AuroraMiddleware/skia2
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Add FenceSync support to Metal

Browse Source 
Bug: skia:8243
Change-Id: Iadd2445efe86e796b23ea20edbe49d684f626d9a
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/212270
Reviewed-by: Brian Salomon <bsalomon@google.com>
Commit-Queue: Jim Van Verth <jvanverth@google.com>
This commit is contained in:
Jim Van Verth 2019-05-08 11:40:55 -04:00 committed by Skia Commit-Bot
parent 8cdf5b6fa4
commit f5ca01a69a
2 changed files with 82 additions and 64 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

14
src/gpu/mtl/GrMtlUtil.h
View File

@ -16,6 +16,20 @@
class GrMtlGpu;
class GrSurface;
#if defined(SK_BUILD_FOR_MAC)
#if __MAC_OS_X_VERSION_MAX_ALLOWED >= 101400
#define GR_METAL_SDK_VERSION 200
#else
#define GR_METAL_SDK_VERSION 100
#endif
#else
#if __IPHONE_OS_VERSION_MAX_ALLOWED >= 120000 || __TV_OS_VERSION_MAX_ALLOWED >= 120000
#define GR_METAL_SDK_VERSION 200
#else
#define GR_METAL_SDK_VERSION 100
#endif
#endif
/**
* Returns the Metal texture format for the given GrPixelConfig
*/

130
tools/gpu/mtl/MtlTestContext.mm
View File

@ -10,6 +10,8 @@
#include "include/gpu/GrContext.h"
#include "include/gpu/GrContextOptions.h"
#include "src/gpu/mtl/GrMtlUtil.h"
#ifdef SK_METAL
#import <Metal/Metal.h>
@ -19,97 +21,82 @@
#define SK_END_AUTORELEASE_BLOCK }
namespace {
#if GR_METAL_SDK_VERSION >= 200
/**
* Implements sk_gpu_test::FenceSync for Metal.
*
* Uses a single MTLSharedEvent, and inserts a GPU command to increment the value
* each time we call insertFence(). On the CPU side we use a MTLSharedEventListener to
* wait for the new value to be signaled on the GPU. Since the event listener is handled
* on a separate thread, we communicate completion to the main thread via a semaphore.
*/
// TODO
#if 0
class MtlFenceSync : public sk_gpu_test::FenceSync {
public:
MtlFenceSync(sk_sp<const GrVkInterface> vk, VkDevice device, VkQueue queue,
uint32_t queueFamilyIndex)
: fVk(std::move(vk))
, fDevice(device)
, fQueue(queue) {
MtlFenceSync(id<MTLDevice> device, id<MTLCommandQueue> queue)
: fDevice(device)
, fQueue(queue)
, fLatestEvent(0) {
SkDEBUGCODE(fUnfinishedSyncs = 0;)
VkCommandPoolCreateInfo createInfo;
createInfo.sType = VK_STRUCTURE_TYPE_COMMAND_POOL_CREATE_INFO;
createInfo.pNext = nullptr;
createInfo.flags = 0;
createInfo.queueFamilyIndex = queueFamilyIndex;
GR_VK_CALL_ERRCHECK(fVk, CreateCommandPool(fDevice, &createInfo, nullptr, &fCommandPool));
VkCommandBufferAllocateInfo allocateInfo;
allocateInfo.sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_ALLOCATE_INFO;
allocateInfo.pNext = nullptr;
allocateInfo.commandBufferCount = 1;
allocateInfo.commandPool = fCommandPool;
allocateInfo.level = VK_COMMAND_BUFFER_LEVEL_PRIMARY;
GR_VK_CALL_ERRCHECK(fVk, AllocateCommandBuffers(fDevice, &allocateInfo, &fCommandBuffer));
VkCommandBufferBeginInfo beginInfo;
beginInfo.sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO;
beginInfo.pNext = nullptr;
beginInfo.flags = VK_COMMAND_BUFFER_USAGE_SIMULTANEOUS_USE_BIT;
beginInfo.pInheritanceInfo = nullptr;
GR_VK_CALL_ERRCHECK(fVk, BeginCommandBuffer(fCommandBuffer, &beginInfo));
GR_VK_CALL_ERRCHECK(fVk, EndCommandBuffer(fCommandBuffer));
SK_BEGIN_AUTORELEASE_BLOCK
fSharedEvent = [fDevice newSharedEvent];
dispatch_queue_t queue = dispatch_queue_create("MTLFenceSync", NULL);
fSharedEventListener = [[MTLSharedEventListener alloc] initWithDispatchQueue:queue];
SK_END_AUTORELEASE_BLOCK
}
~VkFenceSync() override {
~MtlFenceSync() override {
SkASSERT(!fUnfinishedSyncs);
// If the above assertion is true then the command buffer should not be in flight.
GR_VK_CALL(fVk, FreeCommandBuffers(fDevice, fCommandPool, 1, &fCommandBuffer));
GR_VK_CALL(fVk, DestroyCommandPool(fDevice, fCommandPool, nullptr));
// ARC should take care of these:
fSharedEventListener = nil;
fSharedEvent = nil;
}
sk_gpu_test::PlatformFence SK_WARN_UNUSED_RESULT insertFence() const override {
VkFence fence;
VkFenceCreateInfo info;
info.sType = VK_STRUCTURE_TYPE_FENCE_CREATE_INFO;
info.pNext = nullptr;
info.flags = 0;
GR_VK_CALL_ERRCHECK(fVk, CreateFence(fDevice, &info, nullptr, &fence));
VkSubmitInfo submitInfo;
submitInfo.sType = VK_STRUCTURE_TYPE_SUBMIT_INFO;
submitInfo.pNext = nullptr;
submitInfo.waitSemaphoreCount = 0;
submitInfo.pWaitSemaphores = nullptr;
submitInfo.pWaitDstStageMask = nullptr;
submitInfo.commandBufferCount = 1;
submitInfo.pCommandBuffers = &fCommandBuffer;
submitInfo.signalSemaphoreCount = 0;
submitInfo.pSignalSemaphores = nullptr;
GR_VK_CALL_ERRCHECK(fVk, QueueSubmit(fQueue, 1, &submitInfo, fence));
SK_BEGIN_AUTORELEASE_BLOCK
id<MTLCommandBuffer> cmdBuffer = [fQueue commandBuffer];
++fLatestEvent;
[cmdBuffer encodeSignalEvent:fSharedEvent value:fLatestEvent];
[cmdBuffer commit];
SK_END_AUTORELEASE_BLOCK
SkDEBUGCODE(++fUnfinishedSyncs;)
return (sk_gpu_test::PlatformFence)fence;
return (sk_gpu_test::PlatformFence)fLatestEvent;
}
bool waitFence(sk_gpu_test::PlatformFence opaqueFence) const override {
VkFence fence = (VkFence)opaqueFence;
static constexpr uint64_t kForever = ~((uint64_t)0);
auto result = GR_VK_CALL(fVk, WaitForFences(fDevice, 1, &fence, true, kForever));
return result != VK_TIMEOUT;
uint64_t value = (uint64_t)opaqueFence;
dispatch_semaphore_t semaphore = dispatch_semaphore_create(0);
// Add listener for this particular value or greater
__block dispatch_semaphore_t block_sema = semaphore;
[fSharedEvent notifyListener: fSharedEventListener
atValue: value
block: ^(id<MTLSharedEvent> sharedEvent, uint64_t value) {
dispatch_semaphore_signal(block_sema);
}];
long result = dispatch_semaphore_wait(semaphore, DISPATCH_TIME_FOREVER);
return !result;
}
void deleteFence(sk_gpu_test::PlatformFence opaqueFence) const override {
VkFence fence = (VkFence)opaqueFence;
GR_VK_CALL(fVk, DestroyFence(fDevice, fence, nullptr));
// Nothing to delete
SkDEBUGCODE(--fUnfinishedSyncs;)
}
private:
sk_sp<const GrVkInterface> fVk;
VkDevice fDevice;
VkQueue fQueue;
VkCommandPool fCommandPool;
VkCommandBuffer fCommandBuffer;
id<MTLDevice> fDevice;
id<MTLCommandQueue> fQueue;
id<MTLSharedEvent> fSharedEvent;
MTLSharedEventListener* fSharedEventListener;
mutable uint64_t fLatestEvent;
SkDEBUGCODE(mutable int fUnfinishedSyncs;)
typedef sk_gpu_test::FenceSync INHERITED;
};
GR_STATIC_ASSERT(sizeof(VkFence) <= sizeof(sk_gpu_test::PlatformFence));
GR_STATIC_ASSERT(sizeof(uint64_t) <= sizeof(sk_gpu_test::PlatformFence));
#endif
class MtlTestContextImpl : public sk_gpu_test::MtlTestContext {
@ -152,8 +139,25 @@ public:
private:
MtlTestContextImpl(id<MTLDevice> device, id<MTLCommandQueue> queue)
: INHERITED(), fDevice(device), fQueue(queue) {
#if GR_METAL_SDK_VERSION >= 200
// TODO: I believe we can just check whether creating a MTLSharedEvent returns nil,
// but this needs to be tested on an old OS.
NSOperatingSystemVersion osVersion = [[NSProcessInfo processInfo] operatingSystemVersion];
#ifdef SK_BUILD_FOR_MAC
bool supportsFenceSync = (osVersion.majorVersion > 10 ||
(osVersion.majorVersion == 10 && osVersion.minorVersion >= 14));
#else
bool supportsFenceSync = (osVersion.majorVersion >= 12);
#endif
if (supportsFenceSync ) {
fFenceSync.reset(new MtlFenceSync(device, queue));
} else {
fFenceSync.reset(nullptr);
}
#else
fFenceSync.reset(nullptr);
#endif
}
void onPlatformMakeCurrent() const override {}
std::function<void()> onPlatformGetAutoContextRestore() const override { return nullptr; }