1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-igemm/sse-load1.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_3x8__sse_load1(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_3x8__sse_load1(
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 <= 3);
33   assert(nc != 0);
34   assert(kc != 0);
35   assert(kc % sizeof(float) == 0);
36   assert(ks != 0);
37   assert(ks % (3 * 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   float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
45   if XNN_UNPREDICTABLE(mr < 2) {
46     c1 = c0;
47   }
48   float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
49   if XNN_UNPREDICTABLE(mr <= 2) {
50     c2 = c1;
51   }
52 
53   do {
54     __m128 vacc0x0123 = _mm_load_ps(w);
55     __m128 vacc0x4567 = _mm_load_ps(w + 4);
56     __m128 vacc1x0123 = vacc0x0123;
57     __m128 vacc1x4567 = vacc0x4567;
58     __m128 vacc2x0123 = vacc0x0123;
59     __m128 vacc2x4567 = vacc0x4567;
60     w += 8;
61 
62     size_t p = ks;
63     do {
64       const float* restrict a0 = a[0];
65       assert(a0 != NULL);
66       if XNN_UNPREDICTABLE(a0 != zero) {
67         a0 = (const float*) ((uintptr_t) a0 + a_offset);
68       }
69       const float* restrict a1 = a[1];
70       assert(a1 != NULL);
71       if XNN_UNPREDICTABLE(a1 != zero) {
72         a1 = (const float*) ((uintptr_t) a1 + a_offset);
73       }
74       const float* restrict a2 = a[2];
75       assert(a2 != NULL);
76       if XNN_UNPREDICTABLE(a2 != zero) {
77         a2 = (const float*) ((uintptr_t) a2 + a_offset);
78       }
79       a += 3;
80 
81       size_t k = kc;
82       do {
83         const __m128 vb0123 = _mm_load_ps(w);
84         const __m128 vb4567 = _mm_load_ps(w + 4);
85         w += 8;
86 
87         const __m128 va0 = _mm_load1_ps(a0);
88         a0 += 1;
89         const __m128 va1 = _mm_load1_ps(a1);
90         a1 += 1;
91         const __m128 va2 = _mm_load1_ps(a2);
92         a2 += 1;
93 
94         vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0, vb0123));
95         vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0, vb4567));
96         vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1, vb0123));
97         vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1, vb4567));
98         vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2, vb0123));
99         vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2, vb4567));
100         k -= sizeof(float);
101       } while (k != 0);
102       p -= 3 * sizeof(void*);
103     } while (p != 0);
104 
105     const __m128 vmax = _mm_load_ps(params->sse.max);
106     vacc0x0123 = _mm_min_ps(vacc0x0123, vmax);
107     vacc1x0123 = _mm_min_ps(vacc1x0123, vmax);
108     vacc2x0123 = _mm_min_ps(vacc2x0123, vmax);
109     vacc0x4567 = _mm_min_ps(vacc0x4567, vmax);
110     vacc1x4567 = _mm_min_ps(vacc1x4567, vmax);
111     vacc2x4567 = _mm_min_ps(vacc2x4567, vmax);
112 
113     const __m128 vmin = _mm_load_ps(params->sse.min);
114     vacc0x0123 = _mm_max_ps(vacc0x0123, vmin);
115     vacc1x0123 = _mm_max_ps(vacc1x0123, vmin);
116     vacc2x0123 = _mm_max_ps(vacc2x0123, vmin);
117     vacc0x4567 = _mm_max_ps(vacc0x4567, vmin);
118     vacc1x4567 = _mm_max_ps(vacc1x4567, vmin);
119     vacc2x4567 = _mm_max_ps(vacc2x4567, vmin);
120 
121     if XNN_LIKELY(nc >= 8) {
122       _mm_storeu_ps(c2, vacc2x0123);
123       _mm_storeu_ps(c2 + 4, vacc2x4567);
124       c2 = (float*) ((uintptr_t) c2 + cn_stride);
125       _mm_storeu_ps(c1, vacc1x0123);
126       _mm_storeu_ps(c1 + 4, vacc1x4567);
127       c1 = (float*) ((uintptr_t) c1 + cn_stride);
128       _mm_storeu_ps(c0, vacc0x0123);
129       _mm_storeu_ps(c0 + 4, vacc0x4567);
130       c0 = (float*) ((uintptr_t) c0 + cn_stride);
131 
132       a = (const float**restrict) ((uintptr_t) a - ks);
133       nc -= 8;
134     } else {
135       if (nc & 4) {
136         _mm_storeu_ps(c2, vacc2x0123);
137         _mm_storeu_ps(c1, vacc1x0123);
138         _mm_storeu_ps(c0, vacc0x0123);
139 
140         vacc2x0123 = vacc2x4567;
141         vacc1x0123 = vacc1x4567;
142         vacc0x0123 = vacc0x4567;
143 
144         c2 += 4;
145         c1 += 4;
146         c0 += 4;
147       }
148       if (nc & 2) {
149         _mm_storel_pi((__m64*) c2, vacc2x0123);
150         _mm_storel_pi((__m64*) c1, vacc1x0123);
151         _mm_storel_pi((__m64*) c0, vacc0x0123);
152 
153         vacc2x0123 = _mm_movehl_ps(vacc2x0123, vacc2x0123);
154         vacc1x0123 = _mm_movehl_ps(vacc1x0123, vacc1x0123);
155         vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123);
156 
157         c2 += 2;
158         c1 += 2;
159         c0 += 2;
160       }
161       if (nc & 1) {
162         _mm_store_ss(c2, vacc2x0123);
163         _mm_store_ss(c1, vacc1x0123);
164         _mm_store_ss(c0, vacc0x0123);
165       }
166 
167       nc = 0;
168     }
169   } while (nc != 0);
170 }
171