Reland "Use GrVkMemoryAllocator for vulkan memory allocations in ganesh."
This is a reland of 331c266ed716526478a10885aff66181cec64486
Original change's description:
> Use GrVkMemoryAllocator for vulkan memory allocations in ganesh.
>
> Besides using the new allocator, the big logical change is that map
> and unmap calls form GrVkMemory are specc'd to map the entire GrVkAlloc
> instead of a specific offset and size as they did before. As a
> consequence of this, we move the handling of non-coherent alignment
> for flush/invalidate calls to GrVkMemory instead of the callers.
>
> Bug: skia:
> Change-Id: I794d713106602f27aa7e808c306bbb69fd2b67be
> Reviewed-on: https://skia-review.googlesource.com/130021
> Commit-Queue: Greg Daniel <egdaniel@google.com>
> Reviewed-by: Jim Van Verth <jvanverth@google.com>
Bug: skia:
Change-Id: Ia9a4192d344449fb444d2adaa1d62ff1ede4b21d
Reviewed-on: https://skia-review.googlesource.com/131083
Reviewed-by: Jim Van Verth <jvanverth@google.com>
Commit-Queue: Greg Daniel <egdaniel@google.com>
diff --git a/src/gpu/vk/GrVkGpu.cpp b/src/gpu/vk/GrVkGpu.cpp
index 56d0b95..2525c5c 100644
--- a/src/gpu/vk/GrVkGpu.cpp
+++ b/src/gpu/vk/GrVkGpu.cpp
@@ -17,6 +17,7 @@
#include "GrRenderTargetPriv.h"
#include "GrTexturePriv.h"
+#include "GrVkAMDMemoryAllocator.h"
#include "GrVkCommandBuffer.h"
#include "GrVkGpuCommandBuffer.h"
#include "GrVkImage.h"
@@ -92,6 +93,7 @@
sk_sp<const GrVkBackendContext> backendCtx)
: INHERITED(context)
, fBackendContext(std::move(backendCtx))
+ , fMemoryAllocator(fBackendContext->fMemoryAllocator)
, fDevice(fBackendContext->fDevice)
, fQueue(fBackendContext->fQueue)
, fResourceProvider(this)
@@ -118,6 +120,12 @@
}
#endif
+ if (!fMemoryAllocator) {
+ // We were not given a memory allocator at creation
+ fMemoryAllocator.reset(new GrVkAMDMemoryAllocator(fBackendContext->fPhysicalDevice,
+ fDevice, fBackendContext->fInterface));
+ }
+
fCompiler = new SkSL::Compiler();
fVkCaps.reset(new GrVkCaps(options, this->vkInterface(), fBackendContext->fPhysicalDevice,
@@ -142,17 +150,6 @@
fCurrentCmdBuffer = fResourceProvider.findOrCreatePrimaryCommandBuffer();
SkASSERT(fCurrentCmdBuffer);
fCurrentCmdBuffer->begin(this);
-
- // set up our heaps
- fHeaps[kLinearImage_Heap].reset(new GrVkHeap(this, GrVkHeap::kSubAlloc_Strategy, 16*1024*1024));
- fHeaps[kOptimalImage_Heap].reset(new GrVkHeap(this, GrVkHeap::kSubAlloc_Strategy, 64*1024*1024));
- fHeaps[kSmallOptimalImage_Heap].reset(new GrVkHeap(this, GrVkHeap::kSubAlloc_Strategy, 2*1024*1024));
- fHeaps[kVertexBuffer_Heap].reset(new GrVkHeap(this, GrVkHeap::kSingleAlloc_Strategy, 0));
- fHeaps[kIndexBuffer_Heap].reset(new GrVkHeap(this, GrVkHeap::kSingleAlloc_Strategy, 0));
- fHeaps[kUniformBuffer_Heap].reset(new GrVkHeap(this, GrVkHeap::kSubAlloc_Strategy, 256*1024));
- fHeaps[kTexelBuffer_Heap].reset(new GrVkHeap(this, GrVkHeap::kSingleAlloc_Strategy, 0));
- fHeaps[kCopyReadBuffer_Heap].reset(new GrVkHeap(this, GrVkHeap::kSingleAlloc_Strategy, 0));
- fHeaps[kCopyWriteBuffer_Heap].reset(new GrVkHeap(this, GrVkHeap::kSubAlloc_Strategy, 16*1024*1024));
}
void GrVkGpu::destroyResources() {
@@ -562,7 +559,6 @@
0, // arraySlice
};
VkSubresourceLayout layout;
- VkResult err;
const GrVkInterface* interface = this->vkInterface();
@@ -573,28 +569,14 @@
int texTop = kBottomLeft_GrSurfaceOrigin == texOrigin ? tex->height() - top - height : top;
const GrVkAlloc& alloc = tex->alloc();
- VkDeviceSize offset = alloc.fOffset + texTop*layout.rowPitch + left*bpp;
- VkDeviceSize offsetDiff = 0;
+ VkDeviceSize offset = texTop*layout.rowPitch + left*bpp;
VkDeviceSize size = height*layout.rowPitch;
- // For Noncoherent buffers we want to make sure the range that we map, both offset and size,
- // are aligned to the nonCoherentAtomSize limit. We may have to move the initial offset back to
- // meet the alignment requirements. So we track how far we move back and then adjust the mapped
- // ptr back up so that this is opaque to the caller.
- if (SkToBool(alloc.fFlags & GrVkAlloc::kNoncoherent_Flag)) {
- VkDeviceSize alignment = this->physicalDeviceProperties().limits.nonCoherentAtomSize;
- offsetDiff = offset & (alignment - 1);
- offset = offset - offsetDiff;
- // Make size of the map aligned to nonCoherentAtomSize
- size = (size + alignment - 1) & ~(alignment - 1);
- }
- SkASSERT(offset >= alloc.fOffset);
- SkASSERT(size <= alloc.fOffset + alloc.fSize);
- void* mapPtr;
- err = GR_VK_CALL(interface, MapMemory(fDevice, alloc.fMemory, offset, size, 0, &mapPtr));
- if (err) {
+ SkASSERT(size + offset <= alloc.fSize);
+ void* mapPtr = GrVkMemory::MapAlloc(this, alloc);
+ if (!mapPtr) {
return false;
}
- mapPtr = reinterpret_cast<char*>(mapPtr) + offsetDiff;
+ mapPtr = reinterpret_cast<char*>(mapPtr) + offset;
if (kBottomLeft_GrSurfaceOrigin == texOrigin) {
// copy into buffer by rows
@@ -611,7 +593,7 @@
}
GrVkMemory::FlushMappedAlloc(this, alloc, offset, size);
- GR_VK_CALL(interface, UnmapMemory(fDevice, alloc.fMemory));
+ GrVkMemory::UnmapAlloc(this, alloc);
return true;
}
@@ -1147,33 +1129,14 @@
bool copy_testing_data(GrVkGpu* gpu, const void* srcData, const GrVkAlloc& alloc,
size_t bufferOffset, size_t srcRowBytes, size_t dstRowBytes, int h) {
- // For Noncoherent buffers we want to make sure the range that we map, both offset and size,
- // are aligned to the nonCoherentAtomSize limit. We may have to move the initial offset back to
- // meet the alignment requirements. So we track how far we move back and then adjust the mapped
- // ptr back up so that this is opaque to the caller.
- VkDeviceSize mapSize = dstRowBytes * h;
- VkDeviceSize mapOffset = alloc.fOffset + bufferOffset;
- VkDeviceSize offsetDiff = 0;
- if (SkToBool(alloc.fFlags & GrVkAlloc::kNoncoherent_Flag)) {
- VkDeviceSize alignment = gpu->physicalDeviceProperties().limits.nonCoherentAtomSize;
- offsetDiff = mapOffset & (alignment - 1);
- mapOffset = mapOffset - offsetDiff;
- // Make size of the map aligned to nonCoherentAtomSize
- mapSize = (mapSize + alignment - 1) & ~(alignment - 1);
- }
- SkASSERT(mapOffset >= alloc.fOffset);
- SkASSERT(mapSize + mapOffset <= alloc.fOffset + alloc.fSize);
- void* mapPtr;
- VkResult err = GR_VK_CALL(gpu->vkInterface(), MapMemory(gpu->device(),
- alloc.fMemory,
- mapOffset,
- mapSize,
- 0,
- &mapPtr));
- mapPtr = reinterpret_cast<char*>(mapPtr) + offsetDiff;
- if (err) {
+ VkDeviceSize size = dstRowBytes * h;
+ VkDeviceSize offset = bufferOffset;
+ SkASSERT(size + offset <= alloc.fSize);
+ void* mapPtr = GrVkMemory::MapAlloc(gpu, alloc);
+ if (!mapPtr) {
return false;
}
+ mapPtr = reinterpret_cast<char*>(mapPtr) + offset;
if (srcData) {
// If there is no padding on dst we can do a single memcopy.
@@ -1192,8 +1155,8 @@
}
}
}
- GrVkMemory::FlushMappedAlloc(gpu, alloc, mapOffset, mapSize);
- GR_VK_CALL(gpu->vkInterface(), UnmapMemory(gpu->device(), alloc.fMemory));
+ GrVkMemory::FlushMappedAlloc(gpu, alloc, offset, size);
+ GrVkMemory::UnmapAlloc(gpu, alloc);
return true;
}
@@ -2017,7 +1980,7 @@
this->submitCommandBuffer(kForce_SyncQueue);
void* mappedMemory = transferBuffer->map();
const GrVkAlloc& transAlloc = transferBuffer->alloc();
- GrVkMemory::InvalidateMappedAlloc(this, transAlloc, transAlloc.fOffset, VK_WHOLE_SIZE);
+ GrVkMemory::InvalidateMappedAlloc(this, transAlloc, 0, transAlloc.fSize);
if (copyFromOrigin) {
uint32_t skipRows = region.imageExtent.height - height;