[WIP] forward GEMM workloads to GEMV when one argument is actually a vector #4708
Conversation
|
obviously I don't really intend to kick out the recent Loongson patch here - this first draft was thrown together off-grid in an outdated fork |
CodSpeed Performance ReportMerging #4708 will not alter performanceComparing Summary
|
There was a problem hiding this comment.
Thank you for posting this WIP PR. I was able to confirm the failures CI sees:
******* FATAL ERROR - COMPUTED RESULT IS LESS THAN HALF ACCURATE *******
EXPECTED RESULT COMPUTED RESULT
1 0.00000 -0.252747
******* SGEMM FAILED ON CALL NUMBER:
3403: SGEMM ('N','N', 1, 1, 0, 0.0, A, 2, B, 1, 0.0, C, 2).
I haven't had time to track it down, but I appreciate the starting point!
| char *NT=(char*)malloc(2*sizeof(char)); | ||
| if (transb&1)strcpy(NT,"T"); | ||
| else NT="N"; |
There was a problem hiding this comment.
I think we can skip the malloc and go straight to the NT="N" or NT="T" assignment yes?
Actually, we don't need null termination so it should just be char NT; NT='N'; ... &NT.
| // fprintf(stderr,"G E M V ! ! ! lda=%d ldb=%d ldc=%d\n",args.lda,args.ldb,args.ldc); | ||
| GEMV(NT, &args.m ,&args.k, args.alpha, args.a, &args.lda, args.b, &args.n, args.beta, args.c, &args.n); | ||
| //SUBROUTINE SGEMV(TRANS,M,N,ALPHA,A,LDA,X,INCX,BETA,Y,INCY) |
There was a problem hiding this comment.
I was thinking about this a little more. For cases when n = 1, seems like this should be pretty straightforward, and parameters should just directly translate from GEMM to GEMV.
For cases when m = 1 I think there is a challenge. in gemm (C:=AB) A is the vector, but GEMV (y:=Bx). so I think we need to re-arrange the equation by transposing (or, un-transposing) B.
|
@martin-frbg , |
Yes in principle, but I am not convinced we actually have to transform the storage of A for that. (Note that the rough draft I posted here may not even compile. I need to update it and flesh it out when I have time) |
fixes #4580 and fixes #528