1 // Auto-generated file. Do not edit!
2 // Template: src/f32-igemm/MRx2c4-psimd.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 <psimd.h>
13
14 #include <xnnpack/igemm.h>
15
16
xnn_f32_igemm_ukernel_4x2c4__psimd(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__psimd(
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 psimd_f32 vacc0x0c4 = psimd_load1_f32(w);
59 psimd_f32 vacc0x1c4 = psimd_load1_f32(w + 1);
60 psimd_f32 vacc1x0c4 = vacc0x0c4;
61 psimd_f32 vacc1x1c4 = vacc0x1c4;
62 psimd_f32 vacc2x0c4 = vacc0x0c4;
63 psimd_f32 vacc2x1c4 = vacc0x1c4;
64 psimd_f32 vacc3x0c4 = vacc0x0c4;
65 psimd_f32 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 psimd_f32 va0 = psimd_load_f32(a0);
95 a0 += 4;
96 const psimd_f32 va1 = psimd_load_f32(a1);
97 a1 += 4;
98 const psimd_f32 va2 = psimd_load_f32(a2);
99 a2 += 4;
100 const psimd_f32 va3 = psimd_load_f32(a3);
101 a3 += 4;
102
103 const psimd_f32 vb0 = psimd_load_f32(w);
104 const psimd_f32 vb1 = psimd_load_f32(w + 4);
105 w += 8;
106
107 vacc0x0c4 = psimd_qfma_f32(vacc0x0c4, va0, vb0);
108 vacc0x1c4 = psimd_qfma_f32(vacc0x1c4, va0, vb1);
109 vacc1x0c4 = psimd_qfma_f32(vacc1x0c4, va1, vb0);
110 vacc1x1c4 = psimd_qfma_f32(vacc1x1c4, va1, vb1);
111 vacc2x0c4 = psimd_qfma_f32(vacc2x0c4, va2, vb0);
112 vacc2x1c4 = psimd_qfma_f32(vacc2x1c4, va2, vb1);
113 vacc3x0c4 = psimd_qfma_f32(vacc3x0c4, va3, vb0);
114 vacc3x1c4 = psimd_qfma_f32(vacc3x1c4, va3, vb1);
115 }
116 if XNN_UNLIKELY(k != 0) {
117 const psimd_f32 va0 = psimd_load_f32(a0);
118 const psimd_f32 va1 = psimd_load_f32(a1);
119 const psimd_f32 va2 = psimd_load_f32(a2);
120 const psimd_f32 va3 = psimd_load_f32(a3);
121
122 const psimd_f32 vb0 = psimd_load_f32(w);
123 const psimd_f32 vb1 = psimd_load_f32(w + 4);
124 w += 8;
125
126 const psimd_f32 vzero = psimd_splat_f32(0.0f);
127 const psimd_s32 vmask0 = vb0 != vzero;
128 const psimd_s32 vmask1 = vb1 != vzero;
129
130 vacc0x0c4 = psimd_qfma_f32(vacc0x0c4, psimd_andmask_f32(vmask0, va0), vb0);
131 vacc0x1c4 = psimd_qfma_f32(vacc0x1c4, psimd_andmask_f32(vmask1, va0), vb1);
132 vacc1x0c4 = psimd_qfma_f32(vacc1x0c4, psimd_andmask_f32(vmask0, va1), vb0);
133 vacc1x1c4 = psimd_qfma_f32(vacc1x1c4, psimd_andmask_f32(vmask1, va1), vb1);
134 vacc2x0c4 = psimd_qfma_f32(vacc2x0c4, psimd_andmask_f32(vmask0, va2), vb0);
135 vacc2x1c4 = psimd_qfma_f32(vacc2x1c4, psimd_andmask_f32(vmask1, va2), vb1);
136 vacc3x0c4 = psimd_qfma_f32(vacc3x0c4, psimd_andmask_f32(vmask0, va3), vb0);
137 vacc3x1c4 = psimd_qfma_f32(vacc3x1c4, psimd_andmask_f32(vmask1, va3), vb1);
138 }
139 p -= 4 * sizeof(void*);
140 } while (p != 0);
141
142 const psimd_f32 vacc0x01c2 = psimd_add_f32(psimd_interleave_lo_f32(vacc0x0c4, vacc0x1c4), psimd_interleave_hi_f32(vacc0x0c4, vacc0x1c4));
143 const psimd_f32 vacc1x01c2 = psimd_add_f32(psimd_interleave_lo_f32(vacc1x0c4, vacc1x1c4), psimd_interleave_hi_f32(vacc1x0c4, vacc1x1c4));
144 const psimd_f32 vacc2x01c2 = psimd_add_f32(psimd_interleave_lo_f32(vacc2x0c4, vacc2x1c4), psimd_interleave_hi_f32(vacc2x0c4, vacc2x1c4));
145 const psimd_f32 vacc3x01c2 = psimd_add_f32(psimd_interleave_lo_f32(vacc3x0c4, vacc3x1c4), psimd_interleave_hi_f32(vacc3x0c4, vacc3x1c4));
146
147 psimd_f32 vacc01x01 = psimd_add_f32(psimd_concat_lo_f32(vacc0x01c2, vacc1x01c2), psimd_concat_hi_f32(vacc0x01c2, vacc1x01c2));
148 psimd_f32 vacc23x01 = psimd_add_f32(psimd_concat_lo_f32(vacc2x01c2, vacc3x01c2), psimd_concat_hi_f32(vacc2x01c2, vacc3x01c2));
149
150 const psimd_f32 vmax = psimd_load_splat_f32(¶ms->scalar.max);
151 vacc01x01 = psimd_min_f32(vacc01x01, vmax);
152 vacc23x01 = psimd_min_f32(vacc23x01, vmax);
153
154 const psimd_f32 vmin = psimd_load_splat_f32(¶ms->scalar.min);
155 vacc01x01 = psimd_max_f32(vacc01x01, vmin);
156 vacc23x01 = psimd_max_f32(vacc23x01, vmin);
157
158 if XNN_LIKELY(nc >= 2) {
159 psimd_store2_f32(c3, psimd_concat_hi_f32(vacc23x01, vacc23x01));
160 c3 = (float*) ((uintptr_t) c3 + cn_stride);
161 psimd_store2_f32(c2, vacc23x01);
162 c2 = (float*) ((uintptr_t) c2 + cn_stride);
163 psimd_store2_f32(c1, psimd_concat_hi_f32(vacc01x01, vacc01x01));
164 c1 = (float*) ((uintptr_t) c1 + cn_stride);
165 psimd_store2_f32(c0, vacc01x01);
166 c0 = (float*) ((uintptr_t) c0 + cn_stride);
167
168 a = (const float**restrict) ((uintptr_t) a - ks);
169 nc -= 2;
170 } else {
171 assert(nc == 1);
172 psimd_store1_f32(c3, psimd_concat_hi_f32(vacc23x01, vacc23x01));
173 psimd_store1_f32(c2, vacc23x01);
174 psimd_store1_f32(c1, psimd_concat_hi_f32(vacc01x01, vacc01x01));
175 psimd_store1_f32(c0, vacc01x01);
176
177 nc = 0;
178 }
179 } while (nc != 0);
180 }
181