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
Currently we extract the read mode from the ihip*impl() calls, which is not correct. We should be getting it from the texture itself directly.
Change-Id: Idf6449fefa395a887138a252e8ea937a6897e600
reinterpret_cast<> doesn't create an object, so the texref is actually unitiliazed. This may lead to garbage data in some of its struct members.
Initialize it by performing a placement new. The constructer should set all of its members to default values. There's no way currently to extract the channel type, so use single channel char for now.
Change-Id: I41b305a75bb3f30130324de785099f55b3e130c7
HIP assumes that image width is in bytes, but VDI assumes that image width in pixels. Need to perform byte -> pixel coversion before doing anything.
Change-Id: Ia9fd1f46d05db3fbe8049add10b4d7e5118a2b9a
~45% to 50% of Performance drop on rocBLAS_int8 test
Enable cudaSetDeviceFlags() api call. Use active wait by default
for all devices.
Change-Id: Ifc2ebe3dd9b0aa3fdbfbc9cb5c2cd8b3b726124f
Incoming changes from upstream split the struct hipMemcpy3DParms into two separate ones - hipMemcpy3DParms and HIP_MEMCPY3D, which are cudaMemcpy3DParms and CUDA_MEMCPY3D equivalents respectively.
Note that HIP_MEMCPY3D is missing half the members of CUDA_MEMCPY3D (this should be fixed in PR#1887). Work around this by using a substitute _HIP_MEMCPY3D struct for now.
Change-Id: Ic15134e6deb260189b662b3804d2309a9b8473e9
Add hip::syncStreams(dev) to sync blocking streams on a given device.
hip::syncStreams(void) should only sync streams on the current device.
Change-Id: Ib6b0735215fa0ed12c646ebd029e9763ee3712ce
If a user passes a ptr that was allocated by hipMalloc(), the offset is guaranteed to be 0 and NULL may be passed as the offset parameter. We shouldn't return an error in this case.
Change-Id: I4a8d645121e5a17d5e2861a0629356a3599de9ee
Also make D2H and H2D keep track of the chain of events
when we need to use a different HostQueue.
Change-Id: I1c5da6ea6104b37ad7aac00f0eb8ea9371e6ba1c
The current texture implementation is based off the one for HIP-HCC. There's a lot of problems with it - only creating images from buffers, hard coding logic and ignoring user parameters. This leads to a whole lot of UB even with simple examples (as seen with RedShift's code).
This CL is aimed to bring the HIP-VDI texture implementation closer to what is described by Cuda.
hipMemcpyAtoA() - image to image copy.
hipMemcpyHtoA()/hipMemcpyDtoA() - buffer to image copy.
hipMemcpyAtoH()/hipMemcpyAtoD() - image to buffer copy.
hipArrayCreate()/hipArray3DCreate()/hipMallocArray()/hipMalloc3DArray() - creates 1D/2D/3D/1D Array/2D Array images.
hipCreateTextureObject() - creates sampler, (optional) creates 1D/2D image from buffer, (optional) creates image views.
hipBindTexture() - creates 1D image from buffer (should create a typed buffer, however this is not compatible with HIP-HCC).
hipBindTexture2D() - creates 2D image form buffer.
hipBindTextureToArray() - creates image view.
hipTexRefSetAddress() - creates 1D image from buffer (should create a typed buffer, however this is not compatible with HIP-HCC).
hipTexRefSetAddress2D() - creates 2D image from buffer.
hipTexRefSetArray() - creates image view.
There are still a lot of TODOs in the code, here's a few important ones:
1. VDI doesn't support a lot of sampler flags.
2. VDI doesn't support device to image 2D/3D copy.
3. Mipmaps implementation is incomplete.
4. Image view implementation is incomplete.
Change-Id: Ia374ee27aa14f76451fee7667495036f4419a487