diff options
Diffstat (limited to 'intern/cycles/device/device_cuda.cpp')
| -rw-r--r-- | intern/cycles/device/device_cuda.cpp | 212 |
1 files changed, 126 insertions, 86 deletions
diff --git a/intern/cycles/device/device_cuda.cpp b/intern/cycles/device/device_cuda.cpp index 2d404918a38..a85f34082db 100644 --- a/intern/cycles/device/device_cuda.cpp +++ b/intern/cycles/device/device_cuda.cpp @@ -249,121 +249,161 @@ public: return DebugFlags().cuda.adaptive_compile; } + /* Common NVCC flags which stays the same regardless of shading model, + * kernel sources md5 and only depends on compiler or compilation settings. + */ + string compile_kernel_get_common_cflags( + const DeviceRequestedFeatures& requested_features) + { + const int cuda_version = cuewCompilerVersion(); + const int machine = system_cpu_bits(); + const string kernel_path = path_get("kernel"); + const string include = kernel_path; + string cflags = string_printf("-m%d " + "--ptxas-options=\"-v\" " + "--use_fast_math " + "-DNVCC " + "-D__KERNEL_CUDA_VERSION__=%d " + "-I\"%s\"", + machine, + cuda_version, + include.c_str()); + if(use_adaptive_compilation()) { + cflags += " " + requested_features.get_build_options(); + } + const char *extra_cflags = getenv("CYCLES_CUDA_EXTRA_CFLAGS"); + if(extra_cflags) { + cflags += string(" ") + string(extra_cflags); + } +#ifdef WITH_CYCLES_DEBUG + cflags += " -D__KERNEL_DEBUG__"; +#endif + return cflags; + } + + bool compile_check_compiler() { + const char *nvcc = cuewCompilerPath(); + if(nvcc == NULL) { + cuda_error_message("CUDA nvcc compiler not found. " + "Install CUDA toolkit in default location."); + return false; + } + const int cuda_version = cuewCompilerVersion(); + VLOG(1) << "Found nvcc " << nvcc + << ", CUDA version " << cuda_version + << "."; + const int major = cuda_version / 10, minor = cuda_version & 10; + if(cuda_version == 0) { + cuda_error_message("CUDA nvcc compiler version could not be parsed."); + return false; + } + if(cuda_version < 75) { + printf("Unsupported CUDA version %d.%d detected, " + "you need CUDA 7.5 or newer.\n", + major, minor); + return false; + } + else if(cuda_version != 75 && cuda_version != 80) { + printf("CUDA version %d.%d detected, build may succeed but only " + "CUDA 7.5 and 8.0 are officially supported.\n", + major, minor); + } + return true; + } + string compile_kernel(const DeviceRequestedFeatures& requested_features) { /* Compute cubin name. */ int major, minor; cuDeviceComputeCapability(&major, &minor, cuDevId); - string cubin; - - /* Adaptive Compile. - * If enabled, always use that */ - bool use_adaptive_compile = use_adaptive_compilation(); /* Attempt to use kernel provided with Blender. */ - if(!use_adaptive_compile) { - cubin = path_get(string_printf("lib/kernel_sm_%d%d.cubin", major, minor)); - VLOG(1) << "Testing for pre-compiled kernel " << cubin; + if(!use_adaptive_compilation()) { + const string cubin = path_get(string_printf("lib/kernel_sm_%d%d.cubin", + major, minor)); + VLOG(1) << "Testing for pre-compiled kernel " << cubin << "."; if(path_exists(cubin)) { - VLOG(1) << "Using precompiled kernel"; + VLOG(1) << "Using precompiled kernel."; return cubin; } } + const string common_cflags = + compile_kernel_get_common_cflags(requested_features); + /* Try to use locally compiled kernel. */ - string kernel_path = path_get("kernel"); - string md5 = path_files_md5_hash(kernel_path); - - string feature_build_options; - if(use_adaptive_compile) { - feature_build_options = requested_features.get_build_options(); - string device_md5 = util_md5_string(feature_build_options); - cubin = string_printf("cycles_kernel_%s_sm%d%d_%s.cubin", - device_md5.c_str(), - major, minor, - md5.c_str()); - } - else { - cubin = string_printf("cycles_kernel_sm%d%d_%s.cubin", major, minor, md5.c_str()); - } + const string kernel_path = path_get("kernel"); + const string kernel_md5 = path_files_md5_hash(kernel_path); + + /* We include cflags into md5 so changing cuda toolkit or changing other + * compiler command line arguments makes sure cubin gets re-built. + */ + const string cubin_md5 = util_md5_string(kernel_md5 + common_cflags); - cubin = path_user_get(path_join("cache", cubin)); - VLOG(1) << "Testing for locally compiled kernel " << cubin; - /* If exists already, use it. */ + const string cubin_file = string_printf("cycles_kernel_sm%d%d_%s.cubin", + major, minor, + cubin_md5.c_str()); + const string cubin = path_user_get(path_join("cache", cubin_file)); + VLOG(1) << "Testing for locally compiled kernel " << cubin << "."; if(path_exists(cubin)) { - VLOG(1) << "Using locally compiled kernel"; + VLOG(1) << "Using locally compiled kernel."; return cubin; } #ifdef _WIN32 if(have_precompiled_kernels()) { - if(major < 2) - cuda_error_message(string_printf("CUDA device requires compute capability 2.0 or up, found %d.%d. Your GPU is not supported.", major, minor)); - else - cuda_error_message(string_printf("CUDA binary kernel for this graphics card compute capability (%d.%d) not found.", major, minor)); + if(major < 2) { + cuda_error_message(string_printf( + "CUDA device requires compute capability 2.0 or up, " + "found %d.%d. Your GPU is not supported.", + major, minor)); + } + else { + cuda_error_message(string_printf( + "CUDA binary kernel for this graphics card compute " + "capability (%d.%d) not found.", + major, minor)); + } return ""; } #endif - /* If not, find CUDA compiler. */ - const char *nvcc = cuewCompilerPath(); - - if(nvcc == NULL) { - cuda_error_message("CUDA nvcc compiler not found. Install CUDA toolkit in default location."); - return ""; - } - - int cuda_version = cuewCompilerVersion(); - VLOG(1) << "Found nvcc " << nvcc << ", CUDA version " << cuda_version; - - if(cuda_version == 0) { - cuda_error_message("CUDA nvcc compiler version could not be parsed."); - return ""; - } - if(cuda_version < 60) { - printf("Unsupported CUDA version %d.%d detected, you need CUDA 7.5.\n", cuda_version/10, cuda_version%10); + /* Compile. */ + if(!compile_check_compiler()) { return ""; } - else if(cuda_version != 75) - printf("CUDA version %d.%d detected, build may succeed but only CUDA 7.5 is officially supported.\n", cuda_version/10, cuda_version%10); - - /* Compile. */ - string kernel = path_join(kernel_path, path_join("kernels", path_join("cuda", "kernel.cu"))); - string include = kernel_path; - const int machine = system_cpu_bits(); - + const char *nvcc = cuewCompilerPath(); + const string kernel = path_join(kernel_path, + path_join("kernels", + path_join("cuda", "kernel.cu"))); double starttime = time_dt(); printf("Compiling CUDA kernel ...\n"); path_create_directories(cubin); - string command = string_printf("\"%s\" -arch=sm_%d%d -m%d --cubin \"%s\" " - "-o \"%s\" --ptxas-options=\"-v\" --use_fast_math -I\"%s\" " - "-DNVCC -D__KERNEL_CUDA_VERSION__=%d", - nvcc, major, minor, machine, kernel.c_str(), cubin.c_str(), include.c_str(), cuda_version); - - if(use_adaptive_compile) - command += " " + feature_build_options; - - const char* extra_cflags = getenv("CYCLES_CUDA_EXTRA_CFLAGS"); - if(extra_cflags) { - command += string(" ") + string(extra_cflags); - } - -#ifdef WITH_CYCLES_DEBUG - command += " -D__KERNEL_DEBUG__"; -#endif + string command = string_printf("\"%s\" " + "-arch=sm_%d%d " + "--cubin \"%s\" " + "-o \"%s\" " + "%s ", + nvcc, + major, minor, + kernel.c_str(), + cubin.c_str(), + common_cflags.c_str()); printf("%s\n", command.c_str()); if(system(command.c_str()) == -1) { - cuda_error_message("Failed to execute compilation command, see console for details."); + cuda_error_message("Failed to execute compilation command, " + "see console for details."); return ""; } /* Verify if compilation succeeded */ if(!path_exists(cubin)) { - cuda_error_message("CUDA kernel compilation failed, see console for details."); + cuda_error_message("CUDA kernel compilation failed, " + "see console for details."); return ""; } @@ -964,11 +1004,11 @@ public: 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; } @@ -1000,9 +1040,9 @@ public: glBufferData(GL_PIXEL_UNPACK_BUFFER, pmem.w*pmem.h*sizeof(GLhalf)*4, NULL, GL_DYNAMIC_DRAW); else glBufferData(GL_PIXEL_UNPACK_BUFFER, pmem.w*pmem.h*sizeof(uint8_t)*4, NULL, GL_DYNAMIC_DRAW); - + glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0); - + glGenTextures(1, &pmem.cuTexId); glBindTexture(GL_TEXTURE_2D, pmem.cuTexId); if(mem.data_type == TYPE_HALF) @@ -1012,7 +1052,7 @@ public: 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); - + CUresult result = cuGraphicsGLRegisterBuffer(&pmem.cuPBOresource, pmem.cuPBO, CU_GRAPHICS_MAP_RESOURCE_FLAGS_NONE); if(result == CUDA_SUCCESS) { @@ -1114,9 +1154,9 @@ public: else glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, w, h, GL_RGBA, GL_UNSIGNED_BYTE, (void*)offset); glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0); - + glEnable(GL_TEXTURE_2D); - + if(transparent) { glEnable(GL_BLEND); glBlendFunc(GL_ONE, GL_ONE_MINUS_SRC_ALPHA); @@ -1181,7 +1221,7 @@ public: if(transparent) glDisable(GL_BLEND); - + glBindTexture(GL_TEXTURE_2D, 0); glDisable(GL_TEXTURE_2D); @@ -1197,12 +1237,12 @@ public: { if(task->type == DeviceTask::PATH_TRACE) { RenderTile tile; - + bool branched = task->integrator_branched; /* Upload Bindless Mapping */ load_bindless_mapping(); - + /* keep rendering tiles until done */ while(task->acquire_tile(this, tile)) { int start_sample = tile.start_sample; @@ -1339,7 +1379,7 @@ void device_cuda_info(vector<DeviceInfo>& devices) fprintf(stderr, "CUDA cuDeviceGetCount: %s\n", cuewErrorString(result)); return; } - + vector<DeviceInfo> display_devices; for(int num = 0; num < count; num++) { |