1040
|
1 // $HeadURL$
|
|
2 // $Date$
|
|
3 // $Author$
|
|
4
|
|
5 module dstress.run.a.asm_rsqrtps_01_A;
|
|
6
|
|
7 version(D_InlineAsm_X86){
|
|
8 version = runTest;
|
|
9 }else version(D_InlineAsm_X86_64){
|
|
10 version = runTest;
|
|
11 }
|
|
12
|
|
13 version(runTest){
|
|
14 import addon.cpuinfo;
|
|
15
|
|
16 int main(){
|
|
17 haveSSE!()();
|
|
18
|
1265
|
19 float* a = (new float[4]).ptr;
|
1085
|
20 a[0] = 1.0f;
|
|
21 a[1] = 2.0f;
|
|
22 a[2] = 3.0f;
|
|
23 a[3] = 4.0f;
|
|
24
|
1265
|
25 float* b = (new float[4]).ptr;
|
1085
|
26 b[0] = -1.0f;
|
|
27 b[1] = 1.0f;
|
|
28 b[2] = 4.0f;
|
|
29 b[3] = -4.0f;
|
|
30
|
1265
|
31 float* c = (new float[4]).ptr;
|
1040
|
32
|
1265
|
33 static if(size_t.sizeof == 4){
|
|
34 asm{
|
|
35 mov EAX, a;
|
|
36 movups XMM0, [EAX];
|
|
37 mov EAX, b;
|
|
38 movups XMM1, [EAX];
|
|
39 rsqrtps XMM0, XMM1;
|
|
40 mov EAX, c;
|
|
41 movups [EAX], XMM0;
|
|
42 }
|
|
43 }else static if(size_t.sizeof == 8){
|
|
44 asm{
|
|
45 mov RAX, a;
|
|
46 movups XMM0, [RAX];
|
|
47 mov RAX, b;
|
|
48 movups XMM1, [RAX];
|
|
49 rsqrtps XMM0, XMM1;
|
|
50 mov RAX, c;
|
|
51 movups [RAX], XMM0;
|
|
52 }
|
|
53 }else{
|
|
54 static assert(0, "unhandled pointer size");
|
1040
|
55 }
|
|
56
|
|
57 c[0] += 1.0f;
|
|
58 if(c[0] < 0.0f){
|
|
59 c[0] = -c[0];
|
|
60 }
|
|
61 if(c[0] > 1.0f / 4096.0f){
|
|
62 assert(0);
|
|
63 }
|
1044
|
64
|
1040
|
65 c[1] -= 1.0f;
|
|
66 if(c[1] < 0.0f){
|
|
67 c[1] = -c[1];
|
|
68 }
|
|
69 if(c[1] > 1.0f / 4096.0f){
|
|
70 assert(0);
|
|
71 }
|
|
72
|
|
73 c[2] -= 0.5f;
|
|
74 if(c[2] < 0.0f){
|
|
75 c[2] = -c[2];
|
|
76 }
|
|
77 if(c[2] > 0.25f / 4096.0f){
|
|
78 assert(0);
|
|
79 }
|
1044
|
80
|
1040
|
81 c[3] += 0.5f;
|
|
82 if(c[3] < 0.0f){
|
|
83 c[3] = -c[3];
|
|
84 }
|
|
85 if(c[3] > 0.25f / 4096.0f){
|
|
86 assert(0);
|
|
87 }
|
|
88
|
|
89 return 0;
|
|
90 }
|
|
91 }else{
|
|
92 pragma(msg, "DSTRESS{XFAIL}: no inline ASM support");
|
|
93 static assert(0);
|
|
94 }
|