Fix cuda cache same key different value check

An attemp to cache a kernel that already exists in the cache with
the same key (options, input shapes, ...) now properly throws if
the stored values (block/grid sizes, cuda code, ...) are different

Author: Theodoros Theodoridis

tc/core/cuda/cuda_compilation_cache.cc

@@ -198,9 +198,9 @@ void CudaCache::cacheKernel(CudaCachedEntry&& entry) {
       entry.key.inputs,
       entry.key.outputs);
   if (retrievedEntry) {
-    if (retrievedEntry->values.cudaSource == entry.values.cudaSource or
-        retrievedEntry->values.grid == entry.values.grid or
-        retrievedEntry->values.block == entry.values.block) {
+    if (retrievedEntry->values.cudaSource != entry.values.cudaSource or
+        retrievedEntry->values.grid != entry.values.grid or
+        retrievedEntry->values.block != entry.values.block) {
       throw CacheEntrySameKeyDifferentValue(
           "CudaCache::CacheKernel: a kernel matching the id, options and "
           "inputs was previously cached with different cuda source or block "

tc/core/cuda/cuda_mapping_options.h

@@ -105,6 +105,10 @@ class CudaDim {
     return view == other.view;
   }
 
+  inline bool operator!=(const CudaDim& other) const {
+    return not(*this == other);
+  }
+
  private:
   CudaDimProto ownedProto_;
 
@@ -122,6 +126,7 @@ class Block : public CudaDim {
   Block(std::vector<uint64_t> il) : CudaDim(il) {}
 
   using CudaDim::operator=;
+  using CudaDim::operator!=;
 };
 
 /// Specializing CudaDim to differentiate between Block and Grid sizes.
@@ -134,6 +139,7 @@ class Grid : public CudaDim {
   Grid(std::vector<uint64_t> il) : CudaDim(il) {}
 
   using CudaDim::operator=;
+  using CudaDim::operator!=;
 };
 
 class CudaMappingOptions {

test/cuda/test_compilation_cache.cc

@@ -69,6 +69,42 @@ class CudaCacheTest : public ::testing::Test {
   }
 };
 
+TEST_F(CudaCacheTest, EntrySameKeyDifferentValue) {
+  auto options = tc::CudaMappingOptions::makeNaiveCudaMappingOptions();
+  auto inputPtrs = InputPtrs();
+  auto outputPtrs = InputPtrs();
+
+  tc::CudaCache::getCache()->cacheKernel(tc::CudaCachedEntry(
+      "kernel0",
+      "ker000",
+      {0, 0, 1},
+      tc::Grid(std::vector<size_t>{1, 1, 1}),
+      tc::Block(std::vector<size_t>{1, 1, 1}),
+      options,
+      inputPtrs,
+      outputPtrs,
+      "source0",
+      tc::CudaGPUInfo::GPUInfo().GetCudaDeviceStr()));
+  ASSERT_THROW(
+      tc::CudaCache::getCache()->cacheKernel(tc::CudaCachedEntry(
+          "kernel0",
+          "ker000",
+          {0, 0, 1},
+          tc::Grid(std::vector<size_t>{2, 1, 1}),
+          tc::Block(std::vector<size_t>{1, 2, 1}),
+          options,
+          inputPtrs,
+          outputPtrs,
+          "source1",
+          tc::CudaGPUInfo::GPUInfo().GetCudaDeviceStr())),
+      tc::CacheEntrySameKeyDifferentValue);
+
+  ASSERT_EQ(tc::CudaCache::getCache()->size(), 1);
+  ASSERT_EQ(tc::CudaCache::getCache()->numberAttemptedRetrievals, 0);
+  ASSERT_EQ(tc::CudaCache::getCache()->numberSuccessfulRetrievals, 0);
+  ASSERT_EQ(tc::CudaCache::getCache()->numberCacheAttemps, 2);
+}
+
 TEST_F(CudaCacheTest, DifferentIDs) {
   auto options = tc::CudaMappingOptions::makeNaiveCudaMappingOptions();
   auto inputPtrs = InputPtrs();