Use uint64_t for sizes to support large sizes
Signed-off-by: Jaime Arteaga <jaime.a.arteaga.molina@intel.com>
This commit is contained in:
parent
392b1c8b4d
commit
b48192bcfe
|
@ -152,8 +152,8 @@ struct CommandListCoreFamily : CommandListImp {
|
|||
MOCKABLE_VIRTUAL ze_result_t appendMemoryCopyKernelWithGA(void *dstPtr, NEO::GraphicsAllocation *dstPtrAlloc,
|
||||
uint64_t dstOffset, void *srcPtr,
|
||||
NEO::GraphicsAllocation *srcPtrAlloc,
|
||||
uint64_t srcOffset, uint32_t size,
|
||||
uint32_t elementSize, Builtin builtin,
|
||||
uint64_t srcOffset, uint64_t size,
|
||||
uint64_t elementSize, Builtin builtin,
|
||||
ze_event_handle_t hSignalEvent);
|
||||
|
||||
MOCKABLE_VIRTUAL ze_result_t appendMemoryCopyBlit(uintptr_t dstPtr,
|
||||
|
@ -161,7 +161,7 @@ struct CommandListCoreFamily : CommandListImp {
|
|||
uint64_t dstOffset, uintptr_t srcPtr,
|
||||
NEO::GraphicsAllocation *srcPtrAlloc,
|
||||
uint64_t srcOffset,
|
||||
uint32_t size);
|
||||
uint64_t size);
|
||||
|
||||
MOCKABLE_VIRTUAL ze_result_t appendMemoryCopyBlitRegion(NEO::GraphicsAllocation *srcAlloc,
|
||||
NEO::GraphicsAllocation *dstAlloc,
|
||||
|
|
|
@ -704,8 +704,8 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendMemoryCopyKernelWithGA(v
|
|||
void *srcPtr,
|
||||
NEO::GraphicsAllocation *srcPtrAlloc,
|
||||
uint64_t srcOffset,
|
||||
uint32_t size,
|
||||
uint32_t elementSize,
|
||||
uint64_t size,
|
||||
uint64_t elementSize,
|
||||
Builtin builtin,
|
||||
ze_event_handle_t hSignalEvent) {
|
||||
|
||||
|
@ -727,12 +727,12 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendMemoryCopyKernelWithGA(v
|
|||
builtinFunction->setArgBufferWithAlloc(0u, *reinterpret_cast<uintptr_t *>(dstPtr), dstPtrAlloc);
|
||||
builtinFunction->setArgBufferWithAlloc(1u, *reinterpret_cast<uintptr_t *>(srcPtr), srcPtrAlloc);
|
||||
|
||||
uint32_t elems = size / elementSize;
|
||||
uint64_t elems = size / elementSize;
|
||||
builtinFunction->setArgumentValue(2, sizeof(elems), &elems);
|
||||
builtinFunction->setArgumentValue(3, sizeof(dstOffset), &dstOffset);
|
||||
builtinFunction->setArgumentValue(4, sizeof(srcOffset), &srcOffset);
|
||||
|
||||
uint32_t groups = (size + ((groupSizeX * elementSize) - 1)) / (groupSizeX * elementSize);
|
||||
uint32_t groups = static_cast<uint32_t>((size + ((static_cast<uint64_t>(groupSizeX) * elementSize) - 1)) / (static_cast<uint64_t>(groupSizeX) * elementSize));
|
||||
ze_group_count_t dispatchFuncArgs{groups, 1u, 1u};
|
||||
|
||||
return CommandListCoreFamily<gfxCoreFamily>::appendLaunchKernelSplit(builtinFunction->toHandle(), &dispatchFuncArgs, hSignalEvent);
|
||||
|
@ -744,7 +744,7 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendMemoryCopyBlit(uintptr_t
|
|||
uint64_t dstOffset, uintptr_t srcPtr,
|
||||
NEO::GraphicsAllocation *srcPtrAlloc,
|
||||
uint64_t srcOffset,
|
||||
uint32_t size) {
|
||||
uint64_t size) {
|
||||
dstOffset += ptrDiff<uintptr_t>(dstPtr, dstPtrAlloc->getGpuAddress());
|
||||
srcOffset += ptrDiff<uintptr_t>(srcPtr, srcPtrAlloc->getGpuAddress());
|
||||
|
||||
|
@ -925,12 +925,12 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendMemoryCopy(void *dstptr,
|
|||
ret = isCopyOnly() ? appendMemoryCopyBlit(dstAllocationStruct.alignedAllocationPtr,
|
||||
dstAllocationStruct.alloc, dstAllocationStruct.offset,
|
||||
srcAllocationStruct.alignedAllocationPtr,
|
||||
srcAllocationStruct.alloc, srcAllocationStruct.offset, static_cast<uint32_t>(leftSize))
|
||||
srcAllocationStruct.alloc, srcAllocationStruct.offset, leftSize)
|
||||
: appendMemoryCopyKernelWithGA(reinterpret_cast<void *>(&dstAllocationStruct.alignedAllocationPtr),
|
||||
dstAllocationStruct.alloc, dstAllocationStruct.offset,
|
||||
reinterpret_cast<void *>(&srcAllocationStruct.alignedAllocationPtr),
|
||||
srcAllocationStruct.alloc, srcAllocationStruct.offset,
|
||||
static_cast<uint32_t>(leftSize), 1,
|
||||
leftSize, 1UL,
|
||||
Builtin::CopyBufferToBufferSide,
|
||||
hSignalEvent);
|
||||
}
|
||||
|
@ -939,13 +939,13 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendMemoryCopy(void *dstptr,
|
|||
ret = isCopyOnly() ? appendMemoryCopyBlit(dstAllocationStruct.alignedAllocationPtr,
|
||||
dstAllocationStruct.alloc, leftSize + dstAllocationStruct.offset,
|
||||
srcAllocationStruct.alignedAllocationPtr,
|
||||
srcAllocationStruct.alloc, leftSize + srcAllocationStruct.offset, static_cast<uint32_t>(middleSizeBytes))
|
||||
srcAllocationStruct.alloc, leftSize + srcAllocationStruct.offset, middleSizeBytes)
|
||||
: appendMemoryCopyKernelWithGA(reinterpret_cast<void *>(&dstAllocationStruct.alignedAllocationPtr),
|
||||
dstAllocationStruct.alloc, leftSize + dstAllocationStruct.offset,
|
||||
reinterpret_cast<void *>(&srcAllocationStruct.alignedAllocationPtr),
|
||||
srcAllocationStruct.alloc, leftSize + srcAllocationStruct.offset,
|
||||
static_cast<uint32_t>(middleSizeBytes),
|
||||
static_cast<uint32_t>(middleElSize),
|
||||
middleSizeBytes,
|
||||
middleElSize,
|
||||
Builtin::CopyBufferToBufferMiddle,
|
||||
hSignalEvent);
|
||||
}
|
||||
|
@ -954,12 +954,12 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendMemoryCopy(void *dstptr,
|
|||
ret = isCopyOnly() ? appendMemoryCopyBlit(dstAllocationStruct.alignedAllocationPtr,
|
||||
dstAllocationStruct.alloc, leftSize + middleSizeBytes + dstAllocationStruct.offset,
|
||||
srcAllocationStruct.alignedAllocationPtr,
|
||||
srcAllocationStruct.alloc, leftSize + middleSizeBytes + srcAllocationStruct.offset, static_cast<uint32_t>(rightSize))
|
||||
srcAllocationStruct.alloc, leftSize + middleSizeBytes + srcAllocationStruct.offset, rightSize)
|
||||
: appendMemoryCopyKernelWithGA(reinterpret_cast<void *>(&dstAllocationStruct.alignedAllocationPtr),
|
||||
dstAllocationStruct.alloc, leftSize + middleSizeBytes + dstAllocationStruct.offset,
|
||||
reinterpret_cast<void *>(&srcAllocationStruct.alignedAllocationPtr),
|
||||
srcAllocationStruct.alloc, leftSize + middleSizeBytes + srcAllocationStruct.offset,
|
||||
static_cast<uint32_t>(rightSize), 1u,
|
||||
rightSize, 1UL,
|
||||
Builtin::CopyBufferToBufferSide,
|
||||
hSignalEvent);
|
||||
}
|
||||
|
|
|
@ -40,8 +40,8 @@ class MockCommandListHw : public WhiteBox<::L0::CommandListCoreFamily<gfxCoreFam
|
|||
void *srcPtr,
|
||||
NEO::GraphicsAllocation *srcPtrAlloc,
|
||||
uint64_t srcOffset,
|
||||
uint32_t size,
|
||||
uint32_t elementSize,
|
||||
uint64_t size,
|
||||
uint64_t elementSize,
|
||||
Builtin builtin,
|
||||
ze_event_handle_t hSignalEvent) override {
|
||||
appendMemoryCopyKernelWithGACalledTimes++;
|
||||
|
@ -52,7 +52,7 @@ class MockCommandListHw : public WhiteBox<::L0::CommandListCoreFamily<gfxCoreFam
|
|||
uint64_t dstOffset, uintptr_t srcPtr,
|
||||
NEO::GraphicsAllocation *srcPtrAlloc,
|
||||
uint64_t srcOffset,
|
||||
uint32_t size) override {
|
||||
uint64_t size) override {
|
||||
appendMemoryCopyBlitCalledTimes++;
|
||||
return ZE_RESULT_SUCCESS;
|
||||
}
|
||||
|
@ -1062,7 +1062,7 @@ class MockCommandListForMemFill : public WhiteBox<::L0::CommandListCoreFamily<gf
|
|||
ze_result_t appendMemoryCopyBlit(NEO::GraphicsAllocation *dstPtrAlloc,
|
||||
uint64_t dstOffset,
|
||||
NEO::GraphicsAllocation *srcPtrAlloc,
|
||||
uint64_t srcOffset, uint32_t size) override {
|
||||
uint64_t srcOffset, uint64_t size) override {
|
||||
appendMemoryCopyBlitCalledTimes++;
|
||||
return ZE_RESULT_SUCCESS;
|
||||
}
|
||||
|
|
|
@ -33,7 +33,7 @@ class MockCommandListForMemFill : public WhiteBox<::L0::CommandListCoreFamily<gf
|
|||
uint64_t dstOffset, uintptr_t srcPtr,
|
||||
NEO::GraphicsAllocation *srcPtrAlloc,
|
||||
uint64_t srcOffset,
|
||||
uint32_t size) override {
|
||||
uint64_t size) override {
|
||||
appendMemoryCopyBlitCalledTimes++;
|
||||
return ZE_RESULT_SUCCESS;
|
||||
}
|
||||
|
@ -161,7 +161,7 @@ HWTEST2_F(AppendMemoryCopy, givenCopyOnlyCommandListThenDcFlushIsNotAddedAfterBl
|
|||
uintptr_t dstPtr = 0x7001;
|
||||
uint64_t srcOffset = 0x101;
|
||||
uint64_t dstOffset = 0x201;
|
||||
uint32_t copySize = 0x301;
|
||||
uint64_t copySize = 0x301;
|
||||
NEO::MockGraphicsAllocation mockAllocationSrc(0, NEO::GraphicsAllocation::AllocationType::INTERNAL_HOST_MEMORY,
|
||||
reinterpret_cast<void *>(srcPtr), 0x1000, 0, sizeof(uint32_t),
|
||||
MemoryPool::System4KBPages);
|
||||
|
|
Loading…
Reference in New Issue