Metal: use staging buffer for updating non-dynamic buffers
For non-dynamic (i.e. GPU-resident) buffers, onUpdateData() uses a
staging buffer rather than map() and unmap(). In addition, map() and
unmap() do nothing for non-dynamic buffers.
Also removes fOffset member, as it was never anything but 0.
Change-Id: I8bcd36c10c14d21395693686696462cca79a0e70
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/433678
Reviewed-by: Greg Daniel <egdaniel@google.com>
Commit-Queue: Jim Van Verth <jvanverth@google.com>
diff --git a/src/gpu/mtl/GrMtlBuffer.h b/src/gpu/mtl/GrMtlBuffer.h
index 5d89909..5716779 100644
--- a/src/gpu/mtl/GrMtlBuffer.h
+++ b/src/gpu/mtl/GrMtlBuffer.h
@@ -24,7 +24,6 @@
~GrMtlBuffer() override;
id<MTLBuffer> mtlBuffer() const { return fMtlBuffer; }
- size_t offset() const { return fOffset; }
protected:
GrMtlBuffer(GrMtlGpu*, size_t size, GrGpuBufferType intendedType, GrAccessPattern);
@@ -48,8 +47,6 @@
bool fIsDynamic;
id<MTLBuffer> fMtlBuffer;
- size_t fOffset; // offset into shared buffer for dynamic buffers
- id<MTLBuffer> fMappedBuffer; // buffer used by static buffers for uploads
using INHERITED = GrGpuBuffer;
};
diff --git a/src/gpu/mtl/GrMtlBuffer.mm b/src/gpu/mtl/GrMtlBuffer.mm
index d942235..ea54618 100644
--- a/src/gpu/mtl/GrMtlBuffer.mm
+++ b/src/gpu/mtl/GrMtlBuffer.mm
@@ -7,6 +7,7 @@
#include "include/private/GrTypesPriv.h"
#include "src/gpu/GrGpuResourcePriv.h"
+#include "src/gpu/GrStagingBufferManager.h"
#include "src/gpu/mtl/GrMtlBuffer.h"
#include "src/gpu/mtl/GrMtlCommandBuffer.h"
#include "src/gpu/mtl/GrMtlGpu.h"
@@ -46,8 +47,7 @@
GrMtlBuffer::GrMtlBuffer(GrMtlGpu* gpu, size_t size, GrGpuBufferType intendedType,
GrAccessPattern accessPattern)
: INHERITED(gpu, size, intendedType, accessPattern)
- , fIsDynamic(accessPattern != kStatic_GrAccessPattern)
- , fOffset(0) {
+ , fIsDynamic(accessPattern != kStatic_GrAccessPattern) {
NSUInteger options = 0;
if (@available(macOS 10.11, iOS 9.0, *)) {
if (fIsDynamic) {
@@ -76,34 +76,54 @@
}
GrMtlBuffer::~GrMtlBuffer() {
- SkASSERT(fMtlBuffer == nil);
- SkASSERT(fMappedBuffer == nil);
- SkASSERT(fMapPtr == nullptr);
+ SkASSERT(!fMtlBuffer);
+ SkASSERT(!fMapPtr);
}
-bool GrMtlBuffer::onUpdateData(const void* src, size_t srcInBytes) {
- if (!fIsDynamic) {
- if (fMtlBuffer == nil) {
- return false;
- }
- if (srcInBytes > fMtlBuffer.length) {
- return false;
- }
- }
- VALIDATE();
-
- this->internalMap(srcInBytes);
- if (fMapPtr == nil) {
+bool GrMtlBuffer::onUpdateData(const void* src, size_t sizeInBytes) {
+ if (this->wasDestroyed()) {
return false;
}
- SkASSERT(fMappedBuffer);
- if (!fIsDynamic) {
- SkASSERT(SkAlign4(srcInBytes) == fMappedBuffer.length);
- }
- memcpy(fMapPtr, src, srcInBytes);
- this->internalUnmap(srcInBytes);
- VALIDATE();
+ if (sizeInBytes > this->size()) {
+ return false;
+ }
+
+ if (fIsDynamic) {
+ this->internalMap(sizeInBytes);
+ if (!fMapPtr) {
+ return false;
+ }
+ memcpy(fMapPtr, src, sizeInBytes);
+ this->internalUnmap(sizeInBytes);
+ } else {
+ // copy data to gpu buffer
+ GrStagingBufferManager::Slice slice;
+ #ifdef SK_BUILD_FOR_MAC
+ // Mac requires 4-byte alignment for copies
+ // TODO: true for Apple Silicon?
+ static const size_t kMinAlignment = 4;
+ #else
+ static const size_t kMinAlignment = 1;
+ #endif
+ slice = this->mtlGpu()->stagingBufferManager()->allocateStagingBufferSlice(sizeInBytes,
+ kMinAlignment);
+ if (!slice.fBuffer) {
+ return false;
+ }
+ memcpy(slice.fOffsetMapPtr, src, sizeInBytes);
+
+ GrMtlCommandBuffer* cmdBuffer = this->mtlGpu()->commandBuffer();
+ id<MTLBlitCommandEncoder> GR_NORETAIN blitCmdEncoder = cmdBuffer->getBlitCommandEncoder();
+ GrMtlBuffer* mtlBuffer = static_cast<GrMtlBuffer*>(slice.fBuffer);
+ id<MTLBuffer> transferBuffer = mtlBuffer->mtlBuffer();
+ [blitCmdEncoder copyFromBuffer: transferBuffer
+ sourceOffset: slice.fOffset
+ toBuffer: fMtlBuffer
+ destinationOffset: 0
+ size: sizeInBytes];
+ }
+
return true;
}
@@ -114,7 +134,6 @@
void GrMtlBuffer::onAbandon() {
fMtlBuffer = nil;
- fMappedBuffer = nil;
fMapPtr = nullptr;
VALIDATE();
INHERITED::onAbandon();
@@ -124,7 +143,6 @@
if (!this->wasDestroyed()) {
VALIDATE();
fMtlBuffer = nil;
- fMappedBuffer = nil;
fMapPtr = nullptr;
VALIDATE();
}
@@ -132,73 +150,38 @@
}
void GrMtlBuffer::internalMap(size_t sizeInBytes) {
- if (this->wasDestroyed()) {
- return;
- }
- VALIDATE();
- SkASSERT(!this->isMapped());
if (fIsDynamic) {
- fMappedBuffer = fMtlBuffer;
- fMapPtr = static_cast<char*>(fMtlBuffer.contents) + fOffset;
- } else {
- SkASSERT(fMtlBuffer);
- SkASSERT(fMappedBuffer == nil);
- NSUInteger options = 0;
- if (@available(macOS 10.11, iOS 9.0, *)) {
- options |= MTLResourceStorageModeShared;
- }
-#ifdef SK_BUILD_FOR_MAC
- // Mac requires 4-byte alignment for copies so we pad this out
- sizeInBytes = SkAlign4(sizeInBytes);
-#endif
- fMappedBuffer =
- [this->mtlGpu()->device() newBufferWithLength: sizeInBytes
- options: options];
- fMapPtr = fMappedBuffer.contents;
+ VALIDATE();
+ SkASSERT(sizeInBytes <= this->size());
+ SkASSERT(!this->isMapped());
+ fMapPtr = static_cast<char*>(fMtlBuffer.contents);
+ VALIDATE();
}
- VALIDATE();
}
void GrMtlBuffer::internalUnmap(size_t sizeInBytes) {
SkASSERT(fMtlBuffer);
- if (this->wasDestroyed()) {
- return;
- }
- VALIDATE();
- SkASSERT(this->isMapped());
- if (fMtlBuffer == nil) {
- fMappedBuffer = nil;
- fMapPtr = nullptr;
- return;
- }
-#ifdef SK_BUILD_FOR_MAC
- // In both cases the size needs to be 4-byte aligned on Mac
- sizeInBytes = SkAlign4(sizeInBytes);
-#endif
if (fIsDynamic) {
+ VALIDATE();
+ SkASSERT(sizeInBytes <= this->size());
+ SkASSERT(this->isMapped());
#ifdef SK_BUILD_FOR_MAC
- SkASSERT(0 == (fOffset & 0x3)); // should be 4-byte aligned
- [fMtlBuffer didModifyRange: NSMakeRange(fOffset, sizeInBytes)];
+ [fMtlBuffer didModifyRange: NSMakeRange(0, sizeInBytes)];
#endif
- } else {
- GrMtlCommandBuffer* cmdBuffer = this->mtlGpu()->commandBuffer();
- id<MTLBlitCommandEncoder> GR_NORETAIN blitCmdEncoder = cmdBuffer->getBlitCommandEncoder();
- [blitCmdEncoder copyFromBuffer: fMappedBuffer
- sourceOffset: 0
- toBuffer: fMtlBuffer
- destinationOffset: 0
- size: sizeInBytes];
+ fMapPtr = nullptr;
}
- fMappedBuffer = nil;
- fMapPtr = nullptr;
}
void GrMtlBuffer::onMap() {
- this->internalMap(this->size());
+ if (!this->wasDestroyed()) {
+ this->internalMap(this->size());
+ }
}
void GrMtlBuffer::onUnmap() {
- this->internalUnmap(this->size());
+ if (!this->wasDestroyed()) {
+ this->internalUnmap(this->size());
+ }
}
#ifdef SK_DEBUG
@@ -210,8 +193,7 @@
this->intendedType() == GrGpuBufferType::kXferGpuToCpu ||
this->intendedType() == GrGpuBufferType::kDrawIndirect ||
this->intendedType() == GrGpuBufferType::kUniform);
- SkASSERT(fMappedBuffer == nil || fMtlBuffer == nil ||
- fMappedBuffer.length <= fMtlBuffer.length);
+ SkASSERT((fMapPtr && fMtlBuffer) || !fMapPtr);
}
#endif
diff --git a/src/gpu/mtl/GrMtlGpu.h b/src/gpu/mtl/GrMtlGpu.h
index 9a1b5a4..b6d31df 100644
--- a/src/gpu/mtl/GrMtlGpu.h
+++ b/src/gpu/mtl/GrMtlGpu.h
@@ -48,6 +48,8 @@
GrMtlResourceProvider& resourceProvider() { return fResourceProvider; }
+ GrStagingBufferManager* stagingBufferManager() override { return &fStagingBufferManager; }
+
GrMtlCommandBuffer* commandBuffer();
enum SyncQueue {
@@ -124,7 +126,6 @@
void xferBarrier(GrRenderTarget*, GrXferBarrierType) override {}
- GrStagingBufferManager* stagingBufferManager() override { return &fStagingBufferManager; }
void takeOwnershipOfBuffer(sk_sp<GrGpuBuffer>) override;
GrBackendTexture onCreateBackendTexture(SkISize dimensions,
diff --git a/src/gpu/mtl/GrMtlGpu.mm b/src/gpu/mtl/GrMtlGpu.mm
index 5d22155..fcb44c4 100644
--- a/src/gpu/mtl/GrMtlGpu.mm
+++ b/src/gpu/mtl/GrMtlGpu.mm
@@ -1419,7 +1419,7 @@
[blitCmdEncoder pushDebugGroup:@"onTransferPixelsTo"];
#endif
[blitCmdEncoder copyFromBuffer: mtlBuffer
- sourceOffset: offset + grMtlBuffer->offset()
+ sourceOffset: offset
sourceBytesPerRow: rowBytes
sourceBytesPerImage: rowBytes*rect.height()
sourceSize: MTLSizeMake(rect.width(), rect.height(), 1)
@@ -1465,7 +1465,7 @@
rect,
bufferColorType,
grMtlBuffer->mtlBuffer(),
- offset + grMtlBuffer->offset(),
+ offset,
transBufferImageBytes,
transBufferRowBytes);
}
diff --git a/src/gpu/mtl/GrMtlOpsRenderPass.mm b/src/gpu/mtl/GrMtlOpsRenderPass.mm
index 95353ae..ed7aa7d 100644
--- a/src/gpu/mtl/GrMtlOpsRenderPass.mm
+++ b/src/gpu/mtl/GrMtlOpsRenderPass.mm
@@ -353,7 +353,7 @@
fCurrentVertexStride * baseVertex, 0);
auto mtlIndexBuffer = static_cast<const GrMtlBuffer*>(fActiveIndexBuffer.get());
- size_t indexOffset = mtlIndexBuffer->offset() + sizeof(uint16_t) * baseIndex;
+ size_t indexOffset = sizeof(uint16_t) * baseIndex;
id<MTLBuffer> indexBuffer = mtlIndexBuffer->mtlBuffer();
fActiveRenderCmdEncoder->drawIndexedPrimitives(fActivePrimitiveType, indexCount,
MTLIndexTypeUInt16, indexBuffer, indexOffset);
@@ -405,7 +405,7 @@
this->setVertexBuffer(fActiveRenderCmdEncoder, fActiveVertexBuffer.get(), 0, 0);
auto mtlIndexBuffer = static_cast<const GrMtlBuffer*>(fActiveIndexBuffer.get());
- size_t indexOffset = mtlIndexBuffer->offset() + sizeof(uint16_t) * baseIndex;
+ size_t indexOffset = sizeof(uint16_t) * baseIndex;
if (@available(macOS 10.11, iOS 9.0, *)) {
fActiveRenderCmdEncoder->drawIndexedPrimitives(fActivePrimitiveType, indexCount,
MTLIndexTypeUInt16,
@@ -473,7 +473,7 @@
auto mtlIndexBuffer = static_cast<const GrMtlBuffer*>(fActiveIndexBuffer.get());
auto mtlIndirectBuffer = static_cast<const GrMtlBuffer*>(drawIndirectBuffer);
- size_t indexOffset = mtlIndexBuffer->offset();
+ size_t indexOffset = 0;
const size_t stride = sizeof(GrDrawIndexedIndirectCommand);
while (drawCount >= 1) {
@@ -512,7 +512,7 @@
auto mtlBuffer = static_cast<const GrMtlBuffer*>(buffer);
id<MTLBuffer> mtlVertexBuffer = mtlBuffer->mtlBuffer();
SkASSERT(mtlVertexBuffer);
- size_t offset = mtlBuffer->offset() + vertexOffset;
+ size_t offset = vertexOffset;
encoder->setVertexBuffer(mtlVertexBuffer, offset, index);
}
