Differential D3108 Diff 12789 intern/cycles/device/opencl/opencl_base.cpp

Changeset View

Standalone View

intern/cycles/device/opencl/opencl_base.cpp

Show First 20 Lines • Show All 1,085 Lines • ▼ Show 20 Lines	void OpenCLDeviceBase::shader(DeviceTask& task)
cl_int d_shader_eval_type = task.shader_eval_type;		cl_int d_shader_eval_type = task.shader_eval_type;
cl_int d_shader_filter = task.shader_filter;		cl_int d_shader_filter = task.shader_filter;
cl_int d_shader_x = task.shader_x;		cl_int d_shader_x = task.shader_x;
cl_int d_shader_w = task.shader_w;		cl_int d_shader_w = task.shader_w;
cl_int d_offset = task.offset;		cl_int d_offset = task.offset;

cl_kernel kernel;		cl_kernel kernel;

if(task.shader_eval_type >= SHADER_EVAL_BAKE) {		if(task.shader_eval_type == SHADER_EVAL_DISPLACE) {
kernel = base_program(ustring("bake"));
}
else if(task.shader_eval_type == SHADER_EVAL_DISPLACE) {
kernel = base_program(ustring("displace"));		kernel = base_program(ustring("displace"));
}		}
else {		else {
kernel = base_program(ustring("background"));		kernel = base_program(ustring("background"));
}		}

cl_uint start_arg_index =		cl_uint start_arg_index =
kernel_set_args(kernel,		kernel_set_args(kernel,
Show All 28 Lines	for(int sample = 0; sample < task.num_samples; sample++) {
enqueue_kernel(kernel, task.shader_w, 1);		enqueue_kernel(kernel, task.shader_w, 1);

clFinish(cqCommandQueue);		clFinish(cqCommandQueue);

task.update_progress(NULL);		task.update_progress(NULL);
}		}
}		}

		void OpenCLDeviceBase::bake(DeviceTask& task, RenderTile& rtile)
		{
		scoped_timer timer(&rtile.buffers->render_time);

		/* Cast arguments to cl types. */
		cl_mem d_data = CL_MEM_PTR(const_mem_map["__data"]->device_pointer);
		cl_mem d_buffer = CL_MEM_PTR(rtile.buffer);
		cl_int d_x = rtile.x;
		cl_int d_y = rtile.y;
		cl_int d_w = rtile.w;
		cl_int d_h = rtile.h;
		cl_int d_offset = rtile.offset;
		cl_int d_stride = rtile.stride;

		cl_kernel kernel = base_program(ustring("bake"));

		cl_uint start_arg_index =
		kernel_set_args(kernel,
		0,
		d_data,
		d_buffer);

		set_kernel_arg_buffers(kernel, &start_arg_index);

		start_arg_index += kernel_set_args(kernel,
		start_arg_index,
		d_x,
		d_y,
		d_w,
		d_h,
		d_offset,
		d_stride);


		int start_sample = rtile.start_sample;
		int end_sample = rtile.start_sample + rtile.num_samples;

		for(int sample = start_sample; sample < end_sample; sample++) {
		if(task.get_cancel()) {
		if(task.need_finish_queue == false)
		break;
		}

		kernel_set_args(kernel, start_arg_index, sample);

		enqueue_kernel(kernel, d_w, d_h);

		rtile.sample = sample + 1;

		task.update_progress(&rtile, rtile.w*rtile.h);
		}

		clFinish(cqCommandQueue);
		}

string OpenCLDeviceBase::kernel_build_options(const string *debug_src)		string OpenCLDeviceBase::kernel_build_options(const string *debug_src)
{		{
string build_options = "-cl-no-signed-zeros -cl-mad-enable ";		string build_options = "-cl-no-signed-zeros -cl-mad-enable ";

if(platform_name == "NVIDIA CUDA") {		if(platform_name == "NVIDIA CUDA") {
build_options += "-D__KERNEL_OPENCL_NVIDIA__ "		build_options += "-D__KERNEL_OPENCL_NVIDIA__ "
"-cl-nv-maxrregcount=32 "		"-cl-nv-maxrregcount=32 "
"-cl-nv-verbose ";		"-cl-nv-verbose ";
▲ Show 20 Lines • Show All 201 Lines • Show Last 20 Lines