[Mesa-dev] [PATCH] radv: Change memory type order for GPUs without dedicated VRAM

[Koenig, Christian]

Am 02.06.19 um 12:32 schrieb Alex Smith:
> Put the uncached GTT type at a higher index than the visible VRAM type,
> rather than having GTT first.
>
> When we don't have dedicated VRAM, we don't have a non-visible VRAM
> type, and the property flags for GTT and visible VRAM are identical.
> According to the spec, for types with identical flags, we should give
> the one with better performance a lower index.
>
> Previously, apps which follow the spec guidance for choosing a memory
> type would have picked the GTT type in preference to visible VRAM (all
> Feral games will do this), and end up with lower performance.
>
> On a Ryzen 5 2500U laptop (Raven Ridge), this improves average FPS in
> the Rise of the Tomb Raider benchmark by up to ~30%. Tested a couple of
> other (Feral) games and saw similar improvement on those as well.

Well that patch doesn't looks like a good idea to me.

Using VRAM over uncached GTT should have something between no and only 
minimal performance difference on APU.

To make things even worse VRAM is still needed for scanout and newer 
laptops have only a very very low default setting (32 or 16MB). So you 
can end up in VRAM clashing on those systems.

Can you check some kernel statistics to figure out what exactly is going 
on here?

Regards,
Christian.

>
> Signed-off-by: Alex Smith <asmith at feralinteractive.com>
> ---
> I noticed that the memory types advertised on my Raven laptop looked a
> bit odd so played around with it and found this. I'm not sure if it is
> actually expected that the performance difference between visible VRAM
> and GTT is so large, seeing as it's not dedicated VRAM, but the results
> are clear (and consistent, tested multiple times).
> ---
>   src/amd/vulkan/radv_device.c | 18 +++++++++++++++---
>   1 file changed, 15 insertions(+), 3 deletions(-)
>
> diff --git a/src/amd/vulkan/radv_device.c b/src/amd/vulkan/radv_device.c
> index 3cf050ed220..d36ee226ebd 100644
> --- a/src/amd/vulkan/radv_device.c
> +++ b/src/amd/vulkan/radv_device.c
> @@ -171,12 +171,11 @@ radv_physical_device_init_mem_types(struct radv_physical_device *device)
>   			.heapIndex = vram_index,
>   		};
>   	}
> -	if (gart_index >= 0) {
> +	if (gart_index >= 0 && device->rad_info.has_dedicated_vram) {
>   		device->mem_type_indices[type_count] = RADV_MEM_TYPE_GTT_WRITE_COMBINE;
>   		device->memory_properties.memoryTypes[type_count++] = (VkMemoryType) {
>   			.propertyFlags = VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
> -			VK_MEMORY_PROPERTY_HOST_COHERENT_BIT |
> -			(device->rad_info.has_dedicated_vram ? 0 : VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT),
> +			VK_MEMORY_PROPERTY_HOST_COHERENT_BIT,
>   			.heapIndex = gart_index,
>   		};
>   	}
> @@ -189,6 +188,19 @@ radv_physical_device_init_mem_types(struct radv_physical_device *device)
>   			.heapIndex = visible_vram_index,
>   		};
>   	}
> +	if (gart_index >= 0 && !device->rad_info.has_dedicated_vram) {
> +		/* Put GTT after visible VRAM for GPUs without dedicated VRAM
> +		 * as they have identical property flags, and according to the
> +		 * spec, for types with identical flags, the one with greater
> +		 * performance must be given a lower index. */
> +		device->mem_type_indices[type_count] = RADV_MEM_TYPE_GTT_WRITE_COMBINE;
> +		device->memory_properties.memoryTypes[type_count++] = (VkMemoryType) {
> +			.propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
> +			VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
> +			VK_MEMORY_PROPERTY_HOST_COHERENT_BIT,
> +			.heapIndex = gart_index,
> +		};
> +	}
>   	if (gart_index >= 0) {
>   		device->mem_type_indices[type_count] = RADV_MEM_TYPE_GTT_CACHED;
>   		device->memory_properties.memoryTypes[type_count++] = (VkMemoryType) {