wu-arch/kernels
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Set BM==BN==64, update doc

Browse Source
This commit is contained in:
Hansung Kim
2024-07-19 16:37:15 -07:00
parent 02feb36b12
commit 1f844fa9e9
1 changed files with 6 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
tests/regression/sgemm_tcore/util.hpp
View File

@@ -19,7 +19,7 @@
// * Combining BM * BK >= (BM*BN) / (TM*TN) == threadblock yields // * Combining BM * BK >= (BM*BN) / (TM*TN) == threadblock yields
// BM <= BK*TM*TN // BM <= BK*TM*TN
#define BM 64 #define BM 64
#define BN 32 #define BN 64
#define BK 64 #define BK 64
#define WM 16 #define WM 16
#define WN 8 #define WN 8
@@ -33,13 +33,15 @@
// number of loop around the inner 0..TCK..BK loop to simulate perfect-DRAM // number of loop around the inner 0..TCK..BK loop to simulate perfect-DRAM
// scenario // scenario
#define BK_LOOP 1 #define BK_LOOP 1
// whether to transpose smem A tile at GMEM->SMEM (produce), or SMEM->RF // Whether to transpose smem A tile at GMEM->SMEM (produce), or SMEM->RF
// (consume). This is because the tensor core expects the A tile to be stored // (consume). This is because the tensor core expects the A tile to be stored
// in column-major order in SMEM. // in column-major order in SMEM, whereas it is stored row-major in GMEM.
// //
// For correctness, only one of either should be 1. To model the case where // For correctness, only one of either should be 1. To model the case where
// the entire A matrix is already stored transposed in GMEM ("TN" kernel), set // the A matrix is already stored transposed in GMEM ("TN" kernel), set
// both to 0. // both to 0.
//
// For reference, PRODUCE 1 CONSUME 0 generates the performant NN kernel.
#define TRANSPOSE_AT_PRODUCE 1 #define TRANSPOSE_AT_PRODUCE 1
#define TRANSPOSE_AT_CONSUME 0 #define TRANSPOSE_AT_CONSUME 0
// GMEM_COALESCED sets bank conflict-free accesses for // GMEM_COALESCED sets bank conflict-free accesses for