Move the bits we want to share between generations from fd3_program to
ir3_shader. So overall structure is:
fdN_shader_stateobj -> ir3_shader -> ir3_shader_variant -> ir3
|- ...
\- ir3_shader_variant -> ir3
So the ir3_shader becomes the topmost generation neutral object, which
manages the set of variants each of which generates, compiles, and
assembles it's own ir.
There is a bit of additional renaming to s/fd3_compiler/ir3_compiler/,
etc.
Keep the split between the gallium level stateobj and the shader helper
object because it might be a good idea to pre-compute some generation
specific register values (ie. anything that is independent of linking).
Signed-off-by: Rob Clark <robclark@freedesktop.org>
First step of reoganization split out compiler (so it can be shared
between a3xx and a4xx). Rename ir3_shader -> ir3 (since we'll want
the name ir3_shader for a higher level object).
Signed-off-by: Rob Clark <robclark@freedesktop.org>
The scheduler also needs to be aware of predicate register (p0) in
addition to address register (a0).
Signed-off-by: Rob Clark <robclark@freedesktop.org>
It seems like for the most part, different behaviors, workarounds, etc,
should be conditional on GPU patch revision (ie. a320.0 vs a320.2)
rather than GPU id (a320 vs a330).
Signed-off-by: Rob Clark <robclark@freedesktop.org>
The fixed size heap is a remnant of the fdre-a3xx assembler. Yet it is
convenient for being able to free the entire data structure in one shot
without worrying about leaking nodes.
Change it to dynamically grow the heap size (adding chunks) as needed so
we don't have an artificial upper limit on shader size (other than hw
limits) and don't always have to allocate worst-case size.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
The GBM_DRIVERS_PATH environment variable is not documented, and only
used to set the location of gbm drivers, while LIBGL_DRIVERS_PATH is
used for everything else, and is documented.
Generally this split leads to confusion as to why gbm doesn't work.
This patch will read LIBGL_DRIVERS_PATH as a fallback if
GBM_DRIVERS_PATH is not set.
The comments clearly indicate that using LIBGL_DRIVERS_PATH is
preferred over GBM_DRIVERS_PATH.
v2: - Use GBM_DRIVERS_PATH as a fallback
v3: [jordan.l.justen@intel.com] - Make LIBGL_DRIVERS_PATH the fallback
Signed-off-by: Dylan Baker <baker.dylan.c@gmail.com>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Kristian Høgsberg <krh@bitplanet.net>
According to a quick micro-benchmark, this new version is 20% faster on my
Haswell laptop.
v2: Removed the XXX note about x86_64 from the comment
v3: Use an intrinsic instead of an __asm__ block. This should give us MSVC
support for free.
v4: Enable it for all x86_64 builds, not just with USE_X86_64_ASM
Signed-off-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
... to eliminate an ELSE instruction followed immediately by an ENDIF.
instructions in affected programs: 704 -> 700 (-0.57%)
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
In a situation where double-register values are used, the phi nodes can
still end up being u32 values. They all get merged into one RA node
though. When fixing up the merge (which comes after the phi node), the
phi node's def would get fixed, but not its sources which would remain
at the low register value.
This maintains the invariant that a phi node's defs and sources are
allocated the same register.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Since intel is always going to be little-endian,
GL_UNSIGNED_INT_8_8_8_8_REV is the same as GL_UNSIGNED_BYTE for RGBA and
BGRA textures, so the same acceleration code will work. We might as well
use it.
Signed-off-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
In Piglit's EXT_framebuffer_multisample/alpha-to-coverage-dual-src-blend
test, key->nr_color_regions == 2, but the dual source blend FB write has
ir->target set to 0. So we failed to set "Last Render Target Select" on
any FB write message.
We only emit one FB write per render target, so my comment about setting
LastRT on every FB write directed at the last color region is a bit...
misinformed. According to the documentation, depth buffer writes and
scoreboard updates happen on the FB write with LastRT set, so I believe
we want to set it only once.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Cc: "10.2" <mesa-stable@lists.freedesktop.org>
This will be useful for INTEL_DEBUG=optimizer in the vec4 backend, which
needs to know whether it's currently processing a VS or GS. It isn't
worth adding virtual methods for this case.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Dropping this helps most lines fit in an 80 column terminal. The
absence of WE_normal also helps call attention to WE_all, where
something unusual is going on.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Don't assert (debug builds) or assign random uninitialized value for
predicate register (p0).. that screws up kill, etc.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
Accuracy of some operations was recently improved in the R600 backend,
at the cost of slower code. This is required for compute shaders,
but not for graphics shaders. Add unsafe-fp-math hint to make LLVM
generate faster but possibly less accurate code.
Piglit didn't indicate any regressions.
Reviewed-by: Tom Stellard <thomas.stellard@amd.com>
Now that we know that the pool is defragmented, we positively know
that allocated + unallocated will be the total size of the
current pool plus all the items that will be promoted. So we only
need to grow the pool once.
This will allow us to just add the new items to the end of the
item_list without the need of looking for a place to the new item.
Reviewed-by: Tom Stellard <thomas.stellard@amd.com>
This way we can avoid defragmenting the pool, even if it is needed
to defragment it, and looping again through the list of unallocated
items.
Reviewed-by: Tom Stellard <thomas.stellard@amd.com>
This patch adds a new member to the pool to track its status.
For now it is used only for the 'fragmented' status, but if
needed it could be used for more statuses.
The pool will be considered fragmented if: An item that isn't
the last is freed or demoted.
This 'strategy' has a problem, although it shouldn't cause any bug.
If for example we have two items, A and B. We choose to free A first,
now the pool will have the 'fragmented' status. If we now free B,
the pool will retain its 'fragmented' status even if it isn't
fragmented.
Reviewed-by: Tom Stellard <thomas.stellard@amd.com>
This new function will move items forward in the pool, so that
there's no gap between them, effectively defragmenting the pool.
For now this function is a bit dumb as it just moves items
forward without trying to see if other items in the pool could
fit in the gaps.
Reviewed-by: Tom Stellard <thomas.stellard@amd.com>
This function will be used in the future by compute_memory_defrag
to move items forward in the pool.
It does so by first checking for overlaping ranges, if the ranges
don't overlap it will copy the contents directly. If they overlap
it will try first to make a temporary buffer, if this buffer fails
to allocate, it will finally fall back to a mapping.
Note that it will only be needed to move items forward, it only
checks for overlapping ranges in that case. If needed, it can
easily be added by changing the first if.
Reviewed-by: Tom Stellard <thomas.stellard@amd.com>
