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_4x8__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_4x8__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 <= 4);
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 const float* a3 = (const float*) ((uintptr_t) a2 + a_stride);
55 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
56 if XNN_UNPREDICTABLE(mr != 4) {
57 a3 = a2;
58 c3 = c2;
59 }
60
61 do {
62 __m128 vacc0x0123 = _mm_load_ps(acc + 0);
63 __m128 vacc0x4567 = _mm_load_ps(acc + 4);
64 __m128 vacc1x0123 = _mm_load_ps(acc + 8);
65 __m128 vacc1x4567 = _mm_load_ps(acc + 12);
66 __m128 vacc2x0123 = _mm_load_ps(acc + 16);
67 __m128 vacc2x4567 = _mm_load_ps(acc + 20);
68 __m128 vacc3x0123 = _mm_load_ps(acc + 24);
69 __m128 vacc3x4567 = _mm_load_ps(acc + 28);
70 acc += 32;
71
72 size_t k = kc;
73 do {
74 const __m128 va0 = _mm_load1_ps(a0);
75 a0 += 1;
76 const __m128 va1 = _mm_load1_ps(a1);
77 a1 += 1;
78 const __m128 va2 = _mm_load1_ps(a2);
79 a2 += 1;
80 const __m128 va3 = _mm_load1_ps(a3);
81 a3 += 1;
82
83 const __m128 vb0123 = _mm_load_ps(w);
84 const __m128 vb4567 = _mm_load_ps(w + 4);
85 w += 8;
86
87 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0, vb0123));
88 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1, vb0123));
89 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2, vb0123));
90 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3, vb0123));
91 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0, vb4567));
92 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1, vb4567));
93 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2, vb4567));
94 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3, vb4567));
95
96 k -= sizeof(float);
97 } while (k != 0);
98
99 const __m128 vmax = _mm_load_ps(params->sse.max);
100 vacc0x0123 = _mm_min_ps(vacc0x0123, vmax);
101 vacc1x0123 = _mm_min_ps(vacc1x0123, vmax);
102 vacc2x0123 = _mm_min_ps(vacc2x0123, vmax);
103 vacc3x0123 = _mm_min_ps(vacc3x0123, vmax);
104 vacc0x4567 = _mm_min_ps(vacc0x4567, vmax);
105 vacc1x4567 = _mm_min_ps(vacc1x4567, vmax);
106 vacc2x4567 = _mm_min_ps(vacc2x4567, vmax);
107 vacc3x4567 = _mm_min_ps(vacc3x4567, vmax);
108
109 const __m128 vmin = _mm_load_ps(params->sse.min);
110 vacc0x0123 = _mm_max_ps(vacc0x0123, vmin);
111 vacc1x0123 = _mm_max_ps(vacc1x0123, vmin);
112 vacc2x0123 = _mm_max_ps(vacc2x0123, vmin);
113 vacc3x0123 = _mm_max_ps(vacc3x0123, vmin);
114 vacc0x4567 = _mm_max_ps(vacc0x4567, vmin);
115 vacc1x4567 = _mm_max_ps(vacc1x4567, vmin);
116 vacc2x4567 = _mm_max_ps(vacc2x4567, vmin);
117 vacc3x4567 = _mm_max_ps(vacc3x4567, vmin);
118
119 if XNN_LIKELY(nc >= 8) {
120 _mm_storeu_ps(c3, vacc3x0123);
121 _mm_storeu_ps(c3 + 4, vacc3x4567);
122 c3 = (float*) ((uintptr_t) c3 + cn_stride);
123 _mm_storeu_ps(c2, vacc2x0123);
124 _mm_storeu_ps(c2 + 4, vacc2x4567);
125 c2 = (float*) ((uintptr_t) c2 + cn_stride);
126 _mm_storeu_ps(c1, vacc1x0123);
127 _mm_storeu_ps(c1 + 4, vacc1x4567);
128 c1 = (float*) ((uintptr_t) c1 + cn_stride);
129 _mm_storeu_ps(c0, vacc0x0123);
130 _mm_storeu_ps(c0 + 4, vacc0x4567);
131 c0 = (float*) ((uintptr_t) c0 + cn_stride);
132
133 a3 = (const float*) ((uintptr_t) a3 - kc);
134 a2 = (const float*) ((uintptr_t) a2 - kc);
135 a1 = (const float*) ((uintptr_t) a1 - kc);
136 a0 = (const float*) ((uintptr_t) a0 - kc);
137
138 nc -= 8;
139 } else {
140 if (nc & 4) {
141 _mm_storeu_ps(c3, vacc3x0123);
142 _mm_storeu_ps(c2, vacc2x0123);
143 _mm_storeu_ps(c1, vacc1x0123);
144 _mm_storeu_ps(c0, vacc0x0123);
145
146 vacc3x0123 = vacc3x4567;
147 vacc2x0123 = vacc2x4567;
148 vacc1x0123 = vacc1x4567;
149 vacc0x0123 = vacc0x4567;
150
151 c3 += 4;
152 c2 += 4;
153 c1 += 4;
154 c0 += 4;
155 }
156 if (nc & 2) {
157 _mm_storel_pi((__m64*) c3, vacc3x0123);
158 _mm_storel_pi((__m64*) c2, vacc2x0123);
159 _mm_storel_pi((__m64*) c1, vacc1x0123);
160 _mm_storel_pi((__m64*) c0, vacc0x0123);
161
162 vacc3x0123 = _mm_movehl_ps(vacc3x0123, vacc3x0123);
163 vacc2x0123 = _mm_movehl_ps(vacc2x0123, vacc2x0123);
164 vacc1x0123 = _mm_movehl_ps(vacc1x0123, vacc1x0123);
165 vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123);
166
167 c3 += 2;
168 c2 += 2;
169 c1 += 2;
170 c0 += 2;
171 }
172 if (nc & 1) {
173 _mm_store_ss(c3, vacc3x0123);
174 _mm_store_ss(c2, vacc2x0123);
175 _mm_store_ss(c1, vacc1x0123);
176 _mm_store_ss(c0, vacc0x0123);
177 }
178
179 nc = 0;
180 }
181 } while (nc != 0);
182 }
183