Add a new memory property: eDeviceAndLocal (#396)

This is memory which is device local, and also host visible and host coherent. Signed-off-by: Robert Quill <robert.quill@imgtec.com>
KomputeProject · Sep 12, 2024 · f190212 · f190212
1 parent a242929
commit f190212
Show file tree

Hide file tree

Showing 10 changed files with 130 additions and 3 deletions.
diff --git a/src/Image.cpp b/src/Image.cpp
@@ -442,6 +442,7 @@ Image::getPrimaryImageUsageFlags()
     switch (this->mMemoryType) {
         case MemoryTypes::eDevice:
         case MemoryTypes::eHost:
+        case MemoryTypes::eDeviceAndHost:
             return vk::ImageUsageFlagBits::eStorage |
                    vk::ImageUsageFlagBits::eTransferSrc |
                    vk::ImageUsageFlagBits::eTransferDst;

diff --git a/src/Memory.cpp b/src/Memory.cpp
@@ -157,7 +157,8 @@ Memory::mapRawData()
 
     std::shared_ptr<vk::DeviceMemory> hostVisibleMemory = nullptr;
 
-    if (this->mMemoryType == MemoryTypes::eHost) {
+    if (this->mMemoryType == MemoryTypes::eHost ||
+        this->mMemoryType == MemoryTypes::eDeviceAndHost) {
         hostVisibleMemory = this->mPrimaryMemory;
     } else if (this->mMemoryType == MemoryTypes::eDevice) {
         hostVisibleMemory = this->mStagingMemory;
@@ -187,7 +188,8 @@ Memory::unmapRawData()
 
     std::shared_ptr<vk::DeviceMemory> hostVisibleMemory = nullptr;
 
-    if (this->mMemoryType == MemoryTypes::eHost) {
+    if (this->mMemoryType == MemoryTypes::eHost ||
+        this->mMemoryType == MemoryTypes::eDeviceAndHost) {
         hostVisibleMemory = this->mPrimaryMemory;
     } else if (this->mMemoryType == MemoryTypes::eDevice) {
         hostVisibleMemory = this->mStagingMemory;
@@ -226,6 +228,10 @@ Memory::getPrimaryMemoryPropertyFlags()
             return vk::MemoryPropertyFlagBits::eHostVisible |
                    vk::MemoryPropertyFlagBits::eHostCoherent;
             break;
+        case MemoryTypes::eDeviceAndHost:
+            return vk::MemoryPropertyFlagBits::eDeviceLocal |
+                   vk::MemoryPropertyFlagBits::eHostVisible |
+                   vk::MemoryPropertyFlagBits::eHostCoherent;
         case MemoryTypes::eStorage:
             return vk::MemoryPropertyFlagBits::eDeviceLocal;
             break;

diff --git a/src/Tensor.cpp b/src/Tensor.cpp
@@ -282,12 +282,14 @@ Tensor::constructDescriptorSet(vk::DescriptorSet descriptorSet,
 
     return writeDesciptorSet;
 }
+
 vk::BufferUsageFlags
 Tensor::getPrimaryBufferUsageFlags()
 {
     switch (this->mMemoryType) {
         case MemoryTypes::eDevice:
         case MemoryTypes::eHost:
+        case MemoryTypes::eDeviceAndHost:
             return vk::BufferUsageFlagBits::eStorageBuffer |
                    vk::BufferUsageFlagBits::eTransferSrc |
                    vk::BufferUsageFlagBits::eTransferDst;

diff --git a/src/include/kompute/Image.hpp b/src/include/kompute/Image.hpp
@@ -120,7 +120,8 @@ class Image : public Memory
               "Custom data types are not supported for Kompute Images");
         }
 
-        if (memoryType == MemoryTypes::eHost) {
+        if (memoryType == MemoryTypes::eHost ||
+            memoryType == MemoryTypes::eDeviceAndHost) {
             // Host-accessible memory must be linear-tiled.
             tiling = vk::ImageTiling::eLinear;
         } else if (memoryType == MemoryTypes::eDevice ||

diff --git a/src/include/kompute/Memory.hpp b/src/include/kompute/Memory.hpp
@@ -29,6 +29,8 @@ class Memory
         eDevice = 0,  ///< Type is device memory, source and destination
         eHost = 1,    ///< Type is host memory, source and destination
         eStorage = 2, ///< Type is Device memory (only)
+        eDeviceAndHost =
+          3, ///< Type is host-visible and host-coherent device memory
     };
 
     enum class DataTypes

diff --git a/test/TestOpCopyImage.cpp b/test/TestOpCopyImage.cpp
@@ -374,3 +374,27 @@ TEST(TestOpCopyImage, CopyDeviceToDeviceImage2DMismatchedSizes)
     EXPECT_THROW(mgr.sequence()->eval<kp::OpCopy>({ imageA, imageB }),
                  std::runtime_error);
 }
+
+TEST(TestOpImageCopy, CopyDeviceAndHostToDeviceAndHostImage)
+{
+    kp::Manager mgr;
+
+    std::vector<float> testVecA{ 1, 2, 3 };
+    std::vector<float> testVecB{ 0, 0, 0 };
+
+    std::shared_ptr<kp::ImageT<float>> imageA =
+      mgr.image(testVecA, 3, 1, 1, kp::Memory::MemoryTypes::eDeviceAndHost);
+    std::shared_ptr<kp::ImageT<float>> imageB =
+      mgr.image(testVecB, 3, 1, 1, kp::Memory::MemoryTypes::eDeviceAndHost);
+
+    EXPECT_TRUE(imageA->isInit());
+    EXPECT_TRUE(imageB->isInit());
+
+    mgr.sequence()
+      ->eval<kp::OpSyncDevice>({ imageA, imageB })
+      ->eval<kp::OpCopy>({ imageA, imageB })
+      ->eval<kp::OpSyncLocal>({ imageA, imageB });
+
+    // Making sure the GPU holds the same vector
+    EXPECT_EQ(imageA->vector(), imageB->vector());
+}
diff --git a/test/TestOpCopyImageToTensor.cpp b/test/TestOpCopyImageToTensor.cpp
@@ -212,3 +212,29 @@ TEST(TestOpCopyImageToTensor, CopyDeviceToDeviceImageUninitialised)
     // Making sure the GPU holds the same vector
     EXPECT_EQ(imageA->vector(), tensorB->vector());
 }
+
+TEST(TestOpImageCopyToTensor, CopyDeviceAndHostToDeviceAndHostTensor)
+{
+    kp::Manager mgr;
+
+    std::vector<float> testVecA{ 1, 2, 3 };
+    std::vector<float> testVecB{ 0, 0, 0 };
+
+    std::shared_ptr<kp::ImageT<float>> imageA = mgr.image(
+      testVecA, testVecA.size(), 1, 1, kp::Memory::MemoryTypes::eDeviceAndHost);
+    std::shared_ptr<kp::TensorT<float>> tensorB =
+      mgr.tensor(testVecB, kp::Memory::MemoryTypes::eDeviceAndHost);
+
+    EXPECT_TRUE(imageA->isInit());
+    EXPECT_TRUE(tensorB->isInit());
+
+    mgr.sequence()
+      ->eval<kp::OpSyncDevice>({ imageA })
+      ->eval<kp::OpSyncDevice>({ tensorB })
+      ->eval<kp::OpCopy>({ imageA, tensorB })
+      ->eval<kp::OpSyncLocal>({ imageA })
+      ->eval<kp::OpSyncLocal>({ tensorB });
+
+    // Making sure the GPU holds the same vector
+    EXPECT_EQ(imageA->vector(), tensorB->vector());
+}
diff --git a/test/TestOpCopyTensor.cpp b/test/TestOpCopyTensor.cpp
@@ -358,3 +358,27 @@ TEST(TestOpCopyTensor, CopyTensorThroughStorageViaAlgorithmsUninitialisedOutput)
     // Making sure the GPU holds the same vector
     EXPECT_EQ(tensorIn->vector(), tensorOut->vector());
 }
+
+TEST(TestOpTensorCopy, CopyDeviceAndHostToDeviceAndHostTensor)
+{
+    kp::Manager mgr;
+
+    std::vector<float> testVecA{ 1, 2, 3 };
+    std::vector<float> testVecB{ 0, 0, 0 };
+
+    std::shared_ptr<kp::TensorT<float>> tensorA =
+      mgr.tensor(testVecA, kp::Memory::MemoryTypes::eDeviceAndHost);
+    std::shared_ptr<kp::TensorT<float>> tensorB =
+      mgr.tensor(testVecB, kp::Memory::MemoryTypes::eDeviceAndHost);
+
+    EXPECT_TRUE(tensorA->isInit());
+    EXPECT_TRUE(tensorB->isInit());
+
+    mgr.sequence()
+      ->eval<kp::OpSyncDevice>({ tensorA, tensorB })
+      ->eval<kp::OpCopy>({ tensorA, tensorB })
+      ->eval<kp::OpSyncLocal>({ tensorA, tensorB });
+
+    // Making sure the GPU holds the same vector
+    EXPECT_EQ(tensorA->vector(), tensorB->vector());
+}
diff --git a/test/TestOpCopyTensorToImage.cpp b/test/TestOpCopyTensorToImage.cpp
@@ -213,3 +213,29 @@ TEST(TestOpCopyTensorToImage, CopyDeviceToDeviceImageUninitialised)
     // Making sure the GPU holds the same vector
     EXPECT_EQ(tensorA->vector(), imageB->vector());
 }
+
+TEST(TestOpTensorCopyToImage, CopyDeviceAndHostToDeviceAndHostTensor)
+{
+    kp::Manager mgr;
+
+    std::vector<float> testVecA{ 1, 2, 3 };
+    std::vector<float> testVecB{ 0, 0, 0 };
+
+    std::shared_ptr<kp::TensorT<float>> tensorA =
+      mgr.tensor(testVecA, kp::Memory::MemoryTypes::eDeviceAndHost);
+    std::shared_ptr<kp::ImageT<float>> imageB = mgr.image(
+      testVecB, testVecB.size(), 1, 1, kp::Memory::MemoryTypes::eDeviceAndHost);
+
+    EXPECT_TRUE(tensorA->isInit());
+    EXPECT_TRUE(imageB->isInit());
+
+    mgr.sequence()
+      ->eval<kp::OpSyncDevice>({ tensorA })
+      ->eval<kp::OpSyncDevice>({ imageB })
+      ->eval<kp::OpCopy>({ tensorA, imageB })
+      ->eval<kp::OpSyncLocal>({ tensorA })
+      ->eval<kp::OpSyncLocal>({ imageB });
+
+    // Making sure the GPU holds the same vector
+    EXPECT_EQ(tensorA->vector(), imageB->vector());
+}
diff --git a/test/TestOpImageCreate.cpp b/test/TestOpImageCreate.cpp
@@ -64,6 +64,21 @@ TEST(TestOpImageCreate, ExceptionOnInvalidTiledImage)
 
     kp::Manager mgr;
 
+    try {
+        std::shared_ptr<kp::ImageT<float>> imageA =
+          mgr.image(testVecA,
+                    1,
+                    1,
+                    1,
+                    vk::ImageTiling::eOptimal,
+                    kp::Memory::MemoryTypes::eDeviceAndHost);
+    } catch (const std::runtime_error& err) {
+        // check exception
+        ASSERT_TRUE(std::string(err.what())
+                      .find("optimal tiling is only supported for") !=
+                    std::string::npos);
+    }
+
     try {
         std::shared_ptr<kp::ImageT<float>> imageA =
           mgr.image(testVecA,