Skip to content

Commit

Permalink
[CUDA][Bindless] Address USM normalized type image creation failure a…
Browse files Browse the repository at this point in the history 
…nd functionality

The specialised formats, CU_AD_FORMAT_{U|S}NORM_INT{8|16}X{1|2|4},
used to specify a format that is normalized, cannot be applied to
images backed by USM memory, causing a failure. This commit
addresses that issue by using the standard integer formats and
making sure the CU_TRSF_READ_AS_INTEGER flag is set appropriately.

Another issue fixed here is that users with CUDA toolkits pre 11.5
are unable to enable normalized types due to the specialised
formats being introduced in CUDA 11.5.
  • Loading branch information
@Seanst98
Seanst98 committed Sep 17, 2024
1 parent 4d19115 commit 1e40614
Show file tree
Hide file tree
Showing 2 changed files with 48 additions and 124 deletions.
163 changes: 43 additions & 120 deletions source/adapters/cuda/image.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -65,101 +65,59 @@ ur_result_t urCalculateNumChannels(ur_image_channel_order_t order,
/// format if not nullptr.
/// /param return_pixel_size_bytes will be set to the pixel
/// byte size if not nullptr.
/// /param return_normalized_dtype_flag will be set if the
/// data type is normalized if not nullptr.
ur_result_t
urToCudaImageChannelFormat(ur_image_channel_type_t image_channel_type,
ur_image_channel_order_t image_channel_order,
CUarray_format *return_cuda_format,
size_t *return_pixel_size_bytes) {
size_t *return_pixel_size_bytes,
unsigned int *return_normalized_dtype_flag) {

CUarray_format cuda_format;
CUarray_format cuda_format = CU_AD_FORMAT_UNSIGNED_INT8;
size_t pixel_size_bytes = 0;
unsigned int num_channels = 0;
unsigned int normalized_dtype_flag = 0;
UR_CHECK_ERROR(urCalculateNumChannels(image_channel_order, &num_channels));

switch (image_channel_type) {
#define CASE(FROM, TO, SIZE) \
#define CASE(FROM, TO, SIZE, NORM) \
case FROM: { \
cuda_format = TO; \
pixel_size_bytes = SIZE * num_channels; \
normalized_dtype_flag = NORM; \
break; \
}

CASE(UR_IMAGE_CHANNEL_TYPE_UNSIGNED_INT8, CU_AD_FORMAT_UNSIGNED_INT8, 1)
CASE(UR_IMAGE_CHANNEL_TYPE_SIGNED_INT8, CU_AD_FORMAT_SIGNED_INT8, 1)
CASE(UR_IMAGE_CHANNEL_TYPE_UNSIGNED_INT16, CU_AD_FORMAT_UNSIGNED_INT16, 2)
CASE(UR_IMAGE_CHANNEL_TYPE_SIGNED_INT16, CU_AD_FORMAT_SIGNED_INT16, 2)
CASE(UR_IMAGE_CHANNEL_TYPE_HALF_FLOAT, CU_AD_FORMAT_HALF, 2)
CASE(UR_IMAGE_CHANNEL_TYPE_UNSIGNED_INT32, CU_AD_FORMAT_UNSIGNED_INT32, 4)
CASE(UR_IMAGE_CHANNEL_TYPE_SIGNED_INT32, CU_AD_FORMAT_SIGNED_INT32, 4)
CASE(UR_IMAGE_CHANNEL_TYPE_FLOAT, CU_AD_FORMAT_FLOAT, 4)
CASE(UR_IMAGE_CHANNEL_TYPE_UNSIGNED_INT8, CU_AD_FORMAT_UNSIGNED_INT8, 1, 0)
CASE(UR_IMAGE_CHANNEL_TYPE_SIGNED_INT8, CU_AD_FORMAT_SIGNED_INT8, 1, 0)
CASE(UR_IMAGE_CHANNEL_TYPE_UNSIGNED_INT16, CU_AD_FORMAT_UNSIGNED_INT16, 2,
0)
CASE(UR_IMAGE_CHANNEL_TYPE_SIGNED_INT16, CU_AD_FORMAT_SIGNED_INT16, 2, 0)
CASE(UR_IMAGE_CHANNEL_TYPE_HALF_FLOAT, CU_AD_FORMAT_HALF, 2, 0)
CASE(UR_IMAGE_CHANNEL_TYPE_UNSIGNED_INT32, CU_AD_FORMAT_UNSIGNED_INT32, 4,
0)
CASE(UR_IMAGE_CHANNEL_TYPE_SIGNED_INT32, CU_AD_FORMAT_SIGNED_INT32, 4, 0)
CASE(UR_IMAGE_CHANNEL_TYPE_FLOAT, CU_AD_FORMAT_FLOAT, 4, 0)
CASE(UR_IMAGE_CHANNEL_TYPE_UNORM_INT8, CU_AD_FORMAT_UNSIGNED_INT8, 1, 1)
CASE(UR_IMAGE_CHANNEL_TYPE_SNORM_INT8, CU_AD_FORMAT_SIGNED_INT8, 1, 1)
CASE(UR_IMAGE_CHANNEL_TYPE_UNORM_INT16, CU_AD_FORMAT_UNSIGNED_INT16, 2, 1)
CASE(UR_IMAGE_CHANNEL_TYPE_SNORM_INT16, CU_AD_FORMAT_SIGNED_INT16, 2, 1)

#undef CASE
default:
break;
}

// These new formats were brought in in CUDA 11.5
#if CUDA_VERSION >= 11050

// If none of the above channel types were passed, check those below
if (pixel_size_bytes == 0) {

// We can't use a switch statement here because these single
// UR_IMAGE_CHANNEL_TYPEs can correspond to multiple [u/s]norm CU_AD_FORMATs
// depending on the number of channels. We use a std::map instead to
// retrieve the correct CUDA format

// map < <channel type, num channels> , <CUDA format, data type byte size> >
const std::map<std::pair<ur_image_channel_type_t, uint32_t>,
std::pair<CUarray_format, uint32_t>>
norm_channel_type_map{
{{UR_IMAGE_CHANNEL_TYPE_UNORM_INT8, 1},
{CU_AD_FORMAT_UNORM_INT8X1, 1}},
{{UR_IMAGE_CHANNEL_TYPE_UNORM_INT8, 2},
{CU_AD_FORMAT_UNORM_INT8X2, 2}},
{{UR_IMAGE_CHANNEL_TYPE_UNORM_INT8, 4},
{CU_AD_FORMAT_UNORM_INT8X4, 4}},

{{UR_IMAGE_CHANNEL_TYPE_SNORM_INT8, 1},
{CU_AD_FORMAT_SNORM_INT8X1, 1}},
{{UR_IMAGE_CHANNEL_TYPE_SNORM_INT8, 2},
{CU_AD_FORMAT_SNORM_INT8X2, 2}},
{{UR_IMAGE_CHANNEL_TYPE_SNORM_INT8, 4},
{CU_AD_FORMAT_SNORM_INT8X4, 4}},

{{UR_IMAGE_CHANNEL_TYPE_UNORM_INT16, 1},
{CU_AD_FORMAT_UNORM_INT16X1, 2}},
{{UR_IMAGE_CHANNEL_TYPE_UNORM_INT16, 2},
{CU_AD_FORMAT_UNORM_INT16X2, 4}},
{{UR_IMAGE_CHANNEL_TYPE_UNORM_INT16, 4},
{CU_AD_FORMAT_UNORM_INT16X4, 8}},

{{UR_IMAGE_CHANNEL_TYPE_SNORM_INT16, 1},
{CU_AD_FORMAT_SNORM_INT16X1, 2}},
{{UR_IMAGE_CHANNEL_TYPE_SNORM_INT16, 2},
{CU_AD_FORMAT_SNORM_INT16X2, 4}},
{{UR_IMAGE_CHANNEL_TYPE_SNORM_INT16, 4},
{CU_AD_FORMAT_SNORM_INT16X4, 8}},
};

try {
auto cuda_format_and_size = norm_channel_type_map.at(
std::make_pair(image_channel_type, num_channels));
cuda_format = cuda_format_and_size.first;
pixel_size_bytes = cuda_format_and_size.second;
} catch (const std::out_of_range &) {
return UR_RESULT_ERROR_UNSUPPORTED_IMAGE_FORMAT;
}
}

#endif

if (return_cuda_format) {
*return_cuda_format = cuda_format;
}
if (return_pixel_size_bytes) {
*return_pixel_size_bytes = pixel_size_bytes;
}
if (return_normalized_dtype_flag) {
*return_normalized_dtype_flag = normalized_dtype_flag;
}
return UR_RESULT_SUCCESS;
}

Expand Down Expand Up @@ -189,45 +147,6 @@ cudaToUrImageChannelFormat(CUarray_format cuda_format,
UR_IMAGE_CHANNEL_TYPE_HALF_FLOAT);
CUDA_TO_UR_IMAGE_CHANNEL_TYPE(CU_AD_FORMAT_FLOAT,
UR_IMAGE_CHANNEL_TYPE_FLOAT);
#if CUDA_VERSION >= 11050

// Note that the CUDA UNORM and SNORM formats also encode the number of
// channels.
// Since UR does not encode this, we map different CUDA formats to the same
// UR channel type.
// Since this function is only called from `urBindlessImagesImageGetInfoExp`
// which has access to `CUDA_ARRAY3D_DESCRIPTOR`, we can determine the
// number of channels in the calling function.

CUDA_TO_UR_IMAGE_CHANNEL_TYPE(CU_AD_FORMAT_UNORM_INT8X1,
UR_IMAGE_CHANNEL_TYPE_UNORM_INT8);
CUDA_TO_UR_IMAGE_CHANNEL_TYPE(CU_AD_FORMAT_UNORM_INT8X2,
UR_IMAGE_CHANNEL_TYPE_UNORM_INT8);
CUDA_TO_UR_IMAGE_CHANNEL_TYPE(CU_AD_FORMAT_UNORM_INT8X4,
UR_IMAGE_CHANNEL_TYPE_UNORM_INT8);

CUDA_TO_UR_IMAGE_CHANNEL_TYPE(CU_AD_FORMAT_UNORM_INT16X1,
UR_IMAGE_CHANNEL_TYPE_UNORM_INT16);
CUDA_TO_UR_IMAGE_CHANNEL_TYPE(CU_AD_FORMAT_UNORM_INT16X2,
UR_IMAGE_CHANNEL_TYPE_UNORM_INT16);
CUDA_TO_UR_IMAGE_CHANNEL_TYPE(CU_AD_FORMAT_UNORM_INT16X4,
UR_IMAGE_CHANNEL_TYPE_UNORM_INT16);

CUDA_TO_UR_IMAGE_CHANNEL_TYPE(CU_AD_FORMAT_SNORM_INT8X1,
UR_IMAGE_CHANNEL_TYPE_SNORM_INT8);
CUDA_TO_UR_IMAGE_CHANNEL_TYPE(CU_AD_FORMAT_SNORM_INT8X2,
UR_IMAGE_CHANNEL_TYPE_SNORM_INT8);
CUDA_TO_UR_IMAGE_CHANNEL_TYPE(CU_AD_FORMAT_SNORM_INT8X4,
UR_IMAGE_CHANNEL_TYPE_SNORM_INT8);

CUDA_TO_UR_IMAGE_CHANNEL_TYPE(CU_AD_FORMAT_SNORM_INT16X1,
UR_IMAGE_CHANNEL_TYPE_SNORM_INT16);
CUDA_TO_UR_IMAGE_CHANNEL_TYPE(CU_AD_FORMAT_SNORM_INT16X2,
UR_IMAGE_CHANNEL_TYPE_SNORM_INT16);
CUDA_TO_UR_IMAGE_CHANNEL_TYPE(CU_AD_FORMAT_SNORM_INT16X4,
UR_IMAGE_CHANNEL_TYPE_SNORM_INT16);
#endif
#undef MAP
default:
return UR_RESULT_ERROR_UNSUPPORTED_IMAGE_FORMAT;
}
Expand All @@ -236,6 +155,7 @@ cudaToUrImageChannelFormat(CUarray_format cuda_format,
ur_result_t urTextureCreate(ur_sampler_handle_t hSampler,
const ur_image_desc_t *pImageDesc,
const CUDA_RESOURCE_DESC &ResourceDesc,
const unsigned int normalized_dtype_flag,
ur_exp_image_native_handle_t *phRetImage) {

try {
Expand Down Expand Up @@ -306,8 +226,9 @@ ur_result_t urTextureCreate(ur_sampler_handle_t hSampler,

// CUDA default promotes 8-bit and 16-bit integers to float between [0,1]
// This flag prevents this behaviour.
ImageTexDesc.flags |= CU_TRSF_READ_AS_INTEGER;

if (!normalized_dtype_flag) {
ImageTexDesc.flags |= CU_TRSF_READ_AS_INTEGER;
}
// Cubemap attributes
ur_exp_sampler_cubemap_filter_mode_t CubemapFilterModeProp =
hSampler->getCubemapFilterMode();
Expand Down Expand Up @@ -413,9 +334,9 @@ UR_APIEXPORT ur_result_t UR_APICALL urBindlessImagesImageAllocateExp(
UR_CHECK_ERROR(urCalculateNumChannels(pImageFormat->channelOrder,
&array_desc.NumChannels));

UR_CHECK_ERROR(urToCudaImageChannelFormat(pImageFormat->channelType,
pImageFormat->channelOrder,
&array_desc.Format, nullptr));
UR_CHECK_ERROR(urToCudaImageChannelFormat(
pImageFormat->channelType, pImageFormat->channelOrder, &array_desc.Format,
nullptr, nullptr));

array_desc.Flags = 0; // No flags required
array_desc.Width = pImageDesc->width;
Expand Down Expand Up @@ -534,7 +455,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urBindlessImagesUnsampledImageCreateExp(
size_t PixelSizeBytes;
UR_CHECK_ERROR(urToCudaImageChannelFormat(pImageFormat->channelType,
pImageFormat->channelOrder, &format,
&PixelSizeBytes));
&PixelSizeBytes, nullptr));

try {

Expand Down Expand Up @@ -579,9 +500,10 @@ UR_APIEXPORT ur_result_t UR_APICALL urBindlessImagesSampledImageCreateExp(

CUarray_format format;
size_t PixelSizeBytes;
UR_CHECK_ERROR(urToCudaImageChannelFormat(pImageFormat->channelType,
pImageFormat->channelOrder, &format,
&PixelSizeBytes));
unsigned int normalized_dtype_flag;
UR_CHECK_ERROR(urToCudaImageChannelFormat(
pImageFormat->channelType, pImageFormat->channelOrder, &format,
&PixelSizeBytes, &normalized_dtype_flag));

try {
CUDA_RESOURCE_DESC image_res_desc = {};
Expand Down Expand Up @@ -630,8 +552,8 @@ UR_APIEXPORT ur_result_t UR_APICALL urBindlessImagesSampledImageCreateExp(
return UR_RESULT_ERROR_INVALID_VALUE;
}

UR_CHECK_ERROR(
urTextureCreate(hSampler, pImageDesc, image_res_desc, phImage));
UR_CHECK_ERROR(urTextureCreate(hSampler, pImageDesc, image_res_desc,
normalized_dtype_flag, phImage));

} catch (ur_result_t Err) {
return Err;
Expand Down Expand Up @@ -667,7 +589,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urBindlessImagesImageCopyExp(
// later.
UR_CHECK_ERROR(urToCudaImageChannelFormat(pSrcImageFormat->channelType,
pSrcImageFormat->channelOrder,
nullptr, &PixelSizeBytes));
nullptr, &PixelSizeBytes, nullptr));

try {
ScopedContext Active(hQueue->getDevice());
Expand Down Expand Up @@ -1146,8 +1068,9 @@ UR_APIEXPORT ur_result_t UR_APICALL urBindlessImagesMapExternalArrayExp(
urCalculateNumChannels(pImageFormat->channelOrder, &NumChannels));

CUarray_format format;
UR_CHECK_ERROR(urToCudaImageChannelFormat(
pImageFormat->channelType, pImageFormat->channelOrder, &format, nullptr));
UR_CHECK_ERROR(urToCudaImageChannelFormat(pImageFormat->channelType,
pImageFormat->channelOrder, &format,
nullptr, nullptr));

try {
ScopedContext Active(hDevice);
Expand Down
9 changes: 5 additions & 4 deletions source/adapters/cuda/image.hpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -21,14 +21,15 @@ ur_result_t
urToCudaImageChannelFormat(ur_image_channel_type_t image_channel_type,
ur_image_channel_order_t image_channel_order,
CUarray_format *return_cuda_format,
size_t *return_pixel_types_size_bytes);
size_t *return_pixel_types_size_bytes,
unsigned int *return_normalized_dtype_flag);

ur_result_t
cudaToUrImageChannelFormat(CUarray_format cuda_format,
ur_image_channel_type_t *return_image_channel_type);

ur_result_t urTextureCreate(ur_context_handle_t hContext,
ur_sampler_desc_t SamplerDesc,
ur_result_t urTextureCreate(ur_sampler_handle_t hSampler,
const ur_image_desc_t *pImageDesc,
CUDA_RESOURCE_DESC ResourceDesc,
const CUDA_RESOURCE_DESC &ResourceDesc,
const unsigned int normalized_dtype_flag,
ur_exp_image_native_handle_t *phRetImage);

0 comments on commit 1e40614

Please sign in to comment.