Some Vulkan memory fixes and cleanup
* Switch back to not setting transfer_dst on all buffers
* Add some missing unit tests
* Add tracking of heap usage for debugging purposes
* Fall back to non-device-local memory if device-local allocation fails
BUG=skia:5031
GOLD_TRYBOT_URL= https://gold.skia.org/search?issue=2356343003
Committed: https://skia.googlesource.com/skia/+/c5850e9fdb62cc4ae5ed2b6af51aea92cac07455
Review-Url: https://codereview.chromium.org/2356343003
diff --git a/src/gpu/vk/GrVkMemory.cpp b/src/gpu/vk/GrVkMemory.cpp
index 98b2f89..f517b98 100644
--- a/src/gpu/vk/GrVkMemory.cpp
+++ b/src/gpu/vk/GrVkMemory.cpp
@@ -10,16 +10,24 @@
#include "GrVkGpu.h"
#include "GrVkUtil.h"
+#ifdef SK_DEBUG
+// for simple tracking of how much we're using in each heap
+// last counter is for non-subheap allocations
+VkDeviceSize gHeapUsage[VK_MAX_MEMORY_HEAPS+1] = { 0 };
+#endif
+
static bool get_valid_memory_type_index(const VkPhysicalDeviceMemoryProperties& physDevMemProps,
uint32_t typeBits,
VkMemoryPropertyFlags requestedMemFlags,
- uint32_t* typeIndex) {
+ uint32_t* typeIndex,
+ uint32_t* heapIndex) {
for (uint32_t i = 0; i < physDevMemProps.memoryTypeCount; ++i) {
if (typeBits & (1 << i)) {
uint32_t supportedFlags = physDevMemProps.memoryTypes[i].propertyFlags &
requestedMemFlags;
if (supportedFlags == requestedMemFlags) {
*typeIndex = i;
+ *heapIndex = physDevMemProps.memoryTypes[i].heapIndex;
return true;
}
}
@@ -56,6 +64,7 @@
GR_VK_CALL(iface, GetBufferMemoryRequirements(device, buffer, &memReqs));
uint32_t typeIndex = 0;
+ uint32_t heapIndex = 0;
const VkPhysicalDeviceMemoryProperties& phDevMemProps = gpu->physicalDeviceMemoryProperties();
if (dynamic) {
// try to get cached and ideally non-coherent memory first
@@ -63,12 +72,14 @@
memReqs.memoryTypeBits,
VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
- &typeIndex)) {
+ &typeIndex,
+ &heapIndex)) {
// some sort of host-visible memory type should always be available for dynamic buffers
SkASSERT_RELEASE(get_valid_memory_type_index(phDevMemProps,
memReqs.memoryTypeBits,
VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT,
- &typeIndex));
+ &typeIndex,
+ &heapIndex));
}
VkMemoryPropertyFlags mpf = phDevMemProps.memoryTypes[typeIndex].propertyFlags;
@@ -79,15 +90,22 @@
SkASSERT_RELEASE(get_valid_memory_type_index(phDevMemProps,
memReqs.memoryTypeBits,
VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT,
- &typeIndex));
+ &typeIndex,
+ &heapIndex));
alloc->fFlags = 0x0;
}
GrVkHeap* heap = gpu->getHeap(buffer_type_to_heap(type));
- if (!heap->alloc(memReqs.size, memReqs.alignment, typeIndex, alloc)) {
- SkDebugf("Failed to alloc buffer\n");
- return false;
+ if (!heap->alloc(memReqs.size, memReqs.alignment, typeIndex, heapIndex, alloc)) {
+ // if static, try to allocate from non-host-visible non-device-local memory instead
+ if (dynamic ||
+ !get_valid_memory_type_index(phDevMemProps, memReqs.memoryTypeBits,
+ 0, &typeIndex, &heapIndex) ||
+ !heap->alloc(memReqs.size, memReqs.alignment, typeIndex, heapIndex, alloc)) {
+ SkDebugf("Failed to alloc buffer\n");
+ return false;
+ }
}
// Bind buffer
@@ -130,6 +148,7 @@
GR_VK_CALL(iface, GetImageMemoryRequirements(device, image, &memReqs));
uint32_t typeIndex = 0;
+ uint32_t heapIndex = 0;
GrVkHeap* heap;
const VkPhysicalDeviceMemoryProperties& phDevMemProps = gpu->physicalDeviceMemoryProperties();
if (linearTiling) {
@@ -138,12 +157,14 @@
if (!get_valid_memory_type_index(phDevMemProps,
memReqs.memoryTypeBits,
desiredMemProps,
- &typeIndex)) {
+ &typeIndex,
+ &heapIndex)) {
// some sort of host-visible memory type should always be available
SkASSERT_RELEASE(get_valid_memory_type_index(phDevMemProps,
memReqs.memoryTypeBits,
VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT,
- &typeIndex));
+ &typeIndex,
+ &heapIndex));
}
heap = gpu->getHeap(GrVkGpu::kLinearImage_Heap);
VkMemoryPropertyFlags mpf = phDevMemProps.memoryTypes[typeIndex].propertyFlags;
@@ -154,7 +175,8 @@
SkASSERT_RELEASE(get_valid_memory_type_index(phDevMemProps,
memReqs.memoryTypeBits,
VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT,
- &typeIndex));
+ &typeIndex,
+ &heapIndex));
if (memReqs.size <= kMaxSmallImageSize) {
heap = gpu->getHeap(GrVkGpu::kSmallOptimalImage_Heap);
} else {
@@ -163,9 +185,15 @@
alloc->fFlags = 0x0;
}
- if (!heap->alloc(memReqs.size, memReqs.alignment, typeIndex, alloc)) {
- SkDebugf("Failed to alloc image\n");
- return false;
+ if (!heap->alloc(memReqs.size, memReqs.alignment, typeIndex, heapIndex, alloc)) {
+ // if optimal, try to allocate from non-host-visible non-device-local memory instead
+ if (linearTiling ||
+ !get_valid_memory_type_index(phDevMemProps, memReqs.memoryTypeBits,
+ 0, &typeIndex, &heapIndex) ||
+ !heap->alloc(memReqs.size, memReqs.alignment, typeIndex, heapIndex, alloc)) {
+ SkDebugf("Failed to alloc image\n");
+ return false;
+ }
}
// Bind image
@@ -431,10 +459,13 @@
#endif
}
-GrVkSubHeap::GrVkSubHeap(const GrVkGpu* gpu, uint32_t memoryTypeIndex,
+GrVkSubHeap::GrVkSubHeap(const GrVkGpu* gpu, uint32_t memoryTypeIndex, uint32_t heapIndex,
VkDeviceSize size, VkDeviceSize alignment)
: INHERITED(size, alignment)
, fGpu(gpu)
+#ifdef SK_DEBUG
+ , fHeapIndex(heapIndex)
+#endif
, fMemoryTypeIndex(memoryTypeIndex) {
VkMemoryAllocateInfo allocInfo = {
@@ -450,12 +481,20 @@
&fAlloc));
if (VK_SUCCESS != err) {
this->reset();
+ }
+#ifdef SK_DEBUG
+ else {
+ gHeapUsage[heapIndex] += size;
}
+#endif
}
GrVkSubHeap::~GrVkSubHeap() {
const GrVkInterface* iface = fGpu->vkInterface();
GR_VK_CALL(iface, FreeMemory(fGpu->device(), fAlloc, nullptr));
+#ifdef SK_DEBUG
+ gHeapUsage[fHeapIndex] -= fSize;
+#endif
}
bool GrVkSubHeap::alloc(VkDeviceSize size, GrVkAlloc* alloc) {
@@ -470,7 +509,7 @@
}
bool GrVkHeap::subAlloc(VkDeviceSize size, VkDeviceSize alignment,
- uint32_t memoryTypeIndex, GrVkAlloc* alloc) {
+ uint32_t memoryTypeIndex, uint32_t heapIndex, GrVkAlloc* alloc) {
VkDeviceSize alignedSize = align_size(size, alignment);
// if requested is larger than our subheap allocation, just alloc directly
@@ -491,6 +530,9 @@
}
alloc->fOffset = 0;
alloc->fSize = 0; // hint that this is not a subheap allocation
+#ifdef SK_DEBUG
+ gHeapUsage[VK_MAX_MEMORY_HEAPS] += alignedSize;
+#endif
return true;
}
@@ -520,11 +562,11 @@
// need to allocate a new subheap
SkAutoTDelete<GrVkSubHeap>& subHeap = fSubHeaps.push_back();
- subHeap.reset(new GrVkSubHeap(fGpu, memoryTypeIndex, fSubHeapSize, alignment));
+ subHeap.reset(new GrVkSubHeap(fGpu, memoryTypeIndex, heapIndex, fSubHeapSize, alignment));
// try to recover from failed allocation by only allocating what we need
if (subHeap->size() == 0) {
VkDeviceSize alignedSize = align_size(size, alignment);
- subHeap.reset(new GrVkSubHeap(fGpu, memoryTypeIndex, alignedSize, alignment));
+ subHeap.reset(new GrVkSubHeap(fGpu, memoryTypeIndex, heapIndex, alignedSize, alignment));
if (subHeap->size() == 0) {
return false;
}
@@ -539,7 +581,7 @@
}
bool GrVkHeap::singleAlloc(VkDeviceSize size, VkDeviceSize alignment,
- uint32_t memoryTypeIndex, GrVkAlloc* alloc) {
+ uint32_t memoryTypeIndex, uint32_t heapIndex, GrVkAlloc* alloc) {
VkDeviceSize alignedSize = align_size(size, alignment);
// first try to find an unallocated subheap that fits our allocation request
@@ -568,7 +610,7 @@
// need to allocate a new subheap
SkAutoTDelete<GrVkSubHeap>& subHeap = fSubHeaps.push_back();
- subHeap.reset(new GrVkSubHeap(fGpu, memoryTypeIndex, alignedSize, alignment));
+ subHeap.reset(new GrVkSubHeap(fGpu, memoryTypeIndex, heapIndex, alignedSize, alignment));
fAllocSize += alignedSize;
if (subHeap->alloc(size, alloc)) {
fUsedSize += alloc->fSize;