static const char *
radv_get_compiler_string(struct radv_physical_device *pdevice)
{
- if (pdevice->use_aco) {
+ if (!pdevice->use_llvm) {
/* Some games like SotTR apply shader workarounds if the LLVM
* version is too old or if the LLVM version string is
* missing. This gives 2-5% performance with SotTR and ACO.
device->local_fd = fd;
device->ws->query_info(device->ws, &device->rad_info);
- device->use_aco = instance->perftest_flags & RADV_PERFTEST_ACO;
+ device->use_llvm = instance->debug_flags & RADV_DEBUG_LLVM;
snprintf(device->name, sizeof(device->name),
"AMD RADV %s (%s)",
}
/* These flags affect shader compilation. */
- uint64_t shader_env_flags = (device->use_aco ? 0x2 : 0);
+ uint64_t shader_env_flags = (device->use_llvm ? 0 : 0x2);
/* The gpu id is already embedded in the uuid so we just pass "radv"
* when creating the cache.
disk_cache_format_hex_id(buf, device->cache_uuid, VK_UUID_SIZE * 2);
device->disk_cache = disk_cache_create(device->name, buf, shader_env_flags);
- if (device->rad_info.chip_class < GFX8)
+ if (device->rad_info.chip_class < GFX8 || !device->use_llvm)
fprintf(stderr, "WARNING: radv is not a conformant vulkan implementation, testing use only.\n");
radv_get_driver_uuid(&device->driver_uuid);
device->dcc_msaa_allowed =
(device->instance->perftest_flags & RADV_PERFTEST_DCC_MSAA);
- device->use_shader_ballot = (device->use_aco && device->rad_info.chip_class >= GFX8) ||
- (device->instance->perftest_flags & RADV_PERFTEST_SHADER_BALLOT);
-
device->use_ngg = device->rad_info.chip_class >= GFX10 &&
device->rad_info.family != CHIP_NAVI14 &&
!(device->instance->debug_flags & RADV_DEBUG_NO_NGG);
/* TODO: Implement NGG GS with ACO. */
- device->use_ngg_gs = device->use_ngg && !device->use_aco;
+ device->use_ngg_gs = device->use_ngg && device->use_llvm;
device->use_ngg_streamout = false;
/* Determine the number of threads per wave for all stages. */
{"checkir", RADV_DEBUG_CHECKIR},
{"nothreadllvm", RADV_DEBUG_NOTHREADLLVM},
{"nobinning", RADV_DEBUG_NOBINNING},
- {"noloadstoreopt", RADV_DEBUG_NO_LOAD_STORE_OPT},
{"nongg", RADV_DEBUG_NO_NGG},
- {"noshaderballot", RADV_DEBUG_NO_SHADER_BALLOT},
{"allentrypoints", RADV_DEBUG_ALL_ENTRYPOINTS},
{"metashaders", RADV_DEBUG_DUMP_META_SHADERS},
{"nomemorycache", RADV_DEBUG_NO_MEMORY_CACHE},
+ {"llvm", RADV_DEBUG_LLVM},
{NULL, 0}
};
{"localbos", RADV_PERFTEST_LOCAL_BOS},
{"dccmsaa", RADV_PERFTEST_DCC_MSAA},
{"bolist", RADV_PERFTEST_BO_LIST},
- {"shader_ballot", RADV_PERFTEST_SHADER_BALLOT},
{"tccompatcmask", RADV_PERFTEST_TC_COMPAT_CMASK},
{"cswave32", RADV_PERFTEST_CS_WAVE_32},
{"pswave32", RADV_PERFTEST_PS_WAVE_32},
{"gewave32", RADV_PERFTEST_GE_WAVE_32},
{"dfsm", RADV_PERFTEST_DFSM},
- {"aco", RADV_PERFTEST_ACO},
{NULL, 0}
};
if (!strcmp(name, "DOOM_VFR")) {
/* Work around a Doom VFR game bug */
instance->debug_flags |= RADV_DEBUG_NO_DYNAMIC_BOUNDS;
- } else if (!strcmp(name, "MonsterHunterWorld.exe")) {
- /* Workaround for a WaW hazard when LLVM moves/merges
- * load/store memory operations.
- * See https://reviews.llvm.org/D61313
- */
- if (LLVM_VERSION_MAJOR < 9)
- instance->debug_flags |= RADV_DEBUG_NO_LOAD_STORE_OPT;
- } else if (!strcmp(name, "Wolfenstein: Youngblood")) {
- if (!(instance->debug_flags & RADV_DEBUG_NO_SHADER_BALLOT) &&
- !(instance->perftest_flags & RADV_PERFTEST_ACO)) {
- /* Force enable VK_AMD_shader_ballot because it looks
- * safe and it gives a nice boost (+20% on Vega 56 at
- * this time). It also prevents corruption on LLVM.
- */
- instance->perftest_flags |= RADV_PERFTEST_SHADER_BALLOT;
- }
} else if (!strcmp(name, "Fledge")) {
/*
* Zero VRAM for "The Surge 2"
* rendering issues.
*/
instance->debug_flags |= RADV_DEBUG_ZERO_VRAM;
+ } else if (!strcmp(engine_name, "Quantic Dream Engine")) {
+ /* Fix various artifacts in Detroit: Become Human */
+ instance->debug_flags |= RADV_DEBUG_ZERO_VRAM;
}
}
DRI_CONF_ADAPTIVE_SYNC("true")
DRI_CONF_VK_X11_OVERRIDE_MIN_IMAGE_COUNT(0)
DRI_CONF_VK_X11_STRICT_IMAGE_COUNT("false")
+ DRI_CONF_VK_X11_ENSURE_MIN_IMAGE_COUNT("false")
DRI_CONF_RADV_REPORT_LLVM9_VERSION_STRING("false")
DRI_CONF_RADV_ENABLE_MRT_OUTPUT_NAN_FIXUP("false")
DRI_CONF_RADV_NO_DYNAMIC_BOUNDS("false")
if (instance->apiVersion == 0)
instance->apiVersion = VK_API_VERSION_1_0;
- /* Get secure compile thread count. NOTE: We cap this at 32 */
-#define MAX_SC_PROCS 32
- char *num_sc_threads = getenv("RADV_SECURE_COMPILE_THREADS");
- if (num_sc_threads)
- instance->num_sc_threads = MIN2(strtoul(num_sc_threads, NULL, 10), MAX_SC_PROCS);
-
instance->debug_flags = parse_debug_string(getenv("RADV_DEBUG"),
radv_debug_options);
- /* Disable memory cache when secure compile is set */
- if (radv_device_use_secure_compile(instance))
- instance->debug_flags |= RADV_DEBUG_NO_MEMORY_CACHE;
-
instance->perftest_flags = parse_debug_string(getenv("RADV_PERFTEST"),
radv_perftest_options);
- if (instance->perftest_flags & RADV_PERFTEST_ACO)
- fprintf(stderr, "WARNING: Experimental compiler backend enabled. Here be dragons! Incorrect rendering, GPU hangs and/or resets are likely\n");
-
if (instance->debug_flags & RADV_DEBUG_STARTUP)
radv_logi("Created an instance");
if (idx >= RADV_INSTANCE_EXTENSION_COUNT ||
!radv_instance_extensions_supported.extensions[idx]) {
+ vk_object_base_finish(&instance->base);
vk_free2(&default_alloc, pAllocator, instance);
return vk_error(instance, VK_ERROR_EXTENSION_NOT_PRESENT);
}
result = vk_debug_report_instance_init(&instance->debug_report_callbacks);
if (result != VK_SUCCESS) {
+ vk_object_base_finish(&instance->base);
vk_free2(&default_alloc, pAllocator, instance);
return vk_error(instance, result);
}
f->storageBuffer16BitAccess = true;
f->uniformAndStorageBuffer16BitAccess = true;
f->storagePushConstant16 = true;
- f->storageInputOutput16 = pdevice->rad_info.has_packed_math_16bit && (LLVM_VERSION_MAJOR >= 9 || pdevice->use_aco);
+ f->storageInputOutput16 = pdevice->rad_info.has_packed_math_16bit && (LLVM_VERSION_MAJOR >= 9 || !pdevice->use_llvm);
f->multiview = true;
f->multiviewGeometryShader = true;
f->multiviewTessellationShader = true;
f->storageBuffer8BitAccess = true;
f->uniformAndStorageBuffer8BitAccess = true;
f->storagePushConstant8 = true;
- f->shaderBufferInt64Atomics = LLVM_VERSION_MAJOR >= 9 || pdevice->use_aco;
- f->shaderSharedInt64Atomics = LLVM_VERSION_MAJOR >= 9 || pdevice->use_aco;
+ f->shaderBufferInt64Atomics = LLVM_VERSION_MAJOR >= 9 || !pdevice->use_llvm;
+ f->shaderSharedInt64Atomics = LLVM_VERSION_MAJOR >= 9 || !pdevice->use_llvm;
f->shaderFloat16 = pdevice->rad_info.has_packed_math_16bit;
f->shaderInt8 = true;
case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DEMOTE_TO_HELPER_INVOCATION_FEATURES_EXT: {
VkPhysicalDeviceShaderDemoteToHelperInvocationFeaturesEXT *features =
(VkPhysicalDeviceShaderDemoteToHelperInvocationFeaturesEXT *)ext;
- features->shaderDemoteToHelperInvocation = LLVM_VERSION_MAJOR >= 9 || pdevice->use_aco;
+ features->shaderDemoteToHelperInvocation = LLVM_VERSION_MAJOR >= 9 || !pdevice->use_llvm;
break;
}
case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_FEATURES_EXT: {
features-> pipelineCreationCacheControl = true;
break;
}
+ case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTENDED_DYNAMIC_STATE_FEATURES_EXT: {
+ VkPhysicalDeviceExtendedDynamicStateFeaturesEXT *features =
+ (VkPhysicalDeviceExtendedDynamicStateFeaturesEXT *) ext;
+ features->extendedDynamicState = true;
+ break;
+ }
default:
break;
}
.maxCullDistances = 8,
.maxCombinedClipAndCullDistances = 8,
.discreteQueuePriorities = 2,
- .pointSizeRange = { 0.0, 8192.0 },
- .lineWidthRange = { 0.0, 8192.0 },
+ .pointSizeRange = { 0.0, 8191.875 },
+ .lineWidthRange = { 0.0, 8191.875 },
.pointSizeGranularity = (1.0 / 8.0),
.lineWidthGranularity = (1.0 / 8.0),
.strictLines = false, /* FINISHME */
return result;
}
-static int install_seccomp_filter() {
-
- struct sock_filter filter[] = {
- /* Check arch is 64bit x86 */
- BPF_STMT(BPF_LD + BPF_W + BPF_ABS, (offsetof(struct seccomp_data, arch))),
- BPF_JUMP(BPF_JMP + BPF_JEQ + BPF_K, AUDIT_ARCH_X86_64, 0, 12),
-
- /* Futex is required for mutex locks */
- #if defined __NR__newselect
- BPF_STMT(BPF_LD + BPF_W + BPF_ABS, (offsetof(struct seccomp_data, nr))),
- BPF_JUMP(BPF_JMP + BPF_JEQ + BPF_K, __NR__newselect, 11, 0),
- #elif defined __NR_select
- BPF_STMT(BPF_LD + BPF_W + BPF_ABS, (offsetof(struct seccomp_data, nr))),
- BPF_JUMP(BPF_JMP + BPF_JEQ + BPF_K, __NR_select, 11, 0),
- #else
- BPF_STMT(BPF_LD + BPF_W + BPF_ABS, (offsetof(struct seccomp_data, nr))),
- BPF_JUMP(BPF_JMP + BPF_JEQ + BPF_K, __NR_pselect6, 11, 0),
- #endif
-
- /* Allow system exit calls for the forked process */
- BPF_STMT(BPF_LD + BPF_W + BPF_ABS, (offsetof(struct seccomp_data, nr))),
- BPF_JUMP(BPF_JMP + BPF_JEQ + BPF_K, __NR_exit_group, 9, 0),
-
- /* Allow system read calls */
- BPF_STMT(BPF_LD + BPF_W + BPF_ABS, (offsetof(struct seccomp_data, nr))),
- BPF_JUMP(BPF_JMP + BPF_JEQ + BPF_K, __NR_read, 7, 0),
-
- /* Allow system write calls */
- BPF_STMT(BPF_LD + BPF_W + BPF_ABS, (offsetof(struct seccomp_data, nr))),
- BPF_JUMP(BPF_JMP + BPF_JEQ + BPF_K, __NR_write, 5, 0),
-
- /* Allow system brk calls (we need this for malloc) */
- BPF_STMT(BPF_LD + BPF_W + BPF_ABS, (offsetof(struct seccomp_data, nr))),
- BPF_JUMP(BPF_JMP + BPF_JEQ + BPF_K, __NR_brk, 3, 0),
-
- /* Futex is required for mutex locks */
- BPF_STMT(BPF_LD + BPF_W + BPF_ABS, (offsetof(struct seccomp_data, nr))),
- BPF_JUMP(BPF_JMP + BPF_JEQ + BPF_K, __NR_futex, 1, 0),
-
- /* Return error if we hit a system call not on the whitelist */
- BPF_STMT(BPF_RET + BPF_K, SECCOMP_RET_ERRNO | (EPERM & SECCOMP_RET_DATA)),
-
- /* Allow whitelisted system calls */
- BPF_STMT(BPF_RET + BPF_K, SECCOMP_RET_ALLOW),
- };
-
- struct sock_fprog prog = {
- .len = (unsigned short)(sizeof(filter) / sizeof(filter[0])),
- .filter = filter,
- };
-
- if (prctl(PR_SET_NO_NEW_PRIVS, 1, 0, 0, 0))
- return -1;
-
- if (prctl(PR_SET_SECCOMP, SECCOMP_MODE_FILTER, &prog))
- return -1;
-
- return 0;
-}
-
-/* Helper function with timeout support for reading from the pipe between
- * processes used for secure compile.
- */
-bool radv_sc_read(int fd, void *buf, size_t size, bool timeout)
-{
- fd_set fds;
- struct timeval tv;
-
- FD_ZERO(&fds);
- FD_SET(fd, &fds);
-
- while (true) {
- /* We can't rely on the value of tv after calling select() so
- * we must reset it on each iteration of the loop.
- */
- tv.tv_sec = 5;
- tv.tv_usec = 0;
-
- int rval = select(fd + 1, &fds, NULL, NULL, timeout ? &tv : NULL);
-
- if (rval == -1) {
- /* select error */
- return false;
- } else if (rval) {
- ssize_t bytes_read = read(fd, buf, size);
- if (bytes_read < 0)
- return false;
-
- buf += bytes_read;
- size -= bytes_read;
- if (size == 0)
- return true;
- } else {
- /* select timeout */
- return false;
- }
- }
-}
-
-static bool radv_close_all_fds(const int *keep_fds, int keep_fd_count)
-{
- DIR *d;
- struct dirent *dir;
- d = opendir("/proc/self/fd");
- if (!d)
- return false;
- int dir_fd = dirfd(d);
-
- while ((dir = readdir(d)) != NULL) {
- if (dir->d_name[0] == '.')
- continue;
-
- int fd = atoi(dir->d_name);
- if (fd == dir_fd)
- continue;
-
- bool keep = false;
- for (int i = 0; !keep && i < keep_fd_count; ++i)
- if (keep_fds[i] == fd)
- keep = true;
-
- if (keep)
- continue;
-
- close(fd);
- }
- closedir(d);
- return true;
-}
-
-static bool secure_compile_open_fifo_fds(struct radv_secure_compile_state *sc,
- int *fd_server, int *fd_client,
- unsigned process, bool make_fifo)
-{
- bool result = false;
- char *fifo_server_path = NULL;
- char *fifo_client_path = NULL;
-
- if (asprintf(&fifo_server_path, "/tmp/radv_server_%s_%u", sc->uid, process) == -1)
- goto open_fifo_exit;
-
- if (asprintf(&fifo_client_path, "/tmp/radv_client_%s_%u", sc->uid, process) == -1)
- goto open_fifo_exit;
-
- if (make_fifo) {
- int file1 = mkfifo(fifo_server_path, 0666);
- if(file1 < 0)
- goto open_fifo_exit;
-
- int file2 = mkfifo(fifo_client_path, 0666);
- if(file2 < 0)
- goto open_fifo_exit;
- }
-
- *fd_server = open(fifo_server_path, O_RDWR);
- if(*fd_server < 1)
- goto open_fifo_exit;
-
- *fd_client = open(fifo_client_path, O_RDWR);
- if(*fd_client < 1) {
- close(*fd_server);
- goto open_fifo_exit;
- }
-
- result = true;
-
-open_fifo_exit:
- free(fifo_server_path);
- free(fifo_client_path);
-
- return result;
-}
-
-static void run_secure_compile_device(struct radv_device *device, unsigned process,
- int fd_idle_device_output)
-{
- int fd_secure_input;
- int fd_secure_output;
- bool fifo_result = secure_compile_open_fifo_fds(device->sc_state,
- &fd_secure_input,
- &fd_secure_output,
- process, false);
-
- enum radv_secure_compile_type sc_type;
-
- const int needed_fds[] = {
- fd_secure_input,
- fd_secure_output,
- fd_idle_device_output,
- };
-
- if (!fifo_result || !radv_close_all_fds(needed_fds, ARRAY_SIZE(needed_fds)) ||
- install_seccomp_filter() == -1) {
- sc_type = RADV_SC_TYPE_INIT_FAILURE;
- } else {
- sc_type = RADV_SC_TYPE_INIT_SUCCESS;
- device->sc_state->secure_compile_processes[process].fd_secure_input = fd_secure_input;
- device->sc_state->secure_compile_processes[process].fd_secure_output = fd_secure_output;
- }
-
- write(fd_idle_device_output, &sc_type, sizeof(sc_type));
-
- if (sc_type == RADV_SC_TYPE_INIT_FAILURE)
- goto secure_compile_exit;
-
- while (true) {
- radv_sc_read(fd_secure_input, &sc_type, sizeof(sc_type), false);
-
- if (sc_type == RADV_SC_TYPE_COMPILE_PIPELINE) {
- struct radv_pipeline *pipeline;
- bool sc_read = true;
-
- pipeline = vk_zalloc2(&device->vk.alloc, NULL, sizeof(*pipeline), 8,
- VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
-
- pipeline->device = device;
-
- /* Read pipeline layout */
- struct radv_pipeline_layout layout;
- sc_read = radv_sc_read(fd_secure_input, &layout, sizeof(struct radv_pipeline_layout), true);
- sc_read &= radv_sc_read(fd_secure_input, &layout.num_sets, sizeof(uint32_t), true);
- if (!sc_read)
- goto secure_compile_exit;
-
- for (uint32_t set = 0; set < layout.num_sets; set++) {
- uint32_t layout_size;
- sc_read &= radv_sc_read(fd_secure_input, &layout_size, sizeof(uint32_t), true);
- if (!sc_read)
- goto secure_compile_exit;
-
- layout.set[set].layout = malloc(layout_size);
- layout.set[set].layout->layout_size = layout_size;
- sc_read &= radv_sc_read(fd_secure_input, layout.set[set].layout,
- layout.set[set].layout->layout_size, true);
- }
-
- pipeline->layout = &layout;
-
- /* Read pipeline key */
- struct radv_pipeline_key key;
- sc_read &= radv_sc_read(fd_secure_input, &key, sizeof(struct radv_pipeline_key), true);
-
- /* Read pipeline create flags */
- VkPipelineCreateFlags flags;
- sc_read &= radv_sc_read(fd_secure_input, &flags, sizeof(VkPipelineCreateFlags), true);
-
- /* Read stage and shader information */
- uint32_t num_stages;
- const VkPipelineShaderStageCreateInfo *pStages[MESA_SHADER_STAGES] = { 0, };
- sc_read &= radv_sc_read(fd_secure_input, &num_stages, sizeof(uint32_t), true);
- if (!sc_read)
- goto secure_compile_exit;
-
- for (uint32_t i = 0; i < num_stages; i++) {
-
- /* Read stage */
- gl_shader_stage stage;
- sc_read &= radv_sc_read(fd_secure_input, &stage, sizeof(gl_shader_stage), true);
-
- VkPipelineShaderStageCreateInfo *pStage = calloc(1, sizeof(VkPipelineShaderStageCreateInfo));
-
- /* Read entry point name */
- size_t name_size;
- sc_read &= radv_sc_read(fd_secure_input, &name_size, sizeof(size_t), true);
- if (!sc_read)
- goto secure_compile_exit;
-
- char *ep_name = malloc(name_size);
- sc_read &= radv_sc_read(fd_secure_input, ep_name, name_size, true);
- pStage->pName = ep_name;
-
- /* Read shader module */
- size_t module_size;
- sc_read &= radv_sc_read(fd_secure_input, &module_size, sizeof(size_t), true);
- if (!sc_read)
- goto secure_compile_exit;
-
- struct radv_shader_module *module = malloc(module_size);
- sc_read &= radv_sc_read(fd_secure_input, module, module_size, true);
- pStage->module = radv_shader_module_to_handle(module);
-
- /* Read specialization info */
- bool has_spec_info;
- sc_read &= radv_sc_read(fd_secure_input, &has_spec_info, sizeof(bool), true);
- if (!sc_read)
- goto secure_compile_exit;
-
- if (has_spec_info) {
- VkSpecializationInfo *specInfo = malloc(sizeof(VkSpecializationInfo));
- pStage->pSpecializationInfo = specInfo;
-
- sc_read &= radv_sc_read(fd_secure_input, &specInfo->dataSize, sizeof(size_t), true);
- if (!sc_read)
- goto secure_compile_exit;
-
- void *si_data = malloc(specInfo->dataSize);
- sc_read &= radv_sc_read(fd_secure_input, si_data, specInfo->dataSize, true);
- specInfo->pData = si_data;
-
- sc_read &= radv_sc_read(fd_secure_input, &specInfo->mapEntryCount, sizeof(uint32_t), true);
- if (!sc_read)
- goto secure_compile_exit;
-
- VkSpecializationMapEntry *mapEntries = malloc(sizeof(VkSpecializationMapEntry) * specInfo->mapEntryCount);
- for (uint32_t j = 0; j < specInfo->mapEntryCount; j++) {
- sc_read &= radv_sc_read(fd_secure_input, &mapEntries[j], sizeof(VkSpecializationMapEntry), true);
- if (!sc_read)
- goto secure_compile_exit;
- }
-
- specInfo->pMapEntries = mapEntries;
- }
-
- pStages[stage] = pStage;
- }
-
- /* Compile the shaders */
- VkPipelineCreationFeedbackEXT *stage_feedbacks[MESA_SHADER_STAGES] = { 0 };
-
- /* Not fully to spec but if we're doing sandboxed compilations already this doesn't matter. */
- flags &= ~VK_PIPELINE_CREATE_FAIL_ON_PIPELINE_COMPILE_REQUIRED_BIT_EXT;
-
- radv_create_shaders(pipeline, device, NULL, &key, pStages, flags, NULL, stage_feedbacks);
-
- /* free memory allocated above */
- for (uint32_t set = 0; set < layout.num_sets; set++)
- free(layout.set[set].layout);
-
- for (uint32_t i = 0; i < MESA_SHADER_STAGES; i++) {
- if (!pStages[i])
- continue;
-
- free((void *) pStages[i]->pName);
- free(radv_shader_module_from_handle(pStages[i]->module));
- if (pStages[i]->pSpecializationInfo) {
- free((void *) pStages[i]->pSpecializationInfo->pData);
- free((void *) pStages[i]->pSpecializationInfo->pMapEntries);
- free((void *) pStages[i]->pSpecializationInfo);
- }
- free((void *) pStages[i]);
- }
-
- vk_free(&device->vk.alloc, pipeline);
-
- sc_type = RADV_SC_TYPE_COMPILE_PIPELINE_FINISHED;
- write(fd_secure_output, &sc_type, sizeof(sc_type));
-
- } else if (sc_type == RADV_SC_TYPE_DESTROY_DEVICE) {
- goto secure_compile_exit;
- }
- }
-
-secure_compile_exit:
- close(fd_secure_input);
- close(fd_secure_output);
- close(fd_idle_device_output);
- _exit(0);
-}
-
-static enum radv_secure_compile_type fork_secure_compile_device(struct radv_device *device, unsigned process)
-{
- int fd_secure_input[2];
- int fd_secure_output[2];
-
- /* create pipe descriptors (used to communicate between processes) */
- if (pipe(fd_secure_input) == -1 || pipe(fd_secure_output) == -1)
- return RADV_SC_TYPE_INIT_FAILURE;
-
-
- int sc_pid;
- if ((sc_pid = fork()) == 0) {
- device->sc_state->secure_compile_thread_counter = process;
- run_secure_compile_device(device, process, fd_secure_output[1]);
- } else {
- if (sc_pid == -1)
- return RADV_SC_TYPE_INIT_FAILURE;
-
- /* Read the init result returned from the secure process */
- enum radv_secure_compile_type sc_type;
- bool sc_read = radv_sc_read(fd_secure_output[0], &sc_type, sizeof(sc_type), true);
-
- if (sc_type == RADV_SC_TYPE_INIT_FAILURE || !sc_read) {
- close(fd_secure_input[0]);
- close(fd_secure_input[1]);
- close(fd_secure_output[1]);
- close(fd_secure_output[0]);
- int status;
- waitpid(sc_pid, &status, 0);
-
- return RADV_SC_TYPE_INIT_FAILURE;
- } else {
- assert(sc_type == RADV_SC_TYPE_INIT_SUCCESS);
- write(device->sc_state->secure_compile_processes[process].fd_secure_output, &sc_type, sizeof(sc_type));
-
- close(fd_secure_input[0]);
- close(fd_secure_input[1]);
- close(fd_secure_output[1]);
- close(fd_secure_output[0]);
-
- int status;
- waitpid(sc_pid, &status, 0);
- }
- }
-
- return RADV_SC_TYPE_INIT_SUCCESS;
-}
-
-/* Run a bare bones fork of a device that was forked right after its creation.
- * This device will have low overhead when it is forked again before each
- * pipeline compilation. This device sits idle and its only job is to fork
- * itself.
- */
-static void run_secure_compile_idle_device(struct radv_device *device, unsigned process,
- int fd_secure_input, int fd_secure_output)
-{
- enum radv_secure_compile_type sc_type = RADV_SC_TYPE_INIT_SUCCESS;
- device->sc_state->secure_compile_processes[process].fd_secure_input = fd_secure_input;
- device->sc_state->secure_compile_processes[process].fd_secure_output = fd_secure_output;
-
- write(fd_secure_output, &sc_type, sizeof(sc_type));
-
- while (true) {
- radv_sc_read(fd_secure_input, &sc_type, sizeof(sc_type), false);
-
- if (sc_type == RADV_SC_TYPE_FORK_DEVICE) {
- sc_type = fork_secure_compile_device(device, process);
-
- if (sc_type == RADV_SC_TYPE_INIT_FAILURE)
- goto secure_compile_exit;
-
- } else if (sc_type == RADV_SC_TYPE_DESTROY_DEVICE) {
- goto secure_compile_exit;
- }
- }
-
-secure_compile_exit:
- close(fd_secure_input);
- close(fd_secure_output);
- _exit(0);
-}
-
-static void destroy_secure_compile_device(struct radv_device *device, unsigned process)
-{
- int fd_secure_input = device->sc_state->secure_compile_processes[process].fd_secure_input;
-
- enum radv_secure_compile_type sc_type = RADV_SC_TYPE_DESTROY_DEVICE;
- write(fd_secure_input, &sc_type, sizeof(sc_type));
-
- close(device->sc_state->secure_compile_processes[process].fd_secure_input);
- close(device->sc_state->secure_compile_processes[process].fd_secure_output);
-
- int status;
- waitpid(device->sc_state->secure_compile_processes[process].sc_pid, &status, 0);
-}
-
-static VkResult fork_secure_compile_idle_device(struct radv_device *device)
-{
- device->sc_state = vk_zalloc(&device->vk.alloc,
- sizeof(struct radv_secure_compile_state),
- 8, VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
-
- mtx_init(&device->sc_state->secure_compile_mutex, mtx_plain);
-
- pid_t upid = getpid();
- time_t seconds = time(NULL);
-
- char *uid;
- if (asprintf(&uid, "%ld_%ld", (long) upid, (long) seconds) == -1)
- return VK_ERROR_INITIALIZATION_FAILED;
-
- device->sc_state->uid = uid;
-
- uint8_t sc_threads = device->instance->num_sc_threads;
- int fd_secure_input[MAX_SC_PROCS][2];
- int fd_secure_output[MAX_SC_PROCS][2];
-
- /* create pipe descriptors (used to communicate between processes) */
- for (unsigned i = 0; i < sc_threads; i++) {
- if (pipe(fd_secure_input[i]) == -1 ||
- pipe(fd_secure_output[i]) == -1) {
- return VK_ERROR_INITIALIZATION_FAILED;
- }
- }
-
- device->sc_state->secure_compile_processes = vk_zalloc(&device->vk.alloc,
- sizeof(struct radv_secure_compile_process) * sc_threads, 8,
- VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
-
- for (unsigned process = 0; process < sc_threads; process++) {
- if ((device->sc_state->secure_compile_processes[process].sc_pid = fork()) == 0) {
- device->sc_state->secure_compile_thread_counter = process;
- run_secure_compile_idle_device(device, process, fd_secure_input[process][0], fd_secure_output[process][1]);
- } else {
- if (device->sc_state->secure_compile_processes[process].sc_pid == -1)
- return VK_ERROR_INITIALIZATION_FAILED;
-
- /* Read the init result returned from the secure process */
- enum radv_secure_compile_type sc_type;
- bool sc_read = radv_sc_read(fd_secure_output[process][0], &sc_type, sizeof(sc_type), true);
-
- bool fifo_result;
- if (sc_read && sc_type == RADV_SC_TYPE_INIT_SUCCESS) {
- fifo_result = secure_compile_open_fifo_fds(device->sc_state,
- &device->sc_state->secure_compile_processes[process].fd_server,
- &device->sc_state->secure_compile_processes[process].fd_client,
- process, true);
-
- device->sc_state->secure_compile_processes[process].fd_secure_input = fd_secure_input[process][1];
- device->sc_state->secure_compile_processes[process].fd_secure_output = fd_secure_output[process][0];
- }
-
- if (sc_type == RADV_SC_TYPE_INIT_FAILURE || !sc_read || !fifo_result) {
- close(fd_secure_input[process][0]);
- close(fd_secure_input[process][1]);
- close(fd_secure_output[process][1]);
- close(fd_secure_output[process][0]);
- int status;
- waitpid(device->sc_state->secure_compile_processes[process].sc_pid, &status, 0);
-
- /* Destroy any forks that were created sucessfully */
- for (unsigned i = 0; i < process; i++) {
- destroy_secure_compile_device(device, i);
- }
-
- return VK_ERROR_INITIALIZATION_FAILED;
- }
- }
- }
- return VK_SUCCESS;
-}
-
static void
radv_device_init_dispatch(struct radv_device *device)
{
goto fail;
}
- /* Temporarily disable secure compile while we create meta shaders, etc */
- uint8_t sc_threads = device->instance->num_sc_threads;
- if (sc_threads)
- device->instance->num_sc_threads = 0;
-
device->keep_shader_info = keep_shader_info;
result = radv_device_init_meta(device);
if (result != VK_SUCCESS)
1 << util_logbase2(device->force_aniso));
}
- /* Fork device for secure compile as required */
- device->instance->num_sc_threads = sc_threads;
- if (radv_device_use_secure_compile(device->instance)) {
-
- result = fork_secure_compile_idle_device(device);
- if (result != VK_SUCCESS)
- goto fail_meta;
- }
-
*pDevice = radv_device_to_handle(device);
return VK_SUCCESS;
radv_thread_trace_finish(device);
- if (radv_device_use_secure_compile(device->instance)) {
- for (unsigned i = 0; i < device->instance->num_sc_threads; i++ ) {
- destroy_secure_compile_device(device, i);
- }
- }
-
- if (device->sc_state) {
- free(device->sc_state->uid);
- vk_free(&device->vk.alloc, device->sc_state->secure_compile_processes);
- }
- vk_free(&device->vk.alloc, device->sc_state);
vk_free(&device->vk.alloc, device);
}
if (queue->device->trace_bo)
radv_cs_add_buffer(queue->device->ws, cs, queue->device->trace_bo);
+ if (queue->device->border_color_data.bo)
+ radv_cs_add_buffer(queue->device->ws, cs,
+ queue->device->border_color_data.bo);
+
if (i == 0) {
si_cs_emit_cache_flush(cs,
queue->device->physical_device->rad_info.chip_class,
RADV_CMD_FLAG_START_PIPELINE_STATS, 0);
}
- if (!queue->device->ws->cs_finalize(cs))
+ if (queue->device->ws->cs_finalize(cs) != VK_SUCCESS)
goto fail;
}
}
pthread_mutex_unlock(&queue->pending_mutex);
- queue->device->ws->ctx_wait_idle(queue->hw_ctx,
- radv_queue_family_to_ring(queue->queue_family_index),
- queue->queue_idx);
+ if (!queue->device->ws->ctx_wait_idle(queue->hw_ctx,
+ radv_queue_family_to_ring(queue->queue_family_index),
+ queue->queue_idx))
+ return VK_ERROR_DEVICE_LOST;
+
return VK_SUCCESS;
}
for (unsigned i = 0; i < RADV_MAX_QUEUE_FAMILIES; i++) {
for (unsigned q = 0; q < device->queue_count[i]; q++) {
- radv_QueueWaitIdle(radv_queue_to_handle(&device->queues[i][q]));
+ VkResult result =
+ radv_QueueWaitIdle(radv_queue_to_handle(&device->queues[i][q]));
+
+ if (result != VK_SUCCESS)
+ return result;
}
}
return VK_SUCCESS;
}
-static void radv_free_memory(struct radv_device *device,
- const VkAllocationCallbacks* pAllocator,
- struct radv_device_memory *mem)
+void
+radv_free_memory(struct radv_device *device,
+ const VkAllocationCallbacks* pAllocator,
+ struct radv_device_memory *mem)
{
if (mem == NULL)
return;
return VK_SUCCESS;
}
+static void
+radv_destroy_fence(struct radv_device *device,
+ const VkAllocationCallbacks *pAllocator,
+ struct radv_fence *fence)
+{
+ if (fence->temp_syncobj)
+ device->ws->destroy_syncobj(device->ws, fence->temp_syncobj);
+ if (fence->syncobj)
+ device->ws->destroy_syncobj(device->ws, fence->syncobj);
+ if (fence->fence)
+ device->ws->destroy_fence(fence->fence);
+ if (fence->fence_wsi)
+ fence->fence_wsi->destroy(fence->fence_wsi);
+
+ vk_object_base_finish(&fence->base);
+ vk_free2(&device->vk.alloc, pAllocator, fence);
+}
+
VkResult radv_CreateFence(
VkDevice _device,
const VkFenceCreateInfo* pCreateInfo,
if (device->always_use_syncobj || handleTypes) {
int ret = device->ws->create_syncobj(device->ws, &fence->syncobj);
if (ret) {
- vk_free2(&device->vk.alloc, pAllocator, fence);
+ radv_destroy_fence(device, pAllocator, fence);
return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
}
if (pCreateInfo->flags & VK_FENCE_CREATE_SIGNALED_BIT) {
} else {
fence->fence = device->ws->create_fence();
if (!fence->fence) {
- vk_free2(&device->vk.alloc, pAllocator, fence);
+ radv_destroy_fence(device, pAllocator, fence);
return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
}
fence->syncobj = 0;
if (!fence)
return;
- if (fence->temp_syncobj)
- device->ws->destroy_syncobj(device->ws, fence->temp_syncobj);
- if (fence->syncobj)
- device->ws->destroy_syncobj(device->ws, fence->syncobj);
- if (fence->fence)
- device->ws->destroy_fence(fence->fence);
- if (fence->fence_wsi)
- fence->fence_wsi->destroy(fence->fence_wsi);
-
- vk_object_base_finish(&fence->base);
- vk_free2(&device->vk.alloc, pAllocator, fence);
+ radv_destroy_fence(device, pAllocator, fence);
}
return type_info->semaphoreType;
}
+static void
+radv_destroy_semaphore(struct radv_device *device,
+ const VkAllocationCallbacks *pAllocator,
+ struct radv_semaphore *sem)
+{
+ radv_destroy_semaphore_part(device, &sem->temporary);
+ radv_destroy_semaphore_part(device, &sem->permanent);
+ vk_object_base_finish(&sem->base);
+ vk_free2(&device->vk.alloc, pAllocator, sem);
+}
+
VkResult radv_CreateSemaphore(
VkDevice _device,
const VkSemaphoreCreateInfo* pCreateInfo,
assert (device->physical_device->rad_info.has_syncobj);
int ret = device->ws->create_syncobj(device->ws, &sem->permanent.syncobj);
if (ret) {
- vk_free2(&device->vk.alloc, pAllocator, sem);
+ radv_destroy_semaphore(device, pAllocator, sem);
return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
}
sem->permanent.kind = RADV_SEMAPHORE_SYNCOBJ;
} else {
sem->permanent.ws_sem = device->ws->create_sem(device->ws);
if (!sem->permanent.ws_sem) {
- vk_free2(&device->vk.alloc, pAllocator, sem);
+ radv_destroy_semaphore(device, pAllocator, sem);
return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
}
sem->permanent.kind = RADV_SEMAPHORE_WINSYS;
if (!_semaphore)
return;
- radv_destroy_semaphore_part(device, &sem->temporary);
- radv_destroy_semaphore_part(device, &sem->permanent);
- vk_object_base_finish(&sem->base);
- vk_free2(&device->vk.alloc, pAllocator, sem);
+ radv_destroy_semaphore(device, pAllocator, sem);
}
VkResult
const VkAllocationCallbacks* pAllocator,
struct radv_event *event)
{
- device->ws->buffer_destroy(event->bo);
+ if (event->bo)
+ device->ws->buffer_destroy(event->bo);
+
vk_object_base_finish(&event->base);
vk_free2(&device->vk.alloc, pAllocator, event);
}
RADEON_FLAG_VA_UNCACHED | RADEON_FLAG_CPU_ACCESS | RADEON_FLAG_NO_INTERPROCESS_SHARING,
RADV_BO_PRIORITY_FENCE);
if (!event->bo) {
- vk_free2(&device->vk.alloc, pAllocator, event);
+ radv_destroy_event(device, pAllocator, event);
return vk_error(device->instance, VK_ERROR_OUT_OF_DEVICE_MEMORY);
}
return VK_SUCCESS;
}
+static void
+radv_destroy_buffer(struct radv_device *device,
+ const VkAllocationCallbacks *pAllocator,
+ struct radv_buffer *buffer)
+{
+ if ((buffer->flags & VK_BUFFER_CREATE_SPARSE_BINDING_BIT) && buffer->bo)
+ device->ws->buffer_destroy(buffer->bo);
+
+ vk_object_base_finish(&buffer->base);
+ vk_free2(&device->vk.alloc, pAllocator, buffer);
+}
+
VkResult radv_CreateBuffer(
VkDevice _device,
const VkBufferCreateInfo* pCreateInfo,
4096, 0, RADEON_FLAG_VIRTUAL,
RADV_BO_PRIORITY_VIRTUAL);
if (!buffer->bo) {
- vk_free2(&device->vk.alloc, pAllocator, buffer);
+ radv_destroy_buffer(device, pAllocator, buffer);
return vk_error(device->instance, VK_ERROR_OUT_OF_DEVICE_MEMORY);
}
}
if (!buffer)
return;
- if (buffer->flags & VK_BUFFER_CREATE_SPARSE_BINDING_BIT)
- device->ws->buffer_destroy(buffer->bo);
-
- vk_object_base_finish(&buffer->base);
- vk_free2(&device->vk.alloc, pAllocator, buffer);
+ radv_destroy_buffer(device, pAllocator, buffer);
}
VkDeviceAddress radv_GetBufferDeviceAddress(
if (!syncobj_handle) {
int ret = device->ws->create_syncobj(device->ws, &syncobj_handle);
if (ret) {
- return vk_error(device->instance, VK_ERROR_INVALID_EXTERNAL_HANDLE);
+ return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
}
}
switch(pGetFdInfo->handleType) {
case VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT:
ret = device->ws->export_syncobj(device->ws, syncobj_handle, pFd);
+ if (ret)
+ return vk_error(device->instance, VK_ERROR_TOO_MANY_OBJECTS);
break;
case VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT:
ret = device->ws->export_syncobj_to_sync_file(device->ws, syncobj_handle, pFd);
- if (!ret) {
- if (sem->temporary.kind != RADV_SEMAPHORE_NONE) {
- radv_destroy_semaphore_part(device, &sem->temporary);
- } else {
- device->ws->reset_syncobj(device->ws, syncobj_handle);
- }
+ if (ret)
+ return vk_error(device->instance, VK_ERROR_TOO_MANY_OBJECTS);
+
+ if (sem->temporary.kind != RADV_SEMAPHORE_NONE) {
+ radv_destroy_semaphore_part(device, &sem->temporary);
+ } else {
+ device->ws->reset_syncobj(device->ws, syncobj_handle);
}
break;
default:
unreachable("Unhandled semaphore handle type");
}
- if (ret)
- return vk_error(device->instance, VK_ERROR_INVALID_EXTERNAL_HANDLE);
return VK_SUCCESS;
}
switch(pGetFdInfo->handleType) {
case VK_EXTERNAL_FENCE_HANDLE_TYPE_OPAQUE_FD_BIT:
ret = device->ws->export_syncobj(device->ws, syncobj_handle, pFd);
+ if (ret)
+ return vk_error(device->instance, VK_ERROR_TOO_MANY_OBJECTS);
break;
case VK_EXTERNAL_FENCE_HANDLE_TYPE_SYNC_FD_BIT:
ret = device->ws->export_syncobj_to_sync_file(device->ws, syncobj_handle, pFd);
- if (!ret) {
- if (fence->temp_syncobj) {
- close (fence->temp_syncobj);
- fence->temp_syncobj = 0;
- } else {
- device->ws->reset_syncobj(device->ws, syncobj_handle);
- }
+ if (ret)
+ return vk_error(device->instance, VK_ERROR_TOO_MANY_OBJECTS);
+
+ if (fence->temp_syncobj) {
+ device->ws->destroy_syncobj(device->ws, fence->temp_syncobj);
+ fence->temp_syncobj = 0;
+ } else {
+ device->ws->reset_syncobj(device->ws, syncobj_handle);
}
break;
default:
unreachable("Unhandled fence handle type");
}
- if (ret)
- return vk_error(device->instance, VK_ERROR_INVALID_EXTERNAL_HANDLE);
return VK_SUCCESS;
}