Expand RS vector3 types to vector4.
BUG=5609007
The underlying LLVM implementation for vector3 types does this implicitly. If
RS does not adjust its implementation, we will always be misaligned for any
subsequent data after a vector3 type. We previously inserted padding into the
reflected layers from llvm-rs-cc (hence the skip padding part of this change).
We can safely ignore the padding now that the Java/native code is updated to
use the expanded size. The compiler will also need modification to ensure that
we don't mistakenly skip over any end-of-struct padding.
Fixing the 3 component vector padding problem.
Change-Id: If68af42287deb8f4b28addcd19a9fa314656be44
diff --git a/rsAllocation.cpp b/rsAllocation.cpp
index 35d812d..c1192fe 100644
--- a/rsAllocation.cpp
+++ b/rsAllocation.cpp
@@ -195,6 +195,81 @@
prefix, getPtr(), mHal.state.usageFlags, mHal.state.mipmapControl);
}
+uint32_t Allocation::getPackedSize() const {
+ uint32_t numItems = mHal.state.type->getSizeBytes() / mHal.state.type->getElementSizeBytes();
+ return numItems * mHal.state.type->getElement()->getSizeBytesUnpadded();
+}
+
+void Allocation::writePackedData(const Type *type,
+ uint8_t *dst, const uint8_t *src, bool dstPadded) {
+ const Element *elem = type->getElement();
+ uint32_t unpaddedBytes = elem->getSizeBytesUnpadded();
+ uint32_t paddedBytes = elem->getSizeBytes();
+ uint32_t numItems = type->getSizeBytes() / paddedBytes;
+
+ uint32_t srcInc = !dstPadded ? paddedBytes : unpaddedBytes;
+ uint32_t dstInc = dstPadded ? paddedBytes : unpaddedBytes;
+
+ // no sub-elements
+ uint32_t fieldCount = elem->getFieldCount();
+ if (fieldCount == 0) {
+ for (uint32_t i = 0; i < numItems; i ++) {
+ memcpy(dst, src, unpaddedBytes);
+ src += srcInc;
+ dst += dstInc;
+ }
+ return;
+ }
+
+ // Cache offsets
+ uint32_t *offsetsPadded = new uint32_t[fieldCount];
+ uint32_t *offsetsUnpadded = new uint32_t[fieldCount];
+ uint32_t *sizeUnpadded = new uint32_t[fieldCount];
+
+ for (uint32_t i = 0; i < fieldCount; i++) {
+ offsetsPadded[i] = elem->getFieldOffsetBytes(i);
+ offsetsUnpadded[i] = elem->getFieldOffsetBytesUnpadded(i);
+ sizeUnpadded[i] = elem->getField(i)->getSizeBytesUnpadded();
+ }
+
+ uint32_t *srcOffsets = !dstPadded ? offsetsPadded : offsetsUnpadded;
+ uint32_t *dstOffsets = dstPadded ? offsetsPadded : offsetsUnpadded;
+
+ // complex elements, need to copy subelem after subelem
+ for (uint32_t i = 0; i < numItems; i ++) {
+ for (uint32_t fI = 0; fI < fieldCount; fI++) {
+ memcpy(dst + dstOffsets[fI], src + srcOffsets[fI], sizeUnpadded[fI]);
+ }
+ src += srcInc;
+ dst += dstInc;
+ }
+
+ delete[] offsetsPadded;
+ delete[] offsetsUnpadded;
+ delete[] sizeUnpadded;
+}
+
+void Allocation::unpackVec3Allocation(const void *data, uint32_t dataSize) {
+ const uint8_t *src = (const uint8_t*)data;
+ uint8_t *dst = (uint8_t*)getPtr();
+
+ writePackedData(getType(), dst, src, true);
+}
+
+void Allocation::packVec3Allocation(OStream *stream) const {
+ uint32_t paddedBytes = getType()->getElement()->getSizeBytes();
+ uint32_t unpaddedBytes = getType()->getElement()->getSizeBytesUnpadded();
+ uint32_t numItems = mHal.state.type->getSizeBytes() / paddedBytes;
+
+ const uint8_t *src = (const uint8_t*)getPtr();
+ uint8_t *dst = new uint8_t[numItems * unpaddedBytes];
+
+ writePackedData(getType(), dst, src, false);
+ stream->addByteArray(dst, getPackedSize());
+
+ delete[] dst;
+}
+
void Allocation::serialize(OStream *stream) const {
// Need to identify ourselves
stream->addU32((uint32_t)getClassId());
@@ -207,10 +282,17 @@
mHal.state.type->serialize(stream);
uint32_t dataSize = mHal.state.type->getSizeBytes();
+ // 3 element vectors are padded to 4 in memory, but padding isn't serialized
+ uint32_t packedSize = getPackedSize();
// Write how much data we are storing
- stream->addU32(dataSize);
- // Now write the data
- stream->addByteArray(getPtr(), dataSize);
+ stream->addU32(packedSize);
+ if (dataSize == packedSize) {
+ // Now write the data
+ stream->addByteArray(getPtr(), dataSize);
+ } else {
+ // Now write the data
+ packVec3Allocation(stream);
+ }
}
Allocation *Allocation::createFromStream(Context *rsc, IStream *stream) {
@@ -230,22 +312,30 @@
}
type->compute();
+ Allocation *alloc = Allocation::createAllocation(rsc, type, RS_ALLOCATION_USAGE_SCRIPT);
+ type->decUserRef();
+
// Number of bytes we wrote out for this allocation
uint32_t dataSize = stream->loadU32();
- if (dataSize != type->getSizeBytes()) {
+ // 3 element vectors are padded to 4 in memory, but padding isn't serialized
+ uint32_t packedSize = alloc->getPackedSize();
+ if (dataSize != type->getSizeBytes() &&
+ dataSize != packedSize) {
LOGE("failed to read allocation because numbytes written is not the same loaded type wants\n");
+ ObjectBase::checkDelete(alloc);
ObjectBase::checkDelete(type);
return NULL;
}
- Allocation *alloc = Allocation::createAllocation(rsc, type, RS_ALLOCATION_USAGE_SCRIPT);
alloc->setName(name.string(), name.size());
- type->decUserRef();
- uint32_t count = dataSize / type->getElementSizeBytes();
-
- // Read in all of our allocation data
- alloc->data(rsc, 0, 0, count, stream->getPtr() + stream->getPos(), dataSize);
+ if (dataSize == type->getSizeBytes()) {
+ uint32_t count = dataSize / type->getElementSizeBytes();
+ // Read in all of our allocation data
+ alloc->data(rsc, 0, 0, count, stream->getPtr() + stream->getPos(), dataSize);
+ } else {
+ alloc->unpackVec3Allocation(stream->getPtr() + stream->getPos(), dataSize);
+ }
stream->reset(stream->getPos() + dataSize);
return alloc;
