AlexIndustrial/mesa
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

radv: Fix budget calculations with large BAR.

Browse Source 
If we don't have a non-visible VRAM heap, we should be counting
our non-visible VRAM allocations to the visible-VRAM heap.

CC: mesa-stable
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6827>
This commit is contained in:
Bas Nieuwenhuizen
2020-09-23 02:14:30 +02:00
parent 9948ead3cd
commit 91f9bc18e0
2 changed files with 41 additions and 38 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/amd/vulkan/radv_device.c
+40 -38
View File
@@ -161,6 +161,13 @@ radv_get_vram_size(struct radv_physical_device *device)
return radv_get_adjusted_vram_size(device) - device->rad_info.vram_vis_size;
}
enum radv_heap {
RADV_HEAP_VRAM = 1 << 0,
RADV_HEAP_GTT = 1 << 1,
RADV_HEAP_VRAM_VIS = 1 << 2,
RADV_HEAP_MAX = 1 << 3,
};
static void
radv_physical_device_init_mem_types(struct radv_physical_device *device)
{
@@ -168,11 +175,13 @@ radv_physical_device_init_mem_types(struct radv_physical_device *device)
uint64_t vram_size = radv_get_vram_size(device);
int vram_index = -1, visible_vram_index = -1, gart_index = -1;
device->memory_properties.memoryHeapCount = 0;
device->heaps = 0;
/* Only get a VRAM heap if it is significant, not if it is a 16 MiB
* remainder above visible VRAM. */
if (vram_size > 0 && vram_size * 9 >= visible_vram_size) {
vram_index = device->memory_properties.memoryHeapCount++;
device->heaps |= RADV_HEAP_VRAM;
device->memory_properties.memoryHeaps[vram_index] = (VkMemoryHeap) {
.size = vram_size,
.flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT,
@@ -181,6 +190,7 @@ radv_physical_device_init_mem_types(struct radv_physical_device *device)
if (device->rad_info.gart_size > 0) {
gart_index = device->memory_properties.memoryHeapCount++;
device->heaps |= RADV_HEAP_GTT;
device->memory_properties.memoryHeaps[gart_index] = (VkMemoryHeap) {
.size = device->rad_info.gart_size,
.flags = 0,
@@ -189,6 +199,7 @@ radv_physical_device_init_mem_types(struct radv_physical_device *device)
if (visible_vram_size) {
visible_vram_index = device->memory_properties.memoryHeapCount++;
device->heaps |= RADV_HEAP_VRAM_VIS;
device->memory_properties.memoryHeaps[visible_vram_index] = (VkMemoryHeap) {
.size = visible_vram_size,
.flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT,
@@ -2205,10 +2216,6 @@ radv_get_memory_budget_properties(VkPhysicalDevice physicalDevice,
{
RADV_FROM_HANDLE(radv_physical_device, device, physicalDevice);
VkPhysicalDeviceMemoryProperties *memory_properties = &device->memory_properties;
uint64_t visible_vram_size = radv_get_visible_vram_size(device);
uint64_t vram_size = radv_get_vram_size(device);
uint64_t gtt_size = device->rad_info.gart_size;
uint64_t heap_budget, heap_usage;
/* For all memory heaps, the computation of budget is as follow:
* heap_budget = heap_size - global_heap_usage + app_heap_usage
@@ -2219,44 +2226,39 @@ radv_get_memory_budget_properties(VkPhysicalDevice physicalDevice,
* Note that the application heap usages are not really accurate (eg.
* in presence of shared buffers).
*/
for (int i = 0; i < device->memory_properties.memoryTypeCount; i++) {
uint32_t heap_index = device->memory_properties.memoryTypes[i].heapIndex;
unsigned mask = device->heaps;
unsigned heap = 0;
while (mask) {
uint64_t internal_usage = 0, total_usage = 0;
unsigned type = 1u << u_bit_scan(&mask);
if ((device->memory_domains[i] & RADEON_DOMAIN_VRAM) && (device->memory_flags[i] & RADEON_FLAG_NO_CPU_ACCESS)) {
heap_usage = device->ws->query_value(device->ws,
RADEON_ALLOCATED_VRAM);
heap_budget = vram_size -
MIN2(vram_size, device->ws->query_value(device->ws, RADEON_VRAM_USAGE)) +
heap_usage;
memoryBudget->heapBudget[heap_index] = heap_budget;
memoryBudget->heapUsage[heap_index] = heap_usage;
} else if (device->memory_domains[i] & RADEON_DOMAIN_VRAM) {
heap_usage = device->ws->query_value(device->ws,
RADEON_ALLOCATED_VRAM_VIS);
heap_budget = visible_vram_size -
MIN2(visible_vram_size, device->ws->query_value(device->ws, RADEON_VRAM_VIS_USAGE)) +
heap_usage;
memoryBudget->heapBudget[heap_index] = heap_budget;
memoryBudget->heapUsage[heap_index] = heap_usage;
} else {
assert(device->memory_domains[i] & RADEON_DOMAIN_GTT);
heap_usage = device->ws->query_value(device->ws,
RADEON_ALLOCATED_GTT);
heap_budget = gtt_size -
device->ws->query_value(device->ws, RADEON_GTT_USAGE) +
heap_usage;
memoryBudget->heapBudget[heap_index] = heap_budget;
memoryBudget->heapUsage[heap_index] = heap_usage;
switch(type) {
case RADV_HEAP_VRAM:
internal_usage = device->ws->query_value(device->ws, RADEON_ALLOCATED_VRAM);
total_usage = device->ws->query_value(device->ws, RADEON_VRAM_USAGE);
break;
case RADV_HEAP_VRAM_VIS:
internal_usage = device->ws->query_value(device->ws, RADEON_ALLOCATED_VRAM_VIS);
if (!(device->heaps & RADV_HEAP_VRAM))
internal_usage += device->ws->query_value(device->ws, RADEON_ALLOCATED_VRAM);
total_usage = device->ws->query_value(device->ws, RADEON_VRAM_VIS_USAGE);
break;
case RADV_HEAP_GTT:
internal_usage = device->ws->query_value(device->ws, RADEON_ALLOCATED_GTT);
total_usage = device->ws->query_value(device->ws, RADEON_GTT_USAGE);
break;
}
uint64_t free_space = device->memory_properties.memoryHeaps[heap].size -
MIN2(device->memory_properties.memoryHeaps[heap].size,
total_usage);
memoryBudget->heapBudget[heap] = free_space + internal_usage;
memoryBudget->heapUsage[heap] = internal_usage;
++heap;
}
assert(heap == memory_properties->memoryHeapCount);
/* The heapBudget and heapUsage values must be zero for array elements
* greater than or equal to
* VkPhysicalDeviceMemoryProperties::memoryHeapCount.
src/amd/vulkan/radv_private.h
+1
View File
@@ -319,6 +319,7 @@ struct radv_physical_device {
VkPhysicalDeviceMemoryProperties memory_properties;
enum radeon_bo_domain memory_domains[VK_MAX_MEMORY_TYPES];
enum radeon_bo_flag memory_flags[VK_MAX_MEMORY_TYPES];
unsigned heaps;
drmPciBusInfo bus_info;