diff options
Diffstat (limited to 'intern/cycles/device/device_cuda.cpp')
-rw-r--r-- | intern/cycles/device/device_cuda.cpp | 682 |
1 files changed, 682 insertions, 0 deletions
diff --git a/intern/cycles/device/device_cuda.cpp b/intern/cycles/device/device_cuda.cpp new file mode 100644 index 00000000000..76692ba8657 --- /dev/null +++ b/intern/cycles/device/device_cuda.cpp @@ -0,0 +1,682 @@ +/* + * Copyright 2011, Blender Foundation. + * + * This program is free software; you can redistribute it and/or + * modify it under the terms of the GNU General Public License + * as published by the Free Software Foundation; either version 2 + * of the License, or (at your option) any later version. + * + * This program is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + * GNU General Public License for more details. + * + * You should have received a copy of the GNU General Public License + * along with this program; if not, write to the Free Software Foundation, + * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA. + */ + +#include <stdio.h> +#include <stdlib.h> +#include <string.h> + +#include "device.h" +#include "device_intern.h" + +#include "util_cuda.h" +#include "util_debug.h" +#include "util_map.h" +#include "util_opengl.h" +#include "util_path.h" +#include "util_types.h" + +CCL_NAMESPACE_BEGIN + +class CUDADevice : public Device +{ +public: + CUdevice cuDevice; + CUcontext cuContext; + CUmodule cuModule; + map<device_ptr, bool> tex_interp_map; + int cuDevId; + + struct PixelMem { + GLuint cuPBO; + CUgraphicsResource cuPBOresource; + GLuint cuTexId; + int w, h; + }; + + map<device_ptr, PixelMem> pixel_mem_map; + + CUdeviceptr cuda_device_ptr(device_ptr mem) + { + return (CUdeviceptr)mem; + } + + const char *cuda_error_string(CUresult result) + { + switch(result) { + case CUDA_SUCCESS: return "No errors"; + case CUDA_ERROR_INVALID_VALUE: return "Invalid value"; + case CUDA_ERROR_OUT_OF_MEMORY: return "Out of memory"; + case CUDA_ERROR_NOT_INITIALIZED: return "Driver not initialized"; + case CUDA_ERROR_DEINITIALIZED: return "Driver deinitialized"; + + case CUDA_ERROR_NO_DEVICE: return "No CUDA-capable device available"; + case CUDA_ERROR_INVALID_DEVICE: return "Invalid device"; + + case CUDA_ERROR_INVALID_IMAGE: return "Invalid kernel image"; + case CUDA_ERROR_INVALID_CONTEXT: return "Invalid context"; + case CUDA_ERROR_CONTEXT_ALREADY_CURRENT: return "Context already current"; + case CUDA_ERROR_MAP_FAILED: return "Map failed"; + case CUDA_ERROR_UNMAP_FAILED: return "Unmap failed"; + case CUDA_ERROR_ARRAY_IS_MAPPED: return "Array is mapped"; + case CUDA_ERROR_ALREADY_MAPPED: return "Already mapped"; + case CUDA_ERROR_NO_BINARY_FOR_GPU: return "No binary for GPU"; + case CUDA_ERROR_ALREADY_ACQUIRED: return "Already acquired"; + case CUDA_ERROR_NOT_MAPPED: return "Not mapped"; + case CUDA_ERROR_NOT_MAPPED_AS_ARRAY: return "Mapped resource not available for access as an array"; + case CUDA_ERROR_NOT_MAPPED_AS_POINTER: return "Mapped resource not available for access as a pointer"; + case CUDA_ERROR_ECC_UNCORRECTABLE: return "Uncorrectable ECC error detected"; + case CUDA_ERROR_UNSUPPORTED_LIMIT: return "CUlimit not supported by device"; + + case CUDA_ERROR_INVALID_SOURCE: return "Invalid source"; + case CUDA_ERROR_FILE_NOT_FOUND: return "File not found"; + case CUDA_ERROR_SHARED_OBJECT_SYMBOL_NOT_FOUND: return "Link to a shared object failed to resolve"; + case CUDA_ERROR_SHARED_OBJECT_INIT_FAILED: return "Shared object initialization failed"; + + case CUDA_ERROR_INVALID_HANDLE: return "Invalid handle"; + + case CUDA_ERROR_NOT_FOUND: return "Not found"; + + case CUDA_ERROR_NOT_READY: return "CUDA not ready"; + + case CUDA_ERROR_LAUNCH_FAILED: return "Launch failed"; + case CUDA_ERROR_LAUNCH_OUT_OF_RESOURCES: return "Launch exceeded resources"; + case CUDA_ERROR_LAUNCH_TIMEOUT: return "Launch exceeded timeout"; + case CUDA_ERROR_LAUNCH_INCOMPATIBLE_TEXTURING: return "Launch with incompatible texturing"; + + case CUDA_ERROR_UNKNOWN: return "Unknown error"; + + default: return "Unknown CUDA error value"; + } + } + + static int cuda_align_up(int& offset, int alignment) + { + return (offset + alignment - 1) & ~(alignment - 1); + } + +#ifdef NDEBUG +#define cuda_abort() +#else +#define cuda_abort() abort() +#endif + +#define cuda_assert(stmt) \ + { \ + CUresult result = stmt; \ + \ + if(result != CUDA_SUCCESS) { \ + fprintf(stderr, "CUDA error: %s in %s\n", cuda_error_string(result), #stmt); \ + cuda_abort(); \ + } \ + } + + void cuda_push_context() + { + cuda_assert(cuCtxSetCurrent(cuContext)) + } + + void cuda_pop_context() + { + cuda_assert(cuCtxSetCurrent(NULL)); + } + + CUDADevice(bool background_) + { + int major, minor; + background = background_; + + cuDevId = 0; + + /* intialize */ + cuda_assert(cuInit(0)) + + /* setup device and context */ + cuda_assert(cuDeviceGet(&cuDevice, cuDevId)) + + if(background) + cuda_assert(cuCtxCreate(&cuContext, 0, cuDevice)) + else + cuda_assert(cuGLCtxCreate(&cuContext, 0, cuDevice)) + + /* open module */ + cuDeviceComputeCapability(&major, &minor, cuDevId); + string cubin = string_printf("lib/kernel_sm_%d%d.cubin", major, minor); + cuda_assert(cuModuleLoad(&cuModule, path_get(cubin).c_str())) + + cuda_pop_context(); + } + + ~CUDADevice() + { + cuda_push_context(); + cuda_assert(cuCtxDetach(cuContext)) + } + + string description() + { + /* print device information */ + char deviceName[100]; + + cuda_push_context(); + cuDeviceGetName(deviceName, 256, cuDevId); + cuda_pop_context(); + + return string("CUDA ") + deviceName; + } + + void mem_alloc(device_memory& mem, MemoryType type) + { + cuda_push_context(); + CUdeviceptr device_pointer; + cuda_assert(cuMemAlloc(&device_pointer, mem.memory_size())) + mem.device_pointer = (device_ptr)device_pointer; + cuda_pop_context(); + } + + void mem_copy_to(device_memory& mem) + { + cuda_push_context(); + cuda_assert(cuMemcpyHtoD(cuda_device_ptr(mem.device_pointer), (void*)mem.data_pointer, mem.memory_size())) + cuda_pop_context(); + } + + void mem_copy_from(device_memory& mem, size_t offset, size_t size) + { + /* todo: offset is ignored */ + cuda_push_context(); + cuda_assert(cuMemcpyDtoH((uchar*)mem.data_pointer + offset, + (CUdeviceptr)((uchar*)mem.device_pointer + offset), size)) + cuda_pop_context(); + } + + void mem_zero(device_memory& mem) + { + memset((void*)mem.data_pointer, 0, mem.memory_size()); + + cuda_push_context(); + cuda_assert(cuMemsetD8(cuda_device_ptr(mem.device_pointer), 0, mem.memory_size())) + cuda_pop_context(); + } + + void mem_free(device_memory& mem) + { + if(mem.device_pointer) { + cuda_push_context(); + cuda_assert(cuMemFree(cuda_device_ptr(mem.device_pointer))) + cuda_pop_context(); + + mem.device_pointer = 0; + } + } + + void const_copy_to(const char *name, void *host, size_t size) + { + CUdeviceptr mem; + size_t bytes; + + cuda_push_context(); + cuda_assert(cuModuleGetGlobal(&mem, &bytes, cuModule, name)) + assert(bytes == size); + cuda_assert(cuMemcpyHtoD(mem, host, size)) + cuda_pop_context(); + } + + void tex_alloc(const char *name, device_memory& mem, bool interpolation, bool periodic) + { + /* determine format */ + CUarray_format_enum format; + size_t dsize = datatype_size(mem.data_type); + size_t size = mem.memory_size(); + + switch(mem.data_type) { + case TYPE_UCHAR: format = CU_AD_FORMAT_UNSIGNED_INT8; break; + case TYPE_UINT: format = CU_AD_FORMAT_UNSIGNED_INT32; break; + case TYPE_INT: format = CU_AD_FORMAT_SIGNED_INT32; break; + case TYPE_FLOAT: format = CU_AD_FORMAT_FLOAT; break; + default: assert(0); return; + } + + CUtexref texref; + + cuda_push_context(); + cuda_assert(cuModuleGetTexRef(&texref, cuModule, name)) + + if(interpolation) { + CUarray handle; + CUDA_ARRAY_DESCRIPTOR desc; + + desc.Width = mem.data_width; + desc.Height = mem.data_height; + desc.Format = format; + desc.NumChannels = mem.data_elements; + + cuda_assert(cuArrayCreate(&handle, &desc)) + + if(mem.data_height > 1) { + CUDA_MEMCPY2D param; + memset(¶m, 0, sizeof(param)); + param.dstMemoryType = CU_MEMORYTYPE_ARRAY; + param.dstArray = handle; + param.srcMemoryType = CU_MEMORYTYPE_HOST; + param.srcHost = (void*)mem.data_pointer; + param.srcPitch = mem.data_width*dsize*mem.data_elements; + param.WidthInBytes = param.srcPitch; + param.Height = mem.data_height; + + cuda_assert(cuMemcpy2D(¶m)) + } + else + cuda_assert(cuMemcpyHtoA(handle, 0, (void*)mem.data_pointer, size)) + + cuda_assert(cuTexRefSetArray(texref, handle, CU_TRSA_OVERRIDE_FORMAT)) + + cuda_assert(cuTexRefSetFilterMode(texref, CU_TR_FILTER_MODE_LINEAR)) + cuda_assert(cuTexRefSetFlags(texref, CU_TRSF_NORMALIZED_COORDINATES)) + + mem.device_pointer = (device_ptr)handle; + } + else { + cuda_pop_context(); + + mem_alloc(mem, MEM_READ_ONLY); + mem_copy_to(mem); + + cuda_push_context(); + + cuda_assert(cuTexRefSetAddress(NULL, texref, cuda_device_ptr(mem.device_pointer), size)) + cuda_assert(cuTexRefSetFilterMode(texref, CU_TR_FILTER_MODE_POINT)) + cuda_assert(cuTexRefSetFlags(texref, CU_TRSF_READ_AS_INTEGER)) + } + + if(periodic) { + cuda_assert(cuTexRefSetAddressMode(texref, 0, CU_TR_ADDRESS_MODE_WRAP)) + cuda_assert(cuTexRefSetAddressMode(texref, 1, CU_TR_ADDRESS_MODE_WRAP)) + } + else { + cuda_assert(cuTexRefSetAddressMode(texref, 0, CU_TR_ADDRESS_MODE_CLAMP)) + cuda_assert(cuTexRefSetAddressMode(texref, 1, CU_TR_ADDRESS_MODE_CLAMP)) + } + cuda_assert(cuTexRefSetFormat(texref, format, mem.data_elements)) + + cuda_pop_context(); + + tex_interp_map[mem.device_pointer] = interpolation; + } + + void tex_free(device_memory& mem) + { + if(mem.device_pointer) { + if(tex_interp_map[mem.device_pointer]) { + cuda_push_context(); + cuArrayDestroy((CUarray)mem.device_pointer); + cuda_pop_context(); + + tex_interp_map.erase(tex_interp_map.find(mem.device_pointer)); + mem.device_pointer = 0; + } + else { + tex_interp_map.erase(tex_interp_map.find(mem.device_pointer)); + mem_free(mem); + } + } + } + + void path_trace(DeviceTask& task) + { + cuda_push_context(); + + CUfunction cuPathTrace; + CUdeviceptr d_buffer = cuda_device_ptr(task.buffer); + CUdeviceptr d_rng_state = cuda_device_ptr(task.rng_state); + + /* get kernel function */ + cuda_assert(cuModuleGetFunction(&cuPathTrace, cuModule, "kernel_cuda_path_trace")) + + /* pass in parameters */ + int offset = 0; + + cuda_assert(cuParamSetv(cuPathTrace, offset, &d_buffer, sizeof(d_buffer))) + offset += sizeof(d_buffer); + + cuda_assert(cuParamSetv(cuPathTrace, offset, &d_rng_state, sizeof(d_rng_state))) + offset += sizeof(d_rng_state); + + offset = cuda_align_up(offset, __alignof(task.pass)); + + cuda_assert(cuParamSeti(cuPathTrace, offset, task.pass)) + offset += sizeof(task.pass); + + cuda_assert(cuParamSeti(cuPathTrace, offset, task.x)) + offset += sizeof(task.x); + + cuda_assert(cuParamSeti(cuPathTrace, offset, task.y)) + offset += sizeof(task.y); + + cuda_assert(cuParamSeti(cuPathTrace, offset, task.w)) + offset += sizeof(task.w); + + cuda_assert(cuParamSeti(cuPathTrace, offset, task.h)) + offset += sizeof(task.h); + + cuda_assert(cuParamSetSize(cuPathTrace, offset)) + + /* launch kernel: todo find optimal size, cache config for fermi */ +#ifndef __APPLE__ + int xthreads = 16; + int ythreads = 16; +#else + int xthreads = 8; + int ythreads = 8; +#endif + int xblocks = (task.w + xthreads - 1)/xthreads; + int yblocks = (task.h + ythreads - 1)/ythreads; + + cuda_assert(cuFuncSetCacheConfig(cuPathTrace, CU_FUNC_CACHE_PREFER_L1)) + cuda_assert(cuFuncSetBlockShape(cuPathTrace, xthreads, ythreads, 1)) + cuda_assert(cuLaunchGrid(cuPathTrace, xblocks, yblocks)) + + cuda_pop_context(); + } + + void tonemap(DeviceTask& task) + { + cuda_push_context(); + + CUfunction cuFilmConvert; + CUdeviceptr d_rgba = map_pixels(task.rgba); + CUdeviceptr d_buffer = cuda_device_ptr(task.buffer); + + /* get kernel function */ + cuda_assert(cuModuleGetFunction(&cuFilmConvert, cuModule, "kernel_cuda_tonemap")) + + /* pass in parameters */ + int offset = 0; + + cuda_assert(cuParamSetv(cuFilmConvert, offset, &d_rgba, sizeof(d_rgba))) + offset += sizeof(d_rgba); + + cuda_assert(cuParamSetv(cuFilmConvert, offset, &d_buffer, sizeof(d_buffer))) + offset += sizeof(d_buffer); + + offset = cuda_align_up(offset, __alignof(task.pass)); + + cuda_assert(cuParamSeti(cuFilmConvert, offset, task.pass)) + offset += sizeof(task.pass); + + cuda_assert(cuParamSeti(cuFilmConvert, offset, task.resolution)) + offset += sizeof(task.resolution); + + cuda_assert(cuParamSeti(cuFilmConvert, offset, task.x)) + offset += sizeof(task.x); + + cuda_assert(cuParamSeti(cuFilmConvert, offset, task.y)) + offset += sizeof(task.y); + + cuda_assert(cuParamSeti(cuFilmConvert, offset, task.w)) + offset += sizeof(task.w); + + cuda_assert(cuParamSeti(cuFilmConvert, offset, task.h)) + offset += sizeof(task.h); + + cuda_assert(cuParamSetSize(cuFilmConvert, offset)) + + /* launch kernel: todo find optimal size, cache config for fermi */ +#ifndef __APPLE__ + int xthreads = 16; + int ythreads = 16; +#else + int xthreads = 8; + int ythreads = 8; +#endif + int xblocks = (task.w + xthreads - 1)/xthreads; + int yblocks = (task.h + ythreads - 1)/ythreads; + + cuda_assert(cuFuncSetCacheConfig(cuFilmConvert, CU_FUNC_CACHE_PREFER_L1)) + cuda_assert(cuFuncSetBlockShape(cuFilmConvert, xthreads, ythreads, 1)) + cuda_assert(cuLaunchGrid(cuFilmConvert, xblocks, yblocks)) + + unmap_pixels(task.rgba); + + cuda_pop_context(); + } + + void displace(DeviceTask& task) + { + cuda_push_context(); + + CUfunction cuDisplace; + CUdeviceptr d_input = cuda_device_ptr(task.displace_input); + CUdeviceptr d_offset = cuda_device_ptr(task.displace_offset); + + /* get kernel function */ + cuda_assert(cuModuleGetFunction(&cuDisplace, cuModule, "kernel_cuda_displace")) + + /* pass in parameters */ + int offset = 0; + + cuda_assert(cuParamSetv(cuDisplace, offset, &d_input, sizeof(d_input))) + offset += sizeof(d_input); + + cuda_assert(cuParamSetv(cuDisplace, offset, &d_offset, sizeof(d_offset))) + offset += sizeof(d_offset); + + offset = cuda_align_up(offset, __alignof(task.displace_x)); + + cuda_assert(cuParamSeti(cuDisplace, offset, task.displace_x)) + offset += sizeof(task.displace_x); + + cuda_assert(cuParamSetSize(cuDisplace, offset)) + + /* launch kernel: todo find optimal size, cache config for fermi */ +#ifndef __APPLE__ + int xthreads = 16; +#else + int xthreads = 8; +#endif + int xblocks = (task.displace_w + xthreads - 1)/xthreads; + + cuda_assert(cuFuncSetCacheConfig(cuDisplace, CU_FUNC_CACHE_PREFER_L1)) + cuda_assert(cuFuncSetBlockShape(cuDisplace, xthreads, 1, 1)) + cuda_assert(cuLaunchGrid(cuDisplace, xblocks, 1)) + + cuda_pop_context(); + } + + CUdeviceptr map_pixels(device_ptr mem) + { + if(!background) { + PixelMem pmem = pixel_mem_map[mem]; + CUdeviceptr buffer; + + size_t bytes; + cuda_assert(cuGraphicsMapResources(1, &pmem.cuPBOresource, 0)) + cuda_assert(cuGraphicsResourceGetMappedPointer(&buffer, &bytes, pmem.cuPBOresource)) + + return buffer; + } + + return cuda_device_ptr(mem); + } + + void unmap_pixels(device_ptr mem) + { + if(!background) { + PixelMem pmem = pixel_mem_map[mem]; + + cuda_assert(cuGraphicsUnmapResources(1, &pmem.cuPBOresource, 0)) + } + } + + void pixels_alloc(device_memory& mem) + { + if(!background) { + PixelMem pmem; + + pmem.w = mem.data_width; + pmem.h = mem.data_height; + + cuda_push_context(); + + glGenBuffers(1, &pmem.cuPBO); + glBindBuffer(GL_PIXEL_UNPACK_BUFFER, pmem.cuPBO); + glBufferData(GL_PIXEL_UNPACK_BUFFER, pmem.w*pmem.h*sizeof(GLfloat)*3, NULL, GL_DYNAMIC_DRAW); + + glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0); + + glGenTextures(1, &pmem.cuTexId); + glBindTexture(GL_TEXTURE_2D, pmem.cuTexId); + glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA, pmem.w, pmem.h, 0, GL_RGBA, GL_UNSIGNED_BYTE, NULL); + glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST); + glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST); + glBindTexture(GL_TEXTURE_2D, 0); + + cuda_assert(cuGraphicsGLRegisterBuffer(&pmem.cuPBOresource, pmem.cuPBO, CU_GRAPHICS_MAP_RESOURCE_FLAGS_NONE)) + + cuda_pop_context(); + + mem.device_pointer = pmem.cuTexId; + pixel_mem_map[mem.device_pointer] = pmem; + + return; + } + + Device::pixels_alloc(mem); + } + + void pixels_copy_from(device_memory& mem, int y, int w, int h) + { + if(!background) { + PixelMem pmem = pixel_mem_map[mem.device_pointer]; + + cuda_push_context(); + + glBindBuffer(GL_PIXEL_UNPACK_BUFFER, pmem.cuPBO); + uchar *pixels = (uchar*)glMapBuffer(GL_PIXEL_UNPACK_BUFFER, GL_READ_ONLY); + size_t offset = sizeof(uchar)*4*y*w; + memcpy((uchar*)mem.data_pointer + offset, pixels + offset, sizeof(uchar)*4*w*h); + glUnmapBuffer(GL_PIXEL_UNPACK_BUFFER); + glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0); + + cuda_pop_context(); + + return; + } + + Device::pixels_copy_from(mem, y, w, h); + } + + void pixels_free(device_memory& mem) + { + if(mem.device_pointer) { + if(!background) { + PixelMem pmem = pixel_mem_map[mem.device_pointer]; + + cuda_push_context(); + + cuda_assert(cuGraphicsUnregisterResource(pmem.cuPBOresource)) + glDeleteBuffers(1, &pmem.cuPBO); + glDeleteTextures(1, &pmem.cuTexId); + + cuda_pop_context(); + + pixel_mem_map.erase(pixel_mem_map.find(mem.device_pointer)); + mem.device_pointer = 0; + + return; + } + + Device::pixels_free(mem); + } + } + + void draw_pixels(device_memory& mem, int y, int w, int h, int width, int height) + { + if(!background) { + PixelMem pmem = pixel_mem_map[mem.device_pointer]; + + cuda_push_context(); + + glBindBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB, pmem.cuPBO); + glBindTexture(GL_TEXTURE_2D, pmem.cuTexId); + glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, w, h, GL_RGBA, GL_UNSIGNED_BYTE, 0); + glBindBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB, 0); + + glEnable(GL_TEXTURE_2D); + + glColor3f(1.0f, 1.0f, 1.0f); + + glPushMatrix(); + glTranslatef(0, y, 0.0f); + + glBegin(GL_QUADS); + + glTexCoord2f(0, 0); + glVertex2f(0, 0); + glTexCoord2f((float)w/(float)width, 0); + glVertex2f(width, 0); + glTexCoord2f((float)w/(float)width, (float)h/(float)height); + glVertex2f(width, height); + glTexCoord2f(0, (float)h/(float)height); + glVertex2f(0, height); + + glEnd(); + + glPopMatrix(); + + glBindTexture(GL_TEXTURE_2D, 0); + glDisable(GL_TEXTURE_2D); + + cuda_pop_context(); + + return; + } + + Device::draw_pixels(mem, y, w, h, width, height); + } + + void task_add(DeviceTask& task) + { + if(task.type == DeviceTask::TONEMAP) + tonemap(task); + else if(task.type == DeviceTask::PATH_TRACE) + path_trace(task); + else if(task.type == DeviceTask::DISPLACE) + displace(task); + } + + void task_wait() + { + cuda_push_context(); + + cuda_assert(cuCtxSynchronize()) + + cuda_pop_context(); + } + + void task_cancel() + { + } +}; + +Device *device_cuda_create(bool background) +{ + return new CUDADevice(background); +} + +CCL_NAMESPACE_END + |