#include <stdbool.h>
#include <string.h>
#include <sys/mman.h>
+#include <sys/sysinfo.h>
#include <unistd.h>
#include <fcntl.h>
#include <xf86drm.h>
#include "util/strtod.h"
#include "util/debug.h"
#include "util/build_id.h"
-#include "util/vk_util.h"
+#include "util/mesa-sha1.h"
+#include "vk_util.h"
#include "genxml/gen7_pack.h"
va_end(args);
}
-static bool
-anv_device_get_cache_uuid(void *uuid)
+static VkResult
+anv_compute_heap_size(int fd, uint64_t *heap_size)
+{
+ uint64_t gtt_size;
+ if (anv_gem_get_context_param(fd, 0, I915_CONTEXT_PARAM_GTT_SIZE,
+ >t_size) == -1) {
+ /* If, for whatever reason, we can't actually get the GTT size from the
+ * kernel (too old?) fall back to the aperture size.
+ */
+ anv_perf_warn("Failed to get I915_CONTEXT_PARAM_GTT_SIZE: %m");
+
+ if (anv_gem_get_aperture(fd, >t_size) == -1) {
+ return vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
+ "failed to get aperture size: %m");
+ }
+ }
+
+ /* Query the total ram from the system */
+ struct sysinfo info;
+ sysinfo(&info);
+
+ uint64_t total_ram = (uint64_t)info.totalram * (uint64_t)info.mem_unit;
+
+ /* We don't want to burn too much ram with the GPU. If the user has 4GiB
+ * or less, we use at most half. If they have more than 4GiB, we use 3/4.
+ */
+ uint64_t available_ram;
+ if (total_ram <= 4ull * 1024ull * 1024ull * 1024ull)
+ available_ram = total_ram / 2;
+ else
+ available_ram = total_ram * 3 / 4;
+
+ /* We also want to leave some padding for things we allocate in the driver,
+ * so don't go over 3/4 of the GTT either.
+ */
+ uint64_t available_gtt = gtt_size * 3 / 4;
+
+ *heap_size = MIN2(available_ram, available_gtt);
+
+ return VK_SUCCESS;
+}
+
+static VkResult
+anv_physical_device_init_heaps(struct anv_physical_device *device, int fd)
+{
+ /* The kernel query only tells us whether or not the kernel supports the
+ * EXEC_OBJECT_SUPPORTS_48B_ADDRESS flag and not whether or not the
+ * hardware has actual 48bit address support.
+ */
+ device->supports_48bit_addresses =
+ (device->info.gen >= 8) && anv_gem_supports_48b_addresses(fd);
+
+ uint64_t heap_size;
+ VkResult result = anv_compute_heap_size(fd, &heap_size);
+ if (result != VK_SUCCESS)
+ return result;
+
+ if (heap_size <= 3ull * (1ull << 30)) {
+ /* In this case, everything fits nicely into the 32-bit address space,
+ * so there's no need for supporting 48bit addresses on client-allocated
+ * memory objects.
+ */
+ device->memory.heap_count = 1;
+ device->memory.heaps[0] = (struct anv_memory_heap) {
+ .size = heap_size,
+ .flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT,
+ .supports_48bit_addresses = false,
+ };
+ } else {
+ /* Not everything will fit nicely into a 32-bit address space. In this
+ * case we need a 64-bit heap. Advertise a small 32-bit heap and a
+ * larger 48-bit heap. If we're in this case, then we have a total heap
+ * size larger than 3GiB which most likely means they have 8 GiB of
+ * video memory and so carving off 1 GiB for the 32-bit heap should be
+ * reasonable.
+ */
+ const uint64_t heap_size_32bit = 1ull << 30;
+ const uint64_t heap_size_48bit = heap_size - heap_size_32bit;
+
+ assert(device->supports_48bit_addresses);
+
+ device->memory.heap_count = 2;
+ device->memory.heaps[0] = (struct anv_memory_heap) {
+ .size = heap_size_48bit,
+ .flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT,
+ .supports_48bit_addresses = true,
+ };
+ device->memory.heaps[1] = (struct anv_memory_heap) {
+ .size = heap_size_32bit,
+ .flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT,
+ .supports_48bit_addresses = false,
+ };
+ }
+
+ uint32_t type_count = 0;
+ for (uint32_t heap = 0; heap < device->memory.heap_count; heap++) {
+ uint32_t valid_buffer_usage = ~0;
+
+ /* There appears to be a hardware issue in the VF cache where it only
+ * considers the bottom 32 bits of memory addresses. If you happen to
+ * have two vertex buffers which get placed exactly 4 GiB apart and use
+ * them in back-to-back draw calls, you can get collisions. In order to
+ * solve this problem, we require vertex and index buffers be bound to
+ * memory allocated out of the 32-bit heap.
+ */
+ if (device->memory.heaps[heap].supports_48bit_addresses) {
+ valid_buffer_usage &= ~(VK_BUFFER_USAGE_INDEX_BUFFER_BIT |
+ VK_BUFFER_USAGE_VERTEX_BUFFER_BIT);
+ }
+
+ if (device->info.has_llc) {
+ /* Big core GPUs share LLC with the CPU and thus one memory type can be
+ * both cached and coherent at the same time.
+ */
+ device->memory.types[type_count++] = (struct anv_memory_type) {
+ .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
+ VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
+ VK_MEMORY_PROPERTY_HOST_COHERENT_BIT |
+ VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
+ .heapIndex = heap,
+ .valid_buffer_usage = valid_buffer_usage,
+ };
+ } else {
+ /* The spec requires that we expose a host-visible, coherent memory
+ * type, but Atom GPUs don't share LLC. Thus we offer two memory types
+ * to give the application a choice between cached, but not coherent and
+ * coherent but uncached (WC though).
+ */
+ device->memory.types[type_count++] = (struct anv_memory_type) {
+ .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
+ VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
+ VK_MEMORY_PROPERTY_HOST_COHERENT_BIT,
+ .heapIndex = heap,
+ .valid_buffer_usage = valid_buffer_usage,
+ };
+ device->memory.types[type_count++] = (struct anv_memory_type) {
+ .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
+ VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
+ VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
+ .heapIndex = heap,
+ .valid_buffer_usage = valid_buffer_usage,
+ };
+ }
+ }
+ device->memory.type_count = type_count;
+
+ return VK_SUCCESS;
+}
+
+static VkResult
+anv_physical_device_init_uuids(struct anv_physical_device *device)
{
const struct build_id_note *note = build_id_find_nhdr("libvulkan_intel.so");
- if (!note)
- return false;
+ if (!note) {
+ return vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
+ "Failed to find build-id");
+ }
+
+ unsigned build_id_len = build_id_length(note);
+ if (build_id_len < 20) {
+ return vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
+ "build-id too short. It needs to be a SHA");
+ }
+
+ struct mesa_sha1 sha1_ctx;
+ uint8_t sha1[20];
+ STATIC_ASSERT(VK_UUID_SIZE <= sizeof(sha1));
- unsigned len = build_id_length(note);
- if (len < VK_UUID_SIZE)
- return false;
+ /* The pipeline cache UUID is used for determining when a pipeline cache is
+ * invalid. It needs both a driver build and the PCI ID of the device.
+ */
+ _mesa_sha1_init(&sha1_ctx);
+ _mesa_sha1_update(&sha1_ctx, build_id_data(note), build_id_len);
+ _mesa_sha1_update(&sha1_ctx, &device->chipset_id,
+ sizeof(device->chipset_id));
+ _mesa_sha1_final(&sha1_ctx, sha1);
+ memcpy(device->pipeline_cache_uuid, sha1, VK_UUID_SIZE);
+
+ /* The driver UUID is used for determining sharability of images and memory
+ * between two Vulkan instances in separate processes. People who want to
+ * share memory need to also check the device UUID (below) so all this
+ * needs to be is the build-id.
+ */
+ memcpy(device->driver_uuid, build_id_data(note), VK_UUID_SIZE);
+
+ /* The device UUID uniquely identifies the given device within the machine.
+ * Since we never have more than one device, this doesn't need to be a real
+ * UUID. However, on the off-chance that someone tries to use this to
+ * cache pre-tiled images or something of the like, we use the PCI ID and
+ * some bits of ISL info to ensure that this is safe.
+ */
+ _mesa_sha1_init(&sha1_ctx);
+ _mesa_sha1_update(&sha1_ctx, &device->chipset_id,
+ sizeof(device->chipset_id));
+ _mesa_sha1_update(&sha1_ctx, &device->isl_dev.has_bit6_swizzling,
+ sizeof(device->isl_dev.has_bit6_swizzling));
+ _mesa_sha1_final(&sha1_ctx, sha1);
+ memcpy(device->device_uuid, sha1, VK_UUID_SIZE);
- memcpy(uuid, build_id_data(note), VK_UUID_SIZE);
- return true;
+ return VK_SUCCESS;
}
static VkResult
fprintf(stderr, "WARNING: Ivy Bridge Vulkan support is incomplete\n");
} else if (device->info.gen == 7 && device->info.is_baytrail) {
fprintf(stderr, "WARNING: Bay Trail Vulkan support is incomplete\n");
- } else if (device->info.gen >= 8) {
+ } else if (device->info.gen >= 8 && device->info.gen <= 9) {
/* Broadwell, Cherryview, Skylake, Broxton, Kabylake is as fully
* supported as anything */
} else {
}
}
- if (anv_gem_get_aperture(fd, &device->aperture_size) == -1) {
- result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
- "failed to get aperture size: %m");
- goto fail;
- }
-
if (!anv_gem_get_param(fd, I915_PARAM_HAS_WAIT_TIMEOUT)) {
result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
"kernel missing gem wait");
goto fail;
}
- device->supports_48bit_addresses = anv_gem_supports_48b_addresses(fd);
-
- if (!anv_device_get_cache_uuid(device->uuid)) {
- result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
- "cannot generate UUID");
+ result = anv_physical_device_init_heaps(device, fd);
+ if (result != VK_SUCCESS)
goto fail;
- }
+
+ device->has_exec_async = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_ASYNC);
+
bool swizzled = anv_gem_get_bit6_swizzle(fd, I915_TILING_X);
/* GENs prior to 8 do not support EU/Subslice info */
if (device->info.is_cherryview &&
device->subslice_total > 0 && device->eu_total > 0) {
- /* Logical CS threads = EUs per subslice * 7 threads per EU */
- uint32_t max_cs_threads = device->eu_total / device->subslice_total * 7;
+ /* Logical CS threads = EUs per subslice * num threads per EU */
+ uint32_t max_cs_threads =
+ device->eu_total / device->subslice_total * device->info.num_thread_per_eu;
/* Fuse configurations may give more threads than expected, never less. */
if (max_cs_threads > device->info.max_cs_threads)
device->compiler->shader_debug_log = compiler_debug_log;
device->compiler->shader_perf_log = compiler_perf_log;
+ isl_device_init(&device->isl_dev, &device->info, swizzled);
+
+ result = anv_physical_device_init_uuids(device);
+ if (result != VK_SUCCESS)
+ goto fail;
+
result = anv_init_wsi(device);
if (result != VK_SUCCESS) {
ralloc_free(device->compiler);
goto fail;
}
- isl_device_init(&device->isl_dev, &device->info, swizzled);
-
device->local_fd = fd;
return VK_SUCCESS;
}
static const VkExtensionProperties global_extensions[] = {
+ {
+ .extensionName = VK_KHR_GET_PHYSICAL_DEVICE_PROPERTIES_2_EXTENSION_NAME,
+ .specVersion = 1,
+ },
+ {
+ .extensionName = VK_KHR_GET_SURFACE_CAPABILITIES_2_EXTENSION_NAME,
+ .specVersion = 1,
+ },
{
.extensionName = VK_KHR_SURFACE_EXTENSION_NAME,
.specVersion = 25,
},
+#ifdef VK_USE_PLATFORM_WAYLAND_KHR
+ {
+ .extensionName = VK_KHR_WAYLAND_SURFACE_EXTENSION_NAME,
+ .specVersion = 5,
+ },
+#endif
#ifdef VK_USE_PLATFORM_XCB_KHR
{
.extensionName = VK_KHR_XCB_SURFACE_EXTENSION_NAME,
.specVersion = 6,
},
#endif
-#ifdef VK_USE_PLATFORM_WAYLAND_KHR
- {
- .extensionName = VK_KHR_WAYLAND_SURFACE_EXTENSION_NAME,
- .specVersion = 5,
- },
-#endif
- {
- .extensionName = VK_KHR_GET_PHYSICAL_DEVICE_PROPERTIES_2_EXTENSION_NAME,
- .specVersion = 1,
- },
};
static const VkExtensionProperties device_extensions[] = {
{
- .extensionName = VK_KHR_SWAPCHAIN_EXTENSION_NAME,
- .specVersion = 68,
+ .extensionName = VK_KHR_DESCRIPTOR_UPDATE_TEMPLATE_EXTENSION_NAME,
+ .specVersion = 1,
},
{
- .extensionName = VK_KHR_SAMPLER_MIRROR_CLAMP_TO_EDGE_EXTENSION_NAME,
+ .extensionName = VK_KHR_GET_MEMORY_REQUIREMENTS_2_EXTENSION_NAME,
.specVersion = 1,
},
{
- .extensionName = VK_KHR_MAINTENANCE1_EXTENSION_NAME,
+ .extensionName = VK_KHR_INCREMENTAL_PRESENT_EXTENSION_NAME,
.specVersion = 1,
},
{
- .extensionName = VK_KHR_SHADER_DRAW_PARAMETERS_EXTENSION_NAME,
+ .extensionName = VK_KHR_MAINTENANCE1_EXTENSION_NAME,
.specVersion = 1,
},
{
.specVersion = 1,
},
{
- .extensionName = VK_KHR_DESCRIPTOR_UPDATE_TEMPLATE_EXTENSION_NAME,
+ .extensionName = VK_KHR_SAMPLER_MIRROR_CLAMP_TO_EDGE_EXTENSION_NAME,
.specVersion = 1,
},
{
- .extensionName = VK_KHR_INCREMENTAL_PRESENT_EXTENSION_NAME,
+ .extensionName = VK_KHR_SHADER_DRAW_PARAMETERS_EXTENSION_NAME,
+ .specVersion = 1,
+ },
+ {
+ .extensionName = VK_KHR_SWAPCHAIN_EXTENSION_NAME,
+ .specVersion = 68,
+ },
+ {
+ .extensionName = VK_KHX_MULTIVIEW_EXTENSION_NAME,
.specVersion = 1,
},
};
instance->physicalDeviceCount = 0;
- max_devices = drmGetDevices2(0, devices, sizeof(devices));
+ max_devices = drmGetDevices2(0, devices, ARRAY_SIZE(devices));
if (max_devices < 1)
return VK_ERROR_INCOMPATIBLE_DRIVER;
break;
}
}
+ drmFreeDevices(devices, max_devices);
if (result == VK_SUCCESS)
instance->physicalDeviceCount = 1;
.sampleRateShading = true,
.dualSrcBlend = true,
.logicOp = true,
- .multiDrawIndirect = false,
+ .multiDrawIndirect = true,
.drawIndirectFirstInstance = true,
.depthClamp = true,
.depthBiasClamp = true,
vk_foreach_struct(ext, pFeatures->pNext) {
switch (ext->sType) {
+ case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES_KHX: {
+ VkPhysicalDeviceMultiviewFeaturesKHX *features =
+ (VkPhysicalDeviceMultiviewFeaturesKHX *)ext;
+ features->multiview = true;
+ features->multiviewGeometryShader = true;
+ features->multiviewTessellationShader = true;
+ break;
+ }
+
default:
anv_debug_ignored_stype(ext->sType);
break;
const uint32_t max_raw_buffer_sz = devinfo->gen >= 7 ?
(1ul << 30) : (1ul << 27);
+ const uint32_t max_samplers = (devinfo->gen >= 8 || devinfo->is_haswell) ?
+ 128 : 16;
+
VkSampleCountFlags sample_counts =
isl_device_get_sample_counts(&pdevice->isl_dev);
.bufferImageGranularity = 64, /* A cache line */
.sparseAddressSpaceSize = 0,
.maxBoundDescriptorSets = MAX_SETS,
- .maxPerStageDescriptorSamplers = 64,
+ .maxPerStageDescriptorSamplers = max_samplers,
.maxPerStageDescriptorUniformBuffers = 64,
.maxPerStageDescriptorStorageBuffers = 64,
- .maxPerStageDescriptorSampledImages = 64,
+ .maxPerStageDescriptorSampledImages = max_samplers,
.maxPerStageDescriptorStorageImages = 64,
.maxPerStageDescriptorInputAttachments = 64,
- .maxPerStageResources = 128,
+ .maxPerStageResources = 250,
.maxDescriptorSetSamplers = 256,
.maxDescriptorSetUniformBuffers = 256,
.maxDescriptorSetUniformBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2,
.storageImageSampleCounts = VK_SAMPLE_COUNT_1_BIT,
.maxSampleMaskWords = 1,
.timestampComputeAndGraphics = false,
- .timestampPeriod = devinfo->timebase_scale,
+ .timestampPeriod = 1000000000.0 / devinfo->timestamp_frequency,
.maxClipDistances = 8,
.maxCullDistances = 8,
.maxCombinedClipAndCullDistances = 8,
};
*pProperties = (VkPhysicalDeviceProperties) {
- .apiVersion = VK_MAKE_VERSION(1, 0, 42),
- .driverVersion = 1,
+ .apiVersion = VK_MAKE_VERSION(1, 0, 54),
+ .driverVersion = vk_get_driver_version(),
.vendorID = 0x8086,
.deviceID = pdevice->chipset_id,
.deviceType = VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU,
.sparseProperties = {0}, /* Broadwell doesn't do sparse. */
};
- strcpy(pProperties->deviceName, pdevice->name);
- memcpy(pProperties->pipelineCacheUUID, pdevice->uuid, VK_UUID_SIZE);
+ strncpy(pProperties->deviceName, pdevice->name,
+ VK_MAX_PHYSICAL_DEVICE_NAME_SIZE);
+ memcpy(pProperties->pipelineCacheUUID,
+ pdevice->pipeline_cache_uuid, VK_UUID_SIZE);
}
void anv_GetPhysicalDeviceProperties2KHR(
break;
}
+ case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES_KHX: {
+ VkPhysicalDeviceMultiviewPropertiesKHX *properties =
+ (VkPhysicalDeviceMultiviewPropertiesKHX *)ext;
+ properties->maxMultiviewViewCount = 16;
+ properties->maxMultiviewInstanceIndex = UINT32_MAX / 16;
+ break;
+ }
+
default:
anv_debug_ignored_stype(ext->sType);
break;
VkPhysicalDeviceMemoryProperties* pMemoryProperties)
{
ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
- VkDeviceSize heap_size;
-
- /* Reserve some wiggle room for the driver by exposing only 75% of the
- * aperture to the heap.
- */
- heap_size = 3 * physical_device->aperture_size / 4;
- if (physical_device->info.has_llc) {
- /* Big core GPUs share LLC with the CPU and thus one memory type can be
- * both cached and coherent at the same time.
- */
- pMemoryProperties->memoryTypeCount = 1;
- pMemoryProperties->memoryTypes[0] = (VkMemoryType) {
- .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
- VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
- VK_MEMORY_PROPERTY_HOST_COHERENT_BIT |
- VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
- .heapIndex = 0,
- };
- } else {
- /* The spec requires that we expose a host-visible, coherent memory
- * type, but Atom GPUs don't share LLC. Thus we offer two memory types
- * to give the application a choice between cached, but not coherent and
- * coherent but uncached (WC though).
- */
- pMemoryProperties->memoryTypeCount = 2;
- pMemoryProperties->memoryTypes[0] = (VkMemoryType) {
- .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
- VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
- VK_MEMORY_PROPERTY_HOST_COHERENT_BIT,
- .heapIndex = 0,
- };
- pMemoryProperties->memoryTypes[1] = (VkMemoryType) {
- .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
- VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
- VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
- .heapIndex = 0,
+ pMemoryProperties->memoryTypeCount = physical_device->memory.type_count;
+ for (uint32_t i = 0; i < physical_device->memory.type_count; i++) {
+ pMemoryProperties->memoryTypes[i] = (VkMemoryType) {
+ .propertyFlags = physical_device->memory.types[i].propertyFlags,
+ .heapIndex = physical_device->memory.types[i].heapIndex,
};
}
- pMemoryProperties->memoryHeapCount = 1;
- pMemoryProperties->memoryHeaps[0] = (VkMemoryHeap) {
- .size = heap_size,
- .flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT,
- };
+ pMemoryProperties->memoryHeapCount = physical_device->memory.heap_count;
+ for (uint32_t i = 0; i < physical_device->memory.heap_count; i++) {
+ pMemoryProperties->memoryHeaps[i] = (VkMemoryHeap) {
+ .size = physical_device->memory.heaps[i].size,
+ .flags = physical_device->memory.heaps[i].flags,
+ };
+ }
}
void anv_GetPhysicalDeviceMemoryProperties2KHR(
state = anv_state_pool_alloc(pool, size, align);
memcpy(state.map, p, size);
- anv_state_flush(pool->block_pool->device, state);
+ anv_state_flush(pool->block_pool.device, state);
return state;
}
border_colors);
}
-VkResult
-anv_device_submit_simple_batch(struct anv_device *device,
- struct anv_batch *batch)
-{
- struct drm_i915_gem_execbuffer2 execbuf;
- struct drm_i915_gem_exec_object2 exec2_objects[1];
- struct anv_bo bo, *exec_bos[1];
- VkResult result = VK_SUCCESS;
- uint32_t size;
-
- /* Kernel driver requires 8 byte aligned batch length */
- size = align_u32(batch->next - batch->start, 8);
- result = anv_bo_pool_alloc(&device->batch_bo_pool, &bo, size);
- if (result != VK_SUCCESS)
- return result;
-
- memcpy(bo.map, batch->start, size);
- if (!device->info.has_llc)
- anv_flush_range(bo.map, size);
-
- exec_bos[0] = &bo;
- exec2_objects[0].handle = bo.gem_handle;
- exec2_objects[0].relocation_count = 0;
- exec2_objects[0].relocs_ptr = 0;
- exec2_objects[0].alignment = 0;
- exec2_objects[0].offset = bo.offset;
- exec2_objects[0].flags = 0;
- exec2_objects[0].rsvd1 = 0;
- exec2_objects[0].rsvd2 = 0;
-
- execbuf.buffers_ptr = (uintptr_t) exec2_objects;
- execbuf.buffer_count = 1;
- execbuf.batch_start_offset = 0;
- execbuf.batch_len = size;
- execbuf.cliprects_ptr = 0;
- execbuf.num_cliprects = 0;
- execbuf.DR1 = 0;
- execbuf.DR4 = 0;
-
- execbuf.flags =
- I915_EXEC_HANDLE_LUT | I915_EXEC_NO_RELOC | I915_EXEC_RENDER;
- execbuf.rsvd1 = device->context_id;
- execbuf.rsvd2 = 0;
-
- result = anv_device_execbuf(device, &execbuf, exec_bos);
- if (result != VK_SUCCESS)
- goto fail;
-
- result = anv_device_wait(device, &bo, INT64_MAX);
-
- fail:
- anv_bo_pool_free(&device->batch_bo_pool, &bo);
-
- return result;
-}
-
VkResult anv_CreateDevice(
VkPhysicalDevice physicalDevice,
const VkDeviceCreateInfo* pCreateInfo,
return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
}
+ /* Check enabled features */
+ if (pCreateInfo->pEnabledFeatures) {
+ VkPhysicalDeviceFeatures supported_features;
+ anv_GetPhysicalDeviceFeatures(physicalDevice, &supported_features);
+ VkBool32 *supported_feature = (VkBool32 *)&supported_features;
+ VkBool32 *enabled_feature = (VkBool32 *)pCreateInfo->pEnabledFeatures;
+ unsigned num_features = sizeof(VkPhysicalDeviceFeatures) / sizeof(VkBool32);
+ for (uint32_t i = 0; i < num_features; i++) {
+ if (enabled_feature[i] && !supported_feature[i])
+ return vk_error(VK_ERROR_FEATURE_NOT_PRESENT);
+ }
+ }
+
device = vk_alloc2(&physical_device->instance->alloc, pAllocator,
sizeof(*device), 8,
VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
anv_bo_pool_init(&device->batch_bo_pool, device);
- result = anv_block_pool_init(&device->dynamic_state_block_pool, device,
- 16384);
+ result = anv_bo_cache_init(&device->bo_cache);
if (result != VK_SUCCESS)
goto fail_batch_bo_pool;
- anv_state_pool_init(&device->dynamic_state_pool,
- &device->dynamic_state_block_pool);
+ result = anv_state_pool_init(&device->dynamic_state_pool, device, 16384);
+ if (result != VK_SUCCESS)
+ goto fail_bo_cache;
- result = anv_block_pool_init(&device->instruction_block_pool, device,
- 1024 * 1024);
+ result = anv_state_pool_init(&device->instruction_state_pool, device, 16384);
if (result != VK_SUCCESS)
goto fail_dynamic_state_pool;
- anv_state_pool_init(&device->instruction_state_pool,
- &device->instruction_block_pool);
-
- result = anv_block_pool_init(&device->surface_state_block_pool, device,
- 4096);
+ result = anv_state_pool_init(&device->surface_state_pool, device, 4096);
if (result != VK_SUCCESS)
goto fail_instruction_state_pool;
- anv_state_pool_init(&device->surface_state_pool,
- &device->surface_state_block_pool);
-
result = anv_bo_init_new(&device->workaround_bo, device, 1024);
if (result != VK_SUCCESS)
goto fail_surface_state_pool;
case 9:
result = gen9_init_device_state(device);
break;
+ case 10:
+ result = gen10_init_device_state(device);
+ break;
default:
/* Shouldn't get here as we don't create physical devices for any other
* gens. */
anv_gem_close(device, device->workaround_bo.gem_handle);
fail_surface_state_pool:
anv_state_pool_finish(&device->surface_state_pool);
- anv_block_pool_finish(&device->surface_state_block_pool);
fail_instruction_state_pool:
anv_state_pool_finish(&device->instruction_state_pool);
- anv_block_pool_finish(&device->instruction_block_pool);
fail_dynamic_state_pool:
anv_state_pool_finish(&device->dynamic_state_pool);
- anv_block_pool_finish(&device->dynamic_state_block_pool);
+ fail_bo_cache:
+ anv_bo_cache_finish(&device->bo_cache);
fail_batch_bo_pool:
anv_bo_pool_finish(&device->batch_bo_pool);
pthread_cond_destroy(&device->queue_submit);
anv_gem_close(device, device->workaround_bo.gem_handle);
anv_state_pool_finish(&device->surface_state_pool);
- anv_block_pool_finish(&device->surface_state_block_pool);
anv_state_pool_finish(&device->instruction_state_pool);
- anv_block_pool_finish(&device->instruction_block_pool);
anv_state_pool_finish(&device->dynamic_state_pool);
- anv_block_pool_finish(&device->dynamic_state_block_pool);
+
+ anv_bo_cache_finish(&device->bo_cache);
anv_bo_pool_finish(&device->batch_bo_pool);
*pQueue = anv_queue_to_handle(&device->queue);
}
-VkResult
-anv_device_execbuf(struct anv_device *device,
- struct drm_i915_gem_execbuffer2 *execbuf,
- struct anv_bo **execbuf_bos)
-{
- int ret = anv_gem_execbuffer(device, execbuf);
- if (ret != 0) {
- /* We don't know the real error. */
- device->lost = true;
- return vk_errorf(VK_ERROR_DEVICE_LOST, "execbuf2 failed: %m");
- }
-
- struct drm_i915_gem_exec_object2 *objects =
- (void *)(uintptr_t)execbuf->buffers_ptr;
- for (uint32_t k = 0; k < execbuf->buffer_count; k++)
- execbuf_bos[k]->offset = objects[k].offset;
-
- return VK_SUCCESS;
-}
-
VkResult
anv_device_query_status(struct anv_device *device)
{
return VK_SUCCESS;
}
+VkResult
+anv_device_bo_busy(struct anv_device *device, struct anv_bo *bo)
+{
+ /* Note: This only returns whether or not the BO is in use by an i915 GPU.
+ * Other usages of the BO (such as on different hardware) will not be
+ * flagged as "busy" by this ioctl. Use with care.
+ */
+ int ret = anv_gem_busy(device, bo->gem_handle);
+ if (ret == 1) {
+ return VK_NOT_READY;
+ } else if (ret == -1) {
+ /* We don't know the real error. */
+ device->lost = true;
+ return vk_errorf(VK_ERROR_DEVICE_LOST, "gem wait failed: %m");
+ }
+
+ /* Query for device status after the busy call. If the BO we're checking
+ * got caught in a GPU hang we don't want to return VK_SUCCESS to the
+ * client because it clearly doesn't have valid data. Yes, this most
+ * likely means an ioctl, but we just did an ioctl to query the busy status
+ * so it's no great loss.
+ */
+ return anv_device_query_status(device);
+}
+
VkResult
anv_device_wait(struct anv_device *device, struct anv_bo *bo,
int64_t timeout)
return anv_device_query_status(device);
}
-VkResult anv_QueueSubmit(
- VkQueue _queue,
- uint32_t submitCount,
- const VkSubmitInfo* pSubmits,
- VkFence _fence)
-{
- ANV_FROM_HANDLE(anv_queue, queue, _queue);
- ANV_FROM_HANDLE(anv_fence, fence, _fence);
- struct anv_device *device = queue->device;
-
- /* Query for device status prior to submitting. Technically, we don't need
- * to do this. However, if we have a client that's submitting piles of
- * garbage, we would rather break as early as possible to keep the GPU
- * hanging contained. If we don't check here, we'll either be waiting for
- * the kernel to kick us or we'll have to wait until the client waits on a
- * fence before we actually know whether or not we've hung.
- */
- VkResult result = anv_device_query_status(device);
- if (result != VK_SUCCESS)
- return result;
-
- /* We lock around QueueSubmit for three main reasons:
- *
- * 1) When a block pool is resized, we create a new gem handle with a
- * different size and, in the case of surface states, possibly a
- * different center offset but we re-use the same anv_bo struct when
- * we do so. If this happens in the middle of setting up an execbuf,
- * we could end up with our list of BOs out of sync with our list of
- * gem handles.
- *
- * 2) The algorithm we use for building the list of unique buffers isn't
- * thread-safe. While the client is supposed to syncronize around
- * QueueSubmit, this would be extremely difficult to debug if it ever
- * came up in the wild due to a broken app. It's better to play it
- * safe and just lock around QueueSubmit.
- *
- * 3) The anv_cmd_buffer_execbuf function may perform relocations in
- * userspace. Due to the fact that the surface state buffer is shared
- * between batches, we can't afford to have that happen from multiple
- * threads at the same time. Even though the user is supposed to
- * ensure this doesn't happen, we play it safe as in (2) above.
- *
- * Since the only other things that ever take the device lock such as block
- * pool resize only rarely happen, this will almost never be contended so
- * taking a lock isn't really an expensive operation in this case.
- */
- pthread_mutex_lock(&device->mutex);
-
- for (uint32_t i = 0; i < submitCount; i++) {
- for (uint32_t j = 0; j < pSubmits[i].commandBufferCount; j++) {
- ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer,
- pSubmits[i].pCommandBuffers[j]);
- assert(cmd_buffer->level == VK_COMMAND_BUFFER_LEVEL_PRIMARY);
- assert(!anv_batch_has_error(&cmd_buffer->batch));
-
- result = anv_cmd_buffer_execbuf(device, cmd_buffer);
- if (result != VK_SUCCESS)
- goto out;
- }
- }
-
- if (fence) {
- struct anv_bo *fence_bo = &fence->bo;
- result = anv_device_execbuf(device, &fence->execbuf, &fence_bo);
- if (result != VK_SUCCESS)
- goto out;
-
- /* Update the fence and wake up any waiters */
- assert(fence->state == ANV_FENCE_STATE_RESET);
- fence->state = ANV_FENCE_STATE_SUBMITTED;
- pthread_cond_broadcast(&device->queue_submit);
- }
-
-out:
- if (result != VK_SUCCESS) {
- /* In the case that something has gone wrong we may end up with an
- * inconsistent state from which it may not be trivial to recover.
- * For example, we might have computed address relocations and
- * any future attempt to re-submit this job will need to know about
- * this and avoid computing relocation addresses again.
- *
- * To avoid this sort of issues, we assume that if something was
- * wrong during submission we must already be in a really bad situation
- * anyway (such us being out of memory) and return
- * VK_ERROR_DEVICE_LOST to ensure that clients do not attempt to
- * submit the same job again to this device.
- */
- result = VK_ERROR_DEVICE_LOST;
- device->lost = true;
-
- /* If we return VK_ERROR_DEVICE LOST here, we need to ensure that
- * vkWaitForFences() and vkGetFenceStatus() return a valid result
- * (VK_SUCCESS or VK_ERROR_DEVICE_LOST) in a finite amount of time.
- * Setting the fence status to SIGNALED ensures this will happen in
- * any case.
- */
- if (fence)
- fence->state = ANV_FENCE_STATE_SIGNALED;
- }
-
- pthread_mutex_unlock(&device->mutex);
-
- return result;
-}
-
-VkResult anv_QueueWaitIdle(
- VkQueue _queue)
-{
- ANV_FROM_HANDLE(anv_queue, queue, _queue);
-
- return anv_DeviceWaitIdle(anv_device_to_handle(queue->device));
-}
-
VkResult anv_DeviceWaitIdle(
VkDevice _device)
{
anv_bo_init(bo, gem_handle, size);
- if (device->instance->physicalDevice.supports_48bit_addresses)
- bo->flags |= EXEC_OBJECT_SUPPORTS_48B_ADDRESS;
-
return VK_SUCCESS;
}
VkDeviceMemory* pMem)
{
ANV_FROM_HANDLE(anv_device, device, _device);
+ struct anv_physical_device *pdevice = &device->instance->physicalDevice;
struct anv_device_memory *mem;
- VkResult result;
+ VkResult result = VK_SUCCESS;
assert(pAllocateInfo->sType == VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO);
/* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
assert(pAllocateInfo->allocationSize > 0);
- /* We support exactly one memory heap. */
- assert(pAllocateInfo->memoryTypeIndex == 0 ||
- (!device->info.has_llc && pAllocateInfo->memoryTypeIndex < 2));
+ /* The kernel relocation API has a limitation of a 32-bit delta value
+ * applied to the address before it is written which, in spite of it being
+ * unsigned, is treated as signed . Because of the way that this maps to
+ * the Vulkan API, we cannot handle an offset into a buffer that does not
+ * fit into a signed 32 bits. The only mechanism we have for dealing with
+ * this at the moment is to limit all VkDeviceMemory objects to a maximum
+ * of 2GB each. The Vulkan spec allows us to do this:
+ *
+ * "Some platforms may have a limit on the maximum size of a single
+ * allocation. For example, certain systems may fail to create
+ * allocations with a size greater than or equal to 4GB. Such a limit is
+ * implementation-dependent, and if such a failure occurs then the error
+ * VK_ERROR_OUT_OF_DEVICE_MEMORY should be returned."
+ *
+ * We don't use vk_error here because it's not an error so much as an
+ * indication to the application that the allocation is too large.
+ */
+ if (pAllocateInfo->allocationSize > (1ull << 31))
+ return VK_ERROR_OUT_OF_DEVICE_MEMORY;
/* FINISHME: Fail if allocation request exceeds heap size. */
if (mem == NULL)
return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
- /* The kernel is going to give us whole pages anyway */
- uint64_t alloc_size = align_u64(pAllocateInfo->allocationSize, 4096);
+ assert(pAllocateInfo->memoryTypeIndex < pdevice->memory.type_count);
+ mem->type = &pdevice->memory.types[pAllocateInfo->memoryTypeIndex];
+ mem->map = NULL;
+ mem->map_size = 0;
- result = anv_bo_init_new(&mem->bo, device, alloc_size);
+ result = anv_bo_cache_alloc(device, &device->bo_cache,
+ pAllocateInfo->allocationSize,
+ &mem->bo);
if (result != VK_SUCCESS)
goto fail;
- mem->type_index = pAllocateInfo->memoryTypeIndex;
+ assert(mem->type->heapIndex < pdevice->memory.heap_count);
+ if (pdevice->memory.heaps[mem->type->heapIndex].supports_48bit_addresses)
+ mem->bo->flags |= EXEC_OBJECT_SUPPORTS_48B_ADDRESS;
- mem->map = NULL;
- mem->map_size = 0;
+ if (pdevice->has_exec_async)
+ mem->bo->flags |= EXEC_OBJECT_ASYNC;
*pMem = anv_device_memory_to_handle(mem);
if (mem->map)
anv_UnmapMemory(_device, _mem);
- if (mem->bo.map)
- anv_gem_munmap(mem->bo.map, mem->bo.size);
-
- if (mem->bo.gem_handle != 0)
- anv_gem_close(device, mem->bo.gem_handle);
+ anv_bo_cache_release(device, &device->bo_cache, mem->bo);
vk_free2(&device->alloc, pAllocator, mem);
}
}
if (size == VK_WHOLE_SIZE)
- size = mem->bo.size - offset;
+ size = mem->bo->size - offset;
/* From the Vulkan spec version 1.0.32 docs for MapMemory:
*
* equal to the size of the memory minus offset
*/
assert(size > 0);
- assert(offset + size <= mem->bo.size);
+ assert(offset + size <= mem->bo->size);
/* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
* takes a VkDeviceMemory pointer, it seems like only one map of the memory
* userspace. */
uint32_t gem_flags = 0;
- if (!device->info.has_llc && mem->type_index == 0)
+
+ if (!device->info.has_llc &&
+ (mem->type->propertyFlags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT))
gem_flags |= I915_MMAP_WC;
/* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
/* Let's map whole pages */
map_size = align_u64(map_size, 4096);
- void *map = anv_gem_mmap(device, mem->bo.gem_handle,
+ void *map = anv_gem_mmap(device, mem->bo->gem_handle,
map_offset, map_size, gem_flags);
if (map == MAP_FAILED)
return vk_error(VK_ERROR_MEMORY_MAP_FAILED);
if (ranges[i].offset >= mem->map_size)
continue;
- anv_clflush_range(mem->map + ranges[i].offset,
+ gen_clflush_range(mem->map + ranges[i].offset,
MIN2(ranges[i].size, mem->map_size - ranges[i].offset));
}
}
{
ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
ANV_FROM_HANDLE(anv_device, device, _device);
+ struct anv_physical_device *pdevice = &device->instance->physicalDevice;
/* The Vulkan spec (git aaed022) says:
*
* supported memory type for the resource. The bit `1<<i` is set if and
* only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
* structure for the physical device is supported.
- *
- * We support exactly one memory type on LLC, two on non-LLC.
*/
- pMemoryRequirements->memoryTypeBits = device->info.has_llc ? 1 : 3;
+ uint32_t memory_types = 0;
+ for (uint32_t i = 0; i < pdevice->memory.type_count; i++) {
+ uint32_t valid_usage = pdevice->memory.types[i].valid_buffer_usage;
+ if ((valid_usage & buffer->usage) == buffer->usage)
+ memory_types |= (1u << i);
+ }
pMemoryRequirements->size = buffer->size;
pMemoryRequirements->alignment = 16;
+ pMemoryRequirements->memoryTypeBits = memory_types;
+}
+
+void anv_GetBufferMemoryRequirements2KHR(
+ VkDevice _device,
+ const VkBufferMemoryRequirementsInfo2KHR* pInfo,
+ VkMemoryRequirements2KHR* pMemoryRequirements)
+{
+ anv_GetBufferMemoryRequirements(_device, pInfo->buffer,
+ &pMemoryRequirements->memoryRequirements);
+
+ vk_foreach_struct(ext, pMemoryRequirements->pNext) {
+ switch (ext->sType) {
+ default:
+ anv_debug_ignored_stype(ext->sType);
+ break;
+ }
+ }
}
void anv_GetImageMemoryRequirements(
{
ANV_FROM_HANDLE(anv_image, image, _image);
ANV_FROM_HANDLE(anv_device, device, _device);
+ struct anv_physical_device *pdevice = &device->instance->physicalDevice;
/* The Vulkan spec (git aaed022) says:
*
* only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
* structure for the physical device is supported.
*
- * We support exactly one memory type on LLC, two on non-LLC.
+ * All types are currently supported for images.
*/
- pMemoryRequirements->memoryTypeBits = device->info.has_llc ? 1 : 3;
+ uint32_t memory_types = (1ull << pdevice->memory.type_count) - 1;
pMemoryRequirements->size = image->size;
pMemoryRequirements->alignment = image->alignment;
+ pMemoryRequirements->memoryTypeBits = memory_types;
+}
+
+void anv_GetImageMemoryRequirements2KHR(
+ VkDevice _device,
+ const VkImageMemoryRequirementsInfo2KHR* pInfo,
+ VkMemoryRequirements2KHR* pMemoryRequirements)
+{
+ anv_GetImageMemoryRequirements(_device, pInfo->image,
+ &pMemoryRequirements->memoryRequirements);
+
+ vk_foreach_struct(ext, pMemoryRequirements->pNext) {
+ switch (ext->sType) {
+ default:
+ anv_debug_ignored_stype(ext->sType);
+ break;
+ }
+ }
}
void anv_GetImageSparseMemoryRequirements(
*pSparseMemoryRequirementCount = 0;
}
+void anv_GetImageSparseMemoryRequirements2KHR(
+ VkDevice device,
+ const VkImageSparseMemoryRequirementsInfo2KHR* pInfo,
+ uint32_t* pSparseMemoryRequirementCount,
+ VkSparseImageMemoryRequirements2KHR* pSparseMemoryRequirements)
+{
+ *pSparseMemoryRequirementCount = 0;
+}
+
void anv_GetDeviceMemoryCommitment(
VkDevice device,
VkDeviceMemory memory,
ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
if (mem) {
- buffer->bo = &mem->bo;
+ assert((buffer->usage & mem->type->valid_buffer_usage) == buffer->usage);
+ buffer->bo = mem->bo;
buffer->offset = memoryOffset;
} else {
buffer->bo = NULL;
return vk_error(VK_ERROR_FEATURE_NOT_PRESENT);
}
-VkResult anv_CreateFence(
- VkDevice _device,
- const VkFenceCreateInfo* pCreateInfo,
- const VkAllocationCallbacks* pAllocator,
- VkFence* pFence)
-{
- ANV_FROM_HANDLE(anv_device, device, _device);
- struct anv_bo fence_bo;
- struct anv_fence *fence;
- struct anv_batch batch;
- VkResult result;
-
- assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FENCE_CREATE_INFO);
-
- result = anv_bo_pool_alloc(&device->batch_bo_pool, &fence_bo, 4096);
- if (result != VK_SUCCESS)
- return result;
-
- /* Fences are small. Just store the CPU data structure in the BO. */
- fence = fence_bo.map;
- fence->bo = fence_bo;
-
- /* Place the batch after the CPU data but on its own cache line. */
- const uint32_t batch_offset = align_u32(sizeof(*fence), CACHELINE_SIZE);
- batch.next = batch.start = fence->bo.map + batch_offset;
- batch.end = fence->bo.map + fence->bo.size;
- anv_batch_emit(&batch, GEN7_MI_BATCH_BUFFER_END, bbe);
- anv_batch_emit(&batch, GEN7_MI_NOOP, noop);
-
- if (!device->info.has_llc) {
- assert(((uintptr_t) batch.start & CACHELINE_MASK) == 0);
- assert(batch.next - batch.start <= CACHELINE_SIZE);
- __builtin_ia32_mfence();
- __builtin_ia32_clflush(batch.start);
- }
-
- fence->exec2_objects[0].handle = fence->bo.gem_handle;
- fence->exec2_objects[0].relocation_count = 0;
- fence->exec2_objects[0].relocs_ptr = 0;
- fence->exec2_objects[0].alignment = 0;
- fence->exec2_objects[0].offset = fence->bo.offset;
- fence->exec2_objects[0].flags = 0;
- fence->exec2_objects[0].rsvd1 = 0;
- fence->exec2_objects[0].rsvd2 = 0;
-
- fence->execbuf.buffers_ptr = (uintptr_t) fence->exec2_objects;
- fence->execbuf.buffer_count = 1;
- fence->execbuf.batch_start_offset = batch.start - fence->bo.map;
- fence->execbuf.batch_len = batch.next - batch.start;
- fence->execbuf.cliprects_ptr = 0;
- fence->execbuf.num_cliprects = 0;
- fence->execbuf.DR1 = 0;
- fence->execbuf.DR4 = 0;
-
- fence->execbuf.flags =
- I915_EXEC_HANDLE_LUT | I915_EXEC_NO_RELOC | I915_EXEC_RENDER;
- fence->execbuf.rsvd1 = device->context_id;
- fence->execbuf.rsvd2 = 0;
-
- if (pCreateInfo->flags & VK_FENCE_CREATE_SIGNALED_BIT) {
- fence->state = ANV_FENCE_STATE_SIGNALED;
- } else {
- fence->state = ANV_FENCE_STATE_RESET;
- }
-
- *pFence = anv_fence_to_handle(fence);
-
- return VK_SUCCESS;
-}
-
-void anv_DestroyFence(
- VkDevice _device,
- VkFence _fence,
- const VkAllocationCallbacks* pAllocator)
-{
- ANV_FROM_HANDLE(anv_device, device, _device);
- ANV_FROM_HANDLE(anv_fence, fence, _fence);
-
- if (!fence)
- return;
-
- assert(fence->bo.map == fence);
- anv_bo_pool_free(&device->batch_bo_pool, &fence->bo);
-}
-
-VkResult anv_ResetFences(
- VkDevice _device,
- uint32_t fenceCount,
- const VkFence* pFences)
-{
- for (uint32_t i = 0; i < fenceCount; i++) {
- ANV_FROM_HANDLE(anv_fence, fence, pFences[i]);
- fence->state = ANV_FENCE_STATE_RESET;
- }
-
- return VK_SUCCESS;
-}
-
-VkResult anv_GetFenceStatus(
- VkDevice _device,
- VkFence _fence)
-{
- ANV_FROM_HANDLE(anv_device, device, _device);
- ANV_FROM_HANDLE(anv_fence, fence, _fence);
-
- if (unlikely(device->lost))
- return VK_ERROR_DEVICE_LOST;
-
- switch (fence->state) {
- case ANV_FENCE_STATE_RESET:
- /* If it hasn't even been sent off to the GPU yet, it's not ready */
- return VK_NOT_READY;
-
- case ANV_FENCE_STATE_SIGNALED:
- /* It's been signaled, return success */
- return VK_SUCCESS;
-
- case ANV_FENCE_STATE_SUBMITTED: {
- VkResult result = anv_device_wait(device, &fence->bo, 0);
- switch (result) {
- case VK_SUCCESS:
- fence->state = ANV_FENCE_STATE_SIGNALED;
- return VK_SUCCESS;
- case VK_TIMEOUT:
- return VK_NOT_READY;
- default:
- return result;
- }
- }
- default:
- unreachable("Invalid fence status");
- }
-}
-
-#define NSEC_PER_SEC 1000000000
-#define INT_TYPE_MAX(type) ((1ull << (sizeof(type) * 8 - 1)) - 1)
-
-VkResult anv_WaitForFences(
- VkDevice _device,
- uint32_t fenceCount,
- const VkFence* pFences,
- VkBool32 waitAll,
- uint64_t _timeout)
-{
- ANV_FROM_HANDLE(anv_device, device, _device);
- int ret;
-
- if (unlikely(device->lost))
- return VK_ERROR_DEVICE_LOST;
-
- /* DRM_IOCTL_I915_GEM_WAIT uses a signed 64 bit timeout and is supposed
- * to block indefinitely timeouts <= 0. Unfortunately, this was broken
- * for a couple of kernel releases. Since there's no way to know
- * whether or not the kernel we're using is one of the broken ones, the
- * best we can do is to clamp the timeout to INT64_MAX. This limits the
- * maximum timeout from 584 years to 292 years - likely not a big deal.
- */
- int64_t timeout = MIN2(_timeout, INT64_MAX);
-
- VkResult result = VK_SUCCESS;
- uint32_t pending_fences = fenceCount;
- while (pending_fences) {
- pending_fences = 0;
- bool signaled_fences = false;
- for (uint32_t i = 0; i < fenceCount; i++) {
- ANV_FROM_HANDLE(anv_fence, fence, pFences[i]);
- switch (fence->state) {
- case ANV_FENCE_STATE_RESET:
- /* This fence hasn't been submitted yet, we'll catch it the next
- * time around. Yes, this may mean we dead-loop but, short of
- * lots of locking and a condition variable, there's not much that
- * we can do about that.
- */
- pending_fences++;
- continue;
-
- case ANV_FENCE_STATE_SIGNALED:
- /* This fence is not pending. If waitAll isn't set, we can return
- * early. Otherwise, we have to keep going.
- */
- if (!waitAll) {
- result = VK_SUCCESS;
- goto done;
- }
- continue;
-
- case ANV_FENCE_STATE_SUBMITTED:
- /* These are the fences we really care about. Go ahead and wait
- * on it until we hit a timeout.
- */
- result = anv_device_wait(device, &fence->bo, timeout);
- switch (result) {
- case VK_SUCCESS:
- fence->state = ANV_FENCE_STATE_SIGNALED;
- signaled_fences = true;
- if (!waitAll)
- goto done;
- break;
-
- case VK_TIMEOUT:
- goto done;
-
- default:
- return result;
- }
- }
- }
-
- if (pending_fences && !signaled_fences) {
- /* If we've hit this then someone decided to vkWaitForFences before
- * they've actually submitted any of them to a queue. This is a
- * fairly pessimal case, so it's ok to lock here and use a standard
- * pthreads condition variable.
- */
- pthread_mutex_lock(&device->mutex);
-
- /* It's possible that some of the fences have changed state since the
- * last time we checked. Now that we have the lock, check for
- * pending fences again and don't wait if it's changed.
- */
- uint32_t now_pending_fences = 0;
- for (uint32_t i = 0; i < fenceCount; i++) {
- ANV_FROM_HANDLE(anv_fence, fence, pFences[i]);
- if (fence->state == ANV_FENCE_STATE_RESET)
- now_pending_fences++;
- }
- assert(now_pending_fences <= pending_fences);
-
- if (now_pending_fences == pending_fences) {
- struct timespec before;
- clock_gettime(CLOCK_MONOTONIC, &before);
-
- uint32_t abs_nsec = before.tv_nsec + timeout % NSEC_PER_SEC;
- uint64_t abs_sec = before.tv_sec + (abs_nsec / NSEC_PER_SEC) +
- (timeout / NSEC_PER_SEC);
- abs_nsec %= NSEC_PER_SEC;
-
- /* Avoid roll-over in tv_sec on 32-bit systems if the user
- * provided timeout is UINT64_MAX
- */
- struct timespec abstime;
- abstime.tv_nsec = abs_nsec;
- abstime.tv_sec = MIN2(abs_sec, INT_TYPE_MAX(abstime.tv_sec));
-
- ret = pthread_cond_timedwait(&device->queue_submit,
- &device->mutex, &abstime);
- assert(ret != EINVAL);
-
- struct timespec after;
- clock_gettime(CLOCK_MONOTONIC, &after);
- uint64_t time_elapsed =
- ((uint64_t)after.tv_sec * NSEC_PER_SEC + after.tv_nsec) -
- ((uint64_t)before.tv_sec * NSEC_PER_SEC + before.tv_nsec);
-
- if (time_elapsed >= timeout) {
- pthread_mutex_unlock(&device->mutex);
- result = VK_TIMEOUT;
- goto done;
- }
-
- timeout -= time_elapsed;
- }
-
- pthread_mutex_unlock(&device->mutex);
- }
- }
-
-done:
- if (unlikely(device->lost))
- return VK_ERROR_DEVICE_LOST;
-
- return result;
-}
-
-// Queue semaphore functions
-
-VkResult anv_CreateSemaphore(
- VkDevice device,
- const VkSemaphoreCreateInfo* pCreateInfo,
- const VkAllocationCallbacks* pAllocator,
- VkSemaphore* pSemaphore)
-{
- /* The DRM execbuffer ioctl always execute in-oder, even between different
- * rings. As such, there's nothing to do for the user space semaphore.
- */
-
- *pSemaphore = (VkSemaphore)1;
-
- return VK_SUCCESS;
-}
-
-void anv_DestroySemaphore(
- VkDevice device,
- VkSemaphore semaphore,
- const VkAllocationCallbacks* pAllocator)
-{
-}
-
// Event functions
VkResult anv_CreateEvent(
projects / mesa.git / blobdiff