Vulkan: Make staging buffer per context

Right now staging buffers are per BufferVk. This will make it per
ContextVk so that it can be shared among all objects that needs a
staging buffer.

Bug: b/161846868
Change-Id: I9c436acdacaf429a43cbdfa216927df0796f7a28
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/2310962
Reviewed-by: Tim Van Patten <timvp@google.com>
Reviewed-by: Courtney Goeltzenleuchter <courtneygo@google.com>
Reviewed-by: Ian Elliott <ianelliott@google.com>
Reviewed-by: Jamie Madill <jmadill@chromium.org>
Commit-Queue: Charlie Lao <cclao@google.com>

Author: cclao

src/libANGLE/renderer/vulkan/BufferVk.cpp

@@ -33,33 +33,6 @@ static_assert(gl::isPow2(kBufferSizeGranularity), "use as alignment, must be pow
 // Start with a fairly small buffer size. We can increase this dynamically as we convert more data.
 constexpr size_t kConvertedArrayBufferInitialSize = 1024 * 8;
 
-// Base size for all staging buffers
-constexpr size_t kStagingBufferBaseSize = 1024;
-// Fix the staging buffer size multiplier for unpack buffers, for now
-constexpr size_t kUnpackBufferStagingBufferMultiplier = 1024;
-
-size_t CalculateStagingBufferSize(gl::BufferBinding target, size_t size, size_t alignment)
-{
-    size_t alignedSize = rx::roundUp(size, alignment);
-    int multiplier     = std::max(gl::log2(alignedSize), 1);
-
-    switch (target)
-    {
-        case gl::BufferBinding::Array:
-        case gl::BufferBinding::DrawIndirect:
-        case gl::BufferBinding::ElementArray:
-        case gl::BufferBinding::Uniform:
-            return kStagingBufferBaseSize * multiplier;
-
-        case gl::BufferBinding::PixelUnpack:
-            return std::max(alignedSize,
-                            (kStagingBufferBaseSize * kUnpackBufferStagingBufferMultiplier));
-
-        default:
-            return kStagingBufferBaseSize;
-    }
-}
-
 // Buffers that have a static usage pattern will be allocated in
 // device local memory to speed up access to and from the GPU.
 // Dynamic usage patterns or that are frequently mapped
@@ -163,7 +136,6 @@ void BufferVk::destroy(const gl::Context *context)
 void BufferVk::release(ContextVk *contextVk)
 {
     RendererVk *renderer = contextVk->getRenderer();
-    mStagingBuffer.release(renderer);
     mShadowBuffer.release();
     mBufferPool.release(renderer);
     mBuffer = nullptr;
@@ -174,18 +146,6 @@ void BufferVk::release(ContextVk *contextVk)
     }
 }
 
-void BufferVk::initializeStagingBuffer(ContextVk *contextVk, gl::BufferBinding target, size_t size)
-{
-    RendererVk *rendererVk = contextVk->getRenderer();
-
-    constexpr VkImageUsageFlags kBufferUsageFlags = VK_BUFFER_USAGE_TRANSFER_SRC_BIT;
-    size_t alignment =
-        static_cast<size_t>(rendererVk->getPhysicalDeviceProperties().limits.minMemoryMapAlignment);
-    size_t stagingBufferSize = CalculateStagingBufferSize(target, size, alignment);
-
-    mStagingBuffer.init(rendererVk, kBufferUsageFlags, alignment, stagingBufferSize, true);
-}
-
 angle::Result BufferVk::initializeShadowBuffer(ContextVk *contextVk,
                                                gl::BufferBinding target,
                                                size_t size)
@@ -258,9 +218,6 @@ angle::Result BufferVk::setData(const gl::Context *context,
 
         ANGLE_TRY(acquireBufferHelper(contextVk, size, &mBuffer));
 
-        // Initialize the staging buffer
-        initializeStagingBuffer(contextVk, target, size);
-
         // Initialize the shadow buffer
         ANGLE_TRY(initializeShadowBuffer(contextVk, target, size));
     }
@@ -530,29 +487,23 @@ angle::Result BufferVk::stagedUpdate(ContextVk *contextVk,
                                      size_t offset)
 {
     // Acquire a "new" staging buffer
-    bool needToReleasePreviousBuffers = false;
-    uint8_t *mapPointer               = nullptr;
-    VkDeviceSize stagingBufferOffset  = 0;
+    uint8_t *mapPointer              = nullptr;
+    VkDeviceSize stagingBufferOffset = 0;
 
-    ANGLE_TRY(mStagingBuffer.allocate(contextVk, size, &mapPointer, nullptr, &stagingBufferOffset,
-                                      &needToReleasePreviousBuffers));
-    if (needToReleasePreviousBuffers)
-    {
-        // Release previous staging buffers
-        mStagingBuffer.releaseInFlightBuffers(contextVk);
-    }
+    vk::DynamicBuffer *stagingBuffer = contextVk->getStagingBufferStorage();
+    ANGLE_TRY(stagingBuffer->allocate(contextVk, size, &mapPointer, nullptr, &stagingBufferOffset,
+                                      nullptr));
     ASSERT(mapPointer);
 
     memcpy(mapPointer, data, size);
-    ASSERT(!mStagingBuffer.isCoherent());
-    ANGLE_TRY(mStagingBuffer.flush(contextVk));
-    mStagingBuffer.getCurrentBuffer()->onExternalWrite(VK_ACCESS_HOST_WRITE_BIT);
+    ASSERT(!stagingBuffer->isCoherent());
+    ANGLE_TRY(stagingBuffer->flush(contextVk));
+    stagingBuffer->getCurrentBuffer()->onExternalWrite(VK_ACCESS_HOST_WRITE_BIT);
 
     // Enqueue a copy command on the GPU.
     VkBufferCopy copyRegion = {stagingBufferOffset, offset, size};
     ANGLE_TRY(
-        mBuffer->copyFromBuffer(contextVk, mStagingBuffer.getCurrentBuffer(), 1, &copyRegion));
-    mStagingBuffer.getCurrentBuffer()->retain(&contextVk->getResourceUseList());
+        mBuffer->copyFromBuffer(contextVk, stagingBuffer->getCurrentBuffer(), 1, &copyRegion));
 
     return angle::Result::Continue;
 }

src/libANGLE/renderer/vulkan/BufferVk.h

@@ -121,7 +121,6 @@ class BufferVk : public BufferImpl
                                                 bool hostVisible);
 
   private:
-    void initializeStagingBuffer(ContextVk *contextVk, gl::BufferBinding target, size_t size);
     angle::Result initializeShadowBuffer(ContextVk *contextVk,
                                          gl::BufferBinding target,
                                          size_t size);
@@ -182,9 +181,6 @@ class BufferVk : public BufferImpl
     // Pool of BufferHelpers for mBuffer to acquire from
     vk::DynamicBuffer mBufferPool;
 
-    // All staging buffer support is provided by a DynamicBuffer.
-    vk::DynamicBuffer mStagingBuffer;
-
     // For GPU-read only buffers glMap* latency is reduced by maintaining a copy
     // of the buffer which is writeable only by the CPU. The contents are updated on all
     // glData/glSubData/glCopy calls. With this, a glMap* call becomes a non-blocking

src/libANGLE/renderer/vulkan/ContextVk.cpp

@@ -752,6 +752,7 @@ void ContextVk::onDestroy(const gl::Context *context)
 
     mDefaultUniformStorage.release(mRenderer);
     mEmptyBuffer.release(mRenderer);
+    mStagingBufferStorage.release(mRenderer);
 
     for (vk::DynamicBuffer &defaultBuffer : mDefaultAttribBuffers)
     {
@@ -906,6 +907,13 @@ angle::Result ContextVk::initialize()
     constexpr VkMemoryPropertyFlags kMemoryType = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT;
     ANGLE_TRY(mEmptyBuffer.init(this, emptyBufferInfo, kMemoryType));
 
+    constexpr VkImageUsageFlags kStagingBufferUsageFlags = VK_BUFFER_USAGE_TRANSFER_SRC_BIT;
+    size_t stagingBufferAlignment =
+        static_cast<size_t>(mRenderer->getPhysicalDeviceProperties().limits.minMemoryMapAlignment);
+    constexpr size_t kStagingBufferSize = 1024u * 1024u;  // 1M
+    mStagingBufferStorage.init(mRenderer, kStagingBufferUsageFlags, stagingBufferAlignment,
+                               kStagingBufferSize, true);
+
     return angle::Result::Continue;
 }
 
@@ -3972,6 +3980,7 @@ angle::Result ContextVk::flushImpl(const vk::Semaphore *signalSemaphore)
         driverUniform.dynamicBuffer.releaseInFlightBuffersToResourceUseList(this);
     }
     mDefaultUniformStorage.releaseInFlightBuffersToResourceUseList(this);
+    mStagingBufferStorage.releaseInFlightBuffersToResourceUseList(this);
 
     if (mRenderer->getFeatures().enableCommandProcessingThread.enabled)
     {

src/libANGLE/renderer/vulkan/ContextVk.h

@@ -591,6 +591,7 @@ class ContextVk : public ContextImpl, public vk::Context
     void setDefaultUniformBlocksMinSizeForTesting(size_t minSize);
 
     vk::BufferHelper &getEmptyBuffer() { return mEmptyBuffer; }
+    vk::DynamicBuffer *getStagingBufferStorage() { return &mStagingBufferStorage; }
 
   private:
     // Dirty bits.
@@ -1065,15 +1066,18 @@ class ContextVk : public ContextImpl, public vk::Context
 
     ShareGroupVk *mShareGroupVk;
 
-    // Storage for default uniforms of ProgramVks and ProgramPipelineVks.
-    vk::DynamicBuffer mDefaultUniformStorage;
-
     // This is a special "empty" placeholder buffer for use when we just need a dummy buffer but not
     // the data. Examples are shader that has no uniform or doesn't use all slots in the atomic
     // counter buffer array, or places where there is no vertex buffer since Vulkan does not allow
     // binding a null vertex buffer.
     vk::BufferHelper mEmptyBuffer;
 
+    // Storage for default uniforms of ProgramVks and ProgramPipelineVks.
+    vk::DynamicBuffer mDefaultUniformStorage;
+
+    // All staging buffer support is provided by a DynamicBuffer.
+    vk::DynamicBuffer mStagingBufferStorage;
+
     std::vector<std::string> mCommandBufferDiagnostics;
 };