anv: Implement SSBOs bindings with GPU addresses in the descriptor BO

[mesa.git] / src / intel / vulkan / anv_device.c

diff --git a/src/intel/vulkan/anv_device.c b/src/intel/vulkan/anv_device.c
index 991ca50f7dd3d36fe454bcb11bf60818cd59b50a..de56926d9356522b342c069390b73c4e01ec2e0b 100644 (file)
--- a/src/intel/vulkan/anv_device.c
+++ b/src/intel/vulkan/anv_device.c
@@ -276,6 +276,10 @@ anv_physical_device_init_uuids(struct anv_physical_device *device)
    _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_update(&sha1_ctx, &device->always_use_bindless,
+                     sizeof(device->always_use_bindless));
+   _mesa_sha1_update(&sha1_ctx, &device->has_a64_buffer_access,
+                     sizeof(device->has_a64_buffer_access));
    _mesa_sha1_final(&sha1_ctx, sha1);
    memcpy(device->pipeline_cache_uuid, sha1, VK_UUID_SIZE);
 
@@ -451,6 +455,15 @@ anv_physical_device_init(struct anv_physical_device *device,
    device->has_context_isolation =
       anv_gem_get_param(fd, I915_PARAM_HAS_CONTEXT_ISOLATION);
 
+   device->always_use_bindless =
+      env_var_as_boolean("ANV_ALWAYS_BINDLESS", false);
+
+   /* We first got the A64 messages on broadwell and we can only use them if
+    * we can pass addresses directly into the shader which requires softpin.
+    */
+   device->has_a64_buffer_access = device->info.gen >= 8 &&
+                                   device->use_softpin;
+
    /* Starting with Gen10, the timestamp frequency of the command streamer may
     * vary from one part to another. We can query the value from the kernel.
     */
@@ -956,8 +969,7 @@ void anv_GetPhysicalDeviceFeatures2(
 
       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_DEVICE_ADDRESS_FEATURES_EXT: {
          VkPhysicalDeviceBufferDeviceAddressFeaturesEXT *features = (void *)ext;
-         features->bufferDeviceAddress = pdevice->use_softpin &&
-                                         pdevice->info.gen >= 8;
+         features->bufferDeviceAddress = pdevice->has_a64_buffer_access;
          features->bufferDeviceAddressCaptureReplay = false;
          features->bufferDeviceAddressMultiDevice = false;
          break;
@@ -1093,10 +1105,14 @@ void anv_GetPhysicalDeviceProperties(
    const uint32_t max_raw_buffer_sz = devinfo->gen >= 7 ?
                                       (1ul << 30) : (1ul << 27);
 
+   const uint32_t max_ssbos = pdevice->has_a64_buffer_access ? UINT16_MAX : 64;
    const uint32_t max_samplers = (devinfo->gen >= 8 || devinfo->is_haswell) ?
                                  128 : 16;
 
-   const uint32_t max_images = devinfo->gen < 9 ? MAX_GEN8_IMAGES : MAX_IMAGES;
+   /* The moment we have anything bindless, claim a high per-stage limit */
+   const uint32_t max_per_stage =
+      pdevice->has_a64_buffer_access ? UINT32_MAX :
+                                       MAX_BINDING_TABLE_SIZE - MAX_RTS;
 
    VkSampleCountFlags sample_counts =
       isl_device_get_sample_counts(&pdevice->isl_dev);
@@ -1119,18 +1135,18 @@ void anv_GetPhysicalDeviceProperties(
       .maxBoundDescriptorSets                   = MAX_SETS,
       .maxPerStageDescriptorSamplers            = max_samplers,
       .maxPerStageDescriptorUniformBuffers      = 64,
-      .maxPerStageDescriptorStorageBuffers      = 64,
+      .maxPerStageDescriptorStorageBuffers      = max_ssbos,
       .maxPerStageDescriptorSampledImages       = max_samplers,
-      .maxPerStageDescriptorStorageImages       = max_images,
+      .maxPerStageDescriptorStorageImages       = MAX_IMAGES,
       .maxPerStageDescriptorInputAttachments    = 64,
-      .maxPerStageResources                     = 250,
+      .maxPerStageResources                     = max_per_stage,
       .maxDescriptorSetSamplers                 = 6 * max_samplers, /* number of stages * maxPerStageDescriptorSamplers */
       .maxDescriptorSetUniformBuffers           = 6 * 64,           /* number of stages * maxPerStageDescriptorUniformBuffers */
       .maxDescriptorSetUniformBuffersDynamic    = MAX_DYNAMIC_BUFFERS / 2,
-      .maxDescriptorSetStorageBuffers           = 6 * 64,           /* number of stages * maxPerStageDescriptorStorageBuffers */
+      .maxDescriptorSetStorageBuffers           = 6 * max_ssbos,    /* number of stages * maxPerStageDescriptorStorageBuffers */
       .maxDescriptorSetStorageBuffersDynamic    = MAX_DYNAMIC_BUFFERS / 2,
       .maxDescriptorSetSampledImages            = 6 * max_samplers, /* number of stages * maxPerStageDescriptorSampledImages */
-      .maxDescriptorSetStorageImages            = 6 * max_images,   /* number of stages * maxPerStageDescriptorStorageImages */
+      .maxDescriptorSetStorageImages            = 6 * MAX_IMAGES,   /* number of stages * maxPerStageDescriptorStorageImages */
       .maxDescriptorSetInputAttachments         = 256,
       .maxVertexInputAttributes                 = MAX_VBS,
       .maxVertexInputBindings                   = MAX_VBS,
@@ -2005,6 +2021,8 @@ VkResult anv_CreateDevice(
          high_heap->size;
    }
 
+   list_inithead(&device->memory_objects);
+
    /* As per spec, the driver implementation may deny requests to acquire
     * a priority above the default priority (MEDIUM) if the caller does not
     * have sufficient privileges. In this scenario VK_ERROR_NOT_PERMITTED_EXT
@@ -2118,9 +2136,6 @@ VkResult anv_CreateDevice(
    if (device->info.gen >= 10)
       anv_device_init_hiz_clear_value_bo(device);
 
-   if (physical_device->use_softpin)
-      device->pinned_buffers = _mesa_pointer_set_create(NULL);
-
    anv_scratch_pool_init(device, &device->scratch_pool);
 
    anv_queue_init(device, &device->queue);
@@ -2211,9 +2226,6 @@ void anv_DestroyDevice(
 
    anv_queue_finish(&device->queue);
 
-   if (physical_device->use_softpin)
-      _mesa_set_destroy(device->pinned_buffers, NULL);
-
 #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.
@@ -2698,6 +2710,10 @@ VkResult anv_AllocateMemory(
    }
 
  success:
+   pthread_mutex_lock(&device->mutex);
+   list_addtail(&mem->link, &device->memory_objects);
+   pthread_mutex_unlock(&device->mutex);
+
    *pMem = anv_device_memory_to_handle(mem);
 
    return VK_SUCCESS;
@@ -2789,6 +2805,10 @@ void anv_FreeMemory(
    if (mem == NULL)
       return;
 
+   pthread_mutex_lock(&device->mutex);
+   list_del(&mem->link);
+   pthread_mutex_unlock(&device->mutex);
+
    if (mem->map)
       anv_UnmapMemory(_device, _mem);
 
@@ -3324,12 +3344,6 @@ VkResult anv_CreateBuffer(
    buffer->usage = pCreateInfo->usage;
    buffer->address = ANV_NULL_ADDRESS;
 
-   if (buffer->usage & VK_BUFFER_USAGE_SHADER_DEVICE_ADDRESS_BIT_EXT) {
-      pthread_mutex_lock(&device->mutex);
-      _mesa_set_add(device->pinned_buffers, buffer);
-      pthread_mutex_unlock(&device->mutex);
-   }
-
    *pBuffer = anv_buffer_to_handle(buffer);
 
    return VK_SUCCESS;
@@ -3346,12 +3360,6 @@ void anv_DestroyBuffer(
    if (!buffer)
       return;
 
-   if (buffer->usage & VK_BUFFER_USAGE_SHADER_DEVICE_ADDRESS_BIT_EXT) {
-      pthread_mutex_lock(&device->mutex);
-      _mesa_set_remove_key(device->pinned_buffers, buffer);
-      pthread_mutex_unlock(&device->mutex);
-   }
-
    vk_free2(&device->alloc, pAllocator, buffer);
 }