1 // Auto-generated file. Do not edit!
2 // Template: src/f32-igemm/MRx2c4-sse.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_ukernel_4x2c4__sse(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_output_params params[restrict static1])17 void xnn_f32_igemm_ukernel_4x2c4__sse(
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_output_params params[restrict static 1])
30 {
31 assert(mr != 0);
32 assert(mr <= 4);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(kc % sizeof(float) == 0);
36 assert(ks != 0);
37 assert(ks % (4 * 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 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
53 if XNN_UNPREDICTABLE(mr != 4) {
54 c3 = c2;
55 }
56
57 do {
58 __m128 vacc0x0c4 = _mm_load_ss(w);
59 __m128 vacc0x1c4 = _mm_load_ss(w + 1);
60 __m128 vacc1x0c4 = vacc0x0c4;
61 __m128 vacc1x1c4 = vacc0x1c4;
62 __m128 vacc2x0c4 = vacc0x0c4;
63 __m128 vacc2x1c4 = vacc0x1c4;
64 __m128 vacc3x0c4 = vacc0x0c4;
65 __m128 vacc3x1c4 = vacc0x1c4;
66 w += 2;
67
68 size_t p = ks;
69 do {
70 const float* restrict a0 = a[0];
71 assert(a0 != NULL);
72 if XNN_UNPREDICTABLE(a0 != zero) {
73 a0 = (const float*) ((uintptr_t) a0 + a_offset);
74 }
75 const float* restrict a1 = a[1];
76 assert(a1 != NULL);
77 if XNN_UNPREDICTABLE(a1 != zero) {
78 a1 = (const float*) ((uintptr_t) a1 + a_offset);
79 }
80 const float* restrict a2 = a[2];
81 assert(a2 != NULL);
82 if XNN_UNPREDICTABLE(a2 != zero) {
83 a2 = (const float*) ((uintptr_t) a2 + a_offset);
84 }
85 const float* restrict a3 = a[3];
86 assert(a3 != NULL);
87 if XNN_UNPREDICTABLE(a3 != zero) {
88 a3 = (const float*) ((uintptr_t) a3 + a_offset);
89 }
90 a += 4;
91
92 size_t k = kc;
93 for (; k >= 4 * sizeof(float); k -= 4 * sizeof(float)) {
94 const __m128 va0 = _mm_loadu_ps(a0);
95 a0 += 4;
96 const __m128 va1 = _mm_loadu_ps(a1);
97 a1 += 4;
98 const __m128 va2 = _mm_loadu_ps(a2);
99 a2 += 4;
100 const __m128 va3 = _mm_loadu_ps(a3);
101 a3 += 4;
102
103 const __m128 vb0 = _mm_loadu_ps(w);
104 const __m128 vb1 = _mm_loadu_ps(w + 4);
105 w += 8;
106
107 vacc0x0c4 = _mm_add_ps(vacc0x0c4, _mm_mul_ps(va0, vb0));
108 vacc0x1c4 = _mm_add_ps(vacc0x1c4, _mm_mul_ps(va0, vb1));
109 vacc1x0c4 = _mm_add_ps(vacc1x0c4, _mm_mul_ps(va1, vb0));
110 vacc1x1c4 = _mm_add_ps(vacc1x1c4, _mm_mul_ps(va1, vb1));
111 vacc2x0c4 = _mm_add_ps(vacc2x0c4, _mm_mul_ps(va2, vb0));
112 vacc2x1c4 = _mm_add_ps(vacc2x1c4, _mm_mul_ps(va2, vb1));
113 vacc3x0c4 = _mm_add_ps(vacc3x0c4, _mm_mul_ps(va3, vb0));
114 vacc3x1c4 = _mm_add_ps(vacc3x1c4, _mm_mul_ps(va3, vb1));
115 }
116 if XNN_UNLIKELY(k != 0) {
117 const __m128 va0 = _mm_loadu_ps(a0);
118 const __m128 va1 = _mm_loadu_ps(a1);
119 const __m128 va2 = _mm_loadu_ps(a2);
120 const __m128 va3 = _mm_loadu_ps(a3);
121
122 const __m128 vb0 = _mm_loadu_ps(w);
123 const __m128 vb1 = _mm_loadu_ps(w + 4);
124 w += 8;
125
126 const __m128 vmask0 = _mm_cmpeq_ps(_mm_setzero_ps(), vb0);
127 const __m128 vmask1 = _mm_cmpeq_ps(_mm_setzero_ps(), vb1);
128
129 vacc0x0c4 = _mm_add_ps(vacc0x0c4, _mm_mul_ps(_mm_andnot_ps(vmask0, va0), vb0));
130 vacc0x1c4 = _mm_add_ps(vacc0x1c4, _mm_mul_ps(_mm_andnot_ps(vmask1, va0), vb1));
131 vacc1x0c4 = _mm_add_ps(vacc1x0c4, _mm_mul_ps(_mm_andnot_ps(vmask0, va1), vb0));
132 vacc1x1c4 = _mm_add_ps(vacc1x1c4, _mm_mul_ps(_mm_andnot_ps(vmask1, va1), vb1));
133 vacc2x0c4 = _mm_add_ps(vacc2x0c4, _mm_mul_ps(_mm_andnot_ps(vmask0, va2), vb0));
134 vacc2x1c4 = _mm_add_ps(vacc2x1c4, _mm_mul_ps(_mm_andnot_ps(vmask1, va2), vb1));
135 vacc3x0c4 = _mm_add_ps(vacc3x0c4, _mm_mul_ps(_mm_andnot_ps(vmask0, va3), vb0));
136 vacc3x1c4 = _mm_add_ps(vacc3x1c4, _mm_mul_ps(_mm_andnot_ps(vmask1, va3), vb1));
137 }
138 p -= 4 * sizeof(void*);
139 } while (p != 0);
140
141 const __m128 vacc0x01c2 = _mm_add_ps(_mm_unpacklo_ps(vacc0x0c4, vacc0x1c4), _mm_unpackhi_ps(vacc0x0c4, vacc0x1c4));
142 const __m128 vacc1x01c2 = _mm_add_ps(_mm_unpacklo_ps(vacc1x0c4, vacc1x1c4), _mm_unpackhi_ps(vacc1x0c4, vacc1x1c4));
143 const __m128 vacc2x01c2 = _mm_add_ps(_mm_unpacklo_ps(vacc2x0c4, vacc2x1c4), _mm_unpackhi_ps(vacc2x0c4, vacc2x1c4));
144 const __m128 vacc3x01c2 = _mm_add_ps(_mm_unpacklo_ps(vacc3x0c4, vacc3x1c4), _mm_unpackhi_ps(vacc3x0c4, vacc3x1c4));
145
146 __m128 vacc01x01 = _mm_add_ps(_mm_movelh_ps(vacc0x01c2, vacc1x01c2), _mm_movehl_ps(vacc1x01c2, vacc0x01c2));
147 __m128 vacc23x01 = _mm_add_ps(_mm_movelh_ps(vacc2x01c2, vacc3x01c2), _mm_movehl_ps(vacc3x01c2, vacc2x01c2));
148
149 const __m128 vmax = _mm_load_ps(params->sse.max);
150 vacc01x01 = _mm_min_ps(vacc01x01, vmax);
151 vacc23x01 = _mm_min_ps(vacc23x01, vmax);
152
153 const __m128 vmin = _mm_load_ps(params->sse.min);
154 vacc01x01 = _mm_max_ps(vacc01x01, vmin);
155 vacc23x01 = _mm_max_ps(vacc23x01, vmin);
156
157 if XNN_LIKELY(nc >= 2) {
158 _mm_storeh_pi((__m64*) c3, vacc23x01);
159 c3 = (float*) ((uintptr_t) c3 + cn_stride);
160 _mm_storel_pi((__m64*) c2, vacc23x01);
161 c2 = (float*) ((uintptr_t) c2 + cn_stride);
162 _mm_storeh_pi((__m64*) c1, vacc01x01);
163 c1 = (float*) ((uintptr_t) c1 + cn_stride);
164 _mm_storel_pi((__m64*) c0, vacc01x01);
165 c0 = (float*) ((uintptr_t) c0 + cn_stride);
166
167 a = (const float**restrict) ((uintptr_t) a - ks);
168 nc -= 2;
169 } else {
170 assert(nc == 1);
171 _mm_store_ss(c3, _mm_movehl_ps(vacc23x01, vacc23x01));
172 _mm_store_ss(c2, vacc23x01);
173 _mm_store_ss(c1, _mm_movehl_ps(vacc01x01, vacc01x01));
174 _mm_store_ss(c0, vacc01x01);
175
176 nc = 0;
177 }
178 } while (nc != 0);
179 }
180