diff options
Diffstat (limited to 'intern/cycles/device/opencl/opencl_base.cpp')
| -rw-r--r-- | intern/cycles/device/opencl/opencl_base.cpp | 209 |
1 files changed, 171 insertions, 38 deletions
diff --git a/intern/cycles/device/opencl/opencl_base.cpp b/intern/cycles/device/opencl/opencl_base.cpp index e4ab979dcbf..3db3efd1103 100644 --- a/intern/cycles/device/opencl/opencl_base.cpp +++ b/intern/cycles/device/opencl/opencl_base.cpp @@ -20,6 +20,7 @@ #include "kernel/kernel_types.h" +#include "util/util_algorithm.h" #include "util/util_foreach.h" #include "util/util_logging.h" #include "util/util_md5.h" @@ -28,6 +29,15 @@ CCL_NAMESPACE_BEGIN +struct texture_slot_t { + texture_slot_t(const string& name, int slot) + : name(name), + slot(slot) { + } + string name; + int slot; +}; + bool OpenCLDeviceBase::opencl_error(cl_int err) { if(err != CL_SUCCESS) { @@ -62,7 +72,7 @@ void OpenCLDeviceBase::opencl_assert_err(cl_int err, const char* where) } OpenCLDeviceBase::OpenCLDeviceBase(DeviceInfo& info, Stats &stats, bool background_) -: Device(info, stats, background_) +: Device(info, stats, background_), memory_manager(this) { cpPlatform = NULL; cdDevice = NULL; @@ -70,6 +80,7 @@ OpenCLDeviceBase::OpenCLDeviceBase(DeviceInfo& info, Stats &stats, bool backgrou cqCommandQueue = NULL; null_mem = 0; device_initialized = false; + textures_need_update = true; vector<OpenCLPlatformDevice> usable_devices; OpenCLInfo::get_usable_devices(&usable_devices); @@ -125,6 +136,12 @@ OpenCLDeviceBase::OpenCLDeviceBase(DeviceInfo& info, Stats &stats, bool backgrou return; } + /* Allocate this right away so that texture_descriptors_buffer is placed at offset 0 in the device memory buffers */ + texture_descriptors.resize(1); + texture_descriptors_buffer.resize(1); + texture_descriptors_buffer.data_pointer = (device_ptr)&texture_descriptors[0]; + memory_manager.alloc("texture_descriptors", texture_descriptors_buffer); + fprintf(stderr, "Device init success\n"); device_initialized = true; } @@ -133,6 +150,8 @@ OpenCLDeviceBase::~OpenCLDeviceBase() { task_pool.stop(); + memory_manager.free(); + if(null_mem) clReleaseMemObject(CL_MEM_PTR(null_mem)); @@ -209,7 +228,8 @@ bool OpenCLDeviceBase::load_kernels(const DeviceRequestedFeatures& requested_fea base_program = OpenCLProgram(this, "base", "kernel.cl", build_options_for_base_program(requested_features)); base_program.add_kernel(ustring("convert_to_byte")); base_program.add_kernel(ustring("convert_to_half_float")); - base_program.add_kernel(ustring("shader")); + base_program.add_kernel(ustring("displace")); + base_program.add_kernel(ustring("background")); base_program.add_kernel(ustring("bake")); base_program.add_kernel(ustring("zero_buffer")); @@ -276,6 +296,25 @@ void OpenCLDeviceBase::mem_alloc(const char *name, device_memory& mem, MemoryTyp size_t size = mem.memory_size(); + /* check there is enough memory available for the allocation */ + cl_ulong max_alloc_size = 0; + clGetDeviceInfo(cdDevice, CL_DEVICE_MAX_MEM_ALLOC_SIZE, sizeof(cl_ulong), &max_alloc_size, NULL); + + if(DebugFlags().opencl.mem_limit) { + max_alloc_size = min(max_alloc_size, + cl_ulong(DebugFlags().opencl.mem_limit - stats.mem_used)); + } + + if(size > max_alloc_size) { + string error = "Scene too complex to fit in available memory."; + if(name != NULL) { + error += string_printf(" (allocating buffer %s failed.)", name); + } + set_error(error); + + return; + } + cl_mem_flags mem_flag; void *mem_ptr = NULL; @@ -473,29 +512,35 @@ void OpenCLDeviceBase::const_copy_to(const char *name, void *host, size_t size) void OpenCLDeviceBase::tex_alloc(const char *name, device_memory& mem, - InterpolationType /*interpolation*/, - ExtensionType /*extension*/) + InterpolationType interpolation, + ExtensionType extension) { VLOG(1) << "Texture allocate: " << name << ", " << string_human_readable_number(mem.memory_size()) << " bytes. (" << string_human_readable_size(mem.memory_size()) << ")"; - mem_alloc(NULL, mem, MEM_READ_ONLY); - mem_copy_to(mem); - assert(mem_map.find(name) == mem_map.end()); - mem_map.insert(MemMap::value_type(name, mem.device_pointer)); + + memory_manager.alloc(name, mem); + /* Set the pointer to non-null to keep code that inspects its value from thinking its unallocated. */ + mem.device_pointer = 1; + textures[name] = Texture(&mem, interpolation, extension); + textures_need_update = true; } void OpenCLDeviceBase::tex_free(device_memory& mem) { if(mem.device_pointer) { - foreach(const MemMap::value_type& value, mem_map) { - if(value.second == mem.device_pointer) { - mem_map.erase(value.first); + mem.device_pointer = 0; + + if(memory_manager.free(mem)) { + textures_need_update = true; + } + + foreach(TexturesMap::value_type& value, textures) { + if(value.second.mem == &mem) { + textures.erase(value.first); break; } } - - mem_free(mem); } } @@ -561,6 +606,98 @@ void OpenCLDeviceBase::set_kernel_arg_mem(cl_kernel kernel, cl_uint *narg, const opencl_assert(clSetKernelArg(kernel, (*narg)++, sizeof(ptr), (void*)&ptr)); } +void OpenCLDeviceBase::set_kernel_arg_buffers(cl_kernel kernel, cl_uint *narg) +{ + flush_texture_buffers(); + + memory_manager.set_kernel_arg_buffers(kernel, narg); +} + +void OpenCLDeviceBase::flush_texture_buffers() +{ + if(!textures_need_update) { + return; + } + textures_need_update = false; + + /* Setup slots for textures. */ + int num_slots = 0; + + vector<texture_slot_t> texture_slots; + +#define KERNEL_TEX(type, ttype, name) \ + if(textures.find(#name) != textures.end()) { \ + texture_slots.push_back(texture_slot_t(#name, num_slots)); \ + } \ + num_slots++; +#include "kernel/kernel_textures.h" + + int num_data_slots = num_slots; + + foreach(TexturesMap::value_type& tex, textures) { + string name = tex.first; + + if(string_startswith(name, "__tex_image")) { + int pos = name.rfind("_"); + int id = atoi(name.data() + pos + 1); + texture_slots.push_back(texture_slot_t(name, + num_data_slots + id)); + num_slots = max(num_slots, num_data_slots + id + 1); + } + } + + /* Realloc texture descriptors buffer. */ + memory_manager.free(texture_descriptors_buffer); + + texture_descriptors.resize(num_slots); + texture_descriptors_buffer.resize(num_slots * sizeof(tex_info_t)); + texture_descriptors_buffer.data_pointer = (device_ptr)&texture_descriptors[0]; + + memory_manager.alloc("texture_descriptors", texture_descriptors_buffer); + + /* Fill in descriptors */ + foreach(texture_slot_t& slot, texture_slots) { + Texture& tex = textures[slot.name]; + + tex_info_t& info = texture_descriptors[slot.slot]; + + MemoryManager::BufferDescriptor desc = memory_manager.get_descriptor(slot.name); + + info.offset = desc.offset; + info.buffer = desc.device_buffer; + + if(string_startswith(slot.name, "__tex_image")) { + info.width = tex.mem->data_width; + info.height = tex.mem->data_height; + info.depth = tex.mem->data_depth; + + info.options = 0; + + if(tex.interpolation == INTERPOLATION_CLOSEST) { + info.options |= (1 << 0); + } + + switch(tex.extension) { + case EXTENSION_REPEAT: + info.options |= (1 << 1); + break; + case EXTENSION_EXTEND: + info.options |= (1 << 2); + break; + case EXTENSION_CLIP: + info.options |= (1 << 3); + break; + default: + break; + } + } + } + + /* Force write of descriptors. */ + memory_manager.free(texture_descriptors_buffer); + memory_manager.alloc("texture_descriptors", texture_descriptors_buffer); +} + void OpenCLDeviceBase::film_convert(DeviceTask& task, device_ptr buffer, device_ptr rgba_byte, device_ptr rgba_half) { /* cast arguments to cl types */ @@ -585,10 +722,7 @@ void OpenCLDeviceBase::film_convert(DeviceTask& task, device_ptr buffer, device_ d_rgba, d_buffer); -#define KERNEL_TEX(type, ttype, name) \ -set_kernel_arg_mem(ckFilmConvertKernel, &start_arg_index, #name); -#include "kernel/kernel_textures.h" -#undef KERNEL_TEX + set_kernel_arg_buffers(ckFilmConvertKernel, &start_arg_index); start_arg_index += kernel_set_args(ckFilmConvertKernel, start_arg_index, @@ -849,7 +983,6 @@ bool OpenCLDeviceBase::denoising_divide_shadow(device_ptr a_ptr, cl_kernel ckFilterDivideShadow = denoising_program(ustring("filter_divide_shadow")); - char split_kernel = is_split_kernel()? 1 : 0; kernel_set_args(ckFilterDivideShadow, 0, task->render_buffer.samples, tiles_mem, @@ -860,8 +993,7 @@ bool OpenCLDeviceBase::denoising_divide_shadow(device_ptr a_ptr, buffer_variance_mem, task->rect, task->render_buffer.pass_stride, - task->render_buffer.denoising_data_offset, - split_kernel); + task->render_buffer.denoising_data_offset); enqueue_kernel(ckFilterDivideShadow, task->rect.z-task->rect.x, task->rect.w-task->rect.y); @@ -882,7 +1014,6 @@ bool OpenCLDeviceBase::denoising_get_feature(int mean_offset, cl_kernel ckFilterGetFeature = denoising_program(ustring("filter_get_feature")); - char split_kernel = is_split_kernel()? 1 : 0; kernel_set_args(ckFilterGetFeature, 0, task->render_buffer.samples, tiles_mem, @@ -892,8 +1023,7 @@ bool OpenCLDeviceBase::denoising_get_feature(int mean_offset, variance_mem, task->rect, task->render_buffer.pass_stride, - task->render_buffer.denoising_data_offset, - split_kernel); + task->render_buffer.denoising_data_offset); enqueue_kernel(ckFilterGetFeature, task->rect.z-task->rect.x, task->rect.w-task->rect.y); @@ -983,7 +1113,6 @@ void OpenCLDeviceBase::shader(DeviceTask& task) cl_mem d_data = CL_MEM_PTR(const_mem_map["__data"]->device_pointer); cl_mem d_input = CL_MEM_PTR(task.shader_input); cl_mem d_output = CL_MEM_PTR(task.shader_output); - cl_mem d_output_luma = CL_MEM_PTR(task.shader_output_luma); cl_int d_shader_eval_type = task.shader_eval_type; cl_int d_shader_filter = task.shader_filter; cl_int d_shader_x = task.shader_x; @@ -992,10 +1121,15 @@ void OpenCLDeviceBase::shader(DeviceTask& task) cl_kernel kernel; - if(task.shader_eval_type >= SHADER_EVAL_BAKE) + if(task.shader_eval_type >= SHADER_EVAL_BAKE) { kernel = base_program(ustring("bake")); - else - kernel = base_program(ustring("shader")); + } + else if(task.shader_eval_type >= SHADER_EVAL_DISPLACE) { + kernel = base_program(ustring("displace")); + } + else { + kernel = base_program(ustring("background")); + } cl_uint start_arg_index = kernel_set_args(kernel, @@ -1004,16 +1138,7 @@ void OpenCLDeviceBase::shader(DeviceTask& task) d_input, d_output); - if(task.shader_eval_type < SHADER_EVAL_BAKE) { - start_arg_index += kernel_set_args(kernel, - start_arg_index, - d_output_luma); - } - -#define KERNEL_TEX(type, ttype, name) \ - set_kernel_arg_mem(kernel, &start_arg_index, #name); -#include "kernel/kernel_textures.h" -#undef KERNEL_TEX + set_kernel_arg_buffers(kernel, &start_arg_index); start_arg_index += kernel_set_args(kernel, start_arg_index, @@ -1226,7 +1351,7 @@ void OpenCLDeviceBase::store_cached_kernel( } string OpenCLDeviceBase::build_options_for_base_program( - const DeviceRequestedFeatures& /*requested_features*/) + const DeviceRequestedFeatures& requested_features) { /* TODO(sergey): By default we compile all features, meaning * mega kernel is not getting feature-based optimizations. @@ -1234,6 +1359,14 @@ string OpenCLDeviceBase::build_options_for_base_program( * Ideally we need always compile kernel with as less features * enabled as possible to keep performance at it's max. */ + + /* For now disable baking when not in use as this has major + * impact on kernel build times. + */ + if(!requested_features.use_baking) { + return "-D__NO_BAKING__"; + } + return ""; } |