OpenHarmony/third_party_mesa3d
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

anv: Store CPU-side fence information in the BO

Browse Source 
This reduces the number of allocations a bit and cuts back on memory usage.
Kind-of a micro-optimization but it also makes the error handling a bit
simpler so it seems like a win.
This commit is contained in:
Jason Ekstrand
2016-03-07 13:45:25 -08:00
parent f61d40adc2
commit 2308891ede
1 changed files with 15 additions and 18 deletions
Download Patch File Download Diff File Expand all files Collapse all files

33
src/intel/vulkan/anv_device.c
View File

@@ -1383,29 +1383,31 @@ VkResult anv_CreateFence(
VkFence* pFence) VkFence* pFence)
{ {
ANV_FROM_HANDLE(anv_device, device, _device); ANV_FROM_HANDLE(anv_device, device, _device);
struct anv_bo fence_bo;
struct anv_fence *fence; struct anv_fence *fence;
struct anv_batch batch; struct anv_batch batch;
VkResult result; VkResult result;
assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FENCE_CREATE_INFO); assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FENCE_CREATE_INFO);
fence = anv_alloc2(&device->alloc, pAllocator, sizeof(*fence), 8, result = anv_bo_pool_alloc(&device->batch_bo_pool, &fence_bo);
VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
if (fence == NULL)
return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
result = anv_bo_pool_alloc(&device->batch_bo_pool, &fence->bo);
if (result != VK_SUCCESS) if (result != VK_SUCCESS)
goto fail; return result;
batch.next = batch.start = fence->bo.map; /* Fences are small. Just store the CPU data structure in the BO. */
fence = fence_bo.map;
fence->bo = fence_bo;
/* Place the batch after the CPU data but on its own cache line. */
const uint32_t batch_offset = align_u32(sizeof(*fence), CACHELINE_SIZE);
batch.next = batch.start = fence->bo.map + batch_offset;
batch.end = fence->bo.map + fence->bo.size; batch.end = fence->bo.map + fence->bo.size;
anv_batch_emit(&batch, GEN7_MI_BATCH_BUFFER_END); anv_batch_emit(&batch, GEN7_MI_BATCH_BUFFER_END);
anv_batch_emit(&batch, GEN7_MI_NOOP); anv_batch_emit(&batch, GEN7_MI_NOOP);
if (!device->info.has_llc) { if (!device->info.has_llc) {
assert(((uintptr_t) fence->bo.map & CACHELINE_MASK) == 0); assert(((uintptr_t) batch.start & CACHELINE_MASK) == 0);
assert(batch.next - fence->bo.map <= CACHELINE_SIZE); assert(batch.next - batch.start <= CACHELINE_SIZE);
__builtin_ia32_mfence(); __builtin_ia32_mfence();
__builtin_ia32_clflush(fence->bo.map); __builtin_ia32_clflush(fence->bo.map);
} }
@@ -1421,8 +1423,8 @@ VkResult anv_CreateFence(
fence->execbuf.buffers_ptr = (uintptr_t) fence->exec2_objects; fence->execbuf.buffers_ptr = (uintptr_t) fence->exec2_objects;
fence->execbuf.buffer_count = 1; fence->execbuf.buffer_count = 1;
fence->execbuf.batch_start_offset = 0; fence->execbuf.batch_start_offset = batch.start - fence->bo.map;
fence->execbuf.batch_len = batch.next - fence->bo.map; fence->execbuf.batch_len = batch.next - batch.start;
fence->execbuf.cliprects_ptr = 0; fence->execbuf.cliprects_ptr = 0;
fence->execbuf.num_cliprects = 0; fence->execbuf.num_cliprects = 0;
fence->execbuf.DR1 = 0; fence->execbuf.DR1 = 0;
@@ -1438,11 +1440,6 @@ VkResult anv_CreateFence(
*pFence = anv_fence_to_handle(fence); *pFence = anv_fence_to_handle(fence);
return VK_SUCCESS; return VK_SUCCESS;
fail:
anv_free2(&device->alloc, pAllocator, fence);
return result;
} }
void anv_DestroyFence( void anv_DestroyFence(
@@ -1453,8 +1450,8 @@ void anv_DestroyFence(
ANV_FROM_HANDLE(anv_device, device, _device); ANV_FROM_HANDLE(anv_device, device, _device);
ANV_FROM_HANDLE(anv_fence, fence, _fence); ANV_FROM_HANDLE(anv_fence, fence, _fence);
assert(fence->bo.map == fence);
anv_bo_pool_free(&device->batch_bo_pool, &fence->bo); anv_bo_pool_free(&device->batch_bo_pool, &fence->bo);
anv_free2(&device->alloc, pAllocator, fence);
} }
VkResult anv_ResetFences( VkResult anv_ResetFences(