x64: matmul: amx blocking heuristics #2855
Open
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
github-actions
bot
added
the
platform:cpu-x64
|
make test |
yair-obodovsky
force-pushed
the
yobodovs/amx_blocking_heuristics
branch
3 times, most recently
from
af88559 to
0bf3e0d
Compare
|
make test |
yair-obodovsky
force-pushed
the
yobodovs/amx_blocking_heuristics
branch
2 times, most recently
from
241add3 to
33fd053
Compare
yair-obodovsky
changed the title
tczeszun
reviewed
Mar 12, 2025
|
|
||
| float get_bw(int x) { return linear_interpolation(multicore_bw, x); } | ||
|
|
||
| float linear_interpolation(const std::map<int, float> &points, float x) { |
tczeszun
reviewed
Mar 12, 2025
src/cpu/x64/matmul/brgemm_matmul.cpp
Outdated
| const int m_blk_local = m_blk_idx % get_M_chunk_size(); | ||
|
|
||
| int n_blk_local; | ||
| int m_blk_local; |
There was a problem hiding this comment.
assign these to 0 and change below if-else to enter if only else condition is met
tczeszun
reviewed
Mar 12, 2025
src/cpu/x64/matmul/brgemm_matmul.cpp
Outdated
| return status::unimplemented; | ||
| } | ||
| } | ||
| return status::success; |
tczeszun
reviewed
Mar 12, 2025
| } | ||
| } | ||
|
|
||
| // printf("winner: m_div=%d, k_div=%d, n_div=%d, b_div=%d, score=%f\n", best_blocking.nthr_m_, best_blocking.nthr_k_, best_blocking.nthr_n_,best_blocking.nthr_b_, best_blocking.efficiency_score_); |
tczeszun
reviewed
Mar 12, 2025
| //don't do reduction if c tmp doesn't fit | ||
| //also parallel reduction is not supported for large batch. This conforms with the assert in brgemm_matmul.cpp: | ||
| //assert(IMPLICATION(parallel_reduction_is_used(), | ||
| // bgmmc.batch == 1 && !calculate_compensations_in_copy_routines)); |
tczeszun
reviewed
Mar 12, 2025
| size_t B_chunk_sz = b_dt_sz * k_chunk_elems_ * n_chunk_elems_; | ||
| size_t B_buf_sz = use_buffer_b ? tr_b_dt_sz * n_blk_ * k_chunk_elems_ : 0; | ||
| size_t C_chunk_sz = c_dt_sz * m_chunk_elems_ * n_chunk_elems_; | ||
| size_t C_buf_sz |
There was a problem hiding this comment.
all these variables can be const
tczeszun
reviewed
Mar 12, 2025
| * nthr_mnb_; | ||
| float k_parallel_score = 1.0f; | ||
| if (nthr_k_ > 1) { | ||
| dim_t num_K_chunks = div_up(K, k_chunk_elems_); |
There was a problem hiding this comment.
can be const, same for M, N
tczeszun
reviewed
Mar 12, 2025
| dim_t largest_k_tiles = largest_k / this->k_tmul; | ||
| dim_t k_tiles = div_up(K, this->k_tmul); | ||
| dim_t k_per_thread_tiles = div_up(k_tiles, nthr_k_); | ||
| dim_t num_K_blocks = div_up(k_per_thread_tiles, largest_k_tiles); |
There was a problem hiding this comment.
all variables can be const
yair-obodovsky
force-pushed
the
yobodovs/amx_blocking_heuristics
branch
from
33fd053 to
527e4fc
Compare
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
New blocking heuristics for matrix multiplication using AMX include a new k-loop, utilize blocked C, and employ larger N dimensions in BRGEM with LDB2