0);
103     }
104 
105     const __m128 vmax = _mm_load_ps(params->sse.max);
106     vacc0x0123 = _mm_min_ps(vacc0x0123, vmax);
107     vacc0x4567 = _mm_min_ps(vacc0x4567, vmax);
108 
109     const __m128 vmin = _mm_load_ps(params->sse.min);
110     vacc0x0123 = _mm_max_ps(vacc0x0123, vmin);
111     vacc0x4567 = _mm_max_ps(vacc0x4567, vmin);
112 
113     if XNN_LIKELY(nc >= 8) {
114       _mm_storeu_ps(c0, vacc0x0123);
115       _mm_storeu_ps(c0 + 4, vacc0x4567);
116       c0 = (float*) ((uintptr_t) c0 + cn_stride);
117 
118       a0 = (const float*) ((uintptr_t) a0 - kc);
119 
120       nc -= 8;
121     } else {
122       if (nc & 4) {
123         _mm_storeu_ps(c0, vacc0x0123);
124 
125         vacc0x0123 = vacc0x4567;
126 
127         c0 += 4;
128       }
129       if (nc & 2) {
130         _mm_storel_pi((__m64*) c0, vacc0x0123);
131 
132         vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123);
133 
134         c0 += 2;
135       }
136       if (nc & 1) {
137         _mm_store_ss(c0, vacc0x0123);
138       }
139 
140       nc = 0;
141     }
142   } while (nc != 0);
143 }
144