Merge pull request #1382 from billhollings/coherent-texture-flushing
Reorganize coherent texture flushing on memory map and unmap.
diff --git a/MoltenVK/MoltenVK/Commands/MVKCmdTransfer.mm b/MoltenVK/MoltenVK/Commands/MVKCmdTransfer.mm
index 7d1d379..1aa38f2 100644
--- a/MoltenVK/MoltenVK/Commands/MVKCmdTransfer.mm
+++ b/MoltenVK/MoltenVK/Commands/MVKCmdTransfer.mm
@@ -88,8 +88,6 @@
VkBufferImageCopy vkDstCopies[copyCnt];
size_t tmpBuffSize = 0;
- _srcImage->flushToDevice(0, VK_WHOLE_SIZE);
-
for (uint32_t copyIdx = 0; copyIdx < copyCnt; copyIdx++) {
auto& vkIC = _vkImageCopies[copyIdx];
diff --git a/MoltenVK/MoltenVK/GPUObjects/MVKDevice.mm b/MoltenVK/MoltenVK/GPUObjects/MVKDevice.mm
index 6ca2971..a1bc0d0 100644
--- a/MoltenVK/MoltenVK/GPUObjects/MVKDevice.mm
+++ b/MoltenVK/MoltenVK/GPUObjects/MVKDevice.mm
@@ -3616,7 +3616,7 @@
for (uint32_t i = 0; i < memRangeCount; i++) {
const VkMappedMemoryRange* pMem = &pMemRanges[i];
MVKDeviceMemory* mvkMem = (MVKDeviceMemory*)pMem->memory;
- VkResult r = mvkMem->pullFromDevice(pMem->offset, pMem->size, false, &mvkBlitEnc);
+ VkResult r = mvkMem->pullFromDevice(pMem->offset, pMem->size, &mvkBlitEnc);
if (rslt == VK_SUCCESS) { rslt = r; }
}
if (mvkBlitEnc.mtlBlitEncoder) { [mvkBlitEnc.mtlBlitEncoder endEncoding]; }
diff --git a/MoltenVK/MoltenVK/GPUObjects/MVKDeviceMemory.h b/MoltenVK/MoltenVK/GPUObjects/MVKDeviceMemory.h
index b08aa46..953e3b8 100644
--- a/MoltenVK/MoltenVK/GPUObjects/MVKDeviceMemory.h
+++ b/MoltenVK/MoltenVK/GPUObjects/MVKDeviceMemory.h
@@ -93,17 +93,11 @@
/** Returns whether this device memory is currently mapped to host memory. */
bool isMapped() { return _mappedRange.size > 0; }
- /**
- * If this memory is host-visible, the specified memory range is flushed to the device.
- * Normally, flushing will only occur if the device memory is non-coherent, but flushing
- * to coherent memory can be forced by setting evenIfCoherent to true.
- */
- VkResult flushToDevice(VkDeviceSize offset, VkDeviceSize size, bool evenIfCoherent = false);
+ /** If this memory is host-visible, the specified memory range is flushed to the device. */
+ VkResult flushToDevice(VkDeviceSize offset, VkDeviceSize size);
/**
* If this memory is host-visible, pulls the specified memory range from the device.
- * Normally, pulling will only occur if the device memory is non-coherent, but pulling
- * to coherent memory can be forced by setting evenIfCoherent to true.
*
* If pBlitEnc is not null, it points to a holder for a MTLBlitCommandEncoder and its
* associated MTLCommandBuffer. If this instance has a MTLBuffer using managed memory,
@@ -114,7 +108,6 @@
*/
VkResult pullFromDevice(VkDeviceSize offset,
VkDeviceSize size,
- bool evenIfCoherent = false,
MVKMTLBlitEncoder* pBlitEnc = nullptr);
@@ -172,8 +165,10 @@
id<MTLHeap> _mtlHeap = nil;
void* _pMemory = nullptr;
void* _pHostMemory = nullptr;
- bool _isDedicated = false;
+ VkMemoryPropertyFlags _vkMemProps;
MTLStorageMode _mtlStorageMode;
MTLCPUCacheMode _mtlCPUCacheMode;
+ bool _isDedicated = false;
+
};
diff --git a/MoltenVK/MoltenVK/GPUObjects/MVKDeviceMemory.mm b/MoltenVK/MoltenVK/GPUObjects/MVKDeviceMemory.mm
index eac900f..25686df 100644
--- a/MoltenVK/MoltenVK/GPUObjects/MVKDeviceMemory.mm
+++ b/MoltenVK/MoltenVK/GPUObjects/MVKDeviceMemory.mm
@@ -55,8 +55,11 @@
*ppData = (void*)((uintptr_t)_pMemory + offset);
- // Coherent memory does not require flushing by app, so we must flush now, to handle any texture updates.
- pullFromDevice(offset, size, isMemoryHostCoherent());
+ // Coherent memory does not require flushing by app, so we must flush now
+ // to support Metal textures that actually reside in non-coherent memory.
+ if (mvkIsAnyFlagEnabled(_vkMemProps, VK_MEMORY_PROPERTY_HOST_COHERENT_BIT)) {
+ pullFromDevice(offset, size);
+ }
return VK_SUCCESS;
}
@@ -68,54 +71,57 @@
return;
}
- // Coherent memory does not require flushing by app, so we must flush now.
- flushToDevice(_mappedRange.offset, _mappedRange.size, isMemoryHostCoherent());
+ // Coherent memory does not require flushing by app, so we must flush now
+ // to support Metal textures that actually reside in non-coherent memory.
+ if (mvkIsAnyFlagEnabled(_vkMemProps, VK_MEMORY_PROPERTY_HOST_COHERENT_BIT)) {
+ flushToDevice(_mappedRange.offset, _mappedRange.size);
+ }
_mappedRange.offset = 0;
_mappedRange.size = 0;
}
-VkResult MVKDeviceMemory::flushToDevice(VkDeviceSize offset, VkDeviceSize size, bool evenIfCoherent) {
- // Coherent memory is flushed on unmap(), so it is only flushed if forced
+VkResult MVKDeviceMemory::flushToDevice(VkDeviceSize offset, VkDeviceSize size) {
VkDeviceSize memSize = adjustMemorySize(size, offset);
- if (memSize > 0 && isMemoryHostAccessible() && (evenIfCoherent || !isMemoryHostCoherent()) ) {
+ if (memSize == 0 || !isMemoryHostAccessible()) { return VK_SUCCESS; }
#if MVK_MACOS
- if (_mtlBuffer && _mtlStorageMode == MTLStorageModeManaged) {
- [_mtlBuffer didModifyRange: NSMakeRange(offset, memSize)];
- }
+ if (_mtlBuffer && _mtlStorageMode == MTLStorageModeManaged) {
+ [_mtlBuffer didModifyRange: NSMakeRange(offset, memSize)];
+ }
#endif
- // If we have an MTLHeap object, there's no need to sync memory manually between images and the buffer.
- if (!_mtlHeap) {
- lock_guard<mutex> lock(_rezLock);
- for (auto& img : _imageMemoryBindings) { img->flushToDevice(offset, memSize); }
- for (auto& buf : _buffers) { buf->flushToDevice(offset, memSize); }
- }
+ // If we have an MTLHeap object, there's no need to sync memory manually between resources and the buffer.
+ if ( !_mtlHeap ) {
+ lock_guard<mutex> lock(_rezLock);
+ for (auto& img : _imageMemoryBindings) { img->flushToDevice(offset, memSize); }
+ for (auto& buf : _buffers) { buf->flushToDevice(offset, memSize); }
}
+
return VK_SUCCESS;
}
VkResult MVKDeviceMemory::pullFromDevice(VkDeviceSize offset,
VkDeviceSize size,
- bool evenIfCoherent,
MVKMTLBlitEncoder* pBlitEnc) {
- // Coherent memory is flushed on unmap(), so it is only flushed if forced
VkDeviceSize memSize = adjustMemorySize(size, offset);
- if (memSize > 0 && isMemoryHostAccessible() && (evenIfCoherent || !isMemoryHostCoherent()) && !_mtlHeap) {
+ if (memSize == 0 || !isMemoryHostAccessible()) { return VK_SUCCESS; }
+
+#if MVK_MACOS
+ if (pBlitEnc && _mtlBuffer && _mtlStorageMode == MTLStorageModeManaged) {
+ if ( !pBlitEnc->mtlCmdBuffer) { pBlitEnc->mtlCmdBuffer = _device->getAnyQueue()->getMTLCommandBuffer(kMVKCommandUseInvalidateMappedMemoryRanges); }
+ if ( !pBlitEnc->mtlBlitEncoder) { pBlitEnc->mtlBlitEncoder = [pBlitEnc->mtlCmdBuffer blitCommandEncoder]; }
+ [pBlitEnc->mtlBlitEncoder synchronizeResource: _mtlBuffer];
+ }
+#endif
+
+ // If we have an MTLHeap object, there's no need to sync memory manually between resources and the buffer.
+ if ( !_mtlHeap ) {
lock_guard<mutex> lock(_rezLock);
for (auto& img : _imageMemoryBindings) { img->pullFromDevice(offset, memSize); }
for (auto& buf : _buffers) { buf->pullFromDevice(offset, memSize); }
-
-#if MVK_MACOS
- if (pBlitEnc && _mtlBuffer && _mtlStorageMode == MTLStorageModeManaged) {
- if ( !pBlitEnc->mtlCmdBuffer) { pBlitEnc->mtlCmdBuffer = _device->getAnyQueue()->getMTLCommandBuffer(kMVKCommandUseInvalidateMappedMemoryRanges); }
- if ( !pBlitEnc->mtlBlitEncoder) { pBlitEnc->mtlBlitEncoder = [pBlitEnc->mtlCmdBuffer blitCommandEncoder]; }
- [pBlitEnc->mtlBlitEncoder synchronizeResource: _mtlBuffer];
- }
-#endif
-
}
+
return VK_SUCCESS;
}
@@ -271,9 +277,9 @@
const VkMemoryAllocateInfo* pAllocateInfo,
const VkAllocationCallbacks* pAllocator) : MVKVulkanAPIDeviceObject(device) {
// Set Metal memory parameters
- VkMemoryPropertyFlags vkMemProps = _device->_pMemoryProperties->memoryTypes[pAllocateInfo->memoryTypeIndex].propertyFlags;
- _mtlStorageMode = mvkMTLStorageModeFromVkMemoryPropertyFlags(vkMemProps);
- _mtlCPUCacheMode = mvkMTLCPUCacheModeFromVkMemoryPropertyFlags(vkMemProps);
+ _vkMemProps = _device->_pMemoryProperties->memoryTypes[pAllocateInfo->memoryTypeIndex].propertyFlags;
+ _mtlStorageMode = mvkMTLStorageModeFromVkMemoryPropertyFlags(_vkMemProps);
+ _mtlCPUCacheMode = mvkMTLCPUCacheModeFromVkMemoryPropertyFlags(_vkMemProps);
_allocationSize = pAllocateInfo->allocationSize;
