nvk: Advertise minStorageBufferOffsetAlignment=4 for VKD3D

VKD3D won't enable the VK_EXT_descriptor_buffer path unless we advertise
a minStorageBufferOffsetAlignment of 4.  On the NVIDIA proprietary
driver, they get around the limit by just detecting the driver and
smashing it to 4 internally.  It's better if we set the limit for them
so that we can also adjust NIR's optimizations to not assume higher
alignments.

v2 (mhenning): Rebase and always initialize value of
               instance->ssbo_align_4b

Reviewed-by: Mel Henning <mhenning@darkrefraction.com>

src/nouveau/vulkan/nvk_buffer.c

@@ -7,6 +7,7 @@
 #include "nvk_entrypoints.h"
 #include "nvk_device.h"
 #include "nvk_device_memory.h"
+#include "nvk_instance.h"
 #include "nvk_physical_device.h"
 #include "nvk_queue.h"
 #include "nvkmd/nvkmd.h"
@@ -20,13 +21,14 @@ nvk_get_buffer_alignment(const struct nvk_physical_device *pdev,
                          VkBufferUsageFlags2KHR usage_flags,
                          VkBufferCreateFlags create_flags)
 {
+   const struct nvk_instance *instance = nvk_physical_device_instance(pdev);
    uint32_t alignment = 16;
 
    if (usage_flags & VK_BUFFER_USAGE_2_UNIFORM_BUFFER_BIT_KHR)
       alignment = MAX2(alignment, nvk_min_cbuf_alignment(&pdev->info));
 
    if (usage_flags & VK_BUFFER_USAGE_2_STORAGE_BUFFER_BIT_KHR)
-      alignment = MAX2(alignment, NVK_MIN_SSBO_ALIGNMENT);
+      alignment = MAX2(alignment, nvk_min_ssbo_alignment(instance));
 
    if (usage_flags & (VK_BUFFER_USAGE_2_UNIFORM_TEXEL_BUFFER_BIT_KHR |
                       VK_BUFFER_USAGE_2_STORAGE_TEXEL_BUFFER_BIT_KHR))

src/nouveau/vulkan/nvk_descriptor_set.c

@@ -11,6 +11,7 @@
 #include "nvk_entrypoints.h"
 #include "nvk_format.h"
 #include "nvk_image_view.h"
+#include "nvk_instance.h"
 #include "nvk_physical_device.h"
 #include "nvk_sampler.h"
 #include "nvkmd/nvkmd.h"
@@ -314,12 +315,14 @@ write_dynamic_ubo_desc(struct nvk_descriptor_writer *w,
 }
 
 static union nvk_buffer_descriptor
-ssbo_desc(struct nvk_addr_range addr_range)
+ssbo_desc(const struct nvk_physical_device *pdev,
+          struct nvk_addr_range addr_range)
 {
-   assert(addr_range.addr % NVK_MIN_SSBO_ALIGNMENT == 0);
+   const struct nvk_instance *instance = nvk_physical_device_instance(pdev);
+   const uint32_t min_ssbo_alignment = nvk_min_ssbo_alignment(instance);
    assert(addr_range.range <= UINT32_MAX);
 
-   addr_range.addr = ROUND_DOWN_TO(addr_range.addr, NVK_MIN_SSBO_ALIGNMENT);
+   addr_range.addr = ROUND_DOWN_TO(addr_range.addr, min_ssbo_alignment);
    addr_range.range = align(addr_range.range, NVK_SSBO_BOUNDS_CHECK_ALIGNMENT);
 
    return (union nvk_buffer_descriptor) { .addr = {
@@ -337,7 +340,7 @@ write_ssbo_desc(struct nvk_descriptor_writer *w,
    struct nvk_addr_range addr_range =
       nvk_buffer_addr_range(buffer, info->offset, info->range);
 
-   const union nvk_buffer_descriptor desc = ssbo_desc(addr_range);
+   const union nvk_buffer_descriptor desc = ssbo_desc(w->pdev, addr_range);
    write_desc(w, binding, elem, &desc, sizeof(desc));
 }
 
@@ -353,7 +356,7 @@ write_dynamic_ssbo_desc(struct nvk_descriptor_writer *w,
    const struct nvk_descriptor_set_binding_layout *binding_layout =
       &w->layout->binding[binding];
    w->set->dynamic_buffers[binding_layout->dynamic_buffer_index + elem] =
-      ssbo_desc(addr_range);
+      ssbo_desc(w->pdev, addr_range);
 }
 
 static struct nvk_edb_buffer_view_descriptor
@@ -1207,7 +1210,7 @@ nvk_GetDescriptorEXT(VkDevice _device,
             .range = pDescriptorInfo->data.pStorageBuffer->range,
          };
       }
-      union nvk_buffer_descriptor desc = ssbo_desc(addr_range);
+      union nvk_buffer_descriptor desc = ssbo_desc(pdev, addr_range);
       assert(sizeof(desc) <= dataSize);
       memcpy(pDescriptor, &desc, sizeof(desc));
       break;

src/nouveau/vulkan/nvk_instance.c

@@ -130,6 +130,15 @@ nvk_init_dri_options(struct nvk_instance *instance)
       instance->debug_flags |= NVK_DEBUG_ZERO_MEMORY;
 
    instance->app_layer = driQueryOptionstr(&instance->dri_options, "nvk_app_layer");
+
+   /* VKD3D really wants a min SSBO alignment of 4B instead of 16B because
+    * it's required for structure buffers.  If we don't do this, it'll hack
+    * it behind our back and we don't really want that.  We just have to trust
+    * it to not misalign 64-bit atomics.
+    */
+   instance->ssbo_align_4b =
+      (instance->vk.app_info.engine_name != NULL &&
+       strcmp(instance->vk.app_info.engine_name, "vkd3d") == 0);
 }
 
 VKAPI_ATTR VkResult VKAPI_CALL

src/nouveau/vulkan/nvk_instance.h

@@ -22,8 +22,15 @@ struct nvk_instance {
 
    uint8_t driver_build_sha[BLAKE3_KEY_LEN];
    uint32_t force_vk_vendor;
+   bool ssbo_align_4b;
 };
 
 VK_DEFINE_HANDLE_CASTS(nvk_instance, vk.base, VkInstance, VK_OBJECT_TYPE_INSTANCE)
 
+static inline uint32_t
+nvk_min_ssbo_alignment(const struct nvk_instance *instance)
+{
+   return instance->ssbo_align_4b ? 4 : 16;
+}
+
 #endif

src/nouveau/vulkan/nvk_physical_device.c

@@ -877,7 +877,7 @@ nvk_get_device_properties(const struct nvk_instance *instance,
       .minMemoryMapAlignment = os_page_size,
       .minTexelBufferOffsetAlignment = NVK_MIN_TEXEL_BUFFER_ALIGNMENT,
       .minUniformBufferOffsetAlignment = nvk_min_cbuf_alignment(info),
-      .minStorageBufferOffsetAlignment = NVK_MIN_SSBO_ALIGNMENT,
+      .minStorageBufferOffsetAlignment = nvk_min_ssbo_alignment(instance),
       .minTexelOffset = -8,
       .maxTexelOffset = 7,
       .minTexelGatherOffset = -32,

src/nouveau/vulkan/nvk_physical_device.h

@@ -39,6 +39,7 @@ struct nvk_physical_device {
    struct vk_physical_device vk;
    struct nv_device_info info;
    enum nvk_debug debug_flags;
+   bool ssbo_align_4b;
 
    struct nvkmd_pdev *nvkmd;
 

src/nouveau/vulkan/nvk_private.h

@@ -15,7 +15,6 @@
 #define NVK_MAX_DYNAMIC_BUFFERS 64
 #define NVK_MAX_RTS 8
 #define NVK_MAX_SAMPLES 8
-#define NVK_MIN_SSBO_ALIGNMENT 16
 #define NVK_MIN_TEXEL_BUFFER_ALIGNMENT 16
 #define NVK_MIN_UBO_ALIGNMENT 64
 #define NVK_MAX_VIEWPORTS 16

src/nouveau/vulkan/nvk_shader.c

@@ -7,6 +7,7 @@
 #include "nvk_cmd_buffer.h"
 #include "nvk_descriptor_set_layout.h"
 #include "nvk_device.h"
+#include "nvk_instance.h"
 #include "nvk_mme.h"
 #include "nvk_physical_device.h"
 #include "nvk_sampler.h"
@@ -68,6 +69,7 @@ shared_var_info(const struct glsl_type *type, unsigned *size, unsigned *align)
 uint64_t
 nvk_physical_device_compiler_flags(const struct nvk_physical_device *pdev)
 {
+   const struct nvk_instance *instance = nvk_physical_device_instance(pdev);
    bool no_cbufs = pdev->debug_flags & NVK_DEBUG_NO_CBUF;
    bool use_edb_buffer_views = nvk_use_edb_buffer_views(pdev);
    uint64_t nak_flags = nak_debug_flags(pdev->nak);
@@ -76,6 +78,7 @@ nvk_physical_device_compiler_flags(const struct nvk_physical_device *pdev)
 
    return ((uint64_t)no_cbufs << 12)
       | ((uint64_t)use_edb_buffer_views << 13)
+      | ((uint64_t)instance->ssbo_align_4b << 14)
       | (nak_flags << 48);
 }
 
@@ -138,13 +141,14 @@ nvk_get_spirv_options(struct vk_physical_device *vk_pdev,
 {
    const struct nvk_physical_device *pdev =
       container_of(vk_pdev, struct nvk_physical_device, vk);
+   const struct nvk_instance *instance = nvk_physical_device_instance(pdev);
 
    return (struct spirv_to_nir_options) {
       .ssbo_addr_format = nvk_ssbo_addr_format(pdev, rs),
       .phys_ssbo_addr_format = nir_address_format_64bit_global,
       .ubo_addr_format = nvk_ubo_addr_format(pdev, rs),
       .shared_addr_format = nir_address_format_32bit_offset,
-      .min_ssbo_alignment = NVK_MIN_SSBO_ALIGNMENT,
+      .min_ssbo_alignment = nvk_min_ssbo_alignment(instance),
       .min_ubo_alignment = nvk_min_cbuf_alignment(&pdev->info),
    };
 }