1 // Auto-generated file. Do not edit!
2 // Template: src/f32-gemm/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/gemm.h>
15
16
xnn_f32_gemm_minmax_ukernel_4x2c4__sse(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_4x2c4__sse(
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)]) XNN_DISABLE_TSAN
28 {
29 assert(mr != 0);
30 assert(mr <= 4);
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 const float* a3 = (const float*) ((uintptr_t) a2 + a_stride);
53 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
54 if XNN_UNPREDICTABLE(mr != 4) {
55 a3 = a2;
56 c3 = c2;
57 }
58
59 do {
60 __m128 vacc0x0c4 = _mm_load_ss(w);
61 __m128 vacc0x1c4 = _mm_load_ss(w + 1);
62 __m128 vacc1x0c4 = vacc0x0c4;
63 __m128 vacc1x1c4 = vacc0x1c4;
64 __m128 vacc2x0c4 = vacc0x0c4;
65 __m128 vacc2x1c4 = vacc0x1c4;
66 __m128 vacc3x0c4 = vacc0x0c4;
67 __m128 vacc3x1c4 = vacc0x1c4;
68 w += 2;
69
70 size_t k = kc;
71 for (; k >= 4 * sizeof(float); k -= 4 * sizeof(float)) {
72 const __m128 va0 = _mm_loadu_ps(a0);
73 a0 += 4;
74 const __m128 va1 = _mm_loadu_ps(a1);
75 a1 += 4;
76 const __m128 va2 = _mm_loadu_ps(a2);
77 a2 += 4;
78 const __m128 va3 = _mm_loadu_ps(a3);
79 a3 += 4;
80
81 const __m128 vb0 = _mm_loadu_ps(w);
82 const __m128 vb1 = _mm_loadu_ps(w + 4);
83 w += 8;
84
85 vacc0x0c4 = _mm_add_ps(vacc0x0c4, _mm_mul_ps(va0, vb0));
86 vacc0x1c4 = _mm_add_ps(vacc0x1c4, _mm_mul_ps(va0, vb1));
87 vacc1x0c4 = _mm_add_ps(vacc1x0c4, _mm_mul_ps(va1, vb0));
88 vacc1x1c4 = _mm_add_ps(vacc1x1c4, _mm_mul_ps(va1, vb1));
89 vacc2x0c4 = _mm_add_ps(vacc2x0c4, _mm_mul_ps(va2, vb0));
90 vacc2x1c4 = _mm_add_ps(vacc2x1c4, _mm_mul_ps(va2, vb1));
91 vacc3x0c4 = _mm_add_ps(vacc3x0c4, _mm_mul_ps(va3, vb0));
92 vacc3x1c4 = _mm_add_ps(vacc3x1c4, _mm_mul_ps(va3, vb1));
93 }
94 if XNN_UNLIKELY(k != 0) {
95 const __m128 va0 = _mm_loadu_ps(a0);
96 a0 = (const float*) ((uintptr_t) a0 + k);
97 const __m128 va1 = _mm_loadu_ps(a1);
98 a1 = (const float*) ((uintptr_t) a1 + k);
99 const __m128 va2 = _mm_loadu_ps(a2);
100 a2 = (const float*) ((uintptr_t) a2 + k);
101 const __m128 va3 = _mm_loadu_ps(a3);
102 a3 = (const float*) ((uintptr_t) a3 + k);
103
104 const __m128 vb0 = _mm_loadu_ps(w);
105 const __m128 vb1 = _mm_loadu_ps(w + 4);
106 w += 8;
107
108 const __m128 vmask0 = _mm_cmpeq_ps(_mm_setzero_ps(), vb0);
109 const __m128 vmask1 = _mm_cmpeq_ps(_mm_setzero_ps(), vb1);
110
111 vacc0x0c4 = _mm_add_ps(vacc0x0c4, _mm_mul_ps(_mm_andnot_ps(vmask0, va0), vb0));
112 vacc0x1c4 = _mm_add_ps(vacc0x1c4, _mm_mul_ps(_mm_andnot_ps(vmask1, va0), vb1));
113 vacc1x0c4 = _mm_add_ps(vacc1x0c4, _mm_mul_ps(_mm_andnot_ps(vmask0, va1), vb0));
114 vacc1x1c4 = _mm_add_ps(vacc1x1c4, _mm_mul_ps(_mm_andnot_ps(vmask1, va1), vb1));
115 vacc2x0c4 = _mm_add_ps(vacc2x0c4, _mm_mul_ps(_mm_andnot_ps(vmask0, va2), vb0));
116 vacc2x1c4 = _mm_add_ps(vacc2x1c4, _mm_mul_ps(_mm_andnot_ps(vmask1, va2), vb1));
117 vacc3x0c4 = _mm_add_ps(vacc3x0c4, _mm_mul_ps(_mm_andnot_ps(vmask0, va3), vb0));
118 vacc3x1c4 = _mm_add_ps(vacc3x1c4, _mm_mul_ps(_mm_andnot_ps(vmask1, va3), vb1));
119 }
120
121 const __m128 vacc0x01c2 = _mm_add_ps(_mm_unpacklo_ps(vacc0x0c4, vacc0x1c4), _mm_unpackhi_ps(vacc0x0c4, vacc0x1c4));
122 const __m128 vacc1x01c2 = _mm_add_ps(_mm_unpacklo_ps(vacc1x0c4, vacc1x1c4), _mm_unpackhi_ps(vacc1x0c4, vacc1x1c4));
123 const __m128 vacc2x01c2 = _mm_add_ps(_mm_unpacklo_ps(vacc2x0c4, vacc2x1c4), _mm_unpackhi_ps(vacc2x0c4, vacc2x1c4));
124 const __m128 vacc3x01c2 = _mm_add_ps(_mm_unpacklo_ps(vacc3x0c4, vacc3x1c4), _mm_unpackhi_ps(vacc3x0c4, vacc3x1c4));
125
126 __m128 vacc01x01 = _mm_add_ps(_mm_movelh_ps(vacc0x01c2, vacc1x01c2), _mm_movehl_ps(vacc1x01c2, vacc0x01c2));
127 __m128 vacc23x01 = _mm_add_ps(_mm_movelh_ps(vacc2x01c2, vacc3x01c2), _mm_movehl_ps(vacc3x01c2, vacc2x01c2));
128
129 const __m128 vmax = _mm_load_ps(params->sse.max);
130 vacc01x01 = _mm_min_ps(vacc01x01, vmax);
131 vacc23x01 = _mm_min_ps(vacc23x01, vmax);
132
133 const __m128 vmin = _mm_load_ps(params->sse.min);
134 vacc01x01 = _mm_max_ps(vacc01x01, vmin);
135 vacc23x01 = _mm_max_ps(vacc23x01, vmin);
136
137 if XNN_LIKELY(nc >= 2) {
138 _mm_storel_pi((__m64*) c2, vacc23x01);
139 c2 = (float*) ((uintptr_t) c2 + cn_stride);
140 a2 = (const float*) ((uintptr_t) a2 - kc);
141 _mm_storeh_pi((__m64*) c3, vacc23x01);
142 c3 = (float*) ((uintptr_t) c3 + cn_stride);
143 a3 = (const float*) ((uintptr_t) a3 - kc);
144 _mm_storel_pi((__m64*) c0, vacc01x01);
145 c0 = (float*) ((uintptr_t) c0 + cn_stride);
146 a0 = (const float*) ((uintptr_t) a0 - kc);
147 _mm_storeh_pi((__m64*) c1, vacc01x01);
148 c1 = (float*) ((uintptr_t) c1 + cn_stride);
149 a1 = (const float*) ((uintptr_t) a1 - kc);
150
151 nc -= 2;
152 } else {
153 assert(nc == 1);
154 _mm_store_ss(c2, vacc23x01);
155 _mm_store_ss(c3, _mm_movehl_ps(vacc23x01, vacc23x01));
156 _mm_store_ss(c0, vacc01x01);
157 _mm_store_ss(c1, _mm_movehl_ps(vacc01x01, vacc01x01));
158
159 nc = 0;
160 }
161 } while (nc != 0);
162 }
163