Shaders are allocated contiguously in memory for a pipeline and
the freelist.next pointer is a pointer to the pipeline now.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/14950>
Since shaders are allocated per pipeline, the trap handler shader
was not uploaded at all.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/14950>
For some games that take like 400 MiB of shader binaries, the
number of shader arenas ends up going >1500. Cut that down a bit
by using larger arenas.
8 MiB should still be decent with small BAR and should still cut
things down from ~1500 to ~50 buffers.
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/14591>
This introduces a small NIR pass that exports
VARYING_SLOT_PRIMITIVE_SHADING_RATE if RADV_FORCE_VRS is used,
instead of doing this in both backend compilers.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/14907>
This option only applies to relative shuffles (up/down/xor), and in a
moment we're going to add an option to lower normal shuffles, so rename
it.
While we're here, rename lower_shuffle() to lower_to_shuffle() for
similar reasons.
Reviewed-by: Danylo Piliaiev <dpiliaiev@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/14412>
This creates an internal shader_prim enum, I've fixed up most
users to use it instead of GL types.
don't store the enum in shader_info as it changes size, and confuses
other things.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/14605>
To avoid dragging gl.h into places it has no business being,
defined tessellation primitive mode to an enum.
This has a lot of fallout all over the place.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/14605>
Doing this for ir3 required adding a struct for limits of how much base to
fold in (which NTT wants as well for its case of shared vars), otherwise
the later work to lower to the 1<<9 word limit would emit more
instructions.
The shader-db results are that sometimes the reduction in NIR instruction
count results in the fewer sampler prefetches due to the shader being
estimated to be shorter (dota2, nexuiz):
total instructions in shared programs: 8996651 -> 8996776 (<.01%)
total cat5 in shared programs: 86561 -> 86577 (0.02%)
Reviewed-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/14023>
This splits application-provided FMA in vertex/geometry/tesselation/mesh
shaders.
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/14458>
Adreno GPUs has native instruction for unsigned and mixed dot_4x8 but
not signed dot product.
Signed-off-by: Danylo Piliaiev <dpiliaiev@igalia.com>
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Emma Anholt <emma@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/13986>
Mesh shaders use NGG, but the API allows many compute shader
features such as workgroups and shared memory.
Use the appropriate NIR lowerings for these, then
call ac_nir_lower_ngg_ms.
Signed-off-by: Timur Kristóf <timur.kristof@gmail.com>
Acked-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/13580>
This fixes dEQP-VK.glsl.texture_functions.query.texturequerylod.*.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/14147>
Instead of using module->nir or nir->info->name to determine if it's a
meta shader, use nir->info->internal.
This also has an effect of disabling printing of meta shaders with
NIR_DEBUG=print.
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/14087>
RGP expects shaders to be contiguous in memory, otherwise it explodes
because we have to generate huge captures with lot of holes.
This reduces capture sizes of Cyberpunk 2077 from ~3.5GiB to ~180MiB.
This should also help for future pipeline libraries.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/13690>
Instead of mixing compilation and upload. This will allow us to
upload all shader binaries contiguously in memory and also for future
pipeline libraries work.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/13690>
binary_out is never NULL and binaries are freed from the pipeline
after they are added to the cache.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/13690>
The shader locations are now directly stored in radv_shader_args which
makes sense because they are tied to the arguments. The locations are
then copied to radv_shader_info but they will be moved into a new
radv_shader_binary_info with upcoming changes.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/13542>
If the application first uses nontrivial divisors, the driver emits
the vertex shader VA to the upload BO rather than directly via the
user SGPRs locations. But, if the vertex input dynamic state changes,
the driver might select a different VS prolog that no longer needs
nontrivial divisors.
In this case, the driver needs to re-emit the prolog inputs because
otherwise the VS prolog will jump to the PC that is emitted via the
user SGPR locations, and the previous one was somewhere in the
upload BO...
This fixes a GPU hang with Bioshock and Zink.
Fixes: d9c7a17542 ("radv: enable VK_EXT_vertex_input_dynamic_state")
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/13377>
This doesn't actually use the functionality or implement prolog
compilation yet.
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11717>
Way faster than the previous one, especially with a large number of
shaders.
This doesn't have much of an effect right now, but the previous allocator
was expensive compared to the cost of compiling vertex shader prologs.
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11717>
This will allow us to reduce the size of radv_shader_info which is
stored in the cache entry.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12992>
This will allow us to avoid postprocessing binaries when they are
loaded from the shaders cache.
LLVM binaries already contain the shader config as part of the ELF,
so it's duplicated and increase the cache entry by 48 bytes. Though,
I don't think that should matter for LLVM.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12992>
This commit enables NGG culling on all GFX10.3 GPUs by default.
A new debug flag environment variable RADV_DEBUG=nonggc is added to
disable this feature on GPUs where it is enabled by default.
The previous perf test flag RADV_PERFTEST=nggc will not be needed on
GFX10.3 anymore but it can still be used to enable the feature on
GPUs where it isn't on by default.
Totals from 58239 (45.27% of 128647) affected shaders:
VGPRs: 1989752 -> 2049408 (+3.00%); split: -3.21%, +6.21%
SpillSGPRs: 675 -> 883 (+30.81%); split: -78.07%, +108.89%
CodeSize: 72205968 -> 153572764 (+112.69%)
LDS: 0 -> 227125248 (+inf%)
MaxWaves: 1614598 -> 1646934 (+2.00%); split: +3.08%, -1.08%
Instrs: 14202239 -> 29654042 (+108.80%)
Latency: 87986508 -> 136960419 (+55.66%); split: -0.23%, +55.89%
InvThroughput: 14444832 -> 21141875 (+46.36%); split: -0.01%, +46.37%
VClause: 340794 -> 493067 (+44.68%); split: -1.33%, +46.01%
SClause: 520983 -> 738636 (+41.78%); split: -0.25%, +42.03%
Copies: 775639 -> 2787382 (+259.37%)
Branches: 296911 -> 1225431 (+312.73%)
PreSGPRs: 1316896 -> 2057270 (+56.22%); split: -0.14%, +56.36%
PreVGPRs: 1473558 -> 1658432 (+12.55%); split: -1.44%, +13.99%
Signed-off-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/13086>