1 // Auto-generated file. Do not edit!
2 // Template: src/f32-gemm/MRx2-neon-ld64.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 <arm_neon.h>
13
14 #include <xnnpack/common.h>
15 #include <xnnpack/gemm.h>
16
17
xnn_f32_gemm_minmax_ukernel_6x2__neonfma_lane_ld64(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_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_gemm_minmax_ukernel_6x2__neonfma_lane_ld64(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 const float* restrict a,
23 size_t a_stride,
24 const float*restrict w,
25 float*restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
29 {
30 assert(mr != 0);
31 assert(mr <= 6);
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
39 const float* a0 = a;
40 float* c0 = c;
41 const float* a1 = (const float*) ((uintptr_t) a0 + a_stride);
42 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
43 if XNN_UNPREDICTABLE(mr < 2) {
44 a1 = a0;
45 c1 = c0;
46 }
47 const float* a2 = (const float*) ((uintptr_t) a1 + a_stride);
48 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
49 if XNN_UNPREDICTABLE(mr <= 2) {
50 a2 = a1;
51 c2 = c1;
52 }
53 const float* a3 = (const float*) ((uintptr_t) a2 + a_stride);
54 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
55 if XNN_UNPREDICTABLE(mr < 4) {
56 a3 = a2;
57 c3 = c2;
58 }
59 const float* a4 = (const float*) ((uintptr_t) a3 + a_stride);
60 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
61 if XNN_UNPREDICTABLE(mr <= 4) {
62 a4 = a3;
63 c4 = c3;
64 }
65 const float* a5 = (const float*) ((uintptr_t) a4 + a_stride);
66 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
67 if XNN_UNPREDICTABLE(mr != 6) {
68 a5 = a4;
69 c5 = c4;
70 }
71
72 do {
73 float32x2_t vacc0x01 = vld1_f32(w); w += 2;
74 float32x2_t vacc1x01 = vacc0x01;
75 float32x2_t vacc2x01 = vacc0x01;
76 float32x2_t vacc3x01 = vacc0x01;
77 float32x2_t vacc4x01 = vacc0x01;
78 float32x2_t vacc5x01 = vacc0x01;
79
80 size_t k = kc;
81 for (; k >= 2 * sizeof(float); k -= 2 * sizeof(float)) {
82 const float32x2_t va0 = vld1_f32(a0); a0 += 2;
83 const float32x2_t va1 = vld1_f32(a1); a1 += 2;
84 const float32x2_t va2 = vld1_f32(a2); a2 += 2;
85 const float32x2_t va3 = vld1_f32(a3); a3 += 2;
86 const float32x2_t va4 = vld1_f32(a4); a4 += 2;
87 const float32x2_t va5 = vld1_f32(a5); a5 += 2;
88
89 const float32x2_t vb01c0 = vld1_f32(w); w += 2;
90
91 #if XNN_ARCH_ARM64
92 vacc0x01 = vfma_lane_f32(vacc0x01, vb01c0, va0, 0);
93 vacc1x01 = vfma_lane_f32(vacc1x01, vb01c0, va1, 0);
94 vacc2x01 = vfma_lane_f32(vacc2x01, vb01c0, va2, 0);
95 vacc3x01 = vfma_lane_f32(vacc3x01, vb01c0, va3, 0);
96 vacc4x01 = vfma_lane_f32(vacc4x01, vb01c0, va4, 0);
97 vacc5x01 = vfma_lane_f32(vacc5x01, vb01c0, va5, 0);
98 #else
99 const float32x2_t va0c0 = vdup_lane_f32(va0, 0);
100 const float32x2_t va1c0 = vdup_lane_f32(va1, 0);
101 const float32x2_t va2c0 = vdup_lane_f32(va2, 0);
102 const float32x2_t va3c0 = vdup_lane_f32(va3, 0);
103 const float32x2_t va4c0 = vdup_lane_f32(va4, 0);
104 const float32x2_t va5c0 = vdup_lane_f32(va5, 0);
105 vacc0x01 = vfma_f32(vacc0x01, va0c0, vb01c0);
106 vacc1x01 = vfma_f32(vacc1x01, va1c0, vb01c0);
107 vacc2x01 = vfma_f32(vacc2x01, va2c0, vb01c0);
108 vacc3x01 = vfma_f32(vacc3x01, va3c0, vb01c0);
109 vacc4x01 = vfma_f32(vacc4x01, va4c0, vb01c0);
110 vacc5x01 = vfma_f32(vacc5x01, va5c0, vb01c0);
111 #endif
112 const float32x2_t vb01c1 = vld1_f32(w); w += 2;
113
114 #if XNN_ARCH_ARM64
115 vacc0x01 = vfma_lane_f32(vacc0x01, vb01c1, va0, 1);
116 vacc1x01 = vfma_lane_f32(vacc1x01, vb01c1, va1, 1);
117 vacc2x01 = vfma_lane_f32(vacc2x01, vb01c1, va2, 1);
118 vacc3x01 = vfma_lane_f32(vacc3x01, vb01c1, va3, 1);
119 vacc4x01 = vfma_lane_f32(vacc4x01, vb01c1, va4, 1);
120 vacc5x01 = vfma_lane_f32(vacc5x01, vb01c1, va5, 1);
121 #else
122 const float32x2_t va0c1 = vdup_lane_f32(va0, 1);
123 const float32x2_t va1c1 = vdup_lane_f32(va1, 1);
124 const float32x2_t va2c1 = vdup_lane_f32(va2, 1);
125 const float32x2_t va3c1 = vdup_lane_f32(va3, 1);
126 const float32x2_t va4c1 = vdup_lane_f32(va4, 1);
127 const float32x2_t va5c1 = vdup_lane_f32(va5, 1);
128 vacc0x01 = vfma_f32(vacc0x01, va0c1, vb01c1);
129 vacc1x01 = vfma_f32(vacc1x01, va1c1, vb01c1);
130 vacc2x01 = vfma_f32(vacc2x01, va2c1, vb01c1);
131 vacc3x01 = vfma_f32(vacc3x01, va3c1, vb01c1);
132 vacc4x01 = vfma_f32(vacc4x01, va4c1, vb01c1);
133 vacc5x01 = vfma_f32(vacc5x01, va5c1, vb01c1);
134 #endif
135 }
136 if XNN_UNLIKELY(k != 0) {
137 const float32x2_t va0 = vld1_dup_f32(a0); a0 += 1;
138 const float32x2_t va1 = vld1_dup_f32(a1); a1 += 1;
139 const float32x2_t va2 = vld1_dup_f32(a2); a2 += 1;
140 const float32x2_t va3 = vld1_dup_f32(a3); a3 += 1;
141 const float32x2_t va4 = vld1_dup_f32(a4); a4 += 1;
142 const float32x2_t va5 = vld1_dup_f32(a5); a5 += 1;
143
144 const float32x2_t vb01 = vld1_f32(w); w += 2;
145
146 vacc0x01 = vfma_f32(vacc0x01, va0, vb01);
147 vacc1x01 = vfma_f32(vacc1x01, va1, vb01);
148 vacc2x01 = vfma_f32(vacc2x01, va2, vb01);
149 vacc3x01 = vfma_f32(vacc3x01, va3, vb01);
150 vacc4x01 = vfma_f32(vacc4x01, va4, vb01);
151 vacc5x01 = vfma_f32(vacc5x01, va5, vb01);
152 }
153
154 const float32x2_t vmax = vld1_dup_f32(¶ms->scalar.max);
155 vacc0x01 = vmin_f32(vacc0x01, vmax);
156 vacc1x01 = vmin_f32(vacc1x01, vmax);
157 vacc2x01 = vmin_f32(vacc2x01, vmax);
158 vacc3x01 = vmin_f32(vacc3x01, vmax);
159 vacc4x01 = vmin_f32(vacc4x01, vmax);
160 vacc5x01 = vmin_f32(vacc5x01, vmax);
161
162 const float32x2_t vmin = vld1_dup_f32(¶ms->scalar.min);
163 vacc0x01 = vmax_f32(vacc0x01, vmin);
164 vacc1x01 = vmax_f32(vacc1x01, vmin);
165 vacc2x01 = vmax_f32(vacc2x01, vmin);
166 vacc3x01 = vmax_f32(vacc3x01, vmin);
167 vacc4x01 = vmax_f32(vacc4x01, vmin);
168 vacc5x01 = vmax_f32(vacc5x01, vmin);
169
170 if XNN_LIKELY(nc >= 2) {
171 vst1_f32(c0, vacc0x01);
172 c0 = (float*) ((uintptr_t) c0 + cn_stride);
173 vst1_f32(c1, vacc1x01);
174 c1 = (float*) ((uintptr_t) c1 + cn_stride);
175 vst1_f32(c2, vacc2x01);
176 c2 = (float*) ((uintptr_t) c2 + cn_stride);
177 vst1_f32(c3, vacc3x01);
178 c3 = (float*) ((uintptr_t) c3 + cn_stride);
179 vst1_f32(c4, vacc4x01);
180 c4 = (float*) ((uintptr_t) c4 + cn_stride);
181 vst1_f32(c5, vacc5x01);
182 c5 = (float*) ((uintptr_t) c5 + cn_stride);
183
184 a0 = (const float*) ((uintptr_t) a0 - kc);
185 a1 = (const float*) ((uintptr_t) a1 - kc);
186 a2 = (const float*) ((uintptr_t) a2 - kc);
187 a3 = (const float*) ((uintptr_t) a3 - kc);
188 a4 = (const float*) ((uintptr_t) a4 - kc);
189 a5 = (const float*) ((uintptr_t) a5 - kc);
190
191 nc -= 2;
192 } else {
193 assert(nc == 1);
194 vst1_lane_f32(c0, vacc0x01, 0);
195 vst1_lane_f32(c1, vacc1x01, 0);
196 vst1_lane_f32(c2, vacc2x01, 0);
197 vst1_lane_f32(c3, vacc3x01, 0);
198 vst1_lane_f32(c4, vacc4x01, 0);
199 vst1_lane_f32(c5, vacc5x01, 0);
200
201 nc = 0;
202 }
203 } while (nc != 0);
204 }
205