| Commit message (Collapse) | Author | Age | Files | Lines |
|---|
| |
|
|
|
|
|
|
|
|
| |
This lowering isn't needed for RADV because AMDGCN has two
instructions. It will be disabled for RADV in an upcoming series.
While we are at it, factorize a little bit.
Signed-off-by: Samuel Pitoiset <[email protected]>
Reviewed-by: Jason Ekstrand <[email protected]>
|
| |
|
|
|
|
|
|
|
|
| |
Replace done using:
find ./src -type f -exec sed -i -- \
's/glsl_type_is_struct(/glsl_type_is_struct_or_ifc(/g' {} \;
Acked-by: Karol Herbst <[email protected]>
Acked-by: Jason Ekstrand <[email protected]>
Acked-by: Kenneth Graunke <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Not complete, mostly just adding things as I encounter them in CTS. But
not getting far enough yet to hit most of the OpenCL.std instructions.
Anyway, this is better than nothing and covers the most common builtins.
v2: add hadd proof from Jason
move some of the lowering into opt_algebraic and create new nir opcodes
simplify nextafter lowering
fix normalize lowering for inf
rework upsample to use nir_pack_bits
add missing files to build systems
v3: split lines of iadd/sub_sat expressions
Signed-off-by: Karol Herbst <[email protected]>
Reviewed-by: Jason Ekstrand <[email protected]>
|
| |
|
|
|
|
|
|
|
| |
This crops up both in the actual SPIR-V VectorInsert/Extract opcodes as
well as various places where we deal with vector derefs.
Cc: [email protected]
Reviewed-by: Alejandro Piñeiro <[email protected]>
Reviewed-by: Caio Marcelo de Oliveira Filho <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The 16-bit polynomial execution doesn't meet Khronos precision requirements.
Also, the half-float denorm range starts at 2^(-14) and with asin taking input
values in the range [0, 1], polynomial approximations can lead to flushing
relatively easy.
An alternative is to use the atan2 formula to compute asin, which is the
reference taken by Khronos to determine precision requirements, but that
ends up generating too many additional instructions when compared to the
polynomial approximation. Specifically, for the Intel case, doing this
adds +41 instructions to the program for each asin/acos call, which looks
like an undesirable trade off.
So for now we take the easy way out and fallback to using the 32-bit
polynomial approximation, which is better (faster) than the 16-bit atan2
implementation and gives us better precision that matches Khronos
requirements.
v2:
- Fallback to 32-bit using recursion (Jason).
Reviewed-by: Jason Ekstrand <[email protected]>
|
| |
|
|
| |
Reviewed-by: Jason Ekstrand <[email protected]>
|
| |
|
|
|
|
|
|
|
|
| |
v2:
- use nir_fadd_imm and nir_fmul_imm helpers (Jason)
v3:
- since we need to define one for fsub use it for fdiv too (Jason)
Reviewed-by: Jason Ekstrand <[email protected]>
|
| |
|
|
|
|
|
| |
v2
- use nir_fmul_imm helper (Jason)
Reviewed-by: Jason Ekstrand <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
| |
v2:
- fix huge_val for 16-bit, it was mean't to be 2^14 not 10^14.
v3:
- rebase on top of new bool sized opcodes
- use nir_b2f helper
- use nir_fmul_imm helper
Reviewed-by: Jason Ekstrand <[email protected]>
|
| |
|
|
|
|
|
|
|
| |
v2:
- use nir_fadd_imm and nir_fmul_imm helpers (Jason)
- rebased on top of new sized boolean opcodes
- use nir_b2f helper
Reviewed-by: Jason Ekstrand <[email protected]>
|
| |
|
|
| |
Reviewed-by: Jason Ekstrand <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
| |
v2:
- use nir_fmul_imm and nir_fadd_imm helpers (Jason)
v3:
- missed one case where we need to replace nir_imm_float
with nir_imm_floatN_t (Jason)
Reviewed-by: Jason Ekstrand <[email protected]>
|
| |
|
|
|
|
|
|
|
| |
Instead of a single i2b and b2i, we now have i2b32 and b2iN where N is
one if 8, 16, 32, or 64. This leads to having a few more opcodes but
now everything is consistent and booleans aren't a weird special case
anymore.
Reviewed-by: Connor Abbott <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
| |
also move some of the GLSL builtins over we will need for implementing
some OpenCL builtins
v2: replace NIR_IMM_FP by nir_imm_floatN_t in ported code
fix up changes caused by swizzle rework
Reviewed-by: Jason Ekstrand <[email protected]>
Signed-off-by: Karol Herbst <[email protected]>
|
| |
|
|
|
|
|
|
|
|
| |
There are no fixed sized array arguments in C, those are simply pointers
to unsized arrays and as the size is passed in anyway, just rely on that.
where possible calls are replaced by nir_channel and nir_channels.
Reviewed-by: Jason Ekstrand <[email protected]>
Signed-off-by: Karol Herbst <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
If the glsl is something like this:
in vec4 some_input;
interpolateAtCentroid(some_input[idx])
then it now gets generated as if it were:
interpolateAtCentroid(some_input)[idx]
This is necessary because the index will get generated as a series of
nir_bcsel instructions so it would no longer be an input variable. It
is similar to what is done for GLSL in ca63a5ed3e9efb2bd645b42.
Although I can’t find anything explicit in the Vulkan specs to say
this should be allowed, the SPIR-V spec just says “the operand
interpolant must be a pointer to the Input Storage Class”, which I
guess doesn’t rule out any type of pointer to an input.
This was found using the spec/glsl-4.40/execution/fs-interpolateAt*
Piglit tests with the ARB_gl_spirv branch.
Signed-off-by: Neil Roberts <[email protected]>
Signed-off-by: Alejandro Piñeiro <[email protected]>
v2: update after nir_deref_instr land on master. Implemented by
Alejandro Piñeiro. Special thanks to Jason Ekstrand for guidance
at the new nir_deref_instr world.
Reviewed-by: Jason Ekstrand <[email protected]>
|
| |
|
|
|
|
|
|
|
| |
The only thing still using old-school drefs are function calls.
Acked-by: Rob Clark <[email protected]>
Acked-by: Bas Nieuwenhuizen <[email protected]>
Acked-by: Dave Airlie <[email protected]>
Reviewed-by: Kenneth Graunke <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The SPIR-V spec doesn’t specify a size requirement for these and the
equivalent functions in the GLSL spec have explicit alternatives for
doubles. Refract is a little bit more complicated due to the fact that
the final argument is always supposed to be a scalar 32- or 16- bit
float regardless of the other operands. However in practice it seems
there is a bug in glslang that makes it convert the argument to 64-bit
if you actually try to pass it a 32-bit value while the other
arguments are 64-bit. This adds an optional conversion of the final
argument in order to support any type.
These have been tested against the automatically generated tests of
glsl-4.00/execution/built-in-functions using the ARB_gl_spirv branch
which tests it with quite a large range of combinations.
The issue with glslang has been filed here:
https://github.com/KhronosGroup/glslang/issues/1279
v2: Convert the eta operand of Refract from any size in order to make
it eventually cope with 16-bit floats.
Reviewed-by: Jason Ekstrand <[email protected]>
|
| |
|
|
|
|
|
|
|
|
| |
There is an existing macro that is used to choose between either a
float or a double immediate constant based on the bit size of the
first operand to the builtin. This is now changed to use the new
nir_imm_floatN_t helper function to reduce the number of places that
make this decision.
Reviewed-by: Jason Ekstrand <[email protected]>
|
| |
|
|
|
|
|
|
|
|
| |
The MSVC compiler warns when the function parameter types don't
exactly match with respect to enum vs. uint32_t. Use SpvOp everywhere.
Alternately, uint32_t could be used everywhere. There doesn't seem
to be an advantage to one over the other.
Reviewed-by: Neil Roberts <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Previously bitset.h would include u_math.h to get bitscan.h. u_math.h
lives in src/gallium/auxiliary/util while both bitset.h and bitscan.h
live in src/util. Having the one file directly include another file
that lives in the same directory makes much more sense.
As a side-effect, several files need to directly include standard header
files that were previously indirectly included.
v2: Fix build break in src/amd/common/ac_nir_to_llvm.c.
Signed-off-by: Ian Romanick <[email protected]>
Reviewed-by: Eduardo Lima Mitev <[email protected]>
|
| |
|
|
|
|
|
|
|
|
| |
The implementation is inspired by
lower_instructions_visitor::dfrexp_sig_to_arith.
This has been tested against the arb_gpu_shader_fp64/fs-frexp-dvec4
test using the ARB_gl_spirv branch.
Reviewed-by: Jason Ekstrand <[email protected]>
|
| |
|
|
|
| |
Reviewed-by: Tapani Pälli <[email protected]>
Reviewed-by: Ian Romanick <[email protected]>
|
| |
|
|
|
| |
Reviewed-by: Tapani Pälli <[email protected]>
Reviewed-by: Ian Romanick <[email protected]>
|
| |
|
|
| |
Reviewed-by: Connor Abbott <[email protected]>
|
| |
|
|
|
|
| |
v2: NIR fmax/fmin already handles NaN (Connor).
Reviewed by: Elie Tournier <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
| |
According to GLSL.std.450 spec, SmoothStep expects input to be a
floating-point type, but it does not restrict the bitsize.
Current implementation relies on inputs to be 32-bit.
This commit extends the support to 64-bit size inputs.
Reviewed by: Elie Tournier <[email protected]>
|
| |
|
|
|
|
|
| |
NIR is a typeless IR and the two opcodes, when considered bitwise, do
exactly the same thing. There's no reason to have two versions.
Reviewed-by: Kenneth Graunke <[email protected]>
|
| |
|
|
|
| |
Reviewed-by: Ian Romanick <[email protected]>
Reviewed-by: Juan A. Suarez Romero <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
of zero/infinity.
See "glsl: Rewrite atan2 implementation to fix accuracy and handling
of zero/infinity." for the rationale, but note that the instruction
count benefit discussed there is somewhat less important for the SPIRV
implementation, because the current code already emitted no control
flow instructions -- Still this saves us one hardware instruction per
scalar component on Intel SKL hardware.
Fixes the following Vulkan CTS tests on Intel hardware:
dEQP-VK.glsl.builtin.precision.atan2.highp_compute.scalar
dEQP-VK.glsl.builtin.precision.atan2.highp_compute.vec2
dEQP-VK.glsl.builtin.precision.atan2.highp_compute.vec3
dEQP-VK.glsl.builtin.precision.atan2.highp_compute.vec4
dEQP-VK.glsl.builtin.precision.atan2.mediump_compute.vec2
dEQP-VK.glsl.builtin.precision.atan2.mediump_compute.vec4
Note that most of the test-cases above expect IEEE-compliant handling
of atan2(±∞, ±∞), which this patch doesn't explicitly handle, so
except for the last two the test-cases above weren't expected to pass
yet. The reason they do is that the i965 back-end implementation of
the NIR fmin and fmax instructions is not quite GLSL-compliant (it
complies with IEEE 754 recommendations though), because fmin/fmax of a
NaN and a non-NaN argument currently always return the non-NaN
argument, which causes atan() to flush NaN to one and return the
expected value. The front-end should probably not be relying on this
behavior for correctness though because other back-ends are likely to
behave differently -- A follow-up patch will handle the atan2(±∞, ±∞)
corner cases explicitly.
v2: Fix up argument scaling to take into account the range and
precision of exotic FP24 hardware. Flip coordinate system for
arguments along the vertical line as if they were on the left
half-plane in order to avoid division by zero which may give
unspecified results on non-GLSL 4.1-capable hardware. Sprinkle in
some more comments.
Reviewed-by: Ian Romanick <[email protected]>
|
| |
|
|
|
| |
Signed-off-by: Samuel Iglesias Gonsálvez <[email protected]>
Reviewed-by: Jason Ekstrand <[email protected]>
|
| |
|
|
|
|
|
|
|
| |
We rename it to nir_deref_clone, re-order the sources to match the other
clone functions, and expose nir_deref_var_clone. This past part, in
particular, lets us get rid of quite a few lines since we no longer have
to call nir_copy_deref and wrap it in deref_as_var.
Reviewed-by: Jordan Justen <[email protected]>
|
| |
|
|
|
|
|
|
|
|
| |
The new implementation is more correct because it clamps the incoming value
to 10 to avoid floating-point overflow. It also uses a much reduced
version of the formula which only requires 1 exp() rather than 2. This
fixes all of the dEQP-VK.glsl.builtin.precision.tanh.* tests.
Reviewed-by: Kenneth Graunke <[email protected]>
Cc: "13.0" <[email protected]>
|
| |
|
|
|
| |
Signed-off-by: Jason Ekstrand <[email protected]>
Reviewed-by: Dave Airlie <[email protected]>
|
|
|
While it does rely on NIR, it's not really part of the NIR core. At the
moment, it still builds as part of libnir but that can be changed later if
desired.
|
