Use L3 BLAS in LARFT #799
Use L3 BLAS in LARFT #799
Conversation
tfalders
added
the
noOptimizations
| @@ -4,6 +4,10 @@ | |||
| * Univ. of Tennessee, Univ. of California Berkeley, | |||
| * Univ. of Colorado Denver and NAG Ltd.. | |||
| * December 2016 | |||
| * and | |||
| * Joffrain, Low, Quintana-Ortí, et al. (2006). Accumulating householder | |||
| * transformations, revisited. | |||
There was a problem hiding this comment.
It would be nice if this line were indented.
There was a problem hiding this comment.
done, please check if I did it correctly.
| Fp[j + i * ldf] *= -tp[i]; | ||
| Fp[j + i * ldf] += -tp[i] * Vp[i + j * ldv]; |
There was a problem hiding this comment.
You've got two writes to global memory here, as well as two reads from tp[i]. You may want to cache tp[i] and Fp[j + i * ldf] * -tp[i] in separate local variables, which will hopefully speed up the kernel.
There was a problem hiding this comment.
changed to reduce redundant reads/writes
| @@ -269,13 +390,50 @@ rocblas_status rocsolver_larft_template(rocblas_handle handle, | |||
| rocblas_fill uplo; | |||
| rocblas_operation trans; | |||
|
|
|||
| const bool call_l3 = larft_do_l3<T>(n, direct, storev) && n > k; | |||
There was a problem hiding this comment.
I'm not a fan of these names. Maybe something like use_gemm and larft_use_gemm.
|
|
||
| if(call_l3) | ||
| { | ||
| if(direct == rocblas_forward_direction && storev == rocblas_column_wise) |
There was a problem hiding this comment.
A short comment explaining what the gemm is doing would be appreciated.
| return value[get_index(intervals, max, dim)]; | ||
| } | ||
|
|
||
| template <typename T, typename I, std::enable_if_t<std::is_same_v<T, rocblas_float_complex>, int> = 0> |
There was a problem hiding this comment.
A short comment describing why this exception for float complex exists would be appreciated.
This PR introduces a potential optimization to the LARFT routine. The modification aims to reduce the size of the gemv computations and instead offloads the block part of the computation to a call to gemm. Additionally, in some cases the modified method performs worse than the original so the latter is dispatched instead.
Following are performance LARFT figures with configuration k=64, rocblas_forward_direction and rocblas_column_wise (config for QR factorization).
column_wise and forward_direction has similar performance characteristic while row_wise configurations show improvements for all data types, e.g.,
This also shows that the gains with the row_wise configuration is more significant (~x18 vs ~x3 speedup).
Curiously, GEQRF shows weird performance with this modification. These figures were generated with square matrices.
Although there is mostly performance gains, there are some cases where performance degrades with the modification.