* IN THE SOFTWARE.
*/
+#include <dlfcn.h>
#include <assert.h>
#include <stdbool.h>
#include <string.h>
+#include <sys/mman.h>
+#include <sys/stat.h>
#include <unistd.h>
#include <fcntl.h>
#include "anv_private.h"
-#include "anv_timestamp.h"
#include "util/strtod.h"
#include "util/debug.h"
va_end(args);
}
+static bool
+anv_get_function_timestamp(void *ptr, uint32_t* timestamp)
+{
+ Dl_info info;
+ struct stat st;
+ if (!dladdr(ptr, &info) || !info.dli_fname)
+ return false;
+
+ if (stat(info.dli_fname, &st))
+ return false;
+
+ *timestamp = st.st_mtim.tv_sec;
+ return true;
+}
+
+static bool
+anv_device_get_cache_uuid(void *uuid)
+{
+ uint32_t timestamp;
+
+ memset(uuid, 0, VK_UUID_SIZE);
+ if (!anv_get_function_timestamp(anv_device_get_cache_uuid, ×tamp))
+ return false;
+
+ snprintf(uuid, VK_UUID_SIZE, "anv-%d", timestamp);
+ return true;
+}
+
static VkResult
anv_physical_device_init(struct anv_physical_device *device,
struct anv_instance *instance,
goto fail;
}
+ if (!anv_device_get_cache_uuid(device->uuid)) {
+ result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
+ "cannot generate UUID");
+ goto fail;
+ }
bool swizzled = anv_gem_get_bit6_swizzle(fd, I915_TILING_X);
/* GENs prior to 8 do not support EU/Subslice info */
device->info.max_cs_threads = max_cs_threads;
}
- close(fd);
-
brw_process_intel_debug_variable();
device->compiler = brw_compiler_create(NULL, &device->info);
device->compiler->shader_perf_log = compiler_perf_log;
result = anv_init_wsi(device);
- if (result != VK_SUCCESS)
- goto fail;
+ if (result != VK_SUCCESS) {
+ ralloc_free(device->compiler);
+ goto fail;
+ }
- /* XXX: Actually detect bit6 swizzling */
isl_device_init(&device->isl_dev, &device->info, swizzled);
+ close(fd);
return VK_SUCCESS;
fail:
.extensionName = VK_KHR_SWAPCHAIN_EXTENSION_NAME,
.specVersion = 68,
},
+ {
+ .extensionName = VK_KHR_SAMPLER_MIRROR_CLAMP_TO_EDGE_EXTENSION_NAME,
+ .specVersion = 1,
+ }
};
static void *
*pFeatures = (VkPhysicalDeviceFeatures) {
.robustBufferAccess = true,
.fullDrawIndexUint32 = true,
- .imageCubeArray = false,
+ .imageCubeArray = true,
.independentBlend = true,
.geometryShader = true,
- .tessellationShader = false,
+ .tessellationShader = true,
.sampleRateShading = true,
.dualSrcBlend = true,
.logicOp = true,
.multiDrawIndirect = false,
- .drawIndirectFirstInstance = false,
+ .drawIndirectFirstInstance = true,
.depthClamp = true,
- .depthBiasClamp = false,
+ .depthBiasClamp = true,
.fillModeNonSolid = true,
.depthBounds = false,
.wideLines = true,
.pipelineStatisticsQuery = false,
.fragmentStoresAndAtomics = true,
.shaderTessellationAndGeometryPointSize = true,
- .shaderImageGatherExtended = false,
- .shaderStorageImageExtendedFormats = false,
+ .shaderImageGatherExtended = true,
+ .shaderStorageImageExtendedFormats = true,
.shaderStorageImageMultisample = false,
+ .shaderStorageImageReadWithoutFormat = false,
+ .shaderStorageImageWriteWithoutFormat = false,
.shaderUniformBufferArrayDynamicIndexing = true,
.shaderSampledImageArrayDynamicIndexing = true,
.shaderStorageBufferArrayDynamicIndexing = true,
.shaderStorageImageArrayDynamicIndexing = true,
- .shaderStorageImageReadWithoutFormat = false,
- .shaderStorageImageWriteWithoutFormat = true,
- .shaderClipDistance = false,
- .shaderCullDistance = false,
- .shaderFloat64 = false,
+ .shaderClipDistance = true,
+ .shaderCullDistance = true,
+ .shaderFloat64 = pdevice->info.gen >= 8,
.shaderInt64 = false,
.shaderInt16 = false,
- .alphaToOne = true,
+ .shaderResourceMinLod = false,
.variableMultisampleRate = false,
.inheritedQueries = false,
};
pdevice->compiler->scalar_stage[MESA_SHADER_GEOMETRY];
}
-void
-anv_device_get_cache_uuid(void *uuid)
-{
- memset(uuid, 0, VK_UUID_SIZE);
- snprintf(uuid, VK_UUID_SIZE, "anv-%s", ANV_TIMESTAMP);
-}
-
void anv_GetPhysicalDeviceProperties(
VkPhysicalDevice physicalDevice,
VkPhysicalDeviceProperties* pProperties)
.maxVertexInputAttributeOffset = 2047,
.maxVertexInputBindingStride = 2048,
.maxVertexOutputComponents = 128,
- .maxTessellationGenerationLevel = 0,
- .maxTessellationPatchSize = 0,
- .maxTessellationControlPerVertexInputComponents = 0,
- .maxTessellationControlPerVertexOutputComponents = 0,
- .maxTessellationControlPerPatchOutputComponents = 0,
- .maxTessellationControlTotalOutputComponents = 0,
- .maxTessellationEvaluationInputComponents = 0,
- .maxTessellationEvaluationOutputComponents = 0,
+ .maxTessellationGenerationLevel = 64,
+ .maxTessellationPatchSize = 32,
+ .maxTessellationControlPerVertexInputComponents = 128,
+ .maxTessellationControlPerVertexOutputComponents = 128,
+ .maxTessellationControlPerPatchOutputComponents = 128,
+ .maxTessellationControlTotalOutputComponents = 2048,
+ .maxTessellationEvaluationInputComponents = 128,
+ .maxTessellationEvaluationOutputComponents = 128,
.maxGeometryShaderInvocations = 32,
.maxGeometryInputComponents = 64,
.maxGeometryOutputComponents = 128,
.maxGeometryTotalOutputComponents = 1024,
.maxFragmentInputComponents = 128,
.maxFragmentOutputAttachments = 8,
- .maxFragmentDualSrcAttachments = 2,
+ .maxFragmentDualSrcAttachments = 1,
.maxFragmentCombinedOutputResources = 8,
.maxComputeSharedMemorySize = 32768,
.maxComputeWorkGroupCount = { 65535, 65535, 65535 },
.viewportSubPixelBits = 13, /* We take a float? */
.minMemoryMapAlignment = 4096, /* A page */
.minTexelBufferOffsetAlignment = 1,
- .minUniformBufferOffsetAlignment = 1,
- .minStorageBufferOffsetAlignment = 1,
+ .minUniformBufferOffsetAlignment = 16,
+ .minStorageBufferOffsetAlignment = 4,
.minTexelOffset = -8,
.maxTexelOffset = 7,
- .minTexelGatherOffset = -8,
- .maxTexelGatherOffset = 7,
+ .minTexelGatherOffset = -32,
+ .maxTexelGatherOffset = 31,
.minInterpolationOffset = -0.5,
.maxInterpolationOffset = 0.4375,
.subPixelInterpolationOffsetBits = 4,
.maxFramebufferWidth = (1 << 14),
.maxFramebufferHeight = (1 << 14),
- .maxFramebufferLayers = (1 << 10),
+ .maxFramebufferLayers = (1 << 11),
.framebufferColorSampleCounts = sample_counts,
.framebufferDepthSampleCounts = sample_counts,
.framebufferStencilSampleCounts = sample_counts,
.maxSampleMaskWords = 1,
.timestampComputeAndGraphics = false,
.timestampPeriod = time_stamp_base,
- .maxClipDistances = 0 /* FIXME */,
- .maxCullDistances = 0 /* FIXME */,
- .maxCombinedClipAndCullDistances = 0 /* FIXME */,
+ .maxClipDistances = 8,
+ .maxCullDistances = 8,
+ .maxCombinedClipAndCullDistances = 8,
.discreteQueuePriorities = 1,
.pointSizeRange = { 0.125, 255.875 },
.lineWidthRange = { 0.0, 7.9921875 },
};
strcpy(pProperties->deviceName, pdevice->name);
- anv_device_get_cache_uuid(pProperties->pipelineCacheUUID);
+ memcpy(pProperties->pipelineCacheUUID, pdevice->uuid, VK_UUID_SIZE);
}
void anv_GetPhysicalDeviceQueueFamilyProperties(
return;
}
- assert(*pCount >= 1);
+ /* The spec implicitly allows the incoming count to be 0. From the Vulkan
+ * 1.0.38 spec, Section 4.1 Physical Devices:
+ *
+ * If the value referenced by pQueueFamilyPropertyCount is not 0 [then
+ * do stuff].
+ */
+ if (*pCount == 0)
+ return;
*pQueueFamilyProperties = (VkQueueFamilyProperties) {
.queueFlags = VK_QUEUE_GRAPHICS_BIT |
.timestampValidBits = 36, /* XXX: Real value here */
.minImageTransferGranularity = (VkExtent3D) { 1, 1, 1 },
};
+
+ *pCount = 1;
}
void anv_GetPhysicalDeviceMemoryProperties(
return anv_lookup_entrypoint(&device->info, pName);
}
-static VkResult
+static void
anv_queue_init(struct anv_device *device, struct anv_queue *queue)
{
queue->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
queue->device = device;
queue->pool = &device->surface_state_pool;
-
- return VK_SUCCESS;
}
static void
pthread_mutex_init(&device->mutex, NULL);
+ pthread_condattr_t condattr;
+ pthread_condattr_init(&condattr);
+ pthread_condattr_setclock(&condattr, CLOCK_MONOTONIC);
+ pthread_cond_init(&device->queue_submit, NULL);
+ pthread_condattr_destroy(&condattr);
+
anv_bo_pool_init(&device->batch_bo_pool, device);
anv_block_pool_init(&device->dynamic_state_block_pool, device, 16384);
{
ANV_FROM_HANDLE(anv_device, device, _device);
- anv_queue_finish(&device->queue);
-
anv_device_finish_blorp(device);
+ anv_queue_finish(&device->queue);
+
#ifdef HAVE_VALGRIND
/* We only need to free these to prevent valgrind errors. The backing
* BO will go away in a couple of lines so we don't actually leak.
anv_state_pool_free(&device->dynamic_state_pool, device->border_colors);
#endif
+ anv_scratch_pool_finish(device, &device->scratch_pool);
+
anv_gem_munmap(device->workaround_bo.map, device->workaround_bo.size);
anv_gem_close(device, device->workaround_bo.gem_handle);
- anv_bo_pool_finish(&device->batch_bo_pool);
- anv_state_pool_finish(&device->dynamic_state_pool);
- anv_block_pool_finish(&device->dynamic_state_block_pool);
- anv_state_pool_finish(&device->instruction_state_pool);
- anv_block_pool_finish(&device->instruction_block_pool);
anv_state_pool_finish(&device->surface_state_pool);
anv_block_pool_finish(&device->surface_state_block_pool);
- anv_scratch_pool_finish(device, &device->scratch_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);
- close(device->fd);
+ anv_bo_pool_finish(&device->batch_bo_pool);
+ pthread_cond_destroy(&device->queue_submit);
pthread_mutex_destroy(&device->mutex);
+ anv_gem_destroy_context(device, device->context_id);
+
+ close(device->fd);
+
vk_free(&device->alloc, device);
}
return VK_SUCCESS;
}
- assert(*pPropertyCount >= ARRAY_SIZE(global_extensions));
+ *pPropertyCount = MIN2(*pPropertyCount, ARRAY_SIZE(global_extensions));
+ typed_memcpy(pProperties, global_extensions, *pPropertyCount);
- *pPropertyCount = ARRAY_SIZE(global_extensions);
- memcpy(pProperties, global_extensions, sizeof(global_extensions));
+ if (*pPropertyCount < ARRAY_SIZE(global_extensions))
+ return VK_INCOMPLETE;
return VK_SUCCESS;
}
return VK_SUCCESS;
}
- assert(*pPropertyCount >= ARRAY_SIZE(device_extensions));
+ *pPropertyCount = MIN2(*pPropertyCount, ARRAY_SIZE(device_extensions));
+ typed_memcpy(pProperties, device_extensions, *pPropertyCount);
- *pPropertyCount = ARRAY_SIZE(device_extensions);
- memcpy(pProperties, device_extensions, sizeof(device_extensions));
+ if (*pPropertyCount < ARRAY_SIZE(device_extensions))
+ return VK_INCOMPLETE;
return VK_SUCCESS;
}
return vk_errorf(VK_ERROR_DEVICE_LOST, "execbuf2 failed: %m");
}
- struct drm_i915_gem_exec_object2 *objects = (void *)execbuf->buffers_ptr;
+ 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;
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:
assert(pAllocateInfo->sType == VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO);
- if (pAllocateInfo->allocationSize == 0) {
- /* Apparently, this is allowed */
- *pMem = VK_NULL_HANDLE;
- return VK_SUCCESS;
- }
+ /* 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 ||
mem->type_index = pAllocateInfo->memoryTypeIndex;
+ mem->map = NULL;
+ mem->map_size = 0;
+
*pMem = anv_device_memory_to_handle(mem);
return VK_SUCCESS;
if (mem == NULL)
return;
+ if (mem->map)
+ anv_UnmapMemory(_device, _mem);
+
if (mem->bo.map)
anv_gem_munmap(mem->bo.map, mem->bo.size);
if (size == VK_WHOLE_SIZE)
size = mem->bo.size - offset;
+ /* From the Vulkan spec version 1.0.32 docs for MapMemory:
+ *
+ * * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
+ * assert(size != 0);
+ * * If size is not equal to VK_WHOLE_SIZE, size must be less than or
+ * equal to the size of the memory minus offset
+ */
+ assert(size > 0);
+ 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
* at a time is valid. We could just mmap up front and return an offset
/* Let's map whole pages */
map_size = align_u64(map_size, 4096);
- mem->map = anv_gem_mmap(device, mem->bo.gem_handle,
- map_offset, map_size, gem_flags);
+ 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);
+
+ mem->map = map;
mem->map_size = map_size;
*ppData = mem->map + (offset - map_offset);
return;
anv_gem_munmap(mem->map, mem->map_size);
+
+ mem->map = NULL;
+ mem->map_size = 0;
}
static void
fence->execbuf.rsvd1 = device->context_id;
fence->execbuf.rsvd2 = 0;
- fence->ready = false;
+ 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);
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);
}
{
for (uint32_t i = 0; i < fenceCount; i++) {
ANV_FROM_HANDLE(anv_fence, fence, pFences[i]);
- fence->ready = false;
+ fence->state = ANV_FENCE_STATE_RESET;
}
return VK_SUCCESS;
int64_t t = 0;
int ret;
- if (fence->ready)
- return VK_SUCCESS;
+ 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;
- ret = anv_gem_wait(device, fence->bo.gem_handle, &t);
- if (ret == 0) {
- fence->ready = true;
+ case ANV_FENCE_STATE_SIGNALED:
+ /* It's been signaled, return success */
return VK_SUCCESS;
- }
- return VK_NOT_READY;
+ case ANV_FENCE_STATE_SUBMITTED:
+ /* It's been submitted to the GPU but we don't know if it's done yet. */
+ ret = anv_gem_wait(device, fence->bo.gem_handle, &t);
+ if (ret == 0) {
+ fence->state = ANV_FENCE_STATE_SIGNALED;
+ return VK_SUCCESS;
+ } else {
+ return VK_NOT_READY;
+ }
+ 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)
+ uint64_t _timeout)
{
ANV_FROM_HANDLE(anv_device, device, _device);
+ int ret;
/* DRM_IOCTL_I915_GEM_WAIT uses a signed 64 bit timeout and is supposed
* to block indefinitely timeouts <= 0. Unfortunately, this was broken
* 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.
*/
- if (timeout > INT64_MAX)
- timeout = INT64_MAX;
-
- int64_t t = timeout;
+ int64_t timeout = MIN2(_timeout, INT64_MAX);
+
+ 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)
+ return VK_SUCCESS;
+ 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.
+ */
+ ret = anv_gem_wait(device, fence->bo.gem_handle, &timeout);
+ if (ret == -1 && errno == ETIME) {
+ return VK_TIMEOUT;
+ } else if (ret == -1) {
+ /* We don't know the real error. */
+ return vk_errorf(VK_ERROR_DEVICE_LOST, "gem wait failed: %m");
+ } else {
+ fence->state = ANV_FENCE_STATE_SIGNALED;
+ signaled_fences = true;
+ if (!waitAll)
+ return VK_SUCCESS;
+ continue;
+ }
+ }
+ }
- /* FIXME: handle !waitAll */
+ 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);
+ return VK_TIMEOUT;
+ }
+
+ timeout -= time_elapsed;
+ }
- for (uint32_t i = 0; i < fenceCount; i++) {
- ANV_FROM_HANDLE(anv_fence, fence, pFences[i]);
- int ret = anv_gem_wait(device, fence->bo.gem_handle, &t);
- if (ret == -1 && errno == ETIME) {
- return VK_TIMEOUT;
- } else if (ret == -1) {
- /* We don't know the real error. */
- return vk_errorf(VK_ERROR_OUT_OF_DEVICE_MEMORY,
- "gem wait failed: %m");
+ pthread_mutex_unlock(&device->mutex);
}
}
ANV_FROM_HANDLE(anv_device, device, _device);
ANV_FROM_HANDLE(anv_event, event, _event);
+ if (!event)
+ return;
+
anv_state_pool_free(&device->dynamic_state_pool, event->state);
}
ANV_FROM_HANDLE(anv_device, device, _device);
ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
+ if (!buffer)
+ return;
+
vk_free2(&device->alloc, pAllocator, buffer);
}
ANV_FROM_HANDLE(anv_device, device, _device);
ANV_FROM_HANDLE(anv_sampler, sampler, _sampler);
+ if (!sampler)
+ return;
+
vk_free2(&device->alloc, pAllocator, sampler);
}
ANV_FROM_HANDLE(anv_device, device, _device);
ANV_FROM_HANDLE(anv_framebuffer, fb, _fb);
+ if (!fb)
+ return;
+
vk_free2(&device->alloc, pAllocator, fb);
}