bool Sobel::runHalideGPU(Image input, Image output, const Params& params)
{
#if ENABLE_HALIDE
// Create halide buffers
buffer_t inputBuffer = createHalideBuffer(input);
buffer_t outputBuffer = createHalideBuffer(output);
reportStatus("Running Halide GPU filter");
// Warm-up run
inputBuffer.host_dirty = true;
halide_sobel_gpu(&inputBuffer, &outputBuffer);
halide_dev_sync(NULL);
// Timed runs
startTiming();
for (int i = 0; i < params.iterations; i++)
{
halide_sobel_gpu(&inputBuffer, &outputBuffer);
}
halide_dev_sync(NULL);
stopTiming();
halide_copy_to_host(NULL, &outputBuffer);
halide_release(NULL);
return outputResults(input, output, params);
#else
reportStatus("Halide not enabled during build.");
return false;
#endif
}
WEAK void halide_release() {
// TODO: this is for timing; bad for release-mode performance
#ifdef DEBUG
halide_printf("dev_sync on exit" );
#endif
halide_dev_sync();
// Unload the module
if (__mod) {
CHECK_CALL( clReleaseProgram(__mod), "clReleaseProgram" );
__mod = 0;
}
// Unload context (ref counted).
CHECK_CALL( clReleaseCommandQueue(cl_q), "clReleaseCommandQueue" );
CHECK_CALL( clReleaseContext(cl_ctx), "clReleaseContext" );
}
WEAK void halide_release(void *user_context) {
// TODO: this is for timing; bad for release-mode performance
#ifdef DEBUG
halide_printf(user_context, "dev_sync on exit\n" );
#endif
halide_dev_sync(user_context);
// Unload the module
if (__mod) {
#ifdef DEBUG
halide_printf(user_context, "clReleaseProgram %p\n", __mod);
#endif
CHECK_CALL( clReleaseProgram(__mod), "clReleaseProgram" );
__mod = 0;
}
// TODO: This is not a good solution to deal with this problem (finding out if the
// cl_ctx/cl_q are going to be freed). I think a larger redesign of the global
// context scheme might be necessary.
cl_uint refs = 0;
clGetContextInfo(*cl_ctx, CL_CONTEXT_REFERENCE_COUNT, sizeof(refs), &refs, NULL);
// Unload context (ref counted).
CHECK_CALL( clReleaseCommandQueue(*cl_q), "clReleaseCommandQueue" );
#ifdef DEBUG
halide_printf(user_context, "clReleaseContext %p\n", *cl_ctx);
#endif
CHECK_CALL( clReleaseContext(*cl_ctx), "clReleaseContext" );
// See TODO above...
if (--refs == 0) {
*cl_ctx = NULL;
*cl_q = NULL;
}
}
