From e87dd5bdeda6164b6f908818b7d69b39c436638d Mon Sep 17 00:00:00 2001
From: Olivier BICHLER <olivier.bichler@cea.fr>
Date: Fri, 8 Nov 2024 10:15:06 +0100
Subject: [PATCH] Working 2D tiling
---
unit_tests/recipies/Test_MatMulTiling.cpp | 13 +++++++++++++
1 file changed, 13 insertions(+)
diff --git a/unit_tests/recipies/Test_MatMulTiling.cpp b/unit_tests/recipies/Test_MatMulTiling.cpp
index 4920dc63..46d5418f 100644
--- a/unit_tests/recipies/Test_MatMulTiling.cpp
+++ b/unit_tests/recipies/Test_MatMulTiling.cpp
@@ -22,6 +22,7 @@
#include "aidge/operator/Producer.hpp"
#include "aidge/graph/OpArgs.hpp"
#include "aidge/scheduler/SequentialScheduler.hpp"
+#include "aidge/graph/Matching.hpp"
#include "aidge/utils/TensorUtils.hpp"
using namespace Aidge;
@@ -79,10 +80,22 @@ TEST_CASE("[MatMulTiling]") {
// Tiling
fmt::println("Tiling");
matMulTiling(matmul1, {16, 16});
+ removeIdentity(g1);
g1->setBackend("cpu");
g1->save("MatMulSplitting_graph_split");
+ auto gm = SinglePassGraphMatching(g1);
+ gm.addNodeLambda("16x16", [](const NodePtr& node) {
+ const auto op =
+ std::static_pointer_cast<OperatorTensor>(node->getOperator());
+ const auto dims = op->getOutput(0)->dims();
+ return (dims.end()[-2] == 16 && dims.end()[-1] == 16);
+ });
+
+ const auto results = gm.match("MatMul[16x16]");
+ REQUIRE(results.size() == 25);
+
// Check result
fmt::println("Schedule forward tiled graph");
s1 = SequentialScheduler(g1);
--
GitLab