83 lines
3.4 KiB
HTML
83 lines
3.4 KiB
HTML
<p><strong>Note:</strong> Use the class <code class="language-none">"language-opencl"</code> for OpenCL kernel code.
|
|
Host code is automatically highlighted with the <code class="language-none">"language-c"</code>
|
|
or <code class="language-none">"language-cpp"</code> class.
|
|
</p>
|
|
|
|
<h2>OpenCL host code</h2>
|
|
<pre class="language-cpp"><code>// OpenCL functions, constants, etc. are also highlighted in OpenCL host code in the c or cpp language
|
|
cl::Event KernelFilterImages::runSingle(const cl::Image2D& imgSrc, SPImage2D& imgDst)
|
|
{
|
|
const size_t rows = imgSrc.getImageInfo<CL_IMAGE_HEIGHT>();
|
|
const size_t cols = imgSrc.getImageInfo<CL_IMAGE_WIDTH>();
|
|
|
|
ASSERT(rows > 0 && cols > 0, "The image object seems to be invalid, no rows/cols set");
|
|
ASSERT(imgSrc.getImageInfo<CL_IMAGE_FORMAT>().image_channel_data_type == CL_FLOAT, "Only float type images are supported");
|
|
ASSERT(imgSrc.getInfo<CL_MEM_FLAGS>() == CL_MEM_READ_ONLY || imgSrc.getInfo<CL_MEM_FLAGS>() == CL_MEM_READ_WRITE, "Can't read the input image");
|
|
|
|
imgDst = std::make_shared<cl::Image2D>(*context, CL_MEM_READ_WRITE, cl::ImageFormat(CL_R, CL_FLOAT), cols, rows);
|
|
|
|
cl::Kernel kernel(*program, "filter_single");
|
|
kernel.setArg(0, imgSrc);
|
|
kernel.setArg(1, *imgDst);
|
|
kernel.setArg(2, bufferKernel1);
|
|
kernel.setArg(3, kernel1.rows);
|
|
kernel.setArg(4, kernel1.rows / 2);
|
|
kernel.setArg(5, kernel1.cols);
|
|
kernel.setArg(6, kernel1.cols / 2);
|
|
kernel.setArg(7, border);
|
|
|
|
cl::Event eventFilter;
|
|
const cl::NDRange global(cols, rows);
|
|
queue->enqueueNDRangeKernel(kernel, cl::NullRange, global, cl::NullRange, &events, &eventFilter);
|
|
}</code></pre>
|
|
|
|
<h2>OpenCL kernel code</h2>
|
|
<pre><code>// CLK_ADDRESS_CLAMP_TO_EDGE = aaa|abcd|ddd
|
|
constant sampler_t sampler = CLK_NORMALIZED_COORDS_FALSE | CLK_ADDRESS_CLAMP_TO_EDGE | CLK_FILTER_NEAREST;
|
|
typedef float type_single;
|
|
|
|
type_single filter_sum_single_3x3(read_only image2d_t imgIn,
|
|
constant float* filterKernel,
|
|
const int2 coordBase,
|
|
const int border)
|
|
{
|
|
type_single sum = (type_single)(0.0f);
|
|
const int rows = get_image_height(imgIn);
|
|
const int cols = get_image_width(imgIn);
|
|
int2 coordCurrent;
|
|
int2 coordBorder;
|
|
float color;
|
|
|
|
// Image patch is row-wise accessed
|
|
// Filter kernel is centred in the middle
|
|
#pragma unroll
|
|
for (int y = -ROWS_HALF_3x3; y <= ROWS_HALF_3x3; ++y) // Start at the top left corner of the filter
|
|
{
|
|
coordCurrent.y = coordBase.y + y;
|
|
#pragma unroll
|
|
for (int x = -COLS_HALF_3x3; x <= COLS_HALF_3x3; ++x) // And end at the bottom right corner
|
|
{
|
|
coordCurrent.x = coordBase.x + x;
|
|
coordBorder = borderCoordinate(coordCurrent, rows, cols, border);
|
|
color = read_imagef(imgIn, sampler, coordBorder).x;
|
|
|
|
const int idx = (y + ROWS_HALF_3x3) * COLS_3x3 + x + COLS_HALF_3x3;
|
|
sum += color * filterKernel[idx];
|
|
}
|
|
}
|
|
|
|
return sum;
|
|
}
|
|
|
|
kernel void filter_single_3x3(read_only image2d_t imgIn,
|
|
write_only image2d_t imgOut,
|
|
constant float* filterKernel,
|
|
const int border)
|
|
{
|
|
int2 coordBase = (int2)(get_global_id(0), get_global_id(1));
|
|
|
|
type_single sum = filter_sum_single_3x3(imgIn, filterKernel, coordBase, border);
|
|
|
|
write_imagef(imgOut, coordBase, sum);
|
|
}</code></pre>
|