This increases the wasted memory to 140 MB for DeusExMD, still below
the original number. The advantage is that we now get 2 MB pages for more
buffers and fewer total buffers allocated by the kernel, enabling faster
GPU page translation and slightly lower kernel overhead.
Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8683>
Instead of aligning slab allocations to powers of two (e.g. 129K -> 256K),
implement slab allocations with 3/4 of power of two sizes to reduce
overallocation. (e.g. 129K -> 192K)
The limitation is that the alignment must be 1/3rd of the allocation size.
DeusExMD allocates 2.1 GB of VRAM. Without this, slabs waste 194 MB due
to alignment, i.e. 9.2%. This commit reduces the waste to 102 MB, i.e. 4.9%.
Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8683>
Slabs always allocate the next power of two size from their pools. This
wastes memory if the size is not a power of two.
bo->base.size is overwritten because the default is the allocated power of
two size, but we need the real size to compute the wasted size in
amdgpu_bo_slab_destroy. entry_size is added to the hole in pb_slab_entry
to hold the real entry size.
Like other memory stats, no atomics are used.
Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8683>
Left unchecked, an app that just did an endless series of draws could
result in VSC buffer sizes >4GB, which doesn't work out well.
This limit is semi-arbitrary (ie. it is lower than hw limit, but 32*8MB
seems a bit excessive and not a limit that you'd hit in the real world).
Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8842>
According to VK_KHR_depth_stencil_resolve spec (see
VUID-VkSubpassDescriptionDepthStencilResolve-pDepthStencilResolveAttachment-03182):
"If the VkFormat of pDepthStencilResolveAttachment has a stencil
component, then the VkFormat of pDepthStencilAttachment must have a
stencil component with the same number of bits and numerical
type"
The issue with D32_SFLOAT_S8_UINT format is that it is implemented as
two planes, so we need to execute the separate_stencil path in
tu_emit_blit() to resolve its stencil component into S8_UINT image.
Fixes the following tests:
dEQP-VK.renderpass2.depth_stencil_resolve.image_2d_32_32.samples_2.d32_sfloat_s8_uint.compatibility_depth_zero_stencil_zero_testing_stencil
dEQP-VK.renderpass2.depth_stencil_resolve.image_2d_32_32.samples_2.d32_sfloat_s8_uint_separate_layouts.compatibility_depth_zero_stencil_zero_testing_stencil
Signed-off-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Reviewed-by: Jonathan Marek <jonathan@marek.ca>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8527>
According to VK_KHR_depth_stencil_resolve spec (see
VUID-VkSubpassDescriptionDepthStencilResolve-pDepthStencilResolveAttachment-03182)
"If the VkFormat of pDepthStencilResolveAttachment has a stencil
component, then the VkFormat of pDepthStencilAttachment must have a
stencil component with the same number of bits and numerical type"
That means that we can resolve MSAA depth/stencil to a stencil only
image only if the stencil component matches with same number of bits
and type.
Although the driver only supports VK_RESOLVE_MODE_SAMPLE_ZERO_BIT
resolve mode, it was doing a sample average when resolving a MSAA
D24_UNORM_S8_UINT image to S8_UINT.
Fixes the following tests:
dEQP-VK.renderpass2.depth_stencil_resolve.image_2d_32_32.samples_2.d24_unorm_s8_uint.compatibility_depth_zero_stencil_zero_testing_s
tencil
dEQP-VK.renderpass2.depth_stencil_resolve.image_2d_32_32.samples_2.d24_unorm_s8_uint_separate_layouts.compatibility_depth_zero_stenc
il_zero_testing_stencil
Signed-off-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Reviewed-by: Jonathan Marek <jonathan@marek.ca>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8527>
Vertex attribute bounds checking is supposed to be done per-attribute:
is_oob = index * stride + attrib_offset + attrib_size > buffer_size
but we were obtaining num_records by dividing the buffer size by the
stride, making it per-vertex:
is_oob = index * stride + (stride - 1) >= buffer_size
An example from Dead Cells (Wine) is:
attribute bindings: 0, 1, 2
attribute formats: r32g32, r32g32, r32g32b32a32
attribute offsets: 0, 0, 0
binding buffers: all the same buffer
binding offsets: 0, 8, 16
binding sizes: 128, 120, 112
binding strides: 32, 32, 32
Workaround this issue without switching to per-attribute descriptors by
rounding up the division. This is still incorrect, but it should now no
longer consider in-bounds attributes out-of-bounds.
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/3796
Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/4199
Cc: mesa-stable
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8835>
We can just always specify the stride parameter regardless of whether
an alignment was forced or not. This fixes some issues where it is not
straightforward to detect the need to specify stride by checking the
buffer width (e.g. imported dmabuf to be used as texture).
Signed-off-by: Erico Nunes <nunes.erico@gmail.com>
Reviewed-by: Vasily Khoruzhick <anarsoul@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8549>
Marek Olšák