|
<html><head><title>dlib C++ Library - cuda_utils.h</title></head><body bgcolor='white'><pre> |
|
<font color='#009900'>// Copyright (C) 2015 Davis E. King ([email protected]) |
|
</font><font color='#009900'>// License: Boost Software License See LICENSE.txt for the full license. |
|
</font><font color='#0000FF'>#ifndef</font> DLIB_CUDA_UtILS_H_ |
|
<font color='#0000FF'>#define</font> DLIB_CUDA_UtILS_H_ |
|
|
|
<font color='#0000FF'>#include</font> "<a style='text-decoration:none' href='../algs.h.html'>../algs.h</a>" |
|
|
|
<font color='#0000FF'>#ifndef</font> DLIB_USE_CUDA |
|
<font color='#0000FF'>#error</font> "<font color='#CC0000'>This file shouldn't be #included unless DLIB_USE_CUDA is #defined</font>" |
|
<font color='#0000FF'>#endif</font> |
|
|
|
<font color='#0000FF'>#include</font> "<a style='text-decoration:none' href='cuda_errors.h.html'>cuda_errors.h</a>" |
|
<font color='#0000FF'>#include</font> <font color='#5555FF'><</font>cmath<font color='#5555FF'>></font> |
|
|
|
<font color='#0000FF'>#include</font> <font color='#5555FF'><</font>cuda_runtime.h<font color='#5555FF'>></font> |
|
<font color='#0000FF'>#include</font> <font color='#5555FF'><</font>sstream<font color='#5555FF'>></font> |
|
<font color='#0000FF'>#include</font> <font color='#5555FF'><</font>iostream<font color='#5555FF'>></font> |
|
<font color='#0000FF'>#include</font> <font color='#5555FF'><</font>memory<font color='#5555FF'>></font> |
|
<font color='#0000FF'>#include</font> <font color='#5555FF'><</font>vector<font color='#5555FF'>></font> |
|
<font color='#0000FF'>#include</font> <font color='#5555FF'><</font>type_traits<font color='#5555FF'>></font> |
|
|
|
|
|
<font color='#009900'>// Check the return value of a call to the CUDA runtime for an error condition. |
|
</font><font color='#0000FF'>#define</font> CHECK_CUDA<font face='Lucida Console'>(</font>call<font face='Lucida Console'>)</font> \ |
|
<font color='#0000FF'>do</font><b>{</b> \ |
|
<font color='#0000FF'>const</font> cudaError_t error <font color='#5555FF'>=</font> call; \ |
|
<font color='#0000FF'>if</font> <font face='Lucida Console'>(</font>error <font color='#5555FF'>!</font><font color='#5555FF'>=</font> cudaSuccess<font face='Lucida Console'>)</font> \ |
|
<b>{</b> \ |
|
std::ostringstream sout; \ |
|
sout <font color='#5555FF'><</font><font color='#5555FF'><</font> "<font color='#CC0000'>Error while calling </font>" <font color='#5555FF'><</font><font color='#5555FF'><</font> #call <font color='#5555FF'><</font><font color='#5555FF'><</font> "<font color='#CC0000'> in file </font>" <font color='#5555FF'><</font><font color='#5555FF'><</font> __FILE__ <font color='#5555FF'><</font><font color='#5555FF'><</font> "<font color='#CC0000'>:</font>" <font color='#5555FF'><</font><font color='#5555FF'><</font> __LINE__ <font color='#5555FF'><</font><font color='#5555FF'><</font> "<font color='#CC0000'>. </font>";\ |
|
sout <font color='#5555FF'><</font><font color='#5555FF'><</font> "<font color='#CC0000'>code: </font>" <font color='#5555FF'><</font><font color='#5555FF'><</font> error <font color='#5555FF'><</font><font color='#5555FF'><</font> "<font color='#CC0000'>, reason: </font>" <font color='#5555FF'><</font><font color='#5555FF'><</font> <font color='#BB00BB'>cudaGetErrorString</font><font face='Lucida Console'>(</font>error<font face='Lucida Console'>)</font>;\ |
|
<font color='#0000FF'>throw</font> dlib::<font color='#BB00BB'>cuda_error</font><font face='Lucida Console'>(</font>sout.<font color='#BB00BB'>str</font><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font><font face='Lucida Console'>)</font>; \ |
|
<b>}</b> \ |
|
<b>}</b><font color='#0000FF'>while</font><font face='Lucida Console'>(</font><font color='#979000'>false</font><font face='Lucida Console'>)</font> |
|
|
|
<font color='#009900'>// ---------------------------------------------------------------------------------------- |
|
</font> |
|
<font color='#0000FF'>#ifdef</font> __CUDACC__ |
|
|
|
<font color='#0000FF'>namespace</font> dlib |
|
<b>{</b> |
|
<font color='#0000FF'>namespace</font> cuda |
|
<b>{</b> |
|
|
|
<font color='#009900'>// ------------------------------------------------------------------------------------ |
|
</font> |
|
__inline__ __device__ <font color='#0000FF'><u>size_t</u></font> <b><a name='pack_idx'></a>pack_idx</b> <font face='Lucida Console'>(</font> |
|
<font color='#0000FF'><u>size_t</u></font> dim_size3, |
|
<font color='#0000FF'><u>size_t</u></font> dim_size2, |
|
<font color='#0000FF'><u>size_t</u></font> dim_size1, |
|
<font color='#0000FF'><u>size_t</u></font> idx4, |
|
<font color='#0000FF'><u>size_t</u></font> idx3, |
|
<font color='#0000FF'><u>size_t</u></font> idx2, |
|
<font color='#0000FF'><u>size_t</u></font> idx1 |
|
<font face='Lucida Console'>)</font> |
|
<font color='#009900'>/*! |
|
ensures |
|
- Converts a 4D array index into a 1D index assuming row major layout. To |
|
understand precisely what this function does, imagine we had an array |
|
declared like this: |
|
int ARRAY[anything][dim_size3][dim_size2][dim_size1]; |
|
Then we could index it like this: |
|
ARRAY[idx4][idx3][idx2][idx1] |
|
or equivalently like this: |
|
((int*)ARRAY)[pack_idx(dim_size3,dim_size2,dim_size1, idx4,idx3,idx2,idx1)] |
|
!*/</font> |
|
<b>{</b> |
|
<font color='#0000FF'>return</font> <font face='Lucida Console'>(</font><font face='Lucida Console'>(</font>idx4<font color='#5555FF'>*</font>dim_size3 <font color='#5555FF'>+</font> idx3<font face='Lucida Console'>)</font><font color='#5555FF'>*</font>dim_size2 <font color='#5555FF'>+</font> idx2<font face='Lucida Console'>)</font><font color='#5555FF'>*</font>dim_size1 <font color='#5555FF'>+</font> idx1; |
|
<b>}</b> |
|
|
|
__inline__ __device__ <font color='#0000FF'><u>void</u></font> <b><a name='unpack_idx'></a>unpack_idx</b> <font face='Lucida Console'>(</font> |
|
<font color='#0000FF'><u>size_t</u></font> idx, |
|
<font color='#0000FF'><u>size_t</u></font> dim_size3, |
|
<font color='#0000FF'><u>size_t</u></font> dim_size2, |
|
<font color='#0000FF'><u>size_t</u></font> dim_size1, |
|
<font color='#0000FF'><u>size_t</u></font><font color='#5555FF'>&</font> idx4, |
|
<font color='#0000FF'><u>size_t</u></font><font color='#5555FF'>&</font> idx3, |
|
<font color='#0000FF'><u>size_t</u></font><font color='#5555FF'>&</font> idx2, |
|
<font color='#0000FF'><u>size_t</u></font><font color='#5555FF'>&</font> idx1 |
|
<font face='Lucida Console'>)</font> |
|
<font color='#009900'>/*! |
|
ensures |
|
- This function computes the inverse of pack_idx(). Therefore, |
|
if PACKED == pack_idx(dim_size3,dim_size2,dim_size1, idx4,idx3,idx2,idx1) |
|
then unpack_idx(PACKED,dim_size3,dim_size2,dim_size1, IDX4,IDX3,IDX2,IDX1) |
|
results in: |
|
- IDX1 == idx1 |
|
- IDX2 == idx2 |
|
- IDX3 == idx3 |
|
- IDX4 == idx4 |
|
!*/</font> |
|
<b>{</b> |
|
idx1 <font color='#5555FF'>=</font> idx<font color='#5555FF'>%</font>dim_size1; |
|
|
|
idx <font color='#5555FF'>/</font><font color='#5555FF'>=</font> dim_size1; |
|
idx2 <font color='#5555FF'>=</font> idx<font color='#5555FF'>%</font>dim_size2; |
|
|
|
idx <font color='#5555FF'>/</font><font color='#5555FF'>=</font> dim_size2; |
|
idx3 <font color='#5555FF'>=</font> idx<font color='#5555FF'>%</font>dim_size3; |
|
|
|
idx <font color='#5555FF'>/</font><font color='#5555FF'>=</font> dim_size3; |
|
idx4 <font color='#5555FF'>=</font> idx; |
|
<b>}</b> |
|
|
|
<font color='#009900'>// ------------------------------------------------------------------------------------ |
|
</font> |
|
<font color='#009900'>// This function is from the article: |
|
</font> <font color='#009900'>// http://devblogs.nvidia.com/parallelforall/faster-parallel-reductions-kepler/ |
|
</font> __inline__ __device__ <font color='#0000FF'><u>float</u></font> <b><a name='warp_reduce_sum'></a>warp_reduce_sum</b><font face='Lucida Console'>(</font><font color='#0000FF'><u>float</u></font> val<font face='Lucida Console'>)</font> |
|
<b>{</b> |
|
<font color='#0000FF'>for</font> <font face='Lucida Console'>(</font><font color='#0000FF'><u>int</u></font> offset <font color='#5555FF'>=</font> warpSize<font color='#5555FF'>/</font><font color='#979000'>2</font>; offset <font color='#5555FF'>></font> <font color='#979000'>0</font>; offset <font color='#5555FF'>/</font><font color='#5555FF'>=</font> <font color='#979000'>2</font><font face='Lucida Console'>)</font> |
|
<font color='#0000FF'>#if</font> CUDART_VERSION <font color='#5555FF'>></font><font color='#5555FF'>=</font> <font color='#979000'>9000</font> |
|
val <font color='#5555FF'>+</font><font color='#5555FF'>=</font> <font color='#BB00BB'>__shfl_down_sync</font><font face='Lucida Console'>(</font><font color='#979000'>0xFFFFFFFF</font>,val, offset<font face='Lucida Console'>)</font>; |
|
<font color='#0000FF'>#else</font> |
|
val <font color='#5555FF'>+</font><font color='#5555FF'>=</font> <font color='#BB00BB'>__shfl_down</font><font face='Lucida Console'>(</font>val, offset<font face='Lucida Console'>)</font>; |
|
<font color='#0000FF'>#endif</font> |
|
<font color='#0000FF'>return</font> val; |
|
<b>}</b> |
|
|
|
__inline__ __device__ <font color='#0000FF'><u>bool</u></font> <b><a name='is_first_thread_in_warp'></a>is_first_thread_in_warp</b><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font> |
|
<b>{</b> |
|
<font color='#0000FF'>return</font> <font face='Lucida Console'>(</font>threadIdx.x <font color='#5555FF'>&</font> <font face='Lucida Console'>(</font>warpSize <font color='#5555FF'>-</font> <font color='#979000'>1</font><font face='Lucida Console'>)</font><font face='Lucida Console'>)</font> <font color='#5555FF'>=</font><font color='#5555FF'>=</font> <font color='#979000'>0</font>; |
|
<b>}</b> |
|
|
|
__inline__ __device__ <font color='#0000FF'><u>void</u></font> <b><a name='warp_reduce_atomic_add'></a>warp_reduce_atomic_add</b><font face='Lucida Console'>(</font> |
|
<font color='#0000FF'><u>float</u></font><font color='#5555FF'>&</font> out, |
|
<font color='#0000FF'><u>float</u></font> val |
|
<font face='Lucida Console'>)</font> |
|
<font color='#009900'>/*! |
|
ensures |
|
- Atomically adds all the val variables in the current warp to out. |
|
See this page for an extended discussion: |
|
http://devblogs.nvidia.com/parallelforall/faster-parallel-reductions-kepler/ |
|
!*/</font> |
|
<b>{</b> |
|
val <font color='#5555FF'>=</font> <font color='#BB00BB'>warp_reduce_sum</font><font face='Lucida Console'>(</font>val<font face='Lucida Console'>)</font>; |
|
<font color='#0000FF'>if</font> <font face='Lucida Console'>(</font><font color='#BB00BB'>is_first_thread_in_warp</font><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font><font face='Lucida Console'>)</font> |
|
<font color='#BB00BB'>atomicAdd</font><font face='Lucida Console'>(</font><font color='#5555FF'>&</font>out, val<font face='Lucida Console'>)</font>; |
|
<b>}</b> |
|
|
|
<font color='#009900'>// ------------------------------------------------------------------------------------ |
|
</font> |
|
<font color='#0000FF'>struct</font> <b><a name='max_jobs'></a>max_jobs</b> |
|
<b>{</b> |
|
<b><a name='max_jobs'></a>max_jobs</b><font face='Lucida Console'>(</font><font color='#0000FF'><u>int</u></font> x<font face='Lucida Console'>)</font> : num_x<font face='Lucida Console'>(</font>x<font face='Lucida Console'>)</font> <b>{</b><b>}</b> |
|
<b><a name='max_jobs'></a>max_jobs</b><font face='Lucida Console'>(</font><font color='#0000FF'><u>int</u></font> x, <font color='#0000FF'><u>int</u></font> y<font face='Lucida Console'>)</font> : num_x<font face='Lucida Console'>(</font>x<font face='Lucida Console'>)</font>, num_y<font face='Lucida Console'>(</font>y<font face='Lucida Console'>)</font> <b>{</b><b>}</b> |
|
<font color='#0000FF'><u>int</u></font> num_x; |
|
<font color='#0000FF'><u>int</u></font> num_y <font color='#5555FF'>=</font> <font color='#979000'>1</font>; |
|
<b>}</b>; |
|
|
|
<font color='#0000FF'>template</font> <font color='#5555FF'><</font><font color='#0000FF'>typename</font> Kernel, <font color='#0000FF'>typename</font>... T<font color='#5555FF'>></font> |
|
<font color='#0000FF'><u>void</u></font> <b><a name='launch_kernel'></a>launch_kernel</b> <font face='Lucida Console'>(</font> |
|
Kernel K, |
|
T ...args |
|
<font face='Lucida Console'>)</font> |
|
<font color='#009900'>/*! |
|
ensures |
|
- launches the given kernel K(args...). The point of this function is to |
|
automatically set the kernel launch parameters to something reasonable |
|
based on the properties of the kernel and the current GPU card. |
|
!*/</font> |
|
<b>{</b> |
|
<font color='#0000FF'><u>int</u></font> num_blocks, num_threads; |
|
<font color='#BB00BB'>CHECK_CUDA</font><font face='Lucida Console'>(</font><font color='#BB00BB'>cudaOccupancyMaxPotentialBlockSize</font><font face='Lucida Console'>(</font><font color='#5555FF'>&</font>num_blocks,<font color='#5555FF'>&</font>num_threads,K<font face='Lucida Console'>)</font><font face='Lucida Console'>)</font>; |
|
K<font color='#5555FF'><</font><font color='#5555FF'><</font><font color='#5555FF'><</font>num_blocks,num_threads<font color='#5555FF'>></font><font color='#5555FF'>></font><font color='#5555FF'>></font><font face='Lucida Console'>(</font>args...<font face='Lucida Console'>)</font>; |
|
<b>}</b> |
|
|
|
<font color='#0000FF'>template</font> <font color='#5555FF'><</font><font color='#0000FF'>typename</font> Kernel, <font color='#0000FF'>typename</font>... T<font color='#5555FF'>></font> |
|
<font color='#0000FF'><u>void</u></font> <b><a name='launch_kernel'></a>launch_kernel</b> <font face='Lucida Console'>(</font> |
|
Kernel K, |
|
max_jobs m, |
|
T ...args |
|
<font face='Lucida Console'>)</font> |
|
<font color='#009900'>/*! |
|
ensures |
|
- This function is just like launch_kernel(K,args...) except that you can |
|
additionally supply a max_jobs number that tells it how many possible |
|
total threads could be used. This is useful when launching potentially |
|
small jobs that might not need the number of threads suggested by |
|
launch_kernel(). |
|
!*/</font> |
|
<b>{</b> |
|
<font color='#0000FF'>if</font> <font face='Lucida Console'>(</font>m.num_x <font color='#5555FF'>=</font><font color='#5555FF'>=</font> <font color='#979000'>0</font> <font color='#5555FF'>|</font><font color='#5555FF'>|</font> m.num_y <font color='#5555FF'>=</font><font color='#5555FF'>=</font> <font color='#979000'>0</font><font face='Lucida Console'>)</font> |
|
<font color='#0000FF'>return</font>; |
|
<font color='#0000FF'><u>int</u></font> num_blocks, num_threads; |
|
<font color='#BB00BB'>CHECK_CUDA</font><font face='Lucida Console'>(</font><font color='#BB00BB'>cudaOccupancyMaxPotentialBlockSize</font><font face='Lucida Console'>(</font><font color='#5555FF'>&</font>num_blocks,<font color='#5555FF'>&</font>num_threads,K<font face='Lucida Console'>)</font><font face='Lucida Console'>)</font>; |
|
<font color='#009900'>// Check if the job is really small and we don't really need to launch a kernel |
|
</font> <font color='#009900'>// with this many blocks and threads. |
|
</font> <font color='#0000FF'>if</font> <font face='Lucida Console'>(</font>num_blocks<font color='#5555FF'>*</font>num_threads <font color='#5555FF'>></font> m.num_x<font color='#5555FF'>*</font>m.num_y<font face='Lucida Console'>)</font> |
|
num_blocks <font color='#5555FF'>=</font> <font face='Lucida Console'>(</font>m.num_x<font color='#5555FF'>*</font>m.num_y<font color='#5555FF'>+</font>num_threads<font color='#5555FF'>-</font><font color='#979000'>1</font><font face='Lucida Console'>)</font><font color='#5555FF'>/</font>num_threads; |
|
|
|
<font color='#0000FF'>if</font> <font face='Lucida Console'>(</font>m.num_y <font color='#5555FF'>=</font><font color='#5555FF'>=</font> <font color='#979000'>1</font><font face='Lucida Console'>)</font> |
|
<b>{</b> |
|
K<font color='#5555FF'><</font><font color='#5555FF'><</font><font color='#5555FF'><</font>num_blocks,num_threads<font color='#5555FF'>></font><font color='#5555FF'>></font><font color='#5555FF'>></font><font face='Lucida Console'>(</font>args...<font face='Lucida Console'>)</font>; |
|
<b>}</b> |
|
<font color='#0000FF'>else</font> |
|
<b>{</b> |
|
<font color='#009900'>/* |
|
In general, the reason m.num_y!=1 (i.e. the reason you are in this |
|
code path) is because we are using nested grid-stride loops. There are |
|
two important things to note about what we are doing here. To |
|
illustrate them we will talk about this little CUDA code snippet: |
|
|
|
// initialize out before we begin. |
|
for (auto i : grid_stride_range_y(0, nr)) |
|
for (auto j : grid_stride_range(0, 1)) |
|
out[i] = 0; |
|
|
|
__syncthreads(); // synchronize threads in block |
|
|
|
// loop over some 2D thing and sum and store things into out. |
|
for (auto i : grid_stride_range_y(0, nr)) |
|
{ |
|
float temp = 0; |
|
for (auto j : grid_stride_range(0, nc)) |
|
temp += whatever[i*nc+j]; |
|
|
|
// store the sum into out[i] |
|
warp_reduce_atomic_add(out[i], temp); |
|
} |
|
|
|
First, we make sure the number of x threads is a multiple of 32 so that |
|
you can use warp_reduce_atomic_add() inside the y loop. |
|
|
|
Second, we put the x block size to 1 so inter-block synchronization is |
|
easier. For example, if the number of x blocks wasn't 1 the above code |
|
would have a race condition in it. This is because the execution of |
|
out[i]=0 would be done by blocks with blockIdx.x==0, but then in the |
|
second set of loops, *all* the x blocks use out[i]. Since |
|
__syncthreads() doesn't do any synchronization between blocks some of |
|
the blocks might begin before the out[i]=0 statements finished and that |
|
would be super bad. |
|
*/</font> |
|
|
|
<font color='#009900'>// Try and make sure that the ratio of x to y threads is reasonable based |
|
</font> <font color='#009900'>// on the respective size of our loops. |
|
</font> <font color='#0000FF'><u>int</u></font> x_threads <font color='#5555FF'>=</font> <font color='#979000'>32</font>; |
|
<font color='#0000FF'><u>int</u></font> y_threads <font color='#5555FF'>=</font> num_threads<font color='#5555FF'>/</font><font color='#979000'>32</font>; |
|
<font color='#0000FF'>const</font> <font color='#0000FF'><u>int</u></font> ratio <font color='#5555FF'>=</font> <font color='#0000FF'>static_cast</font><font color='#5555FF'><</font><font color='#0000FF'><u>int</u></font><font color='#5555FF'>></font><font face='Lucida Console'>(</font>std::<font color='#BB00BB'>round</font><font face='Lucida Console'>(</font><font color='#BB00BB'>put_in_range</font><font face='Lucida Console'>(</font><font color='#979000'>1</font>, y_threads, m.num_x<font color='#5555FF'>/</font><font face='Lucida Console'>(</font><font color='#0000FF'><u>double</u></font><font face='Lucida Console'>)</font>m.num_y<font face='Lucida Console'>)</font><font face='Lucida Console'>)</font><font face='Lucida Console'>)</font>; |
|
x_threads <font color='#5555FF'>*</font><font color='#5555FF'>=</font> ratio; |
|
y_threads <font color='#5555FF'>/</font><font color='#5555FF'>=</font> ratio; |
|
|
|
dim3 <font color='#BB00BB'>blocks</font><font face='Lucida Console'>(</font><font color='#979000'>1</font>,num_blocks<font face='Lucida Console'>)</font>; |
|
dim3 <font color='#BB00BB'>threads</font><font face='Lucida Console'>(</font>x_threads,y_threads<font face='Lucida Console'>)</font>; |
|
K<font color='#5555FF'><</font><font color='#5555FF'><</font><font color='#5555FF'><</font>blocks,threads<font color='#5555FF'>></font><font color='#5555FF'>></font><font color='#5555FF'>></font><font face='Lucida Console'>(</font>args...<font face='Lucida Console'>)</font>; |
|
<b>}</b> |
|
<b>}</b> |
|
|
|
<font color='#009900'>// ------------------------------------------------------------------------------------ |
|
</font> |
|
<font color='#0000FF'>class</font> <b><a name='grid_stride_range'></a>grid_stride_range</b> |
|
<b>{</b> |
|
<font color='#009900'>/*! |
|
WHAT THIS OBJECT REPRESENTS |
|
This is a tool for making a for loop that loops over an entire block of |
|
memory inside a kernel, but doing so in a way that parallelizes |
|
appropriately across all the threads in a kernel launch. For example, |
|
the following kernel would add the vector a to the vector b and store |
|
the output in out (assuming all vectors are of dimension n): |
|
__global__ void add_arrays( |
|
const float* a, |
|
const float* b, |
|
float* out, |
|
size_t n |
|
) |
|
{ |
|
for (auto i : grid_stride_range(0, n)) |
|
{ |
|
out[i] = a[i]+b[i]; |
|
} |
|
} |
|
!*/</font> |
|
|
|
<font color='#0000FF'>public</font>: |
|
__device__ <b><a name='grid_stride_range'></a>grid_stride_range</b><font face='Lucida Console'>(</font> |
|
<font color='#0000FF'><u>size_t</u></font> ibegin_, |
|
<font color='#0000FF'><u>size_t</u></font> iend_ |
|
<font face='Lucida Console'>)</font> : |
|
ibegin<font face='Lucida Console'>(</font>ibegin_<font face='Lucida Console'>)</font>, |
|
iend<font face='Lucida Console'>(</font>iend_<font face='Lucida Console'>)</font> |
|
<b>{</b><b>}</b> |
|
|
|
<font color='#0000FF'>class</font> <b><a name='iterator'></a>iterator</b> |
|
<b>{</b> |
|
<font color='#0000FF'>public</font>: |
|
__device__ <b><a name='iterator'></a>iterator</b><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font> <b>{</b><b>}</b> |
|
__device__ <b><a name='iterator'></a>iterator</b><font face='Lucida Console'>(</font><font color='#0000FF'><u>size_t</u></font> pos_<font face='Lucida Console'>)</font> : pos<font face='Lucida Console'>(</font>pos_<font face='Lucida Console'>)</font> <b>{</b><b>}</b> |
|
|
|
__device__ <font color='#0000FF'><u>size_t</u></font> <b><a name='operator'></a>operator</b><font color='#5555FF'>*</font><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font> <font color='#0000FF'>const</font> |
|
<b>{</b> |
|
<font color='#0000FF'>return</font> pos; |
|
<b>}</b> |
|
|
|
__device__ iterator<font color='#5555FF'>&</font> <b><a name='operator'></a>operator</b><font color='#5555FF'>+</font><font color='#5555FF'>+</font><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font> |
|
<b>{</b> |
|
pos <font color='#5555FF'>+</font><font color='#5555FF'>=</font> gridDim.x <font color='#5555FF'>*</font> blockDim.x; |
|
<font color='#0000FF'>return</font> <font color='#5555FF'>*</font><font color='#0000FF'>this</font>; |
|
<b>}</b> |
|
|
|
__device__ <font color='#0000FF'><u>bool</u></font> <b><a name='operator'></a>operator</b><font color='#5555FF'>!</font><font color='#5555FF'>=</font><font face='Lucida Console'>(</font><font color='#0000FF'>const</font> iterator<font color='#5555FF'>&</font> item<font face='Lucida Console'>)</font> <font color='#0000FF'>const</font> |
|
<b>{</b> <font color='#0000FF'>return</font> pos <font color='#5555FF'><</font> item.pos; <b>}</b> |
|
|
|
<font color='#0000FF'>private</font>: |
|
<font color='#0000FF'><u>size_t</u></font> pos; |
|
<b>}</b>; |
|
|
|
__device__ iterator <b><a name='begin'></a>begin</b><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font> <font color='#0000FF'>const</font> |
|
<b>{</b> |
|
<font color='#0000FF'>return</font> <font color='#BB00BB'>iterator</font><font face='Lucida Console'>(</font>ibegin<font color='#5555FF'>+</font>blockDim.x <font color='#5555FF'>*</font> blockIdx.x <font color='#5555FF'>+</font> threadIdx.x<font face='Lucida Console'>)</font>; |
|
<b>}</b> |
|
__device__ iterator <b><a name='end'></a>end</b><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font> <font color='#0000FF'>const</font> |
|
<b>{</b> |
|
<font color='#0000FF'>return</font> <font color='#BB00BB'>iterator</font><font face='Lucida Console'>(</font>iend<font face='Lucida Console'>)</font>; |
|
<b>}</b> |
|
<font color='#0000FF'>private</font>: |
|
|
|
<font color='#0000FF'><u>size_t</u></font> ibegin; |
|
<font color='#0000FF'><u>size_t</u></font> iend; |
|
<b>}</b>; |
|
|
|
<font color='#009900'>// ------------------------------------------------------------------------------------ |
|
</font> |
|
<font color='#0000FF'>class</font> <b><a name='grid_stride_range_y'></a>grid_stride_range_y</b> |
|
<b>{</b> |
|
<font color='#009900'>/*! |
|
WHAT THIS OBJECT REPRESENTS |
|
This object is just like grid_stride_range except that it looks at |
|
CUDA's y thread index (e.g. threadIdx.y) instead of the x index. |
|
Therefore, if you launch a cuda kernel with a statement like: |
|
dim3 blocks(1,10); |
|
dim3 threads(32,32); // You need to have x and y not equal to 1 to get parallelism over both loops. |
|
add_arrays<<<blocks,threads>>>(a,b,out,nr,nc); |
|
You can perform a nested 2D parallel for loop rather than doing just a |
|
1D for loop. |
|
|
|
So the code in the kernel would look like this if you wanted to add two |
|
2D matrices: |
|
__global__ void add_arrays( |
|
const float* a, |
|
const float* b, |
|
float* out, |
|
size_t nr, |
|
size_t nc |
|
) |
|
{ |
|
for (auto r : grid_stride_range_y(0, nr)) |
|
{ |
|
for (auto c : grid_stride_range(0, nc)) |
|
{ |
|
auto i = r*nc+c; |
|
out[i] = a[i]+b[i]; |
|
} |
|
} |
|
} |
|
!*/</font> |
|
|
|
<font color='#0000FF'>public</font>: |
|
__device__ <b><a name='grid_stride_range_y'></a>grid_stride_range_y</b><font face='Lucida Console'>(</font> |
|
<font color='#0000FF'><u>size_t</u></font> ibegin_, |
|
<font color='#0000FF'><u>size_t</u></font> iend_ |
|
<font face='Lucida Console'>)</font> : |
|
ibegin<font face='Lucida Console'>(</font>ibegin_<font face='Lucida Console'>)</font>, |
|
iend<font face='Lucida Console'>(</font>iend_<font face='Lucida Console'>)</font> |
|
<b>{</b><b>}</b> |
|
|
|
<font color='#0000FF'>class</font> <b><a name='iterator'></a>iterator</b> |
|
<b>{</b> |
|
<font color='#0000FF'>public</font>: |
|
__device__ <b><a name='iterator'></a>iterator</b><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font> <b>{</b><b>}</b> |
|
__device__ <b><a name='iterator'></a>iterator</b><font face='Lucida Console'>(</font><font color='#0000FF'><u>size_t</u></font> pos_<font face='Lucida Console'>)</font> : pos<font face='Lucida Console'>(</font>pos_<font face='Lucida Console'>)</font> <b>{</b><b>}</b> |
|
|
|
__device__ <font color='#0000FF'><u>size_t</u></font> <b><a name='operator'></a>operator</b><font color='#5555FF'>*</font><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font> <font color='#0000FF'>const</font> |
|
<b>{</b> |
|
<font color='#0000FF'>return</font> pos; |
|
<b>}</b> |
|
|
|
__device__ iterator<font color='#5555FF'>&</font> <b><a name='operator'></a>operator</b><font color='#5555FF'>+</font><font color='#5555FF'>+</font><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font> |
|
<b>{</b> |
|
pos <font color='#5555FF'>+</font><font color='#5555FF'>=</font> gridDim.y <font color='#5555FF'>*</font> blockDim.y; |
|
<font color='#0000FF'>return</font> <font color='#5555FF'>*</font><font color='#0000FF'>this</font>; |
|
<b>}</b> |
|
|
|
__device__ <font color='#0000FF'><u>bool</u></font> <b><a name='operator'></a>operator</b><font color='#5555FF'>!</font><font color='#5555FF'>=</font><font face='Lucida Console'>(</font><font color='#0000FF'>const</font> iterator<font color='#5555FF'>&</font> item<font face='Lucida Console'>)</font> <font color='#0000FF'>const</font> |
|
<b>{</b> <font color='#0000FF'>return</font> pos <font color='#5555FF'><</font> item.pos; <b>}</b> |
|
|
|
<font color='#0000FF'>private</font>: |
|
<font color='#0000FF'><u>size_t</u></font> pos; |
|
<b>}</b>; |
|
|
|
__device__ iterator <b><a name='begin'></a>begin</b><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font> <font color='#0000FF'>const</font> |
|
<b>{</b> |
|
<font color='#0000FF'>return</font> <font color='#BB00BB'>iterator</font><font face='Lucida Console'>(</font>ibegin<font color='#5555FF'>+</font>blockDim.y <font color='#5555FF'>*</font> blockIdx.y <font color='#5555FF'>+</font> threadIdx.y<font face='Lucida Console'>)</font>; |
|
<b>}</b> |
|
__device__ iterator <b><a name='end'></a>end</b><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font> <font color='#0000FF'>const</font> |
|
<b>{</b> |
|
<font color='#0000FF'>return</font> <font color='#BB00BB'>iterator</font><font face='Lucida Console'>(</font>iend<font face='Lucida Console'>)</font>; |
|
<b>}</b> |
|
<font color='#0000FF'>private</font>: |
|
|
|
<font color='#0000FF'><u>size_t</u></font> ibegin; |
|
<font color='#0000FF'><u>size_t</u></font> iend; |
|
<b>}</b>; |
|
|
|
<font color='#009900'>// ------------------------------------------------------------------------------------ |
|
</font> |
|
<b>}</b> |
|
<b>}</b> |
|
|
|
<font color='#0000FF'>#endif</font> <font color='#009900'>// __CUDACC__ |
|
</font> |
|
<font color='#009900'>// ---------------------------------------------------------------------------------------- |
|
</font> |
|
<font color='#0000FF'>#endif</font> <font color='#009900'>// DLIB_CUDA_UtILS_H_ |
|
</font> |
|
|
|
</pre></body></html> |
