xref: /aosp_15_r20/external/clpeak/src/kernels/compute_dp_kernels.cl (revision 1cd03ba3888297bc945f2c84574e105e3ced3e34)
1*1cd03ba3SJeremy KempMSTRINGIFY(
2*1cd03ba3SJeremy Kemp
3*1cd03ba3SJeremy Kemp// Stringifying requires a new line after hash defines
4*1cd03ba3SJeremy Kemp
5*1cd03ba3SJeremy Kemp\n#if defined(cl_khr_fp64)
6*1cd03ba3SJeremy Kemp\n  #pragma OPENCL EXTENSION cl_khr_fp64 : enable
7*1cd03ba3SJeremy Kemp\n  #define DOUBLE_AVAILABLE
8*1cd03ba3SJeremy Kemp\n#elif defined(cl_amd_fp64)
9*1cd03ba3SJeremy Kemp\n  #pragma OPENCL EXTENSION cl_amd_fp64 : enable
10*1cd03ba3SJeremy Kemp\n  #define DOUBLE_AVAILABLE
11*1cd03ba3SJeremy Kemp\n#endif
12*1cd03ba3SJeremy Kemp
13*1cd03ba3SJeremy Kemp\n#undef MAD_4
14*1cd03ba3SJeremy Kemp\n#undef MAD_16
15*1cd03ba3SJeremy Kemp\n#undef MAD_64
16*1cd03ba3SJeremy Kemp\n
17*1cd03ba3SJeremy Kemp\n#define MAD_4(x, y)     x = mad(y, x, y);   y = mad(x, y, x);   x = mad(y, x, y);   y = mad(x, y, x);
18*1cd03ba3SJeremy Kemp\n#define MAD_16(x, y)    MAD_4(x, y);        MAD_4(x, y);        MAD_4(x, y);        MAD_4(x, y);
19*1cd03ba3SJeremy Kemp\n#define MAD_64(x, y)    MAD_16(x, y);       MAD_16(x, y);       MAD_16(x, y);       MAD_16(x, y);
20*1cd03ba3SJeremy Kemp\n
21*1cd03ba3SJeremy Kemp
22*1cd03ba3SJeremy Kemp\n
23*1cd03ba3SJeremy Kemp\n#ifdef DOUBLE_AVAILABLE
24*1cd03ba3SJeremy Kemp\n
25*1cd03ba3SJeremy Kemp
26*1cd03ba3SJeremy Kemp
27*1cd03ba3SJeremy Kemp__kernel void compute_dp_v1(__global double *ptr, double _A)
28*1cd03ba3SJeremy Kemp{
29*1cd03ba3SJeremy Kemp    double x = _A;
30*1cd03ba3SJeremy Kemp    double y = (double)get_local_id(0);
31*1cd03ba3SJeremy Kemp
32*1cd03ba3SJeremy Kemp    for(int i=0; i<128; i++)
33*1cd03ba3SJeremy Kemp    {
34*1cd03ba3SJeremy Kemp        MAD_16(x, y);
35*1cd03ba3SJeremy Kemp    }
36*1cd03ba3SJeremy Kemp
37*1cd03ba3SJeremy Kemp    ptr[get_global_id(0)] = y;
38*1cd03ba3SJeremy Kemp}
39*1cd03ba3SJeremy Kemp
40*1cd03ba3SJeremy Kemp
41*1cd03ba3SJeremy Kemp__kernel void compute_dp_v2(__global double *ptr, double _A)
42*1cd03ba3SJeremy Kemp{
43*1cd03ba3SJeremy Kemp    double2 x = (double2)(_A, (_A+1));
44*1cd03ba3SJeremy Kemp    double2 y = (double2)get_local_id(0);
45*1cd03ba3SJeremy Kemp
46*1cd03ba3SJeremy Kemp    for(int i=0; i<64; i++)
47*1cd03ba3SJeremy Kemp    {
48*1cd03ba3SJeremy Kemp        MAD_16(x, y);
49*1cd03ba3SJeremy Kemp    }
50*1cd03ba3SJeremy Kemp
51*1cd03ba3SJeremy Kemp    ptr[get_global_id(0)] = (y.S0) + (y.S1);
52*1cd03ba3SJeremy Kemp}
53*1cd03ba3SJeremy Kemp
54*1cd03ba3SJeremy Kemp__kernel void compute_dp_v4(__global double *ptr, double _A)
55*1cd03ba3SJeremy Kemp{
56*1cd03ba3SJeremy Kemp    double4 x = (double4)(_A, (_A+1), (_A+2), (_A+3));
57*1cd03ba3SJeremy Kemp    double4 y = (double4)get_local_id(0);
58*1cd03ba3SJeremy Kemp
59*1cd03ba3SJeremy Kemp    for(int i=0; i<32; i++)
60*1cd03ba3SJeremy Kemp    {
61*1cd03ba3SJeremy Kemp        MAD_16(x, y);
62*1cd03ba3SJeremy Kemp    }
63*1cd03ba3SJeremy Kemp
64*1cd03ba3SJeremy Kemp    ptr[get_global_id(0)] = (y.S0) + (y.S1) + (y.S2) + (y.S3);
65*1cd03ba3SJeremy Kemp}
66*1cd03ba3SJeremy Kemp
67*1cd03ba3SJeremy Kemp
68*1cd03ba3SJeremy Kemp__kernel void compute_dp_v8(__global double *ptr, double _A)
69*1cd03ba3SJeremy Kemp{
70*1cd03ba3SJeremy Kemp    double8 x = (double8)(_A, (_A+1), (_A+2), (_A+3), (_A+4), (_A+5), (_A+6), (_A+7));
71*1cd03ba3SJeremy Kemp    double8 y = (double8)get_local_id(0);
72*1cd03ba3SJeremy Kemp
73*1cd03ba3SJeremy Kemp    for(int i=0; i<16; i++)
74*1cd03ba3SJeremy Kemp    {
75*1cd03ba3SJeremy Kemp        MAD_16(x, y);
76*1cd03ba3SJeremy Kemp    }
77*1cd03ba3SJeremy Kemp
78*1cd03ba3SJeremy Kemp    ptr[get_global_id(0)] = (y.S0) + (y.S1) + (y.S2) + (y.S3) + (y.S4) + (y.S5) + (y.S6) + (y.S7);
79*1cd03ba3SJeremy Kemp}
80*1cd03ba3SJeremy Kemp
81*1cd03ba3SJeremy Kemp__kernel void compute_dp_v16(__global double *ptr, double _A)
82*1cd03ba3SJeremy Kemp{
83*1cd03ba3SJeremy Kemp    double16 x = (double16)(_A, (_A+1), (_A+2), (_A+3), (_A+4), (_A+5), (_A+6), (_A+7),
84*1cd03ba3SJeremy Kemp                    (_A+8), (_A+9), (_A+10), (_A+11), (_A+12), (_A+13), (_A+14), (_A+15));
85*1cd03ba3SJeremy Kemp    double16 y = (double16)get_local_id(0);
86*1cd03ba3SJeremy Kemp
87*1cd03ba3SJeremy Kemp    for(int i=0; i<8; i++)
88*1cd03ba3SJeremy Kemp    {
89*1cd03ba3SJeremy Kemp        MAD_16(x, y);
90*1cd03ba3SJeremy Kemp    }
91*1cd03ba3SJeremy Kemp
92*1cd03ba3SJeremy Kemp    double2 t = (y.S01) + (y.S23) + (y.S45) + (y.S67) + (y.S89) + (y.SAB) + (y.SCD) + (y.SEF);
93*1cd03ba3SJeremy Kemp    ptr[get_global_id(0)] = t.S0 + t.S1;
94*1cd03ba3SJeremy Kemp}
95*1cd03ba3SJeremy Kemp
96*1cd03ba3SJeremy Kemp\n
97*1cd03ba3SJeremy Kemp\n#endif      // DOUBLE_AVAILABLE
98*1cd03ba3SJeremy Kemp\n
99*1cd03ba3SJeremy Kemp
100*1cd03ba3SJeremy Kemp)
101