Move GrMtlBufferManager functionality to better places
This is setting up for handling buffer recycling in the resource provider. Bug: skia:8243 Change-Id: Ie850b68bdcac455128aa3e89ae8adf62f17bddea Reviewed-on: https://skia-review.googlesource.com/c/skia/+/211988 Reviewed-by: Brian Osman <brianosman@google.com> Commit-Queue: Jim Van Verth <jvanverth@google.com>
This commit is contained in:
parent
fd24b1fe84
commit
35a67eba61
@ -54,27 +54,4 @@ private:
|
||||
typedef GrGpuBuffer INHERITED;
|
||||
};
|
||||
|
||||
class GrMtlBufferManager {
|
||||
public:
|
||||
GrMtlBufferManager(GrMtlGpu* gpu)
|
||||
: fGpu(gpu), fBufferAllocation(nil), fAllocationSize(0), fNextOffset(0) {}
|
||||
|
||||
~GrMtlBufferManager() {
|
||||
fBufferAllocation = nil; // Just to be sure
|
||||
}
|
||||
|
||||
id<MTLBuffer> getDynamicAllocation(size_t size, size_t* offset);
|
||||
void setVertexBuffer(id<MTLRenderCommandEncoder>, const GrMtlBuffer*, size_t index);
|
||||
void setFragmentBuffer(id<MTLRenderCommandEncoder>, const GrMtlBuffer*, size_t index);
|
||||
void resetBindings();
|
||||
|
||||
private:
|
||||
GrMtlGpu* fGpu;
|
||||
id<MTLBuffer> fBufferAllocation;
|
||||
size_t fAllocationSize;
|
||||
size_t fNextOffset;
|
||||
static constexpr size_t kNumBindings = GrMtlUniformHandler::kLastUniformBinding + 3;
|
||||
id<MTLBuffer> fBufferBindings[kNumBindings];
|
||||
};
|
||||
|
||||
#endif
|
||||
|
@ -106,7 +106,7 @@ void GrMtlBuffer::internalMap(size_t sizeInBytes) {
|
||||
VALIDATE();
|
||||
SkASSERT(!this->isMapped());
|
||||
if (fIsDynamic) {
|
||||
fMtlBuffer = this->mtlGpu()->bufferManager().getDynamicAllocation(sizeInBytes, &fOffset);
|
||||
fMtlBuffer = this->mtlGpu()->resourceProvider().getDynamicBuffer(sizeInBytes, &fOffset);
|
||||
fMappedBuffer = fMtlBuffer;
|
||||
fMapPtr = static_cast<char*>(fMtlBuffer.contents) + fOffset;
|
||||
} else {
|
||||
@ -171,96 +171,3 @@ void GrMtlBuffer::validate() const {
|
||||
fMappedBuffer.length <= fMtlBuffer.length);
|
||||
}
|
||||
#endif
|
||||
|
||||
id<MTLBuffer> GrMtlBufferManager::getDynamicAllocation(size_t size, size_t* offset) {
|
||||
static size_t kSharedDynamicBufferSize = 16*1024;
|
||||
|
||||
// The idea here is that we create a ring buffer which is used for all dynamic allocations
|
||||
// below a certain size. When a dynamic GrMtlBuffer is mapped, it grabs a portion of this
|
||||
// buffer and uses it. On a subsequent map it will grab a different portion of the buffer.
|
||||
// This prevents the buffer from overwriting itself before it's submitted to the command
|
||||
// stream.
|
||||
|
||||
// Create a new buffer if we need to.
|
||||
// If the requested size is larger than the shared buffer size, then we'll
|
||||
// just make the allocation and the owning GrMtlBuffer will manage it (this
|
||||
// only happens with buffers created by GrBufferAllocPool).
|
||||
//
|
||||
// TODO: By sending addCompletedHandler: to MTLCommandBuffer we can track when buffers
|
||||
// are no longer in use and recycle them rather than creating a new one each time.
|
||||
if (fAllocationSize - fNextOffset < size) {
|
||||
size_t allocSize = (size >= kSharedDynamicBufferSize) ? size : kSharedDynamicBufferSize;
|
||||
id<MTLBuffer> buffer;
|
||||
SK_BEGIN_AUTORELEASE_BLOCK
|
||||
buffer = [fGpu->device() newBufferWithLength: allocSize
|
||||
#ifdef SK_BUILD_FOR_MAC
|
||||
options: MTLResourceStorageModeManaged];
|
||||
#else
|
||||
options: MTLResourceStorageModeShared];
|
||||
#endif
|
||||
SK_END_AUTORELEASE_BLOCK
|
||||
if (nil == buffer) {
|
||||
return nil;
|
||||
}
|
||||
|
||||
if (size >= kSharedDynamicBufferSize) {
|
||||
*offset = 0;
|
||||
return buffer;
|
||||
}
|
||||
|
||||
fBufferAllocation = buffer;
|
||||
fNextOffset = 0;
|
||||
fAllocationSize = kSharedDynamicBufferSize;
|
||||
}
|
||||
|
||||
// Grab the next available block
|
||||
*offset = fNextOffset;
|
||||
fNextOffset += size;
|
||||
// Uniform buffer offsets need to be aligned to the nearest 256-byte boundary.
|
||||
fNextOffset = GrSizeAlignUp(fNextOffset, 256);
|
||||
|
||||
return fBufferAllocation;
|
||||
}
|
||||
|
||||
void GrMtlBufferManager::setVertexBuffer(id<MTLRenderCommandEncoder> encoder,
|
||||
const GrMtlBuffer* buffer,
|
||||
size_t index) {
|
||||
SkASSERT(index < 4);
|
||||
id<MTLBuffer> mtlVertexBuffer = buffer->mtlBuffer();
|
||||
SkASSERT(mtlVertexBuffer);
|
||||
// Apple recommends using setVertexBufferOffset: when changing the offset
|
||||
// for a currently bound vertex buffer, rather than setVertexBuffer:
|
||||
if (fBufferBindings[index] != mtlVertexBuffer) {
|
||||
[encoder setVertexBuffer: mtlVertexBuffer
|
||||
offset: 0
|
||||
atIndex: index];
|
||||
fBufferBindings[index] = mtlVertexBuffer;
|
||||
}
|
||||
[encoder setVertexBufferOffset: buffer->offset()
|
||||
atIndex: index];
|
||||
}
|
||||
|
||||
void GrMtlBufferManager::setFragmentBuffer(id<MTLRenderCommandEncoder> encoder,
|
||||
const GrMtlBuffer* buffer,
|
||||
size_t index) {
|
||||
SkASSERT(index < kNumBindings);
|
||||
id<MTLBuffer> mtlFragmentBuffer = buffer->mtlBuffer();
|
||||
// Apple recommends using setFragmentBufferOffset: when changing the offset
|
||||
// for a currently bound fragment buffer, rather than setFragmentBuffer:
|
||||
if (mtlFragmentBuffer) {
|
||||
if (fBufferBindings[index] != mtlFragmentBuffer) {
|
||||
[encoder setFragmentBuffer: mtlFragmentBuffer
|
||||
offset: 0
|
||||
atIndex: index];
|
||||
fBufferBindings[index] = mtlFragmentBuffer;
|
||||
}
|
||||
[encoder setFragmentBufferOffset: buffer->offset()
|
||||
atIndex: index];
|
||||
}
|
||||
}
|
||||
|
||||
void GrMtlBufferManager::resetBindings() {
|
||||
for (size_t i = 0; i < kNumBindings; ++i) {
|
||||
fBufferBindings[i] = nil;
|
||||
}
|
||||
}
|
||||
|
@ -13,7 +13,6 @@
|
||||
#include "src/gpu/GrGpu.h"
|
||||
#include "src/gpu/GrSemaphore.h"
|
||||
|
||||
#include "src/gpu/mtl/GrMtlBuffer.h"
|
||||
#include "src/gpu/mtl/GrMtlCaps.h"
|
||||
#include "src/gpu/mtl/GrMtlCopyManager.h"
|
||||
#include "src/gpu/mtl/GrMtlResourceProvider.h"
|
||||
@ -49,8 +48,6 @@ public:
|
||||
|
||||
GrMtlResourceProvider& resourceProvider() { return fResourceProvider; }
|
||||
|
||||
GrMtlBufferManager& bufferManager() { return fBufferManager; }
|
||||
|
||||
GrMtlCommandBuffer* commandBuffer();
|
||||
|
||||
enum SyncQueue {
|
||||
@ -230,7 +227,6 @@ private:
|
||||
|
||||
GrMtlCopyManager fCopyManager;
|
||||
GrMtlResourceProvider fResourceProvider;
|
||||
GrMtlBufferManager fBufferManager;
|
||||
|
||||
bool fDisconnected;
|
||||
|
||||
|
@ -101,7 +101,6 @@ GrMtlGpu::GrMtlGpu(GrContext* context, const GrContextOptions& options,
|
||||
, fCompiler(new SkSL::Compiler())
|
||||
, fCopyManager(this)
|
||||
, fResourceProvider(this)
|
||||
, fBufferManager(this)
|
||||
, fDisconnected(false) {
|
||||
fMtlCaps.reset(new GrMtlCaps(options, fDevice, featureSet));
|
||||
fCaps = fMtlCaps;
|
||||
|
@ -16,6 +16,7 @@
|
||||
#import <metal/metal.h>
|
||||
|
||||
typedef uint32_t GrColor;
|
||||
class GrMtlBuffer;
|
||||
class GrMtlPipelineState;
|
||||
class GrMtlRenderTarget;
|
||||
|
||||
@ -121,6 +122,9 @@ private:
|
||||
const GrBuffer* instanceBuffer, int instanceCount,
|
||||
int baseInstance, GrPrimitiveRestart) final;
|
||||
|
||||
void setVertexBuffer(id<MTLRenderCommandEncoder>, const GrMtlBuffer*, size_t index);
|
||||
void resetBufferBindings();
|
||||
|
||||
GrMtlGpu* fGpu;
|
||||
// GrRenderTargetProxy bounds
|
||||
#ifdef SK_DEBUG
|
||||
@ -138,6 +142,9 @@ private:
|
||||
|
||||
CommandBufferInfo fCommandBufferInfo;
|
||||
|
||||
static constexpr size_t kNumBindings = GrMtlUniformHandler::kLastUniformBinding + 3;
|
||||
id<MTLBuffer> fBufferBindings[kNumBindings];
|
||||
|
||||
typedef GrGpuRTCommandBuffer INHERITED;
|
||||
};
|
||||
|
||||
|
@ -232,7 +232,7 @@ void GrMtlGpuRTCommandBuffer::onDraw(const GrPrimitiveProcessor& primProc,
|
||||
}
|
||||
|
||||
fActiveRenderCmdEncoder = nil;
|
||||
fGpu->bufferManager().resetBindings();
|
||||
this->resetBufferBindings();
|
||||
fCommandBufferInfo.fBounds.join(bounds);
|
||||
SK_END_AUTORELEASE_BLOCK
|
||||
}
|
||||
@ -330,16 +330,14 @@ void GrMtlGpuRTCommandBuffer::bindGeometry(const GrBuffer* vertexBuffer,
|
||||
SkASSERT(!static_cast<const GrGpuBuffer*>(vertexBuffer)->isMapped());
|
||||
|
||||
const GrMtlBuffer* grMtlBuffer = static_cast<const GrMtlBuffer*>(vertexBuffer);
|
||||
fGpu->bufferManager().setVertexBuffer(fActiveRenderCmdEncoder, grMtlBuffer,
|
||||
bufferIndex++);
|
||||
this->setVertexBuffer(fActiveRenderCmdEncoder, grMtlBuffer, bufferIndex++);
|
||||
}
|
||||
if (instanceBuffer) {
|
||||
SkASSERT(!instanceBuffer->isCpuBuffer());
|
||||
SkASSERT(!static_cast<const GrGpuBuffer*>(instanceBuffer)->isMapped());
|
||||
|
||||
const GrMtlBuffer* grMtlBuffer = static_cast<const GrMtlBuffer*>(instanceBuffer);
|
||||
fGpu->bufferManager().setVertexBuffer(fActiveRenderCmdEncoder, grMtlBuffer,
|
||||
bufferIndex++);
|
||||
this->setVertexBuffer(fActiveRenderCmdEncoder, grMtlBuffer, bufferIndex++);
|
||||
}
|
||||
}
|
||||
|
||||
@ -393,3 +391,27 @@ void GrMtlGpuRTCommandBuffer::sendIndexedInstancedMeshToGpu(GrPrimitiveType prim
|
||||
baseInstance:baseInstance];
|
||||
fGpu->stats()->incNumDraws();
|
||||
}
|
||||
|
||||
void GrMtlGpuRTCommandBuffer::setVertexBuffer(id<MTLRenderCommandEncoder> encoder,
|
||||
const GrMtlBuffer* buffer,
|
||||
size_t index) {
|
||||
SkASSERT(index < 4);
|
||||
id<MTLBuffer> mtlVertexBuffer = buffer->mtlBuffer();
|
||||
SkASSERT(mtlVertexBuffer);
|
||||
// Apple recommends using setVertexBufferOffset: when changing the offset
|
||||
// for a currently bound vertex buffer, rather than setVertexBuffer:
|
||||
if (fBufferBindings[index] != mtlVertexBuffer) {
|
||||
[encoder setVertexBuffer: mtlVertexBuffer
|
||||
offset: 0
|
||||
atIndex: index];
|
||||
fBufferBindings[index] = mtlVertexBuffer;
|
||||
}
|
||||
[encoder setVertexBufferOffset: buffer->offset()
|
||||
atIndex: index];
|
||||
}
|
||||
|
||||
void GrMtlGpuRTCommandBuffer::resetBufferBindings() {
|
||||
for (size_t i = 0; i < kNumBindings; ++i) {
|
||||
fBufferBindings[i] = nil;
|
||||
}
|
||||
}
|
||||
|
@ -42,6 +42,8 @@ public:
|
||||
// Finds or creates a compatible MTLSamplerState based on the GrSamplerState.
|
||||
GrMtlSampler* findOrCreateCompatibleSampler(const GrSamplerState&, uint32_t maxMipLevel);
|
||||
|
||||
id<MTLBuffer> getDynamicBuffer(size_t size, size_t* offset);
|
||||
|
||||
private:
|
||||
#ifdef SK_DEBUG
|
||||
#define GR_PIPELINE_STATE_CACHE_STATS
|
||||
@ -92,6 +94,14 @@ private:
|
||||
|
||||
SkTDynamicHash<GrMtlSampler, GrMtlSampler::Key> fSamplers;
|
||||
SkTDynamicHash<GrMtlDepthStencil, GrMtlDepthStencil::Key> fDepthStencilStates;
|
||||
|
||||
// Buffer state
|
||||
struct BufferState {
|
||||
id<MTLBuffer> fAllocation;
|
||||
size_t fAllocationSize;
|
||||
size_t fNextOffset;
|
||||
};
|
||||
BufferState fBufferState;
|
||||
};
|
||||
|
||||
#endif
|
||||
|
@ -7,6 +7,7 @@
|
||||
|
||||
#include "src/gpu/mtl/GrMtlResourceProvider.h"
|
||||
|
||||
#include "src/gpu/mtl/GrMtlCommandBuffer.h"
|
||||
#include "src/gpu/mtl/GrMtlCopyManager.h"
|
||||
#include "src/gpu/mtl/GrMtlGpu.h"
|
||||
#include "src/gpu/mtl/GrMtlPipelineState.h"
|
||||
@ -15,7 +16,8 @@
|
||||
#include "src/sksl/SkSLCompiler.h"
|
||||
|
||||
GrMtlResourceProvider::GrMtlResourceProvider(GrMtlGpu* gpu)
|
||||
: fGpu(gpu) {
|
||||
: fGpu(gpu)
|
||||
, fBufferState({nil, 0, 0}) {
|
||||
fPipelineStateCache.reset(new PipelineStateCache(gpu));
|
||||
}
|
||||
|
||||
@ -151,3 +153,53 @@ GrMtlPipelineState* GrMtlResourceProvider::PipelineStateCache::refPipelineState(
|
||||
}
|
||||
return (*entry)->fPipelineState.get();
|
||||
}
|
||||
|
||||
id<MTLBuffer> GrMtlResourceProvider::getDynamicBuffer(size_t size, size_t* offset) {
|
||||
static size_t kSharedDynamicBufferSize = 16*1024;
|
||||
|
||||
// The idea here is that we create a ring buffer which is used for all dynamic allocations
|
||||
// below a certain size. When a dynamic GrMtlBuffer is mapped, it grabs a portion of this
|
||||
// buffer and uses it. On a subsequent map it will grab a different portion of the buffer.
|
||||
// This prevents the buffer from overwriting itself before it's submitted to the command
|
||||
// stream.
|
||||
|
||||
// Create a new buffer if we need to.
|
||||
// If the requested size is larger than the shared buffer size, then we'll
|
||||
// just make the allocation and the owning GrMtlBuffer will manage it (this
|
||||
// only happens with buffers created by GrBufferAllocPool).
|
||||
//
|
||||
// TODO: By sending addCompletedHandler: to MTLCommandBuffer we can track when buffers
|
||||
// are no longer in use and recycle them rather than creating a new one each time.
|
||||
if (fBufferState.fAllocationSize - fBufferState.fNextOffset < size) {
|
||||
size_t allocSize = (size >= kSharedDynamicBufferSize) ? size : kSharedDynamicBufferSize;
|
||||
id<MTLBuffer> buffer;
|
||||
SK_BEGIN_AUTORELEASE_BLOCK
|
||||
buffer = [fGpu->device() newBufferWithLength: allocSize
|
||||
#ifdef SK_BUILD_FOR_MAC
|
||||
options: MTLResourceStorageModeManaged];
|
||||
#else
|
||||
options: MTLResourceStorageModeShared];
|
||||
#endif
|
||||
SK_END_AUTORELEASE_BLOCK
|
||||
if (nil == buffer) {
|
||||
return nil;
|
||||
}
|
||||
|
||||
if (size >= kSharedDynamicBufferSize) {
|
||||
*offset = 0;
|
||||
return buffer;
|
||||
}
|
||||
|
||||
fBufferState.fAllocation = buffer;
|
||||
fBufferState.fNextOffset = 0;
|
||||
fBufferState.fAllocationSize = kSharedDynamicBufferSize;
|
||||
}
|
||||
|
||||
// Grab the next available block
|
||||
*offset = fBufferState.fNextOffset;
|
||||
fBufferState.fNextOffset += size;
|
||||
// Uniform buffer offsets need to be aligned to the nearest 256-byte boundary.
|
||||
fBufferState.fNextOffset = GrSizeAlignUp(fBufferState.fNextOffset, 256);
|
||||
|
||||
return fBufferState.fAllocation;
|
||||
}
|
||||
|
Loading…
Reference in New Issue
Block a user