1berry/qtbase
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

rhi: Reuse the data in buffer ops in res.update batches

Browse Source 
Because having profilers bombarded with mallocs (due to creating
deep copy QByteArrays) is not nice.

Change-Id: I848f41f3465d6dc2a58a193cc863495aacf13d79
Reviewed-by: Andy Nichols <andy.nichols@qt.io>
This commit is contained in:
Laszlo Agocs 2020-09-22 19:54:30 +02:00
parent 6b52ba4286
commit be2635b8dd
6 changed files with 60 additions and 25 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

43
src/gui/rhi/qrhi_p_p.h
View File

@ -289,6 +289,7 @@ public:
QRhiBuffer *buf; QRhiBuffer *buf;
int offset; int offset;
QByteArray data; QByteArray data;
int dataSize; // the real number of currently used bytes in data, not the same as data.size()
int readSize; int readSize;
QRhiBufferReadbackResult *result; QRhiBufferReadbackResult *result;
@ -298,7 +299,9 @@ public:
op.type = DynamicUpdate; op.type = DynamicUpdate;
op.buf = buf; op.buf = buf;
op.offset = offset; op.offset = offset;
op.data = QByteArray(reinterpret_cast<const char *>(data), size ? size : buf->size()); const int effectiveSize = size ? size : buf->size();
op.data = QByteArray(reinterpret_cast<const char *>(data), effectiveSize);
op.dataSize = effectiveSize;
return op; return op;
} }
@ -307,7 +310,29 @@ public:
op->type = DynamicUpdate; op->type = DynamicUpdate;
op->buf = buf; op->buf = buf;
op->offset = offset; op->offset = offset;
op->data = QByteArray(reinterpret_cast<const char *>(data), size ? size : buf->size()); const int effectiveSize = size ? size : buf->size();
// Why the isDetached check? Simply because the cost of detaching
// with a larger allocation may be a lot higher than creating a new
// deep copy bytearray with our (potentially lot smaller) data.
// This reduces the benefits with certain backends (e.g. Vulkan)
// that hold on to the data (implicit sharing!) of host visible
// buffers for the current and next frame (assuming 2 frames in
// flight), but it is still an improvement (enabled by
// nextResourceUpdateBatch's shuffling when choosing a free batch
// from the pool). For other backends (e.g. D3D11) this can reduce
// mallocs (caused by creating new deep copy bytearrays) almost
// completely after a few frames (assuming of course that no
// dynamic elements with larger buffer data appear).
if (op->data.isDetached()) {
if (op->data.size() < effectiveSize)
op->data.resize(effectiveSize);
memcpy(op->data.data(), data, effectiveSize);
} else {
op->data = QByteArray(reinterpret_cast<const char *>(data), effectiveSize);
}
op->dataSize = effectiveSize;
} }
static BufferOp staticUpload(QRhiBuffer *buf, int offset, int size, const void *data) static BufferOp staticUpload(QRhiBuffer *buf, int offset, int size, const void *data)
@ -316,7 +341,9 @@ public:
op.type = StaticUpload; op.type = StaticUpload;
op.buf = buf; op.buf = buf;
op.offset = offset; op.offset = offset;
op.data = QByteArray(reinterpret_cast<const char *>(data), size ? size : buf->size()); const int effectiveSize = size ? size : buf->size();
op.data = QByteArray(reinterpret_cast<const char *>(data), effectiveSize);
op.dataSize = effectiveSize;
return op; return op;
} }
@ -325,7 +352,15 @@ public:
op->type = StaticUpload; op->type = StaticUpload;
op->buf = buf; op->buf = buf;
op->offset = offset; op->offset = offset;
op->data = QByteArray(reinterpret_cast<const char *>(data), size ? size : buf->size()); const int effectiveSize = size ? size : buf->size();
if (op->data.isDetached()) {
if (op->data.size() < effectiveSize)
op->data.resize(effectiveSize);
memcpy(op->data.data(), data, effectiveSize);
} else {
op->data = QByteArray(reinterpret_cast<const char *>(data), effectiveSize);
}
op->dataSize = effectiveSize;
} }
static BufferOp read(QRhiBuffer *buf, int offset, int size, QRhiBufferReadbackResult *result) static BufferOp read(QRhiBuffer *buf, int offset, int size, QRhiBufferReadbackResult *result)

6
src/gui/rhi/qrhid3d11.cpp
View File

@ -1418,12 +1418,12 @@ void QRhiD3D11::enqueueResourceUpdates(QRhiCommandBuffer *cb, QRhiResourceUpdate
if (u.type == QRhiResourceUpdateBatchPrivate::BufferOp::DynamicUpdate) { if (u.type == QRhiResourceUpdateBatchPrivate::BufferOp::DynamicUpdate) {
QD3D11Buffer *bufD = QRHI_RES(QD3D11Buffer, u.buf); QD3D11Buffer *bufD = QRHI_RES(QD3D11Buffer, u.buf);
Q_ASSERT(bufD->m_type == QRhiBuffer::Dynamic); Q_ASSERT(bufD->m_type == QRhiBuffer::Dynamic);
memcpy(bufD->dynBuf.data() + u.offset, u.data.constData(), size_t(u.data.size())); memcpy(bufD->dynBuf.data() + u.offset, u.data.constData(), size_t(u.dataSize));
bufD->hasPendingDynamicUpdates = true; bufD->hasPendingDynamicUpdates = true;
} else if (u.type == QRhiResourceUpdateBatchPrivate::BufferOp::StaticUpload) { } else if (u.type == QRhiResourceUpdateBatchPrivate::BufferOp::StaticUpload) {
QD3D11Buffer *bufD = QRHI_RES(QD3D11Buffer, u.buf); QD3D11Buffer *bufD = QRHI_RES(QD3D11Buffer, u.buf);
Q_ASSERT(bufD->m_type != QRhiBuffer::Dynamic); Q_ASSERT(bufD->m_type != QRhiBuffer::Dynamic);
Q_ASSERT(u.offset + u.data.size() <= bufD->m_size); Q_ASSERT(u.offset + u.dataSize <= bufD->m_size);
QD3D11CommandBuffer::Command cmd; QD3D11CommandBuffer::Command cmd;
cmd.cmd = QD3D11CommandBuffer::Command::UpdateSubRes; cmd.cmd = QD3D11CommandBuffer::Command::UpdateSubRes;
cmd.args.updateSubRes.dst = bufD->buffer; cmd.args.updateSubRes.dst = bufD->buffer;
@ -1437,7 +1437,7 @@ void QRhiD3D11::enqueueResourceUpdates(QRhiCommandBuffer *cb, QRhiResourceUpdate
box.left = UINT(u.offset); box.left = UINT(u.offset);
box.top = box.front = 0; box.top = box.front = 0;
box.back = box.bottom = 1; box.back = box.bottom = 1;
box.right = UINT(u.offset + u.data.size()); // no -1: right, bottom, back are exclusive, see D3D11_BOX doc box.right = UINT(u.offset + u.dataSize); // no -1: right, bottom, back are exclusive, see D3D11_BOX doc
cmd.args.updateSubRes.hasDstBox = true; cmd.args.updateSubRes.hasDstBox = true;
cmd.args.updateSubRes.dstBox = box; cmd.args.updateSubRes.dstBox = box;
cbD->commands.append(cmd); cbD->commands.append(cmd);

10
src/gui/rhi/qrhigles2.cpp
View File

@ -1687,7 +1687,7 @@ void QRhiGles2::enqueueResourceUpdates(QRhiCommandBuffer *cb, QRhiResourceUpdate
QGles2Buffer *bufD = QRHI_RES(QGles2Buffer, u.buf); QGles2Buffer *bufD = QRHI_RES(QGles2Buffer, u.buf);
Q_ASSERT(bufD->m_type == QRhiBuffer::Dynamic); Q_ASSERT(bufD->m_type == QRhiBuffer::Dynamic);
if (bufD->m_usage.testFlag(QRhiBuffer::UniformBuffer)) { if (bufD->m_usage.testFlag(QRhiBuffer::UniformBuffer)) {
memcpy(bufD->ubuf.data() + u.offset, u.data.constData(), size_t(u.data.size())); memcpy(bufD->ubuf.data() + u.offset, u.data.constData(), size_t(u.dataSize));
} else { } else {
trackedBufferBarrier(cbD, bufD, QGles2Buffer::AccessUpdate); trackedBufferBarrier(cbD, bufD, QGles2Buffer::AccessUpdate);
QGles2CommandBuffer::Command cmd; QGles2CommandBuffer::Command cmd;
@ -1695,16 +1695,16 @@ void QRhiGles2::enqueueResourceUpdates(QRhiCommandBuffer *cb, QRhiResourceUpdate
cmd.args.bufferSubData.target = bufD->targetForDataOps; cmd.args.bufferSubData.target = bufD->targetForDataOps;
cmd.args.bufferSubData.buffer = bufD->buffer; cmd.args.bufferSubData.buffer = bufD->buffer;
cmd.args.bufferSubData.offset = u.offset; cmd.args.bufferSubData.offset = u.offset;
cmd.args.bufferSubData.size = u.data.size(); cmd.args.bufferSubData.size = u.dataSize;
cmd.args.bufferSubData.data = cbD->retainData(u.data); cmd.args.bufferSubData.data = cbD->retainData(u.data);
cbD->commands.append(cmd); cbD->commands.append(cmd);
} }
} else if (u.type == QRhiResourceUpdateBatchPrivate::BufferOp::StaticUpload) { } else if (u.type == QRhiResourceUpdateBatchPrivate::BufferOp::StaticUpload) {
QGles2Buffer *bufD = QRHI_RES(QGles2Buffer, u.buf); QGles2Buffer *bufD = QRHI_RES(QGles2Buffer, u.buf);
Q_ASSERT(bufD->m_type != QRhiBuffer::Dynamic); Q_ASSERT(bufD->m_type != QRhiBuffer::Dynamic);
Q_ASSERT(u.offset + u.data.size() <= bufD->m_size); Q_ASSERT(u.offset + u.dataSize <= bufD->m_size);
if (bufD->m_usage.testFlag(QRhiBuffer::UniformBuffer)) { if (bufD->m_usage.testFlag(QRhiBuffer::UniformBuffer)) {
memcpy(bufD->ubuf.data() + u.offset, u.data.constData(), size_t(u.data.size())); memcpy(bufD->ubuf.data() + u.offset, u.data.constData(), size_t(u.dataSize));
} else { } else {
trackedBufferBarrier(cbD, bufD, QGles2Buffer::AccessUpdate); trackedBufferBarrier(cbD, bufD, QGles2Buffer::AccessUpdate);
QGles2CommandBuffer::Command cmd; QGles2CommandBuffer::Command cmd;
@ -1712,7 +1712,7 @@ void QRhiGles2::enqueueResourceUpdates(QRhiCommandBuffer *cb, QRhiResourceUpdate
cmd.args.bufferSubData.target = bufD->targetForDataOps; cmd.args.bufferSubData.target = bufD->targetForDataOps;
cmd.args.bufferSubData.buffer = bufD->buffer; cmd.args.bufferSubData.buffer = bufD->buffer;
cmd.args.bufferSubData.offset = u.offset; cmd.args.bufferSubData.offset = u.offset;
cmd.args.bufferSubData.size = u.data.size(); cmd.args.bufferSubData.size = u.dataSize;
cmd.args.bufferSubData.data = cbD->retainData(u.data); cmd.args.bufferSubData.data = cbD->retainData(u.data);
cbD->commands.append(cmd); cbD->commands.append(cmd);
} }

10
src/gui/rhi/qrhimetal.mm
View File

@ -1703,10 +1703,10 @@ void QRhiMetal::enqueueResourceUpdates(QRhiCommandBuffer *cb, QRhiResourceUpdate
// basically the same. So go through the same pendingUpdates machinery. // basically the same. So go through the same pendingUpdates machinery.
QMetalBuffer *bufD = QRHI_RES(QMetalBuffer, u.buf); QMetalBuffer *bufD = QRHI_RES(QMetalBuffer, u.buf);
Q_ASSERT(bufD->m_type != QRhiBuffer::Dynamic); Q_ASSERT(bufD->m_type != QRhiBuffer::Dynamic);
Q_ASSERT(u.offset + u.data.size() <= bufD->m_size); Q_ASSERT(u.offset + u.dataSize <= bufD->m_size);
for (int i = 0, ie = bufD->d->slotted ? QMTL_FRAMES_IN_FLIGHT : 1; i != ie; ++i) for (int i = 0, ie = bufD->d->slotted ? QMTL_FRAMES_IN_FLIGHT : 1; i != ie; ++i)
bufD->d->pendingUpdates[i].append( bufD->d->pendingUpdates[i].append(
QRhiResourceUpdateBatchPrivate::BufferOp::dynamicUpdate(u.buf, u.offset, u.data.size(), u.data.constData())); QRhiResourceUpdateBatchPrivate::BufferOp::dynamicUpdate(u.buf, u.offset, u.dataSize, u.data.constData()));
} else if (u.type == QRhiResourceUpdateBatchPrivate::BufferOp::Read) { } else if (u.type == QRhiResourceUpdateBatchPrivate::BufferOp::Read) {
QMetalBuffer *bufD = QRHI_RES(QMetalBuffer, u.buf); QMetalBuffer *bufD = QRHI_RES(QMetalBuffer, u.buf);
executeBufferHostWritesForCurrentFrame(bufD); executeBufferHostWritesForCurrentFrame(bufD);
@ -1868,11 +1868,11 @@ void QRhiMetal::executeBufferHostWritesForSlot(QMetalBuffer *bufD, int slot)
int changeEnd = -1; int changeEnd = -1;
for (const QRhiResourceUpdateBatchPrivate::BufferOp &u : qAsConst(bufD->d->pendingUpdates[slot])) { for (const QRhiResourceUpdateBatchPrivate::BufferOp &u : qAsConst(bufD->d->pendingUpdates[slot])) {
Q_ASSERT(bufD == QRHI_RES(QMetalBuffer, u.buf)); Q_ASSERT(bufD == QRHI_RES(QMetalBuffer, u.buf));
memcpy(static_cast<char *>(p) + u.offset, u.data.constData(), size_t(u.data.size())); memcpy(static_cast<char *>(p) + u.offset, u.data.constData(), size_t(u.dataSize));
if (changeBegin == -1 || u.offset < changeBegin) if (changeBegin == -1 || u.offset < changeBegin)
changeBegin = u.offset; changeBegin = u.offset;
if (changeEnd == -1 || u.offset + u.data.size() > changeEnd) if (changeEnd == -1 || u.offset + u.dataSize > changeEnd)
changeEnd = u.offset + u.data.size(); changeEnd = u.offset + u.dataSize;
} }
#ifdef Q_OS_MACOS #ifdef Q_OS_MACOS
if (changeBegin >= 0 && bufD->d->managed) if (changeBegin >= 0 && bufD->d->managed)

2
src/gui/rhi/qrhinull.cpp
View File

@ -465,7 +465,7 @@ void QRhiNull::resourceUpdate(QRhiCommandBuffer *cb, QRhiResourceUpdateBatch *re
|| u.type == QRhiResourceUpdateBatchPrivate::BufferOp::StaticUpload) || u.type == QRhiResourceUpdateBatchPrivate::BufferOp::StaticUpload)
{ {
QNullBuffer *bufD = QRHI_RES(QNullBuffer, u.buf); QNullBuffer *bufD = QRHI_RES(QNullBuffer, u.buf);
memcpy(bufD->data.data() + u.offset, u.data.constData(), size_t(u.data.size())); memcpy(bufD->data.data() + u.offset, u.data.constData(), size_t(u.dataSize));
} else if (u.type == QRhiResourceUpdateBatchPrivate::BufferOp::Read) { } else if (u.type == QRhiResourceUpdateBatchPrivate::BufferOp::Read) {
QRhiBufferReadbackResult *result = u.result; QRhiBufferReadbackResult *result = u.result;
result->data.resize(u.readSize); result->data.resize(u.readSize);

14
src/gui/rhi/qrhivulkan.cpp
View File

@ -2933,7 +2933,7 @@ void QRhiVulkan::enqueueResourceUpdates(QVkCommandBuffer *cbD, QRhiResourceUpdat
} else if (u.type == QRhiResourceUpdateBatchPrivate::BufferOp::StaticUpload) { } else if (u.type == QRhiResourceUpdateBatchPrivate::BufferOp::StaticUpload) {
QVkBuffer *bufD = QRHI_RES(QVkBuffer, u.buf); QVkBuffer *bufD = QRHI_RES(QVkBuffer, u.buf);
Q_ASSERT(bufD->m_type != QRhiBuffer::Dynamic); Q_ASSERT(bufD->m_type != QRhiBuffer::Dynamic);
Q_ASSERT(u.offset + u.data.size() <= bufD->m_size); Q_ASSERT(u.offset + u.dataSize <= bufD->m_size);
if (!bufD->stagingBuffers[currentFrameSlot]) { if (!bufD->stagingBuffers[currentFrameSlot]) {
VkBufferCreateInfo bufferInfo; VkBufferCreateInfo bufferInfo;
@ -2967,9 +2967,9 @@ void QRhiVulkan::enqueueResourceUpdates(QVkCommandBuffer *cbD, QRhiResourceUpdat
qWarning("Failed to map buffer: %d", err); qWarning("Failed to map buffer: %d", err);
continue; continue;
} }
memcpy(static_cast<uchar *>(p) + u.offset, u.data.constData(), size_t(u.data.size())); memcpy(static_cast<uchar *>(p) + u.offset, u.data.constData(), size_t(u.dataSize));
vmaUnmapMemory(toVmaAllocator(allocator), a); vmaUnmapMemory(toVmaAllocator(allocator), a);
vmaFlushAllocation(toVmaAllocator(allocator), a, VkDeviceSize(u.offset), VkDeviceSize(u.data.size())); vmaFlushAllocation(toVmaAllocator(allocator), a, VkDeviceSize(u.offset), VkDeviceSize(u.dataSize));
trackedBufferBarrier(cbD, bufD, 0, trackedBufferBarrier(cbD, bufD, 0,
VK_ACCESS_TRANSFER_WRITE_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT); VK_ACCESS_TRANSFER_WRITE_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT);
@ -2978,7 +2978,7 @@ void QRhiVulkan::enqueueResourceUpdates(QVkCommandBuffer *cbD, QRhiResourceUpdat
memset(&copyInfo, 0, sizeof(copyInfo)); memset(&copyInfo, 0, sizeof(copyInfo));
copyInfo.srcOffset = VkDeviceSize(u.offset); copyInfo.srcOffset = VkDeviceSize(u.offset);
copyInfo.dstOffset = VkDeviceSize(u.offset); copyInfo.dstOffset = VkDeviceSize(u.offset);
copyInfo.size = VkDeviceSize(u.data.size()); copyInfo.size = VkDeviceSize(u.dataSize);
QVkCommandBuffer::Command cmd; QVkCommandBuffer::Command cmd;
cmd.cmd = QVkCommandBuffer::Command::CopyBuffer; cmd.cmd = QVkCommandBuffer::Command::CopyBuffer;
@ -3428,11 +3428,11 @@ void QRhiVulkan::executeBufferHostWritesForSlot(QVkBuffer *bufD, int slot)
int changeEnd = -1; int changeEnd = -1;
for (const QRhiResourceUpdateBatchPrivate::BufferOp &u : qAsConst(bufD->pendingDynamicUpdates[slot])) { for (const QRhiResourceUpdateBatchPrivate::BufferOp &u : qAsConst(bufD->pendingDynamicUpdates[slot])) {
Q_ASSERT(bufD == QRHI_RES(QVkBuffer, u.buf)); Q_ASSERT(bufD == QRHI_RES(QVkBuffer, u.buf));
memcpy(static_cast<char *>(p) + u.offset, u.data.constData(), size_t(u.data.size())); memcpy(static_cast<char *>(p) + u.offset, u.data.constData(), size_t(u.dataSize));
if (changeBegin == -1 || u.offset < changeBegin) if (changeBegin == -1 || u.offset < changeBegin)
changeBegin = u.offset; changeBegin = u.offset;
if (changeEnd == -1 || u.offset + u.data.size() > changeEnd) if (changeEnd == -1 || u.offset + u.dataSize > changeEnd)
changeEnd = u.offset + u.data.size(); changeEnd = u.offset + u.dataSize;
} }
vmaUnmapMemory(toVmaAllocator(allocator), a); vmaUnmapMemory(toVmaAllocator(allocator), a);
if (changeBegin >= 0) if (changeBegin >= 0)