panvk: Add support for storage buffers

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Jason Ekstrand <jason.ekstrand@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15248>

src/panfrost/vulkan/panvk_cmd_buffer.c

@@ -100,7 +100,10 @@ panvk_CmdBindDescriptorSets(VkCommandBuffer commandBuffer,
          pDynamicOffsets += set->layout->num_dynoffsets;
       }
 
-      if (set->layout->num_ubos || set->layout->num_dynoffsets)
+      if (set->layout->num_ssbos)
+         descriptors_state->dirty |= PANVK_DYNAMIC_SSBO;
+
+      if (set->layout->num_ubos || set->layout->num_ssbos || set->layout->num_dynoffsets)
          descriptors_state->ubos = 0;
 
       if (set->layout->num_textures)

src/panfrost/vulkan/panvk_descriptor_set.c

@@ -436,6 +436,7 @@ panvk_descriptor_set_destroy(struct panvk_device *device,
    vk_free(&device->vk.alloc, set->textures);
    vk_free(&device->vk.alloc, set->samplers);
    vk_free(&device->vk.alloc, set->ubos);
+   vk_free(&device->vk.alloc, set->ssbos);
    vk_free(&device->vk.alloc, set->descs);
    vk_object_free(&device->vk, NULL, set);
 }

src/panfrost/vulkan/panvk_private.h

@@ -345,11 +345,18 @@ struct panvk_descriptor {
    };
 };
 
+struct panvk_buffer_desc {
+   struct panvk_buffer *buffer;
+   VkDeviceSize offset;
+   VkDeviceSize size;
+};
+
 struct panvk_descriptor_set {
    struct vk_object_base base;
    struct panvk_descriptor_pool *pool;
    const struct panvk_descriptor_set_layout *layout;
    struct panvk_descriptor *descs;
+   struct panvk_buffer_desc *ssbos;
    void *ubos;
    void *samplers;
    void *textures;
@@ -478,10 +485,12 @@ enum panvk_dynamic_state_bits {
    PANVK_DYNAMIC_STENCIL_WRITE_MASK = 1 << 7,
    PANVK_DYNAMIC_STENCIL_REFERENCE = 1 << 8,
    PANVK_DYNAMIC_DISCARD_RECTANGLE = 1 << 9,
-   PANVK_DYNAMIC_ALL = (1 << 10) - 1,
+   PANVK_DYNAMIC_SSBO = 1 << 10,
+   PANVK_DYNAMIC_ALL = (1 << 11) - 1,
 };
 
 struct panvk_descriptor_state {
+   uint32_t dirty;
    struct {
       const struct panvk_descriptor_set *set;
       struct panfrost_ptr dynoffsets;

src/panfrost/vulkan/panvk_vX_cmd_buffer.c

@@ -293,9 +293,35 @@ panvk_per_arch(cmd_alloc_tls_desc)(struct panvk_cmd_buffer *cmdbuf, bool gfx)
    }
 }
 
+static void
+panvk_sysval_upload_ssbo_info(struct panvk_cmd_buffer *cmdbuf,
+                              unsigned ssbo_id,
+                              struct panvk_cmd_bind_point_state *bind_point_state,
+                              union panvk_sysval_data *data)
+{
+   const struct panvk_pipeline *pipeline = bind_point_state->pipeline;
+   const struct panvk_descriptor_state *desc_state = &bind_point_state->desc_state;
+
+   for (unsigned s = 0; s < pipeline->layout->num_sets; s++) {
+      unsigned ssbo_offset = pipeline->layout->sets[s].ssbo_offset;
+      unsigned num_ssbos = pipeline->layout->sets[s].layout->num_ssbos;
+      const struct panvk_buffer_desc *ssbo = NULL;
+
+      if (ssbo_id >= ssbo_offset && ssbo_id < (ssbo_offset + num_ssbos))
+         ssbo = &desc_state->sets[s].set->ssbos[ssbo_id - ssbo_offset];
+
+      if (ssbo) {
+         data->u64[0] = ssbo->buffer->bo->ptr.gpu + ssbo->offset;
+         data->u32[2] = ssbo->size == VK_WHOLE_SIZE ? ssbo->buffer->size - ssbo->offset : ssbo->size;
+      }
+   }
+}
+
 static void
 panvk_cmd_upload_sysval(struct panvk_cmd_buffer *cmdbuf,
-                        unsigned id, union panvk_sysval_data *data)
+                        unsigned id,
+                        struct panvk_cmd_bind_point_state *bind_point_state,
+                        union panvk_sysval_data *data)
 {
    switch (PAN_SYSVAL_TYPE(id)) {
    case PAN_SYSVAL_VIEWPORT_SCALE:
@@ -311,6 +337,14 @@ panvk_cmd_upload_sysval(struct panvk_cmd_buffer *cmdbuf,
    case PAN_SYSVAL_BLEND_CONSTANTS:
       memcpy(data->f32, cmdbuf->state.blend.constants, sizeof(data->f32));
       break;
+   case PAN_SYSVAL_SSBO:
+      /* This won't work with dynamic SSBO indexing. We might want to
+       * consider storing SSBO mappings in a separate UBO if we need to
+       * support
+       * VkPhysicalDeviceVulkan12Features.shaderStorageBufferArrayNonUniformIndexing.
+       */
+      panvk_sysval_upload_ssbo_info(cmdbuf, PAN_SYSVAL_ID(id), bind_point_state, data);
+      break;
    default:
       unreachable("Invalid static sysval");
    }
@@ -326,11 +360,12 @@ panvk_cmd_prepare_sysvals(struct panvk_cmd_buffer *cmdbuf,
    if (!pipeline->num_sysvals)
       return;
 
+   uint32_t dirty = cmdbuf->state.dirty | desc_state->dirty;
+
    for (unsigned i = 0; i < ARRAY_SIZE(desc_state->sysvals); i++) {
       unsigned sysval_count = pipeline->sysvals[i].ids.sysval_count;
       if (!sysval_count || pipeline->sysvals[i].ubo ||
-          (desc_state->sysvals[i] &&
+          (desc_state->sysvals[i] && !(dirty & pipeline->sysvals[i].dirty_mask)))
-           !(cmdbuf->state.dirty & pipeline->sysvals[i].dirty_mask)))
          continue;
 
       struct panfrost_ptr sysvals =
@@ -339,7 +374,7 @@ panvk_cmd_prepare_sysvals(struct panvk_cmd_buffer *cmdbuf,
 
       for (unsigned s = 0; s < pipeline->sysvals[i].ids.sysval_count; s++) {
          panvk_cmd_upload_sysval(cmdbuf, pipeline->sysvals[i].ids.sysvals[s],
-                                 &data[s]);
+                                 bind_point_state, &data[s]);
       }
 
       desc_state->sysvals[i] = sysvals.gpu;
@@ -782,7 +817,7 @@ panvk_per_arch(CmdDraw)(VkCommandBuffer commandBuffer,
    }
 
    /* Clear the dirty flags all at once */
-   cmdbuf->state.dirty = 0;
+   desc_state->dirty = cmdbuf->state.dirty = 0;
 }
 
 VkResult

src/panfrost/vulkan/panvk_vX_descriptor_set.c

@@ -72,6 +72,14 @@ panvk_per_arch(descriptor_set_create)(struct panvk_device *device,
          goto err_free_set;
    }
 
+   if (layout->num_ssbos) {
+      set->ssbos = vk_zalloc(&device->vk.alloc,
+                            sizeof(*set->ssbos) * layout->num_ssbos, 8,
+                            VK_OBJECT_TYPE_DESCRIPTOR_SET);
+      if (!set->ssbos)
+         goto err_free_set;
+   }
+
    if (layout->num_samplers) {
       set->samplers = vk_zalloc(&device->vk.alloc,
                                 pan_size(SAMPLER) * layout->num_samplers, 8,
@@ -106,6 +114,7 @@ panvk_per_arch(descriptor_set_create)(struct panvk_device *device,
 err_free_set:
    vk_free(&device->vk.alloc, set->textures);
    vk_free(&device->vk.alloc, set->samplers);
+   vk_free(&device->vk.alloc, set->ssbos);
    vk_free(&device->vk.alloc, set->ubos);
    vk_free(&device->vk.alloc, set->descs);
    vk_object_free(&device->vk, NULL, set);
@@ -173,6 +182,17 @@ panvk_set_buffer_info_desc(struct panvk_descriptor *desc,
    desc->buffer_info.range = pBufferInfo->range;
 }
 
+static void
+panvk_set_buffer_desc(struct panvk_buffer_desc *bdesc,
+                      const VkDescriptorBufferInfo *pBufferInfo)
+{
+   VK_FROM_HANDLE(panvk_buffer, buffer, pBufferInfo->buffer);
+
+   bdesc->buffer = buffer;
+   bdesc->offset = pBufferInfo->offset;
+   bdesc->size = pBufferInfo->range;
+}
+
 static void
 panvk_per_arch(set_ubo_desc)(void *ubo,
                              const VkDescriptorBufferInfo *pBufferInfo)
@@ -284,6 +304,11 @@ panvk_per_arch(write_descriptor_set)(struct panvk_device *dev,
          break;
 
       case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER:
+         for (unsigned i = 0; i < ndescs; i++) {
+            unsigned ssbo = binding_layout->ssbo_idx + dest_offset + i;
+            panvk_set_buffer_desc(&set->ssbos[ssbo], &pDescriptorWrite->pBufferInfo[src_offset + i]);
+         }
+         break;
       case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC:
          for (unsigned i = 0; i < ndescs; i++)
             panvk_set_buffer_info_desc(&descs[i], &pDescriptorWrite->pBufferInfo[src_offset + i]);

src/panfrost/vulkan/panvk_vX_pipeline.c

@@ -230,6 +230,9 @@ panvk_pipeline_builder_alloc_static_state_bo(struct panvk_pipeline_builder *buil
          case PAN_SYSVAL_VIEWPORT_OFFSET:
             pipeline->sysvals[i].dirty_mask |= PANVK_DYNAMIC_VIEWPORT;
             break;
+         case PAN_SYSVAL_SSBO:
+            pipeline->sysvals[i].dirty_mask |= PANVK_DYNAMIC_SSBO;
+            break;
          default:
             break;
          }