Add a script for generating the quick-tuning perfconfigs list #1689
Conversation
djramic
force-pushed
the
quicktuning_gen_script
branch
from
bf24e45
to
588592a
Compare
Codecov ReportAttention: Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## develop #1689 +/- ##
===========================================
- Coverage 77.76% 77.59% -0.17%
===========================================
Files 100 100
Lines 27866 27897 +31
Branches 4063 4072 +9
===========================================
- Hits 21671 21648 -23
- Misses 4540 4584 +44
- Partials 1655 1665 +10
Flags with carried forward coverage won't be shown. Click here to find out more. ☔ View full report in Codecov by Sentry. 🚨 Try these New Features:
|
|
|
||
| def get_top_n_perfconfigs_per_problems(self, df, targetColumns): | ||
| """ | ||
| Identifies the top perfcofnigs for each problem based on a threshold |
There was a problem hiding this comment.
perfcofnigs -> perfconfigs
| """ | ||
| Finds the minimal set of perfconfigs that cover all | ||
| problems using set cover optimizaiton. | ||
| Returns : A dictionary containing data types as keys and thier |
There was a problem hiding this comment.
why does this PR change this? and other tests?
| params = {initParameters, nInitParameters}; | ||
| if (opType == KernelType::Gemm) { | ||
| switch (dataTypeA.getIntOrFloatBitWidth()) { | ||
| case 8: |
There was a problem hiding this comment.
this would apply to both f8 and i8 here, is that the goal?
|
|
||
| // BEGIN_GEMM_Wmma_i8_DECS | ||
| static constexpr size_t nInitParametersForward8BitGemm = 15; | ||
| static const InitParamsAccel initParametersForward8BitGemm[nInitParametersForward8BitGemm]; |
There was a problem hiding this comment.
is this only for forward? otherwise we can keep names consistent.
| parser.add_argument("--th", | ||
| required=False, | ||
| type=float, | ||
| default=0.93) |
There was a problem hiding this comment.
is 0.93 the number used here? from what I heard 0.95 is what we consider noise, by using 0.93 we could be getting worse performance?
|
|
||
| # Create coverage matrix | ||
| A = np.zeros((n, m), dtype=int) | ||
| for problem, perfconfig_list in problems_to_perfconfigs.items(): |
There was a problem hiding this comment.
nit: just enumerate here and don't use problem_to_index
| A = np.zeros((n, m), dtype=int) | ||
| for problem, perfconfig_list in problems_to_perfconfigs.items(): | ||
| i = problem_to_index[problem] | ||
| for perfconfig in perfconfig_list: |
| # Linear programming model to minimize the number of perfconfigs | ||
| prob = pulp.LpProblem("SetCoverProblems", pulp.LpMinimize) | ||
| x = pulp.LpVariable.dicts("x", range(m), cat='Binary') | ||
| prob += pulp.lpSum([x[j]] for j in range(m)) |
There was a problem hiding this comment.
nice! could you add a comment here to explain, I think we are setting the objective function here, right?
| x = pulp.LpVariable.dicts("x", range(m), cat='Binary') | ||
| prob += pulp.lpSum([x[j]] for j in range(m)) | ||
| for i in range(n): | ||
| prob += pulp.lpSum([A[i][j] * x[j] |
There was a problem hiding this comment.
same here, I guess these are the constraints?
This PR includes a script for finding quick-tuning perfconfigs in the following way:
Additionally, there are option to automatically generated the QuickTuningPerfconfigs.inc file that contains selected perfconfigs, and its integration into the codebase.
closes : ROCm/rocMLIR-internal#1641
closes : ROCm/rocMLIR-internal#1258
closes : ROCm/rocMLIR-internal#1518