10 #include "warpselect/WarpSelectImpl.cuh"
12 namespace faiss {
namespace gpu {
14 #ifdef FAISS_USE_FLOAT16
25 WARP_SELECT_DECL(half,
true, 1);
26 WARP_SELECT_DECL(half,
true, 32);
27 WARP_SELECT_DECL(half,
true, 64);
28 WARP_SELECT_DECL(half,
true, 128);
29 WARP_SELECT_DECL(half,
true, 256);
30 WARP_SELECT_DECL(half,
true, 512);
31 WARP_SELECT_DECL(half,
true, 1024);
33 WARP_SELECT_DECL(half,
false, 1);
34 WARP_SELECT_DECL(half,
false, 32);
35 WARP_SELECT_DECL(half,
false, 64);
36 WARP_SELECT_DECL(half,
false, 128);
37 WARP_SELECT_DECL(half,
false, 256);
38 WARP_SELECT_DECL(half,
false, 512);
39 WARP_SELECT_DECL(half,
false, 1024);
41 void runWarpSelect(Tensor<half, 2, true>& in,
42 Tensor<half, 2, true>& outK,
43 Tensor<int, 2, true>& outV,
44 bool dir,
int k, cudaStream_t stream) {
45 FAISS_ASSERT(k <= 1024);
49 WARP_SELECT_CALL(half,
true, 1);
51 WARP_SELECT_CALL(half,
true, 32);
53 WARP_SELECT_CALL(half,
true, 64);
54 }
else if (k <= 128) {
55 WARP_SELECT_CALL(half,
true, 128);
56 }
else if (k <= 256) {
57 WARP_SELECT_CALL(half,
true, 256);
58 }
else if (k <= 512) {
59 WARP_SELECT_CALL(half,
true, 512);
60 }
else if (k <= 1024) {
61 WARP_SELECT_CALL(half,
true, 1024);
65 WARP_SELECT_CALL(half,
false, 1);
67 WARP_SELECT_CALL(half,
false, 32);
69 WARP_SELECT_CALL(half,
false, 64);
70 }
else if (k <= 128) {
71 WARP_SELECT_CALL(half,
false, 128);
72 }
else if (k <= 256) {
73 WARP_SELECT_CALL(half,
false, 256);
74 }
else if (k <= 512) {
75 WARP_SELECT_CALL(half,
false, 512);
76 }
else if (k <= 1024) {
77 WARP_SELECT_CALL(half,
false, 1024);