Update/buffer fill do not merge before buffer copy PR #123
Conversation
leonematt
changed the title
sjw36
left a comment
sjw36
left a comment
There was a problem hiding this comment.
A few minor changes for the api.
We shouldn't disable tests at this point. Really need to be adding more tests.
plugins/cuda/cuda_runtime.cpp
Outdated
| } | ||
|
|
||
| // Add the 'const' to match the header | ||
| extern "C" nxs_status NXS_API_CALL nxsFillBuffer(nxs_int buffer_id, const void *fill_value) { |
There was a problem hiding this comment.
This should pass a size of bytes of the fill_value. Assuming float seems arbitrary.
And we should probably have a special case for zero that just uses cudaMemset.
There was a problem hiding this comment.
Is it fine if we just use the memcpy instead of memset given that we will probably end up replacing these cuda calls with a shader kernel call?
leonematt
force-pushed
the
update/buffer_fill
branch
from
d8a0daf to
d8f0182
Compare
leonematt
force-pushed
the
update/buffer_fill
branch
from
d8f0182 to
8923626
Compare
| auto buf1 = dev0.createBuffer(size, vecB.data()); | ||
| auto buf2 = dev0.createBuffer(size, vecResult_GPU.data()); | ||
|
|
||
|
|
||
| // 3. The "Inefficient" but Reliable Method: | ||
| // Calculate how many floats we need to fill the allocated space | ||
| size_t num_elements = size / sizeof(int); |
There was a problem hiding this comment.
size should be the number of bytes of value.
There was a problem hiding this comment.
Also add a check for all zeros in value, then call cudaMemset.
There was a problem hiding this comment.
I see, so not how many bytes of the buffer should be filled, but the size of the input.
There was a problem hiding this comment.
Yes, it should fill the entire buffer.
| size_t num_elements = size / sizeof(int); | ||
|
|
||
| // Create a temporary host buffer and fill it using the CPU | ||
| std::vector<int> host_gold_standard(num_elements, (int)(intptr_t)value); |
There was a problem hiding this comment.
The vector should be bytes (char) and the size should be buffer->size(). You'll need another way to set the value, like std::fill for different size bytes (ie. 1, 2, 4).
There was a problem hiding this comment.
I suppose you will need to handle the case of size not aligned with buffer->size().
2 participants
No description provided.