Add cacheable flag to Ruy Matrix so that caller "opts in" to cache behavior on a per-call basis

PiperOrigin-RevId: 286241942
Change-Id: Ie1320c17f6a50468a03dad2664a1c8645e09f3ce

tensorflow/lite/experimental/ruy/dispatch.h

@@ -382,13 +382,15 @@ struct CompileTimeEnabledReferenceMul</*ReferenceMulIsEnabled=*/false> {
   }
 };
 
-inline void HandlePrepackedCaching(TrMulParams* params, Context* context) {
+inline void HandlePrepackedCaching(TrMulParams* params,
+                                   const SidePair<bool>& cacheable,
+                                   Context* context) {
   if (context->cache_policy == CachePolicy::kNoCache) {
     return;
   }
 
   if (context->cache_policy == CachePolicy::kCacheLHSOnGemV) {
-    if (params->dst.layout.cols != 1) {
+    if (!cacheable[Side::kLhs] || params->dst.layout.cols != 1) {
       return;
     }
     PrepackedCache* prepacked_cache = context->GetPrepackedCache();
@@ -465,7 +467,8 @@ void DispatchMul(const Matrix<LhsScalar>& lhs, const Matrix<RhsScalar>& rhs,
   TrMulParams params;
   CreateTrMulParams<TrMulCompiledPaths>(transposed_lhs, rhs, spec, context, dst,
                                         the_path, &params);
-  HandlePrepackedCaching(&params, context);
+  SidePair<bool> cacheable(lhs.cacheable, rhs.cacheable);
+  HandlePrepackedCaching(&params, cacheable, context);
   TrMul(&params, context);
 }
 

tensorflow/lite/experimental/ruy/matrix.h

@@ -108,6 +108,7 @@ template <typename Scalar>
 struct Matrix final {
   Matrix& operator=(const Matrix& other) {
     data = other.data;
+    cacheable = other.cacheable;
     layout = other.layout;
     zero_point = other.zero_point;
     return *this;
@@ -120,6 +121,10 @@ struct Matrix final {
   // The zero_point, i.e. which Scalar value is to be interpreted as zero.
   // When Scalar is floating-point, this must be 0.
   Scalar zero_point = 0;
+  // Clients of Ruy must set this flag to enable any caching behavior. Doesn't
+  // impact numerical results, but caching can impact observable metrics like
+  // latency, memory usage, power, etc.
+  bool cacheable = false;
 };
 
 inline void MakeSimpleLayout(int rows, int cols, Order order, Layout* layout) {

tensorflow/lite/experimental/ruy/prepacked_cache_test.cc

@@ -25,7 +25,6 @@ namespace ruy {
 namespace {
 
 TEST(PrepackedCacheTest, TestCacheEjection) {
-  ruy::Context* context = new ruy::Context();
   // Create the cache.
   PrepackedCache prepacked_cache(32);
   // Allocate the prepacked matrix.
@@ -54,11 +53,9 @@ TEST(PrepackedCacheTest, TestCacheEjection) {
   // The cache size was exceeded by inserting mat2. Ensure that mat1 was
   // ejected.
   EXPECT_EQ(prepacked_cache.FindAndUpdate(cache_key1), prepacked_cache.cend());
-  delete context;
 }
 
 TEST(PrepackedCacheTest, TestCacheBasic) {
-  ruy::Context* context = new ruy::Context();
   // Create the cache.
   PrepackedCache prepacked_cache(48);
   // Allocate the prepacked matrix.
@@ -83,11 +80,9 @@ TEST(PrepackedCacheTest, TestCacheBasic) {
   // The cache size was not exceeded by inserting mat2. Ensure that mat1 was not
   // ejected.
   EXPECT_NE(prepacked_cache.FindAndUpdate(cache_key1), prepacked_cache.cend());
-  delete context;
 }
 
 TEST(PrepackedCacheTest, TestCacheEjection2) {
-  ruy::Context* context = new ruy::Context();
   // Create the cache.
   PrepackedCache prepacked_cache(73);
   // Allocate the prepacked matrix 1.
@@ -137,7 +132,39 @@ TEST(PrepackedCacheTest, TestCacheEjection2) {
   EXPECT_NE(prepacked_cache.FindAndUpdate(cache_key3), prepacked_cache.cend());
   EXPECT_NE(prepacked_cache.FindAndUpdate(cache_key1), prepacked_cache.cend());
   EXPECT_NE(prepacked_cache.FindAndUpdate(cache_key4), prepacked_cache.cend());
-  delete context;
+}
+
+TEST(PrepackedCacheTest, TestCacheOnCacheable) {
+  // Create context and set the cache policy
+  ruy::Context context;
+  context.cache_policy = ruy::kCacheLHSOnGemV;
+  PrepackedCache* cache = context.GetPrepackedCache();
+  EXPECT_EQ(cache->TotalSize(), 0);
+
+  const float lhs_data[] = {1, 2, 3, 4};
+  const float rhs_data[] = {1, 2};
+  float dst_data[4];
+
+  ruy::Matrix<float> lhs;
+  ruy::MakeSimpleLayout(2, 2, ruy::Order::kRowMajor, &lhs.layout);
+  lhs.data = lhs_data;
+  ruy::Matrix<float> rhs;
+  ruy::MakeSimpleLayout(2, 1, ruy::Order::kColMajor, &rhs.layout);
+  rhs.data = rhs_data;
+  ruy::Matrix<float> dst;
+  ruy::MakeSimpleLayout(2, 1, ruy::Order::kColMajor, &dst.layout);
+  dst.data = dst_data;
+
+  ruy::BasicSpec<float, float> spec;
+  // Perform the multiplication and confirm no caching occured.
+  ruy::Mul<ruy::kAllPaths>(lhs, rhs, spec, &context, &dst);
+  EXPECT_EQ(cache->TotalSize(), 0);
+
+  // Set cacheable for the LHS, repeat the multiplication, and see
+  // that caching did occur.
+  lhs.cacheable = true;
+  ruy::Mul<ruy::kAllPaths>(lhs, rhs, spec, &context, &dst);
+  EXPECT_NE(cache->TotalSize(), 0);
 }
 
 }  // namespace