-
Notifications
You must be signed in to change notification settings - Fork 17
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.
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
CP029 Image accessors with format information #131
base: master
Are you sure you want to change the base?
Changes from 2 commits
File filter
Filter by extension
Conversations
Jump to
Diff view
Diff view
There are no files selected for viewing
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,169 @@ | ||
| Proposal ID | CP029 | | ||
| ---------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------- | | ||
| Name | Image accessors with format information | | ||
| Date of Creation | 2 July 2020 | | ||
| Target | Vendor extension | | ||
| Current Status | *Under Review* | | ||
| Original authors | Steffen Larsen [email protected], Alexander Johnston [email protected] | | ||
| Contributors | Steffen Larsen [email protected], Alexander Johnston [email protected], Victor Lomüller [email protected], Ruyman Reyes [email protected] | | ||
|
||
# Image accessors with format information | ||
|
||
## Motivation | ||
|
||
In SYCL 1.2.1 operations on images are closely tied to the corresponding | ||
operations in OpenCL thus only allowing read and write operations to be done | ||
using 4-element vectors of 32-bit floating points, 32-bit signed integers, | ||
32-bit unsigned integers, or 16-bit floating points, which are in turn converted | ||
to the types of the channels in the image. Because of this, backends that cannot | ||
query the types of the channels in the image cannot determine the necessary | ||
conversions to do. | ||
|
||
## Overview | ||
|
||
This extension adds `sycl::short4`, `sycl::ushort4`, `sycl::schar4`, and | ||
`sycl::uchar4` to the list of types supported for `dataT` for image accessors. | ||
In order to carry information about the format of the image, the | ||
`image_format` enumerator from SYCL 2020 is adopted, given as | ||
|
||
```c++ | ||
namespace sycl { | ||
enum class image_format : unsigned int { | ||
r8g8b8a8_unorm, | ||
r16g16b16a16_unorm, | ||
r8g8b8a8_sint, | ||
r16g16b16a16_sint, | ||
r32b32g32a32_sint, | ||
r8g8b8a8_uint, | ||
r16g16b16a16_uint, | ||
r32b32g32a32_uint, | ||
r16b16g16a16_sfloat, | ||
r32g32b32a32_sfloat, | ||
b8g8r8a8_unorm, | ||
}; | ||
} // namespace sycl | ||
``` | ||
|
||
and mappings from `image_format` to the corresponding accessor data type is | ||
defined as | ||
|
||
```c++ | ||
|
||
namespace sycl { | ||
template<image_format> struct image_format_type; | ||
|
||
template<> struct image_format_type<image_format::r8g8b8a8_unorm> { | ||
using type = float4; | ||
}; | ||
template<> struct image_format_type<image_format::r16g16b16a16_unorm> { | ||
using type = float4; | ||
}; | ||
template<> struct image_format_type<image_format::r8g8b8a8_sint> { | ||
using type = schar4; | ||
}; | ||
template<> struct image_format_type<image_format::r16g16b16a16_sint> { | ||
using type = short4; | ||
}; | ||
template<> struct image_format_type<image_format::r32b32g32a32_sint> { | ||
using type = int4; | ||
}; | ||
template<> struct image_format_type<image_format::r8g8b8a8_uint> { | ||
using type = uchar4; | ||
}; | ||
template<> struct image_format_type<image_format::r16g16b16a16_uint> { | ||
using type = ushort4; | ||
}; | ||
template<> struct image_format_type<image_format::r32b32g32a32_uint> { | ||
using type = uint4; | ||
}; | ||
template<> struct image_format_type<image_format::r16b16g16a16_sfloat> { | ||
using type = half4; | ||
}; | ||
template<> struct image_format_type<image_format::r32g32b32a32_sfloat> { | ||
using type = float4; | ||
}; | ||
template<> struct image_format_type<image_format::b8g8r8a8_unorm> { | ||
using type = float4; | ||
}; | ||
|
||
template<image_format format> | ||
using image_format_t = typename image_format_type<format>::type; | ||
} // namespace sycl | ||
``` | ||
|
||
With these, a new overload of `get_access` is added to `image` defined as | ||
|
||
```c++ | ||
template <image_format format, access::mode accessMode> | ||
accessor<image_format_t<format>, dimensions, accessMode, access::target::image> | ||
get_access(handler & commandGroupHandler); | ||
``` | ||
|
||
Calling `read` or `write` on an image accessor retrieved through the new | ||
`get_access` above causes undefined behaviour unless the `image_channel_type` of | ||
the underlying image and the `image_format` given to `get_access` is one of the | ||
combinations shown in the following table: | ||
|
||
| `image_channel_type ` | `image_format` | | ||
There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. It would be useful to specify the type of There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. I don't think this would be the table for it, but I could add a new table with the mappings if that would be useful? |
||
|-------------------------|-----------------------| | ||
| `snorm_int8` | `r8g8b8a8_snorm` | | ||
| `snorm_int16` | `r16g16b16a16_unorm` | | ||
| `unorm_int8` | `r8g8b8a8_unorm` | | ||
| `unorm_int16` | `r16g16b16a16_unorm` | | ||
| `signed_int8` | `r8g8b8a8_sint` | | ||
| `signed_int16` | `r16g16b16a16_sint` | | ||
| `signed_int32` | `r32b32g32a32_sint` | | ||
| `unsigned_int8` | `r8g8b8a8_uint` | | ||
| `unsigned_int16` | `r16g16b16a16_uint` | | ||
| `unsigned_int32` | `r32b32g32a32_uint` | | ||
| `fp16` | `r16b16g16a16_sfloat` | | ||
| `fp32` | `r32g32b32a32_sfloat` | | ||
|
||
## Example | ||
|
||
```c++ | ||
cl::sycl::image<2> image(hostPtr, cl::sycl::image_channel_order::rgba, | ||
cl::sycl::image_channel_type::unsigned_int8, | ||
cl::sycl::range<2>{4, 4}); | ||
|
||
cl::sycl::buffer<cl::sycl::uchar4, 1> resultDataBuf(&resultData, | ||
cl::sycl::range<1>(1)); | ||
|
||
cl::sycl::queue myQueue(selector); | ||
myQueue.submit([&](cl::sycl::handler &cgh) { | ||
auto imageAcc = | ||
image.get_access<image_format::r8g8b8a8_uint, cl::sycl::access::mode::read>(cgh); | ||
cl::sycl::accessor<cl::sycl::uchar4, 1, cl::sycl::access::mode::write> | ||
resultDataAcc(resultDataBuf, cgh); | ||
|
||
cgh.single_task<unsampled_kernel_class>([=]() { | ||
cl::sycl::uchar4 RetColor = imageAcc.read(coord); | ||
resultDataAcc[0] = RetColor; | ||
}); | ||
}); | ||
``` | ||
|
||
## OpenCL implementation Note | ||
|
||
Given that image read and write operations in OpenCL only support `cl_int4`, | ||
`cl_uint4`, `cl_float4`, and `cl_half4`, which have direct relation to the | ||
previously supported data types for image accessors, namely `sycl::int4`, | ||
`sycl::uint4`, `sycl::float4`, and `sycl::half4` respectively. By adding support | ||
for `sycl::short4`, `sycl::ushort4`, `sycl::schar4`, and `sycl::uchar4` an | ||
OpenCL backend will require conversions to the types corresponding to those | ||
supported by OpenCL after a read and before a write. That is `sycl::short4` and | ||
`sycl::schar4` must be converted to and from `sycl::int4`, whereas | ||
`sycl::ushort4` and `sycl::uchar4` must be converted to and from `sycl::uint4`. | ||
|
||
## SYCL 2020 Note | ||
|
||
Though not included in this proposal, we believe a change in this direction or a | ||
similar one should be made to the SYCL 2020 provisional too. In SYCL 2020 the | ||
data type of an image accessor is also currently restricted to `sycl::int4`, | ||
`sycl::uint4`, `sycl::float4`, and `sycl::half4`. This should be updated to be | ||
deduced from the `image_format` of the image the image accessor is produced | ||
from. This will enable each backend to perform appropriate hardware supported | ||
operations for image reads and writes, and use the appropriate data type for | ||
them. The OpenCL backend will still be able to use the four types listed | ||
above, while the CUDA backend, and others where appropriate, will be able to | ||
take and data types suitable for their hardware operations. |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Could we maybe also define a helper alias:
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Good call. I have added this and changed the data type of the accessor returned by the new
get_accessor
to use this.I'm not completely convinced that
image_format_type
is the best name for this struct, so I'm open to alternatives.There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Perhaps
image_access_type
?There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I think that's a good name, though I'm not fond of the
_type
suffix. I've renamed it toimage_access
. What do you think?