Merge pull request #2217 from billhollings/avoid-managed-mem-on-apple…

…-silicon

On macOS Apple Silicon, avoid managed-memory textures, and resource syncs.

MoltenVK/MoltenVK/API/mvk_datatypes.h

@@ -482,9 +482,6 @@ static inline VkExtent3D mvkVkExtent3DFromMTLSize(MTLSize mtlSize) {
 /** Macro indicating the Vulkan memory type bits corresponding to Metal memoryless memory (not host visible and lazily allocated). */
 #define MVK_VK_MEMORY_TYPE_METAL_MEMORYLESS (VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT | VK_MEMORY_PROPERTY_LAZILY_ALLOCATED_BIT)
 
-/** Returns the Metal storage mode corresponding to the specified Vulkan memory flags. */
-MTLStorageMode mvkMTLStorageModeFromVkMemoryPropertyFlags(VkMemoryPropertyFlags vkFlags);
-
 /** Returns the Metal CPU cache mode corresponding to the specified Vulkan memory flags. */
 MTLCPUCacheMode mvkMTLCPUCacheModeFromVkMemoryPropertyFlags(VkMemoryPropertyFlags vkFlags);
 

MoltenVK/MoltenVK/GPUObjects/MVKBuffer.mm

@@ -81,7 +81,11 @@
 
 #if MVK_MACOS
  if (_deviceMemory) {
- _isHostCoherentTexelBuffer = !_device->_pMetalFeatures->sharedLinearTextures && _deviceMemory->isMemoryHostCoherent() && mvkIsAnyFlagEnabled(_usage, VK_BUFFER_USAGE_UNIFORM_TEXEL_BUFFER_BIT | VK_BUFFER_USAGE_STORAGE_TEXEL_BUFFER_BIT);
+ _isHostCoherentTexelBuffer = (!isUnifiedMemoryGPU() &&
+ !_device->_pMetalFeatures->sharedLinearTextures &&
+ _deviceMemory->isMemoryHostCoherent() &&
+ mvkIsAnyFlagEnabled(_usage, (VK_BUFFER_USAGE_UNIFORM_TEXEL_BUFFER_BIT |
+ VK_BUFFER_USAGE_STORAGE_TEXEL_BUFFER_BIT)));
  }
 #endif
 
@@ -118,7 +122,8 @@
 // buffer and host memory for the purpose of the host reading texture memory.
 bool MVKBuffer::needsHostReadSync(MVKPipelineBarrier& barrier) {
 #if MVK_MACOS
- return (mvkIsAnyFlagEnabled(barrier.dstStageMask, (VK_PIPELINE_STAGE_HOST_BIT)) &&
+ return (!isUnifiedMemoryGPU() &&
+ mvkIsAnyFlagEnabled(barrier.dstStageMask, (VK_PIPELINE_STAGE_HOST_BIT)) &&
  mvkIsAnyFlagEnabled(barrier.dstAccessMask, (VK_ACCESS_HOST_READ_BIT)) &&
  isMemoryHostAccessible() && (!isMemoryHostCoherent() || _isHostCoherentTexelBuffer));
 #else
@@ -138,9 +143,7 @@
  return false;
 }
 
-#if MVK_MACOS
-bool MVKBuffer::shouldFlushHostMemory() { return _isHostCoherentTexelBuffer; }
-#endif
+bool MVKBuffer::shouldFlushHostMemory() { return !isUnifiedMemoryGPU() && _isHostCoherentTexelBuffer; }
 
 // Flushes the device memory at the specified memory range into the MTLBuffer.
 VkResult MVKBuffer::flushToDevice(VkDeviceSize offset, VkDeviceSize size) {

MoltenVK/MoltenVK/GPUObjects/MVKDevice.h

@@ -331,9 +331,6 @@ class MVKPhysicalDevice : public MVKDispatchableVulkanAPIObject {
  */
  uint32_t getLazilyAllocatedMemoryTypes() { return _lazilyAllocatedMemoryTypes; }
 
- /** Returns whether this is a unified memory device. */
- bool getHasUnifiedMemory();
-
  /** Returns the external memory properties supported for buffers for the handle type. */
  VkExternalMemoryProperties& getExternalBufferProperties(VkExternalMemoryHandleTypeFlagBits handleType);
 
@@ -363,6 +360,9 @@ class MVKPhysicalDevice : public MVKDispatchableVulkanAPIObject {
  /** Returns whether native texture atomics are supported and should be used. */
  bool useNativeTextureAtomics() { return _metalFeatures.nativeTextureAtomics; }
 
+ /** Returns the MTLStorageMode that matches the Vulkan memory property flags. */
+ MTLStorageMode getMTLStorageModeFromVkMemoryPropertyFlags(VkMemoryPropertyFlags vkFlags);
+
 
 #pragma mark Construction
 
@@ -388,6 +388,7 @@ class MVKPhysicalDevice : public MVKDispatchableVulkanAPIObject {
 
 protected:
  friend class MVKDevice;
+ friend class MVKDeviceTrackingMixin;
 
  void propagateDebugName() override {}
  MTLFeatureSet getMaximalMTLFeatureSet();
@@ -443,6 +444,8 @@ class MVKPhysicalDevice : public MVKDispatchableVulkanAPIObject {
  uint32_t _hostCoherentMemoryTypes;
  uint32_t _privateMemoryTypes;
  uint32_t _lazilyAllocatedMemoryTypes;
+ bool _hasUnifiedMemory = true;
+ bool _isAppleGPU = true;
 };
 
 
@@ -887,6 +890,8 @@ class MVKDevice : public MVKDispatchableVulkanAPIObject {
  }
 
 protected:
+ friend class MVKDeviceTrackingMixin;
+
  void propagateDebugName() override {}
  MVKBuffer* addBuffer(MVKBuffer* mvkBuff);
  MVKBuffer* removeBuffer(MVKBuffer* mvkBuff);
@@ -956,6 +961,12 @@ class MVKDeviceTrackingMixin {
  /** Returns the underlying Metal device. */
  id<MTLDevice> getMTLDevice() { return _device->getMTLDevice(); }
 
+ /** Returns whether the GPU is a unified memory device. */
+ bool isUnifiedMemoryGPU() { return getPhysicalDevice()->_hasUnifiedMemory; }
+
+ /** Returns whether the GPU is Apple Silicon. */
+ bool isAppleGPU() { return getPhysicalDevice()->_isAppleGPU; }
+
  /** Returns info about the pixel format supported by the physical device. */
  MVKPixelFormats* getPixelFormats() { return _device->getPixelFormats(); }
 

MoltenVK/MoltenVK/GPUObjects/MVKDevice.mm

@@ -1765,9 +1765,7 @@
 // wild temporary changes, particularly during initial queries before much GPU activity has occurred.
 // On Apple GPUs, CPU & GPU timestamps are the same, and timestamp period never changes.
 void MVKPhysicalDevice::updateTimestampPeriod() {
- if (_properties.vendorID != kAppleVendorId &&
- [_mtlDevice respondsToSelector: @selector(sampleTimestamps:gpuTimestamp:)]) {
-
+ if ( !_isAppleGPU && [_mtlDevice respondsToSelector: @selector(sampleTimestamps:gpuTimestamp:)]) {
  MTLTimestamp earlierCPUTs = _prevCPUTimestamp;
  MTLTimestamp earlierGPUTs = _prevGPUTimestamp;
  [_mtlDevice sampleTimestamps: &_prevCPUTimestamp gpuTimestamp: &_prevGPUTimestamp];
@@ -1804,7 +1802,7 @@
  auto* budgetProps = (VkPhysicalDeviceMemoryBudgetPropertiesEXT*)next;
  mvkClear(budgetProps->heapBudget, VK_MAX_MEMORY_HEAPS);
  mvkClear(budgetProps->heapUsage, VK_MAX_MEMORY_HEAPS);
- if (!getHasUnifiedMemory()) {
+ if ( !_hasUnifiedMemory ) {
  budgetProps->heapBudget[1] = (VkDeviceSize)mvkGetAvailableMemorySize();
  budgetProps->heapUsage[1] = (VkDeviceSize)mvkGetUsedMemorySize();
  }
@@ -1833,11 +1831,11 @@
  _supportedExtensions(this, true),
  _pixelFormats(this) { // Set after _mtlDevice
 
- initMTLDevice();
- initProperties(); // Call first.
- initMetalFeatures(); // Call second.
- initFeatures(); // Call third.
- initLimits(); // Call fourth.
+ initMTLDevice(); // Call first.
+ initProperties(); // Call second.
+ initMetalFeatures(); // Call third.
+ initFeatures(); // Call fourth.
+ initLimits(); // Call fifth.
  initExtensions();
  initMemoryProperties();
  initExternalMemoryProperties();
@@ -1847,12 +1845,21 @@
 }
 
 void MVKPhysicalDevice::initMTLDevice() {
-#if MVK_XCODE_14_3 && MVK_MACOS && !MVK_MACCAT
+#if MVK_MACOS
+ _isAppleGPU = supportsMTLGPUFamily(Apple1);
+
+ // Apple Silicon will respond false to isLowPower, but never hits it.
+ _hasUnifiedMemory = ([_mtlDevice respondsToSelector: @selector(hasUnifiedMemory)]
+ ? _mtlDevice.hasUnifiedMemory : _mtlDevice.isLowPower);
+
+#if MVK_XCODE_14_3 && !MVK_MACCAT
  if ([_mtlDevice respondsToSelector: @selector(setShouldMaximizeConcurrentCompilation:)]) {
  [_mtlDevice setShouldMaximizeConcurrentCompilation: getMVKConfig().shouldMaximizeConcurrentCompilation];
  MVKLogInfoIf(getMVKConfig().debugMode, "maximumConcurrentCompilationTaskCount %lu", _mtlDevice.maximumConcurrentCompilationTaskCount);
  }
 #endif
+
+#endif // MVK_MACOS
 }
 
 // Initializes the physical device properties (except limits).
@@ -2968,16 +2975,14 @@ static uint32_t mvkGetEntryProperty(io_registry_entry_t entry, CFStringRef prope
 }
 
 void MVKPhysicalDevice::initGPUInfoProperties() {
-
- bool isIntegrated = getHasUnifiedMemory();
- _properties.deviceType = isIntegrated ? VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU : VK_PHYSICAL_DEVICE_TYPE_DISCRETE_GPU;
+ _properties.deviceType = _hasUnifiedMemory ? VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU : VK_PHYSICAL_DEVICE_TYPE_DISCRETE_GPU;
  strlcpy(_properties.deviceName, _mtlDevice.name.UTF8String, VK_MAX_PHYSICAL_DEVICE_NAME_SIZE);
 
  // For Apple Silicon, the Device ID is determined by the highest
  // GPU capability, which is a combination of OS version and GPU type.
  // We determine Apple Silicon directly from the GPU, instead
  // of from the build, in case we are running Rosetta2.
- if (supportsMTLGPUFamily(Apple1)) {
+ if (_isAppleGPU) {
  _properties.vendorID = kAppleVendorId;
  _properties.deviceID = getHighestGPUCapability();
  return;
@@ -3012,9 +3017,9 @@ static uint32_t mvkGetEntryProperty(io_registry_entry_t entry, CFStringRef prope
  if (mvkGetEntryProperty(entry, CFSTR("class-code")) == 0x30000) { // 0x30000 : DISPLAY_VGA
 
  // The Intel GPU will always be marked as integrated.
- // Return on a match of either Intel && low power, or non-Intel and non-low-power.
+ // Return on a match of either Intel && unified memory, or non-Intel and non-unified memory.
  uint32_t vendorID = mvkGetEntryProperty(entry, CFSTR("vendor-id"));
- if ( (vendorID == kIntelVendorId) == isIntegrated) {
+ if ( (vendorID == kIntelVendorId) == _hasUnifiedMemory) {
  isFound = true;
  _properties.vendorID = vendorID;
  _properties.deviceID = mvkGetEntryProperty(entry, CFSTR("device-id"));
@@ -3168,7 +3173,7 @@ static uint32_t mvkGetEntryProperty(io_registry_entry_t entry, CFStringRef prope
  // Optional second heap for shared memory
  uint32_t sharedHeapIdx;
  VkMemoryPropertyFlags sharedTypePropFlags;
- if (getHasUnifiedMemory()) {
+ if (_hasUnifiedMemory) {
  // Shared memory goes in the single main heap in unified memory, and per Vulkan spec must be marked local
  sharedHeapIdx = mainHeapIdx;
  sharedTypePropFlags = MVK_VK_MEMORY_TYPE_METAL_SHARED | VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT;
@@ -3194,12 +3199,14 @@ static uint32_t mvkGetEntryProperty(io_registry_entry_t entry, CFStringRef prope
  setMemoryType(typeIdx, sharedHeapIdx, sharedTypePropFlags);
  typeIdx++;
 
- // Managed storage
+ // Managed storage. On all Apple Silicon, use Shared instead.
  uint32_t managedBit = 0;
 #if MVK_MACOS
- managedBit = 1 << typeIdx;
- setMemoryType(typeIdx, mainHeapIdx, MVK_VK_MEMORY_TYPE_METAL_MANAGED);
- typeIdx++;
+ if ( !_isAppleGPU ) {
+ managedBit = 1 << typeIdx;
+ setMemoryType(typeIdx, mainHeapIdx, MVK_VK_MEMORY_TYPE_METAL_MANAGED);
+ typeIdx++;
+ }
 #endif
 
  // Memoryless storage
@@ -3235,17 +3242,33 @@ static uint32_t mvkGetEntryProperty(io_registry_entry_t entry, CFStringRef prope
  _allMemoryTypes = privateBit | sharedBit | managedBit | memlessBit;
 }
 
-bool MVKPhysicalDevice::getHasUnifiedMemory() {
+MVK_PUBLIC_SYMBOL MTLStorageMode MVKPhysicalDevice::getMTLStorageModeFromVkMemoryPropertyFlags(VkMemoryPropertyFlags vkFlags) {
+
+ // If not visible to the host, use Private, or Memoryless if available and lazily allocated.
+ if ( !mvkAreAllFlagsEnabled(vkFlags, VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT) ) {
+#if MVK_APPLE_SILICON
+ if (mvkAreAllFlagsEnabled(vkFlags, VK_MEMORY_PROPERTY_LAZILY_ALLOCATED_BIT)) {
+ return MTLStorageModeMemoryless;
+ }
+#endif
+ return MTLStorageModePrivate;
+ }
+
+ // If visible to the host and coherent: Shared
+ if (mvkAreAllFlagsEnabled(vkFlags, VK_MEMORY_PROPERTY_HOST_COHERENT_BIT)) {
+ return MTLStorageModeShared;
+ }
+
+ // If visible to the host, but not coherent: Shared on Apple Silicon, Managed on other GPUs.
 #if MVK_MACOS
- return ([_mtlDevice respondsToSelector: @selector(hasUnifiedMemory)]
- ? _mtlDevice.hasUnifiedMemory : _mtlDevice.isLowPower);
+ return _isAppleGPU ? MTLStorageModeShared : MTLStorageModeManaged;
 #else
- return true;
+ return MTLStorageModeShared;
 #endif
 }
 
 uint64_t MVKPhysicalDevice::getVRAMSize() {
- if (getHasUnifiedMemory()) {
+ if (_hasUnifiedMemory) {
  return mvkGetSystemMemorySize();
  } else {
  // There's actually no way to query the total physical VRAM on the device in Metal.
@@ -3408,7 +3431,7 @@ static uint32_t mvkGetEntryProperty(io_registry_entry_t entry, CFStringRef prope
  switch (getMVKConfig().semaphoreSupportStyle) {
  case MVK_CONFIG_VK_SEMAPHORE_SUPPORT_STYLE_METAL_EVENTS_WHERE_SAFE: {
  bool isNVIDIA = _properties.vendorID == kNVVendorId;
- bool isRosetta2 = _properties.vendorID == kAppleVendorId && !MVK_APPLE_SILICON;
+ bool isRosetta2 = _isAppleGPU && !MVK_APPLE_SILICON;
  if (_metalFeatures.events && !(isRosetta2 || isNVIDIA)) { _vkSemaphoreStyle = MVKSemaphoreStyleUseMTLEvent; }
  break;
  }

MoltenVK/MoltenVK/GPUObjects/MVKDeviceMemory.mm

@@ -84,7 +84,7 @@
  if (memSize == 0 || !isMemoryHostAccessible()) { return VK_SUCCESS; }
 
 #if MVK_MACOS
- if (_mtlBuffer && _mtlStorageMode == MTLStorageModeManaged) {
+ if ( !isUnifiedMemoryGPU() && _mtlBuffer && _mtlStorageMode == MTLStorageModeManaged) {
  [_mtlBuffer didModifyRange: NSMakeRange(offset, memSize)];
  }
 #endif
@@ -106,7 +106,7 @@
  if (memSize == 0 || !isMemoryHostAccessible()) { return VK_SUCCESS; }
 
 #if MVK_MACOS
- if (pBlitEnc && _mtlBuffer && _mtlStorageMode == MTLStorageModeManaged) {
+ if ( !isUnifiedMemoryGPU() && 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];
@@ -285,7 +285,7 @@
  // Set Metal memory parameters
  _vkMemAllocFlags = 0;
  _vkMemPropFlags = _device->_pMemoryProperties->memoryTypes[pAllocateInfo->memoryTypeIndex].propertyFlags;
- _mtlStorageMode = mvkMTLStorageModeFromVkMemoryPropertyFlags(_vkMemPropFlags);
+ _mtlStorageMode = getPhysicalDevice()->getMTLStorageModeFromVkMemoryPropertyFlags(_vkMemPropFlags);
  _mtlCPUCacheMode = mvkMTLCPUCacheModeFromVkMemoryPropertyFlags(_vkMemPropFlags);
 
  _allocationSize = pAllocateInfo->allocationSize;

MoltenVK/MoltenVK/GPUObjects/MVKImage.mm

@@ -469,18 +469,18 @@
 // texture and host memory for the purpose of the host reading texture memory.
 bool MVKImageMemoryBinding::needsHostReadSync(MVKPipelineBarrier& barrier) {
 #if MVK_MACOS
- return ((barrier.newLayout == VK_IMAGE_LAYOUT_GENERAL) &&
+ return ( !isUnifiedMemoryGPU() && (barrier.newLayout == VK_IMAGE_LAYOUT_GENERAL) &&
  mvkIsAnyFlagEnabled(barrier.dstAccessMask, (VK_ACCESS_HOST_READ_BIT | VK_ACCESS_MEMORY_READ_BIT)) &&
  isMemoryHostAccessible() && (!_device->_pMetalFeatures->sharedLinearTextures || !isMemoryHostCoherent()));
 #else
- return false;
+ return false;
 #endif
 }
 
 bool MVKImageMemoryBinding::shouldFlushHostMemory() { return isMemoryHostAccessible() && (!_mtlTexelBuffer || _ownsTexelBuffer); }
 
-// Flushes the device memory at the specified memory range into the MTLTexture. Updates
-// all subresources that overlap the specified range and are in an updatable layout state.
+// Flushes the memory at the specified memory range into the MTLTexture. 
+// Updates all subresources that overlap the specified range and are in an updatable layout state.
 VkResult MVKImageMemoryBinding::flushToDevice(VkDeviceSize offset, VkDeviceSize size) {
  if (shouldFlushHostMemory()) {
  for(uint8_t planeIndex = beginPlaneIndex(); planeIndex < endPlaneIndex(); planeIndex++) {
@@ -501,7 +501,7 @@
  return VK_SUCCESS;
 }
 
-// Pulls content from the MTLTexture into the device memory at the specified memory range.
+// Pulls content from the MTLTexture into memory at the specified memory range.
 // Pulls from all subresources that overlap the specified range and are in an updatable layout state.
 VkResult MVKImageMemoryBinding::pullFromDevice(VkDeviceSize offset, VkDeviceSize size) {
  if (shouldFlushHostMemory()) {
@@ -715,7 +715,7 @@ static MTLRegion getMTLRegion(const ImgRgn& imgRgn) {
 #if MVK_MACOS
  // On macOS, if the device doesn't have unified memory, and the texture is using managed memory, we need
  // to sync the managed memory from the GPU, so the texture content is accessible to be copied by the CPU.
- if ( !getPhysicalDevice()->getHasUnifiedMemory() && getMTLStorageMode() == MTLStorageModeManaged ) {
+ if ( !isUnifiedMemoryGPU() && getMTLStorageMode() == MTLStorageModeManaged ) {
  @autoreleasepool {
  id<MTLCommandBuffer> mtlCmdBuff = getDevice()->getAnyQueue()->getMTLCommandBuffer(kMVKCommandUseCopyImageToMemory);
  id<MTLBlitCommandEncoder> mtlBlitEnc = [mtlCmdBuff blitCommandEncoder];
@@ -858,9 +858,9 @@ static MTLRegion getMTLRegion(const ImgRgn& imgRgn) {
  pMemoryRequirements->memoryTypeBits = (_isDepthStencilAttachment)
  ? mvkPD->getPrivateMemoryTypes()
  : mvkPD->getAllMemoryTypes();
+ // Metal on non-Apple GPUs does not provide native support for host-coherent memory, but Vulkan requires it for Linear images
 #if MVK_MACOS
- // Metal on macOS does not provide native support for host-coherent memory, but Vulkan requires it for Linear images
- if ( !_isLinear ) {
+ if ( !isAppleGPU() && !_isLinear ) {
  mvkDisableFlags(pMemoryRequirements->memoryTypeBits, mvkPD->getHostCoherentMemoryTypes());
  }
 #endif
@@ -1052,6 +1052,7 @@ static MTLRegion getMTLRegion(const ImgRgn& imgRgn) {
 
 #if MVK_MACOS
  // For macOS prior to 10.15.5, textures cannot use Shared storage mode, so change to Managed storage mode.
+ // All Apple GPUs support shared linear textures, so this only applies to other GPUs.
  if (stgMode == MTLStorageModeShared && !_device->_pMetalFeatures->sharedLinearTextures) {
  stgMode = MTLStorageModeManaged;
  }

MoltenVK/MoltenVK/Vulkan/mvk_datatypes.mm

@@ -882,31 +882,6 @@ MVK_PUBLIC_SYMBOL CGRect mvkCGRectFromVkRectLayerKHR(VkRectLayerKHR vkRect) {
 #pragma mark -
 #pragma mark Memory options
 
-MVK_PUBLIC_SYMBOL MTLStorageMode mvkMTLStorageModeFromVkMemoryPropertyFlags(VkMemoryPropertyFlags vkFlags) {
-
- // If not visible to the host, use Private, or Memoryless if available and lazily allocated.
- if ( !mvkAreAllFlagsEnabled(vkFlags, VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT) ) {
-#if MVK_APPLE_SILICON
- if (mvkAreAllFlagsEnabled(vkFlags, VK_MEMORY_PROPERTY_LAZILY_ALLOCATED_BIT)) {
- return MTLStorageModeMemoryless;
- }
-#endif
- return MTLStorageModePrivate;
- }
-
- // If visible to the host and coherent: Shared
- if (mvkAreAllFlagsEnabled(vkFlags, VK_MEMORY_PROPERTY_HOST_COHERENT_BIT)) {
- return MTLStorageModeShared;
- }
-
- // If visible to the host, and not coherent: Managed on macOS, Shared on iOS
-#if MVK_MACOS
- return MTLStorageModeManaged;
-#else
- return MTLStorageModeShared;
-#endif
-}
-
 MVK_PUBLIC_SYMBOL MTLCPUCacheMode mvkMTLCPUCacheModeFromVkMemoryPropertyFlags(VkMemoryPropertyFlags vkFlags) {
  return MTLCPUCacheModeDefaultCache;
 }