Because we must maintain an exec_mask even if there's currently nothing
on the mask stack, we can still have an exec_mask at the end of the program.
Effectively, this mask should be set back to default when returning from main.
Without relying on END/RET opcode (I think it's valid to have neither) it is
actually difficult to do this, as there doesn't seem any reasonable place to
do it, so instead let's just say the exec_mask is invalid outside main (which
it really is effectively).
The problem is that geometry shader called end_primitive outside the shader
(in the epilogue), and as a result used a bogus mask, leading to bugs if we
had to set the (somewhat misnamed) ret_in_main bit anywhere. So just avoid
the mask combining function when called from outside the shader.
Reviewed-by: Zack Rusin <zackr@vmware.com>
Instead of reducing masks to 0/1 simply use the mask directly as -1.
Also use some signed comparison instead of unsigned (as far as I understand
these values have to be (very) small and signed means llvm doesn't have to
apply additional logic to do the unsigned comparisons the cpu can't do).
Saves a couple of instructions in some test geometry shader here.
v2: that was a bit to much optimization, don't skip combining the masks...
Reviewed-by: Zack Rusin <zackr@vmware.com>
My previous attempt at doing so double-failed miserably (minification of
zero still gives one, and even if it would not the value was never written
anyway).
While here also rename the confusingly named int_vec bld as we have int vecs
of different sizes, and rename need_nr_mips (as this also changes out-of-bounds
behavior) to is_sviewinfo too.
Reviewed-by: Zack Rusin <zackr@vmware.com>
d3d10 has no notion of distinct array resources neither at the resource nor
sampler view level. However, shader dcl of resources certainly has, and
d3d10 expects resinfo to return the values according to that - in particular
a resource might have been a 1d texture with some array layers, then the
sampler view might have only used 1 layer so it can be accessed both as 1d
or 1d array texture (I think - the former definitely works). resinfo of a
resource decleared as array needs to return number of array layers but
non-array resource needs to return 0 (and not 1). Hence fix this by passing
the target from the shader decl to emit_size_query and use that (in case of
OpenGL the target will come from the instruction itself).
Could probably do the same for actual sampling, though it may not matter there
(as the bogus components will essentially get clamped away), possibly could
wreak havoc though if it REALLY doesn't match (which is of course an error
but still).
Reviewed-by: Zack Rusin <zackr@vmware.com>
Specifically, must return 0 for non-existent mip levels (and non-existent
textures which is an unsolved problem) for everything but total mip count.
Reviewed-by: Zack Rusin <zackr@vmware.com>
This is wrong both for OpenGL and d3d. (In fact clamping is a side effect
of converting to depth format, so this should really do quantization too
at least in d3d10 for the comparisons to be truly correct.)
Reviewed-by: Zack Rusin <zackr@vmware.com>
Clearly the returned values need to be per-element if the lod is per element.
Does not actually change behavior yet.
Reviewed-by: Zack Rusin <zackr@vmware.com>
For d3d10 and ARB_robust_buffer_access_behavior, we are required to return
0 for out-of-bounds coordinates (for which we can just enable the code already
there was just disabled). Additionally, also need to return 0 for
out-of-bounds mip level and out-of-bounds layer. This changes the logic
so instead of clamping the level/layer, an out-of-bound mask is computed
instead in this case (actual clamping then can be omitted just like with
coordinates, since we set the fetch offset to zero if that happens anyway).
Reviewed-by: Zack Rusin <zackr@vmware.com>
Since llvm -3.4svn r187618, TargetOptions doesn't provide
RealignStack, so only enable it with llvm<3.4
This option must now be specified using function attributes, see LLVM
commit r187618
Reviewed-by: Tom Stellard <thomas.stellard@amd.com>
llvm shifts are undefined for shift counts exceeding (or matching) bit width,
so need to apply a mask for the tgsi shift instructions.
v2: only use mask for the tgsi shift instructions, not for the build shift
helpers. None of the internal callers need this behavior, and while llvm can
optimize away the masking for constants there are legitimate cases where it
might not be able to do so even if we know that shift count must be smaller
than type width (currently all such callers do not use the build shift
helpers).
Reviewed-by: Zack Rusin <zackr@vmware.com>
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Previously we were using truncation, which gives the correct result
only for numbers in [0.5-1.0] range (because there's no mantissa bits
to do any rounding there).
This is frequently hit (and probably only used there) when converting
fragment depth to depth format (d24s8 etc.) or otherwise dealing with
depth format.
v2: as spotted by Jose, get rid of extra type (src_type is already unsigned).
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Commit 8c3d3622d9 introduced a new assertion,
but since it causes lp_test_conv failures remove it again and let's hope
we don't really hit bugs caused by the potentially bogus code (it is possible
the assert() caught some cases which work correctly too).
Just like the UNORM case we need to use round to nearest, not trunc.
(There's also another problem, we're using the formula for SNORM->float
which will produce a value below -1.0 for the most negative value which
according to both OpenGL and d3d10 would need clamping. However, no actual
failures have been observed due to that hence keep cheating on that.)
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
unlike OpenGL, the texel swizzle is embedded in the instruction, so honor
that.
(Technically we now execute both the sampler_view swizzle and the
per-instruction swizzle but this should be quite ok.)
v2: add documentation note as it's not obvious.
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
TargetOptions::NoFramePointerElimNonLeaf was removed in LLVM 3.4
r187093.
Signed-off-by: Vinson Lee <vlee@freedesktop.org>
Reviewed-by: José Fonseca <jfonseca@vmware.com>
There are earlier returns for PIPE_FUNC_NEVER and PIPE_FUNC_ALWAYS. The
switch value of 'func' cannot be either of those values.
Fixes "Logically dead code" defects reported by Coverity.
Signed-off-by: Vinson Lee <vlee@freedesktop.org>
Reviewed-by: José Fonseca <jfonseca@vmware.com>
Same as log2_safe, which means that it can handle infs, 0s and
nans.
Signed-off-by: Zack Rusin <zackr@vmware.com>
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Only the floating point operarators change everything else
is the same so it makes sense to share the code.
Signed-off-by: Zack Rusin <zackr@vmware.com>
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
sin/cos for anything not finite is nan and everything else has
to be between [-1, 1].
Signed-off-by: Zack Rusin <zackr@vmware.com>
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
That means that if input is:
* - less than zero (to and including -inf) then NaN will be returned
* - equal to zero (-denorm, -0, +0 or +denorm), then -inf will be returned
* - +infinity, then +infinity will be returned
* - NaN, then NaN will be returned
It's a separate function because the checks are a little bit costly
and in most cases are likely unnecessary.
Signed-off-by: Zack Rusin <zackr@vmware.com>
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
exp(0) has to be exactly 1, exp(-inf) has to be 0, exp(inf) has
to be inf and exp(nan) has to be nan, this fixes all of those
cases.
Signed-off-by: Zack Rusin <zackr@vmware.com>
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Both D3D10 and OpenCL say that if one the inputs is nan then
the other should be returned. To preserve that behavior
the patch fixes both the sse and the non-sse paths in both
functions and adds helper code for handling nans.
Signed-off-by: Zack Rusin <zackr@vmware.com>
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Use "or" instead of "add" (this is a classic select sequence, which at
least newer llvm versions can actually recognize (3.2+?), and the "add"
might prevent that - and we really don't want an add instead of an or with
avx if it isn't recognized (even without avx logic ops might be cheaper)).
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
MCJIT is the only supported LLVM JIT on AArch64 and ARM (the regular
JIT has bit-rotted badly on ARM and doesn't exist on AArch64.)
Signed-off-by: Kyle McMartin <kyle@redhat.com>
Signed-off-by: Dave Airlie <airlied@gmail.com>
Just use the new conversion functions to do the work. The way it's plugged
in into the blend code is quite hacktastic but follows all the same hacks
as used by packed float format already.
Only support 4x8bit srgb formats (rgba/rgbx plus swizzle), 24bit formats never
worked anyway in the blend code and are thus disabled, and I don't think anyone
is interested in L8/L8A8. Would need even more hacks otherwise.
Unless I'm missing something, this is the last feature except MSAA needed for
OpenGL 3.0, and for OpenGL 3.1 as well I believe.
v2: prettify a bit, use separate function for packing.
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Some lame compilers can't do exp2f() and as far as I can tell they can't do
exp2() (with doubles) neither so instead of providing some workaround for
that (wouldn't actually be too bad just replace with pow) and since it is
used with a constant only just use the precalculated constant.
srgb-to-linear is using 3rd degree polynomial for now which should be _just_
good enough. Reverse is using some rational polynomials and is quite accurate,
though not hooked into llvmpipe's blend code yet and hence unused (untested).
Using a table might also be an option (for srgb-to-linear especially).
This does not enable any new features yet because EXT_texture_srgb was already
supported via util_format fallbacks, but performance was lacking probably due
to the external function call (the table used by the util_format_srgb code may
not be all that much slower on its own).
Some performance figures (taken from modified gloss, replaced both base and
sphere texture to use GL_SRGB instead of GL_RGB, measured on 1Ghz Sandy Bridge,
the numbers aren't terribly accurate):
normal gloss, aos, 8-wide: 47 fps
normal gloss, aos, 4-wide: 48 fps
normal gloss, forced to soa, 8-wide: 48 fps
normal gloss, forced to soa, 4-wide: 47 fps
patched gloss, old code, soa, 8-wide: 21 fps
patched gloss, old code, soa, 4-wide: 24 fps
patched gloss, new code, soa, 8-wide: 41 fps
patched gloss, new code, soa, 4-wide: 38 fps
So there's a performance hit but it seems acceptable, certainly better
than using the fallback.
Note the new code only works for 4x8bit srgb formats, others (L8/L8A8) will
continue to use the old util_format fallback, because I can't be bothered
to write code for formats noone uses anyway (as decoding is done as part of
lp_build_unpack_rgba_soa which can only handle block type width of 32).
Compressed srgb formats should get their own path though eventually (it is
going to be expensive in any case, first decompress, then convert).
No piglit regressions.
v2: use lp_build_polynomial instead of ad-hoc polynomial construction, also
since keeping both linear to srgb functions for now make sure both are
compiled (since they share quite some code just integrate into the same
function).
v3: formatting fixes and bugfix in the complicated (disabled) linear-to-srgb
path.
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
We had to disable fast rsqrt before because it wasn't precise enough etc.
However in situations when we know we're not going to need more precision
we can still use a fast rsqrt (which can be several times faster than
the quite expensive sqrt). Hence introduce a new helper which does exactly
that - it is probably not useful calling it in some situations if there's
no fast rsqrt available so make it queryable if it's available too.
v2: use fast_rsqrt consistently instead of rsqrt_fast, fix indentation,
let rsqrt use fast_rsqrt.
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
GLSL spec says that rsq is undefined for src<=0, but the D3D10
spec says it needs to be a NaN, so lets stop taking an absolute
value of the source which completely breaks that behavior. For
the gl program we can simply insert an extra abs instrunction
which produces the desired behavior there.
Signed-off-by: Zack Rusin <zackr@vmware.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
TGSI_OPCODE_KIL and KILP had confusing names. The former was conditional
kill (if any src component < 0). The later was unconditional kill.
At one time KILP was supposed to work with NV-style condition
codes/predicates but we never had that in TGSI.
This patch renames both opcodes:
TGSI_OPCODE_KIL -> KILL_IF (kill if src.xyzw < 0)
TGSI_OPCODE_KILP -> KILL (unconditional kill)
Note: I didn't just transpose the opcode names to help ensure that I
didn't miss updating any code anywhere.
I believe I've updated all the relevant code and comments but I'm
not 100% sure that some drivers had this right in the first place.
For example, the radeon driver might have llvm.AMDGPU.kill and
llvm.AMDGPU.kilp mixed up. Driver authors should review their code.
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
KILP is really unconditional fragment kill.
We've had KIL and KILP transposed forever. I'll fix that next.
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
The D3D10 spec is very explicit about treatment of denorm floats and
the behavior is exactly the same for them as it would be for -0 or
+0. This makes our shading code match that behavior, since OpenGL
doesn't care and on a few cpu's it's faster (worst case the same).
Float16 conversions will likely break but we'll fix them in a follow
up commit.
Signed-off-by: Zack Rusin <zackr@vmware.com>
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
The logic for choosing number of lods was bogus.
(The code should ultimately handle the case of only one lod even with multiple
quads but currently can't.)
It is perfectly valid for the swizzle to be bigger than 2. For example the
texel offsets could be
SAMPLE ..., IMM[0].zzz
What is not correct is for chan_index to be bigger than 2.
Trivial.
The assertion was always broken but the code unused until enabling the
per-element lod code. Fixes piglit texelFetch vs isampler1D and similar
tests (only run with GL 3.0 version override).
d3d10 requires per-pixel lod calculations for explicit lod, lod bias and
explicit derivatives, and we should probably do it for OpenGL too - at least
if they are used from vertex or geometry shaders (so doesn't apply to lod
bias) this doesn't just affect neighboring pixels.
Some code was already there to handle this so fix it up and enable it.
There will no doubt be a performance hit unfortunately, we could do better
if we'd knew we had a real vector shift instruction (with variable shift
count) but this requires AVX2 on x86 (or a AMD Bulldozer family cpu).
Don't do anything for lod bias and explicit derivatives yet, though
no special magic should be needed for them neither.
Likewise, the size query is still broken just the same.
v2: Use information if lod is a (broadcast) scalar or not. The idea would be
to base this on the actual value, for now just pretend it's a scalar in fs
and not a scalar otherwise (so, per-pixel lod is only used in gs/vs but same
code is generated for fs as before).
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
If reg->Register.Indirect is true then the immediate is not truly a
constant LLVM expression.
There is no performance regression in using LLVMBuildBitCast, as it will
fallback to LLVMConstBitCast internally when the argument is a constant.
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Reviewed-by: Zack Rusin <zackr@vmware.com>
Squashed commit of the following:
commit 0857a7e105bfcbc4d1431b2cc56612094c747ca3
Author: Richard Sandiford <r.sandiford@uk.ibm.com>
Date: Tue Jun 18 12:25:07 2013 -0400
gallivm: Fix lp_build_rgba8_to_fi32_soa for big endian
Reviewed-by: Adam Jackson <ajax@redhat.com>
Signed-off-by: Richard Sandiford <r.sandiford@uk.ibm.com>
commit 0d65131649a8aa140e2db228ba779d685c4333e3
Author: Richard Sandiford <r.sandiford@uk.ibm.com>
Date: Tue Jun 18 12:25:07 2013 -0400
gallivm: Fix big-endian machines
This adds a bit-shift count to the format table, and adds the concept of
vector or bitwise alignment on gathers.
Reviewed-by: Adam Jackson <ajax@redhat.com>
Signed-off-by: Richard Sandiford <r.sandiford@uk.ibm.com>
commit 9740bda9b7dc894b629ed38be9b51059ce90818f
Author: Richard Sandiford <r.sandiford@uk.ibm.com>
Date: Tue Jun 18 12:25:07 2013 -0400
llvmpipe: Fix convert_to_blend_type on big-endian
Reviewed-by: Adam Jackson <ajax@redhat.com>
Signed-off-by: Richard Sandiford <r.sandiford@uk.ibm.com>
commit ae037c2de0f029e4e99371c0de25560484f0d8df
Author: Richard Sandiford <r.sandiford@uk.ibm.com>
Date: Tue Jun 18 12:25:06 2013 -0400
util: Convert color pack to packed formats
This fixes them on big-endian.
Reviewed-by: Adam Jackson <ajax@redhat.com>
Signed-off-by: Richard Sandiford <r.sandiford@uk.ibm.com>
commit 5b05ac0c89ae092ea8ba5bba9f739708d7396b5c
Author: Richard Sandiford <r.sandiford@uk.ibm.com>
Date: Tue Jun 18 12:25:06 2013 -0400
graw-xlib: Convert to packed formats
Reviewed-by: Adam Jackson <ajax@redhat.com>
Signed-off-by: Richard Sandiford <r.sandiford@uk.ibm.com>
commit 51396e7d098cb6ff794391cf11afe4dbf86dbea0
Author: Richard Sandiford <r.sandiford@uk.ibm.com>
Date: Tue Jun 18 12:25:06 2013 -0400
format: Convert to packed formats
Reviewed-by: Adam Jackson <ajax@redhat.com>
Signed-off-by: Richard Sandiford <r.sandiford@uk.ibm.com>
commit 417b60bc66eb450e68a92ab0e47f76e292b385e6
Author: Adam Jackson <ajax@redhat.com>
Date: Tue Jun 18 12:25:06 2013 -0400
st/dri: Convert to packed formats
Reviewed-by: Adam Jackson <ajax@redhat.com>
Signed-off-by: Richard Sandiford <r.sandiford@uk.ibm.com>
commit 0934b2e022a5e0847d312c40734e2b44cac52fd8
Author: Richard Sandiford <r.sandiford@uk.ibm.com>
Date: Tue Jun 18 12:25:06 2013 -0400
st/xlib: Convert to packed formats
Reviewed-by: Adam Jackson <ajax@redhat.com>
Signed-off-by: Richard Sandiford <r.sandiford@uk.ibm.com>
commit a307ea3c3716a706963acce7966b5e405ba11db9
Author: Richard Sandiford <r.sandiford@uk.ibm.com>
Date: Tue Jun 18 12:25:06 2013 -0400
gbm: Convert to packed formats
Reviewed-by: Adam Jackson <ajax@redhat.com>
Signed-off-by: Richard Sandiford <r.sandiford@uk.ibm.com>
commit 53eebdd253e1960a645ea278f31d7ef6a6cf4aeb
Author: Richard Sandiford <r.sandiford@uk.ibm.com>
Date: Tue Jun 18 12:25:06 2013 -0400
tests: Convert to packed formats
Reviewed-by: Adam Jackson <ajax@redhat.com>
Signed-off-by: Richard Sandiford <r.sandiford@uk.ibm.com>
commit 2f77fe3ee524945eacd546efcac34f7799fb3124
Author: Adam Jackson <ajax@redhat.com>
Date: Tue Jun 18 13:07:37 2013 -0400
gallium: Document packed formats
Signed-off-by: Adam Jackson <ajax@redhat.com>
commit 1f1017159ce951f922210a430de9229f91f62714
Author: Richard Sandiford <r.sandiford@uk.ibm.com>
Date: Tue Jun 18 12:25:06 2013 -0400
gallium: Introduce 32-bit packed format names
These are for interacting with buffers natively described in terms of
bit shifts, like X11 visuals:
uint32_t xyzw8888 = (x << 0) | (y << 8) | (z << 16) | (w << 24);
Define these in terms of (endian-dependent) aliases to the array-style
format names.
Reviewed-by: Adam Jackson <ajax@redhat.com>
Signed-off-by: Richard Sandiford <r.sandiford@uk.ibm.com>
commit 6cc7ab1ee66ed668da78c1d951dfd7782b4e786a
Author: Adam Jackson <ajax@redhat.com>
Date: Mon Jun 3 12:10:32 2013 -0400
gallium: Document format name conventions
v2:
- Fix a channel name thinko (Michel Dänzer)
- Elaborate on SCALED versus INT
- Add links to DirectX and FOURCC docs
Signed-off-by: Adam Jackson <ajax@redhat.com>
commit df4d269e7fb62051a3c029b84147465001e5776e
Author: Adam Jackson <ajax@redhat.com>
Date: Tue Jun 18 12:25:06 2013 -0400
gallivm: Remove all notion of byte-swapping
Signed-off-by: Adam Jackson <ajax@redhat.com>
Signed-off-by: Adam Jackson <ajax@redhat.com>
lp_build_add and lp_build_sub have fallback code for cases
that cannot be handled by known intrinsics. For UNORM formats,
this code was using modulo rather than saturating arithmetic.
This fixes some rendering issues for a gnome session on System z.
It also fixes various piglit tests on z, such as
spec/ARB_color_buffer_float/GL_RGBA8-render.
The patch deliberately doesn't tackle the more complicated
SNORM case.
Tested against piglit on x86_64 and System z with no regressions.
Reviewed-by: Adam Jackson <ajax@redhat.com>
Signed-off-by: Richard Sandiford <rsandifo@linux.vnet.ibm.com>
We use 128bit vector interleave for untwiddling in the blend code (with
256bit vectors). llvm generates terrible code for this for some reason,
so instead of generating a shuffle for 2 128bit vectors use a
extract/insert shuffle instead (it only seems to matter we're not using
128bit wide vectors for the shuffle). This decreases instruction count of
the blend code generated for a rgba8 render target without blending from
169 to 113 with llvm 3.1 and from 136 to 114 in llvm 3.2/3.3, and I got
a ~8% (llvm 3.1) and ~5% (3.2/3.3) performance improvement in gears.
(The generated code is still not terribly good as we could actually avoid
the interleaving completely but llvm can't know this.)
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
This change came from the discovery that the STATIC_ASSERT to check that
the number of register file strings didn't actually work.
Similar changes could be made for the other string arrays in tgsi_string.c
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
The main change is to use MCJIT rather than the old JIT, which will never
be supported for System z. The endianness part is by example since the
patch was tested on a glibc system.
Signed-off-by: Richard Sandiford <rsandifo@linux.vnet.ibm.com>
Signed-off-by: Brian Paul <brianp@vmware.com>
There's no good reason why it can't handle 2x4f->1x8ub, 1x4f->1x4ub and
1x8f->1x8ub cases, there might be legitimate reasons why we don't have
enough input vectors for a full destination vector, and using pack
intrinsics should still be much better than using generic conversion
(it looks like convert_alpha from the blend code might hit this though
I suspect it could be avoided).
v2: add another test vector format to lp_test_conv so this gets tested.
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
The code was designed to handle no-op concat but failed (unless the
caller was using same pointer for src and dst).
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Surprising this bug survived so long, we were missing a clamp (in the
linear filtering version).
(Valgrind complained a lot about invalid reads with piglit texwrap,
I've also seen spurios failures in this test which might have
happened due to this. Valgrind probably didn't complain before the
alignment reduction in llvmpipe to 4x4 since the test is using tiny
textures so the reads were still always well within allocated area.)
While here, also do an effective clamp (after half subtraction)
of [0,length-0.5] instead of [0, length-1] which saves an instruction
(the filtering weight could be different due to this, but only if
both texels point to the same max texel so it doesn't matter).
(Both changes are borrowed from PIPE_TEX_CLAMP_TO_EDGE case.)
Note: This is a candidate for the stable branches.
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
We need to split up the depth and stencil values in this case, and there's
some new logic required to handle float depth and stencil simultaneously.
Also make sure we get the 64bit zs clear values and masks propagated
correctly.
Since we can only sample either depth or stencil but not both only load
the required bits which makes things a bit easier (it requires special
handling since the format doesn't fit into 32bit).
The logic for deciding if depth or stencil should be sampled is a bit odd,
but seems to be what other drivers and statetrackers do: if it's a format with
both depth and stencil (or just with depth) then sample depth, for sampling
stencil a sampler view format with only stencil is required.
Also while here fix up stencil sampling for other formats as well, though
this isn't supported by mesa (ARB_stencil_texturing), and while blits would
use it they don't work neither since they'd also need stencil export.
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
This change was meant as a stepping stone to use PMADDUBSW SSSE3
instruction, but actually this refactoring by itself yields a 10%
speedup on texture intensive shaders (e.g, Google Earth's ocean water
w/o S3TC on a Ivy Bridge machine), while giving yielding exactly the
same results, whereas PMADDUBSW only gave an extra 5%, at the expense of
2bits of precision in the interpolation.
I belive that the speedup of this change comes from the reduced register
pressure (as 8.8 fixed point intermediates take twice the space of 8bit
unorm).
Also, not dealing with 8.8 simplifies lp_bld_sample_aos.c code
substantially -- it's no longer necessary to have code duplicated for
low and high register halfs.
Note about lp_build_sample_mipmap(): the path for num_quads > 1 is never
executed (as it is faster on AVX to split the 256bit wide texture
computation into two 128bit chunks, in order to leverage integer
opcodes). This path might be useful in the future, so in order to
verify this change did not break that path I had to apply this change:
@@ -1662,11 +1662,11 @@ lp_build_sample_soa(struct gallivm_state *gallivm,
/*
* we only try 8-wide sampling with soa as it appears to
* be a loss with aos with AVX (but it should work).
* (It should be faster if we'd support avx2)
*/
- if (num_quads == 1 || !use_aos) {
+ if (/* num_quads == 1 || ! */ use_aos) {
if (num_quads > 1) {
if (mip_filter == PIPE_TEX_MIPFILTER_NONE) {
LLVMValueRef index0 = lp_build_const_int32(gallivm, 0);
/*
and then run texfilt mesademo:
LP_NATIVE_VECTOR_WIDTH=256 ./texfilt
Ran whole piglit without regressions.
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
The support is analogous to the way we handle indirect addressing
in temporaries, except that we don't have to worry about storing
(after declarations) and thus we'll able to keep using the old
code when indirect addressing isn't used. In other words we're
still using constants directly, unless the instruction has
immediate register with indirect addressing.
Signed-off-by: Zack Rusin <zackr@vmware.com>
Reviewed-by: José Fonseca <jfonseca@vmware.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
It helps a bit with vertex shader performance on i915g
(a couple percent faster with openarena).
I have tried most other passes, and they weren't showing
any measurable improvement. Note that my vertex shaders
didn't have loops, so maybe the loop optimizations could
still be useful in the future.
Reviewed-by: Brian Paul <brianp@vmware.com>
It should be TGSI_TYPE_UNSIGNED, not TGSI_TYPE_FLOAT.
Fixed also gallivm not_emit_cpu() to use uint build context.
Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
Acked-by: Roland Scheidegger <sroland@vmware.com>
This does not solve all of the problems with using LLVM in a
multithreaded enivronment, but it should help in some cases.
Reviewed-by: Mathias.Froehlich@web.de
It's valid because we reuse certain arithmetic operations
for both signed and unsigned types (e.g. uadd, umad, which
have a bit unfortunate naming)
Signed-off-by: Zack Rusin <zackr@vmware.com>
Reviewed-by: José Fonseca <jfonseca@vmware.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
we weren't adding the soa offsets when constructing the indices
for the gather functions. That meant that we were always returning
the data in the first vertex/primitive/pixel in the SoA structure
and not correctly fetching from all structures.
Signed-off-by: Zack Rusin <zackr@vmware.com>
Reviewed-by: José Fonseca <jfonseca@vmware.com>
Courtesy of clang:
src/gallium/auxiliary/gallivm/lp_bld_sample.c:1483:10: warning: array index of '2' indexes past the end of an array (that contains 2 elements) [-Warray-bounds]
tmp[2] = lp_build_swizzle_aos(coord_bld, ddx_ddy[1], swizzle02);
^ ~
src/gallium/auxiliary/gallivm/lp_bld_sample.c:1430:10: note: array 'tmp' declared here
LLVMValueRef ddx_ddy[2], tmp[2], rho_vec;
^
src/gallium/auxiliary/gallivm/lp_bld_sample.c:1487:56: warning: array index of '2' indexes past the end of an array (that contains 2 elements) [-Warray-bounds]
rho_vec = lp_build_add(coord_bld, rho_vec, tmp[2]);
^ ~
src/gallium/auxiliary/gallivm/lp_bld_sample.c:1430:10: note: array 'tmp' declared here
LLVMValueRef ddx_ddy[2], tmp[2], rho_vec;
^
src/gallium/auxiliary/gallivm/lp_bld_sample.c:1491:56: warning: array index of '2' indexes past the end of an array (that contains 2 elements) [-Warray-bounds]
rho_vec = lp_build_max(coord_bld, rho_vec, tmp[2]);
^ ~
src/gallium/auxiliary/gallivm/lp_bld_sample.c:1430:10: note: array 'tmp' declared here
LLVMValueRef ddx_ddy[2], tmp[2], rho_vec;
^
TEMP is not the only register file that accept unsigned. OUT too.
Actually, what determines the appropriate type of the destination value is
not the opcode, but rather the register.
Also cleanup/simplify code. Add a few more asserts, but also make
code more robust by handling graceful if assert fails.
This fixes segfault / assertion in the included vert-uadd.sh graw shader.
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
2.7 was a particularly trouble ridden release.
Furthermore, the bug no longer can be reproduced ever since the
first_level state was taken in account.
Reviewed-by: Brian Paul <brianp@vmware.com>
They are supported on LLVM 3.1, at least on x86. (I haven't tested on PPC
though.)
Actually lp_build_linear_mip_levels() already has been emitting them for
some time.
This avoids intrinsics, which tend to be an obstacle for certain
optimization passes.
Reviewed-by: Brian Paul <brianp@vmware.com>
Should be able to handle all things which make this tricky to implement.
Fallthroughs, including most notably into/out of default, should be handled
correctly but are quite a mess.
If we see largely unoptimized switches in the wild should probably think
about some "real" switch optimization pass, e.g. things like this:
switch
case1
someinst
brk
case2
default
case3
someinst
brk
case4
someinst
endswitch
are legal, but the pointless case2/case3 statements not only cause condition
evaluation but will turn this into a "fake" fallthrough case (because
mask and defaultmask are already updated for case2 when default is
encountered) requiring executing code twice.
If default is at the end though, there's never any code re-execution, and
if that's not the case if there's no fallthrough in (not even a fake one)
and out of default there's no code re-execution neither.
v2: add comments, and use enum for break type instead of magic boolean.
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
It seems there was a typo in gallivm breakc handling (I am actually still
not sure it is really needed but otherwise that statement really should go
away). Also fix the wrong src argument type, even though they weren't really
used.
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
This is still not really correct, since at least for sm 4.0
the nesting limit is 64 per subroutine, and subroutine nesting itself
has a limit of 32, so since we have a flat stack we'd need 32*64.
But this should probably be better fixed with per-subroutine stacks,
since otherwise these structures get really big (like 100kB for the
lp_exec_mask).
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Turns out the previous "fix" for handling per-pixel face selection and
derivatives didn't work out that well - the derivatives were wrong by
quite a bit, in theory transformation of the derivatives into cube space
should work, but would be _a lot_ more work than the "simplified" transform
used.
So, for explicit derivatives, I'm just giving up and go back to not honoring
them.
For implicit derivatives (and the fake explicit ones) however we try
something a little different, we just calculate rho as we would for a 3d
texture, that is after scaling the coords by the inverse major axis.
This gives the same results as calculating the derivs after projection of
the coords to the same face as long as all pixels hit the same face (and
only without rho_no_opt, otherwise it should be a bit worse). And when
not all pixels are hitting the same face, the results aren't so hot but
not catastrophically bad (I believe not off by more than a factor of 2 without
no_rho_approx and not more than sqrt(2) with no_rho_approx). I think this is
better than just picking the wrong face but who knows...
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
This will calculate rho correctly as
sqrt(max((ds/dx)^2 + (dt/dx)^2 + (dr/dx)^2), (ds/dx)^2 + (dt/dx)^2 + (dr/dx)^2))
instead of max(|ds/dx|,|dt/dx|,|dr/dx|,|ds/dy|,|dt/dy,|dr/dy|)
(for 3 coords - 2 coords work analogous, for 1 coord there's no point doing
the exact version), for both implicit and explicit derivatives.
While such approximation seems to be allowed in OpenGL some APIs may be less
forgiving, and the error can be quite large (sqrt(2) for 2 coords, sqrt(3) for
3 coords so wrong by nearly one mip level in the latter case).
This also helps to single out "real" bugs from "expected" ones, so it is debug
only (though at least combined with no_brilinear I didn't really see much of a
performance difference but only tested with a debug build - at least with
implicit mipmaps the instruction count is almost exactly the same though the
instructions are more complex (1 sqrt and mul/adds instead of and/max mostly).
The code when the option isn't set stays exactly the same.
v2: rename no_rho_opt to no_rho_approx.
Reviewed-by: Brian Paul <brianp@vmware.com>
We were always treating the vertex index as a scalar but when the
shader is using indirect addressing it will be a vector of indices
for each channel. This was causing some nasty crashes insides
LLVM.
Signed-off-by: Zack Rusin <zackr@vmware.com>
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
The specification says that the geometry shader should exit if the
number of emitted vertices is bigger or equal to max_output_vertices and
we can't do that because we're running in the SoA mode, which means that
our storing routines will keep getting called on channels that have
overflown (even though they will be masked out, but we just can't skip
them).
So we need some scratch area where we can keep writing the overflown
vertices without overwriting anything important or crashing.
Signed-off-by: Zack Rusin <zackr@vmware.com>
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
The issue with SOA execution and end_primitive opcode is that it
can be executed both when we haven't emitted any vertices, in
which case we don't want to emit an empty primitive, and when
the execution mask is zero and the execution should be skipped. We
handled only the latter of those conditions. Now we're combining the
execution mask with a mask created from emitted vertices to handle
both cases. As a result we don't need the pending_end_primitive
flag which was broken because it was static and could be affected
by both above mentioned conditions at run-time.
Signed-off-by: Zack Rusin <zackr@vmware.com>
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
TGSI_OPCODE_IF condition had two possible interpretations:
- src.x != 0.0f
- Mesa statetracker when PIPE_SHADER_CAP_INTEGERS was false either for
vertex and fragment shaders
- gallivm/llvmpipe
- postprocess
- vl state tracker
- vega state tracker
- most old drivers
- old internal state trackers
- many graw examples
- src.x != 0U
- Mesa statetracker when PIPE_SHADER_CAP_INTEGERS was true for both
vertex and fragment shaders
- tgsi_exec/softpipe
- r600
- radeonsi
- nv50
And drivers that use draw module also were a mess (because Mesa would
emit float IFs, but draw module supports native integers so it would
interpret IF arg as integers...)
This sort of works if the source argument is limited to float +0.0f or
+1.0f, integer 0, but would fail if source is float -0.0f, or integer in
the float NaN range. It could also fail if source is integer 1, and
hardware flushes denormalized numbers to zero.
But with this change there are now two opcodes, IF and UIF, with clear
meaning.
Drivers that do not support native integers do not need to worry about
UIF. However, for backwards compatibility with old state trackers and
examples, it is advisable that native integer capable drivers also
support the float IF opcode.
I tried to implement this for r600 and radeonsi based on the surrounding
code. I couldn't do this for nouveau, so I just shunted IF/UIF
together, which matches the current behavior.
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Reviewed-by: Marek Olšák <maraeo@gmail.com>
v2:
- Incorporate Roland's feedback.
- Fix r600_shader.c merge conflict.
- Fix typo in radeon, spotted by Michel Dänzer.
- Incorporte Christoph Bumiller's patch to handle TGSI_OPCODE_IF(float)
properly in nv50/ir.
Inserting the value for the second quad in the wrong place for the
following shuffle. This meant the row or image stride was undefined which is
quite catastrophic, can lead to bogus texels fetched or just segfault.
This code is only hit for SoA path currently, still surprising it
didn't crash more or caused more visible issues (I think llvm used a
broadcast shuffle for the undefined parts of the vector, hence the undefined
value for the second quad was just the same as that from the first quad,
so as long as both quads hit the same mip level everything was fine, and since
lower mips always have the same large stride it made it less likely to
hit out-of-bound memory in case of differing lods).
Note: this is a candidate for stable branches.
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
both mov and ucmp can be used to move variables of any type.
correctly note that about ucmp in the tgsi_info and make
sure gallivm can handle that by correctly casting the untyped
moves.
Signed-off-by: Zack Rusin <zackr@vmware.com>
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
We were using simple temporaries, without using alloca or phi
nodes which meant that on every iteration of the loop our
temporaries, which were holding the number of vertices and
primitives which were emitted, were being reset to zero. Now
we're using alloca to allocate those variables to preserve
them across conditionals.
Signed-off-by: Zack Rusin <zackr@vmware.com>
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
We want to both make sure we never divide by zero to not generate
sigfpe and that divide by zero is guaranteed to return 0xffffffff.
Based on José idea.
Signed-off-by: Zack Rusin <zackr@vmware.com>
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
we break when the mask values are 0 not, 1, plus it's bit comparison
not a floating point comparison. This fixes both.
Signed-off-by: Zack Rusin <zackr@vmware.com>
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
The ar_ge_as_at variable was just very very confusing since the condition
was actually the other way around (as_at_ge_ar). So change the condition
(and the selects depending on it) to match the variable name.
And also change the chosen major axis in case the coord values are the
same. OpenGL doesn't care one bit which one is chosen in this case but
it looks like dx10 would require z chosen over y, and y chosen over x
(previously did x chosen over y, y chosen over z). Since it's all the
same effort just honor dx10's wishes. (Though actually, for some prefered
orderings, we could save one (or two with derivatives) selects since the
tnewx and tnewz (and the corresponding dmax values) are the same.)
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
This is trivial now, though need to make sure we pass all the necessary
derivative values (which is 3 each for ddx/ddy not 2).
Passes piglit arb_shader_texture_lod-texgradcube test.
v2: add the forgotten abs() for all incoming derivatives (discovered
by new piglit arb_shader_texture_lod-texgradcube test, though more by
luck as it was failing only for exactly one pixel...).
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
This proved to be tricky, the problem is that after selection/mirroring
we cannot calculate reasonable derivatives (if not all pixels in a quad
end up on the same face the derivatives could get "randomly" exceedingly
large).
However, it is actually quite easy to simply calculate the derivatives
before selection/mirroring and then transform them similar to
the cube coordinates (they only need selection/projection, but not
mirroring as we're not interested in the sign bit, of course). While
there is a tiny bit more work to do (need to calculate derivs for 3
coords instead of 2, and additional selects) it also simplifies things
somewhat for the coord selection itself (as we save some broadcast aos
shuffles, and we don't need to calculate the average vector) - hence if
derivatives aren't needed this should actually be faster.
Also, this has the benefit that this will (trivially) work for explicit
derivatives too, which we completely ignored before that (will be in a
separate commit for better trackability).
Note that while the way for getting rho looks very different, it should
result in "nearly" the same values as before (the "nearly" is only because
before the code would choose the face based on an "average" vector and hence
the derivatives calculated according to this face, where now (for implicit
derivatives) the derivatives are projected on the face selected for the
first (top-left) pixel in a quad, so not necessarly the same face).
The transformation done might not quite be state-of-the-art, calculating
length(dx,dy) as max(dx,dy) certainly isn't neither but this stays the
same as before (that is I think a better transform would _somehow_ take
the "derivative major axis" into account so that derivative changes in
the major axis wouldn't get ignored).
Should solve some accuracy problems with cubemaps (can easily be seen with
the cubemap demo when switching wrapping/filtering), though we still don't
do seamless filtering to fix it completely (so not per-sample but per-pixel
is certainly better than per-quad and already sufficient for accurate
results with nearest tex filter).
As for performance, it seems to be a tiny bit faster too (maybe 3% or so
with cubemap demo). Which I'd have expected with nearest/nearest filtering
where this will be less instructions, but the difference seems to actually
be larger with linear/linear_mipmap_linear where it is slightly more
instructions, probably the code appears less serialized allowing better
scheduling (on a sandy bridge cpu). It actually seems to be now at least
as fast as the old path using a conditional when using 128bit vectors too
(that is probably more a result of testing with a newer cpu though), for now
that old path is still there but unused.
No piglit regressions.
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Using a different packing for the single coord case should save a shuffle.
Plus some minor style fixes.
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Should be way faster of course on cpus supporting this (includes AMD
Bulldozer and Jaguar cores, Intel Ivy Bridge and up (except budget models)).
Passes piglit fbo-blending-formats GL_ARB_texture_float -auto on Ivy Bridge.
Reviewed-by: Brian Paul <brianp@vmware.com>
Instead of void pointers use a base interface.
Signed-off-by: Zack Rusin <zackr@vmware.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: José Fonseca <jfonseca@vmware.com>
Conceptually the same as previously done in float_to_half.
Should cut down number of instructions from 14 to 10 or so, but
will promote some NaNs to Infs, so it's disabled.
It gets a bit tricky though handling all the cases correctly...
Passes basic tests either way (though there are no tests testing special
cases, but some manual tests injecting them seemed promising).
v2: style and comment fixes suggested by Jose
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
This replaces the existing float-to-half implementation.
There are definitely a couple of differences - the old implementation
had unspecified(?) rounding behavior, and could at least in theory
construct Inf values out of NaNs. NaNs and Infs should now always be
properly propagated, and rounding behavior is now towards zero
(note this means too large but non-Infinity values get propagated to max
representable value, not Infinity).
The implementation will definitely not match util code, however (which
does nearest rounding, which also means too large values will get
propagated to Infinity).
Also fix a bogus round mask probably leading to rounding bugs...
v2: fix a logic bug in handling infs/nans.
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Similar enough that we can try to use shared code.
v2: fix a stupid bug using wrong variable causing mayhem with Inf and NaNs.
Reviewed-by: Jose Fonseca <jfonseca@vmware.com
Required by more modern examples. Like BRK but with a condition.
Signed-off-by: Zack Rusin <zackr@vmware.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: José Fonseca <jfonseca@vmware.com>
TGSI semantics currently require an implicit endprim at the end
of GS if an ending primitive hasn't been emitted.
Signed-off-by: Zack Rusin <zackr@vmware.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: José Fonseca <jfonseca@vmware.com>
This commits implements code generation of the geometry shaders in
the SOA paths. All the code is there but bugs are likely present.
Signed-off-by: Zack Rusin <zackr@vmware.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: José Fonseca <jfonseca@vmware.com>
And use this (and the code for r11g11b10 packed float to float conversion)
in the soa texturing code (the generated code looks quite good).
Should be an order of magnitude faster probably than using the fallback
(not measured).
Tested with piglit texwrap GL_EXT_packed_float and
GL_EXT_texture_shared_exponent respectively (didn't find much else using
it).
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
New conversion code to handle conversion from/to r11g11b10 AoS to/from
SoA floats, and also add code for conversion from rgb9e5 AoS to float SoA
(which works pretty much the same as r11g11b10 except for the packing).
(This code should also be used for texture sampling instead of
relying on u_format conversion but it's not yet, so rgb9e5 is unused.)
Unfortunately a crazy amount of hacks is necessary to get the conversion
code running in llvmpipe's generate_unswizzled_blend, which isn't well
suited for formats where the storage representation has nothing to do
with what's needed for blending (moreover, the conversion will convert
from packed AoS values, which is the storage format, to float SoA values,
because this is much more natural for the conversion, and likewise from
SoA values to packed AoS values - but the "blend" (which includes
trivial things like partial mask) works on AoS values, so incoming fs
values will go SoA->AoS, values from destination will go packed
AoS->SoA->AoS, then do blend, then AoS->SoA->packed AoS which probably
isn't the most efficient way though the shuffles are probably bearable).
Passes piglit fbo-blending-formats (with GL_EXT_packed_float parameter),
still need to verify Inf/NaNs (where most of the complexity in the
conversion comes from actually).
v2: drop the (very bogus) rgb9e5 part, and do component extraction
in the helper code for r11g11b10 to float conversion, making the code
slightly more compact (suggested by Jose), now that there are no other
callers left this works quite well. (Could do the same for the
opposite way but it's less than ideal there, final part of packing
needs to be done in caller anyway and there'd be another conditional.)
v3: minor style and comment fixes. Also fix a potential issue with
negative zero being potentially returned by max(src, zero) as we
don't have well-defined min/max behavior (fortunately no additonal cost).
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
If we're in some conditional or loop we must not return, or the code
after the condition is never executed.
(v2): And, we also can't just continue as nothing happened, since the
mask update code would later check if we actually have a mask, so we
need to remember that there was a return in main where we didn't exit
(to illustrate this, a ret in a if clause would cause a mask update
which is still ok as we're in a conditional, but after the endif the
mask update code would drop the mask hence bringing execution back to
pixels which should have their execution mask set to zero by the ret).
Thanks to Christoph Bumiller for figuring this out.
This fixes https://bugs.freedesktop.org/show_bug.cgi?id=62357.
Note: This is a candidate for the stable branches.
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
To further improve the optimization of source and destination
indirect addressing we need the ability to store a reference
to the declaration of the addressed operands.
Since most of the fields in tgsi_src_register doesn't apply for
an indirect addressing operand replace it with a separate
tgsi_ind_register structure and so make room for extra information.
v2: rename Declaration to ArrayID, put the ArrayID into () instead of []
Signed-off-by: Christian König <christian.koenig@amd.com>
Previously, the derivatives were calculated and passed in a packed form
to the sample code (for implicit derivatives, explicit derivatives were
packed to the same format).
There's several reasons why this wasn't such a good idea:
1) the derivatives may not even be needed (not as bad as it sounds since
llvm will just throw the calculations needed for them away but still)
2) the special packing format really shouldn't be part of the sampler
interface
3) depending what the sample code actually does the derivatives will
be processed differently, hence there is no "ideal" packing. For cube
maps with explicit derivatives (which we don't do yet) for instance the
packing looked downright useless, and for non-isotropic filtering we'd
need different calculations too.
So, instead just pass the derivatives as is (for explicit derivatives),
or let the rho calculating sample code calculate them itself. This still
does exactly the same packing stuff for implicit derivatives for now,
though explicit ones are handled in a more straightforward manner (quick
estimates show performance should be quite similar, though it is much
easier to follow and also does the rho calculation per-pixel until the
end, which we eventually need for spec compliance anyway).
No piglit changes.
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
This was previously only handled for texelFetch (much easier).
Depending on the wrap mode this works slightly differently (for somewhat
efficient implementation), hence have to do that separately in all roughly
137 places - it is easy if we use fixed point coords for wrapping, however
some wrapping modes are near impossible with fixed point (the repeat stuff)
hence we have to normalize the offsets if we can't do the wrapping in
unnormalized space (which is a division which is slow but should still be
much better than the alternative, which would be integer modulo for wrapping
which is just unusable). This should still give accurate results in all
cases that really matter, though it might be not quite conformant behavior
for some apis (but we have much worse problems there anyway even without
using offsets).
(Untested, no piglit test.)
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Since with llvm execution parts of sampler view and sampler state is baked into
the shader, we need to revalidate otherwise the wrong shader might get used.
(Not completely sure but I think this would not be required for non-llvm case,
along with everything else in these functions.)
This caused bugs in piglit arb_texture_buffer_object-formats, because we never
noticed that the view format changed.
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
This also fixes not honoring first/last_layer view parameters for array
textures, plus not honoring last_level view parameter for all textures
(neither is really used by OpenGL).
This mostly passes piglit arb_texture_buffer_object tests (it needs, however,
glsl 140 version override, plus GL 3.1 override, the latter only because
mesa does not allow ARB_tbo in non-core contexts).
Most arb_texture_buffer_object tests pass, with the exception of
arb_texture_buffer_object-formats. With "arb" parameter it passes most weirdo
formats before it segfaults in the state tracker, this looks to be some issue
with using legacy formats in core context (fails the same in softpipe).
With "core" parameter it passes with "fs", however fails with "vs" (for most
formats). This will be fixed later (debugging shows we're completely missing
the shader recompile depending on format).
v2: based on Jose's feedback, fix comments, variable/function names.
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
For constant and temporary register fetches, the bitcasts weren't done
correctly for the indirect case, leading to crashes due to type mismatches.
Simply do the bitcasts after fetching (much simpler than fixing up the load
pointer for the various cases).
This fixes https://bugs.freedesktop.org/show_bug.cgi?id=61036
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
We need to encode them as Texture instructions since the NumOffsets field
is encoded there. However, we don't encode the actual target in there, this
is derived from the sampler view src later.
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Need to take the type into account. Also, if we want to allow
mov's with modifiers we need to pick a type (assume float).
v2: don't allow all modifiers on all type, in particular don't allow
absolute on non-float types and don't allow negate on unsigned.
Also treat UADD as signed (despite the name) since it is used
for handling both signed and unsigned integer arguments and otherwise
modifiers don't work.
Also add tgsi docs clarifying this.
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
The emulation of these if there's no rounding instruction available
is a bit more complicated than what the code did.
In particular, doing fp-to-int/int-to-fp will not work if the exponent
is large enough (and with NaNs, Infs). Hence such values need to be filtered
out and the original value returned in this case (which fortunately should
always be exact). This comes at the expense of performance (if your cpu
doesn't support rounding instructions).
Furthermore, floor/ifloor/ceil/iceil were affected by precision issues for
values near negative (for floor) or positive (for ceil) zero, fix that as well
(fixing this issue might not actually be slower except for ceil/iceil if the
type is not signed which is probably rare - note iceil has no callers left
in any case).
Also add some new rounding test values in lp_test_arit to actually test
for that stuff (which previously would have failed without sse41).
This fixes https://bugs.freedesktop.org/show_bug.cgi?id=59701.
It looks like using coord.w as explicit lod value is a mistake, most likely
because some dx10 docs had it specified that way. Seems this was changed though:
http://msdn.microsoft.com/en-us/library/windows/desktop/hh447229%28v=vs.85%29.aspx
- let's just hope it doesn't depend on runtime build version or something.
Not only would this need translation (so go against the stated goal these
opcodes should be close to dx10 semantics) but it would prevent usage of this
opcode with cube arrays, which is apparently possible:
http://msdn.microsoft.com/en-us/library/windows/desktop/bb509699%28v=vs.85%29.aspx
(Note not only does this show cube arrays using explicit lod, but also the
confusion with this opcode: it lists an explicit lod parameter value, but then
states last component of location is used as lod).
(For "true" hw drivers, only nv50 had code to handle it, and it appears the
code was already right for the new semantics, though fix up the seemingly
wrong c/d arguments while there.)
v2: fix comment, separate out other changes.
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Need to calculate the number of mip levels (if it would be worthwile could
store it in dynamic state).
While here, the query code also used chan 2 for the lod value.
This worked with mesa state tracker but it seems safer to use chan 0.
Still passes piglit textureSize (with some handwaving), though the non-GL
parts are (largely) untested.
v2: clarify and expect the sviewinfo opcode to return ints, not floats,
just like the OpenGL textureSize (dx10 supports dst modifiers with resinfo).
Also simplify some code.
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
They are similar to old-style tex opcodes but with separate sampler and
texture units (and other arguments in different places).
Also adjust the debug tgsi dump code.
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
The signed case didn't do what the comment indicated. Should increase rounding
precision (at the expense of performance since the former code was effectively
a no-op).
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Make it obvious what "unit" this is (no change in functionality).
draw still uses "unit" in places where it changes the shader by adding
texture sampling itself - it seems like this can't work with shaders
using dx10-style sample opcodes (can't mix gl-style and dx10-style
sample instructions in a shader).
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Split the sampler interface to use separate sampler and texture (sampler_view)
state. This is needed to support dx10-style sampling instructions.
This is not quite complete since both draw/llvmpipe don't really track
textures/samplers independently yet, as well as the gallivm code not quite
using the right sampler or texture index respectively (but it should work
for the sampling codes used by opengl).
We are however losing some optimizations in the process, apply_max_lod will
no longer work, and we potentially could end up with more (unnecessary)
recompiles (if switching textures with/without mipmaps only so it shouldn't
be too bad).
v2: don't use different callback structs for sampler/sampler view functions
(which just complicates things), fix up sampling code to actually use the
right texture or sampler index, and similar for llvmpipe/draw actually
distinguish between samplers and sampler views.
v3: fix more of PIPE_MAX_SAMPLER / PIPE_MAX_SHADER_SAMPLER_VIEWS mismatches
(both in draw and llvmpipe), based on feedback from José get rid of unneeded
static sampler derived state.(which also fixes the only 2 piglit regressions
due to a forgotten assignment), fix comments based on Brian's feedback.
v4: remove some accidental unrelated whitespace changes
Reviewed-by: José Fonseca <jfonseca@vmware.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Need to bitcast the float border color (luckily we already get
the color as int just disguised as float).
Fixes piglit texwrap GL_EXT_texture_integer bordercolor.
Reviewed-by: José Fonseca <jfonseca@vmware.com>
Change the texel type to int/uint instead of float throughout the sampling
code which makes it easier to catch errors (as llvm will complain about wrong
types if we mistakenly treat these values as real floats somewhere).
This should also get things like e.g. sampler swizzles (for unused channels)
right.
This fixes piglit texture_integer_glsl130 test.
Border color not working (crashing) yet.
(These formats are not exposed yet in llvmpipe.)
v2: couple cleanups according to José's comments
Reviewed-by: José Fonseca <jfonseca@vmware.com>
Since the idea is to just expand or shrink the bit width but not otherwise do
conversion we also need to adjust the sign bit according to src, otherwise
the conversion code will incorrectly clamp the values. (Since this only works
for casting to ordinary floats the norm and fixed bits should always be fine.)
This fixes the remaining piglit attribs GL3 failures.
Reviewed-by: José Fonseca <jfonseca@vmware.com>
a460aea3f1 wasn't entirely correct,
since all coords are already ints hence need to skip the iround.
Passes piglit texelFetch with sampler1DArray/sampler2DArray.
Reviewed-by: Dave Airlie <airlied@redhat.com>
Since we don't call lp_build_sample_common() in the texel fetch path we missed
the layer fixup code. If someone would have tried to do texelFetch with array
textures it would have crashed for sure.
Not really tested (can't run the piglit test being able to use texelFetch with
array samplers for now with llvmpipe).
Reviewed-by: José Fonseca <jfonseca@vmware.com>
This fixes a number of crashes on r600g due to the fact that
lp_build_mul assumes vector types when optimizing mul to bit shifts.
This bug was uncovered by 0ad1fefd69
Noticed would fail, we were doing two things wrong
a) 1d arrays require the layers in height
b) minifying the layers field.
v2: don't change height code, fixup completely inside txq
as suggested by Roland.
v3: just add minify before texture array size
v1: Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Just enough for draw module to work ok.
This improves "piglit attribs GL3", though something fishy is still
happening with certain unsigned integer values.
Reviewed-by: Brian Paul <brianp@vmware.com>
The ADDR file is cumbersome for native integer capable drivers. We
should consider deprecating it eventually, but this just adds support
for indirection from TEMP registers.
Reviewed-by: Brian Paul <brianp@vmware.com>
Support 16 (defined in LP_MAX_TGSI_CONST_BUFFERS) as opposed to 32 (as
defined by PIPE_MAX_CONSTANT_BUFFERS) because that would make the jit
context become unnecessarily large.
v2: Bump limit from 4 to 16 to cover ARB_uniform_buffer_object needs,
per Dave Airlie.
Reviewed-by: Brian Paul <brianp@vmware.com>
Several issues actually:
- Fix a regression in unsigned normalized in the rescaling
[0, 255] to [0, 256]
- Ensure we use signed shifts where appropriate (instead of
unsigned shifts)
- Refactor the code slightly -- move all the logic inside
lp_build_lerp_simple().
This change, plus an adjustment in the tolerance of signed normalized
results in piglit fbo-blending-formats fixes bug 57903
Reviewed-by: Brian Paul <brianp@vmware.com>
They need to be converted to the native integer type to prevent garbage
in higher order bits from being printed.
Reviewed-by: Brian Paul <brianp@vmware.com>
This fixes fdo bug 57755 and most of the failures of piglit fbo-blending-formats
GL_EXT_texture_snorm.
GL_INTENSITY_SNORM is still failing, but problem is probably elsewhere,
as GL_R8_SNORM works fine.
we need to rely on util code for fetching those, just like before
9f06061d50.
Fixes bugs 57699 and 57756.
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: José Fonseca <jfonseca@vmware.com>
The border clamping code is unnecessary, since we don't care if a wrapped
coord value is -1 or <-1 (same for length vs. >length), in either case the
border handling code will mask out the offset and replace the texel value with
the border color.
Note that technically this is not entirely correct. Omitting clamping on the
float coords means that flt->int conversion may result in undefined values for
values of very large magnitude.
However there's no reason we should honor this here since:
a) we don't care for that for ordinary wrap modes in the aos code when
converting coords and the problem is worse there (as we've got only
effectively 24 instead of 32bits)
b) at least in some cases the clamping was done already in int space hence
doing nothing to fix that problem.
c) with sse2 flt->int conversion with such values results in 0x80000000 which
is just perfect (for clamp to border - not so much for the ordinary clamp to
edge).
Reviewed-by: Brian Paul <brianp@vmware.com>
The current implementation was close by not fully correct: several
operations that should be done in floating point were being done in
integer.
Fixes piglit fbo-clear-formats GL_ARB_texture_float
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
untested (couldn't get the piglit test to run even with version overrides)
but seemed blatantly wrong.
In any case it would only affect an error case which when it would happen
probably all hope is lost anyway.
Reviewed-by: José Fonseca <jfonseca@vmware.com>
Support 1d and 2d array textures (including shadow samplers),
and (as a side effect mostly) also shadow cube samplers.
Seems to pass the relevant piglit tests both for sampling and rendering
to (though some require version overrides).
Since we don't support render target indices rendering to array textures
is still restricted to a single layer at a time.
Also, the min/max layer in the sampler view (which is unnecessary for GL)
is ignored (always use all layers).
Reviewed-by: José Fonseca <jfonseca@vmware.com>
It is buggy (it was giving wrong results for some of the formats with
padding), and util_format_description::is_array already does precisely
what's intended.
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
This patch fixes various format manipulation for big-endian
architectures.
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
This patch adds two more functions in type conversions header:
* lp_build_bswap: construct a call to llvm.bswap intrinsic for an
element
* lp_build_bswap_vec: byte swap every element in a vector base on the
input and output types.
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
This patch fixes the vector constant generation used for vector shuffle
for big-endian machines.
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
This patch enforces the clear of NJ bit in VSCR Altivec register so
denormal numbers are handles as expected by IEEE standards.
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
This patch adds Altivec intrinsics for float vector types. It changes
the SSE specific definitions to a platform neutral and adds the calls
to Altivec intrinsic builder.
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
This patch add correct vector addition and substraction intrisics when
using Altivec with PPC. Current code uses default path and LLVM backend
ends up issuing carry-out arithmetic instruction while it is expected
saturated ones.
It also includes a fix for PowerPC where char are unsigned by default,
resulting in bogus values for vector shifting.
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
This patch adds PPC Altivec support for pack/unpack operations using Altivec
supported vector type (8xi8, 16xi16, 4xi32, 4xf32).
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
No longer have to split fetching into quads dynamically if mip levels
are not the same for all quads (aos sampling still always splits due
to performance reasons).
Instead handle multiple mip levels further down, minification etc. takes
this into account.
Reviewed-by: José Fonseca <jfonseca@vmware.com>
This also adds some code to handle per-quad lods for more than 4-wide fetches,
because otherwise I'd have to integrate the texelFetch function into
the splitting stuff... (but it is not used yet outside texelFetch).
passes piglit fs-texelFetch-2D, fails fs-texelFetchOffset-2D due to I believe
a test error (results are undefined for out-of-bounds fetches, we return
whatever is at offset 0, whereas the test expects [0,0,0,1]).
Texel offsets are only handled by texelFetch for now, though the interface
can handle it for everything.
Reviewed-by: José Fonseca <jfonseca@vmware.com>
This might have a slight overhead but handling mip offsets more like
the width (and image) strides should make some things easier (mip level
being just part of the offset calculation) later.
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: José Fonseca <jfonseca@vmware.com>
LLVM 3.1+ haven't more "extern unsigned llvm::StackAlignmentOverride"
and friends for configuring code generation options, like stack
alignment.
So I restrict assiging of lvm::StackAlignmentOverride and other
variables to LLVM 3.0 only, and wrote similiar code using
TargetOptions.
This patch fix segfaulting of WINE using llvmpipe built with LLVM 3.1
Signed-off-by: Alexander V. Nikolaev <avn@daemon.hole.ru>
Signed-off-by: José Fonseca <jose.r.fonseca@gmail.com>
Per commentary and direction in the LLVM community, support for ppc64 is
going into MCJIT rather than the old JIT. There is no existing support
in prior llvm versions, so no need to specify LLVM version numbers.
Signed-off-by: Will Schmidt <will_schmidt@vnet.ibm.com>
Signed-off-by: José Fonseca <jfonseca@vmware.com>
lp_build_rsqrt initially did not do any newton-raphson step. This meant that
precision was only ~11 bits, but this handled both input 0.0 and +infinity
correctly. It did not however handle input 1.0 accurately, and denormals
always generated infinity result.
Doing a newton-raphson step increased precision significantly (but notably
input 1.0 still doesn't give output 1.0), however this fails for inputs
0.0 and infinity (both result in NaNs).
Try to fix this up by using cmp/select but since this is all quite fishy
(and still doesn't handle denormals) disable for now. Note that even with
workarounds it should still have been faster since the fallback uses sqrt/div
(which both use the usually unpipelined and slow divider hw).
Also add some more test values to lp_test_arit and test lp_build_rcp() too while
there.
v2: based on José's feedback, avoid hacky infinity definition which doesn't
work with msvc (unfortunately using INFINITY won't cut it neither on non-c99
compilers) in lp_build_rsqrt, and while here fix up the input infinity case
too (it's disabled anyway). Only test infinity input case if we have c99,
and use float cast for calculating reference rsqrt value so we really get
what we expect.
Reviewed-by: José Fonseca <jfonseca@vmware.com>
Fixes WebGL conformance/uniforms/uniform-default-values.html crash.
We need to check for the null view pointer before accessing view->texture.
Fixes http://bugs.freedesktop.org/show_bug.cgi?id=53317
Note: This is a candidate for the 8.0 branch.
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Fixes uninitialized scalar field defect reported by Coverity.
Signed-off-by: Vinson Lee <vlee@freedesktop.org>
Reviewed-by: José Fonseca <jfonseca@vmware.com>
Testing shows that the standard JIT engine retrofited with AVX support is quite
stable and as capable to handle AVX instructions as MC-JIT is.
And the old JIT is much more memory efficient, as we don't need to
allocate one engine instance per shader, as we do for MC-JIT due to its
incompleteness.
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Squashed commit of the following:
commit 7acb7b4f60dc505af3dd00dcff744f80315d5b0e
Author: José Fonseca <jfonseca@vmware.com>
Date: Mon Jul 9 17:46:31 2012 +0100
draw: Don't use dynamically sized arrays.
Not supported by MSVC.
commit 5810c28c83647612cb372d1e763fd9d7780df3cb
Author: José Fonseca <jfonseca@vmware.com>
Date: Mon Jul 9 17:44:16 2012 +0100
gallivm,llvmpipe: Don't use expressions with PIPE_ALIGN_VAR().
MSVC doesn't accept exceptions in _declspec(align(...)). Use a
define instead.
commit 8aafd1457ba572a02b289b3f3411e99a3c056072
Author: José Fonseca <jfonseca@vmware.com>
Date: Mon Jul 9 17:41:56 2012 +0100
gallium/util: Make u_cpu_detect.h header C++ safe.
commit 5795248350771f899cfbfc1a3a58f1835eb2671d
Author: José Fonseca <jfonseca@vmware.com>
Date: Mon Jul 2 12:08:01 2012 +0100
gallium/util: Add ULL suffix to large constants.
As suggested by Andy Furniss: it looks like some old gcc versions
require it.
commit 4c66c22727eff92226544c7d43c4eb94de359e10
Author: José Fonseca <jfonseca@vmware.com>
Date: Fri Jun 29 13:39:07 2012 +0100
gallium/util: Truly disable INF/NAN tests on MSVC.
Thanks to Brian for spotting this.
commit 8bce274c7fad578d7eb656d9a1413f5c0844c94e
Author: José Fonseca <jfonseca@vmware.com>
Date: Fri Jun 29 13:39:07 2012 +0100
gallium/util: Disable INF/NAN tests on MSVC.
Somehow they are not recognized as constants.
commit 6868649cff8d7fd2e2579c28d0b74ef6dd4f9716
Author: Roland Scheidegger <sroland@vmware.com>
Date: Thu Jul 5 15:05:24 2012 +0200
gallivm: Cleanup the 2 x 8 float -> 16 ub special path in lp_build_conv.
No behaviour change intended, like 7b98455fb40c2df84cfd3cdb1eb7650f67c8a751.
Reviewed-by: José Fonseca <jfonseca@vmware.com>
commit 5147a0949c4407e8bce9e41d9859314b4a9ccf77
Author: Roland Scheidegger <sroland@vmware.com>
Date: Thu Jul 5 14:28:19 2012 +0200
gallivm: (trivial) fix issues with multiple-of-4 texture fetch
Some formats can't handle non-multiple of 4 fetches I believe, but
everything must support length 1 and multiples of 4.
So avoid going to scalar fetch (which is very costly) just because length
isn't 4.
Also extend the hack to not use shift with variable count for yuv formats to
arbitrary length (larger than 1) - doesn't matter how many elements we
have we always want to avoid it unless we have variable shift count
instruction (which we should get with avx2).
Reviewed-by: José Fonseca <jfonseca@vmware.com>
commit 87ebcb1bd71fa4c739451ec8ca89a7f29b168c08
Author: Roland Scheidegger <sroland@vmware.com>
Date: Wed Jul 4 02:09:55 2012 +0200
gallivm: (trivial) fix typo for wrap repeat mode in linear filtering aos code
This would lead to bogus coordinates at the edges.
(undetected by piglit because this path is only taken for block-based
formats).
Signed-off-by: José Fonseca <jfonseca@vmware.com>
commit 3a42717101b1619874c8932a580c0b9e6896b557
Author: José Fonseca <jfonseca@vmware.com>
Date: Tue Jul 3 19:42:49 2012 +0100
gallivm: Fix TGSI integer translation with AVX.
commit d71ff104085c196b16426081098fb0bde128ce4f
Author: José Fonseca <jfonseca@vmware.com>
Date: Fri Jun 29 15:17:41 2012 +0100
llvmpipe: Fix LLVM JIT linear path.
It was not working properly because it was looking at the JIT function
before it was actually compiled.
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
commit a94df0386213e1f5f9a6ed470c535f9688ec0a1b
Author: José Fonseca <jfonseca@vmware.com>
Date: Thu Jun 28 18:07:10 2012 +0100
gallivm: Refactor lp_build_broadcast(_scalar) to share code.
Doesn't really change the generated assembly, but produces more compact IR,
and of course, makes code more consistent.
Reviewed-by: Brian Paul <brianp@vmware.com>
commit 66712ba2731fc029fa246d4fc477d61ab785edb5
Author: José Fonseca <jfonseca@vmware.com>
Date: Wed Jun 27 17:30:13 2012 +0100
gallivm: Make LLVMContextRef a singleton.
There are any places inside LLVM that depend on it. Too many to attempt
to fix.
Reviewed-by: Brian Paul <brianp@vmware.com>
commit ff5fb7897495ac263f0b069370fab701b70dccef
Author: Roland Scheidegger <sroland@vmware.com>
Date: Thu Jun 28 18:15:27 2012 +0200
gallivm: don't use 8-wide texture fetch in aos path
This appears to be a slight loss usually.
There are probably several reasons for that:
- fetching itself is scalar
- filtering is pure int code hence needs splitting anyway, same
for the final texel offset calculations
- texture wrap related code, which can be done 8-wide, is slightly more
complex with floats (with clamp_to_edge) and float operations generally
more costly hence probably not much faster overall
- the code needed to split when encountering different mip levels for the
quads, adding complexity
So, just split always for aos path (but leave it 8-wide for soa, since we
do 8-wide filtering there when possible).
This should certainly be revisited if we'd have avx2 support.
Reviewed-by: José Fonseca <jfonseca@vmware.com>
commit ce8032b43dcd8e8d816cbab6428f54b0798f945d
Author: Roland Scheidegger <sroland@vmware.com>
Date: Wed Jun 27 18:41:19 2012 +0200
gallivm: (trivial) don't extract fparts variable if not needed
Did not have any consequences but unnecessary.
commit aaa9aaed8f80dc282492f62aa583a7ee23a4c6d5
Author: Roland Scheidegger <sroland@vmware.com>
Date: Wed Jun 27 18:09:06 2012 +0200
gallivm: fix precision issue in aos linear int wrap code
now not just passes at a quick glance but also with piglit...
If we do the wrapping with floats, we also need to set the
weights accordingly. We can potentially end up with different
(integer) coordinates than what the integer calculations would
have chosen, which means the integer weights calculated previously
in this case are completely wrong. Well at least that's what I think
happens, at least recalculating the weights helps.
(Some day really should refactor all the wrapping, so we do whatever is
fastest independent of 16bit int aos or 32bit float soa filtering.)
Reviewed-by: José Fonseca <jfonseca@vmware.com>
commit fd6f18588ced7ac8e081892f3bab2916623ad7a2
Author: José Fonseca <jfonseca@vmware.com>
Date: Wed Jun 27 11:15:53 2012 +0100
gallium/util: Fix parsing of options with underscore.
For example
GALLIVM_DEBUG=no_brilinear
which was being parsed as two options, "no" and "brilinear".
commit 09a8f809088178a03e49e409fa18f1ac89561837
Author: James Benton <jbenton@vmware.com>
Date: Tue Jun 26 15:00:14 2012 +0100
gallivm: Added a generic lp_build_print_value which prints a LLVMValueRef.
Updated lp_build_printf to share common code.
Removed specific lp_build_print_vecX.
Reviewed-by: José Fonseca <jfonseca@vmware.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
commit e59bdcc2c075931bfba2a84967a5ecd1dedd6eb0
Author: José Fonseca <jfonseca@vmware.com>
Date: Wed May 16 15:00:23 2012 +0100
draw,llvmpipe: Avoid named struct types on LLVM 3.0 and later.
Starting with LLVM 3.0, named structures are meant not for debugging, but
for recursive data types, previously also known as opaque types.
The recursive nature of these types leads to several memory management
difficulties. Given that we don't actually need recursive types, avoid
them altogether.
This is an attempt to address fdo bugs 41791 and 44466. The issue is
somewhat random so there's no easy way to check how effective this is.
Cherry-picked from 9af1ba565d
commit df6070f618a203c7a876d984c847cde4cbc26bdb
Author: Roland Scheidegger <sroland@vmware.com>
Date: Wed Jun 27 14:42:53 2012 +0200
gallivm: (trivial) fix typo in faster aos linear int wrap code
no longer crashes, now REALLY tested.
commit d8f98dce452c867214e6782e86dc08562643c862
Author: Roland Scheidegger <sroland@vmware.com>
Date: Tue Jun 26 18:20:58 2012 +0200
llvmpipe: (trivial) remove bogus optimization for float aos repeat wrap
This optimization for nearest filtering on the linear path generated
likely bogus results, and the int path didn't have any optimizations
there since the only shader using force_nearest apparently uses
clamp_to_edge not repeat wrap anyway.
Reviewed-by: José Fonseca <jfonseca@vmware.com>
commit c4e271a0631087c795e756a5bb6b046043b5099d
Author: Roland Scheidegger <sroland@vmware.com>
Date: Tue Jun 26 23:01:52 2012 +0200
gallivm: faster repeat wrap for linear aos path too
Even if we already have scaled integer coords, it's way faster to use
the original float coord (plus some conversions) rather than use URem.
The choice of what to do for texture wrapping is not really tied to int
aos or float soa filtering though for some modes there can be some gains
(because of easier weight calculations).
Reviewed-by: José Fonseca <jfonseca@vmware.com>
commit 1174a75b1806e92aee4264ffe0ffe7e70abbbfa3
Author: Roland Scheidegger <sroland@vmware.com>
Date: Tue Jun 26 14:39:22 2012 +0200
gallivm: improve npot tex wrap repeat in linear soa path
URem gets translated into series of scalar divisions so
just about anything else is faster.
Reviewed-by: José Fonseca <jfonseca@vmware.com>
commit f849ffaa499ed96fa0efd3594fce255c7f22891b
Author: Roland Scheidegger <sroland@vmware.com>
Date: Tue Jun 26 00:40:35 2012 +0100
gallivm: (trivial) fix near-invisible shift-space typo
I blame the keyboard.
commit 5298a0b19fe672aebeb70964c0797d5921b51cf0
Author: Roland Scheidegger <sroland@vmware.com>
Date: Mon Jun 25 16:24:28 2012 +0200
gallivm: add new intrinsic helper to deal with arbitrary vector length
This helper will split vectors which are too large for the hw, or expand
them if they are too small, so a caller of a function using intrinsics which
uses such sizes need not split (or expand) the vectors manually and the
function will still use the intrinsic instead of dropping back to generic
llvm code. It can also accept scalars for use with pseudo-vector intrinsics
(only useful for float arguments, all x86 scalar simd float intrinsics use
4vf32).
Only used for lp_build_min/max() for now (also added the scalar float case
for these while there). (Other basic binary functions could use it easily,
whereas functions with a different interface would need different helpers.)
Expanding vectors isn't widely used, because we always try to use
build contexts with native hw vector sizes. But it might (or not) be nicer
if this wouldn't need to be done, the generated code should in theory stay
the same (it does get hit by lp_build_rho though already since we
didn't have a intrinsic for the scalar lp_build_max case before).
v2: incorporated Brian's feedback, and also made the scalar min/max case work
instead of crash (all scalar simd float intrinsics take 4vf32 as argument,
probably the reason why it wasn't used before).
Moved to lp_bld_intr based on José's request, and passing intrinsic size
instead of length.
Ideally we'd derive the source type info from the passed in llvm value refs
and process some llvmtype return type so we could handle intrinsics where
the source and destination type isn't the same (like float/int conversions,
packing instructions) but that's a bit too complicated for now.
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: José Fonseca <jfonseca@vmware.com>
commit 01aa760b99ec0b2dc8ce57a43650e83f8c1becdf
Author: Roland Scheidegger <sroland@vmware.com>
Date: Mon Jun 25 16:19:18 2012 +0200
gallivm: (trivial) increase max code size for shader disassembly
64kB was just short of what I needed (which caused a crash) hence
increase to 96kB (should probably be smarter about that).
commit 74aa739138d981311ce13076388382b5e89c6562
Author: Roland Scheidegger <sroland@vmware.com>
Date: Mon Jun 25 11:53:29 2012 +0100
gallivm: simplify aos float tex wrap repeat nearest
just handle pot and npot the same. The previous pot handling
ended up with exactly the same instructions plus 2 more (leave it
in the soa path though since it is probably still cheaper there).
While here also fix a issue which would cause a crash after an assert.
commit 0e1e755645e9e49cfaa2025191e3245ccd723564
Author: Roland Scheidegger <sroland@vmware.com>
Date: Mon Jun 25 11:29:24 2012 +0100
gallivm: (trivial) skip floor rounding in ifloor when not signed
This was only done for the non-sse41 case before, but even with
sse41 this is obviously unnecessary (some callers already call
itrunc in this case anyway but some might not).
commit 7f01a62f27dcb1d52597b24825931e88bae76f33
Author: Roland Scheidegger <sroland@vmware.com>
Date: Mon Jun 25 11:23:12 2012 +0100
gallivm: (trivial) fix bogus comments
commit 5c85be25fd82e28490274c468ce7f3e6e8c1d416
Author: José Fonseca <jfonseca@vmware.com>
Date: Wed Jun 20 11:51:57 2012 +0100
translate: Free elt8_func/elt16_func too.
These were leaking.
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
commit 0ad498f36fb6f7458c7cffa73b6598adceee0a6c
Author: Roland Scheidegger <sroland@vmware.com>
Date: Tue Jun 19 15:55:34 2012 +0200
gallivm: fix bug for tex wrap repeat with linear sampling in aos float path
The comparison needs to be against length not length_minus_one, otherwise
the max texel is never chosen (for the second coordinate).
Fixes piglit texwrap-1D-npot-proj (and 2D/3D versions).
Reviewed-by: José Fonseca <jfonseca@vmware.com>
commit d1ad65937c5b76407dc2499b7b774ab59341209e
Author: Roland Scheidegger <sroland@vmware.com>
Date: Tue Jun 19 16:13:43 2012 +0200
gallivm: simplify soa tex wrap repeat with npot textures and no mip filtering
Similar to what is already done in aos sampling for the float path (but not
the int path since we don't get normalized float coordinates there).
URem is expensive and the calculation is done trivially with
normalized floats instead (at least with sse41-capable cpus).
(Some day should probably do the same for the mip filter path but it's much
more complicated there hence the gain is smaller.)
Reviewed-by: José Fonseca <jfonseca@vmware.com>
commit e1e23f57ba9b910295c306d148f15643acc3fc83
Author: Roland Scheidegger <sroland@vmware.com>
Date: Mon Jun 18 20:38:56 2012 +0200
llvmpipe: (trivial) remove duplicated function declaration
Reviewed-by: José Fonseca <jfonseca@vmware.com>
commit 07ca57eb09e04c48a157733255427ef5de620861
Author: Roland Scheidegger <sroland@vmware.com>
Date: Mon Jun 18 20:37:34 2012 +0200
llvmpipe: destroy setup variants on context destruction
lp_delete_setup_variants() used to be called in garbage collection,
but this no longer exists hence the setup shaders never got freed.
Reviewed-by: José Fonseca <jfonseca@vmware.com>
commit ed0003c633859a45f9963a479f4c15ae0ef1dca3
Author: Roland Scheidegger <sroland@vmware.com>
Date: Mon Jun 18 16:25:29 2012 +0100
gallivm: handle different ilod parts for multiple quad sampling
This fixes filtering when the integer part of the lod is not the same
for all quads. I'm not fully convinced of that solution yet as it just
splits the vector if the levels to be sampled from are different.
But otherwise we'd need to do things like some minify steps, and getting
mip level base address separately anyway hence it wouldn't really look
like much of a win (and making the code even more complex).
This should now give identical results to single quad sampling.
commit 8580ac4cfc43a64df55e84ac71ce1a774d33c0d2
Author: Roland Scheidegger <sroland@vmware.com>
Date: Thu Jun 14 18:14:47 2012 +0200
gallivm: de-duplicate sample code common to soa and aos sampling
There doesn't seem to be any reason why this code dealing with cube face
selection, lod and mip level calculation is separate in aos and
soa sampling, and I am sick of having it to change in both places.
commit fb541e5f957408ce305b272100196f1e12e5b1e8
Author: Roland Scheidegger <sroland@vmware.com>
Date: Thu Jun 14 18:15:41 2012 +0200
gallivm: do mip filtering with per quad lod_fpart
This gives better results for mip filtering, though the generated code might
not be optimal. For now it also creates some artifacts if the lod_ipart isn't
the same for all quads, since instead of using the same mip weight for all
quads as previously (which just caused non-smooth gradients) this now will
use the right weights but with the wrong mip level in this case (can easily
be seen with things like texfilt, mipmap_tunnel).
v2: use logic helper suggested by José, and fix issue with negative lod_fpart
values
commit f1cc84eef7d826a20fab6cd8ccef9a275ff78967
Author: Roland Scheidegger <sroland@vmware.com>
Date: Wed Jun 13 18:35:25 2012 +0200
gallivm: (trivial) fix bogus assert in lp_build_unpack_broadcast_aos_scalars
commit 7c17dbae8ae290df9ce0f50781a09e8ed640c044
Author: James Benton <jbenton@vmware.com>
Date: Tue Jun 12 12:11:14 2012 +0100
util: Reimplement half <-> float conversions.
Removed u_half.py used to generate the table for previous method.
Previous implementation of float to half conversion was faulty for
denormalised and NaNs and would require extra logic to fix,
thus making the speedup of using tables irrelevant.
commit 7762f59274070e1dd4b546f5cb431c2eb71ae5c3
Author: James Benton <jbenton@vmware.com>
Date: Tue Jun 12 12:12:16 2012 +0100
tests: Updated tests to properly handle NaN for half floats.
commit fa94c135aea5911fd93d5dfb6e6f157fb40dce5e
Author: Roland Scheidegger <sroland@vmware.com>
Date: Mon Jun 11 18:33:10 2012 +0200
gallivm: do mip level calculations per quad
This is the final piece which shouldn't change the rendering output yet.
Reviewed-by: José Fonseca <jfonseca@vmware.com>
commit 23cbeaddfe03c09ca18c45d28955515317ffcf4c
Author: Roland Scheidegger <sroland@vmware.com>
Date: Sat Jun 9 00:54:21 2012 +0200
gallivm: do per-quad cube face selection
Doesn't quite fix the piglit cubemap test (not sure why actually)
but doing per-quad face selection is doing the right thing and
definitely an improvement.
Reviewed-by: José Fonseca <jfonseca@vmware.com>
commit abfb372b3702ac97ac8b5aa80ad1b94a2cc39d33
Author: Roland Scheidegger <sroland@vmware.com>
Date: Mon Jun 11 18:22:59 2012 +0200
gallivm: do all lod calculations per quad
Still no functional change but lod is now converted to scalar after
lod calculations.
Reviewed-by: José Fonseca <jfonseca@vmware.com>
commit 519368632747ae03feb5bca9c655eccbc5b751b4
Author: James Benton <jbenton@vmware.com>
Date: Tue May 22 16:46:10 2012 +0100
gallivm: Added support for half-float to float conversion in lp_build_conv.
Updated various utility functions to support this change.
commit 135b4d683a4c95f7577ba27b9bffa4a6fbd2c2e7
Author: James Benton <jbenton@vmware.com>
Date: Tue May 22 16:02:46 2012 +0100
gallivm: Added function for half-float to float conversion.
Updated lp_build_format_aos_array to support half-float source.
commit 37d648827406a20c5007abeb177698723ed86673
Author: James Benton <jbenton@vmware.com>
Date: Tue May 22 14:55:18 2012 +0100
util: Updated u_format_tests to rigidly test half-float boundary values.
commit 2ad18165d96e578aa9046df7c93cb1c3284d8c6b
Author: James Benton <jbenton@vmware.com>
Date: Tue May 22 14:54:16 2012 +0100
llvmpipe: Updated lp_test_format to properly handle Inf/NaN results.
commit 78740acf25aeba8a7d146493dd5c966e22c27b73
Author: James Benton <jbenton@vmware.com>
Date: Tue May 22 14:53:30 2012 +0100
util: Added functions for checking NaN / Inf for double and half-floats.
commit 35e9f640ae01241f9e0d67fe893bbbf564c05809
Author: Roland Scheidegger <sroland@vmware.com>
Date: Thu May 24 21:05:13 2012 +0200
gallivm: Fix calculating rho for 3d textures for the single-quad case
Discovered by accident, this looks like a very old typo bug.
commit fc1220c636326536fd0541913154e62afa7cd1d8
Author: Roland Scheidegger <sroland@vmware.com>
Date: Thu May 24 21:04:59 2012 +0200
gallivm: do calcs per-quad in lp_build_rho
Still convert to scalar at the end of the function.
commit 50a887ffc550bf310a6988fa2cea5c24d38c1a41
Author: Roland Scheidegger <sroland@vmware.com>
Date: Mon May 21 23:21:50 2012 +0200
gallivm: (trivial) return scalar in lp_build_extract_range for length 1 vectors
Our type system on top of llvm's one doesn't generally support vectors of
length 1, instead using scalars. So we should return a scalar from this
function instead of having to bitcast the vector with length 1 later elsewhere.
commit 80c71c621f9391f0f9230460198d861643324876
Author: James Benton <jbenton@vmware.com>
Date: Tue May 22 17:49:15 2012 +0100
draw: Fixed bad merge error
commit c47401cfad0c9167de20ff560654f533579f452c
Author: James Benton <jbenton@vmware.com>
Date: Tue May 22 15:29:30 2012 +0100
draw: Updated store_clip to store whole vectors instead of individual elements.
commit 2d9c1ad74b0b0b41861fffcecde39f09cc27f1cf
Author: James Benton <jbenton@vmware.com>
Date: Tue May 22 15:28:32 2012 +0100
gallivm: Added lp_build_fetch_rgba_aos_array.
A version of lp_build_fetch_rgba_aos which is targeted at simple array formats.
Reads the whole vector from memory in one, instead of reading each element
individually.
Tested with mesa tests and demos.
commit ff7805dc2b6ef6d8b11ec4e54aab1633aef29ac8
Author: James Benton <jbenton@vmware.com>
Date: Tue May 22 15:27:40 2012 +0100
gallivm: Added lp_build_pad_vector.
This function pads a vector with undef to a desired length.
commit 701f50acef24a2791dabf4730e5b5687d6eb875d
Author: James Benton <jbenton@vmware.com>
Date: Fri May 18 17:27:19 2012 +0100
util: Added util_format_is_array.
This function checks whether a format description is in a simple array format.
commit 5e0a7fa543dcd009de26f34a7926674190fa6246
Author: James Benton <jbenton@vmware.com>
Date: Fri May 18 19:13:47 2012 +0100
draw: Removed draw_llvm_translate_from and draw/draw_llvm_translate.c.
This is "replaced" by adding an optimised path in lp_build_fetch_rgba_aos
in an upcoming patch.
commit 8c886d6a7dd3fb464ecf031de6f747cb33e5361d
Author: James Benton <jbenton@vmware.com>
Date: Wed May 16 15:02:31 2012 +0100
draw: Modified store_aos to write the vector as one, not individual elements.
commit 37337f3d657e21dfd662c7b26d61cb0f8cfa6f17
Author: James Benton <jbenton@vmware.com>
Date: Wed May 16 14:16:23 2012 +0100
draw: Changed aos_to_soa to use lp_build_transpose_aos.
commit bd2b69ce5d5c94b067944d1dcd5df9f8e84548f1
Author: James Benton <jbenton@vmware.com>
Date: Fri May 18 19:14:27 2012 +0100
draw: Changed soa_to_aos to use lp_build_transpose_aos.
commit 0b98a950d29a116e82ce31dfe7b82cdadb632f2b
Author: James Benton <jbenton@vmware.com>
Date: Fri May 18 18:57:45 2012 +0100
gallivm: Added lp_build_transpose_aos which converts between aos and soa.
commit 69ea84531ad46fd145eb619ed1cedbe97dde7cb5
Author: James Benton <jbenton@vmware.com>
Date: Fri May 18 18:57:01 2012 +0100
gallivm: Added lp_build_interleave2_half aimed at AVX unpack instructions.
commit 7a4cb1349dd35c18144ad5934525cfb9436792f9
Author: José Fonseca <jfonseca@vmware.com>
Date: Tue May 22 11:54:14 2012 +0100
gallivm: Fix build on Windows.
MC-JIT not yet supported there.
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
commit afd105fc16bb75d874e418046b80d9cc578818a1
Author: James Benton <jbenton@vmware.com>
Date: Fri May 18 16:17:26 2012 +0100
llvmpipe: Added a error counter to lp_test_conv.
Useful for keeping track of progress when fixing errors!
Signed-off-by: José Fonseca <jfonseca@vmware.com>
commit b644907d08c10a805657841330fc23db3963d59c
Author: James Benton <jbenton@vmware.com>
Date: Fri May 18 16:16:46 2012 +0100
llvmpipe: Changed known failures in lp_test_conv.
To comply with the recent fixes to lp_bld_conv.
Signed-off-by: José Fonseca <jfonseca@vmware.com>
commit d7061507bd94f6468581e218e61261b79c760d4f
Author: James Benton <jbenton@vmware.com>
Date: Fri May 18 16:14:38 2012 +0100
llvmpipe: Added fixed point types tests to lp_test_conv.
Signed-off-by: José Fonseca <jfonseca@vmware.com>
commit 146b3ea39b4726dbe125ac666bd8902ea3d6ca8c
Author: James Benton <jbenton@vmware.com>
Date: Fri May 18 16:26:35 2012 +0100
llvmpipe: Changed lp_test_conv src/dst alignment to be correct.
Now based on the define rather than a fixed number.
Signed-off-by: José Fonseca <jfonseca@vmware.com>
commit f3b57441f834833a4b142a951eb98df0aa874536
Author: James Benton <jbenton@vmware.com>
Date: Fri May 18 16:06:44 2012 +0100
gallivm: Fixed erroneous optimisation in lp_build_min/max.
Previously assumed normalised was 0 to 1, but it can be -1 to 1
if type is signed.
Tested with lp_test_conv and lp_test_format, reduced errors.
Signed-off-by: José Fonseca <jfonseca@vmware.com>
commit a0613382e5a215cd146bb277646a6b394d376ae4
Author: James Benton <jbenton@vmware.com>
Date: Fri May 18 16:04:49 2012 +0100
gallivm: Compensate for lp_const_offset in lp_build_conv.
Fixing a /*FIXME*/ to remove errors in integer conversion in lp_build_conv.
Tested using lp_test_conv and lp_test_format, reduced errors.
Signed-off-by: José Fonseca <jfonseca@vmware.com>
commit a3d2bf15ea345bc8a0664f8f441276fd566566f3
Author: James Benton <jbenton@vmware.com>
Date: Fri May 18 16:01:25 2012 +0100
gallivm: Fixed overflow in lp_build_clamped_float_to_unsigned_norm.
Tested with lp_test_conv and lp_test_format, reduced errors.
Signed-off-by: José Fonseca <jfonseca@vmware.com>
commit e7b1e76fe237613731fa6003b5e1601a2e506207
Author: José Fonseca <jfonseca@vmware.com>
Date: Mon May 21 20:07:51 2012 +0100
gallivm: Fix build with LLVM 2.6
Trivial, and useful.
commit d3c6bbe5c7f5ba1976710831281ab1b6a631082d
Author: José Fonseca <jfonseca@vmware.com>
Date: Tue May 15 17:15:59 2012 +0100
gallivm: Enable MCJIT/AVX with vanilla LLVM 3.1.
Add the necessary C++ glue, so that we don't need any modifications
to the soon to be released LLVM 3.1.
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
commit 724a019a14d40fdbed21759a204a2bec8a315636
Author: José Fonseca <jfonseca@vmware.com>
Date: Mon May 14 22:04:06 2012 +0100
gallivm: Use HAVE_LLVM 0x0301 consistently.
commit af6991e2a3868e40ad599b46278551b794839748
Author: José Fonseca <jfonseca@vmware.com>
Date: Mon May 14 21:49:06 2012 +0100
gallivm: Add MCRegisterInfo.h to silence benign warnings about missing implementation.
Trivial.
commit 6f8a1d75458daae2503a86c6b030ecc4bb494e23
Author: Vinson Lee <vlee@freedesktop.org>
Date: Mon Apr 2 22:14:15 2012 -0700
gallivm: Pass in a MCInstrInfo to createMCInstPrinter on llvm-3.1.
llvm-3.1svn r153860 makes MCInstrInfo available to the MCInstPrinter.
Signed-off-by: Vinson Lee <vlee@freedesktop.org>
Reviewed-by: Brian Paul <brianp@vmware.com>
commit 62555b6ed8760545794f83064e27cddcb3ce5284
Author: Vinson Lee <vlee@freedesktop.org>
Date: Tue Mar 27 21:51:17 2012 -0700
gallivm: Fix method overriding in raw_debug_ostream.
Use matching type qualifers to avoid method hiding.
Signed-off-by: Vinson Lee <vlee@freedesktop.org>
Reviewed-by: José Fonseca <jfonseca@vmware.com>
commit 6a9bd784f4ac68ad0a731dcd39e5a3c39989f2be
Author: Vinson Lee <vlee@freedesktop.org>
Date: Tue Mar 13 22:40:52 2012 -0700
gallivm: Fix createOProfileJITEventListener namespace with llvm-3.1.
llvm-3.1svn r152620 refactored the OProfile profiling code.
createOProfileJITEventListener was moved from the llvm namespace to the
llvm::JITEventListener namespace.
Signed-off-by: Vinson Lee <vlee@freedesktop.org>
Reviewed-by: Brian Paul <brianp@vmware.com>
commit b674955d39adae272a779be85aa1bd665de24e3e
Author: Vinson Lee <vlee@freedesktop.org>
Date: Mon Mar 5 22:00:40 2012 -0800
gallivm: Pass in a MCRegisterInfo to MCInstPrinter on llvm-3.1.
llvm-3.1svn r152043 changes createMCInstPrinter to take an additional
MCRegisterInfo argument.
Signed-off-by: Vinson Lee <vlee@freedesktop.org>
Reviewed-by: Brian Paul <brianp@vmware.com>
commit 11ab69971a8a31c62f6de74905dbf8c02884599f
Author: Vinson Lee <vlee@freedesktop.org>
Date: Wed Feb 29 21:20:53 2012 -0800
Revert "gallivm: Change getExtent and readByte to non-const with llvm-3.1."
This reverts commit d5a6c17254.
llvm-3.1svn r151687 makes MemoryObject accessor members const again.
Signed-off-by: Vinson Lee <vlee@freedesktop.org>
Reviewed-by: Brian Paul <brianp@vmware.com>
commit 339960c82d2a9f5c928ee9035ed31dadb7f45537
Author: Roland Scheidegger <sroland@vmware.com>
Date: Mon May 14 16:19:56 2012 +0200
gallivm: (trivial) fix assertion failure for mipmapped 1d textures
In lp_build_rho, we may end up with a 1-element vector (for mipmapped 1d
textures), but in this case we require the type to be a non-vector type,
so need a cast.
commit 9d73edb727bd6d196030dc3026b7bf0c574b3e19
Author: Roland Scheidegger <sroland@vmware.com>
Date: Thu May 10 18:12:07 2012 +0200
gallivm: prepare for per-quad lod calculations for large vectors
to be able to handle multiple quads at once in texture sampling and still
do lod calculations per quad, it is necessary to get the per-quad derivatives
into the lp_build_rho function.
Until now these derivative values were just scalars, which isn't going to work.
So we now use vectors, and since the interface needs to change we also do some
different (slightly more efficient) packing of the values.
For 8-wide vectors the packed derivative values for 3 coords would look like
this, this scales to a arbitrary (multiple of 4) vector size:
ds1dx ds1dy dt1dx dt1dy ds2dx ds2dy dt2dx dt2dy
dr1dx dr1dy _____ _____ dr2dx dr2dy _____ _____
The second vector will be unused for 1d and 2d textures.
To facilitate future changes the derivative values are put into a struct, since
quite some functions just pass these values through.
The generated code seems to be very slightly better for 2d textures (with
4-wide vectors) than before with sse2 (if you have a cpu with physical 128bit
simd units - otherwise it's probably not a win).
v2: suggestions from José, rename variables, add comments, use swizzle helper
commit 0aa21de0d31466dac77b05c97005722e902517b8
Author: Roland Scheidegger <sroland@vmware.com>
Date: Thu May 10 18:10:31 2012 +0200
gallivm: add undefined swizzle handling to lp_build_swizzle_aos
This is useful for vectors with "holes", it lets llvm choose the most
efficient shuffle instructions if some elements aren't needed without having to
worry what elements to manually pick otherwise.
commit 00faf3f370e7ce92f5ef51002b0ea42ef856e181
Author: José Fonseca <jfonseca@vmware.com>
Date: Fri May 4 17:25:16 2012 +0100
gallivm: Get the LLVM IR optimization passes before JIT compilation.
MC-JIT engine compiles the module immediately on creation, so the optimization
passes were being run too late.
So now we create a target data layout from a string, that matches the
ABI parameters reported by the compiler.
The backend optimization passes were always been run, so the performance
improvement is modest (3% on multiarb mesa demo).
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
commit 40a43f4e2ce3074b5ce9027179d657ebba68800a
Author: Roland Scheidegger <sroland@vmware.com>
Date: Wed May 2 16:03:54 2012 +0200
gallivm: (trivial) fix wrong define used in lp_build_pack2
should fix stack-smashing crashes.
commit e6371d0f4dffad4eb3b7a9d906c23f1c88a2ab9e
Author: Roland Scheidegger <sroland@vmware.com>
Date: Mon Apr 30 21:25:29 2012 +0200
gallivm: add perf warnings when not using intrinsics with 256bit vectors
Helper functions using integer sse2 intrinsics could split the vectors with AVX
instead of using generic fallback (which should be faster).
We don't actually expect to hit these paths (hence don't fix them up to actually
do the vector splitting) so just emit warnings (for those functions where it's
obvious doing split/intrinsic is faster than using generic path).
Only emit warnings for 256bit vectors since we _really_ don't expect to hit
arbitrary large vectors which would affect a lot more functions.
The warnings do not actually depend on avx since the same logic applies to
plain sse2 too (but of course again there's _really_ no reason we should hit
these functions with 256bit vectors without avx).
commit 8a9ea701ea7295181e846c6383bf66a5f5e47637
Author: Roland Scheidegger <sroland@vmware.com>
Date: Tue May 1 20:37:07 2012 +0200
gallivm: split vectors manually for avx in lp_build_pack2 (v2)
There's 2 reasons for this:
First, there's a llvm bug (fixed in 3.1) which generates tons of byte
inserts/extracts otherwise, and second, more importantly, we want to use
pack intrinsics instead of shuffles.
We do this in lp_build_pack2 and not the calling code (aos sample path)
because potentially other callers might find that useful too, even if
for larger sequences of code using non-native vector sizes it might be
better to manually split vectors.
This should boost texture performance in the aos path considerably.
v2: fix issues with intrinsics types with old llvm
commit 27ac5b48fa1f2ea3efeb5248e2ce32264aba466e
Author: Roland Scheidegger <sroland@vmware.com>
Date: Tue May 1 20:26:22 2012 +0200
llvmpipe: refactor lp_build_pack2 (v2)
prettify, and it's unnecessary to assert when there's no intrinsic due to
unsupported bit width - the shuffle path will work regardless.
In contrast lp_build_packs2, should only rely on lp_build_pack2 doing the
clamping for element sizes for which there is a sse2 intrinsic.
v2: fix bug spotted by Jose regarding the intrinsic type for packusdw
on old llvm versions.
commit ddf279031f0111de4b18eaf783bdc0a1e47813c8
Author: Roland Scheidegger <sroland@vmware.com>
Date: Tue May 1 20:13:59 2012 +0200
gallivm: add src width check in lp_build_packs2()
not doing so would skip clamping even if no sse2 pack instruction is
available, which is incorrect (in theory only, such widths would also always
hit a (unnecessary) assertion in lp_build_pack2().
commit e7f0ad7fe079975eae7712a6e0c54be4fae0114b
Author: Roland Scheidegger <sroland@vmware.com>
Date: Fri Apr 27 15:57:00 2012 +0200
gallivm: (trivial) fix crash-causing typo for npot textures with avx
commit 28a9d7f6f655b6ec508c8a3aa6ffefc1e79793a0
Author: Roland Scheidegger <sroland@vmware.com>
Date: Wed Apr 25 19:38:45 2012 +0200
gallivm: (trivial) remove code mistakenly added twice.
commit d5926537316f8ff67ad0a52e7242f7c5478d919b
Author: Roland Scheidegger <sroland@vmware.com>
Date: Tue Apr 24 21:16:15 2012 +0200
gallivm: add a new avx aos sample path (v2)
Try to avoid mixing float and int address calculations. This does texture wrap
modes with floats, and then the offset calculations still with ints (because
of lack of precision with floats, though we could do some effort to make it work
with not too large (16MB) textures).
This also handles wrap repeat mode with npot-sized textures differently than
either the old soa or aos int path (likely way faster but untested).
Otherwise the actual address wrap code is largely similar to the soa path (not
quite the same as this one also has some int code), it should get used by avx
soa sampling later as well but doesn't handle more complex address modes yet
(this will also have the benefit that we can use aos sampling path for all
texture address modes).
Generated code for that looks reasonable, but still does not split vectors
explicitly for fetch/filter which means still get hit by llvm (fixed upstream)
which generates hundreds of pinsrb/pextrb instead of two shuffles.
It is not obvious though if it's much of a win over just doing address calcs
4-wide but with ints, even if it is definitely much less instructions on avx.
piglit's texwrap seems to look exactly the same but doesn't test
neither the non-normalized nor the npot cases.
v2: fix comments, prettify based on Brian's and Jose's feedback.
commit bffecd22dea66fb416ecff8cffd10dd4bdb73fce
Author: Roland Scheidegger <sroland@vmware.com>
Date: Thu Apr 19 01:58:29 2012 +0200
gallivm: refactor aos lp_build_sample_image_nearest/linear
split them up to separate address calculations and fetching/filtering.
Need this for being able to do 8-wide float address calcs and 4-wide
fetch/filter later (for avx). Plus the functions were very big scary monsters
anyway (in particular lp_build_sample_image_linear).
commit a80b325c57529adddcfa367f96f03557725c4773
Author: Roland Scheidegger <sroland@vmware.com>
Date: Mon Apr 16 17:17:18 2012 +0200
gallivm: fix lp_build_resize when truncating width but expanding vector size
Missed this case which I thought was impossible - the assertion for it was
right after the division by zero...
(AoS) texture sampling may ask us to do this, for things like 8 4x32int
vectors to 1 32x8int vector conversion (eventually, we probably don't want
this to happen).
commit f9c8337caa3eb185830d18bce8b95676a065b1d7
Author: Roland Scheidegger <sroland@vmware.com>
Date: Sat Apr 14 18:00:59 2012 +0200
gallivm: fix cube maps with larger vectors
This makes the branchless cube face selection code work with larger vectors.
Because the complexity is quite high (cannot really be improved it seems,
per-face selection would reduce complexity a lot but this leads to errors
unless the derivatives are calculated all from the same face which almost
doubles the work to be done) it is still slower than the branching version,
hence only enable this with large vectors.
It doesn't actually do per-quad face selection yet (only makes sense with
matching lod selection, in fact it will select the same face for all pixels
based on the average of the first four pixels for now) but only different
shuffles are required to make it work (the branching version actually should
work with larger vectors too now thanks to the improved horizontal add but of
course it cannot be extended to really select the face per-quad unless doing
branching per quad).
commit 7780c58869fc9a00af4f23209902db7e058e8a66
Author: Roland Scheidegger <sroland@vmware.com>
Date: Fri Mar 30 21:11:12 2012 +0100
llvmpipe: (trivial) fix compiler warning
and also clarify comment regarding availability of popcnt instruction.
commit a266dccf477df6d29a611154e988e8895892277e
Author: Roland Scheidegger <sroland@vmware.com>
Date: Fri Mar 30 14:21:07 2012 +0100
gallivm: remove unneeded members in lp_build_sample_context
Minor cleanup, the texture width, height, depth aren't accessed in their
scalar form anywhere. Makes it more obvious those values should probably be
fetched already vectorized (but this requires more invasive changes)...
commit b678c57fb474e14f05e25658c829fc04d2792fff
Author: Roland Scheidegger <sroland@vmware.com>
Date: Thu Mar 29 15:53:55 2012 +0100
gallivm: add a helper for concatenating vectors
Similar to the extract_range helper intended to get around slow code generated
by llvm for 128bit insertelements.
Concatenating two 128bit vectors this way will result in a single vinsertf128
operation rather than two 64bit stores plus one 128bit load, though it might be
mildly useful for other purposes as well.
commit 415ff228bcd0cf5e44a4c15350a661f0f5520029
Author: Roland Scheidegger <sroland@vmware.com>
Date: Wed Mar 28 19:41:15 2012 +0100
gallivm: add a custom 2x8f->1x16ub avx conversion path
Similar to the existing 4x4f->1x16ub sse2 path, shaves off a couple
instructions (min/max mostly) because it relies on pack intrinsics clamping.
commit 78c08fc89f8fbcc6dba09779981b1e873e2a0299
Author: Roland Scheidegger <sroland@vmware.com>
Date: Wed Mar 28 18:44:07 2012 +0100
gallivm: add avx arithmetic intrinsics
Add all avx intrinsics for arithmetic functions (with the exception
of the horizontal add function which needs another look).
Seems to pass basic tests.
Reviewed-by: José Fonseca <jfonseca@vmware.com>
commit a586caa2800aa5ce54c173f7c0d4fc48153dbc4e
Author: Roland Scheidegger <sroland@vmware.com>
Date: Wed Mar 28 15:31:35 2012 +0100
gallivm: add avx logic intrinsics
Add the blend intrinsics for 8-wide float and 4-wide double vectors.
Since we lack 256bit int instructions these are used for int vectors as well,
though obviously not for byte or word element values.
The comparison intrinsics aren't extended for avx since these are only used
for pre-2.7 llvm versions.
commit 70275e4c13c89315fc2560a4c488c0e6935d5caf
Author: Roland Scheidegger <sroland@vmware.com>
Date: Wed Mar 28 00:40:53 2012 +0100
gallivm: new helper function for extract shuffles.
Based on José's idea as we can need that in a couple places.
Note that such shuffles should not be used lightly, since data layout
of <4 x i8> is different to <16 x i8> for instance, hence might cause
data rearrangement.
commit 4d586dbae1b0c55915dda1759d2faea631c0a1c2
Author: Roland Scheidegger <sroland@vmware.com>
Date: Tue Mar 27 18:27:25 2012 +0100
gallivm: (trivial) don't overallocate shuffle variable
using wrong define meant huge array...
commit 06b0ec1f6d665d98c135f9573ddf4ba04b2121ad
Author: Roland Scheidegger <sroland@vmware.com>
Date: Tue Mar 27 17:54:20 2012 +0100
gallivm: don't do per-element extract/insert for vector element resize
Instead of doing per-element extract/insert if the src vectors
and dst vector differ in total size (which generates atrocious code)
first change the src vectors size by using shuffles to destination
vector size.
We can still do better than that on AVX for packing to color buffer
(by exploiting pack intrinsics characteristics hence eleminating the
need for some clamps) but this already generates much better code.
v2: incorporate feedback from José, Keith and use shuffle instead of
bitcasts/extracts. Due to llvm deficiencies the latter cause all data
to get moved to GPRs and back in pieces (even though the data in the
regs actually stays the same...).
commit c9970d70e05f95d3f52fe7d2cd794176a52693aa
Author: Roland Scheidegger <sroland@vmware.com>
Date: Fri Mar 23 19:33:19 2012 +0000
gallivm: fix bug in simple position interpolation
Accidental use of position attribute instead of just pixel coordinates.
Caused failures in piglit glsl-fs-ceil and glsl-fs-floor.
commit d0b6fcdb008d04d7f73d3d725615321544da5a7e
Author: Roland Scheidegger <sroland@vmware.com>
Date: Fri Mar 23 15:31:14 2012 +0000
gallivm: fix emission of ceil opcode
lp_build_ceil seems more appropriate than lp_build_trunc.
This seems to be never hit though someone performs some ceil
to floor magic.
commit d97fafed7e62ffa6bf76560a92ea246a1a26d256
Author: Roland Scheidegger <sroland@vmware.com>
Date: Thu Mar 22 11:46:52 2012 +0000
gallivm: new vectorized path for cubemap calculations
should be faster when adapted to multiple quads as only selection masks need to be different.
The code is more or less a per-pixel version adapted to only do it per quad.
A per pixel version would be much simpler (could drop 2 selects, 6 broadcasts and the messy
horizontal add of 3 vectors at the expense of only 2 more absolute value instructions -
would also just work for arbitary large vectors).
This version doesn't yet work with larger vectors because the horizontal add isn't adjusted
to be able to work with 2x4 vectors (and also because face selection wouldn't be done per
quad just per block though that would be only a correctness issue just as with lod selection).
The downside is this code is quite a bit slower. On a Core2 it can be sped up by disabling the
hw blend instructions for selection and using logicop fallbacks instead, but it is still slower
than the old code, hence leave that in for now. Probably will chose one or the other version
based on vector length in the end.
commit b375fbb18a3fd46859b7fdd42f3e9908ea4ff9a3
Author: Roland Scheidegger <sroland@vmware.com>
Date: Wed Mar 21 14:42:29 2012 +0000
gallivm: fix optimized occlusion query intrinsic name
commit a9ba0a3b611e48efbb0e79eb09caa85033dbe9a2
Author: José Fonseca <jfonseca@vmware.com>
Date: Wed Mar 21 16:19:43 2012 +0000
draw,gallivm,llvmpipe: Call gallivm_verify_function everywhere.
commit f94c2238d2bc7383e088b8845b7410439a602071
Author: Roland Scheidegger <sroland@vmware.com>
Date: Tue Mar 20 18:54:10 2012 +0000
gallivm: optimize calculations for cube maps a bit
this does some more vectorized calculations and uses horizontal adds if possible.
A definite win with sse3 otherwise it doesn't seem to make much of a difference.
In any case this is arithmetically identical, cannot handle larger vectors.
Should be useful as a reference point against larger vector version later...
commit 21a2c1cf3c8e1ac648ff49e59fdc0e3be77e2ebb
Author: Roland Scheidegger <sroland@vmware.com>
Date: Tue Mar 20 15:16:27 2012 +0000
llvmpipe: slight optimization of occlusion queries
using movmskps when available.
While this is slightly better for cpus without popcnt we should
really sum the vectors ourselves (it is also possible to cast to i4 before
doing the popcnt but that doesn't help that much neither since llvm
is using some optimized popcnt version for i32)
commit 5ab5a35f216619bcdf55eed52b0db275c4a06c1b
Author: Roland Scheidegger <sroland@vmware.com>
Date: Tue Mar 20 13:32:11 2012 +0000
llvmpipe: fix occlusion queries with larger vectors
need to adjust casts etc.
commit ff95e6fdf5f16d4ef999ffcf05ea6e8c7160b0d5
Author: José Fonseca <jfonseca@vmware.com>
Date: Mon Mar 19 20:15:25 2012 +0000
gallivm: Restore optimization passes.
commit 57b05b4b36451e351659e98946dae27be0959832
Author: Roland Scheidegger <sroland@vmware.com>
Date: Mon Mar 19 19:34:22 2012 +0000
llvmpipe: use existing min2 macro
commit bc9a20e19b4f600a439f45679451f2e87cd4b299
Author: Roland Scheidegger <sroland@vmware.com>
Date: Mon Mar 19 19:07:27 2012 +0000
llvmpipe: add some safeguards against really large vectors
As per José's suggestion, prevent things from blowing up if some cpu
would have 1024bit or larger vectors.
commit 0e2b525e5ca1c5bbaa63158bde52ad1c1564a3a9
Author: Roland Scheidegger <sroland@vmware.com>
Date: Mon Mar 19 18:31:08 2012 +0000
llvmpipe: fix mask generation for uberwide vectors
this was the only piece preventing 16-wide vectors from working
(apart from the LP_MAX_VECTOR_WIDTH define that is), which is the maximum
as we don't get more pixels in the fragment shader at once.
Hence adjust that so things could be tested properly with that size
even though there seems to be no practical value.
commit 3c8334162211c97f3a11c7f64e9e5a2a91ad9656
Author: Roland Scheidegger <sroland@vmware.com>
Date: Mon Mar 19 18:19:41 2012 +0000
llvmpipe: fix the simple interpolation method with larger vectors
so both methods actually _really_ work now. Makes textures look
nice with larger vectors...
commit 1cb0464ef8871be1778d43b0c56adf9c06843e2d
Author: Roland Scheidegger <sroland@vmware.com>
Date: Mon Mar 19 17:26:35 2012 +0000
llvmpipe: fix mask generation and position interpolation with 8-wide vectors
trivial bugs, with these things start to look somewhat reasonable.
Textures though have some swizzling issues it seems.
commit 168277a63ef5b72542cf063c337f2d701053ff4b
Author: Roland Scheidegger <sroland@vmware.com>
Date: Mon Mar 19 16:04:03 2012 +0000
llvmpipe: don't overallocate variables
we never have more than 16 (stamp size) / 4 (minimum possible vector size).
(With larger vectors those variables are still overallocated a bit.)
commit 409b54b30f81ed0aa9ed0b01affe15c72de9abd2
Author: Roland Scheidegger <sroland@vmware.com>
Date: Mon Mar 19 15:56:48 2012 +0000
llvmpipe: add some 32f8 formats to lp_test_conv
Also add the ability to handle different sized vectors.
commit 55dcd3af8366ebdac0af3cdb22c2588f24aa18ce
Author: Roland Scheidegger <sroland@vmware.com>
Date: Mon Mar 19 15:47:27 2012 +0000
gallivm: handle different sized vectors in conversion / pack
only fully generic path for now (extract/insert per element).
commit 9c040f78c54575fcd94a8808216cf415fe8868f6
Author: Roland Scheidegger <sroland@vmware.com>
Date: Sun Mar 18 00:58:28 2012 +0100
llvmpipe: fix harmless use of unitialized values
commit 551e9d5468b92fc7d5aa2265db9a52bb1e368a36
Author: Roland Scheidegger <sroland@vmware.com>
Date: Fri Mar 16 23:31:21 2012 +0100
gallivm: drop special path in extract_broadcast with different sized vectors
Not needed, llvm can handle shuffles with different sized result vector just
fine. Should hopefully generate the same code in the end, but simpler IR.
commit 44da531119ffa07a421eaa041f63607cec88f6f8
Author: Roland Scheidegger <sroland@vmware.com>
Date: Fri Mar 16 23:28:49 2012 +0100
llvmpipe: adapt interpolation for handling multiple quads at once
this is still WIP there are actually two methods possible not quite
sure what makes the most sense, so there's code for both for now:
1) the iterative method as used before (compute attrib values at upper left
corner of stamp and upper left corner of each quad initially).
It is improved to handle more than one quad at once, and also do some more vectorized
calculations initially for slightly better code - newer cpus have full throughput with
4 wide float vectors, hence don't try to code up a path which might be faster if there's
just one channel active per attribute.
2) just do straight interpolation for each pixel.
Method 2) is more work per quad, but less initially - if all quads are executed
significantly more overall though. But this might change with larger vector lengths.
This method would also be needed if we'd do some kind of active quad merging when
operating on multiple quads at once.
This path contains some hack to force llvm to generate better code, it is still far
from ideal though, still generates far too many unnecessary register spills/reloads.
Both methods should work with different sized vectors.
Not very well tested yet, still seems to work with four-wide vectors, need changes
elsewhere to be able to test with wider vectors.
commit be5d3e82e2fe14ad0a46529ab79f65bf2276cd28
Author: José Fonseca <jfonseca@vmware.com>
Date: Fri Mar 16 20:59:37 2012 +0000
draw: Cleanup.
commit f85bc12c7fbacb3de2a94e88c6cd2d5ee0ec0e8d
Author: José Fonseca <jfonseca@vmware.com>
Date: Fri Mar 16 20:43:30 2012 +0000
gallivm: More module compilation refactoring.
commit d76f093198f2a06a93b2204857e6fea5fd0b3ece
Author: José Fonseca <jfonseca@vmware.com>
Date: Thu Mar 15 21:29:11 2012 +0000
llvmpipe: Use gallivm_compile/free_function() in linear code.
Should had been done before.
commit 122e1adb613ce083ad739b153ced1cde61dfc8c0
Author: Roland Scheidegger <sroland@vmware.com>
Date: Tue Mar 13 14:47:10 2012 +0100
llvmpipe: generate partial pixel mask for multiple quads
still works with one quad, cannot be tested yet with more
At least for now always fixed order with multiple quads.
commit 4c4f15081d75ed585a01392cd2dcce0ad10e0ea8
Author: Roland Scheidegger <sroland@vmware.com>
Date: Thu Mar 8 22:09:24 2012 +0100
llvmpipe: refactor state setup a bit
Refactor to make it easier to emit (and potentially later fetch in fs)
coefficients for multiple attributes at once.
Need to think more about how to make this actually happen however, the
problem is different attributes can have different interpolation modes,
requiring different handling in both setup and fs (though linear and
perspective handling is close).
commit 9363e49722ff47094d688a4be6f015a03fba9c79
Author: Roland Scheidegger <sroland@vmware.com>
Date: Thu Mar 8 19:23:23 2012 +0100
llvmpipe: vectorize tri offset calc
cuts number of instructions in quad-offset-factor from 107 to 75.
This code actually duplicated the (scalar) code calculating the determinant
except it used different vertex order (leading to different sign but it doesn't
matter) hence llvm could not have figured out it's the same (of course with
determinant vectorized in the other place that wouldn't have worked any longer
neither).
Note this particular piece doesn't actually vectorize well, not many arithmetic
instructions left but tons of shuffle instructions...
Probably would need to work on n tris at a time for better vectorization.
commit 63169dcb9dd445c94605625bf86d85306e2b4297
Author: Roland Scheidegger <sroland@vmware.com>
Date: Thu Mar 8 03:11:37 2012 +0100
llvmpipe: vectorize some scalar code in setup
reduces number of arithmetic instructions, and avoids loading
vector x,y values twice (once as scalars once as vectors).
Results in a reduction of instructions from 76 to 64 in fs setup for glxgears
(16%) on a cpu with sse41.
Since this code uses vec2 disguised as vec4, on old cpus which had physical
64bit sse units (pre-Core2) it probably is less of a win in practice (and if
you have no vectors you can only hope llvm eliminates the arithmetic for
unneeded elements).
commit 732ecb877f951ab89bf503ac5e35ab8d838b58a1
Author: Roland Scheidegger <sroland@vmware.com>
Date: Wed Mar 7 00:32:24 2012 +0100
draw: fix clipping
bug introduced by 4822fea3f0440b5205e957cd303838c3b128419c broke
clipping pretty badly (verified with lineclip test)
commit ef5d90b86d624c152d200c7c4056f47c3c6d2688
Author: Roland Scheidegger <sroland@vmware.com>
Date: Tue Mar 6 23:38:59 2012 +0100
draw: don't store vertex header per attribute
storing the vertex header once per attribute is totally unnecessary.
Some quick look at the generated assembly says llvm in fact cannot optimize
away the additional stores (maybe due to potentially aliasing pointers
somewhere).
Plus, this makes the code cleaner and also allows using a vector "or"
instead of scalar ones.
commit 6b3a5a57b0b9850854cfbd7b586e4e50102dda71
Author: Roland Scheidegger <sroland@vmware.com>
Date: Tue Mar 6 19:11:01 2012 +0100
draw: do the per-vertex "boolean" clipmask "or" with vectors
no point extracting the values and doing it per component.
Doesn't help that much since we still extract the values elsewhere anyway.
commit 36519caf1af40e4480251cc79a2d527350b7c61f
Author: Roland Scheidegger <sroland@vmware.com>
Date: Fri Mar 2 22:27:01 2012 +0100
gallivm: fix lp_build_extract_broadcast with different sized vectors
Fix the obviously wrong argument, so it doesn't blow up.
commit 76d0ac3ad85066d6058486638013afd02b069c58
Author: José Fonseca <jfonseca@vmware.com>
Date: Fri Mar 2 12:16:23 2012 +0000
draw: Compile per module and not per function (WIP).
Enough to get gears w/ LLVM draw + softpipe to work on AVX doing:
GALLIUM_DRIVER=softpipe SOFTPIPE_USE_LLVM=yes glxgears
But still hackish -- will need to rethink and refactor this.
commit 78e32b247d2a7a771be9a1a07eb000d1e54ea8bd
Author: José Fonseca <jfonseca@vmware.com>
Date: Wed Feb 29 12:01:05 2012 +0000
llvmpipe: Remove lp_state_setup_fallback.
Never used.
commit 6895d5e40d19b4972c361e8b83fdb7eecda3c225
Author: José Fonseca <jfonseca@vmware.com>
Date: Mon Feb 27 19:14:27 2012 +0000
llvmpipe: Don't emit EMMS on x86
We already take precautions to ensure that LLVM never emits MMX code.
commit 4822fea3f0440b5205e957cd303838c3b128419c
Author: Roland Scheidegger <sroland@vmware.com>
Date: Wed Feb 29 15:58:19 2012 +0100
draw: modifications for larger vector sizes
We want to be able to use larger vectors especially for running the vertex
shader. With this patch we build soa vectors which might have a different
length than 4.
Note that aos structures really remain the same, only when aos structures
are converted to soa potentially different sized vectors are used.
Samplers probably don't work yet, didn't look at them.
Testing done:
glxgears works with both 128bit and 256bit vectors.
commit f4950fc1ea784680ab767d3dd0dce589f4e70603
Author: José Fonseca <jfonseca@vmware.com>
Date: Wed Feb 29 15:51:57 2012 +0100
gallivm: override native vector width with LP_NATIVE_VECTOR_WIDTH env var for debug
commit 6ad6dbf0c92f3bf68ae54e5f2aca035d19b76e53
Author: José Fonseca <jfonseca@vmware.com>
Date: Wed Feb 29 15:51:24 2012 +0100
draw: allocate storage with alignment according to native vector width
commit 7bf0e3e7c9bd2469ae7279cabf4c5229ae9880c1
Author: José Fonseca <jfonseca@vmware.com>
Date: Fri Feb 24 19:06:08 2012 +0000
gallivm: Fix comment grammar.
Was missing several words. Spotted by Roland.
commit b20f1b28eb890b2fa2de44a0399b9b6a0d453c52
Author: José Fonseca <jfonseca@vmware.com>
Date: Thu Feb 23 19:22:09 2012 +0000
gallivm: Use MC-JIT on LLVM 3.1 + (i.e, SVN)
MC-JIT
Note: MC-JIT is still WIP. For this to work correctly it requires
LLVM changes which are not yet upstream.
commit b1af4dfcadfc241fd4023f4c3f823a1286d452c0
Author: Roland Scheidegger <sroland@vmware.com>
Date: Thu Feb 23 20:03:15 2012 +0100
llvmpipe: use new lp_type_width() helper in lp_test_blend
commit 04e0a37e888237d4db2298f31973af459ef9c95f
Author: Roland Scheidegger <sroland@vmware.com>
Date: Thu Feb 23 19:50:34 2012 +0100
llvmpipe: clean up lp_test_blend a little
Using variables just sized and aligned right makes it a bit more obvious
what's going on.
The test still only tests vector length 4.
For AoS anything else probably isn't going to work.
For SoA other lengths should work (at least with floats).
commit e61c393d3ec392ddee0a3da170e985fda885a823
Author: José Fonseca <jfonseca@vmware.com>
Date: Thu Feb 23 17:48:30 2012 +0000
gallivm: Ensure vector width consistency.
Instead of assuming that everything is the max native size.
commit 330081ac7bc41c5754a92825e51456d231bf84dd
Author: José Fonseca <jfonseca@vmware.com>
Date: Thu Feb 23 17:44:14 2012 +0000
draw: More simd vector width consistency fixes.
commit d90ca002753596269e37297e2e6c139b19f29f03
Author: José Fonseca <jfonseca@vmware.com>
Date: Thu Feb 23 17:43:00 2012 +0000
gallivm: Remove unused lp_build_int32_vec4_type() helper.
commit cae23417824d75869c202aaf897808d73a2c1db0
Author: Roland Scheidegger <sroland@vmware.com>
Date: Thu Feb 23 17:32:16 2012 +0100
gallivm: use global variable for native vector width instead of define
We do not know the simd extensions (and hence the simd width we should use)
available at compile time.
At least for now keep a define for maximum vector width, since a global
variable obviously can't be used to adjust alignment of automatic stack
variables.
Leave the runtime-determined value at 128 for now in all cases.
commit 51270ace6349acc2c294fc6f34c025c707be538a
Author: José Fonseca <jfonseca@vmware.com>
Date: Thu Feb 23 15:41:02 2012 +0000
gallivm: Add a hunk inadvertedly lost when rebasing.
commit bf256df9cfdd0236637a455cbaece949b1253e98
Author: José Fonseca <jfonseca@vmware.com>
Date: Thu Feb 23 14:24:23 2012 +0000
llvmpipe: Use consistent vector width in depth/stencil test.
commit 5543b0901677146662c44be2cfba655fd55da94b
Author: José Fonseca <jfonseca@vmware.com>
Date: Thu Feb 23 14:19:59 2012 +0000
draw: Use a consistent the vector register width.
Instead of 4x32 sometimes, LP_NATIVE_VECTOR_WIDTH other times.
commit eada8bbd22a3a61f549f32fe2a7e408222e5c824
Author: José Fonseca <jfonseca@vmware.com>
Date: Thu Feb 23 12:08:04 2012 +0000
gallivm: Remove garbagge collection.
MC-JIT will require one compilation per module (as opposed to one
compilation per function), therefore no state will be shared,
eliminating the need to do garbagge collection.
commit 556697ea0ed72e0641851e4fbbbb862c470fd7eb
Author: José Fonseca <jfonseca@vmware.com>
Date: Thu Feb 23 10:33:41 2012 +0000
gallivm: Move all native target initialization to lp_set_target_options().
commit c518e8f3f2649d5dc265403511fab4bcbe2cc5c8
Author: José Fonseca <jfonseca@vmware.com>
Date: Thu Feb 23 09:52:32 2012 +0000
llvmpipe: Create one gallivm instance for each test.
commit 90f10af8920ec6be6f2b1e7365cfc477a0cb111d
Author: José Fonseca <jfonseca@vmware.com>
Date: Thu Feb 23 09:48:08 2012 +0000
gallivm: Avoid LLVMAddGlobalMapping() in lp_bld_assert().
Brittle, complex, and unecesary. Just use function pointer constant.
commit 98fde550b33401e3fe006af59db4db628bcbf476
Author: José Fonseca <jfonseca@vmware.com>
Date: Thu Feb 23 09:21:26 2012 +0000
gallivm: Add a lp_build_const_func_pointer() helper.
To be reused in all places where we want to call C code.
commit 6cfedadb62c2ce5af8d75969bc95a607f3ece118
Author: José Fonseca <jfonseca@vmware.com>
Date: Thu Feb 23 09:44:41 2012 +0000
gallivm: Cleanup/simplify lp_build_const_string_variable.
- Move to lp_bld_const where it belongs
- Rename to lp_build_const_string
- take the length from the argument (and don't count the zero terminator twice)
- bitcast the constant to generic i8 *
commit db1d4018c0f1fa682a9da93c032977659adfb68c
Author: José Fonseca <jfonseca@vmware.com>
Date: Thu Feb 23 11:52:17 2012 +0000
gallivm: Set NoFramePointerElimNonLeaf to true where supported.
commit 088614164aa915baaa5044fede728aa898483183
Author: Roland Scheidegger <sroland@vmware.com>
Date: Wed Feb 22 19:38:47 2012 +0100
llvmpipe: pass in/out pointers rather scalar floats in lp_bld_arit
we don't want llvm to potentially optimize away the vectors (though it doesn't
seem to currently), plus we want to be able to handle in/out vectors of arbitrary
length.
commit 3f5c4e04af8a7592fdffa54938a277c34ae76b51
Author: Roland Scheidegger <sroland@vmware.com>
Date: Tue Feb 21 23:22:55 2012 +0100
gallivm: fix lp_build_sqrt() for vector length 1
since we optimize away vectors with length 1 need to emit intrinsic
without vector type.
commit 79d94e5f93ed8ba6757b97e2026722ea31d32c06
Author: José Fonseca <jfonseca@vmware.com>
Date: Wed Feb 22 17:00:46 2012 +0000
llvmpipe: Remove lp_test_round.
commit 81f41b5aeb3f4126e06453cfc78990086b85b78d
Author: Roland Scheidegger <sroland@vmware.com>
Date: Tue Feb 21 23:56:24 2012 +0100
llvmpipe: subsume lp_test_round into lp_test_arit
Much simpler, and since the arguments aren't passed as 128bit values can run
on any arch.
This also uses the float instead of the double versions of the c functions
(which probably was the intention anyway).
In contrast to lp_test_round the output is much less verbose however.
Tested vector width of 32 to 512 bits - all pass except 32 (length 1) which
crashes in lp_build_sqrt() due to wrong type.
Signed-off-by: José Fonseca <jfonseca@vmware.com>
commit 945b338b421defbd274481d8c4f7e0910fd0e7eb
Author: José Fonseca <jfonseca@vmware.com>
Date: Wed Feb 22 09:55:03 2012 +0000
gallivm: Centralize the function compilation logic.
This simplifies a lot of code.
Also doing this in a central place will make it easier to carry out the
changes necessary to use MC-JIT in the future.
gallivm: Fix typo in explicit derivative shuffle.
Trivial.
draw: make DEBUG_STORE work again
adapt to lp_build_printf() interface changes
Reviewed-by: José Fonseca <jfonseca@vmware.com>
draw: get rid of vecnf_from_scalar()
just use lp_build_broadcast directly (cannot assign a name but don't really
need it, vecnf_from_scalar() was producing much uglier IR due to using
repeated insertelement instead of insertelement+shuffle).
Reviewed-by: José Fonseca <jfonseca@vmware.com>
llvmpipe: fix typo in complex interpolation code
Fixes position interpolation when using complex mode
(piglit fp-fragment-position and similar)
Reviewed-by: José Fonseca <jfonseca@vmware.com>
draw: fix clipvertex/position storing again
This appears to be the result of a bad merge.
Fixes piglit tests relying on clipping, like a lot of the interpolation tests.
Reviewed-by: José Fonseca <jfonseca@vmware.com>
gallivm: Fix explicit derivative manipulation.
Same counter variable was being used in two nested loops. Use more
meanigful variable names for the counter to fix and avoid this.
gallivm: Prevent buffer overflow in repeat wrap mode for NPOT.
Based on Roland's patch, discussion, and review .
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
gallivm: Fix dims for TGSI_TEXTURE_1D in emit_tex.
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
gallivm: Fix explicit volume texture derivatives.
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
gallivm: fix 1d shadow texture sampling
Always r coordinate is used, hence need 3 coords not two
(the second one is unused).
Reviewed-by: José Fonseca <jfonseca@vmware.com>
gallivm: Enable AVX support without MCJIT, where available.
For now, this just enables AVX on Windows for testing. If the code is
stable then we might consider prefering the old JIT wherever possible.
No change elsewhere.
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
- Use LLVM limits when LLVM is being used, instead of TGSI limits
- Provide draw_get_shader_param_no_llvm for when llvm is never used (softpipe)
- Eliminate several of the hacks around draw shader caps in several drivers
Unfortunately the hack for PIPE_MAX_VERTEX_SAMPLERS is still necessary.
Reviewed-by: Marek Olšák <maraeo@gmail.com>
Doesn't really change the generated assembly, but produces more compact IR,
and of course, makes code more consistent.
Reviewed-by: Brian Paul <brianp@vmware.com>
Updated lp_build_printf to share common code.
Removed specific lp_build_print_vecX.
Reviewed-by: José Fonseca <jfonseca@vmware.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
The type is the destination type (i.e. float vector) and not the
source type. Fixes piglit fs-{in,de}crement-uint.
Signed-off-by: Olivier Galibert <galibert@pobox.com>
Signed-off-by: José Fonseca <jfonseca@vmware.com>