Move blob from compiler/ to util/
[mesa.git] / src / intel / vulkan / anv_batch_chain.c
index a41305bc6d134d5f2f9b55a6b43fc5bdc7a393d8..3b1f713ef5b1ae170be27ba4bc6a07d24fb8c9a5 100644 (file)
@@ -418,9 +418,9 @@ anv_batch_bo_link(struct anv_cmd_buffer *cmd_buffer,
                   struct anv_batch_bo *next_bbo,
                   uint32_t next_bbo_offset)
 {
-   MAYBE_UNUSED const uint32_t bb_start_offset =
+   const uint32_t bb_start_offset =
       prev_bbo->length - GEN8_MI_BATCH_BUFFER_START_length * 4;
-   MAYBE_UNUSED const uint32_t *bb_start = prev_bbo->bo.map + bb_start_offset;
+   ASSERTED const uint32_t *bb_start = prev_bbo->bo.map + bb_start_offset;
 
    /* Make sure we're looking at a MI_BATCH_BUFFER_START */
    assert(((*bb_start >> 29) & 0x07) == 0);
@@ -500,7 +500,7 @@ anv_cmd_buffer_surface_base_address(struct anv_cmd_buffer *cmd_buffer)
 {
    struct anv_state *bt_block = u_vector_head(&cmd_buffer->bt_block_states);
    return (struct anv_address) {
-      .bo = &anv_binding_table_pool(cmd_buffer->device)->block_pool.bo,
+      .bo = anv_binding_table_pool(cmd_buffer->device)->block_pool.bo,
       .offset = bt_block->offset,
    };
 }
@@ -678,8 +678,8 @@ anv_cmd_buffer_alloc_binding_table(struct anv_cmd_buffer *cmd_buffer,
       return (struct anv_state) { 0 };
 
    state.offset = cmd_buffer->bt_next;
-   state.map = anv_binding_table_pool(device)->block_pool.map +
-      bt_block->offset + state.offset;
+   state.map = anv_block_pool_map(&anv_binding_table_pool(device)->block_pool,
+                                  bt_block->offset + state.offset);
 
    cmd_buffer->bt_next += state.alloc_size;
 
@@ -1035,6 +1035,12 @@ _compare_bo_handles(const void *_bo1, const void *_bo2)
    return (*bo1)->gem_handle - (*bo2)->gem_handle;
 }
 
+static VkResult
+anv_execbuf_add_bo_set(struct anv_execbuf *exec,
+                       struct set *deps,
+                       uint32_t extra_flags,
+                       const VkAllocationCallbacks *alloc);
+
 static VkResult
 anv_execbuf_add_bo(struct anv_execbuf *exec,
                    struct anv_bo *bo,
@@ -1124,36 +1130,46 @@ anv_execbuf_add_bo(struct anv_execbuf *exec,
          }
       }
 
-      if (relocs->deps && relocs->deps->entries > 0) {
-         const uint32_t entries = relocs->deps->entries;
-         struct anv_bo **bos =
-            vk_alloc(alloc, entries * sizeof(*bos),
-                     8, VK_SYSTEM_ALLOCATION_SCOPE_COMMAND);
-         if (bos == NULL)
-            return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+      return anv_execbuf_add_bo_set(exec, relocs->deps, extra_flags, alloc);
+   }
 
-         struct anv_bo **bo = bos;
-         set_foreach(relocs->deps, entry) {
-            *bo++ = (void *)entry->key;
-         }
+   return VK_SUCCESS;
+}
 
-         qsort(bos, entries, sizeof(struct anv_bo*), _compare_bo_handles);
+/* Add BO dependencies to execbuf */
+static VkResult
+anv_execbuf_add_bo_set(struct anv_execbuf *exec,
+                       struct set *deps,
+                       uint32_t extra_flags,
+                       const VkAllocationCallbacks *alloc)
+{
+   if (!deps || deps->entries <= 0)
+      return VK_SUCCESS;
 
-         VkResult result = VK_SUCCESS;
-         for (bo = bos; bo < bos + entries; bo++) {
-            result = anv_execbuf_add_bo(exec, *bo, NULL, extra_flags, alloc);
-            if (result != VK_SUCCESS)
-               break;
-         }
+   const uint32_t entries = deps->entries;
+   struct anv_bo **bos =
+      vk_alloc(alloc, entries * sizeof(*bos),
+               8, VK_SYSTEM_ALLOCATION_SCOPE_COMMAND);
+   if (bos == NULL)
+      return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 
-         vk_free(alloc, bos);
+   struct anv_bo **bo = bos;
+   set_foreach(deps, entry) {
+      *bo++ = (void *)entry->key;
+   }
 
-         if (result != VK_SUCCESS)
-            return result;
-      }
+   qsort(bos, entries, sizeof(struct anv_bo*), _compare_bo_handles);
+
+   VkResult result = VK_SUCCESS;
+   for (bo = bos; bo < bos + entries; bo++) {
+      result = anv_execbuf_add_bo(exec, *bo, NULL, extra_flags, alloc);
+      if (result != VK_SUCCESS)
+         break;
    }
 
-   return VK_SUCCESS;
+   vk_free(alloc, bos);
+
+   return result;
 }
 
 static VkResult
@@ -1227,7 +1243,7 @@ adjust_relocations_to_state_pool(struct anv_state_pool *pool,
     * relocations that point to the pool bo with the correct offset.
     */
    for (size_t i = 0; i < relocs->num_relocs; i++) {
-      if (relocs->reloc_bos[i] == &pool->block_pool.bo) {
+      if (relocs->reloc_bos[i] == pool->block_pool.bo) {
          /* Adjust the delta value in the relocation to correctly
           * correspond to the new delta.  Initially, this value may have
           * been negative (if treated as unsigned), but we trust in
@@ -1335,7 +1351,7 @@ relocate_cmd_buffer(struct anv_cmd_buffer *cmd_buffer,
     * given time.  The only option is to always relocate them.
     */
    anv_reloc_list_apply(cmd_buffer->device, &cmd_buffer->surface_relocs,
-                        &cmd_buffer->device->surface_state_pool.block_pool.bo,
+                        cmd_buffer->device->surface_state_pool.block_pool.bo,
                         true /* always relocate surface states */);
 
    /* Since we own all of the batch buffers, we know what values are stored
@@ -1364,11 +1380,64 @@ setup_execbuf_for_cmd_buffer(struct anv_execbuf *execbuf,
 
    adjust_relocations_from_state_pool(ss_pool, &cmd_buffer->surface_relocs,
                                       cmd_buffer->last_ss_pool_center);
-   VkResult result = anv_execbuf_add_bo(execbuf, &ss_pool->block_pool.bo,
-                                        &cmd_buffer->surface_relocs, 0,
-                                        &cmd_buffer->device->alloc);
-   if (result != VK_SUCCESS)
-      return result;
+   VkResult result;
+   struct anv_bo *bo;
+   if (cmd_buffer->device->instance->physicalDevice.use_softpin) {
+      anv_block_pool_foreach_bo(bo, &ss_pool->block_pool) {
+         result = anv_execbuf_add_bo(execbuf, bo, NULL, 0,
+                                     &cmd_buffer->device->alloc);
+         if (result != VK_SUCCESS)
+            return result;
+      }
+      /* Add surface dependencies (BOs) to the execbuf */
+      anv_execbuf_add_bo_set(execbuf, cmd_buffer->surface_relocs.deps, 0,
+                             &cmd_buffer->device->alloc);
+
+      /* Add the BOs for all memory objects */
+      list_for_each_entry(struct anv_device_memory, mem,
+                          &cmd_buffer->device->memory_objects, link) {
+         result = anv_execbuf_add_bo(execbuf, mem->bo, NULL, 0,
+                                     &cmd_buffer->device->alloc);
+         if (result != VK_SUCCESS)
+            return result;
+      }
+
+      struct anv_block_pool *pool;
+      pool = &cmd_buffer->device->dynamic_state_pool.block_pool;
+      anv_block_pool_foreach_bo(bo, pool) {
+         result = anv_execbuf_add_bo(execbuf, bo, NULL, 0,
+                                     &cmd_buffer->device->alloc);
+         if (result != VK_SUCCESS)
+            return result;
+      }
+
+      pool = &cmd_buffer->device->instruction_state_pool.block_pool;
+      anv_block_pool_foreach_bo(bo, pool) {
+         result = anv_execbuf_add_bo(execbuf, bo, NULL, 0,
+                                     &cmd_buffer->device->alloc);
+         if (result != VK_SUCCESS)
+            return result;
+      }
+
+      pool = &cmd_buffer->device->binding_table_pool.block_pool;
+      anv_block_pool_foreach_bo(bo, pool) {
+         result = anv_execbuf_add_bo(execbuf, bo, NULL, 0,
+                                     &cmd_buffer->device->alloc);
+         if (result != VK_SUCCESS)
+            return result;
+      }
+   } else {
+      /* Since we aren't in the softpin case, all of our STATE_BASE_ADDRESS BOs
+       * will get added automatically by processing relocations on the batch
+       * buffer.  We have to add the surface state BO manually because it has
+       * relocations of its own that we need to be sure are processsed.
+       */
+      result = anv_execbuf_add_bo(execbuf, ss_pool->block_pool.bo,
+                                  &cmd_buffer->surface_relocs, 0,
+                                  &cmd_buffer->device->alloc);
+      if (result != VK_SUCCESS)
+         return result;
+   }
 
    /* First, we walk over all of the bos we've seen and add them and their
     * relocations to the validate list.
@@ -1525,6 +1594,7 @@ anv_cmd_buffer_execbuf(struct anv_device *device,
                        VkFence _fence)
 {
    ANV_FROM_HANDLE(anv_fence, fence, _fence);
+   UNUSED struct anv_physical_device *pdevice = &device->instance->physicalDevice;
 
    struct anv_execbuf execbuf;
    anv_execbuf_init(&execbuf);
@@ -1539,6 +1609,7 @@ anv_cmd_buffer_execbuf(struct anv_device *device,
 
       switch (impl->type) {
       case ANV_SEMAPHORE_TYPE_BO:
+         assert(!pdevice->has_syncobj);
          result = anv_execbuf_add_bo(&execbuf, impl->bo, NULL,
                                      0, &device->alloc);
          if (result != VK_SUCCESS)
@@ -1546,6 +1617,7 @@ anv_cmd_buffer_execbuf(struct anv_device *device,
          break;
 
       case ANV_SEMAPHORE_TYPE_SYNC_FILE:
+         assert(!pdevice->has_syncobj);
          if (in_fence == -1) {
             in_fence = impl->fd;
          } else {
@@ -1595,6 +1667,7 @@ anv_cmd_buffer_execbuf(struct anv_device *device,
 
       switch (impl->type) {
       case ANV_SEMAPHORE_TYPE_BO:
+         assert(!pdevice->has_syncobj);
          result = anv_execbuf_add_bo(&execbuf, impl->bo, NULL,
                                      EXEC_OBJECT_WRITE, &device->alloc);
          if (result != VK_SUCCESS)
@@ -1602,6 +1675,7 @@ anv_cmd_buffer_execbuf(struct anv_device *device,
          break;
 
       case ANV_SEMAPHORE_TYPE_SYNC_FILE:
+         assert(!pdevice->has_syncobj);
          need_out_fence = true;
          break;
 
@@ -1636,6 +1710,7 @@ anv_cmd_buffer_execbuf(struct anv_device *device,
 
       switch (impl->type) {
       case ANV_FENCE_TYPE_BO:
+         assert(!pdevice->has_syncobj_wait);
          result = anv_execbuf_add_bo(&execbuf, &impl->bo.bo, NULL,
                                      EXEC_OBJECT_WRITE, &device->alloc);
          if (result != VK_SUCCESS)
@@ -1655,10 +1730,20 @@ anv_cmd_buffer_execbuf(struct anv_device *device,
       }
    }
 
-   if (cmd_buffer)
+   if (cmd_buffer) {
+      if (unlikely(INTEL_DEBUG & DEBUG_BATCH)) {
+         struct anv_batch_bo **bo = u_vector_head(&cmd_buffer->seen_bbos);
+
+         device->cmd_buffer_being_decoded = cmd_buffer;
+         gen_print_batch(&device->decoder_ctx, (*bo)->bo.map,
+                         (*bo)->bo.size, (*bo)->bo.offset, false);
+         device->cmd_buffer_being_decoded = NULL;
+      }
+
       result = setup_execbuf_for_cmd_buffer(&execbuf, cmd_buffer);
-   else
+   } else {
       result = setup_empty_execbuf(&execbuf, device);
+   }
 
    if (result != VK_SUCCESS)
       return result;
@@ -1699,6 +1784,7 @@ anv_cmd_buffer_execbuf(struct anv_device *device,
    }
 
    if (fence && fence->permanent.type == ANV_FENCE_TYPE_BO) {
+      assert(!pdevice->has_syncobj_wait);
       /* BO fences can't be shared, so they can't be temporary. */
       assert(fence->temporary.type == ANV_FENCE_TYPE_NONE);
 
@@ -1716,6 +1802,7 @@ anv_cmd_buffer_execbuf(struct anv_device *device,
    }
 
    if (result == VK_SUCCESS && need_out_fence) {
+      assert(!pdevice->has_syncobj_wait);
       int out_fence = execbuf.execbuf.rsvd2 >> 32;
       for (uint32_t i = 0; i < num_out_semaphores; i++) {
          ANV_FROM_HANDLE(anv_semaphore, semaphore, out_semaphores[i]);