intel_decoder_init() initializes intel_batch_decode_ctx so later
we can call decode functions but it depends on data stored in
brw/elk_isa_info but that was being allocated in stack
of intel_decoder_init() then when the decode functions were executed
it was accessing garbage at the brw/elk_isa_info memory.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Fixes: ec2d20a70d ("intel/tools: Add helpers for decoder_init/disasm")
Signed-off-by: José Roberto de Souza <jose.souza@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/34776>
Or current render target cache setting is to key on the binding table
index, meaning the HW associates a number in the range [0, 7] to a
RENDER_SURFACE_STATE description. If you want change the render target
0 between 2 draw calls, you need to insert a PIPE_CONTROL in between
the 2 draw calls with pb-stall + rt-flush in order to flush an writes
to a previous RENDER_SURFACE_STATE that has now becomed disassociated
with the [0, 7] number.
This PIPE_CONTROL taking care of the flush is dealt with in
cmd_buffer_maybe_flush_rt_writes(). This function diffs the current
BTI setup for render targets (first 0 to 7 BTIs) with what the next
fragment shader wants.
The issue here is we might have a render pass with 0 color attachments
and yet in 98cdb9349a we added one pointing to the render target 0,
but in the emit_binding_table() when we finally program the BTI, we
check the render pass color count and program a null surface state
instead of an actual surface state. And this leads to hangs because
the render target cache will end up with inconsistent state data.
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Fixes: 98cdb9349a ("anv: ensure null-rt bit in compiler isn't used when there is ds attachment")
Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/12955
Reviewed-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/34603>
ANV_BO_ALLOC_MAPPED are internal allocated bos that need mmap() but as
internally we don't do any cflush() we need to make sure those are also
ANV_BO_ALLOC_HOST_COHERENT.
Checking for ANV_BO_ALLOC_HOST_CACHED could lead a cached+uncoherent
bo being allocated internally with ANV_BO_ALLOC_MAPPED.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Signed-off-by: José Roberto de Souza <jose.souza@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/34778>
It was failing in the first assert of anv_device_alloc_bo() because
it has ANV_BO_ALLOC_MAPPED but it don't have ANV_BO_ALLOC_HOST_COHERENT or
ANV_BO_ALLOC_HOST_CACHED(this second one is wrong and fixed in the next
patch).
LMEM is always write-combine, even SMEM on discrete GPU is always
write-back + coherent because the PCI bus protocol snooping at CPU
caches and that behavior can't be disabled.
So we can add this coherent flag without any side effects.
The ANV_BO_ALLOC_MAPPED is needed for ANV_BO_SLAB_HEAP_LMEM_SMEM
because to trigger SMEM+LMEM in anv_device_alloc_bo() we need
ANV_BO_ALLOC_MAPPED or ANV_BO_ALLOC_LOCAL_MEM_CPU_VISIBLE but the
second one is mostly used with small PCI bar discrete GPUs.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Fixes: dabb012423 ("anv: Implement anv_slab_bo and enable memory pool")
Signed-off-by: José Roberto de Souza <jose.souza@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/34778>
On a7xx, the max constlen for compute is increased to 512 vec4s or 8KB,
however the size of the LB was not increased beyond 40KB. A quick
calculation shows that 8KB of consts multiplied by 2 banks plus the
API maximum of 32KB shared memory would exceed 40KB. This means that
we can't always use a constlen of 512, and sometimes have to fall back
to 256 when a lot of shared memory is in use.
In the future, we can use similar calculations to figure out how much
"extra" shared memory is available for the backend to spill to, but we
currently don't support spilling to shared memory.
Fixes: 5879eaac18 ("ir3: Increase compute const size on a7xx")
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/34746>
This change relocates the fdot lowering from the generic NIR to the lima,
since lima is the only consumer of this particular lowering. This avoids
potential conflicts with the similar fdot lowering already present in
nir_lower_alu_width.
Signed-off-by: Christian Gmeiner <cgmeiner@igalia.com>
Reviewed-by: Erico Nunes <nunes.erico@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/34757>
Round up allocations to nearest 2MB interval if this increases
the allocation by no more than 1.33x. This reduces page count but
at the cost of extra memory consumption. Optimization only applied
to MTL(Xe KMD only)/LNL platforms, which are particularly impacted by
page misses.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Signed-off-by: Felix DeGrood <felix.j.degrood@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/33558>
BOs larger than 1MB don't go memory pool due the size but applications
tend to use a lot of VkMemory with size larger than 1MB so to reduce
the number of pages and improve performance here I'm aligning the size
of BOs larger than 1MB to 64kb, this allows 64kb pages to be used at
least on Xe KMD.
This bring substantial perfomance benefit in exchange of a small
memory waste.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Signed-off-by: José Roberto de Souza <jose.souza@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/33558>
MTL and newer integrated platforms has a performance gain when using
transparent huge pages, because of the fixed address requirement
we can't use slab for this case but we can change the initial pool
size to 2MB so all allocations get the transparent huge page
optimization.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Signed-off-by: José Roberto de Souza <jose.souza@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/33558>
Karol Herbst