1 /* { dg-do run { target { powerpc*-*-* && { lp64 && p9vector_hw } } } } */
2 /* { dg-options "-mdejagnu-cpu=power9 -O2 " } */
3
4 /* PR83677: This test case used to fail due to mis-generation of the
5 xxpermr instruction. It requires inlining to create enough register
6 pressure that we generate xxpermr rather than vpermr. */
7
8 #include <altivec.h>
9
10 void v_expand_u8(vector unsigned char* a, vector unsigned short* b0, vector unsigned short* b1)
11 {
12 #if __LITTLE_ENDIAN__
13 *b0 = (vector unsigned short)vec_mergeh(*a, vec_splats((unsigned char)0));
14 *b1 = (vector unsigned short)vec_mergel(*a, vec_splats((unsigned char)0));
15 #else
16 *b0 = (vector unsigned short)vec_mergeh(vec_splats((unsigned char)0), *a);
17 *b1 = (vector unsigned short)vec_mergel(vec_splats((unsigned char)0), *a);
18 #endif
19 }
20
21 void v_expand_u16(vector unsigned short* a, vector unsigned int* b0, vector unsigned int* b1)
22 {
23 #if __LITTLE_ENDIAN__
24 *b0 = (vector unsigned int)vec_mergeh(*a, vec_splats((unsigned short)0));
25 *b1 = (vector unsigned int)vec_mergel(*a, vec_splats((unsigned short)0));
26 #else
27 *b0 = (vector unsigned int)vec_mergeh(vec_splats((unsigned short)0), *a);
28 *b1 = (vector unsigned int)vec_mergel(vec_splats((unsigned short)0), *a);
29 #endif
30 }
31
32 void v_load_deinterleave_u8(unsigned char *ptr, vector unsigned char* a, vector unsigned char* b, vector unsigned char* c)
33 {
34 vector unsigned char v1 = vec_xl( 0, ptr);
35 vector unsigned char v2 = vec_xl(16, ptr);
36 vector unsigned char v3 = vec_xl(32, ptr);
37
38 static const vector unsigned char a12_perm = {0, 3, 6, 9, 12, 15, 18, 21, 24, 27, 30, 0, 0, 0, 0, 0};
39 static const vector unsigned char a123_perm = {0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 17, 20, 23, 26, 29};
40 *a = vec_perm(vec_perm(v1, v2, a12_perm), v3, a123_perm);
41
42 static const vector unsigned char b12_perm = {1, 4, 7, 10, 13, 16, 19, 22, 25, 28, 31, 0, 0, 0, 0, 0};
43 static const vector unsigned char b123_perm = {0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 18, 21, 24, 27, 30};
44 *b = vec_perm(vec_perm(v1, v2, b12_perm), v3, b123_perm);
45
46 static const vector unsigned char c12_perm = {2, 5, 8, 11, 14, 17, 20, 23, 26, 29, 0, 0, 0, 0, 0, 0};
47 static const vector unsigned char c123_perm = {0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 16, 19, 22, 25, 28, 31};
48 *c = vec_perm(vec_perm(v1, v2, c12_perm), v3, c123_perm);
49 }
50
51 void v_load_deinterleave_f32(float *ptr, vector float* a, vector float* b, vector float* c)
52 {
53 vector float v1 = vec_xl( 0, ptr);
54 vector float v2 = vec_xl(16, ptr);
55 vector float v3 = vec_xl(32, ptr);
56
57 #if __LITTLE_ENDIAN__
58 vector float t1 = vec_sld(v3, v2, 8);
59 vector float t2 = vec_sld(v1, v3, 8);
60 vector float t3 = vec_sld(v2, v1, 8);
61 #else
62 vector float t1 = vec_sld(v2, v3, 8);
63 vector float t2 = vec_sld(v3, v1, 8);
64 vector float t3 = vec_sld(v1, v2, 8);
65 #endif
66
67 static const vector unsigned char flp = {0, 1, 2, 3, 12, 13, 14, 15, 16, 17, 18, 19, 28, 29, 30, 31};
68 *a = vec_perm(v1, t1, flp);
69
70 static const vector unsigned char flp2 = {28, 29, 30, 31, 0, 1, 2, 3, 12, 13, 14, 15, 16, 17, 18, 19};
71 *b = vec_perm(v2, t2, flp2);
72
73 *c = vec_perm(t3, v3, flp);
74 }
75
76 void v_store_interleave_f32(float *ptr, vector float a, vector float b, vector float c)
77 {
78 vector float hbc = vec_mergeh(b, c);
79
80 static const vector unsigned char ahbc = {0, 1, 2, 3, 16, 17, 18, 19, 20, 21, 22, 23, 4, 5, 6, 7};
81 vec_xst(vec_perm(a, hbc, ahbc), 0, ptr);
82
83 vector float lab = vec_mergel(a, b);
84 #if __LITTLE_ENDIAN__
85 vec_xst(vec_sld(lab, hbc, 8), 16, ptr);
86 #else
87 vec_xst(vec_sld(hbc, lab, 8), 16, ptr);
88 #endif
89
90 static const vector unsigned char clab = {8, 9, 10, 11, 24, 25, 26, 27, 28, 29, 30, 31, 12, 13, 14, 15};
91 vec_xst(vec_perm(c, lab, clab), 32, ptr);
92 }
93
94 vector float v_cvt_f32(vector unsigned int a)
95 {
96 return (vector float)vec_ctf(a, 0);
97 }
98
99 void acc_simd_(const unsigned char* src, float* dst, const unsigned char* mask, int len)
100 {
101 int x = 0;
102 const int cVectorWidth = 16;
103
104 for ( ; x <= len - cVectorWidth; x += cVectorWidth)
105 {
106 vector unsigned char v_mask = vec_xl(0, mask + x);
107 v_mask = (vector unsigned char)vec_cmpeq(vec_splats((unsigned char)0), v_mask);
108 v_mask = (vector unsigned char)vec_nor(v_mask, v_mask);
109 vector unsigned char v_src0, v_src1, v_src2;
110 v_load_deinterleave_u8((unsigned char *)(src + (x * 3)), &v_src0, &v_src1, &v_src2);
111 v_src0 = v_src0 & v_mask;
112 v_src1 = v_src1 & v_mask;
113 v_src2 = v_src2 & v_mask;
114
115 /* expand 16 uchar to 4 vectors which contains 4 uint */
116 vector unsigned short v_src00, v_src01, v_src10, v_src11, v_src20, v_src21;
117 v_expand_u8(&v_src0, &v_src00, &v_src01);
118 v_expand_u8(&v_src1, &v_src10, &v_src11);
119 v_expand_u8(&v_src2, &v_src20, &v_src21);
120 vector unsigned int v_src000, v_src001, v_src010, v_src011;
121 vector unsigned int v_src100, v_src101, v_src110, v_src111;
122 vector unsigned int v_src200, v_src201, v_src210, v_src211;
123 v_expand_u16(&v_src00, &v_src000, &v_src001);
124 v_expand_u16(&v_src01, &v_src010, &v_src011);
125 v_expand_u16(&v_src10, &v_src100, &v_src101);
126 v_expand_u16(&v_src11, &v_src110, &v_src111);
127 v_expand_u16(&v_src20, &v_src200, &v_src201);
128 v_expand_u16(&v_src21, &v_src210, &v_src211);
129
130 vector float v_dst000, v_dst001, v_dst010, v_dst011;
131 vector float v_dst100, v_dst101, v_dst110, v_dst111;
132 vector float v_dst200, v_dst201, v_dst210, v_dst211;
133 v_load_deinterleave_f32(dst + (x * 3), &v_dst000, &v_dst100, &v_dst200);
134 v_load_deinterleave_f32(dst + ((x + 4) * 3), &v_dst001, &v_dst101, &v_dst201);
135 v_load_deinterleave_f32(dst + ((x + 8) * 3), &v_dst010, &v_dst110, &v_dst210);
136 v_load_deinterleave_f32(dst + ((x + 12) * 3), &v_dst011, &v_dst111, &v_dst211);
137
138 v_store_interleave_f32(dst + (x * 3), vec_add(v_dst000, v_cvt_f32(v_src000)), vec_add(v_dst100, v_cvt_f32(v_src100)), vec_add(v_dst200, v_cvt_f32(v_src200)));
139 v_store_interleave_f32(dst + ((x + 4) * 3), vec_add(v_dst001, v_cvt_f32(v_src001)), vec_add(v_dst101, v_cvt_f32(v_src101)), vec_add(v_dst201, v_cvt_f32(v_src201)));
140 v_store_interleave_f32(dst + ((x + 8) * 3), vec_add(v_dst010, v_cvt_f32(v_src010)), vec_add(v_dst110, v_cvt_f32(v_src110)), vec_add(v_dst210, v_cvt_f32(v_src210)));
141 v_store_interleave_f32(dst + ((x + 12) * 3), vec_add(v_dst011, v_cvt_f32(v_src011)), vec_add(v_dst111, v_cvt_f32(v_src111)), vec_add(v_dst211, v_cvt_f32(v_src211)));
142 }
143 return;
144 }
145
146 void acc_(const unsigned char* src, float* dst, const unsigned char* mask, int len)
147 {
148 int x = 0;
149 src += (x * 3);
150 dst += (x * 3);
151 for( ; x < len; x++, src += 3, dst += 3 )
152 {
153 if( mask[x] ) /* if mask, R/G/B dst[] += src[] */
154 {
155 for( int k = 0; k < 3; k++ )
156 {
157 dst[k] += src[k];
158 }
159 }
160 }
161 return;
162 }
163
164 #define N 16
165
166 int main(int argc, char *argv[])
167 {
168 unsigned char __attribute__ ((aligned (16) )) mask[] = {0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 1, 0, 0, 1};
169 unsigned char __attribute__ ((aligned (16) )) src[3*N];
170 float __attribute__ ((aligned (16) )) dst[3*N];
171 float __attribute__ ((aligned (16) )) exp[3*N];
172
173 int i;
174
175 /* initialize src and dst */
176 for (i=0; i<3*N; i++) src[i] = (unsigned char)(i*3);
177 for (i=0; i<3*N; i++) {dst[i] = i * 1.0f; exp[i] = dst[i];}
178
179 acc_(src, exp, mask, N);
180 acc_simd_(src, dst, mask, N);
181
182 for (i=0; i<N; i++)
183 {
184 if ((dst[3*i] != exp[3*i]) || (dst[3*i+1] != exp[3*i+1]) || (dst[3*i+2] != exp[3*i+2]))
185 __builtin_abort ();
186 }
187
188 return 0;
189 }