1 // Auto-generated file. Do not edit!
2 // Template: src/f32-gemm/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/gemm.h>
15
16
xnn_f32_gemminc_minmax_ukernel_3x8__sse_load1(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_3x8__sse_load1(
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 <= 3);
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 const float* a1 = (const float*) ((uintptr_t) a0 + a_stride);
43 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
44 if XNN_UNPREDICTABLE(mr < 2) {
45 a1 = a0;
46 c1 = c0;
47 }
48 const float* a2 = (const float*) ((uintptr_t) a1 + a_stride);
49 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
50 if XNN_UNPREDICTABLE(mr <= 2) {
51 a2 = a1;
52 c2 = c1;
53 }
54
55 do {
56 __m128 vacc0x0123 = _mm_load_ps(acc + 0);
57 __m128 vacc0x4567 = _mm_load_ps(acc + 4);
58 __m128 vacc1x0123 = _mm_load_ps(acc + 8);
59 __m128 vacc1x4567 = _mm_load_ps(acc + 12);
60 __m128 vacc2x0123 = _mm_load_ps(acc + 16);
61 __m128 vacc2x4567 = _mm_load_ps(acc + 20);
62 acc += 24;
63
64 size_t k = kc;
65 do {
66 const __m128 va0 = _mm_load1_ps(a0);
67 a0 += 1;
68 const __m128 va1 = _mm_load1_ps(a1);
69 a1 += 1;
70 const __m128 va2 = _mm_load1_ps(a2);
71 a2 += 1;
72
73 const __m128 vb0123 = _mm_load_ps(w);
74 const __m128 vb4567 = _mm_load_ps(w + 4);
75 w += 8;
76
77 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0, vb0123));
78 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1, vb0123));
79 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2, vb0123));
80 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0, vb4567));
81 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1, vb4567));
82 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2, vb4567));
83
84 k -= sizeof(float);
85 } while (k != 0);
86
87 const __m128 vmax = _mm_load_ps(params->sse.max);
88 vacc0x0123 = _mm_min_ps(vacc0x0123, vmax);
89 vacc1x0123 = _mm_min_ps(vacc1x0123, vmax);
90 vacc2x0123 = _mm_min_ps(vacc2x0123, vmax);
91 vacc0x4567 = _mm_min_ps(vacc0x4567, vmax);
92 vacc1x4567 = _mm_min_ps(vacc1x4567, vmax);
93 vacc2x4567 = _mm_min_ps(vacc2x4567, vmax);
94
95 const __m128 vmin = _mm_load_ps(params->sse.min);
96 vacc0x0123 = _mm_max_ps(vacc0x0123, vmin);
97 vacc1x0123 = _mm_max_ps(vacc1x0123, vmin);
98 vacc2x0123 = _mm_max_ps(vacc2x0123, vmin);
99 vacc0x4567 = _mm_max_ps(vacc0x4567, vmin);
100 vacc1x4567 = _mm_max_ps(vacc1x4567, vmin);
101 vacc2x4567 = _mm_max_ps(vacc2x4567, vmin);
102
103 if XNN_LIKELY(nc >= 8) {
104 _mm_storeu_ps(c2, vacc2x0123);
105 _mm_storeu_ps(c2 + 4, vacc2x4567);
106 c2 = (float*) ((uintptr_t) c2 + cn_stride);
107 _mm_storeu_ps(c1, vacc1x0123);
108 _mm_storeu_ps(c1 + 4, vacc1x4567);
109 c1 = (float*) ((uintptr_t) c1 + cn_stride);
110 _mm_storeu_ps(c0, vacc0x0123);
111 _mm_storeu_ps(c0 + 4, vacc0x4567);
112 c0 = (float*) ((uintptr_t) c0 + cn_stride);
113
114 a2 = (const float*) ((uintptr_t) a2 - kc);
115 a1 = (const float*) ((uintptr_t) a1 - kc);
116 a0 = (const float*) ((uintptr_t) a0 - kc);
117
118 nc -= 8;
119 } else {
120 if (nc & 4) {
121 _mm_storeu_ps(c2, vacc2x0123);
122 _mm_storeu_ps(c1, vacc1x0123);
123 _mm_storeu_ps(c0, vacc0x0123);
124
125 vacc2x0123 = vacc2x4567;
126 vacc1x0123 = vacc1x4567;
127 vacc0x0123 = vacc0x4567;
128
129 c2 += 4;
130 c1 += 4;
131 c0 += 4;
132 }
133 if (nc & 2) {
134 _mm_storel_pi((__m64*) c2, vacc2x0123);
135 _mm_storel_pi((__m64*) c1, vacc1x0123);
136 _mm_storel_pi((__m64*) c0, vacc0x0123);
137
138 vacc2x0123 = _mm_movehl_ps(vacc2x0123, vacc2x0123);
139 vacc1x0123 = _mm_movehl_ps(vacc1x0123, vacc1x0123);
140 vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123);
141
142 c2 += 2;
143 c1 += 2;
144 c0 += 2;
145 }
146 if (nc & 1) {
147 _mm_store_ss(c2, vacc2x0123);
148 _mm_store_ss(c1, vacc1x0123);
149 _mm_store_ss(c0, vacc0x0123);
150 }
151
152 nc = 0;
153 }
154 } while (nc != 0);
155 }
156