Todo.xml

<?xml-stylesheet type="text/xsl" href="Admin/XSL/Todo.xsl" ?> 
<topics
	 xmlns:r="http://www.r-project.org"
	 xmlns:omg="http://www.omegahat.org">

<topic>
<title>
 RCUDA Todo List
</title>

<items>

<item status="done">
tests/doubleFloat.R - failing with the floats. Getting the wrong answer.
Need .numericAsDouble to cover all parameters, not just the numeric ones.
So needed one for the length(x) as well as the x and out.
</item>



<item>
cuDeviceGetAttribute - 
CU_DEVICE_ATTRIBUTE_COMPUTE_MODE returns an enum. We need to map it to the R name-value pair.
<item>
cuDeviceGetAttribute - put the name of the attibute queried on the result.
Unfortunately, it is auto created so we'd have to intercede there, i.e., using typeMap 
or hardcoding how to convert these types of requests in cuda.createNativeProxy().
</item>
<item>
Set the default device to 1 for cuDeviceGetName() and others like it.
</item>
<item>
Make the names of the functions consistent
getDeviceProperties versus cuDeviceGetAttributes.
getDeviceProperties uses a deprecated function, so just kill off.
Different choices for approach.
</item>

<item>
R_cuModuleGetFunction doesn't seem to return the function.
Instead it seems to expect a CUfunction pointer when this is probably the return value.
The function isOutPointerType is the culprit and we explicitly identify pointers to CUfunction as not out arguments in
that function! (See the return() expression.)
<br/>
What routines have a parameter that is a pointer to a CUfunction?
<r:code>
rc = r [ grep("/cuda/", sapply(r, getFileName)) ]
</r:code>
<r:code><![CDATA[
isCUfunction = function(ty) isPointerType(ty) && getName(getPointeeType(ty)) == "CUfunction"
i = sapply(rc, function(f) any(sapply(f@params, function(cur) isCUfunction(getType(cur)))))
]]></r:code>
So only one!  The other routines take a CUfunction value, not a pointer to one. So this is an out argument.
<br/>
If we remove the CUfunction from the vector of special names,
<r:expr>cuda.createNativeProxy(rc$cuModuleGetFunction)</r:expr>
almost behaves. The CUfunction is allocated on the stack. Is this okay?
<br/>
But what about the other routines.
<r:code>
isCUfunction = function(ty) getName(ty) == "CUfunction"
i = sapply(rc, function(f) any(sapply(f@params, function(cur) isCUfunction(getType(cur)))))
lapply(rc[i], cuda.createNativeProxy)
</r:code>
<br/>
A CUfunction is already a pointer to a CUfunc_st. But this is irrelevant.
</item>

<item>
Finish off generating the documents for the functions and classes (enums) not documented.
<br/>
Enhance the return/value to mirror what we do in the R code, not the  C declarations and their documentation.
</item>

  <!-- From TODO -->
<item status="done"> Turn on USE_S4_ENUMS in src/Makevars.in</item>

<item>Process only the routines and data structures related to the host/CPU
   ignore the ones for use on the device/GPU
   Are all of the routines in the library for the host?
</item>

<item>Should we remove the cuda/cublas prefix from the names of the R functions.</item>

<item status="done"> Copy the converter code from RAutoGenRuntime into the package.</item>

<item>Change the class of the bitwise enum types that are wrong.
   only CUarray_format_enum
</item>

<item>Generate routines.
<br/>
See TU/clang.R
</item>

<item>
Remove the code from inst/generateCode.
</item>

<item status="finish">
Put a version check on the enums.
<br/>
In configure.in and onLoad.R
</item>


<item status="complete">
global option to use double rather than float.
And use it everywhere. Connect to strict for cudaMalloc.
</item>

<item status="done">
copy to device for a numeric comes back with size of the elements being 8
and type double. Should this be 4 and float.
See tests/doubleFloat.R
</item>

<item status="done">
If GPU supports double, then don't use float.
(Change the type in cudaMalloc and add code for copying doubles.)
</item>

<item>
seg faulting when quitting out of R, at least on lipschitz.
<br/>
Could it be the version of gcc, R, etc. No, looks like CUDA library issue.
<br/>
If we allocate memory, we don't get the segfault.
So I've added a C routine we call in .onLoad() 
to create the context and allocate a byte.
<br/>
Looks like a clean up issue in libcuruntime.
Similar issue in pycuda but with libcudarand.
</item>

<item status="done">
test synchronization.
See tests/async.R
</item>

<item>
Finalizer on the stream.
</item>

<item>
Copy asynchronously.
Does this make sense in R.
Use pinned memory allocated with cudaMallocHost().
</item>

<item status="working">
streams - queues of tasks.
Example - distances and then  clustering. Get the gputools kernels working
for the hierarchical clustering.
</item>

<item status="done">
Events on streams. Do they make sense in R, i.e. single thread?
Yes.
Can use a single routine and a call the R function as user data for callbacks.
See tests/event.R tests/eventSync.R
</item>

<item>
Implement cuCtxGetStreamPriorityRange and any other routines.
Just have the one for the Device, not the Ctx
</item>

<item status="check">
cudaMalloc() etc. should return an object derived from cudaPtrWithLength
so we know the size.
</item>

<item>
Problem when checking tag on a reference. 
Specifically void and voidPtr.
See inst/doc/distPitch.R.
<r:code>
library(RCUDA)
m = matrix(as.numeric(1:20), 5, 4)
mem = cudaMallocPitch(ncol(m) * 4L, nrow(m))
ref = convertToPtr(t(m), "float")

cudaMemcpy2D(mem[[1]], mem[[2]], ref, ncol(m)*4L, ncol(m)*4L, nrow(m), cudaMemcpyHostToDevice)
cudaMemcpy2D(ref, mem[[2]], mem[[1]], ncol(m)*4L, ncol(m)*4L, nrow(m), cudaMemcpyDeviceToHost)
</r:code>
<br/>
We can disable this test on the tag but it would be better  to get it right and consistent, not a special case
or string comparisons on the tag.
</item>

<item status="done">
cudaMemcpy2D should coerce to a voidPtr, not a void.
Fixed in the  makeCoerceArg function in the RCodeGen package.
</item>

<item status="check">
Need a mechanism to convert R object to pointer in cudaMemcpy2D, i.e. the src
See convertToPtr.
Checked for float.
</item>

<item>
AB = matrix(1:(300*299), 300, 299)
mem = cudaMallocPitch( ncol(AB) * 4L,  nrow(AB))
RCUDA:::cudaMemcpy2D(mem[[1]], mem[[2]], t(AB), nrow(AB)*4L, nrow(AB)*4L, ncol(AB), RCUDA:::cudaMemcpyHostToDevice)
<br/>
Fails trying to coerce the matrix to a voidPtr.
Should this be a void. But still no method to coerce R object to void.
</item>

<item status="done">
[Check]  Higher-level R-like functions for cudaMallocPitch &amp; cudaMemcpy2D
See inst/doc/distPitch.R
</item>

<item status="check">
cudaMemcpy2D should raise an error if the C routine doesn't return 0.
The C routine returns an object of class cudaError_t.
This is different CUresult.
So we have to generate the code differently to understand which
error type it is getting.
</item>


<item>
Make the code that expects a device number consistently us as(, "CUDeviceNum") or as (, "CUdevice").
Auto-generated code such as cudaDeviceGetPCIBusId causes problems as the declaration is integer.
Also, cuDeviceGet() gets back the correct number but then calls as(num, "CUdevice") which decrements
the value so it is wrong. When we get back from C code, we should leave it as is.
So new("CUdevice", value) or have the C code do it. 
When the device is an argument, then do the subtraction.

<item status="check">
integer to CUdevice should subtract 1.
And when reutrn a CUdevice, put a class on it
so that we keep it as is when we pass it to an
R function/C routine.
Put this in the typeMap.
</item>

</item>

<item status="optimize">
Subsetting and assigning to parts of a cudaPtrWithLength.
subsetting done - integer, logical indices.
Subset assignment not done
</item>

<item status="low">
Show how to use structs in PTX code and pass them from R as inputs.
</item>

<item status="low">
Allow obj[] to take a routine to copy each element to a SEXP.
Caller specifies a native symbol.
</item>

<item status="low">
cudaMalloc should allow specification of the device.
Implicit in the current context?
</item>

<item status="low">
Allow .device/.gpu argument in .gpu/.cuda function to switch
to a specific device. Probably too much overhead for common use.
</item>

<item status="done">
In cuGetContext(), check if cuCtxGetCurrent returns a NULL pointer, not a NULL object.
<br/>
Move C code and R function to RAutoGenRunTime.
Isn't this already done - isNativeNull?
</item>

<item status="check">
Make certain to clean up so don't run out of memory across sessions.
If quit R, want to release resources.
</item>

<item status="done">
Examples - perhaps taken from gputools or rgpu
but done directly  from R code, not with C wrappers.
<br/>
See dist stuff in sampleKernels and  Paper/
</item>

<item status="test">
Configure script.
</item>

<item status="InProgress">
Generate bindings via TU and Clang
<br/>
Ignore the deprecated ones. How can we tell in clang.
<br/>
Add default arguments for types such as the device to be 1L or new("CUdevice", 0L).
</item>

<item>
Functions to manipulate module.
Anyway  to find names in a module?
In a C++ API - http://adsm.googlecode.com/svn/trunk/libgmac/src/api/cudadrv/Module.h.
</item>

<item status="low">
Write function for reading profiler in key=value  form.
</item>

<item status="done">
Profiler
</item>

<item status="check">
Class information on the cudaAlloc() returns so that
we know how to retrieve the result later.
Put length information on it also.
Make them RC++Reference and not just external pointers.
</item>


<item status="done">
Find out what the problem is with cubin files and not being able to load them.
<br/>
Need to get the nvcc flags to generate code for the correct device.
</item>


</items>
</topic>
</topics>