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_gemm_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 union xnn_f32_output_params params[restrict static1])17 void xnn_f32_gemm_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 union xnn_f32_output_params params[restrict static 1])
28 {
29 assert(mr != 0);
30 assert(mr <= 1);
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
41 do {
42 psimd_f32 vacc0x0123 = psimd_load_f32(w + 0);
43 psimd_f32 vacc0x4567 = psimd_load_f32(w + 4);
44 w += 8;
45
46 size_t k = kc;
47 while (k >= 4 * sizeof(float)) {
48 psimd_f32 va0 = psimd_load_f32(a0);
49 a0 += 4;
50
51
52 const psimd_f32 vb0123c0 = psimd_load_f32(w + 0);
53 const psimd_f32 vb4567c0 = psimd_load_f32(w + 4);
54
55 vacc0x0123 = psimd_qfma_f32(vacc0x0123, va0, vb0123c0);
56 vacc0x4567 = psimd_qfma_f32(vacc0x4567, va0, vb4567c0);
57
58 #ifdef __clang__
59 va0 = __builtin_shufflevector(va0, va0, 1, 2, 3, 0);
60 #else
61 va0 = __builtin_shuffle(va0, va0, (psimd_s32) { 1, 2, 3, 0 });
62 #endif
63
64 const psimd_f32 vb0123c1 = psimd_load_f32(w + 8);
65 const psimd_f32 vb4567c1 = psimd_load_f32(w + 12);
66
67 vacc0x0123 = psimd_qfma_f32(vacc0x0123, va0, vb0123c1);
68 vacc0x4567 = psimd_qfma_f32(vacc0x4567, va0, vb4567c1);
69
70 #ifdef __clang__
71 va0 = __builtin_shufflevector(va0, va0, 1, 2, 3, 0);
72 #else
73 va0 = __builtin_shuffle(va0, va0, (psimd_s32) { 1, 2, 3, 0 });
74 #endif
75
76 const psimd_f32 vb0123c2 = psimd_load_f32(w + 16);
77 const psimd_f32 vb4567c2 = psimd_load_f32(w + 20);
78
79 vacc0x0123 = psimd_qfma_f32(vacc0x0123, va0, vb0123c2);
80 vacc0x4567 = psimd_qfma_f32(vacc0x4567, va0, vb4567c2);
81
82 #ifdef __clang__
83 va0 = __builtin_shufflevector(va0, va0, 1, 2, 3, 0);
84 #else
85 va0 = __builtin_shuffle(va0, va0, (psimd_s32) { 1, 2, 3, 0 });
86 #endif
87
88 const psimd_f32 vb0123c3 = psimd_load_f32(w + 24);
89 const psimd_f32 vb4567c3 = psimd_load_f32(w + 28);
90
91 vacc0x0123 = psimd_qfma_f32(vacc0x0123, va0, vb0123c3);
92 vacc0x4567 = psimd_qfma_f32(vacc0x4567, va0, vb4567c3);
93
94
95 w += 32;
96 k -= 4 * sizeof(float);
97 }
98 if XNN_UNLIKELY(k != 0) {
99 do {
100 const psimd_f32 va0 = psimd_load_splat_f32(a0);
101 a0 += 1;
102
103 const psimd_f32 vb0123 = psimd_load_f32(w);
104 const psimd_f32 vb4567 = psimd_load_f32(w + 4);
105 w += 8;
106
107 vacc0x0123 = psimd_qfma_f32(vacc0x0123, va0, vb0123);
108 vacc0x4567 = psimd_qfma_f32(vacc0x4567, va0, vb4567);
109
110 k -= sizeof(float);
111 } while (k != 0);
112 }
113
114 const psimd_f32 vmax = psimd_load_splat_f32(¶ms->scalar.max);
115 vacc0x0123 = psimd_min_f32(vacc0x0123, vmax);
116 vacc0x4567 = psimd_min_f32(vacc0x4567, vmax);
117
118 const psimd_f32 vmin = psimd_load_splat_f32(¶ms->scalar.min);
119 vacc0x0123 = psimd_max_f32(vacc0x0123, vmin);
120 vacc0x4567 = psimd_max_f32(vacc0x4567, vmin);
121
122 if XNN_LIKELY(nc >= 8) {
123 psimd_store_f32(c0, vacc0x0123);
124 psimd_store_f32(c0 + 4, vacc0x4567);
125 c0 = (float*) ((uintptr_t) c0 + cn_stride);
126
127 a0 = (const float*) ((uintptr_t) a0 - kc);
128
129 nc -= 8;
130 } else {
131 if (nc & 4) {
132 psimd_store_f32(c0, vacc0x0123);
133
134 vacc0x0123 = vacc0x4567;
135
136 c0 += 4;
137 }
138 if (nc & 2) {
139 psimd_store2_f32(c0, vacc0x0123);
140
141 vacc0x0123 = psimd_concat_hi_f32(vacc0x0123, vacc0x0123);
142
143 c0 += 2;
144 }
145 if (nc & 1) {
146 psimd_store1_f32(c0, vacc0x0123);
147 }
148
149 nc = 0;
150 }
151 } while (nc != 0);
152 }
153