1 // Auto-generated file. Do not edit!
2 // Template: src/f32-gemm/psimd-s4.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_1x8s4__psimd(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_1x8s4__psimd(
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 psimd_f32 va0 = psimd_load_f32(a0);
51 a0 += 4;
52
53
54 const psimd_f32 vb0123c0 = psimd_load_f32(w + 0);
55 const psimd_f32 vb4567c0 = psimd_load_f32(w + 4);
56
57 vacc0x0123 = psimd_qfma_f32(vacc0x0123, va0, vb0123c0);
58 vacc0x4567 = psimd_qfma_f32(vacc0x4567, va0, vb4567c0);
59
60 #ifdef __clang__
61 va0 = __builtin_shufflevector(va0, va0, 1, 2, 3, 0);
62 #else
63 va0 = __builtin_shuffle(va0, va0, (psimd_s32) { 1, 2, 3, 0 });
64 #endif
65
66 const psimd_f32 vb0123c1 = psimd_load_f32(w + 8);
67 const psimd_f32 vb4567c1 = psimd_load_f32(w + 12);
68
69 vacc0x0123 = psimd_qfma_f32(vacc0x0123, va0, vb0123c1);
70 vacc0x4567 = psimd_qfma_f32(vacc0x4567, va0, vb4567c1);
71
72 #ifdef __clang__
73 va0 = __builtin_shufflevector(va0, va0, 1, 2, 3, 0);
74 #else
75 va0 = __builtin_shuffle(va0, va0, (psimd_s32) { 1, 2, 3, 0 });
76 #endif
77
78 const psimd_f32 vb0123c2 = psimd_load_f32(w + 16);
79 const psimd_f32 vb4567c2 = psimd_load_f32(w + 20);
80
81 vacc0x0123 = psimd_qfma_f32(vacc0x0123, va0, vb0123c2);
82 vacc0x4567 = psimd_qfma_f32(vacc0x4567, va0, vb4567c2);
83
84 #ifdef __clang__
85 va0 = __builtin_shufflevector(va0, va0, 1, 2, 3, 0);
86 #else
87 va0 = __builtin_shuffle(va0, va0, (psimd_s32) { 1, 2, 3, 0 });
88 #endif
89
90 const psimd_f32 vb0123c3 = psimd_load_f32(w + 24);
91 const psimd_f32 vb4567c3 = psimd_load_f32(w + 28);
92
93 vacc0x0123 = psimd_qfma_f32(vacc0x0123, va0, vb0123c3);
94 vacc0x4567 = psimd_qfma_f32(vacc0x4567, va0, vb4567c3);
95
96
97 w += 32;
98 k -= 4 * sizeof(float);
99 }
100 if XNN_UNLIKELY(k != 0) {
101 do {
102 const psimd_f32 va0 = psimd_load_splat_f32(a0);
103 a0 += 1;
104
105 const psimd_f32 vb0123 = psimd_load_f32(w);
106 const psimd_f32 vb4567 = psimd_load_f32(w + 4);
107 w += 8;
108
109 vacc0x0123 = psimd_qfma_f32(vacc0x0123, va0, vb0123);
110 vacc0x4567 = psimd_qfma_f32(vacc0x4567, va0, vb4567);
111
112 k -= sizeof(float);
113 } while (k != 0);
114 }
115
116 const psimd_f32 vmax = psimd_load_splat_f32(¶ms->scalar.max);
117 vacc0x0123 = psimd_min_f32(vacc0x0123, vmax);
118 vacc0x4567 = psimd_min_f32(vacc0x4567, vmax);
119
120 const psimd_f32 vmin = psimd_load_splat_f32(¶ms->scalar.min);
121 vacc0x0123 = psimd_max_f32(vacc0x0123, vmin);
122 vacc0x4567 = psimd_max_f32(vacc0x4567, vmin);
123
124 if XNN_LIKELY(nc >= 8) {
125 psimd_store_f32(c0, vacc0x0123);
126 psimd_store_f32(c0 + 4, vacc0x4567);
127 c0 = (float*) ((uintptr_t) c0 + cn_stride);
128
129 a0 = (const float*) ((uintptr_t) a0 - kc);
130
131 nc -= 8;
132 } else {
133 if (nc & 4) {
134 psimd_store_f32(c0, vacc0x0123);
135
136 vacc0x0123 = vacc0x4567;
137
138 c0 += 4;
139 }
140 if (nc & 2) {
141 psimd_store2_f32(c0, vacc0x0123);
142
143 vacc0x0123 = psimd_concat_hi_f32(vacc0x0123, vacc0x0123);
144
145 c0 += 2;
146 }
147 if (nc & 1) {
148 psimd_store1_f32(c0, vacc0x0123);
149 }
150
151 nc = 0;
152 }
153 } while (nc != 0);
154 }
155