This is a better, more fine-grained way of lowering if statements. Fixes the
game And Yet It Moves on nv50.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Using multiply and reciprocal for integer division involves potentially
lossy floating point conversions. This is okay for older GPUs that
represent integers as floating point, but undesirable for GPUs with
native integer division instructions.
TGSI, for example, has UDIV/IDIV instructions for integer division,
so it makes sense to handle this directly. Likewise for i965.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Signed-off-by: Bryan Cain <bryancain3@gmail.com>
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
!a && b occurs frequently when nexted if-statements have been
flattened. It should also be possible use a MAD for (a && b) || c,
though that would require a MAD_SAT.
Reviewed-by: Eric Anholt <eric@anholt.net>
The operation ir_binop_all_equal is !(a.x != b.x || a.y != b.y || a.z
!= b.z || a.w != b.w). Logical-or is implemented using addition
(followed by clampling to [0,1]) on values of 0.0 and 1.0. Replacing
the logical-or operators with addition gives !bool((int(a.x != b.x) +
int(a.y == b.y) + int(a.z == b.z) + int(a.w == b.w)). This can be
implemented using a dot-product with a vector of all 1.0. After the
dot-product, the value will be an integer on the range [0,4].
Previously a SEQ instruction was used to clamp the resulting logic
value to [0,1] and invert the result. Using an SGE instruction on the
negation of the dot-product result has the same effect. Many older
shader architectures do not support the SEQ instruction. It must be
emulated using two SGE instructions and a MUL. On these
architectures, the single SGE saves two instructions.
Reviewed-by: Eric Anholt <eric@anholt.net>
The operation ir_binop_any_nequal is (a.x != b.x) || (a.y != b.y) ||
(a.z != b.z) || (a.w != b.w), and that is the same as any(bvec4(a.x !=
b.x, a.y != b.y, a.z != b.z, a.w != b.w)). Implement the any() part
the same way the regular ir_unop_any is implemented.
Reviewed-by: Eric Anholt <eric@anholt.net>
This is just like the ir_binop_logic_or case. The operation
ir_unop_any is (a.x || a.y || a.z || a.w). Logical-or is implemented
using addition (followed by clampling to [0,1]) on values of 0.0 and
1.0. Replacing the logical-or operators with addition gives (a.x +
a.y + a.z + a.w). This can be implemented using a dot-product with a
vector of all 1.0.
Previously a SNE instruction was used to clamp the resulting logic
value to [0,1]. In a fragment shader, using a saturate on the
dot-product has the same effect. Adding the saturate to the
dot-product is free, so (at least) one instruction is saved.
In a vertex shader, using an SLT on the negation of the dot-product
result has the same effect. Many older shader architectures do not
support the SNE instruction. It must be emulated using two SLT
instructions and an ADD. On these architectures, the single SLT saves
two instructions.
Reviewed-by: Eric Anholt <eric@anholt.net>
Logical-or is implemented using addition (followed by clampling to
[0,1]) on values of 0.0 and 1.0. Replacing the logical-or operators
with addition gives a + b which has a result on the range [0, 2].
Previously a SNE instruction was used to clamp the resulting logic
value to [0,1]. In a fragment shader, using a saturate on the add has
the same effect. Adding the saturate to the add is free, so (at
least) one instruction is saved.
In a vertex shader, using an SLT on the negation of the add result has
the same effect. Many older shader architectures do not support the
SNE instruction. It must be emulated using two SLT instructions and
an ADD. On these architectures, the single SLT saves two
instructions.
Reviewed-by: Eric Anholt <eric@anholt.net>
Rely on the driver to do the right thing. This probably means falling
back to software. Page 88 of the OpenGL 2.1 spec specifically says:
"A shader should not fail to compile, and a program object should
not fail to link due to lack of instruction space or lack of
temporary variables. Implementations should ensure that all valid
shaders and program objects may be successfully compiled, linked
and executed."
There is no provision for saying "No" to a valid shader that is
difficult for the hardware to handle, so stop doing that.
On i915 this causes a large number of piglit tests to change from FAIL
to WARN. The warning is because the driver still emits messages to
stderr like "i915_program_error: Unsupported opcode: BGNLOOP".
It also fixes ES2 conformance CorrectFull_frag and CorrectParse1_frag
on i915 (and probably other hardware that can't handle loops).
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
The functionality is not used by anything yet, and the glUniform functions will
need to be reworked before this can reach its full usefulness. It is
nonetheless a step towards integer support in the state tracker and classic drivers.
This fixes many cases of accessing arrays of matrices using
non-constant indices at each level.
Fixes i965 piglit:
vs-temp-array-mat[234]-index-col-rd
vs-temp-array-mat[234]-index-col-row-rd
vs-temp-array-mat[234]-index-col-wr
vs-uniform-array-mat[234]-index-col-rd
Fixes swrast piglit:
fs-temp-array-mat[234]-index-col-rd
fs-temp-array-mat[234]-index-col-row-rd
fs-temp-array-mat[234]-index-col-wr
fs-uniform-array-mat[234]-index-col-rd
fs-uniform-array-mat[234]-index-col-row-rd
fs-varying-array-mat[234]-index-col-rd
fs-varying-array-mat[234]-index-col-row-rd
vs-temp-array-mat[234]-index-col-rd
vs-temp-array-mat[234]-index-col-row-rd
vs-temp-array-mat[234]-index-col-wr
vs-uniform-array-mat[234]-index-col-rd
vs-uniform-array-mat[234]-index-col-row-rd
vs-varying-array-mat[234]-index-col-rd
vs-varying-array-mat[234]-index-col-row-rd
vs-varying-array-mat[234]-index-col-wr
Reviewed-by: Eric Anholt <eric@anholt.net>
And don't delete them. Let ralloc clean them up. Deleting the
temporary IR leaves dangling references in the prog_instruction. That
results in a bad dereference when printing the IR with MESA_GLSL=dump.
NOTE: This is a candidate for the 7.10 and 7.11 branches.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=38584
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Mesa IR actually stores all numbers as floating point, so this is
totally a farce, but we may as well keep it going.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
Mesa already supports this because of NV_fragment_program.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Tested-by: Marek Olšák <maraeo@gmail.com>
There's really no need for a prefix on member functions, and overloading
takes care of the _op1/_op2 distinction quite nicely. Eric already made
a similar change in the i965 FS backend.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Rather than ir_to_mesa_dst_reg_from_src and ir_to_mesa_src_reg_from_dst.
The new constructors are marked 'explicit' so that the compiler can
catch cases where source and destination registers were accidentally
interchanged.
This also necessitated using constructors to initialize the undef and
address registers, as well as adding a default constructor.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Both classes are completely private to ir_to_mesa.cpp, so there won't be
any name conflicts with other parts of Mesa. The prefix simply makes it
harder to read.
Also, use a class rather than typedef structs.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
This is in preparation from removing the "ir_to_mesa_" prefix on the
src_reg and dst_reg types, which would cause a naming conflict.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
The code would previously handle the projection, then swizzle the
shadow comparitor into place. However, when the projection is done
"by hand," as in the TXB case, the unprojected shadow comparitor would
over-write the projected shadow comparitor.
Shadow comparison with projection and LOD is an extremely rare case in
real application code, so it shouldn't matter that we don't handle
that case with the greatest efficiency.
NOTE: This is a candidate for the stable branches.
Reviewed-by: Brian Paul <brianp@vmware.com>
References: https://bugs.freedesktop.org/show_bug.cgi?id=32395
This should be the last bit of infrastructure changes before
generating GLSL IR for assembly shaders.
This commit leaves some odd code formatting in ir_to_mesa and brw_fs.
This was done to minimize whitespace changes / reindentation in some
loops. The following commit will restore formatting sanity.
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Chad Versace <chad.versace@intel.com>
The r300 compiler can eliminate unused uniforms and remap uniform locations
if their number surpasses hardware limits, so the limit is actually
NumParameters + NumUnusedParameters. This is important for some apps
under Wine to run.
Wine sometimes declares a uniform array of 256 vec4's and some Wine-specific
constants on top of that, so in total there is more uniforms than r300 can
handle. This was the main motivation for implementing the elimination
of unused constants.
We should allow drivers to implement fail & recovery paths where it makes
sense, so giving up too early especially when comes to uniforms is not
so good idea, though I agree there should be some hard limit for all drivers.
This patch fixes:
- glsl-fs-uniform-array-5
- glsl-vs-large-uniform-array
on drivers which can eliminate unused uniforms.
Since we're compiling/linking GLSL shaders we should check against
the shader uniform limits, not the legacy vertex/fragment program
parameter limits which are usually lower.
The gl_program_constants struct is for limits that are applicable to
any/all shader stages. Move the geometry shader-only fields into the
gl_constants struct.
Remove redundant MaxGeometryUniformComponents field too.
Addresses excessive TEMP allocation in vertex shaders where all CONSTs are
stored into TEMPS at the start, but copy propagation was failing due to
the presence of IFs.
We could do something about loops, but ifs are easy enough.
