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_gemm_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 union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_gemm_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 union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
28 {
29 assert(mr != 0);
30 assert(mr <= 3);
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 const float* a1 = (const float*) ((uintptr_t) a0 + a_stride);
41 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
42 if XNN_UNPREDICTABLE(mr < 2) {
43 a1 = a0;
44 c1 = c0;
45 }
46 const float* a2 = (const float*) ((uintptr_t) a1 + a_stride);
47 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
48 if XNN_UNPREDICTABLE(mr <= 2) {
49 a2 = a1;
50 c2 = c1;
51 }
52
53 do {
54 __m128 vacc0x0123 = _mm_load_ps(w + 0);
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 k = kc;
63 do {
64 const __m128 va0 = _mm_load1_ps(a0);
65 a0 += 1;
66 const __m128 va1 = _mm_load1_ps(a1);
67 a1 += 1;
68 const __m128 va2 = _mm_load1_ps(a2);
69 a2 += 1;
70
71 const __m128 vb0123 = _mm_load_ps(w);
72 const __m128 vb4567 = _mm_load_ps(w + 4);
73 w += 8;
74
75 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0, vb0123));
76 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1, vb0123));
77 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2, vb0123));
78 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0, vb4567));
79 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1, vb4567));
80 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2, vb4567));
81
82 k -= sizeof(float);
83 } while (k != 0);
84
85 const __m128 vmax = _mm_load_ps(params->sse.max);
86 vacc0x0123 = _mm_min_ps(vacc0x0123, vmax);
87 vacc1x0123 = _mm_min_ps(vacc1x0123, vmax);
88 vacc2x0123 = _mm_min_ps(vacc2x0123, vmax);
89 vacc0x4567 = _mm_min_ps(vacc0x4567, vmax);
90 vacc1x4567 = _mm_min_ps(vacc1x4567, vmax);
91 vacc2x4567 = _mm_min_ps(vacc2x4567, vmax);
92
93 const __m128 vmin = _mm_load_ps(params->sse.min);
94 vacc0x0123 = _mm_max_ps(vacc0x0123, vmin);
95 vacc1x0123 = _mm_max_ps(vacc1x0123, vmin);
96 vacc2x0123 = _mm_max_ps(vacc2x0123, vmin);
97 vacc0x4567 = _mm_max_ps(vacc0x4567, vmin);
98 vacc1x4567 = _mm_max_ps(vacc1x4567, vmin);
99 vacc2x4567 = _mm_max_ps(vacc2x4567, vmin);
100
101 if XNN_LIKELY(nc >= 8) {
102 _mm_storeu_ps(c2, vacc2x0123);
103 _mm_storeu_ps(c2 + 4, vacc2x4567);
104 c2 = (float*) ((uintptr_t) c2 + cn_stride);
105 _mm_storeu_ps(c1, vacc1x0123);
106 _mm_storeu_ps(c1 + 4, vacc1x4567);
107 c1 = (float*) ((uintptr_t) c1 + cn_stride);
108 _mm_storeu_ps(c0, vacc0x0123);
109 _mm_storeu_ps(c0 + 4, vacc0x4567);
110 c0 = (float*) ((uintptr_t) c0 + cn_stride);
111
112 a2 = (const float*) ((uintptr_t) a2 - kc);
113 a1 = (const float*) ((uintptr_t) a1 - kc);
114 a0 = (const float*) ((uintptr_t) a0 - kc);
115
116 nc -= 8;
117 } else {
118 if (nc & 4) {
119 _mm_storeu_ps(c2, vacc2x0123);
120 _mm_storeu_ps(c1, vacc1x0123);
121 _mm_storeu_ps(c0, vacc0x0123);
122
123 vacc2x0123 = vacc2x4567;
124 vacc1x0123 = vacc1x4567;
125 vacc0x0123 = vacc0x4567;
126
127 c2 += 4;
128 c1 += 4;
129 c0 += 4;
130 }
131 if (nc & 2) {
132 _mm_storel_pi((__m64*) c2, vacc2x0123);
133 _mm_storel_pi((__m64*) c1, vacc1x0123);
134 _mm_storel_pi((__m64*) c0, vacc0x0123);
135
136 vacc2x0123 = _mm_movehl_ps(vacc2x0123, vacc2x0123);
137 vacc1x0123 = _mm_movehl_ps(vacc1x0123, vacc1x0123);
138 vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123);
139
140 c2 += 2;
141 c1 += 2;
142 c0 += 2;
143 }
144 if (nc & 1) {
145 _mm_store_ss(c2, vacc2x0123);
146 _mm_store_ss(c1, vacc1x0123);
147 _mm_store_ss(c0, vacc0x0123);
148 }
149
150 nc = 0;
151 }
152 } while (nc != 0);
153 }
154