Skip to content

[SYCL] Use correct units when specifying image offset and extent #19936

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 7 commits into from
Sep 16, 2025
Merged

[SYCL] Use correct units when specifying image offset and extent #19936

Show file tree
Hide file tree
Changes from 4 commits
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
880fd69
WIP
AlexeySachkov Sep 1, 2025
7837acf
Fix CUDA & HIP (untested locally)
AlexeySachkov Aug 27, 2025
dbcdff5
Fix CUDA adapter build
AlexeySachkov Aug 28, 2025
7b99dc8
Limit the change to bindless images only
AlexeySachkov Sep 3, 2025
0a255cf
Merge remote-tracking branch 'origin/sycl' into private/asachkov/use-…
AlexeySachkov Sep 8, 2025
6cdf678
Update sycl/source/handler.cpp
AlexeySachkov Sep 15, 2025
0d4f37f
Merge remote-tracking branch 'origin/sycl' into private/asachkov/use-…
AlexeySachkov Sep 15, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
17 changes: 13 additions & 4 deletions sycl/source/handler.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -262,18 +262,27 @@ fill_copy_args(detail::handler_impl *impl,
auto ZCopyExtentComponent = detail::fill_image_type(SrcImgDesc, UrSrcDesc);
detail::fill_image_type(DestImgDesc, UrDestDesc);

impl->MSrcOffset = {SrcOffset[0], SrcOffset[1], SrcOffset[2]};
impl->MDestOffset = {DestOffset[0], DestOffset[1], DestOffset[2]};
// Copy args computed here are directly passed to UR. Various offsets and
// extents end up passed as ur_rect_offset_t and ur_rect_region_t. Both those
// structs expect theirfirst component to be in bytes, not in pixels
size_t SrcPixelSize = SrcImgDesc.num_channels * get_channel_size(SrcImgDesc);
size_t DestPixelSize =
DestImgDesc.num_channels * get_channel_size(DestImgDesc);

impl->MSrcOffset = {SrcOffset[0] * SrcPixelSize, SrcOffset[1], SrcOffset[2]};
impl->MDestOffset = {DestOffset[0] * DestPixelSize, DestOffset[1],
DestOffset[2]};
impl->MSrcImageDesc = UrSrcDesc;
impl->MDstImageDesc = UrDestDesc;
impl->MSrcImageFormat = UrSrcFormat;
impl->MDstImageFormat = UrDestFormat;
impl->MImageCopyFlags = ImageCopyFlags;

if (CopyExtent.size() != 0) {
impl->MCopyExtent = {CopyExtent[0], CopyExtent[1], CopyExtent[2]};
impl->MCopyExtent = {CopyExtent[0] * SrcPixelSize, CopyExtent[1],
CopyExtent[2]};
} else {
impl->MCopyExtent = {SrcImgDesc.width, SrcImgDesc.height,
impl->MCopyExtent = {SrcImgDesc.width * SrcPixelSize, SrcImgDesc.height,
ZCopyExtentComponent};
}

Expand Down
98 changes: 47 additions & 51 deletions unified-runtime/source/adapters/cuda/image.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -645,7 +645,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urBindlessImagesImageCopyExp(
};

unsigned int NumChannels = 0;
size_t PixelSizeBytes = 0;
[[maybe_unused]] size_t PixelSizeBytes = 0;

UR_CALL(urCalculateNumChannels(pSrcImageFormat->channelOrder, &NumChannels));

Expand Down Expand Up @@ -673,19 +673,18 @@ UR_APIEXPORT ur_result_t UR_APICALL urBindlessImagesImageCopyExp(
cuPointerGetAttribute(&memType, CU_POINTER_ATTRIBUTE_MEMORY_TYPE,
(CUdeviceptr)pDst) != CUDA_SUCCESS;

size_t CopyExtentBytes = PixelSizeBytes * pCopyRegion->copyExtent.width;
const char *SrcWithOffset = static_cast<const char *>(pSrc) +
(pCopyRegion->srcOffset.x * PixelSizeBytes);
size_t CopyExtentBytes = pCopyRegion->copyExtent.width;
const char *SrcWithOffset =
static_cast<const char *>(pSrc) + pCopyRegion->srcOffset.x;

if (isCudaArray) {
UR_CHECK_ERROR(cuMemcpyHtoAAsync(
(CUarray)pDst, pCopyRegion->dstOffset.x * PixelSizeBytes,
static_cast<const void *>(SrcWithOffset), CopyExtentBytes,
Stream));
UR_CHECK_ERROR(
cuMemcpyHtoAAsync((CUarray)pDst, pCopyRegion->dstOffset.x,
static_cast<const void *>(SrcWithOffset),
CopyExtentBytes, Stream));
} else if (memType == CU_MEMORYTYPE_DEVICE) {
void *DstWithOffset =
static_cast<void *>(static_cast<char *>(pDst) +
(PixelSizeBytes * pCopyRegion->dstOffset.x));
void *DstWithOffset = static_cast<void *>(static_cast<char *>(pDst) +
pCopyRegion->dstOffset.x);
UR_CHECK_ERROR(
cuMemcpyHtoDAsync((CUdeviceptr)DstWithOffset,
static_cast<const void *>(SrcWithOffset),
Expand All @@ -698,11 +697,11 @@ UR_APIEXPORT ur_result_t UR_APICALL urBindlessImagesImageCopyExp(
CUDA_MEMCPY2D cpy_desc = {};
cpy_desc.srcMemoryType = CUmemorytype_enum::CU_MEMORYTYPE_HOST;
cpy_desc.srcHost = pSrc;
cpy_desc.srcXInBytes = pCopyRegion->srcOffset.x * PixelSizeBytes;
cpy_desc.srcXInBytes = pCopyRegion->srcOffset.x;
cpy_desc.srcY = pCopyRegion->srcOffset.y;
cpy_desc.dstXInBytes = pCopyRegion->dstOffset.x * PixelSizeBytes;
cpy_desc.dstXInBytes = pCopyRegion->dstOffset.x;
cpy_desc.dstY = pCopyRegion->dstOffset.y;
cpy_desc.WidthInBytes = PixelSizeBytes * pCopyRegion->copyExtent.width;
cpy_desc.WidthInBytes = pCopyRegion->copyExtent.width;
cpy_desc.Height = pCopyRegion->copyExtent.height;
cpy_desc.srcPitch = pSrcImageDesc->rowPitch;
if (pDstImageDesc->rowPitch == 0) {
Expand All @@ -717,10 +716,10 @@ UR_APIEXPORT ur_result_t UR_APICALL urBindlessImagesImageCopyExp(
UR_CHECK_ERROR(cuMemcpy2DAsync(&cpy_desc, Stream));
} else if (pDstImageDesc->type == UR_MEM_TYPE_IMAGE3D) {
CUDA_MEMCPY3D cpy_desc = {};
cpy_desc.srcXInBytes = pCopyRegion->srcOffset.x * PixelSizeBytes;
cpy_desc.srcXInBytes = pCopyRegion->srcOffset.x;
cpy_desc.srcY = pCopyRegion->srcOffset.y;
cpy_desc.srcZ = pCopyRegion->srcOffset.z;
cpy_desc.dstXInBytes = pCopyRegion->dstOffset.x * PixelSizeBytes;
cpy_desc.dstXInBytes = pCopyRegion->dstOffset.x;
cpy_desc.dstY = pCopyRegion->dstOffset.y;
cpy_desc.dstZ = pCopyRegion->dstOffset.z;
cpy_desc.srcMemoryType = CUmemorytype_enum::CU_MEMORYTYPE_HOST;
Expand All @@ -729,18 +728,18 @@ UR_APIEXPORT ur_result_t UR_APICALL urBindlessImagesImageCopyExp(
cpy_desc.srcHeight = pSrcImageDesc->height;
cpy_desc.dstMemoryType = CUmemorytype_enum::CU_MEMORYTYPE_ARRAY;
cpy_desc.dstArray = (CUarray)pDst;
cpy_desc.WidthInBytes = PixelSizeBytes * pCopyRegion->copyExtent.width;
cpy_desc.WidthInBytes = pCopyRegion->copyExtent.width;
cpy_desc.Height = pCopyRegion->copyExtent.height;
cpy_desc.Depth = pCopyRegion->copyExtent.depth;
UR_CHECK_ERROR(cuMemcpy3DAsync(&cpy_desc, Stream));
} else if (pDstImageDesc->type == UR_MEM_TYPE_IMAGE1D_ARRAY ||
pDstImageDesc->type == UR_MEM_TYPE_IMAGE2D_ARRAY ||
pDstImageDesc->type == UR_MEM_TYPE_IMAGE_CUBEMAP_EXP) {
CUDA_MEMCPY3D cpy_desc = {};
cpy_desc.srcXInBytes = pCopyRegion->srcOffset.x * PixelSizeBytes;
cpy_desc.srcXInBytes = pCopyRegion->srcOffset.x;
cpy_desc.srcY = pCopyRegion->srcOffset.y;
cpy_desc.srcZ = pCopyRegion->srcOffset.z;
cpy_desc.dstXInBytes = pCopyRegion->dstOffset.x * PixelSizeBytes;
cpy_desc.dstXInBytes = pCopyRegion->dstOffset.x;
cpy_desc.dstY = pCopyRegion->dstOffset.y;
cpy_desc.dstZ = pCopyRegion->dstOffset.z;
cpy_desc.srcMemoryType = CUmemorytype_enum::CU_MEMORYTYPE_HOST;
Expand All @@ -749,7 +748,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urBindlessImagesImageCopyExp(
cpy_desc.srcHeight = std::max(uint64_t{1}, pSrcImageDesc->height);
cpy_desc.dstMemoryType = CUmemorytype_enum::CU_MEMORYTYPE_ARRAY;
cpy_desc.dstArray = (CUarray)pDst;
cpy_desc.WidthInBytes = PixelSizeBytes * pCopyRegion->copyExtent.width;
cpy_desc.WidthInBytes = pCopyRegion->copyExtent.width;
cpy_desc.Height = std::max(uint64_t{1}, pCopyRegion->copyExtent.height);
cpy_desc.Depth = pCopyRegion->copyExtent.depth;
UR_CHECK_ERROR(cuMemcpy3DAsync(&cpy_desc, Stream));
Expand All @@ -764,20 +763,17 @@ UR_APIEXPORT ur_result_t UR_APICALL urBindlessImagesImageCopyExp(
cuPointerGetAttribute(&memType, CU_POINTER_ATTRIBUTE_MEMORY_TYPE,
(CUdeviceptr)pSrc) != CUDA_SUCCESS;

size_t CopyExtentBytes = PixelSizeBytes * pCopyRegion->copyExtent.width;
void *DstWithOffset =
static_cast<void *>(static_cast<char *>(pDst) +
(PixelSizeBytes * pCopyRegion->dstOffset.x));
size_t CopyExtentBytes = pCopyRegion->copyExtent.width;
void *DstWithOffset = static_cast<void *>(static_cast<char *>(pDst) +
pCopyRegion->dstOffset.x);

if (isCudaArray) {
UR_CHECK_ERROR(
cuMemcpyAtoHAsync(DstWithOffset, as_CUArray(pSrc),
PixelSizeBytes * pCopyRegion->srcOffset.x,
CopyExtentBytes, Stream));
UR_CHECK_ERROR(cuMemcpyAtoHAsync(DstWithOffset, as_CUArray(pSrc),
pCopyRegion->srcOffset.x,
CopyExtentBytes, Stream));
} else if (memType == CU_MEMORYTYPE_DEVICE) {
const char *SrcWithOffset =
static_cast<const char *>(pSrc) +
(pCopyRegion->srcOffset.x * PixelSizeBytes);
static_cast<const char *>(pSrc) + pCopyRegion->srcOffset.x;
UR_CHECK_ERROR(cuMemcpyDtoHAsync(DstWithOffset,
(CUdeviceptr)SrcWithOffset,
CopyExtentBytes, Stream));
Expand All @@ -787,11 +783,11 @@ UR_APIEXPORT ur_result_t UR_APICALL urBindlessImagesImageCopyExp(
}
} else if (pSrcImageDesc->type == UR_MEM_TYPE_IMAGE2D) {
CUDA_MEMCPY2D cpy_desc = {};
cpy_desc.srcXInBytes = pCopyRegion->srcOffset.x * PixelSizeBytes;
cpy_desc.srcXInBytes = pCopyRegion->srcOffset.x;
cpy_desc.srcY = pCopyRegion->srcOffset.y;
cpy_desc.dstXInBytes = pCopyRegion->dstOffset.x * PixelSizeBytes;
cpy_desc.dstXInBytes = pCopyRegion->dstOffset.x;
cpy_desc.dstY = pCopyRegion->dstOffset.y;
cpy_desc.WidthInBytes = PixelSizeBytes * pCopyRegion->copyExtent.width;
cpy_desc.WidthInBytes = pCopyRegion->copyExtent.width;
cpy_desc.Height = pCopyRegion->copyExtent.height;
cpy_desc.dstPitch = pDstImageDesc->rowPitch;
cpy_desc.dstMemoryType = CUmemorytype_enum::CU_MEMORYTYPE_HOST;
Expand All @@ -808,10 +804,10 @@ UR_APIEXPORT ur_result_t UR_APICALL urBindlessImagesImageCopyExp(
UR_CHECK_ERROR(cuMemcpy2DAsync(&cpy_desc, Stream));
} else if (pSrcImageDesc->type == UR_MEM_TYPE_IMAGE3D) {
CUDA_MEMCPY3D cpy_desc = {};
cpy_desc.srcXInBytes = pCopyRegion->srcOffset.x * PixelSizeBytes;
cpy_desc.srcXInBytes = pCopyRegion->srcOffset.x;
cpy_desc.srcY = pCopyRegion->srcOffset.y;
cpy_desc.srcZ = pCopyRegion->srcOffset.z;
cpy_desc.dstXInBytes = pCopyRegion->dstOffset.x * PixelSizeBytes;
cpy_desc.dstXInBytes = pCopyRegion->dstOffset.x;
cpy_desc.dstY = pCopyRegion->dstOffset.y;
cpy_desc.dstZ = pCopyRegion->dstOffset.z;
cpy_desc.srcMemoryType = CUmemorytype_enum::CU_MEMORYTYPE_ARRAY;
Expand All @@ -820,18 +816,18 @@ UR_APIEXPORT ur_result_t UR_APICALL urBindlessImagesImageCopyExp(
cpy_desc.dstHost = pDst;
cpy_desc.dstPitch = pDstImageDesc->rowPitch;
cpy_desc.dstHeight = pDstImageDesc->height;
cpy_desc.WidthInBytes = PixelSizeBytes * pCopyRegion->copyExtent.width;
cpy_desc.WidthInBytes = pCopyRegion->copyExtent.width;
cpy_desc.Height = pCopyRegion->copyExtent.height;
cpy_desc.Depth = pCopyRegion->copyExtent.depth;
UR_CHECK_ERROR(cuMemcpy3DAsync(&cpy_desc, Stream));
} else if (pSrcImageDesc->type == UR_MEM_TYPE_IMAGE1D_ARRAY ||
pSrcImageDesc->type == UR_MEM_TYPE_IMAGE2D_ARRAY ||
pSrcImageDesc->type == UR_MEM_TYPE_IMAGE_CUBEMAP_EXP) {
CUDA_MEMCPY3D cpy_desc = {};
cpy_desc.srcXInBytes = pCopyRegion->srcOffset.x * PixelSizeBytes;
cpy_desc.srcXInBytes = pCopyRegion->srcOffset.x;
cpy_desc.srcY = pCopyRegion->srcOffset.y;
cpy_desc.srcZ = pCopyRegion->srcOffset.z;
cpy_desc.dstXInBytes = pCopyRegion->dstOffset.x * PixelSizeBytes;
cpy_desc.dstXInBytes = pCopyRegion->dstOffset.x;
cpy_desc.dstY = pCopyRegion->dstOffset.y;
cpy_desc.dstZ = pCopyRegion->dstOffset.z;
cpy_desc.srcMemoryType = CUmemorytype_enum::CU_MEMORYTYPE_ARRAY;
Expand All @@ -840,7 +836,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urBindlessImagesImageCopyExp(
cpy_desc.dstHost = pDst;
cpy_desc.dstPitch = pDstImageDesc->rowPitch;
cpy_desc.dstHeight = std::max(uint64_t{1}, pDstImageDesc->height);
cpy_desc.WidthInBytes = PixelSizeBytes * pCopyRegion->copyExtent.width;
cpy_desc.WidthInBytes = pCopyRegion->copyExtent.width;
cpy_desc.Height = std::max(uint64_t{1}, pCopyRegion->copyExtent.height);
cpy_desc.Depth = pCopyRegion->copyExtent.depth;
UR_CHECK_ERROR(cuMemcpy3DAsync(&cpy_desc, Stream));
Expand Down Expand Up @@ -874,11 +870,11 @@ UR_APIEXPORT ur_result_t UR_APICALL urBindlessImagesImageCopyExp(
(CUdeviceptr)pDst) != CUDA_SUCCESS;

CUDA_MEMCPY2D cpy_desc = {};
cpy_desc.srcXInBytes = pCopyRegion->srcOffset.x * PixelSizeBytes;
cpy_desc.srcXInBytes = pCopyRegion->srcOffset.x;
cpy_desc.srcY = 0;
cpy_desc.dstXInBytes = pCopyRegion->dstOffset.x * PixelSizeBytes;
cpy_desc.dstXInBytes = pCopyRegion->dstOffset.x;
cpy_desc.dstY = 0;
cpy_desc.WidthInBytes = PixelSizeBytes * pCopyRegion->copyExtent.width;
cpy_desc.WidthInBytes = pCopyRegion->copyExtent.width;
cpy_desc.Height = 1;
if (isSrcCudaArray) {
cpy_desc.srcMemoryType = CUmemorytype_enum::CU_MEMORYTYPE_ARRAY;
Expand All @@ -897,11 +893,11 @@ UR_APIEXPORT ur_result_t UR_APICALL urBindlessImagesImageCopyExp(
UR_CHECK_ERROR(cuMemcpy2DAsync(&cpy_desc, Stream));
} else if (pSrcImageDesc->type == UR_MEM_TYPE_IMAGE2D) {
CUDA_MEMCPY2D cpy_desc = {};
cpy_desc.srcXInBytes = pCopyRegion->srcOffset.x * PixelSizeBytes;
cpy_desc.srcXInBytes = pCopyRegion->srcOffset.x;
cpy_desc.srcY = pCopyRegion->srcOffset.y;
cpy_desc.dstXInBytes = pCopyRegion->dstOffset.x * PixelSizeBytes;
cpy_desc.dstXInBytes = pCopyRegion->dstOffset.x;
cpy_desc.dstY = pCopyRegion->dstOffset.y;
cpy_desc.WidthInBytes = PixelSizeBytes * pCopyRegion->copyExtent.width;
cpy_desc.WidthInBytes = pCopyRegion->copyExtent.width;
cpy_desc.Height = pCopyRegion->copyExtent.height;
if (pSrcImageDesc->rowPitch == 0) {
cpy_desc.srcMemoryType = CUmemorytype_enum::CU_MEMORYTYPE_ARRAY;
Expand All @@ -924,35 +920,35 @@ UR_APIEXPORT ur_result_t UR_APICALL urBindlessImagesImageCopyExp(
UR_CHECK_ERROR(cuMemcpy2DAsync(&cpy_desc, Stream));
} else if (pSrcImageDesc->type == UR_MEM_TYPE_IMAGE3D) {
CUDA_MEMCPY3D cpy_desc = {};
cpy_desc.srcXInBytes = pCopyRegion->srcOffset.x * PixelSizeBytes;
cpy_desc.srcXInBytes = pCopyRegion->srcOffset.x;
cpy_desc.srcY = pCopyRegion->srcOffset.y;
cpy_desc.srcZ = pCopyRegion->srcOffset.z;
cpy_desc.dstXInBytes = pCopyRegion->dstOffset.x * PixelSizeBytes;
cpy_desc.dstXInBytes = pCopyRegion->dstOffset.x;
cpy_desc.dstY = pCopyRegion->dstOffset.y;
cpy_desc.dstZ = pCopyRegion->dstOffset.z;
cpy_desc.srcMemoryType = CUmemorytype_enum::CU_MEMORYTYPE_ARRAY;
cpy_desc.srcArray = as_CUArray(pSrc);
cpy_desc.dstMemoryType = CUmemorytype_enum::CU_MEMORYTYPE_ARRAY;
cpy_desc.dstArray = (CUarray)pDst;
cpy_desc.WidthInBytes = PixelSizeBytes * pCopyRegion->copyExtent.width;
cpy_desc.WidthInBytes = pCopyRegion->copyExtent.width;
cpy_desc.Height = pCopyRegion->copyExtent.height;
cpy_desc.Depth = pCopyRegion->copyExtent.depth;
UR_CHECK_ERROR(cuMemcpy3DAsync(&cpy_desc, Stream));
} else if (pSrcImageDesc->type == UR_MEM_TYPE_IMAGE1D_ARRAY ||
pSrcImageDesc->type == UR_MEM_TYPE_IMAGE2D_ARRAY ||
pSrcImageDesc->type == UR_MEM_TYPE_IMAGE_CUBEMAP_EXP) {
CUDA_MEMCPY3D cpy_desc = {};
cpy_desc.srcXInBytes = pCopyRegion->srcOffset.x * PixelSizeBytes;
cpy_desc.srcXInBytes = pCopyRegion->srcOffset.x;
cpy_desc.srcY = pCopyRegion->srcOffset.y;
cpy_desc.srcZ = pCopyRegion->srcOffset.z;
cpy_desc.dstXInBytes = pCopyRegion->dstOffset.x * PixelSizeBytes;
cpy_desc.dstXInBytes = pCopyRegion->dstOffset.x;
cpy_desc.dstY = pCopyRegion->dstOffset.y;
cpy_desc.dstZ = pCopyRegion->dstOffset.z;
cpy_desc.srcMemoryType = CUmemorytype_enum::CU_MEMORYTYPE_ARRAY;
cpy_desc.srcArray = as_CUArray(pSrc);
cpy_desc.dstMemoryType = CUmemorytype_enum::CU_MEMORYTYPE_ARRAY;
cpy_desc.dstArray = (CUarray)pDst;
cpy_desc.WidthInBytes = PixelSizeBytes * pCopyRegion->copyExtent.width;
cpy_desc.WidthInBytes = pCopyRegion->copyExtent.width;
cpy_desc.Height = std::max(uint64_t{1}, pCopyRegion->copyExtent.height);
cpy_desc.Depth = pCopyRegion->copyExtent.depth;
UR_CHECK_ERROR(cuMemcpy3DAsync(&cpy_desc, Stream));
Expand Down
Loading
Loading