From fae88d8791b2903771c85958903ee81080b66aec Mon Sep 17 00:00:00 2001 From: Lionel Landwerlin Date: Mon, 2 May 2022 12:38:16 +0300 Subject: [PATCH] anv: make use of the new smallbar uAPI Instead of having 2 VkMemoryType pointing to the same VkMemoryHeap, we have each VkMemoryType with VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT (one host visible, the other not) point to its own VkMemoryHeap. For the local heap that is host visible, we'll use the I915_GEM_CREATE_EXT_FLAG_NEEDS_CPU_ACCESS flag at GEM BO creation. When the smallbar uAPI is not available we fallback to a single heap and do not use I915_GEM_CREATE_EXT_FLAG_NEEDS_CPU_ACCESS. v2: Handle probed_cpu_visible_size == probed_size (Matthew) v3: * Jordan: Use region info from devinfo v4: Also make the vram host visible heap as local (Ken) Signed-off-by: Lionel Landwerlin Signed-off-by: Jordan Justen Reviewed-by: Kenneth Graunke Part-of: --- src/intel/vulkan/anv_allocator.c | 18 +++++---- src/intel/vulkan/anv_device.c | 64 ++++++++++++++++++++++++++------ src/intel/vulkan/anv_gem.c | 6 ++- src/intel/vulkan/anv_gem_stubs.c | 2 +- src/intel/vulkan/anv_private.h | 17 ++++++++- 5 files changed, 84 insertions(+), 23 deletions(-) diff --git a/src/intel/vulkan/anv_allocator.c b/src/intel/vulkan/anv_allocator.c index 652c20d8c79..f6aeb5cb5cc 100644 --- a/src/intel/vulkan/anv_allocator.c +++ b/src/intel/vulkan/anv_allocator.c @@ -1696,23 +1696,27 @@ anv_device_alloc_bo(struct anv_device *device, /* If we have vram size, we have multiple memory regions and should choose * one of them. */ - if (device->physical->vram.size > 0) { + if (anv_physical_device_has_vram(device->physical)) { struct drm_i915_gem_memory_class_instance regions[2]; uint32_t nregions = 0; if (alloc_flags & ANV_BO_ALLOC_LOCAL_MEM) { - regions[nregions++] = device->physical->vram.region; + /* vram_non_mappable & vram_mappable actually are the same region. */ + regions[nregions++] = device->physical->vram_non_mappable.region; } else { regions[nregions++] = device->physical->sys.region; } - /* TODO: Add I915_GEM_CREATE_EXT_FLAG_NEEDS_CPU_ACCESS to flags for - * after small BAR uapi is stabilized. - */ - assert(intel_vram_all_mappable(&device->info)); + uint32_t flags = 0; + if (alloc_flags & ANV_BO_ALLOC_LOCAL_MEM_CPU_VISIBLE) { + assert(alloc_flags & ANV_BO_ALLOC_LOCAL_MEM); + /* We're required to add smem as a region when using mappable vram. */ + regions[nregions++] = device->physical->sys.region; + flags |= I915_GEM_CREATE_EXT_FLAG_NEEDS_CPU_ACCESS; + } gem_handle = anv_gem_create_regions(device, size + ccs_size, - nregions, regions); + flags, nregions, regions); } else { gem_handle = anv_gem_create(device, size + ccs_size); } diff --git a/src/intel/vulkan/anv_device.c b/src/intel/vulkan/anv_device.c index 03cbea51f02..63309be61dd 100644 --- a/src/intel/vulkan/anv_device.c +++ b/src/intel/vulkan/anv_device.c @@ -378,11 +378,18 @@ anv_init_meminfo(struct anv_physical_device *device, int fd) anv_compute_sys_heap_size(device, devinfo->mem.sram.mappable.size); device->sys.available = devinfo->mem.sram.mappable.free; - device->vram.region.memory_class = devinfo->mem.vram.mem_class; - device->vram.region.memory_instance = + device->vram_mappable.region.memory_class = devinfo->mem.vram.mem_class; + device->vram_mappable.region.memory_instance = devinfo->mem.vram.mem_instance; - device->vram.size = devinfo->mem.vram.mappable.size; - device->vram.available = devinfo->mem.vram.mappable.free; + device->vram_mappable.size = devinfo->mem.vram.mappable.size; + device->vram_mappable.available = devinfo->mem.vram.mappable.free; + + device->vram_non_mappable.region.memory_class = + devinfo->mem.vram.mem_class; + device->vram_non_mappable.region.memory_instance = + devinfo->mem.vram.mem_instance; + device->vram_non_mappable.size = devinfo->mem.vram.unmappable.size; + device->vram_non_mappable.available = devinfo->mem.vram.unmappable.free; return VK_SUCCESS; } @@ -395,7 +402,8 @@ anv_update_meminfo(struct anv_physical_device *device, int fd) const struct intel_device_info *devinfo = &device->info; device->sys.available = devinfo->mem.sram.mappable.free; - device->vram.available = devinfo->mem.vram.mappable.free; + device->vram_mappable.available = devinfo->mem.vram.mappable.free; + device->vram_non_mappable.available = devinfo->mem.vram.unmappable.free; } @@ -408,13 +416,19 @@ anv_physical_device_init_heaps(struct anv_physical_device *device, int fd) assert(device->sys.size != 0); - if (device->vram.size > 0) { - /* We can create 2 different heaps when we have local memory support, - * first heap with local memory size and second with system memory size. + if (anv_physical_device_has_vram(device)) { + /* We can create 2 or 3 different heaps when we have local memory + * support, first heap with local memory size and second with system + * memory size and the third is added only if part of the vram is + * mappable to the host. */ device->memory.heap_count = 2; device->memory.heaps[0] = (struct anv_memory_heap) { - .size = device->vram.size, + /* If there is a vram_non_mappable, use that for the device only + * heap. Otherwise use the vram_mappable. + */ + .size = device->vram_non_mappable.size != 0 ? + device->vram_non_mappable.size : device->vram_mappable.size, .flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT, .is_local_mem = true, }; @@ -423,6 +437,17 @@ anv_physical_device_init_heaps(struct anv_physical_device *device, int fd) .flags = 0, .is_local_mem = false, }; + /* Add an additional smaller vram mappable heap if we can't map all the + * vram to the host. + */ + if (device->vram_non_mappable.size > 0) { + device->memory.heap_count++; + device->memory.heaps[2] = (struct anv_memory_heap) { + .size = device->vram_mappable.size, + .flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT, + .is_local_mem = true, + }; + } device->memory.type_count = 3; device->memory.types[0] = (struct anv_memory_type) { @@ -439,7 +464,11 @@ anv_physical_device_init_heaps(struct anv_physical_device *device, int fd) .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT | VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT | VK_MEMORY_PROPERTY_HOST_COHERENT_BIT, - .heapIndex = 0, + /* This memory type either comes from heaps[0] if there is only + * mappable vram region, or from heaps[2] if there is both mappable & + * non-mappable vram regions. + */ + .heapIndex = device->vram_non_mappable.size > 0 ? 2 : 0, }; } else if (device->info.has_llc) { device->memory.heap_count = 1; @@ -843,7 +872,8 @@ anv_physical_device_try_create(struct anv_instance *instance, device->gtt_size > (4ULL << 30 /* GiB */); /* Initialize memory regions struct to 0. */ - memset(&device->vram, 0, sizeof(device->vram)); + memset(&device->vram_non_mappable, 0, sizeof(device->vram_non_mappable)); + memset(&device->vram_mappable, 0, sizeof(device->vram_mappable)); memset(&device->sys, 0, sizeof(device->sys)); result = anv_physical_device_init_heaps(device, fd); @@ -2741,7 +2771,7 @@ anv_get_memory_budget(VkPhysicalDevice physicalDevice, if (device->memory.heaps[i].is_local_mem) { total_heaps_size = total_vram_heaps_size; - mem_available = device->vram.available; + mem_available = device->vram_non_mappable.available; } else { total_heaps_size = total_sys_heaps_size; mem_available = device->sys.available; @@ -3832,6 +3862,16 @@ VkResult anv_AllocateMemory( if (device->physical->has_implicit_ccs && device->info.has_aux_map) alloc_flags |= ANV_BO_ALLOC_IMPLICIT_CCS; + /* If i915 reported a mappable/non_mappable vram regions and the + * application want lmem mappable, then we need to use the + * I915_GEM_CREATE_EXT_FLAG_NEEDS_CPU_ACCESS flag to create our BO. + */ + if (pdevice->vram_mappable.size > 0 && + pdevice->vram_non_mappable.size > 0 && + (mem_type->propertyFlags & VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT) && + (mem_type->propertyFlags & VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT)) + alloc_flags |= ANV_BO_ALLOC_LOCAL_MEM_CPU_VISIBLE; + if (vk_flags & VK_MEMORY_ALLOCATE_DEVICE_ADDRESS_BIT) alloc_flags |= ANV_BO_ALLOC_CLIENT_VISIBLE_ADDRESS; diff --git a/src/intel/vulkan/anv_gem.c b/src/intel/vulkan/anv_gem.c index 852b94cb957..61896b402a2 100644 --- a/src/intel/vulkan/anv_gem.c +++ b/src/intel/vulkan/anv_gem.c @@ -66,9 +66,12 @@ anv_gem_close(struct anv_device *device, uint32_t gem_handle) uint32_t anv_gem_create_regions(struct anv_device *device, uint64_t anv_bo_size, - uint32_t num_regions, + uint32_t flags, uint32_t num_regions, struct drm_i915_gem_memory_class_instance *regions) { + /* Check for invalid flags */ + assert((flags & ~I915_GEM_CREATE_EXT_FLAG_NEEDS_CPU_ACCESS) == 0); + struct drm_i915_gem_create_ext_memory_regions ext_regions = { .base = { .name = I915_GEM_CREATE_EXT_MEMORY_REGIONS }, .num_regions = num_regions, @@ -78,6 +81,7 @@ anv_gem_create_regions(struct anv_device *device, uint64_t anv_bo_size, struct drm_i915_gem_create_ext gem_create = { .size = anv_bo_size, .extensions = (uintptr_t) &ext_regions, + .flags = flags, }; int ret = intel_ioctl(device->fd, DRM_IOCTL_I915_GEM_CREATE_EXT, diff --git a/src/intel/vulkan/anv_gem_stubs.c b/src/intel/vulkan/anv_gem_stubs.c index b42d8bc8bf1..0a0f4c6e479 100644 --- a/src/intel/vulkan/anv_gem_stubs.c +++ b/src/intel/vulkan/anv_gem_stubs.c @@ -47,7 +47,7 @@ anv_gem_close(struct anv_device *device, uint32_t gem_handle) uint32_t anv_gem_create_regions(struct anv_device *device, uint64_t anv_bo_size, - uint32_t num_regions, + uint32_t flags, uint32_t num_regions, struct drm_i915_gem_memory_class_instance *regions) { return 0; diff --git a/src/intel/vulkan/anv_private.h b/src/intel/vulkan/anv_private.h index a00ac622942..5cfce532622 100644 --- a/src/intel/vulkan/anv_private.h +++ b/src/intel/vulkan/anv_private.h @@ -1040,7 +1040,11 @@ struct anv_physical_device { bool need_clflush; } memory; - struct anv_memregion vram; + /* Either we have a single vram region and it's all mappable, or we have + * both mappable & non-mappable parts. System memory is always available. + */ + struct anv_memregion vram_mappable; + struct anv_memregion vram_non_mappable; struct anv_memregion sys; uint8_t driver_build_sha1[20]; uint8_t pipeline_cache_uuid[VK_UUID_SIZE]; @@ -1066,6 +1070,12 @@ struct anv_physical_device { struct intel_measure_device measure_device; }; +static inline bool +anv_physical_device_has_vram(const struct anv_physical_device *device) +{ + return device->vram_mappable.size > 0; +} + struct anv_app_info { const char* app_name; uint32_t app_version; @@ -1368,6 +1378,9 @@ enum anv_bo_alloc_flags { /** This buffer is allocated from local memory */ ANV_BO_ALLOC_LOCAL_MEM = (1 << 10), + + /** This buffer is allocated from local memory and should be cpu visible */ + ANV_BO_ALLOC_LOCAL_MEM_CPU_VISIBLE = (1 << 11), }; VkResult anv_device_alloc_bo(struct anv_device *device, @@ -1431,7 +1444,7 @@ void anv_gem_munmap(struct anv_device *device, void *p, uint64_t size); uint32_t anv_gem_create(struct anv_device *device, uint64_t size); void anv_gem_close(struct anv_device *device, uint32_t gem_handle); uint32_t anv_gem_create_regions(struct anv_device *device, uint64_t anv_bo_size, - uint32_t num_regions, + uint32_t flags, uint32_t num_regions, struct drm_i915_gem_memory_class_instance *regions); uint32_t anv_gem_userptr(struct anv_device *device, void *mem, size_t size); int anv_gem_busy(struct anv_device *device, uint32_t gem_handle);