1 // Auto-generated file. Do not edit!
2 // Template: src/f32-igemm/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/igemm.h>
15
16
xnn_f32_igemm_ukernel_1x8s4__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_1x8s4__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 <= 1);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(kc % sizeof(float) == 0);
36 assert(ks != 0);
37 assert(ks % (1 * 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
45 do {
46 psimd_f32 vacc0x0123 = psimd_load_f32(w);
47 psimd_f32 vacc0x4567 = psimd_load_f32(w + 4);
48 w += 8;
49
50 size_t p = ks;
51 do {
52 const float* restrict a0 = a[0];
53 assert(a0 != NULL);
54 if XNN_UNPREDICTABLE(a0 != zero) {
55 a0 = (const float*) ((uintptr_t) a0 + a_offset);
56 }
57 a += 1;
58
59 size_t k = kc;
60 while (k >= 4 * sizeof(float)) {
61 psimd_f32 va0 = psimd_load_f32(a0);
62 a0 += 4;
63
64
65 const psimd_f32 vb0123c0 = psimd_load_f32(w + 0);
66 const psimd_f32 vb4567c0 = psimd_load_f32(w + 4);
67
68 vacc0x0123 = psimd_qfma_f32(vacc0x0123, va0, vb0123c0);
69 vacc0x4567 = psimd_qfma_f32(vacc0x4567, va0, vb4567c0);
70
71 #ifdef __clang__
72 va0 = __builtin_shufflevector(va0, va0, 1, 2, 3, 0);
73 #else
74 va0 = __builtin_shuffle(va0, va0, (psimd_s32) { 1, 2, 3, 0 });
75 #endif
76
77 const psimd_f32 vb0123c1 = psimd_load_f32(w + 8);
78 const psimd_f32 vb4567c1 = psimd_load_f32(w + 12);
79
80 vacc0x0123 = psimd_qfma_f32(vacc0x0123, va0, vb0123c1);
81 vacc0x4567 = psimd_qfma_f32(vacc0x4567, va0, vb4567c1);
82
83 #ifdef __clang__
84 va0 = __builtin_shufflevector(va0, va0, 1, 2, 3, 0);
85 #else
86 va0 = __builtin_shuffle(va0, va0, (psimd_s32) { 1, 2, 3, 0 });
87 #endif
88
89 const psimd_f32 vb0123c2 = psimd_load_f32(w + 16);
90 const psimd_f32 vb4567c2 = psimd_load_f32(w + 20);
91
92 vacc0x0123 = psimd_qfma_f32(vacc0x0123, va0, vb0123c2);
93 vacc0x4567 = psimd_qfma_f32(vacc0x4567, va0, vb4567c2);
94
95 #ifdef __clang__
96 va0 = __builtin_shufflevector(va0, va0, 1, 2, 3, 0);
97 #else
98 va0 = __builtin_shuffle(va0, va0, (psimd_s32) { 1, 2, 3, 0 });
99 #endif
100
101 const psimd_f32 vb0123c3 = psimd_load_f32(w + 24);
102 const psimd_f32 vb4567c3 = psimd_load_f32(w + 28);
103
104 vacc0x0123 = psimd_qfma_f32(vacc0x0123, va0, vb0123c3);
105 vacc0x4567 = psimd_qfma_f32(vacc0x4567, va0, vb4567c3);
106
107
108 w += 32;
109 k -= 4 * sizeof(float);
110 }
111 if XNN_UNLIKELY(k != 0) {
112 do {
113 const psimd_f32 vb0123 = psimd_load_f32(w);
114 const psimd_f32 vb4567 = psimd_load_f32(w + 4);
115 w += 8;
116
117 const psimd_f32 va0 = psimd_load_splat_f32(a0);
118 a0 += 1;
119
120 vacc0x0123 = psimd_qfma_f32(vacc0x0123, va0, vb0123);
121 vacc0x4567 = psimd_qfma_f32(vacc0x4567, va0, vb4567);
122 k -= sizeof(float);
123 } while (k != 0);
124 }
125 p -= 1 * sizeof(void*);
126 } while (p != 0);
127
128 const psimd_f32 vmax = psimd_load_splat_f32(¶ms->scalar.max);
129 vacc0x0123 = psimd_min_f32(vacc0x0123, vmax);
130 vacc0x4567 = psimd_min_f32(vacc0x4567, vmax);
131
132 const psimd_f32 vmin = psimd_load_splat_f32(¶ms->scalar.min);
133 vacc0x0123 = psimd_max_f32(vacc0x0123, vmin);
134 vacc0x4567 = psimd_max_f32(vacc0x4567, vmin);
135
136 if XNN_LIKELY(nc >= 8) {
137 psimd_store_f32(c0, vacc0x0123);
138 psimd_store_f32(c0 + 4, vacc0x4567);
139 c0 = (float*) ((uintptr_t) c0 + cn_stride);
140
141 a = (const float**restrict) ((uintptr_t) a - ks);
142 nc -= 8;
143 } else {
144 if (nc & 4) {
145 psimd_store_f32(c0, vacc0x0123);
146
147 vacc0x0123 = vacc0x4567;
148
149 c0 += 4;
150 }
151 if (nc & 2) {
152 psimd_store2_f32(c0, vacc0x0123);
153
154 vacc0x0123 = psimd_concat_hi_f32(vacc0x0123, vacc0x0123);
155
156 c0 += 2;
157 }
158 if (nc & 1) {
159 psimd_store1_f32(c0, vacc0x0123);
160 }
161
162 nc = 0;
163 }
164 } while (nc != 0);
165 }
166