Support hipLaunchCooperativeKernelMultiDevice()
- Add validation logic for MGPU launches to pass a cuda test
Change-Id: Iccca7fde43493fc3bc6685512d39202271ae3e92
[ROCm/hip commit: 5fe91ccb1b]
Support hipLaunchCooperativeKernelMultiDevice()
- Add hipCooperativeLaunchMultiDeviceNoPreSync and
hipCooperativeLaunchMultiDeviceNoPostSync support to pass a cuda test
Change-Id: If518f11ef2636a2235e5df9e77f879d8ced68102
[ROCm/hip commit: da1444bfc8]
These fixes address regressions caused by http://gerrit-git.amd.com/c/compute/ec/hip/+/337601
Currently we're converting a 1D offset into a 3D offset, which doesn't make much sense once you consider the fact that this offset is relative to a different origin than our current 3D offset.
I traced through our blit kernels in VDI - the copy buffer rect path is able to handle immediate offsets in the 3D buffer via the amd::BufferRect::start_ parameter.
Instead of adjusting the offset, simply adjust the start of the region.
Change-Id: Ic8797a2c8ac0ad106f246f61ff06ca1ca03d3058
[ROCm/hip commit: 1bd640b659]
Even though the runtime and driver texture object API is one to one, the structs used by these APIs are not. See hipResourceDesc vs HIP_RESOURCE_DESC differences.
These differences are not trivial and most likely won't be able to handled by hipify, so we need new API entry points.
Change-Id: Id4bcb1ad0ae15378dbdb5a2ed07e5ea30f320082
[ROCm/hip commit: aea688b79c]
This patch is a workaround to support user pitch for hipMemcpy{2D/3D}.
Historically OpenCL didn't support pitch with clEnqueueFillBuffer(), so neither did we in VDI. Adding it now will be slightly nontrivial, since the fill kernel and runtime in many places will need to be modified.
As a temporary workaround for cases when pitch > width, we can just enqueue a fill for each row separately. This implementation is slow, but it satisfies the correctness criteria.
Change-Id: Idfeca349288b51d6ff84a7cf001fb63c6a66818a
[ROCm/hip commit: 77223a8eca]
- Use symbol value as the qeury key. Compared to the symbol name, the
symbol value is more robust as developers may use unqualified or
qualified identifiers. It also removes the mangling and/or demangling
requirement for the runtime API.
Change-Id: I9d4259f3842612c7cc98551269fc2092d8b5c19e
[ROCm/hip commit: b72196613a]
Adjust the origin of the copy if the user passes a pointer that wasn't allocated by the runtime.
Change-Id: I0aeb20195ed730857a461a53f537626ec2573fd1
[ROCm/hip commit: 6ed73f50f7]
The only requirment from hipMallocPitch() is that the returned pitch is aligned to the HW image pitch alignment. There is no restriction on the size of the allocation, since the memory might not be used for images.
Change-Id: I97438e5fe4012ca4721b14b85f514dbac803c17c
[ROCm/hip commit: a91b82f00e]
Instead of using the sampler field force_degamma to perform sRGB->linear conversion during pixel sampling, we use an appropriate image format instead. The overhead of this is having to create an image view when creating a texture object from an array.
Change-Id: I1ca368c312c1fd4b6f784a3a1b35b5eeb28070ff
[ROCm/hip commit: 8ddeeb4551]
This avoids the use of extra casts when obtaining a texture object handle.
Change-Id: I42df22bdad0ab9ac6c33cb8b282dee65fe7cfd6e
[ROCm/hip commit: c1475f948e]
The struct consists of a union - only the active object should be read.
Change-Id: I1c40965b61518acd91a2dcbae92a015ac9be346a
[ROCm/hip commit: fbe30090a2]
This also adds declarations of all the missing texture APIs.
hipTexRefSet*() functions need to take a textureReference as a ptr for type erasure to work. Runtime has been modified to accomodate this.
This change only applies to VDI.
Change-Id: Icf43cc5bd44dfc2c39084b7fe56d5a793bf7319f
[ROCm/hip commit: 2028b6eb29]
We don't program the numChannels and format members (these are HCC specific), so printing these will only display garbage.
Change-Id: I83dc8be9a3cae2659c64f4594d07c05330d2dd14
[ROCm/hip commit: 39558dc1b9]
All we have to do is align the ptr to HW requirments an if it's not zero, then return the offset to the user.
We currently don't have anywhere to store this offset, so hipGetTextureAlignmentOffset() will still always return 0.
Change-Id: If31998127d99a2a3222a026d88249519d6102505
[ROCm/hip commit: ea701777d8]
Change-Id: If419d2fad490d0ed50eb1315af809fc1deda1ce3
SWDEV-227875: Add a lock in streams to lock when the callback is call so we make sure things aren't moving forward in the stream
[ROCm/hip commit: 653277bd3f]
Change-Id: I1f107fc8a5c586cd571f0280ed8716c5f89d25b7
SWDEV-227875: Need to add a dummy marker in case the stream is empty.
[ROCm/hip commit: 3dfbfc408b]
What Cuda refers to "linear texture memory" is the OpenCL equivalent of CL_MEM_OBJECT_IMAGE1D_BUFFER. For these types of allocations we should create a typed buffer instead of an image.
Currently there is no check in the texture fetch functions as to what kind of SRD is written into the texture object, so any kind of incorrect programming will cause the TA to hang. Fortunately for us, every one writes correct code :)
Change-Id: I80dab85a992f2c0754ebf303d40ac6b5e045c7c1
[ROCm/hip commit: 4829a7c215]
These are artifacts left from HIP-HCC and now are not needed by HIP-VDI.
Change-Id: Ib25a1081fe6146c8a89659395151e9d5bdaf7519
[ROCm/hip commit: 2bad9e2821]
German Andryeyev
Vladislav Sytchenko
Michael LIAO
Christophe Paquot
Saleel Kudchadker
kjayapra-amd
Anusha Godavarthy Surya
Payam