-
Notifications
You must be signed in to change notification settings - Fork 62
Tensor::gemm: arena scale fast paths serve both operand orientations #582
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Merged
+197
−31
Merged
Changes from all commits
Commits
Show all changes
3 commits
Select commit
Hold shift + click to select a range
File filter
Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
There are no files selected for viewing
This file contains hidden or 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
This file contains hidden or 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
This file contains hidden or 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
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,128 @@ | ||
| // tests/arena_mixed_product_order.cpp | ||
| // | ||
| // DistArray-level regression for MIXED products (an arena tensor-of- | ||
| // tensors operand times a plain tensor operand), found on the CSV-CC | ||
| // per-batch intermediate (occ,occ,PAO;PNO) * (PAO,PAO,DF) -> | ||
| // (occ,occ,PAO,DF;PNO): | ||
| // | ||
| // 1. operand order: the expression layer's cost of a mixed product depended | ||
| // on which operand came first (nested-first ~2.5x faster on two tiles | ||
| // per mode, ~4.5x on one tile, regardless of the result order or an | ||
| // imposed shape). Cause: Tensor::gemm's two arena "scale" strided-GEMM | ||
| // fast paths (nested-left rows, nested-right columns) accepted only the | ||
| // NoTranspose/NoTranspose orientation, and the permutation optimizer | ||
| // hands a nested-right operand whose contracted mode trails to the GEMM | ||
| // as an implicit transpose -- so that order fell to the per-cell AXPY | ||
| // loop. Both paths now serve either orientation (the plain matrix's | ||
| // transpose goes to BLAS as its op; a transposed nested tile only changes | ||
| // which cells form a row/column). The test asserts that all spellings | ||
| // (both orders, plain and shaped, both result orders) agree and reports | ||
| // the wall time of each (informational; no timing assertion). | ||
| // | ||
| // (A second case, a permuted arena operand consumed inside a binary | ||
| // expression, lands with the fix for that defect.) | ||
|
|
||
| #include "TiledArray/tensor/arena_einsum.h" | ||
| #include "TiledArray/tensor/arena_tensor.h" | ||
| #include "tiledarray.h" | ||
| #include "unit_test_config.h" | ||
|
|
||
| #include <chrono> | ||
| #include <cmath> | ||
| #include <iostream> | ||
| #include <string> | ||
|
|
||
| namespace TA = TiledArray; | ||
|
|
||
| namespace { | ||
|
|
||
| using ArenaInner = TA::ArenaTensor<double, TA::Range>; | ||
| using ArenaOuter = TA::Tensor<ArenaInner>; | ||
| using ToTArray = TA::DistArray<ArenaOuter, TA::SparsePolicy>; | ||
| using FlatArray = TA::DistArray<TA::Tensor<double>, TA::SparsePolicy>; | ||
|
|
||
| struct MixedFixture { | ||
| long const I = 8, P = 24, K = 10, A = 6; | ||
| TA::TiledRange1 const tr_i{0l, I / 2, I}, tr_p{0l, P / 2, P}, tr_k{0l, K}; | ||
| TA::TiledRange const ij_p{tr_i, tr_i, tr_p}; | ||
| TA::TiledRange const q_p_k{tr_p, tr_p, tr_k}; | ||
| TA::TiledRange const ijqk{tr_i, tr_i, tr_p, tr_k}; | ||
| std::string const ta = "i,j,p;a", tb = "q,p,k", tc = "i,j,q,k;a", | ||
| tc_alt = "q,k,i,j;a"; | ||
| ToTArray tot; | ||
| FlatArray flat; | ||
|
|
||
| MixedFixture() { | ||
| auto& world = TA::get_default_world(); | ||
| tot = ToTArray(world, ij_p); | ||
| tot.init_tiles([&](TA::Range const& tr) { | ||
| ArenaOuter t = TA::detail::arena_outer_init<ArenaOuter>( | ||
| tr, 1, [A = this->A](std::size_t) { return TA::Range{A}; }); | ||
| for (std::size_t o = 0; o < t.range().volume(); ++o) { | ||
| ArenaInner& c = t.data()[o]; | ||
| if (!c) continue; | ||
| for (long a = 0; a < A; ++a) | ||
| c.data()[a] = 1.0 + 0.001 * static_cast<double>((o * 7 + a) % 13); | ||
| } | ||
| return t; | ||
| }); | ||
| flat = FlatArray(world, q_p_k); | ||
| flat.init_tiles([&](TA::Range const& tr) { | ||
| TA::Tensor<double> t(tr); | ||
| for (std::size_t o = 0; o < t.range().volume(); ++o) | ||
| t.data()[o] = 0.5 + 0.001 * static_cast<double>(o % 17); | ||
| return t; | ||
| }); | ||
| world.gop.fence(); | ||
| } | ||
|
|
||
| template <typename F> | ||
| static double timed(F&& f) { | ||
| auto& world = TA::get_default_world(); | ||
| world.gop.fence(); | ||
| auto const t0 = std::chrono::steady_clock::now(); | ||
| f(); | ||
| world.gop.fence(); | ||
| return std::chrono::duration<double>(std::chrono::steady_clock::now() - t0) | ||
| .count(); | ||
| } | ||
|
|
||
| double diff(ToTArray const& x, ToTArray const& ref, std::string const& xa) { | ||
| ToTArray d; | ||
| d(tc) = x(xa) - ref(tc); | ||
| return TA::norm2(d); | ||
| } | ||
| }; | ||
|
|
||
| } // namespace | ||
|
|
||
| BOOST_FIXTURE_TEST_SUITE(arena_mixed_product_order_suite, MixedFixture, | ||
| TA_UT_LABEL_SERIAL) | ||
|
|
||
| BOOST_AUTO_TEST_CASE(operand_order_is_canonicalized) { | ||
| ToTArray ref, ff, nf_sh, ff_sh, nf_alt, ff_alt; | ||
| double const t_nf = timed([&] { ref(tc) = tot(ta) * flat(tb); }); | ||
| double const t_ff = timed([&] { ff(tc) = flat(tb) * tot(ta); }); | ||
| TA::Tensor<float> norms(ijqk.tiles_range(), 1.0f); | ||
| TA::SparseShape<float> const shape(norms, ijqk, /*do_not_scale=*/true); | ||
| double const t_nf_sh = | ||
| timed([&] { nf_sh(tc) = (tot(ta) * flat(tb)).set_shape(shape); }); | ||
| double const t_ff_sh = | ||
| timed([&] { ff_sh(tc) = (flat(tb) * tot(ta)).set_shape(shape); }); | ||
| double const t_nf_alt = timed([&] { nf_alt(tc_alt) = tot(ta) * flat(tb); }); | ||
| double const t_ff_alt = timed([&] { ff_alt(tc_alt) = flat(tb) * tot(ta); }); | ||
| double const n = TA::norm2(ref); | ||
| BOOST_REQUIRE(n > 0.0); | ||
| double const tol = 1e-12 * n; | ||
| BOOST_CHECK_SMALL(diff(ff, ref, tc), tol); | ||
| BOOST_CHECK_SMALL(diff(nf_sh, ref, tc), tol); | ||
| BOOST_CHECK_SMALL(diff(ff_sh, ref, tc), tol); | ||
| BOOST_CHECK_SMALL(diff(nf_alt, ref, tc_alt), tol); | ||
| BOOST_CHECK_SMALL(diff(ff_alt, ref, tc_alt), tol); | ||
| BOOST_TEST_MESSAGE("mixed product wall time (s): nested*flat=" | ||
| << t_nf << " flat*nested=" << t_ff | ||
| << " +shape: " << t_nf_sh << " / " << t_ff_sh | ||
| << " result q,k,i,j: " << t_nf_alt << " / " << t_ff_alt); | ||
| } | ||
|
|
||
| BOOST_AUTO_TEST_SUITE_END() | ||
Oops, something went wrong.
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.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
cafebd6 adds the explicit include (line 31). With the second case moved to #583 nothing in this file calls std::isfinite any more, so the include is now only precautionary.