1 // Auto-generated file. Do not edit!
2 // Template: src/f32-gemm/psimd-splat.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/gemm.h>
15
16
xnn_f32_gemminc_ukernel_1x8__psimd_splat(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_output_params params[restrict static1])17 void xnn_f32_gemminc_ukernel_1x8__psimd_splat(
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_output_params params[restrict static 1])
29 {
30 assert(mr != 0);
31 assert(mr <= 1);
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
43 do {
44 psimd_f32 vacc0x0123 = psimd_load_f32(acc + 0);
45 psimd_f32 vacc0x4567 = psimd_load_f32(acc + 4);
46 acc += 8;
47
48 size_t k = kc;
49 while (k >= 4 * sizeof(float)) {
50 const psimd_f32 va0 = psimd_load_f32(a0);
51 a0 += 4;
52
53 const psimd_f32 va0c0 = psimd_splat0_f32(va0);
54
55 const psimd_f32 vb0123c0 = psimd_load_f32(w + 0);
56 const psimd_f32 vb4567c0 = psimd_load_f32(w + 4);
57
58 vacc0x0123 = psimd_qfma_f32(vacc0x0123, va0c0, vb0123c0);
59 vacc0x4567 = psimd_qfma_f32(vacc0x4567, va0c0, vb4567c0);
60 const psimd_f32 va0c1 = psimd_splat1_f32(va0);
61
62 const psimd_f32 vb0123c1 = psimd_load_f32(w + 8);
63 const psimd_f32 vb4567c1 = psimd_load_f32(w + 12);
64
65 vacc0x0123 = psimd_qfma_f32(vacc0x0123, va0c1, vb0123c1);
66 vacc0x4567 = psimd_qfma_f32(vacc0x4567, va0c1, vb4567c1);
67 const psimd_f32 va0c2 = psimd_splat2_f32(va0);
68
69 const psimd_f32 vb0123c2 = psimd_load_f32(w + 16);
70 const psimd_f32 vb4567c2 = psimd_load_f32(w + 20);
71
72 vacc0x0123 = psimd_qfma_f32(vacc0x0123, va0c2, vb0123c2);
73 vacc0x4567 = psimd_qfma_f32(vacc0x4567, va0c2, vb4567c2);
74 const psimd_f32 va0c3 = psimd_splat3_f32(va0);
75
76 const psimd_f32 vb0123c3 = psimd_load_f32(w + 24);
77 const psimd_f32 vb4567c3 = psimd_load_f32(w + 28);
78
79 vacc0x0123 = psimd_qfma_f32(vacc0x0123, va0c3, vb0123c3);
80 vacc0x4567 = psimd_qfma_f32(vacc0x4567, va0c3, vb4567c3);
81
82 w += 32;
83 k -= 4 * sizeof(float);
84 }
85 if XNN_UNLIKELY(k != 0) {
86 do {
87 const psimd_f32 va0 = psimd_load_splat_f32(a0);
88 a0 += 1;
89
90 const psimd_f32 vb0123 = psimd_load_f32(w);
91 const psimd_f32 vb4567 = psimd_load_f32(w + 4);
92 w += 8;
93
94 vacc0x0123 = psimd_qfma_f32(vacc0x0123, va0, vb0123);
95 vacc0x4567 = psimd_qfma_f32(vacc0x4567, va0, vb4567);
96
97 k -= sizeof(float);
98 } while (k != 0);
99 }
100
101 const psimd_f32 vmax = psimd_load_splat_f32(¶ms->scalar.max);
102 vacc0x0123 = psimd_min_f32(vacc0x0123, vmax);
103 vacc0x4567 = psimd_min_f32(vacc0x4567, vmax);
104
105 const psimd_f32 vmin = psimd_load_splat_f32(¶ms->scalar.min);
106 vacc0x0123 = psimd_max_f32(vacc0x0123, vmin);
107 vacc0x4567 = psimd_max_f32(vacc0x4567, vmin);
108
109 if XNN_LIKELY(nc >= 8) {
110 psimd_store_f32(c0, vacc0x0123);
111 psimd_store_f32(c0 + 4, vacc0x4567);
112 c0 = (float*) ((uintptr_t) c0 + cn_stride);
113
114 a0 = (const float*) ((uintptr_t) a0 - kc);
115
116 nc -= 8;
117 } else {
118 if (nc & 4) {
119 psimd_store_f32(c0, vacc0x0123);
120
121 vacc0x0123 = vacc0x4567;
122
123 c0 += 4;
124 }
125 if (nc & 2) {
126 psimd_store2_f32(c0, vacc0x0123);
127
128 vacc0x0123 = psimd_concat_hi_f32(vacc0x0123, vacc0x0123);
129
130 c0 += 2;
131 }
132 if (nc & 1) {
133 psimd_store1_f32(c0, vacc0x0123);
134 }
135
136 nc = 0;
137 }
138 } while (nc != 0);
139 }
140