<tr style="height: 56px;">
<td style="padding-left: 0.5em;">
<div id="projectname">Compute Library
-  <span id="projectnumber">17.12</span>
+  <span id="projectnumber">18.05</span>
</div>
</td>
</tr>
<li class="level2"><a href="#S4_7_3_memory_manager_function_support">Function support</a></li>
</ul>
</li>
+<li class="level1"><a href="#S4_8_opencl_tuner">OpenCL Tuner</a></li>
</ul>
</div>
<div class="textblock"><h1><a class="anchor" id="S4_1"></a>
<div class="fragment"><div class="line"><span class="comment">//Create a kernel object:</span></div><div class="line">MyKernel kernel;</div><div class="line"><span class="comment">// Initialize the kernel with the input/output and options you want to use:</span></div><div class="line">kernel.configure( input, output, option0, option1);</div><div class="line"><span class="comment">// Retrieve the execution window of the kernel:</span></div><div class="line"><span class="keyword">const</span> Window& max_window = kernel.window();</div><div class="line"><span class="comment">// Run the whole kernel in the current thread:</span></div><div class="line">kernel.run( max_window ); <span class="comment">// Run the kernel on the full window</span></div></div><!-- fragment --><h2><a class="anchor" id="S4_2_3"></a>
Multi-threading</h2>
<p>The previous section shows how to run a NEON / CPP kernel in the current thread, however if your system has several CPU cores, you will probably want the kernel to use several cores. Here is how this can be done:</p>
-<div class="fragment"><div class="line"> ThreadInfo <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a096668313a9a819d54a2e65ec21ff0cc">info</a>;</div><div class="line"> info.cpu_info = _info;</div><div class="line"></div><div class="line"> <span class="keyword">const</span> Window &max_window = kernel->window();</div><div class="line"> <span class="keyword">const</span> <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> num_iterations = max_window.num_iterations(split_dimension);</div><div class="line"> info.num_threads = <a class="code" href="namespacearm__compute_1_1test_1_1fixed__point__arithmetic_1_1detail.xhtml#aabcf39e3917f842dbc5fbb0d802f24d5">std::min</a>(num_iterations, _num_threads);</div><div class="line"></div><div class="line"> <span class="keywordflow">if</span>(num_iterations == 0)</div><div class="line"> {</div><div class="line"> <span class="keywordflow">return</span>;</div><div class="line"> }</div><div class="line"></div><div class="line"> <span class="keywordflow">if</span>(!kernel->is_parallelisable() || info.num_threads == 1)</div><div class="line"> {</div><div class="line"> kernel->run(max_window, info);</div><div class="line"> }</div><div class="line"> <span class="keywordflow">else</span></div><div class="line"> {</div><div class="line"> <span class="keywordtype">int</span> t = 0;</div><div class="line"> <span class="keyword">auto</span> thread_it = _threads.begin();</div><div class="line"></div><div class="line"> <span class="keywordflow">for</span>(; t < info.num_threads - 1; ++t, ++thread_it)</div><div class="line"> {</div><div class="line"> Window win = max_window.split_window(split_dimension, t, info.num_threads);</div><div class="line"> info.thread_id = t;</div><div class="line"> thread_it->start(kernel, win, info);</div><div class="line"> }</div><div class="line"></div><div class="line"> <span class="comment">// Run last part on main thread</span></div><div class="line"> Window win = max_window.split_window(split_dimension, t, info.num_threads);</div><div class="line"> info.thread_id = t;</div><div class="line"> kernel->run(win, info);</div><div class="line"></div><div class="line"> <span class="keywordflow">try</span></div><div class="line"> {</div><div class="line"> <span class="keywordflow">for</span>(<span class="keyword">auto</span> &thread : _threads)</div><div class="line"> {</div><div class="line"> thread.wait();</div><div class="line"> }</div><div class="line"> }</div><div class="line"> <span class="keywordflow">catch</span>(<span class="keyword">const</span> std::system_error &e)</div><div class="line"> {</div><div class="line"> std::cerr << <span class="stringliteral">"Caught system_error with code "</span> << e.code() << <span class="stringliteral">" meaning "</span> << e.what() << <span class="charliteral">'\n'</span>;</div><div class="line"> }</div><div class="line"> }</div></div><!-- fragment --><p> This is the very basic implementation used in the NEON runtime library by all the NEON functions.</p>
+<div class="fragment"><div class="line"> ThreadInfo <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a096668313a9a819d54a2e65ec21ff0cc">info</a>;</div><div class="line"> info.cpu_info =
&_cpu_info;</div><div class="line"></div><div class="line"> <span class="keyword">const</span> Window &max_window = kernel->window();</div><div class="line"> <span class="keyword">const</span> <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> num_iterations = max_window.num_iterations(split_dimension);</div><div class="line"> info.num_threads = <a class="code" href="namespacearm__compute_1_1test_1_1fixed__point__arithmetic_1_1detail.xhtml#aabcf39e3917f842dbc5fbb0d802f24d5">std::min</a>(num_iterations, _num_threads);</div><div class="line"></div><div class="line"> <span class="keywordflow">if</span>(num_iterations == 0)</div><div class="line"> {</div><div class="line"> <span class="keywordflow">return</span>;</div><div class="line"> }</div><div class="line"></div><div class="line"> <span class="keywordflow">if</span>(!kernel->is_parallelisable() || info.num_threads == 1)</div><div class="line"> {</div><div class="line"> kernel->run(max_window, info);</div><div class="line"> }</div><div class="line"> <span class="keywordflow">else</span></div><div class="line"> {</div><div class="line"> <span class="keywordtype">int</span> t = 0;</div><div class="line"> <span class="keyword">auto</span> thread_it = _threads.begin();</div><div class="line"></div><div class="line"> <span class="keywordflow">for</span>(; t < info.num_threads - 1; ++t, ++thread_it)</div><div class="line"> {</div><div class="line"> Window win = max_window.split_window(split_dimension, t, info.num_threads);</div><div class="line"> info.thread_id = t;</div><div class="line"> thread_it->start(kernel, win, info);</div><div class="line"> }</div><div class="line"></div><div class="line"> <span class="comment">// Run last part on main thread</span></div><div class="line"> Window win = max_window.split_window(split_dimension, t, info.num_threads);</div><div class="line"> info.thread_id = t;</div><div class="line"> kernel->run(win, info);</div><div class="line"></div><div class="line"> <span class="keywordflow">try</span></div><div class="line"> {</div><div class="line"> <span class="keywordflow">for</span>(<span class="keyword">auto</span> &thread : _threads)</div><div class="line"> {</div><div class="line"> thread.wait();</div><div class="line"> }</div><div class="line"> }</div><div class="line"> <span class="keywordflow">catch</span>(<span class="keyword">const</span> std::system_error &e)</div><div class="line"> {</div><div class="line"> std::cerr << <span class="stringliteral">"Caught system_error with code "</span> << e.code() << <span class="stringliteral">" meaning "</span> << e.what() << <span class="charliteral">'\n'</span>;</div><div class="line"> }</div><div class="line"> }</div></div><!-- fragment --><p> This is the very basic implementation used in the NEON runtime library by all the NEON functions.</p>
<dl class="section see"><dt>See also</dt><dd><a class="el" href="classarm__compute_1_1_c_p_p_scheduler.xhtml" title="C++11 implementation of a pool of threads to automatically split a kernel's execution among several t...">CPPScheduler</a>.</dd></dl>
<dl class="section note"><dt>Note</dt><dd>Some kernels like for example <a class="el" href="classarm__compute_1_1_n_e_histogram_kernel.xhtml">NEHistogramKernel</a> need some local temporary buffer to perform their calculations. In order to avoid memory corruption between threads, the local buffer must be of size: <code>memory_needed_per_thread * num_threads</code> and a unique thread_id between 0 and num_threads must be assigned to the <a class="el" href="structarm__compute_1_1_thread_info.xhtml">ThreadInfo</a> object passed to the <code>run</code> function.</dd></dl>
<h2><a class="anchor" id="S4_2_4"></a>
<h2><a class="anchor" id="S4_4_2_events_sync"></a>
OpenCL events and synchronization</h2>
<p>In order to block until all the jobs in the <a class="el" href="classarm__compute_1_1_c_l_scheduler.xhtml" title="Provides global access to a CL context and command queue. ">CLScheduler</a>'s command queue are done executing the user can call <a class="el" href="classarm__compute_1_1_c_l_scheduler.xhtml#ad55f80ed3cd8b6c4f247763b747016af">CLScheduler::sync()</a> or create a sync event using <a class="el" href="classarm__compute_1_1_c_l_scheduler.xhtml#a6096d26e8e29e1a7f01b0f083fb7e33f">CLScheduler::enqueue_sync_event()</a></p>
-<p>For example: </p><div class="fragment"><div class="line">
PPMLoader ppm;</div><div class="line"> <a class="code" href="namespacearm__compute.xhtml#aae712f442eae2d56d17529ca6cb9cb77">CLImage</a> src, tmp_scale_median, tmp_median_gauss, <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#adbf67dcee294e673cf796f1ed8aeb6a4">dst</a>;</div><div class="line"> constexpr <span class="keywordtype">int</span> scale_factor = 2;</div><div class="line"></div><div class="line"> <a class="code" href="classarm__compute_1_1_c_l_scheduler.xhtml#a60f9a6836b628a7171914c4afe43b4a7">CLScheduler::get</a>().<a class="code" href="classarm__compute_1_1_c_l_scheduler.xhtml#a46ecf9ef0fe80ba2ed35acfc29856b7d">default_init</a>();</div><div class="line"></div><div class="line"> <span class="keywordflow">if</span>(argc < 2)</div><div class="line"> {</div><div class="line"> <span class="comment">// Print help</span></div><div class="line"> std::cout << <span class="stringliteral">"Usage: ./build/cl_events [input_image.ppm]\n\n"</span>;</div><div class="line"> std::cout << <span class="stringliteral">"No input_image provided, creating a dummy 640x480 image\n"</span>;</div><div class="line"> <span class="comment">// Create an empty grayscale 640x480 image</span></div><div class="line"> src.allocator()->init(TensorInfo(640, 480, <a class="code" href="namespacearm__compute.xhtml#ab4e88c89b3b7ea1735996cc4def22d58a6669348b484e3008dca2bfa8e85e40b5">Format::U8</a>));</div><div class="line"> }</div><div class="line"> <span class="keywordflow">else</span></div><div class="line"> {</div><div class="line"> ppm.open(argv[1]);</div><div class="line"> ppm.init_image(src, <a class="code" href="namespacearm__compute.xhtml#ab4e88c89b3b7ea1735996cc4def22d58a6669348b484e3008dca2bfa8e85e40b5">Format::U8</a>);</div><div class="line"> }</div><div class="line"></div><div class="line"> <span class="comment">// Declare and configure the functions to create the following pipeline: scale -> median -> gauss</span></div><div class="line"> CLScale <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#acec6d8ad52a28972fa74e071c1a63b6a">scale</a>;</div><div class="line"> CLMedian3x3 median;</div><div class="line"> CLGaussian5x5 gauss;</div><div class="line"></div><div class="line"> TensorInfo dst_info(src.info()->dimension(0) / scale_factor, src.info()->dimension(1) / scale_factor, <a class="code" href="namespacearm__compute.xhtml#ab4e88c89b3b7ea1735996cc4def22d58a6669348b484e3008dca2bfa8e85e40b5">Format::U8</a>);</div><div class="line"></div><div class="line"> <span class="comment">// Configure the temporary and destination images</span></div><div class="line"> dst.allocator()->init(dst_info);</div><div class="line"> tmp_scale_median.allocator()->init(dst_info);</div><div class="line"> tmp_median_gauss.allocator()->init(dst_info);</div><div class="line"></div><div class="line"> <span class="comment">//Configure the functions:</span></div><div class="line"> scale.configure(&src, &tmp_scale_median, <a class="code" href="namespacearm__compute.xhtml#a966a9c417ce5e94dca08d9b5e745c0c9a7f5ccbc3d30c2cd3fd04d567946cbde2">InterpolationPolicy::NEAREST_NEIGHBOR</a>, <a class="code" href="namespacearm__compute.xhtml#a15a05537a472ee742404821851529327a4ef59320fbe90fe47d40f1f71e4c5daa">BorderMode::REPLICATE</a>);</div><div class="line"> median.configure(&tmp_scale_median, &tmp_median_gauss, <a class="code" href="namespacearm__compute.xhtml#a15a05537a472ee742404821851529327a4ef59320fbe90fe47d40f1f71e4c5daa">BorderMode::REPLICATE</a>);</div><div class="line"> gauss.configure(&tmp_median_gauss, &dst, <a class="code" href="namespacearm__compute.xhtml#a15a05537a472ee742404821851529327a4ef59320fbe90fe47d40f1f71e4c5daa">BorderMode::REPLICATE</a>);</div><div class="line"></div><div class="line"> <span class="comment">// Allocate all the images</span></div><div class="line"> src.allocator()->allocate();</div><div class="line"> dst.allocator()->allocate();</div><div class="line"> tmp_scale_median.allocator()->allocate();</div><div class="line"> tmp_median_gauss.allocator()->allocate();</div><div class="line"> <span class="comment">// Fill the input image with the content of the PPM image if a filename was provided:</span></div><div class="line"> <span class="keywordflow">if</span>(ppm.is_open())</div><div class="line"> {</div><div class="line"> ppm.fill_image(src);</div><div class="line"> }</div><div class="line"></div><div class="line"> <span class="comment">// Enqueue and flush the scale OpenCL kernel:</span></div><div class="line"> scale.run();</div><div class="line"> <span class="comment">// Create a synchronisation event between scale and median:</span></div><div class="line"> cl::Event scale_event = <a class="code" href="classarm__compute_1_1_c_l_scheduler.xhtml#a60f9a6836b628a7171914c4afe43b4a7">CLScheduler::get</a>().<a class="code" href="classarm__compute_1_1_c_l_scheduler.xhtml#a6096d26e8e29e1a7f01b0f083fb7e33f">enqueue_sync_event</a>();</div><div class="line"> <span class="comment">// Enqueue and flush the median OpenCL kernel:</span></div><div class="line"> median.run();</div><div class="line"> <span class="comment">// Enqueue and flush the Gaussian OpenCL kernel:</span></div><div class="line"> gauss.run();</div><div class="line"></div><div class="line"> <span class="comment">//Make sure all the OpenCL jobs are done executing:</span></div><div class="line"> scale_event.wait(); <span class="comment">// Block until Scale is done executing (Median3x3 and Gaussian5x5 might still be running)</span></div><div class="line"> <a class="code" href="classarm__compute_1_1_c_l_scheduler.xhtml#a60f9a6836b628a7171914c4afe43b4a7">CLScheduler::get</a>().<a class="code" href="classarm__compute_1_1_c_l_scheduler.xhtml#ad55f80ed3cd8b6c4f247763b747016af">sync</a>(); <span class="comment">// Block until Gaussian5x5 is done executing</span></div><div class="line"></div><div class="line"> <span class="comment">// Save the result to file:</span></div><div class="line"> <span class="keywordflow">if</span>(ppm.is_open())</div><div class="line"> {</div><div class="line"> <span class="keyword">const</span> std::string output_filename = std::string(argv[1]) + <span class="stringliteral">"_out.ppm"</span>;</div><div class="line"> <a class="code" href="namespacearm__compute_1_1utils.xhtml#a301d0b7bfd70f73fc1924f4281938d08">save_to_ppm</a>(dst, output_filename); <span class="comment">// save_to_ppm maps and unmaps the image to store as PPM</span></div><div class="line"> }</div></div><!-- fragment --> <h2><a class="anchor" id="S4_4_2_cl_neon"></a>
+<p>For example: </p><div class="fragment"><div class="line">
PPMLoader ppm;</div><div class="line"> constexpr <span class="keywordtype">int</span> scale_factor = 2;</div><div class="line"></div><div class="line"> <a class="code" href="classarm__compute_1_1_c_l_scheduler.xhtml#a60f9a6836b628a7171914c4afe43b4a7">CLScheduler::get</a>().<a class="code" href="classarm__compute_1_1_c_l_scheduler.xhtml#a46ecf9ef0fe80ba2ed35acfc29856b7d">default_init</a>();</div><div class="line"></div><div class="line"> <span class="keywordflow">if</span>(argc < 2)</div><div class="line"> {</div><div class="line"> <span class="comment">// Print help</span></div><div class="line"> std::cout << <span class="stringliteral">"Usage: ./build/cl_events [input_image.ppm]\n\n"</span>;</div><div class="line"> std::cout << <span class="stringliteral">"No input_image provided, creating a dummy 640x480 image\n"</span>;</div><div class="line"> <span class="comment">// Create an empty grayscale 640x480 image</span></div><div class="line"> <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a6743f0a130e8311e6f5b1a23df102472">src</a>.allocator()->init(TensorInfo(640, 480, <a class="code" href="namespacearm__compute.xhtml#ab4e88c89b3b7ea1735996cc4def22d58a6669348b484e3008dca2bfa8e85e40b5">Format::U8</a>));</div><div class="line"> }</div><div class="line"> <span class="keywordflow">else</span></div><div class="line"> {</div><div class="line"> ppm.open(argv[1]);</div><div class="line"> ppm.init_image(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a6743f0a130e8311e6f5b1a23df102472">src</a>, <a class="code" href="namespacearm__compute.xhtml#ab4e88c89b3b7ea1735996cc4def22d58a6669348b484e3008dca2bfa8e85e40b5">Format::U8</a>);</div><div class="line"> }</div><div class="line"></div><div class="line"> TensorInfo dst_info(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a6743f0a130e8311e6f5b1a23df102472">src</a>.info()->dimension(0) / scale_factor, <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a6743f0a130e8311e6f5b1a23df102472">src</a>.info()->dimension(1) / scale_factor, <a class="code" href="namespacearm__compute.xhtml#ab4e88c89b3b7ea1735996cc4def22d58a6669348b484e3008dca2bfa8e85e40b5">Format::U8</a>);</div><div class="line"></div><div class="line"> <span class="comment">// Configure the temporary and destination images</span></div><div class="line"> <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#adbf67dcee294e673cf796f1ed8aeb6a4">dst</a>.<a class="code" href="classarm__compute_1_1_c_l_tensor.xhtml#a256b18d4e6fdbbff14937b4b9089bdd3">allocator</a>()-><a class="code" href="classarm__compute_1_1_i_tensor_allocator.xhtml#aa72161e0e3c0f6b2da20f835de6af680">init</a>(dst_info);</div><div class="line"> tmp_scale_median.allocator()->init(dst_info);</div><div class="line"> tmp_median_gauss.allocator()->init(dst_info);</div><div class="line"></div><div class="line"> <span class="comment">//Configure the functions:</span></div><div class="line"> <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#acec6d8ad52a28972fa74e071c1a63b6a">scale</a>.configure(&<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a6743f0a130e8311e6f5b1a23df102472">src</a>, &tmp_scale_median, <a class="code" href="namespacearm__compute.xhtml#a966a9c417ce5e94dca08d9b5e745c0c9a7f5ccbc3d30c2cd3fd04d567946cbde2">InterpolationPolicy::NEAREST_NEIGHBOR</a>, <a class="code" href="namespacearm__compute.xhtml#a15a05537a472ee742404821851529327a4ef59320fbe90fe47d40f1f71e4c5daa">BorderMode::REPLICATE</a>);</div><div class="line"> median.configure(&tmp_scale_median, &tmp_median_gauss, <a class="code" href="namespacearm__compute.xhtml#a15a05537a472ee742404821851529327a4ef59320fbe90fe47d40f1f71e4c5daa">BorderMode::REPLICATE</a>);</div><div class="line"> gauss.configure(&tmp_median_gauss, &<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#adbf67dcee294e673cf796f1ed8aeb6a4">dst</a>, <a class="code" href="namespacearm__compute.xhtml#a15a05537a472ee742404821851529327a4ef59320fbe90fe47d40f1f71e4c5daa">BorderMode::REPLICATE</a>);</div><div class="line"></div><div class="line"> <span class="comment">// Allocate all the images</span></div><div class="line"> <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a6743f0a130e8311e6f5b1a23df102472">src</a>.allocator()->allocate();</div><div class="line"> <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#adbf67dcee294e673cf796f1ed8aeb6a4">dst</a>.<a class="code" href="classarm__compute_1_1_c_l_tensor.xhtml#a256b18d4e6fdbbff14937b4b9089bdd3">allocator</a>()-><a class="code" href="classarm__compute_1_1_c_l_tensor_allocator.xhtml#a6e509c2a177b0b29e9e2369535094dee">allocate</a>();</div><div class="line"> tmp_scale_median.allocator()->allocate();</div><div class="line"> tmp_median_gauss.allocator()->allocate();</div><div class="line"></div><div class="line"> <span class="comment">// Fill the input image with the content of the PPM image if a filename was provided:</span></div><div class="line"> <span class="keywordflow">if</span>(ppm.is_open())</div><div class="line"> {</div><div class="line"> ppm.fill_image(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a6743f0a130e8311e6f5b1a23df102472">src</a>);</div><div class="line"> output_filename = std::string(argv[1]) + <span class="stringliteral">"_out.ppm"</span>;</div><div class="line"> }</div></div><!-- fragment --> <h2><a class="anchor" id="S4_4_2_cl_neon"></a>
OpenCL / NEON interoperability</h2>
<p>You can mix OpenCL and NEON kernels and functions. However it is the user's responsibility to handle the mapping/unmapping of OpenCL objects, for example:</p>
-<div class="fragment"><div class="line"> PPMLoader ppm;</div><div class="line"> <a class="code" href="namespacearm__compute.xhtml#aae712f442eae2d56d17529ca6cb9cb77">CLImage</a> src, scale_median, median_gauss, <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#adbf67dcee294e673cf796f1ed8aeb6a4">dst</a>;</div><div class="line"></div><div class="line"> <a class="code" href="classarm__compute_1_1_c_l_scheduler.xhtml#a60f9a6836b628a7171914c4afe43b4a7">CLScheduler::get</a>().<a class="code" href="classarm__compute_1_1_c_l_scheduler.xhtml#a46ecf9ef0fe80ba2ed35acfc29856b7d">default_init</a>();</div><div class="line"></div><div class="line"> <span class="keywordflow">if</span>(argc < 2)</div><div class="line"> {</div><div class="line"> <span class="comment">// Print help</span></div><div class="line"> std::cout << <span class="stringliteral">"Usage: ./build/cl_convolution [input_image.ppm]\n\n"</span>;</div><div class="line"> std::cout << <span class="stringliteral">"No input_image provided, creating a dummy 640x480 image\n"</span>;</div><div class="line"> <span class="comment">// Create an empty grayscale 640x480 image</span></div><div class="line"> src.allocator()->init(TensorInfo(640, 480, <a class="code" href="namespacearm__compute.xhtml#ab4e88c89b3b7ea1735996cc4def22d58a6669348b484e3008dca2bfa8e85e40b5">Format::U8</a>));</div><div class="line"> }</div><div class="line"> <span class="keywordflow">else</span></div><div class="line"> {</div><div class="line"> ppm.open(argv[1]);</div><div class="line"> ppm.init_image(src, <a class="code" href="namespacearm__compute.xhtml#ab4e88c89b3b7ea1735996cc4def22d58a6669348b484e3008dca2bfa8e85e40b5">Format::U8</a>);</div><div class="line"> }</div><div class="line"></div><div class="line"> TensorInfo scale_median_info(TensorInfo(src.info()->dimension(0) / 2, src.info()->dimension(1) / 2, <a class="code" href="namespacearm__compute.xhtml#ab4e88c89b3b7ea1735996cc4def22d58a6669348b484e3008dca2bfa8e85e40b5">Format::U8</a>));</div><div class="line"></div><div class="line"> <span class="comment">// Configure the temporary and destination images</span></div><div class="line"> scale_median.allocator()->init(scale_median_info);</div><div class="line"> median_gauss.allocator()->init(scale_median_info);</div><div class="line"> dst.allocator()->init(scale_median_info);</div><div class="line"></div><div class="line"> <span class="comment">// Declare and configure the functions to create the following pipeline: scale -> median -> gauss</span></div><div class="line"> CLScale <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#acec6d8ad52a28972fa74e071c1a63b6a">scale</a>;</div><div class="line"> NEMedian3x3 median;</div><div class="line"> CLGaussian5x5 gauss;</div><div class="line"></div><div class="line"> scale.configure(&src, &scale_median, <a class="code" href="namespacearm__compute.xhtml#a966a9c417ce5e94dca08d9b5e745c0c9a7f5ccbc3d30c2cd3fd04d567946cbde2">InterpolationPolicy::NEAREST_NEIGHBOR</a>, <a class="code" href="namespacearm__compute.xhtml#a15a05537a472ee742404821851529327a4ef59320fbe90fe47d40f1f71e4c5daa">BorderMode::REPLICATE</a>);</div><div class="line"> median.configure(&scale_median, &median_gauss, <a class="code" href="namespacearm__compute.xhtml#a15a05537a472ee742404821851529327a4ef59320fbe90fe47d40f1f71e4c5daa">BorderMode::REPLICATE</a>);</div><div class="line"> gauss.configure(&median_gauss, &dst, <a class="code" href="namespacearm__compute.xhtml#a15a05537a472ee742404821851529327a4ef59320fbe90fe47d40f1f71e4c5daa">BorderMode::REPLICATE</a>);</div><div class="line"></div><div class="line"> <span class="comment">// Allocate all the images</span></div><div class="line"> src.allocator()->allocate();</div><div class="line"> scale_median.allocator()->allocate();</div><div class="line"> median_gauss.allocator()->allocate();</div><div class="line"> dst.allocator()->allocate();</div><div class="line"></div><div class="line"> <span class="comment">// Fill the input image with the content of the PPM image if a filename was provided:</span></div><div class="line"> <span class="keywordflow">if</span>(ppm.is_open())</div><div class="line"> {</div><div class="line"> ppm.fill_image(src);</div><div class="line"> }</div><div class="line"></div><div class="line"> <span class="comment">// Enqueue and flush the OpenCL kernel:</span></div><div class="line"> scale.run();</div><div class="line"></div><div class="line"> <span class="comment">// Do a blocking map of the input and output buffers of the NEON function:</span></div><div class="line"> scale_median.map();</div><div class="line"> median_gauss.map();</div><div class="line"></div><div class="line"> <span class="comment">// Run the NEON function:</span></div><div class="line"> median.run();</div><div class="line"></div><div class="line"> <span class="comment">// Unmap the output buffer before it's used again by OpenCL:</span></div><div class="line"> scale_median.unmap();</div><div class="line"> median_gauss.unmap();</div><div class="line"></div><div class="line"> <span class="comment">// Run the final OpenCL function:</span></div><div class="line"> gauss.run();</div><div class="line"></div><div class="line"> <span class="comment">// Make sure all the OpenCL jobs are done executing:</span></div><div class="line"> <a class="code" href="classarm__compute_1_1_c_l_scheduler.xhtml#a60f9a6836b628a7171914c4afe43b4a7">CLScheduler::get</a>().<a class="code" href="classarm__compute_1_1_c_l_scheduler.xhtml#ad55f80ed3cd8b6c4f247763b747016af">sync</a>();</div><div class="line"></div><div class="line"> <span class="comment">// Save the result to file:</span></div><div class="line"> <span class="keywordflow">if</span>(ppm.is_open())</div><div class="line"> {</div><div class="line"> <span class="keyword">const</span> std::string output_filename = std::string(argv[1]) + <span class="stringliteral">"_out.ppm"</span>;</div><div class="line"> <a class="code" href="namespacearm__compute_1_1utils.xhtml#a301d0b7bfd70f73fc1924f4281938d08">save_to_ppm</a>(dst, output_filename); <span class="comment">// save_to_ppm maps and unmaps the image to store as PPM</span></div><div class="line"> }</div></div><!-- fragment --> <dl class="section see"><dt>See also</dt><dd><a class="el" href="neoncl__scale__median__gaussian_8cpp.xhtml#a4003cb8b626a6604e2f51b8e17f8bb3d" title="Example demonstrating how to use both CL and NEON functions in the same pipeline. ...">main_neoncl_scale_median_gaussian</a></dd></dl>
+<div class="fragment"><div class="line"> PPMLoader ppm;</div><div class="line"></div><div class="line"> <a class="code" href="classarm__compute_1_1_c_l_scheduler.xhtml#a60f9a6836b628a7171914c4afe43b4a7">CLScheduler::get</a>().<a class="code" href="classarm__compute_1_1_c_l_scheduler.xhtml#a46ecf9ef0fe80ba2ed35acfc29856b7d">default_init</a>();</div><div class="line"></div><div class="line"> <span class="keywordflow">if</span>(argc < 2)</div><div class="line"> {</div><div class="line"> <span class="comment">// Print help</span></div><div class="line"> std::cout << <span class="stringliteral">"Usage: ./build/cl_convolution [input_image.ppm]\n\n"</span>;</div><div class="line"> std::cout << <span class="stringliteral">"No input_image provided, creating a dummy 640x480 image\n"</span>;</div><div class="line"> <span class="comment">// Create an empty grayscale 640x480 image</span></div><div class="line"> <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a6743f0a130e8311e6f5b1a23df102472">src</a>.allocator()->init(TensorInfo(640, 480, <a class="code" href="namespacearm__compute.xhtml#ab4e88c89b3b7ea1735996cc4def22d58a6669348b484e3008dca2bfa8e85e40b5">Format::U8</a>));</div><div class="line"> }</div><div class="line"> <span class="keywordflow">else</span></div><div class="line"> {</div><div class="line"> ppm.open(argv[1]);</div><div class="line"> ppm.init_image(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a6743f0a130e8311e6f5b1a23df102472">src</a>, <a class="code" href="namespacearm__compute.xhtml#ab4e88c89b3b7ea1735996cc4def22d58a6669348b484e3008dca2bfa8e85e40b5">Format::U8</a>);</div><div class="line"> }</div><div class="line"></div><div class="line"> TensorInfo scale_median_info(TensorInfo(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a6743f0a130e8311e6f5b1a23df102472">src</a>.info()->dimension(0) / 2, <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a6743f0a130e8311e6f5b1a23df102472">src</a>.info()->dimension(1) / 2, <a class="code" href="namespacearm__compute.xhtml#ab4e88c89b3b7ea1735996cc4def22d58a6669348b484e3008dca2bfa8e85e40b5">Format::U8</a>));</div><div class="line"></div><div class="line"> <span class="comment">// Configure the temporary and destination images</span></div><div class="line"> scale_median.allocator()->init(scale_median_info);</div><div class="line"> median_gauss.allocator()->init(scale_median_info);</div><div class="line"> <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#adbf67dcee294e673cf796f1ed8aeb6a4">dst</a>.<a class="code" href="classarm__compute_1_1_c_l_tensor.xhtml#a256b18d4e6fdbbff14937b4b9089bdd3">allocator</a>()-><a class="code" href="classarm__compute_1_1_i_tensor_allocator.xhtml#aa72161e0e3c0f6b2da20f835de6af680">init</a>(scale_median_info);</div><div class="line"></div><div class="line"> <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#acec6d8ad52a28972fa74e071c1a63b6a">scale</a>.configure(&<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a6743f0a130e8311e6f5b1a23df102472">src</a>, &scale_median, <a class="code" href="namespacearm__compute.xhtml#a966a9c417ce5e94dca08d9b5e745c0c9a7f5ccbc3d30c2cd3fd04d567946cbde2">InterpolationPolicy::NEAREST_NEIGHBOR</a>, <a class="code" href="namespacearm__compute.xhtml#a15a05537a472ee742404821851529327a4ef59320fbe90fe47d40f1f71e4c5daa">BorderMode::REPLICATE</a>);</div><div class="line"> median.configure(&scale_median, &median_gauss, <a class="code" href="namespacearm__compute.xhtml#a15a05537a472ee742404821851529327a4ef59320fbe90fe47d40f1f71e4c5daa">BorderMode::REPLICATE</a>);</div><div class="line"> gauss.configure(&median_gauss, &<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#adbf67dcee294e673cf796f1ed8aeb6a4">dst</a>, <a class="code" href="namespacearm__compute.xhtml#a15a05537a472ee742404821851529327a4ef59320fbe90fe47d40f1f71e4c5daa">BorderMode::REPLICATE</a>);</div><div class="line"></div><div class="line"> <span class="comment">// Allocate all the images</span></div><div class="line"> <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a6743f0a130e8311e6f5b1a23df102472">src</a>.allocator()->allocate();</div><div class="line"> scale_median.allocator()->allocate();</div><div class="line"> median_gauss.allocator()->allocate();</div><div class="line"> <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#adbf67dcee294e673cf796f1ed8aeb6a4">dst</a>.<a class="code" href="classarm__compute_1_1_c_l_tensor.xhtml#a256b18d4e6fdbbff14937b4b9089bdd3">allocator</a>()-><a class="code" href="classarm__compute_1_1_c_l_tensor_allocator.xhtml#a6e509c2a177b0b29e9e2369535094dee">allocate</a>();</div><div class="line"></div><div class="line"> <span class="comment">// Fill the input image with the content of the PPM image if a filename was provided:</span></div><div class="line"> <span class="keywordflow">if</span>(ppm.is_open())</div><div class="line"> {</div><div class="line"> ppm.fill_image(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a6743f0a130e8311e6f5b1a23df102472">src</a>);</div><div class="line"> <span class="keyword">const</span> std::string output_filename = std::string(argv[1]) + <span class="stringliteral">"_out.ppm"</span>;</div><div class="line"> }</div></div><!-- fragment --> <dl class="section see"><dt>See also</dt><dd>main_neoncl_scale_median_gaussian</dd></dl>
<h1><a class="anchor" id="S4_5_algorithms"></a>
Algorithms</h1>
<p>All computer vision algorithms in this library have been implemented following the <a href="https://www.khronos.org/registry/vx/specs/1.1/html/">OpenVX 1.1 specifications</a>. Please refer to the Khronos documentation for more information.</p>
<ul>
<li>Accurate padding:</li>
</ul>
-<div class="fragment"><div class="line">
PPMLoader ppm;</div><div class="line"> <a class="code" href="struct_image.xhtml">Image</a> src, tmp, <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#adbf67dcee294e673cf796f1ed8aeb6a4">dst</a>;</div><div class="line"></div><div class="line"> <span class="keywordflow">if</span>(argc < 2)</div><div class="line"> {</div><div class="line"> <span class="comment">// Print help</span></div><div class="line"> std::cout << <span class="stringliteral">"Usage: ./build/neon_convolution [input_image.ppm]\n\n"</span>;</div><div class="line"> std::cout << <span class="stringliteral">"No input_image provided, creating a dummy 640x480 image\n"</span>;</div><div class="line"> <span class="comment">// Initialize just the dimensions and format of your buffers:</span></div><div class="line"> src.allocator()->init(TensorInfo(640, 480, <a class="code" href="namespacearm__compute.xhtml#ab4e88c89b3b7ea1735996cc4def22d58a6669348b484e3008dca2bfa8e85e40b5">Format::U8</a>));</div><div class="line"> }</div><div class="line"> <span class="keywordflow">else</span></div><div class="line"> {</div><div class="line"> ppm.open(argv[1]);</div><div class="line"> <span class="comment">// Initialize just the dimensions and format of your buffers:</span></div><div class="line"> ppm.init_image(src, <a class="code" href="namespacearm__compute.xhtml#ab4e88c89b3b7ea1735996cc4def22d58a6669348b484e3008dca2bfa8e85e40b5">Format::U8</a>);</div><div class="line"> }</div><div class="line"></div><div class="line"> <span class="comment">// Initialize just the dimensions and format of the temporary and destination images:</span></div><div class="line"> tmp.allocator()->init(*src.info());</div><div class="line"> dst.allocator()->init(*src.info());</div><div class="line"></div><div class="line"> NEConvolution3x3 conv3x3;</div><div class="line"> <a class="code" href="namespacearm__compute.xhtml#adbc7771d367ba8f51da1450d3602e5c0">NEConvolution5x5</a> conv5x5;</div><div class="line"></div><div class="line"> <span class="comment">// Apply a Gaussian 3x3 filter to the source image followed by a Gaussian 5x5:</span></div><div class="line"> <span class="comment">// The function will automatically update the padding information inside input and output to match its requirements</span></div><div class="line"> conv3x3.configure(&src, &tmp, <a class="code" href="cl__convolution_8cpp.xhtml#a741ba5321da40184f8653e0a50ace070">gaussian3x3</a>, 0 <span class="comment">/* Let arm_compute calculate the scale */</span>, <a class="code" href="namespacearm__compute.xhtml#a15a05537a472ee742404821851529327a0db45d2a4141101bdfe48e3314cfbca3">BorderMode::UNDEFINED</a>);</div><div class="line"> conv5x5.configure(&tmp, &dst, <a class="code" href="cl__convolution_8cpp.xhtml#a565013cf7e49a591bacd548571951f94">gaussian5x5</a>, 0 <span class="comment">/* Let arm_compute calculate the scale */</span>, <a class="code" href="namespacearm__compute.xhtml#a15a05537a472ee742404821851529327a0db45d2a4141101bdfe48e3314cfbca3">BorderMode::UNDEFINED</a>);</div><div class="line"></div><div class="line"> <span class="comment">// Now that the padding requirements are known we can allocate the images:</span></div><div class="line"> src.allocator()->allocate();</div><div class="line"> tmp.allocator()->allocate();</div><div class="line"> dst.allocator()->allocate();</div><div class="line"></div><div class="line"> <span class="comment">// Fill the input image with the content of the PPM image if a filename was provided:</span></div><div class="line"> <span class="keywordflow">if</span>(ppm.is_open())</div><div class="line"> {</div><div class="line"> ppm.fill_image(src);</div><div class="line"> }</div><div class="line"></div><div class="line"> <span class="comment">//Execute the functions:</span></div><div class="line"> conv3x3.run();</div><div class="line"> conv5x5.run();</div><div class="line"></div><div class="line"> <span class="comment">// Save the result to file:</span></div><div class="line"> <span class="keywordflow">if</span>(ppm.is_open())</div><div class="line"> {</div><div class="line"> <span class="keyword">const</span> std::string output_filename = std::string(argv[1]) + <span class="stringliteral">"_out.ppm"</span>;</div><div class="line"> <a class="code" href="namespacearm__compute_1_1utils.xhtml#a301d0b7bfd70f73fc1924f4281938d08">save_to_ppm</a>(dst, output_filename);</div><div class="line"> }</div></div><!-- fragment --> <dl class="section note"><dt>Note</dt><dd>It's important to call allocate <b>after</b> the function is configured: if the image / tensor is already allocated then the function will shrink its execution window instead of increasing the padding. (See below for more details).</dd></dl>
+<div class="fragment"><div class="line">
PPMLoader ppm;</div><div class="line"></div><div class="line"> <span class="keywordflow">if</span>(argc < 2)</div><div class="line"> {</div><div class="line"> <span class="comment">// Print help</span></div><div class="line"> std::cout << <span class="stringliteral">"Usage: ./build/neon_convolution [input_image.ppm]\n\n"</span>;</div><div class="line"> std::cout << <span class="stringliteral">"No input_image provided, creating a dummy 640x480 image\n"</span>;</div><div class="line"> <span class="comment">// Initialize just the dimensions and format of your buffers:</span></div><div class="line"> <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a6743f0a130e8311e6f5b1a23df102472">src</a>.allocator()->init(TensorInfo(640, 480, <a class="code" href="namespacearm__compute.xhtml#ab4e88c89b3b7ea1735996cc4def22d58a6669348b484e3008dca2bfa8e85e40b5">Format::U8</a>));</div><div class="line"> }</div><div class="line"> <span class="keywordflow">else</span></div><div class="line"> {</div><div class="line"> ppm.open(argv[1]);</div><div class="line"> <span class="comment">// Initialize just the dimensions and format of your buffers:</span></div><div class="line"> ppm.init_image(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a6743f0a130e8311e6f5b1a23df102472">src</a>, <a class="code" href="namespacearm__compute.xhtml#ab4e88c89b3b7ea1735996cc4def22d58a6669348b484e3008dca2bfa8e85e40b5">Format::U8</a>);</div><div class="line"> }</div><div class="line"></div><div class="line"> <span class="comment">// Initialize just the dimensions and format of the temporary and destination images:</span></div><div class="line"> tmp.allocator()->init(*<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a6743f0a130e8311e6f5b1a23df102472">src</a>.info());</div><div class="line"> <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#adbf67dcee294e673cf796f1ed8aeb6a4">dst</a>.<a class="code" href="classarm__compute_1_1_c_l_tensor.xhtml#a256b18d4e6fdbbff14937b4b9089bdd3">allocator</a>()-><a class="code" href="classarm__compute_1_1_i_tensor_allocator.xhtml#aa72161e0e3c0f6b2da20f835de6af680">init</a>(*<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a6743f0a130e8311e6f5b1a23df102472">src</a>.info());</div><div class="line"></div><div class="line"> <span class="comment">// Apply a Gaussian 3x3 filter to the source image followed by a Gaussian 5x5:</span></div><div class="line"> <span class="comment">// The function will automatically update the padding information inside input and output to match its requirements</span></div><div class="line"> conv3x3.configure(&<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a6743f0a130e8311e6f5b1a23df102472">src</a>, &tmp, <a class="code" href="cl__convolution_8cpp.xhtml#a741ba5321da40184f8653e0a50ace070">gaussian3x3</a>, 0 <span class="comment">/* Let arm_compute calculate the scale */</span>, <a class="code" href="namespacearm__compute.xhtml#a15a05537a472ee742404821851529327a0db45d2a4141101bdfe48e3314cfbca3">BorderMode::UNDEFINED</a>);</div><div class="line"> conv5x5.configure(&tmp, &<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#adbf67dcee294e673cf796f1ed8aeb6a4">dst</a>, <a class="code" href="cl__convolution_8cpp.xhtml#a565013cf7e49a591bacd548571951f94">gaussian5x5</a>, 0 <span class="comment">/* Let arm_compute calculate the scale */</span>, <a class="code" href="namespacearm__compute.xhtml#a15a05537a472ee742404821851529327a0db45d2a4141101bdfe48e3314cfbca3">BorderMode::UNDEFINED</a>);</div><div class="line"></div><div class="line"> <span class="comment">// Now that the padding requirements are known we can allocate the images:</span></div><div class="line"> <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a6743f0a130e8311e6f5b1a23df102472">src</a>.allocator()->allocate();</div><div class="line"> tmp.allocator()->allocate();</div><div class="line"> <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#adbf67dcee294e673cf796f1ed8aeb6a4">dst</a>.<a class="code" href="classarm__compute_1_1_c_l_tensor.xhtml#a256b18d4e6fdbbff14937b4b9089bdd3">allocator</a>()-><a class="code" href="classarm__compute_1_1_c_l_tensor_allocator.xhtml#a6e509c2a177b0b29e9e2369535094dee">allocate</a>();</div><div class="line"></div><div class="line"> <span class="comment">// Fill the input image with the content of the PPM image if a filename was provided:</span></div><div class="line"> <span class="keywordflow">if</span>(ppm.is_open())</div><div class="line"> {</div><div class="line"> ppm.fill_image(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a6743f0a130e8311e6f5b1a23df102472">src</a>);</div><div class="line"> output_filename = std::string(argv[1]) + <span class="stringliteral">"_out.ppm"</span>;</div><div class="line"> }</div></div><!-- fragment --> <dl class="section note"><dt>Note</dt><dd>It's important to call allocate <b>after</b> the function is configured: if the image / tensor is already allocated then the function will shrink its execution window instead of increasing the padding. (See below for more details).</dd></dl>
<ul>
<li>Manual padding / no padding / auto padding: You can allocate your images / tensors up front (before configuring your functions). In that case the function will use whatever padding is available and will shrink its execution window if there isn't enough padding available (which translates into a smaller valid region for the output). See also <a class="el" href="architecture.xhtml#valid_region">Valid regions</a>). If you don't want to manually set the padding but still want to allocate your objects upfront then you can use auto_padding. It guarantees that the allocation will have enough padding to run any of the provided functions.</li>
</ul>
-<div class="fragment"><div class="line"><a class="code" href="struct_image.xhtml">Image</a>
src, <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#adbf67dcee294e673cf796f1ed8aeb6a4">dst</a>;</div><div class="line"></div><div class="line"><span class="comment">// Use auto padding for the input:</span></div><div class="line">src.info()->init_auto_padding(TensorShape(640u,480u), <a class="code" href="namespacearm__compute.xhtml#ab4e88c89b3b7ea1735996cc4def22d58a6669348b484e3008dca2bfa8e85e40b5">Format::U8</a>);</div><div class="line"></div><div class="line"><span class="comment">// Use manual padding for the destination image</span></div><div class="line">dst.info()->init(src.info()->tensor_shape(), <a class="code" href="namespacearm__compute.xhtml#ab4e88c89b3b7ea1735996cc4def22d58a6669348b484e3008dca2bfa8e85e40b5">Format::U8</a>, strides_in_bytes, offset_first_element_in_bytes, total_size_in_bytes);</div><div class="line"></div><div class="line"><span class="comment">// Allocate all the images</span></div><div class="line">src.allocator()->allocate();</div><div class="line">dst.allocator()->allocate();</div><div class="line"><span class="comment">// Fill the input image with the content of the PPM image if a filename was provided:</span></div><div class="line">fill_image(src);</div><div class="line"></div><div class="line">NEGaussian3x3 gauss;</div><div class="line"></div><div class="line"><span class="comment">// Apply a Gaussian 3x3 filter to the source image (Note: if the padding provided is not enough then the execution window and valid region of the output will be shrunk)</span></div><div class="line">gauss.configure(&src, &dst, <a class="code" href="namespacearm__compute.xhtml#a15a05537a472ee742404821851529327a0db45d2a4141101bdfe48e3314cfbca3">BorderMode::UNDEFINED</a>);</div><div class="line"></div><div class="line"><span class="comment">//Execute the functions:</span></div><div class="line">gauss.run();</div></div><!-- fragment --><dl class="section warning"><dt>Warning</dt><dd>Some kernels need up to 3 neighbor values to calculate the value of a given pixel. Therefore, to be safe, we use a 4-pixel padding all around the image. In addition, some kernels read and write up to 32 pixels at the same time. To cover that case as well we add an extra 32 pixels of padding at the end of each row. As a result auto padded buffers waste a lot of memory and are less cache friendly. It is therefore recommended to use accurate padding or manual padding wherever possible.</dd></dl>
+<div class="fragment"><div class="line"><a class="code" href="struct_image.xhtml">Image</a>
<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a6743f0a130e8311e6f5b1a23df102472">src</a>, <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#adbf67dcee294e673cf796f1ed8aeb6a4">dst</a>;</div><div class="line"></div><div class="line"><span class="comment">// Use auto padding for the input:</span></div><div class="line">src.info()->init_auto_padding(TensorShape(640u,480u), <a class="code" href="namespacearm__compute.xhtml#ab4e88c89b3b7ea1735996cc4def22d58a6669348b484e3008dca2bfa8e85e40b5">Format::U8</a>);</div><div class="line"></div><div class="line"><span class="comment">// Use manual padding for the destination image</span></div><div class="line">dst.info()->init(src.info()->tensor_shape(), <a class="code" href="namespacearm__compute.xhtml#ab4e88c89b3b7ea1735996cc4def22d58a6669348b484e3008dca2bfa8e85e40b5">Format::U8</a>, strides_in_bytes, offset_first_element_in_bytes, total_size_in_bytes);</div><div class="line"></div><div class="line"><span class="comment">// Allocate all the images</span></div><div class="line">src.allocator()->allocate();</div><div class="line">dst.allocator()->allocate();</div><div class="line"><span class="comment">// Fill the input image with the content of the PPM image if a filename was provided:</span></div><div class="line">fill_image(src);</div><div class="line"></div><div class="line">NEGaussian3x3 gauss;</div><div class="line"></div><div class="line"><span class="comment">// Apply a Gaussian 3x3 filter to the source image (Note: if the padding provided is not enough then the execution window and valid region of the output will be shrunk)</span></div><div class="line">gauss.configure(&src, &dst, <a class="code" href="namespacearm__compute.xhtml#a15a05537a472ee742404821851529327a0db45d2a4141101bdfe48e3314cfbca3">BorderMode::UNDEFINED</a>);</div><div class="line"></div><div class="line"><span class="comment">//Execute the functions:</span></div><div class="line">gauss.run();</div></div><!-- fragment --><dl class="section warning"><dt>Warning</dt><dd>Some kernels need up to 3 neighbor values to calculate the value of a given pixel. Therefore, to be safe, we use a 4-pixel padding all around the image. In addition, some kernels read and write up to 32 pixels at the same time. To cover that case as well we add an extra 32 pixels of padding at the end of each row. As a result auto padded buffers waste a lot of memory and are less cache friendly. It is therefore recommended to use accurate padding or manual padding wherever possible.</dd></dl>
<h3><a class="anchor" id="valid_region"></a>
Valid regions</h3>
<p>Some kernels (like edge detectors for example) need to read values of neighboring pixels to calculate the value of a given pixel, it is therefore not possible to calculate the values of the pixels on the edges.</p>
Unless specified otherwise in the kernel's or function's documentation the number of channels for tensors is expected to be 1 (For images, the number of channels is inferred from the <a class="el" href="namespacearm__compute.xhtml#ab4e88c89b3b7ea1735996cc4def22d58">Format</a>).</dd></dl>
<dl class="section attention"><dt>Attention</dt><dd>Regardless of the <a class="el" href="namespacearm__compute.xhtml#ad8ed01ff3ff33333d8e19db4d2818bb6">DataType</a> used by a tensor the <a class="el" href="classarm__compute_1_1_i_tensor.xhtml#ab988210662dbd3bf32fd563c7dd1bdbf">ITensor::buffer()</a> method will always return a uint8_t pointer, and all the metadata in <a class="el" href="classarm__compute_1_1_tensor_info.xhtml">TensorInfo</a> will be expressed in bytes. It is the user's responsibility to cast the pointer to the correct type.</dd></dl>
<p>For example, to read the element located at the coordinates (x,y) of a float tensor:</p>
-<div class="fragment"><div class="line"><span class="keywordtype">float</span>
<a class="code" href="hwc_8hpp.xhtml#a0f61d63b009d0880a89c843bd50d8d76">value</a> = *<span class="keyword">reinterpret_cast<</span><span class="keywordtype">float</span>*<span class="keyword">></span>(input.buffer() + input.info()->offset_element_in_bytes(Coordinates(x,y)));</div></div><!-- fragment --><h2><a class="anchor" id="S4_6_4_working_with_objects"></a>
+<div class="fragment"><div class="line"><span class="keywordtype">float</span>
value = *<span class="keyword">reinterpret_cast<</span><span class="keywordtype">float</span>*<span class="keyword">></span>(input.buffer() + input.info()->offset_element_in_bytes(Coordinates(x,y)));</div></div><!-- fragment --><h2><a class="anchor" id="S4_6_4_working_with_objects"></a>
Working with Images and Tensors using iterators</h2>
<p>The library provides some iterators to access objects' data. Iterators are created by associating a data object (An image or a tensor for example) with an iteration window.</p>
<p>Iteration windows are defined by an array of dimensions, each of which consists of a start, end and step.</p>
<p>The <a class="el" href="namespacearm__compute.xhtml#a6c0dcc38187027dcb89cd9724bc5a823">execute_window_loop</a> function takes an execution window, a lambda function and one or more iterators. It will iterate through every element of the execution window and for each element it will update the iterators accordingly and call the lambda function.</p>
<p>Here are a couple of examples of how to use the iterators to fill / read tensors:</p>
-<div class="fragment"><div class="line">
constexpr <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> width = 4;</div><div class="line"> constexpr <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> height = 3;</div><div class="line"> constexpr <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> batch = 2;</div><div class="line"></div><div class="line"> <span class="keyword">auto</span> *src_data = <span class="keyword">new</span> <span class="keywordtype">float</span>[width * height * batch];</div><div class="line"> <span class="keyword">auto</span> *dst_data = <span class="keyword">new</span> <span class="keywordtype">float</span>[width * height * batch];</div><div class="line"></div><div class="line"> <span class="comment">// Fill src_data with dummy values:</span></div><div class="line"> <span class="keywordflow">for</span>(<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a7b8004eef325a40dd43eb80755610fff">b</a> = 0; <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a7b8004eef325a40dd43eb80755610fff">b</a> < batch; <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a7b8004eef325a40dd43eb80755610fff">b</a>++)</div><div class="line"> {</div><div class="line"> <span class="keywordflow">for</span>(<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> h = 0; h < height; h++)</div><div class="line"> {</div><div class="line"> <span class="keywordflow">for</span>(<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> w = 0; w < width; w++)</div><div class="line"> {</div><div class="line"> src_data[<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a7b8004eef325a40dd43eb80755610fff">b</a> * (width * height) + h * width + w] = static_cast<float>(100 * <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a7b8004eef325a40dd43eb80755610fff">b</a> + 10 * h + w);</div><div class="line"> }</div><div class="line"> }</div><div class="line"> }</div><div class="line"></div><div class="line"> Tensor input, output;</div><div class="line"> NESoftmaxLayer softmax;</div><div class="line"></div><div class="line"> <span class="comment">// Initialize the tensors dimensions and type:</span></div><div class="line"> <span class="keyword">const</span> TensorShape <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a45cde9abb508c62d67c3bb2b9bf566a5">shape</a>(width, height, batch);</div><div class="line"> input.allocator()->init(TensorInfo(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a45cde9abb508c62d67c3bb2b9bf566a5">shape</a>, 1, <a class="code" href="namespacearm__compute.xhtml#ab4e88c89b3b7ea1735996cc4def22d58a44ad4ef5a76e6aa6fb3e3fa079a54fda">DataType::F32</a>));</div><div class="line"> output.allocator()->init(TensorInfo(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a45cde9abb508c62d67c3bb2b9bf566a5">shape</a>, 1, <a class="code" href="namespacearm__compute.xhtml#ab4e88c89b3b7ea1735996cc4def22d58a44ad4ef5a76e6aa6fb3e3fa079a54fda">DataType::F32</a>));</div><div class="line"></div><div class="line"> <span class="comment">// Configure softmax:</span></div><div class="line"> softmax.configure(&input, &output);</div><div class="line"></div><div class="line"> <span class="comment">// Allocate the input / output tensors:</span></div><div class="line"> input.allocator()->allocate();</div><div class="line"> output.allocator()->allocate();</div><div class="line"></div><div class="line"> <span class="comment">// Fill the input tensor:</span></div><div class="line"> <span class="comment">// Simplest way: create an iterator to iterate through each element of the input tensor:</span></div><div class="line"> Window input_window;</div><div class="line"> input_window.use_tensor_dimensions(input.info()->tensor_shape());</div><div class="line"> std::cout << <span class="stringliteral">" Dimensions of the input's iterator:\n"</span>;</div><div class="line"> std::cout << <span class="stringliteral">" X = [start="</span> << input_window.x().start() << <span class="stringliteral">", end="</span> << input_window.x().end() << <span class="stringliteral">", step="</span> << input_window.x().step() << <span class="stringliteral">"]\n"</span>;</div><div class="line"> std::cout << <span class="stringliteral">" Y = [start="</span> << input_window.y().start() << <span class="stringliteral">", end="</span> << input_window.y().end() << <span class="stringliteral">", step="</span> << input_window.y().step() << <span class="stringliteral">"]\n"</span>;</div><div class="line"> std::cout << <span class="stringliteral">" Z = [start="</span> << input_window.z().start() << <span class="stringliteral">", end="</span> << input_window.z().end() << <span class="stringliteral">", step="</span> << input_window.z().step() << <span class="stringliteral">"]\n"</span>;</div><div class="line"></div><div class="line"> <span class="comment">// Create an iterator:</span></div><div class="line"> Iterator input_it(&input, input_window);</div><div class="line"></div><div class="line"> <span class="comment">// Iterate through the elements of src_data and copy them one by one to the input tensor:</span></div><div class="line"> <span class="comment">// This is equivalent to:</span></div><div class="line"> <span class="comment">// for( unsigned int z = 0; z < batch; ++z)</span></div><div class="line"> <span class="comment">// {</span></div><div class="line"> <span class="comment">// for( unsigned int y = 0; y < height; ++y)</span></div><div class="line"> <span class="comment">// {</span></div><div class="line"> <span class="comment">// for( unsigned int x = 0; x < width; ++x)</span></div><div class="line"> <span class="comment">// {</span></div><div class="line"> <span class="comment">// *reinterpret_cast<float*>( input.buffer() + input.info()->offset_element_in_bytes(Coordinates(x,y,z))) = src_data[ z * (width*height) + y * width + x];</span></div><div class="line"> <span class="comment">// }</span></div><div class="line"> <span class="comment">// }</span></div><div class="line"> <span class="comment">// }</span></div><div class="line"> <span class="comment">// Except it works for an arbitrary number of dimensions</span></div><div class="line"> <a class="code" href="namespacearm__compute.xhtml#a6c0dcc38187027dcb89cd9724bc5a823">execute_window_loop</a>(input_window, [&](<span class="keyword">const</span> Coordinates & <span class="keywordtype">id</span>)</div><div class="line"> {</div><div class="line"> std::cout << <span class="stringliteral">"Setting item ["</span> << <span class="keywordtype">id</span>.x() << <span class="stringliteral">","</span> << <span class="keywordtype">id</span>.y() << <span class="stringliteral">","</span> << <span class="keywordtype">id</span>.z() << <span class="stringliteral">"]\n"</span>;</div><div class="line"> *<span class="keyword">reinterpret_cast<</span><span class="keywordtype">float</span> *<span class="keyword">></span>(input_it.ptr()) = src_data[<span class="keywordtype">id</span>.z() * (width * height) + <span class="keywordtype">id</span>.y() * width + <span class="keywordtype">id</span>.x()];</div><div class="line"> },</div><div class="line"> input_it);</div><div class="line"></div><div class="line"> <span class="comment">// Run NEON softmax:</span></div><div class="line"> softmax.run();</div><div class="line"></div><div class="line"> <span class="comment">// More efficient way: create an iterator to iterate through each row (instead of each element) of the output tensor:</span></div><div class="line"> Window output_window;</div><div class="line"> output_window.use_tensor_dimensions(output.info()->tensor_shape(), <span class="comment">/* first_dimension =*/</span><a class="code" href="classarm__compute_1_1_window.xhtml#ad2d402364fa822b0b7775081291eeca9">Window::DimY</a>); <span class="comment">// Iterate through the rows (not each element)</span></div><div class="line"> std::cout << <span class="stringliteral">" Dimensions of the output's iterator:\n"</span>;</div><div class="line"> std::cout << <span class="stringliteral">" X = [start="</span> << output_window.x().start() << <span class="stringliteral">", end="</span> << output_window.x().end() << <span class="stringliteral">", step="</span> << output_window.x().step() << <span class="stringliteral">"]\n"</span>;</div><div class="line"> std::cout << <span class="stringliteral">" Y = [start="</span> << output_window.y().start() << <span class="stringliteral">", end="</span> << output_window.y().end() << <span class="stringliteral">", step="</span> << output_window.y().step() << <span class="stringliteral">"]\n"</span>;</div><div class="line"> std::cout << <span class="stringliteral">" Z = [start="</span> << output_window.z().start() << <span class="stringliteral">", end="</span> << output_window.z().end() << <span class="stringliteral">", step="</span> << output_window.z().step() << <span class="stringliteral">"]\n"</span>;</div><div class="line"></div><div class="line"> <span class="comment">// Create an iterator:</span></div><div class="line"> Iterator output_it(&output, output_window);</div><div class="line"></div><div class="line"> <span class="comment">// Iterate through the rows of the output tensor and copy them to dst_data:</span></div><div class="line"> <span class="comment">// This is equivalent to:</span></div><div class="line"> <span class="comment">// for( unsigned int z = 0; z < batch; ++z)</span></div><div class="line"> <span class="comment">// {</span></div><div class="line"> <span class="comment">// for( unsigned int y = 0; y < height; ++y)</span></div><div class="line"> <span class="comment">// {</span></div><div class="line"> <span class="comment">// memcpy( dst_data + z * (width*height) + y * width, input.buffer() + input.info()->offset_element_in_bytes(Coordinates(0,y,z)), width * sizeof(float));</span></div><div class="line"> <span class="comment">// }</span></div><div class="line"> <span class="comment">// }</span></div><div class="line"> <span class="comment">// Except it works for an arbitrary number of dimensions</span></div><div class="line"> <a class="code" href="namespacearm__compute.xhtml#a6c0dcc38187027dcb89cd9724bc5a823">execute_window_loop</a>(output_window, [&](<span class="keyword">const</span> Coordinates & <span class="keywordtype">id</span>)</div><div class="line"> {</div><div class="line"> std::cout << <span class="stringliteral">"Copying one row starting from ["</span> << <span class="keywordtype">id</span>.x() << <span class="stringliteral">","</span> << <span class="keywordtype">id</span>.y() << <span class="stringliteral">","</span> << <span class="keywordtype">id</span>.z() << <span class="stringliteral">"]\n"</span>;</div><div class="line"> <span class="comment">// Copy one whole row:</span></div><div class="line"> memcpy(dst_data + <span class="keywordtype">id</span>.z() * (width * height) + <span class="keywordtype">id</span>.y() * width, output_it.ptr(), width * <span class="keyword">sizeof</span>(float));</div><div class="line"> },</div><div class="line"> output_it);</div><div class="line"></div><div class="line"> <span class="keyword">delete</span>[] src_data;</div><div class="line"> <span class="keyword">delete</span>[] dst_data;</div></div><!-- fragment --> <h1><a class="anchor" id="S4_7_memory_manager"></a>
+<div class="fragment"><div class="line">
constexpr <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> width = 4;</div><div class="line"> constexpr <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> height = 3;</div><div class="line"> constexpr <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> batch = 2;</div><div class="line"></div><div class="line"> src_data = <span class="keyword">new</span> <span class="keywordtype">float</span>[width * height * batch];</div><div class="line"> dst_data = <span class="keyword">new</span> <span class="keywordtype">float</span>[width * height * batch];</div><div class="line"></div><div class="line"> <span class="comment">// Fill src_data with dummy values:</span></div><div class="line"> <span class="keywordflow">for</span>(<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a7b8004eef325a40dd43eb80755610fff">b</a> = 0; <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a7b8004eef325a40dd43eb80755610fff">b</a> < batch; <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a7b8004eef325a40dd43eb80755610fff">b</a>++)</div><div class="line"> {</div><div class="line"> <span class="keywordflow">for</span>(<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> h = 0; h < height; h++)</div><div class="line"> {</div><div class="line"> <span class="keywordflow">for</span>(<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> w = 0; w < width; w++)</div><div class="line"> {</div><div class="line"> src_data[<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a7b8004eef325a40dd43eb80755610fff">b</a> * (width * height) + h * width + w] = static_cast<float>(100 * <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a7b8004eef325a40dd43eb80755610fff">b</a> + 10 * h + w);</div><div class="line"> }</div><div class="line"> }</div><div class="line"> }</div><div class="line"></div><div class="line"> <span class="comment">// Initialize the tensors dimensions and type:</span></div><div class="line"> <span class="keyword">const</span> TensorShape <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a45cde9abb508c62d67c3bb2b9bf566a5">shape</a>(width, height, batch);</div><div class="line"> input.allocator()->init(TensorInfo(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a45cde9abb508c62d67c3bb2b9bf566a5">shape</a>, 1, <a class="code" href="namespacearm__compute.xhtml#ab4e88c89b3b7ea1735996cc4def22d58a44ad4ef5a76e6aa6fb3e3fa079a54fda">DataType::F32</a>));</div><div class="line"> output.allocator()->init(TensorInfo(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a45cde9abb508c62d67c3bb2b9bf566a5">shape</a>, 1, <a class="code" href="namespacearm__compute.xhtml#ab4e88c89b3b7ea1735996cc4def22d58a44ad4ef5a76e6aa6fb3e3fa079a54fda">DataType::F32</a>));</div><div class="line"></div><div class="line"> <span class="comment">// Configure softmax:</span></div><div class="line"> softmax.configure(&input, &output);</div><div class="line"></div><div class="line"> <span class="comment">// Allocate the input / output tensors:</span></div><div class="line"> input.allocator()->allocate();</div><div class="line"> output.allocator()->allocate();</div><div class="line"></div><div class="line"> <span class="comment">// Fill the input tensor:</span></div><div class="line"> <span class="comment">// Simplest way: create an iterator to iterate through each element of the input tensor:</span></div><div class="line"> Window input_window;</div><div class="line"> input_window.use_tensor_dimensions(input.info()->tensor_shape());</div><div class="line"> std::cout << <span class="stringliteral">" Dimensions of the input's iterator:\n"</span>;</div><div class="line"> std::cout << <span class="stringliteral">" X = [start="</span> << input_window.x().start() << <span class="stringliteral">", end="</span> << input_window.x().end() << <span class="stringliteral">", step="</span> << input_window.x().step() << <span class="stringliteral">"]\n"</span>;</div><div class="line"> std::cout << <span class="stringliteral">" Y = [start="</span> << input_window.y().start() << <span class="stringliteral">", end="</span> << input_window.y().end() << <span class="stringliteral">", step="</span> << input_window.y().step() << <span class="stringliteral">"]\n"</span>;</div><div class="line"> std::cout << <span class="stringliteral">" Z = [start="</span> << input_window.z().start() << <span class="stringliteral">", end="</span> << input_window.z().end() << <span class="stringliteral">", step="</span> << input_window.z().step() << <span class="stringliteral">"]\n"</span>;</div><div class="line"></div><div class="line"> <span class="comment">// Create an iterator:</span></div><div class="line"> Iterator input_it(&input, input_window);</div><div class="line"></div><div class="line"> <span class="comment">// Iterate through the elements of src_data and copy them one by one to the input tensor:</span></div><div class="line"> <span class="comment">// This is equivalent to:</span></div><div class="line"> <span class="comment">// for( unsigned int z = 0; z < batch; ++z)</span></div><div class="line"> <span class="comment">// {</span></div><div class="line"> <span class="comment">// for( unsigned int y = 0; y < height; ++y)</span></div><div class="line"> <span class="comment">// {</span></div><div class="line"> <span class="comment">// for( unsigned int x = 0; x < width; ++x)</span></div><div class="line"> <span class="comment">// {</span></div><div class="line"> <span class="comment">// *reinterpret_cast<float*>( input.buffer() + input.info()->offset_element_in_bytes(Coordinates(x,y,z))) = src_data[ z * (width*height) + y * width + x];</span></div><div class="line"> <span class="comment">// }</span></div><div class="line"> <span class="comment">// }</span></div><div class="line"> <span class="comment">// }</span></div><div class="line"> <span class="comment">// Except it works for an arbitrary number of dimensions</span></div><div class="line"> <a class="code" href="namespacearm__compute.xhtml#a6c0dcc38187027dcb89cd9724bc5a823">execute_window_loop</a>(input_window, [&](<span class="keyword">const</span> Coordinates & <span class="keywordtype">id</span>)</div><div class="line"> {</div><div class="line"> std::cout << <span class="stringliteral">"Setting item ["</span> << <span class="keywordtype">id</span>.x() << <span class="stringliteral">","</span> << <span class="keywordtype">id</span>.y() << <span class="stringliteral">","</span> << <span class="keywordtype">id</span>.z() << <span class="stringliteral">"]\n"</span>;</div><div class="line"> *<span class="keyword">reinterpret_cast<</span><span class="keywordtype">float</span> *<span class="keyword">></span>(input_it.ptr()) = src_data[<span class="keywordtype">id</span>.z() * (width * height) + <span class="keywordtype">id</span>.y() * width + <span class="keywordtype">id</span>.x()];</div><div class="line"> },</div><div class="line"> input_it);</div><div class="line"></div><div class="line"> <span class="comment">// More efficient way: create an iterator to iterate through each row (instead of each element) of the output tensor:</span></div><div class="line"> Window output_window;</div><div class="line"> output_window.use_tensor_dimensions(output.info()->tensor_shape(), <span class="comment">/* first_dimension =*/</span><a class="code" href="classarm__compute_1_1_window.xhtml#ad2d402364fa822b0b7775081291eeca9">Window::DimY</a>); <span class="comment">// Iterate through the rows (not each element)</span></div><div class="line"> std::cout << <span class="stringliteral">" Dimensions of the output's iterator:\n"</span>;</div><div class="line"> std::cout << <span class="stringliteral">" X = [start="</span> << output_window.x().start() << <span class="stringliteral">", end="</span> << output_window.x().end() << <span class="stringliteral">", step="</span> << output_window.x().step() << <span class="stringliteral">"]\n"</span>;</div><div class="line"> std::cout << <span class="stringliteral">" Y = [start="</span> << output_window.y().start() << <span class="stringliteral">", end="</span> << output_window.y().end() << <span class="stringliteral">", step="</span> << output_window.y().step() << <span class="stringliteral">"]\n"</span>;</div><div class="line"> std::cout << <span class="stringliteral">" Z = [start="</span> << output_window.z().start() << <span class="stringliteral">", end="</span> << output_window.z().end() << <span class="stringliteral">", step="</span> << output_window.z().step() << <span class="stringliteral">"]\n"</span>;</div><div class="line"></div><div class="line"> <span class="comment">// Create an iterator:</span></div><div class="line"> Iterator output_it(&output, output_window);</div><div class="line"></div><div class="line"> <span class="comment">// Iterate through the rows of the output tensor and copy them to dst_data:</span></div><div class="line"> <span class="comment">// This is equivalent to:</span></div><div class="line"> <span class="comment">// for( unsigned int z = 0; z < batch; ++z)</span></div><div class="line"> <span class="comment">// {</span></div><div class="line"> <span class="comment">// for( unsigned int y = 0; y < height; ++y)</span></div><div class="line"> <span class="comment">// {</span></div><div class="line"> <span class="comment">// memcpy( dst_data + z * (width*height) + y * width, input.buffer() + input.info()->offset_element_in_bytes(Coordinates(0,y,z)), width * sizeof(float));</span></div><div class="line"> <span class="comment">// }</span></div><div class="line"> <span class="comment">// }</span></div><div class="line"> <span class="comment">// Except it works for an arbitrary number of dimensions</span></div><div class="line"> <a class="code" href="namespacearm__compute.xhtml#a6c0dcc38187027dcb89cd9724bc5a823">execute_window_loop</a>(output_window, [&](<span class="keyword">const</span> Coordinates & <span class="keywordtype">id</span>)</div><div class="line"> {</div><div class="line"> std::cout << <span class="stringliteral">"Copying one row starting from ["</span> << <span class="keywordtype">id</span>.x() << <span class="stringliteral">","</span> << <span class="keywordtype">id</span>.y() << <span class="stringliteral">","</span> << <span class="keywordtype">id</span>.z() << <span class="stringliteral">"]\n"</span>;</div><div class="line"> <span class="comment">// Copy one whole row:</span></div><div class="line"> memcpy(dst_data + <span class="keywordtype">id</span>.z() * (width * height) + <span class="keywordtype">id</span>.y() * width, output_it.ptr(), width * <span class="keyword">sizeof</span>(float));</div><div class="line"> },</div><div class="line"> output_it);</div><div class="line"></div></div><!-- fragment --> <h1><a class="anchor" id="S4_7_memory_manager"></a>
MemoryManager</h1>
<p><a class="el" href="classarm__compute_1_1_i_memory_manager.xhtml">IMemoryManager</a> is a memory managing interface that can be used to reduce the memory requirements of a given pipeline by recycling temporary buffers.</p>
<h2><a class="anchor" id="S4_7_1_memory_manager_components"></a>
<h2><a class="anchor" id="S4_7_3_memory_manager_function_support"></a>
Function support</h2>
<p>Most of the library's function have been ported to use <a class="el" href="classarm__compute_1_1_i_memory_manager.xhtml">IMemoryManager</a> for their internal temporary buffers.</p>
-<p>If that is the case, a memory manager can be passed to them during construction to reuse memory among these functions. </p><div class="fragment"><div class="line"><span class="comment">// Setup Memory Manager</span></div><div class="line">CLBufferAllocator allocator{}; <span class="comment">// Create an allocator to use for the backing memory allocation</span></div><div class="line"><span class="keyword">auto</span> lifetime_mgr = std::make_shared<BlobLifetimeManager>(); <span class="comment">// Create Lifetime Manager</span></div><div class="line"><span class="keyword">auto</span> pool_mgr = std::make_shared<PoolManager>(); <span class="comment">// Create Pool Manager</span></div><div class="line"><span class="keyword">auto</span> mm = std::make_shared<MemoryManagerOnDemand>(lifetime_mgr, pool_mgr); <span class="comment">// Create Memory Manager</span></div><div class="line"></div><div class="line"><span class="comment">// Create two convolution layers and use the memory manager to manager their internal temporary buffers</span></div><div class="line">CLConvolutionLayer conv1(mm), conv2(mm);</div><div class="line"></div><div class="line"><span class="comment">// Configure layers</span></div><div class="line">conv1.configure(...);</div><div class="line">conv2.configure(...);</div><div class="line"></div><div class="line"><span class="comment">// Finalize memory manager</span></div><div class="line">mm->set_allocator(&allocator); <span class="comment">// Set allocator to use</span></div><div class="line">mm->set_set_num_pools(1); <span class="comment">// Set number of pools to create in case parallel operations can be run</span></div><div class="line">mm->finalize(); <span class="comment">// Finalize memory manager (Object lifetime check, Memory pool creation etc)</span></div><div class="line"></div><div class="line"><span class="comment">// Run layers (Memory will be recycled for internal buffers for conv1 and conv2</span></div><div class="line">conv1.run();</div><div class="line">conv2.run();</div></div><!-- fragment --> </div></div><!-- contents -->
+<p>If that is the case, a memory manager can be passed to them during construction to reuse memory among these functions. </p><div class="fragment"><div class="line"><span class="comment">// Setup Memory Manager</span></div><div class="line">CLBufferAllocator allocator{}; <span class="comment">// Create an allocator to use for the backing memory allocation</span></div><div class="line"><span class="keyword">auto</span> lifetime_mgr = std::make_shared<BlobLifetimeManager>(); <span class="comment">// Create Lifetime Manager</span></div><div class="line"><span class="keyword">auto</span> pool_mgr = std::make_shared<PoolManager>(); <span class="comment">// Create Pool Manager</span></div><div class="line"><span class="keyword">auto</span> mm = std::make_shared<MemoryManagerOnDemand>(lifetime_mgr, pool_mgr); <span class="comment">// Create Memory Manager</span></div><div class="line"></div><div class="line"><span class="comment">// Create two convolution layers and use the memory manager to manager their internal temporary buffers</span></div><div class="line">CLConvolutionLayer conv1(mm), conv2(mm);</div><div class="line"></div><div class="line"><span class="comment">// Configure layers</span></div><div class="line">conv1.configure(...);</div><div class="line">conv2.configure(...);</div><div class="line"></div><div class="line"><span class="comment">// Finalize memory manager</span></div><div class="line">mm->set_allocator(&allocator); <span class="comment">// Set allocator to use</span></div><div class="line">mm->set_set_num_pools(1); <span class="comment">// Set number of pools to create in case parallel operations can be run</span></div><div class="line">mm->finalize(); <span class="comment">// Finalize memory manager (Object lifetime check, Memory pool creation etc)</span></div><div class="line"></div><div class="line"><span class="comment">// Run layers (Memory will be recycled for internal buffers for conv1 and conv2</span></div><div class="line">conv1.run();</div><div class="line">conv2.run();</div></div><!-- fragment --><h1><a class="anchor" id="S4_8_opencl_tuner"></a>
+OpenCL Tuner</h1>
+<p>OpenCL kernels when dispatched to the GPU take two arguments:</p><ul>
+<li>The Global Workgroup Size (GWS): That's the number of times to run an OpenCL kernel to process all the elements we want to process.</li>
+<li>The Local Workgroup Size (LWS): That's the number of elements we want to run in parallel on a GPU core at a given point in time.</li>
+</ul>
+<p>The LWS can be required by an algorithm (For example if it contains memory barriers or uses local memory) but it can also be used for performance reasons to tweak the performance of a kernel: the execution time of the overall kernel might vary significantly depending on how the GWS is broken down.</p>
+<p>However, there is no universal rule regarding which LWS is best for a given kernel, so instead we created the <a class="el" href="classarm__compute_1_1_c_l_tuner.xhtml">CLTuner</a>.</p>
+<p>When the <a class="el" href="classarm__compute_1_1_c_l_tuner.xhtml">CLTuner</a> is enabled ( Target = 2 for the graph examples), the first time an OpenCL kernel is executed the Compute Library will try to run it with a variety of LWS values and will remember which one performed best for subsequent runs. At the end of the run the <a class="el" href="classarm__compute_1_1graph_1_1_graph.xhtml">graph::Graph</a> will try to save these tuning parameters to a file.</p>
+<p>However this process takes quite a lot of time, which is why it cannot be enabled all the time.</p>
+<p>But, when the <a class="el" href="classarm__compute_1_1_c_l_tuner.xhtml">CLTuner</a> is disabled ( Target = 1 for the graph examples), the <a class="el" href="classarm__compute_1_1graph_1_1_graph.xhtml">graph::Graph</a> will try to reload the file containing the tuning parameters, then for each executed kernel the Compute Library will use the fine tuned LWS if it was present in the file or use a default LWS value if it's not. </p>
+</div></div><!-- contents -->
</div><!-- doc-content -->
<!-- start footer part -->
<div id="nav-path" class="navpath"><!-- id is needed for treeview function! -->
<ul>
- <li class="footer">Generated on Thu Dec 14 2017 23:48:34 for Compute Library by
+ <li class="footer">Generated on Wed May 23 2018 11:36:39 for Compute Library by
<a href="http://www.doxygen.org/index.html">
<img class="footer" src="doxygen.png" alt="doxygen"/></a> 1.8.11 </li>
</ul>
projects / platform / upstream / armcl.git / blobdiff