mesa.git
8 years agoutil: Add ATTRIBUTE_RETURNS_NONNULL.
Matt Turner [Fri, 13 May 2016 20:17:02 +0000 (13:17 -0700)]
util: Add ATTRIBUTE_RETURNS_NONNULL.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
8 years agoclover: grid_offset should be padded with 0 not 1
Jan Vesely [Mon, 16 May 2016 00:08:09 +0000 (20:08 -0400)]
clover: grid_offset should be padded with 0 not 1

Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
8 years agoi965: Expose OpenGL 4.0 for gen8+
Iago Toral Quiroga [Tue, 23 Feb 2016 11:15:36 +0000 (12:15 +0100)]
i965: Expose OpenGL 4.0 for gen8+

ARB_gpu_shader_fp64 was the only feature missing.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
8 years agodocs: Mark ARB_gpu_shader_fp64 as done for i965/gen8+
Iago Toral Quiroga [Tue, 23 Feb 2016 11:14:11 +0000 (12:14 +0100)]
docs: Mark ARB_gpu_shader_fp64 as done for i965/gen8+

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
8 years agoi965: Enable ARB_gpu_shader_fp64 for gen8+
Iago Toral Quiroga [Tue, 23 Feb 2016 11:14:52 +0000 (12:14 +0100)]
i965: Enable ARB_gpu_shader_fp64 for gen8+

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
8 years agoi965/tes/scalar: Fix load input for doubles
Iago Toral Quiroga [Mon, 9 May 2016 07:43:56 +0000 (09:43 +0200)]
i965/tes/scalar: Fix load input for doubles

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
8 years agoi965/tcs/scalar: fix store output for doubles
Iago Toral Quiroga [Mon, 9 May 2016 13:23:34 +0000 (15:23 +0200)]
i965/tcs/scalar: fix store output for doubles

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
8 years agoi965/tcs/scalar: fix load input for doubles
Iago Toral Quiroga [Mon, 9 May 2016 08:31:50 +0000 (10:31 +0200)]
i965/tcs/scalar: fix load input for doubles

v2: do not write to the original indirect_offset since that is
    an expression that could be used somewhere else (Ken)

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
8 years agoi965/fs: fix nir_intrinsic_store_output for doubles
Iago Toral Quiroga [Mon, 9 May 2016 08:14:48 +0000 (10:14 +0200)]
i965/fs: fix nir_intrinsic_store_output for doubles

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
8 years agoi965/fs: fix number of output components for doubles
Iago Toral Quiroga [Mon, 9 May 2016 08:14:18 +0000 (10:14 +0200)]
i965/fs: fix number of output components for doubles

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
8 years agoi965/vec4: handle doubles in type_size_vec4()
Iago Toral Quiroga [Mon, 9 May 2016 13:21:25 +0000 (15:21 +0200)]
i965/vec4: handle doubles in type_size_vec4()

The scalar backend uses this to check URB input sizes.

v2: Removed redundant break after return (Curro)

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
8 years agoi965/fs: support doubles with shared variable stores
Iago Toral Quiroga [Tue, 26 Jan 2016 09:30:39 +0000 (10:30 +0100)]
i965/fs: support doubles with shared variable stores

This is pretty much the same we do with SSBOs.

v2: do not shuffle in-place, it is not safe since the original 64-bit data
    could be used after the write, instead use a temporary like we do
    for SSBO stores (Iago)

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
8 years agoi965/fs: support doubles with ssbo stores
Iago Toral Quiroga [Mon, 25 Jan 2016 12:42:19 +0000 (13:42 +0100)]
i965/fs: support doubles with ssbo stores

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
8 years agoi965/fs: add shuffle_64bit_data_for_32bit_write helper
Iago Toral Quiroga [Mon, 25 Jan 2016 12:37:50 +0000 (13:37 +0100)]
i965/fs: add shuffle_64bit_data_for_32bit_write helper

This does the inverse operation of shuffle_32bit_load_result_to_64bit_data
and we will use it when we need to write 64-bit data in the layout expected
by untyped write messages.

v2 (curro):
- Use subscript() instead of stride()
- Assert on the input types rather than silently retyping.
- Use offset() instead of horiz_offset(), drop the multiplier definition.
- Drop the temporary vgrf and force_writemask_all.
- Make component_i const.
- Move to brw_fs_nir.cpp

v3 (curro):
- Pass dst and src by reference.
- Simplify allocation of tmp register.
- Move to brw_fs_nir.cpp.
- Get rid of the temporary.

v3 (Iago):
- Check that the src and dst regions do not overlap, since that would
  typically be a bug in the caller.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
8 years agoi965/fs: support doubles with SSBO loads
Iago Toral Quiroga [Mon, 25 Jan 2016 09:58:59 +0000 (10:58 +0100)]
i965/fs: support doubles with SSBO loads

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
8 years agoi965/fs: support doubles with shared variable loads
Iago Toral Quiroga [Thu, 5 May 2016 10:58:19 +0000 (12:58 +0200)]
i965/fs: support doubles with shared variable loads

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
8 years agoi965/fs: Add do_untyped_vector_read helper
Iago Toral Quiroga [Thu, 5 May 2016 10:55:44 +0000 (12:55 +0200)]
i965/fs: Add do_untyped_vector_read helper

We are going to need the same logic for anything that reads
doubles via untyped messages (CS shared variables and SSBOs). Add a
helper function with that logic so that we can reuse it.

v2:
- Make this a static function instead of a method of fs_visitor (Iago)
- We only support types with a size of 4 or 8 (Curro)
- Avoid retypes by using a separate vgrf for the packed result (Curro)
- Put dst parameter before source parameters (Curro)

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
8 years agoi965/fs: support doubles with UBO loads
Iago Toral Quiroga [Wed, 13 Jan 2016 09:17:10 +0000 (10:17 +0100)]
i965/fs: support doubles with UBO loads

UBO loads with constant offset use the UNIFORM_PULL_CONSTANT_LOAD
instruction, which reads 16 bytes (a vec4) of data from memory. For dvec
types this only provides components x and y. Thus, if we are reading
more than 2 components we need to issue a second load at offset+16 to
read the next 16-byte chunk with components w and z.

UBO loads with non-constant offset emit a load for each component
in the vector (and rely in CSE to fix redundant loads), so we only
need to consider the size of the data type when computing the offset
of each element in a vector.

v2 (Sam):
- Adapt the code to use component() (Curro).

v3 (Sam):
- Use type_sz(dest.type) in VARYING_PULL_CONSTANT_LOAD() call (Curro).
- Add asserts to ensure std140 vector alignment rules are followed
  (Curro).

Signed-off-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
8 years agoi965/fs: fix pull constant load component selection for doubles
Iago Toral Quiroga [Mon, 18 Jan 2016 12:09:31 +0000 (13:09 +0100)]
i965/fs: fix pull constant load component selection for doubles

UNIFORM_PULL_CONSTANT_LOAD is used to load a contiguous vec4 starting at a
constant offset that is 16-byte aligned. If we need to access an unaligned
offset we emit a load with an aligned offset and use the remaining constant
offset to select the component into the vec4 result that we are interested
in. This component must be computed in units of the type size, since that
is what fs_reg::set_smear expects.

This patch does this change in the two places where we use this message:
In demote_pull_constants when we lower uniform access with constant offset
into the pull constant buffer and in UBO loads with constant offset.

v2 (Sam):
- Fix set_smear() in fs_visitor::lower_constant_loads(), take into account
source type instead and remove MAX2 (Curro).
- Improve changes to nir_intrinsic_load_ubo case in nir_emit_intrinsic()
(Curro).

Signed-off-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
8 years agoi965/fs: Fix and document component().
Francisco Jerez [Wed, 11 May 2016 19:54:26 +0000 (12:54 -0700)]
i965/fs: Fix and document component().

This fixes a number of bugs of component() by reimplementing it in
terms of horiz_offset(): Handling of base registers starting at a
non-zero subreg_offset, handling of strided registers and overflow of
subreg_offset into reg_offset.

Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
8 years agoi965/fs: Fix fs_visitor::VARYING_PULL_CONSTANT_LOAD for doubles
Iago Toral Quiroga [Thu, 14 Jan 2016 07:55:28 +0000 (08:55 +0100)]
i965/fs: Fix fs_visitor::VARYING_PULL_CONSTANT_LOAD for doubles

v2 (Curro):
   - Assert on scale == 1 when shuffling 64-bit data.
   - Remove type_slots, use type_sz(vec4_result.type) instead.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
8 years agoi965/fs: add shuffle_32bit_load_result_to_64bit_data helper
Iago Toral Quiroga [Fri, 22 Jan 2016 13:00:38 +0000 (14:00 +0100)]
i965/fs: add shuffle_32bit_load_result_to_64bit_data helper

There will be a few places where we need to shuffle the result of a 32-bit
load into valid 64-bit data, so extract this logic into a separate helper
that we can reuse.

v2 (Curro):
- Use subscript() instead of stride()
- Assert on the input types rather than retyping.
- Use offset() instead of horiz_offset(), drop the multiplier definition.
- Don't use  force_writemask_all.
- Mark component_i as const.
- Make the function name lower case.

v3 (Curro):
- Pass src and dst by reference.
- Move to brw_fs_nir.cpp

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
8 years agoi965/fs: Stop using the LOAD_PAYLOAD instruction in lower_simd_width.
Francisco Jerez [Wed, 4 May 2016 04:26:13 +0000 (21:26 -0700)]
i965/fs: Stop using the LOAD_PAYLOAD instruction in lower_simd_width.

Instead of using the LOAD_PAYLOAD instruction (emitted through the
emit_transpose() helper that is no longer useful and this commit
removes) which had to be marked force_writemask_all in some cases,
emit a series of moves to apply proper channel enable signals to the
destination.  Until now lower_simd_width() had mainly been used to
lower things that invariably had a basic block-local temporary as
destination so it didn't seem like a big deal, but I found it to be
the reason for several Piglit regressions in my SIMD32 branch and
Igalia discovered the same issue independently while working on FP64
support.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
8 years agoi965/fs: fix copy/constant propagation regioning checks
Iago Toral Quiroga [Fri, 11 Mar 2016 07:46:36 +0000 (08:46 +0100)]
i965/fs: fix copy/constant propagation regioning checks

We were not accounting for subreg_offset in the check for the start
of the region.

Also, fs_reg::regs_read() already takes the stride into account, so we
should not multiply its result by the stride again. This was making
copy-propagation fail to copy-propagate cases that would otherwise be
safe to copy-propagate. Again, this was observed in fp64 code, since
there we use stride > 1 often.

v2 (Sam):
- Rename function and add comment (Jason, Curro).
- Assert that register files and number are the same (Jason).
- Fix code to take into account the assumption that src.subreg_offset
is strictly less than the reg_offset unit (Curro).
- Don't pass the registers by value to the function, use
'const fs_reg &' instead (Curro).
- Remove obsolete comment in the commit log (Curro).

v3 (Sam):
- Remove the assert and put the condition in the return (Curro).
- Fix function name (Curro).

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
8 years agoi965/fs: fix copy propagation from load payload
Iago Toral Quiroga [Tue, 12 Apr 2016 11:53:24 +0000 (13:53 +0200)]
i965/fs: fix copy propagation from load payload

We were not considering the case where the load payload is writing to
a destination with a reg_offset > 0.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
8 years agoi965/fs: fix copy propagation of partially invalidated entries
Iago Toral Quiroga [Fri, 11 Mar 2016 13:35:07 +0000 (14:35 +0100)]
i965/fs: fix copy propagation of partially invalidated entries

We were not invalidating entries with a src that reads more than one register
when we find writes that overwrite any register read by entry->src after
the first. This leads to incorrect copy propagation because we re-use
entries from the ACP that have been partially invalidated. Same thing for
entries with a dst that writes to more than one register.

v2 (Sam):
- Improve code by defining regions_overlap() and using it instead of a
loop (Curro).

Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
8 years agoi965/fs: Reindent register offset calculation of try_copy_propagate().
Francisco Jerez [Tue, 10 May 2016 23:03:36 +0000 (16:03 -0700)]
i965/fs: Reindent register offset calculation of try_copy_propagate().

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
8 years agoi965/fs: Simplify and fix register offset calculation of try_copy_propagate().
Francisco Jerez [Tue, 10 May 2016 23:01:56 +0000 (16:01 -0700)]
i965/fs: Simplify and fix register offset calculation of try_copy_propagate().

try_copy_propagate() was special-casing UNIFORM registers (the
BAD_FILE, ARF and FIXED_GRF cases are dead, see the assertion at the
top of the function) and then failing to take into account the
possibility of the instruction reading from a non-zero offset of the
destination of the copy.  The VGRF/ATTR handling takes it into account
correctly, and there is no reason we couldn't use the exact same logic
for the UNIFORM file aside from the fact that uniforms represent
reg_offset in different units.  We can work around that easily by
defining an additional constant with the right unit reg_offset is
expressed in.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
8 years agoi965/fs: disallow type change in copy-propagation if types have different sizes
Iago Toral Quiroga [Wed, 23 Mar 2016 11:02:21 +0000 (12:02 +0100)]
i965/fs: disallow type change in copy-propagation if types have different sizes

Because the semantics of source modifiers are type-dependent, the type of the
original source of the copy must be kept unmodified while propagating it into
some instruction, which implies that we need to have the guarantee that the
meaning of the instruction is going to remain the same after we have changed
the types. Whenthe size of the new type is different from the size of the old
type the new and old instructions cannot possibly be equivalent because the new
instruction will be reading more data than the old one was.

Prevents that we turn this:

load_payload(8) vgrf17:DF, |vgrf4+0.0|:DF 1sthalf
mov(8) vgrf18:DF, vgrf17:DF 1sthalf
load_payload(8) vgrf5:DF, vgrf18:DF, vgrf20:DF NoMask 1sthalf WE_all
load_payload(8) vgrf21:UD, vgrf5+0.4<2>:UD 1sthalf
mov(8) vgrf22:UD, vgrf21:UD 1sthalf

into:

load_payload(8) vgrf17:DF, |vgrf4+0.0|:DF 1sthalf
mov(8) vgrf18:DF, |vgrf4+0.0|:DF 1sthalf
load_payload(8) vgrf5:DF, |vgrf4+0.0|:DF, |vgrf4+2.0|:DF NoMask 1sthalf WE_all
load_payload(8) vgrf21:UD, vgrf5+0.4<2>:UD 1sthalf
mov(8) vgrf22:DF, |vgrf4+0.4|<2>:DF 1sthalf

where the semantics of the last instruccion have changed.

v2 (Curro):
  - Update commit log and add comment to explain the problem better.
  - Simplify the condition.

Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
8 years agoi965/fs: Fix copy propagation of load payload for double operands
Iago Toral Quiroga [Mon, 18 Jan 2016 09:44:20 +0000 (10:44 +0100)]
i965/fs: Fix copy propagation of load payload for double operands

Specifically, consider the size of the data type of the operand to compute
the number of registers written.

v2 (Sam):
- Fix line width (Jordan).
- Add an assert (Jordan).
- Use REG_SIZE in the calculation of regs_written (Curro)

v3 (Sam):
- Fix assert and calculation of regs_written (Curro).

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
8 years agoi965/fs: Fix propagation of copies with strided source.
Francisco Jerez [Mon, 25 Apr 2016 22:40:05 +0000 (15:40 -0700)]
i965/fs: Fix propagation of copies with strided source.

This has likely been broken since we started propagating copies not
matching the offset of the instruction exactly
(1728e74957a62b1b4b9fbb62a7de2c12b77c8a75).  The copy source stride
needs to be taken into account to find out the offset at the origin
that corresponds to the offset at the destination of the copy which is
being read by the instruction.  This has led to program miscompilation
on both my SIMD32 branch and Igalia's FP64 branch.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
8 years agoi965/fs: fix subreg_offset overflow in byte_offset()
Iago Toral Quiroga [Wed, 6 Apr 2016 08:27:14 +0000 (10:27 +0200)]
i965/fs: fix subreg_offset overflow in byte_offset()

This can happen if the register already has a non-zero subreg_offset
when byte_offset() is called.

v2 (Sam):
- Refactor byte_offset() (Jordan).

Signed-off-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
8 years agoi965: Fix JIP to skip over sibling do...while loops.
Kenneth Graunke [Sun, 15 May 2016 06:54:48 +0000 (23:54 -0700)]
i965: Fix JIP to skip over sibling do...while loops.

We've apparently always been botching JIP for sequences such as:

do
    cmp.f0.0 ...
    (+f0.0) break
    ...
    do
        ...
    while
    ...
while

Because the "do" instruction doesn't actually exist, the inner "while"
is at the same depth as the "break".  brw_find_next_block_end() thus
mistook the inner "while" as the end of the loop containing the "break",
and set the "break" to point to the wrong place.

Only "while" instructions that jump before our instruction are relevant.
We need to ignore the rest, as they're sibling control flow nodes (or
children, but this was already handled by the depth == 0 check).

See also commit 1ac1581f3889d5f7e6e231c05651f44fbd80f0b6.

This prevents channel masks from being screwed up, and fixes GPU
hangs(*) in dEQP-GLES31.functional.shaders.multisample_interpolation.
interpolate_at_sample.centroid_qualified.multisample_texture_16.

The test ended up executing code with no channels enabled, and that
code contained FIND_LIVE_CHANNEL, which returned 8 (out of range for
a SIMD8 program), which then was used in indirect GRF addressing,
which randomly got a boolean value (0xFFFFFFFF), interpreted it as
a sample ID, OR'd it into an indirect send message descriptor,
which corrupted the message length, sending a pixel interpolator
message with mlen 15, which is illegal.  Whew :)

(*) Technically, the test doesn't GPU hang currently, but only
    because another bug prevents it from issuing pixel interpolator
    messages entirely...with that fixed, it hangs.

Cc: mesa-stable@lists.freedesktop.org
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
8 years agoi965: Make a "does this while jump before our instruction?" helper.
Kenneth Graunke [Sun, 15 May 2016 06:53:19 +0000 (23:53 -0700)]
i965: Make a "does this while jump before our instruction?" helper.

I need to use this in an additional place.

Cc: mesa-stable@lists.freedesktop.org
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
8 years agoi965: Send the minimal number of STATE_BASE_ADDRESS packets.
Kenneth Graunke [Wed, 27 Apr 2016 16:35:03 +0000 (09:35 -0700)]
i965: Send the minimal number of STATE_BASE_ADDRESS packets.

STATE_BASE_ADDRESS stalls the whole pipeline, and the documentation
cautions us to emit it as little as possible for better performance.

We recently put some hacks in BLORP to try and avoid emitting it
if it was already set correctly.  However, this wasn't quite minimal:
if BLORP is the first operation (i.e. glClear()), then it would emit
it, and subsequent draw calls would emit it again.

This caused a small drop in performance in GPUTest Triangle when
switching from Meta to BLORP.

Unlike most packets, STATE_BASE_ADDRESS isn't influenced by GL state:
it needs to be emitted once per batch, before most other commands, or
whenever we change the program cache BO.  It's also valid in both the
3D and compute pipelines, which makes it even more unique.

This patch removes it from the atom mechanism and instead directly
calls it as part of every draw, compute dispatch, or BLORP operation.
We introduce a new flag indicating that STATE_BASE_ADDRESS has already
been emitted this batch, and if so, skip doing it again.  When we make
a new program cache BO, we simply reset the flag, so the next operation
will emit it again.  When we flush/reset the batch, we reset the flag.

This guarantees that we'll emit STATE_BASE_ADDRESS only when we have to.
It's also less code than the old atom mechanism.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
8 years agoi965: Combine Gen4-7 and Gen8+ state base address emitters.
Kenneth Graunke [Wed, 27 Apr 2016 16:30:34 +0000 (09:30 -0700)]
i965: Combine Gen4-7 and Gen8+ state base address emitters.

We're about to start calling it directly, and this means the callers
won't have to think about generations.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
8 years agoi965: Move Gen4-5 programs to brw_upload_programs() too.
Kenneth Graunke [Wed, 27 Apr 2016 18:25:26 +0000 (11:25 -0700)]
i965: Move Gen4-5 programs to brw_upload_programs() too.

This way all the programs are in one place again, and it also should
make some future STATE_BASE_ADDRESS related changes possible.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
8 years agoi965: Mark brw const in brw_state_dirty and callers.
Kenneth Graunke [Wed, 27 Apr 2016 18:26:08 +0000 (11:26 -0700)]
i965: Mark brw const in brw_state_dirty and callers.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
8 years agoglsl: Don't do constant propagation in opt_constant_folding.
Kenneth Graunke [Fri, 29 Apr 2016 20:13:01 +0000 (13:13 -0700)]
glsl: Don't do constant propagation in opt_constant_folding.

opt_constant_folding is supposed to fold trees of constants into a
single constant.  Surprisingly, it was also propagating constant values
from variables into expression trees - even when the result couldn't be
folded together.  This is opt_constant_propagation's job.

The ir_dereference_variable::constant_expression_value() method returns
a clone of var->constant_value.  So we would replace the dereference
with a constant, propagating it into the tree.

Skip over ir_dereference_variable to avoid this surprising behavior.
However, add code to explicitly continue doing it in the constant
propagation pass, as it's useful to do so.

shader-db statistics on Broadwell:

total instructions in shared programs: 8905349 -> 8905126 (-0.00%)
instructions in affected programs: 30100 -> 29877 (-0.74%)
helped: 93
HURT: 20

total cycles in shared programs: 71017030 -> 71015944 (-0.00%)
cycles in affected programs: 132456 -> 131370 (-0.82%)
helped: 54
HURT: 45

The only hurt programs are by a single instruction, while the helped
ones are helped by 1-4 instructions.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
8 years agoglsl: Avoid excess tree walking when folding ir_dereference_arrays.
Kenneth Graunke [Fri, 29 Apr 2016 20:19:33 +0000 (13:19 -0700)]
glsl: Avoid excess tree walking when folding ir_dereference_arrays.

If an ir_dereference_array has non-constant components, there's no
point in trying to evaluate its value (which involves walking down
the tree and possibly allocating memory for portions of the subtree
which are constant).

This also removes convoluted tree walking in opt_constant_folding(),
which tries to fold constants while walking up the tree.  No need to
walk down, then up, then down again.

We did this for swizzles and expressions already, but I was lazy
back in the day and didn't do this for ir_dereference_array.

No change in shader-db.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
8 years agoglsl: Consolidate duplicate copies of constant folding.
Kenneth Graunke [Fri, 29 Apr 2016 21:40:26 +0000 (14:40 -0700)]
glsl: Consolidate duplicate copies of constant folding.

We could probably clean this up more (maybe make it a method), but at
least there's only one copy of this code now, and that's a start.

No change in shader-db.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
8 years agoglsl: Remove bonus tree walking in opt_constant_folding().
Kenneth Graunke [Fri, 29 Apr 2016 21:04:17 +0000 (14:04 -0700)]
glsl: Remove bonus tree walking in opt_constant_folding().

It looks like this was missed when converting opt_constant_folding()
from a hierarchical visitor to an rvalue visitor in 6606fde3.

ir_rvalue_visitor already processes values on the way back up the tree,
so we will have already visited every child node.  There's no point in
doing it again.

No change in shader-db.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
8 years agoglsl: Make opt_constant_variable() bail in useless cases.
Kenneth Graunke [Fri, 29 Apr 2016 19:53:03 +0000 (12:53 -0700)]
glsl: Make opt_constant_variable() bail in useless cases.

The pass ultimately skips over any entries with assignment_count != 1,
so there's no need to do further work once we've determined that there
are multiple assignments.

The constant value could be a large array (i.e. uvec4[327]), at which
point skipping the constant_expression_value() call (and the clone()
call within) can save us piles of memory.

No change in shader-db.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
8 years agoi965: Flip interpolateAtOffset's y offset when necessary.
Kenneth Graunke [Fri, 13 May 2016 01:53:23 +0000 (18:53 -0700)]
i965: Flip interpolateAtOffset's y offset when necessary.

Fixes 4 dEQP-GLES31.functional.shaders.multisample_interpolation tests:
- interpolate_at_offset.no_qualifiers.default_framebuffer
- interpolate_at_offset.centroid_qualifier.default_framebuffer
- interpolate_at_offset.sample_qualifier.default_framebuffer
- interpolate_at_offset.array_element.default_framebuffer

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
8 years agonir: Add a nir->info.uses_interp_var_at_offset flag.
Kenneth Graunke [Fri, 13 May 2016 06:53:13 +0000 (23:53 -0700)]
nir: Add a nir->info.uses_interp_var_at_offset flag.

I've added this to nir_gather_info(), but also to glsl_to_nir() as a
temporary measure, since the i965 GL driver today doesn't use
nir_gather_info() yet.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
8 years agoglsl: Drop bad ASSERT_TRUE in gl_CullDistance link_varyings test.
Kenneth Graunke [Sat, 14 May 2016 02:26:37 +0000 (19:26 -0700)]
glsl: Drop bad ASSERT_TRUE in gl_CullDistance link_varyings test.

I don't know what the intention was here, but this function returns
void.  We can't assert anything about its return value.

Fixes "make check" failures.

v2: Also fix prototype for the function (caught by Jordan).

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
8 years agoclover: Handle PIPE_SHADER_IR_NIR in switch
Jan Vesely [Thu, 12 May 2016 20:24:58 +0000 (16:24 -0400)]
clover: Handle PIPE_SHADER_IR_NIR in switch

Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
8 years agofreedreno/ir3: small standalone compiler cleanup
Rob Clark [Sat, 14 May 2016 17:39:57 +0000 (13:39 -0400)]
freedreno/ir3: small standalone compiler cleanup

Don't hard-code the gpu-id anymore.

Signed-off-by: Rob Clark <robclark@freedesktop.org>
8 years agonir: forward-declare 'struct gl_shader_program'
Rob Clark [Sat, 14 May 2016 17:35:54 +0000 (13:35 -0400)]
nir: forward-declare 'struct gl_shader_program'

Drop extra #include which is otherwise unneeded (and makes this header
difficult to include from outside of src/mesa).

Signed-off-by: Rob Clark <robclark@freedesktop.org>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
8 years agonir: return progress from lower_idiv
Rob Clark [Mon, 9 May 2016 16:36:03 +0000 (12:36 -0400)]
nir: return progress from lower_idiv

With algebraic-opt support for lowering div to shift, the driver would
like to be able to run this pass *after* the main opt-loop, and then
conditionally re-run the opt-loop if this pass actually lowered some-
thing.

Signed-off-by: Rob Clark <robclark@freedesktop.org>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
8 years agofreedreno/ir3: lower fdiv
Rob Clark [Sat, 14 May 2016 17:40:48 +0000 (13:40 -0400)]
freedreno/ir3: lower fdiv

Not sure how we didn't hit this already, but since we want fdiv
converted into mul + rcp, we should set this.

Signed-off-by: Rob Clark <robclark@freedesktop.org>
8 years agofreedreno/ir3: handle VARYING_SLOT_PNTC
Rob Clark [Mon, 21 Mar 2016 17:56:04 +0000 (13:56 -0400)]
freedreno/ir3: handle VARYING_SLOT_PNTC

In the glsl->tgsi path, this already gets translated to VAR8, which
matches up with rasterizer->sprite_coord_enable.

Signed-off-by: Rob Clark <robclark@freedesktop.org>
8 years agofreedreno/ir3: disable TGSI specific hacks in nir case
Rob Clark [Mon, 21 Mar 2016 15:42:04 +0000 (11:42 -0400)]
freedreno/ir3: disable TGSI specific hacks in nir case

When we got NIR directly from state tracker (vs using tgsi_to_nir) we
need to realize this and skip some TGSI specific hacks.

Signed-off-by: Rob Clark <robclark@freedesktop.org>
8 years agofreedreno/ir3: add support for NIR as preferred IR
Rob Clark [Mon, 28 Mar 2016 14:28:29 +0000 (10:28 -0400)]
freedreno/ir3: add support for NIR as preferred IR

For now under debug flag, since only suitable for debugging/testing.

Signed-off-by: Rob Clark <robclark@freedesktop.org>
8 years agonir: fix comment typo about f2d/d2f
Rob Clark [Fri, 13 May 2016 19:26:15 +0000 (15:26 -0400)]
nir: fix comment typo about f2d/d2f

Signed-off-by: Rob Clark <robclark@freedesktop.org>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
8 years agonv50/ir: avoid asserts when the state tracker feeds us bogus inputs
Ilia Mirkin [Fri, 13 May 2016 03:42:47 +0000 (23:42 -0400)]
nv50/ir: avoid asserts when the state tracker feeds us bogus inputs

INTERP is defined (by me) to have to have a INPUT source. However the
state tracker does not always obey this. This happens due to varying
packing logic introducing additional mov's which can't always be undone.
Instead of just giving up, we instead try harder to find the original
input. This won't always be possible, for example with indirect
accesses. There's not much we can (easily) do about that though.

This fixes the remaining interpolateAt* failures in dEQP:

dEQP-GLES31.functional.shaders.multisample_interpolation.interpolate_at*

some of which were asserting due to INTERP_* being passed a non-input.

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
8 years agonvc0: don't try to go through the push path for indirect draws
Ilia Mirkin [Sun, 15 May 2016 05:26:37 +0000 (01:26 -0400)]
nvc0: don't try to go through the push path for indirect draws

This fixes

dEQP-GLES31.functional.draw_indirect.draw_elements_indirect.*.default_attribute

These tests were causing a const vbo to be set up, and were small enough
draws that the logic was trying to go via the push path (which emits
data directly into the cmd stream rather than uploading a user vbo).

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Cc: mesa-stable@lists.freedesktop.org
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
8 years agonvc0/ir: make sure to align the second arg of TXD to 4, as we do for TEX
Ilia Mirkin [Sat, 14 May 2016 23:25:15 +0000 (19:25 -0400)]
nvc0/ir: make sure to align the second arg of TXD to 4, as we do for TEX

This was handled in handleTEX(), however the way the logic works, those
extra arguments aren't added on by then, so it did nothing. Instead we
must duplicate that bit here. GK110 appears to complain about
MISALIGNED_GPR, however it's reasonable to believe that GK104 has the
same requirements.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=95403
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Cc: mesa-stable@lists.freedesktop.org
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
8 years agonv50,nvc0: add support for cull distances
Tobias Klausmann [Sun, 8 May 2016 20:44:11 +0000 (22:44 +0200)]
nv50,nvc0: add support for cull distances

Cull distances are just a special case of clip distances as far as the
hardware is concerned. Make sure that the relevant "planes" are enabled,
and flip the clip mode to cull for those.

Signed-off-by: Tobias Klausmann <tobias.johannes.klausmann@mni.thm.de>
[imirkin: add enables on nvc0, add nv50 support]
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Tobias Klausmann <tobias.johannes.klausmann@mni.thm.de>
8 years agost/mesa: disable cull distance for now
Ilia Mirkin [Sat, 14 May 2016 14:19:03 +0000 (10:19 -0400)]
st/mesa: disable cull distance for now

The pass that st/mesa relies on to combine clip and cull distances has
been reverted, so we can't expose ARB_cull_distance until that is
resolved.

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
8 years agoi965: Use blorp for all clears
Jason Ekstrand [Thu, 12 May 2016 21:07:51 +0000 (14:07 -0700)]
i965: Use blorp for all clears

We used to use a meta path on gen8 but we haven't since c7cf17ae758.  We
might as well delete the meta path since blorp works on all gens.

Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
8 years agoi965: Use blorp for all stencil blits
Jason Ekstrand [Thu, 12 May 2016 02:59:38 +0000 (19:59 -0700)]
i965: Use blorp for all stencil blits

We used to use a meta path because blorp didn't support 16x MSAA.  Now it
does, so we don't need the meta paths anymore.

Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
8 years agoi965: Use blorp for all updownsample blits
Jason Ekstrand [Thu, 12 May 2016 00:11:52 +0000 (17:11 -0700)]
i965: Use blorp for all updownsample blits

We used to use a meta path because blorp didn't support 16x MSAA.  Now it
does, so we don't need the meta paths anymore.

Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
8 years agoi965/blorp: Add support for 16x MSAA
Jason Ekstrand [Thu, 12 May 2016 00:11:47 +0000 (17:11 -0700)]
i965/blorp: Add support for 16x MSAA

Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
8 years agoi965: move brw_meta_set_fast_clear_color to brw_meta_util.c
Jason Ekstrand [Thu, 12 May 2016 22:24:40 +0000 (15:24 -0700)]
i965: move brw_meta_set_fast_clear_color to brw_meta_util.c

Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
8 years agoi965; Move brw_meta_get_*_rect to brw_meta_util.c
Jason Ekstrand [Thu, 12 May 2016 22:15:03 +0000 (15:15 -0700)]
i965; Move brw_meta_get_*_rect to brw_meta_util.c

Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
8 years agoi965: Move brw_is_color_fast_clear_compatible to brw_meta_util
Jason Ekstrand [Thu, 12 May 2016 22:03:46 +0000 (15:03 -0700)]
i965: Move brw_is_color_fast_clear_compatible to brw_meta_util

Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
8 years agoi965: Move brw_get_rb_for_slice to brw_meta_util
Jason Ekstrand [Thu, 12 May 2016 03:55:37 +0000 (20:55 -0700)]
i965: Move brw_get_rb_for_slice to brw_meta_util

Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
8 years agoi965/blorp: Get rid of the blorp_prog_data_int() helper
Jason Ekstrand [Wed, 11 May 2016 18:04:19 +0000 (11:04 -0700)]
i965/blorp: Get rid of the blorp_prog_data_int() helper

The helper was initially created to allow us to set reasonable defaults as
we mutated the brw_blorp_prog_data structure in preparation for NIR.  Now
that everything is going through brw_blorp_compile_nir_shader() which fully
fills out the brw_blorp_prog_data structure, we don't need the helper.

Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
8 years agoi965/blorp: Delete the old blorp shader emit code
Jason Ekstrand [Thu, 5 May 2016 21:37:53 +0000 (14:37 -0700)]
i965/blorp: Delete the old blorp shader emit code

Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
8 years agoi965/blorp: Stop doing f2i(i2f(sample_id))
Jason Ekstrand [Fri, 13 May 2016 07:36:25 +0000 (00:36 -0700)]
i965/blorp: Stop doing f2i(i2f(sample_id))

NIR gets kind of awkward when you have a 3-component vector with two floats
and one int.  This led to us accidentally going through float for the
sample index.  It doesn't hurt anything but it also isn't needed.

Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
8 years agoi965/blorp: Refactor coordinate munging
Jason Ekstrand [Thu, 5 May 2016 21:27:23 +0000 (14:27 -0700)]
i965/blorp: Refactor coordinate munging

The original code-flow tried to map original blorp.  This puts things more
where they belong and simplifies some of the logic.

Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
8 years agoi965/blorp: Add bilinear blending support to the NIR path
Jason Ekstrand [Thu, 5 May 2016 18:01:16 +0000 (11:01 -0700)]
i965/blorp: Add bilinear blending support to the NIR path

Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
8 years agoi965/blorp: Add support for averaging resolves to the NIR path
Jason Ekstrand [Tue, 3 May 2016 23:22:46 +0000 (16:22 -0700)]
i965/blorp: Add support for averaging resolves to the NIR path

Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
8 years agoi965/blorp: Add MSAA encode/decode support to the NIR path
Jason Ekstrand [Mon, 2 May 2016 19:13:14 +0000 (12:13 -0700)]
i965/blorp: Add MSAA encode/decode support to the NIR path

Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
8 years agoi965/blorp: Add support for W-[de]tiling to the NIR path
Jason Ekstrand [Mon, 2 May 2016 18:50:06 +0000 (11:50 -0700)]
i965/blorp: Add support for W-[de]tiling to the NIR path

Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
8 years agoi965/blorp: Add support for discard-based bounds checks to the NIR path
Jason Ekstrand [Mon, 2 May 2016 19:30:45 +0000 (12:30 -0700)]
i965/blorp: Add support for discard-based bounds checks to the NIR path

Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
8 years agoi965/blorp: Add initial support for NIR-based blit shaders
Jason Ekstrand [Fri, 29 Apr 2016 19:52:00 +0000 (12:52 -0700)]
i965/blorp: Add initial support for NIR-based blit shaders

Many of the more complex cases still fall back to the old shader builder.

Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
8 years agoi965/blorp: Refactor getting the blit kernel into a helper
Jason Ekstrand [Fri, 29 Apr 2016 19:34:10 +0000 (12:34 -0700)]
i965/blorp: Refactor getting the blit kernel into a helper

Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
8 years agoi965/blorp: Use NIR for clear shaders
Jason Ekstrand [Thu, 28 Apr 2016 00:17:11 +0000 (17:17 -0700)]
i965/blorp: Use NIR for clear shaders

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=95373
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
8 years agoi965/blorp: Create the program key in get_clear_kernel
Jason Ekstrand [Thu, 28 Apr 2016 00:18:02 +0000 (17:18 -0700)]
i965/blorp: Create the program key in get_clear_kernel

There's no reason to be passing a whole struct around just for a single
boolean.  We can create it later when we actually need to use it as a key.

Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
8 years agoi965/blorp: Add a helper for compiling NIR shaders
Jason Ekstrand [Thu, 28 Apr 2016 00:16:30 +0000 (17:16 -0700)]
i965/blorp: Add a helper for compiling NIR shaders

Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
8 years agoblorp: Add initial state setup support for SIMD8 dispatch
Jason Ekstrand [Tue, 10 May 2016 21:57:15 +0000 (14:57 -0700)]
blorp: Add initial state setup support for SIMD8 dispatch

Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
8 years agoi965/blorp: Add a param array to prog_data
Jason Ekstrand [Thu, 28 Apr 2016 00:15:11 +0000 (17:15 -0700)]
i965/blorp: Add a param array to prog_data

This array allows the push constants to be re-arranged on upload.  The
actual arrangement will, eventually, come from the back-end compiler.

Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
8 years agoi965/blorp: Add a prog_data_init helper
Jason Ekstrand [Thu, 28 Apr 2016 04:05:17 +0000 (21:05 -0700)]
i965/blorp: Add a prog_data_init helper

Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
8 years agoi965/fs: Implement the new NIR MCS texturing
Jason Ekstrand [Tue, 3 May 2016 19:34:51 +0000 (12:34 -0700)]
i965/fs: Implement the new NIR MCS texturing

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
8 years agonir: Add texture opcodes and source types for multisample compression
Jason Ekstrand [Tue, 3 May 2016 00:28:38 +0000 (17:28 -0700)]
nir: Add texture opcodes and source types for multisample compression

Intel hardware does a form of multisample compression that involves an
auxilary surface called the MCS.  When an MCS is in use, you have to first
sample from the MCS with a special opcode and then pass the result of that
operation into the next sample instrucion.  Normally, we just do this
ourselves in the back-end, but we want to expose that functionality to NIR
so that we can use MCS values directly in NIR-based blorp.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
8 years agonir/builder: Add a helper for grabbing multiple channels from an ssa def
Jason Ekstrand [Mon, 2 May 2016 23:29:05 +0000 (16:29 -0700)]
nir/builder: Add a helper for grabbing multiple channels from an ssa def

This is similar to nir_channel except that it lets you grab more than one
channel by providing a mask.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
8 years agonir/builder: Generate the alu helpers directly in python
Jason Ekstrand [Fri, 29 Apr 2016 19:26:07 +0000 (12:26 -0700)]
nir/builder: Generate the alu helpers directly in python

There's no reason for having a macro *and* a python generator.  We can
easily just do the whole thing in python.  This has the advantage that we
are no longer definining ALU# macros which conflict with the ones in
brw_fs_builder.h.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
8 years agoi965/fs: Use MRF0 for the repclear message
Jason Ekstrand [Sat, 30 Apr 2016 01:40:35 +0000 (18:40 -0700)]
i965/fs: Use MRF0 for the repclear message

This is what BLORP does.  Making them match cuts down on the noise when
looking at AUB diffs.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
8 years agoi965/blorp: Simplify the sample layout calculation
Jason Ekstrand [Thu, 5 May 2016 00:32:29 +0000 (17:32 -0700)]
i965/blorp: Simplify the sample layout calculation

Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
8 years agoi965/fs: Organize prog_data by ksp number rather than SIMD width
Jason Ekstrand [Thu, 28 Apr 2016 22:37:39 +0000 (15:37 -0700)]
i965/fs: Organize prog_data by ksp number rather than SIMD width

The hardware packets organize kernel pointers and GRF start by slots that
don't map directly to dispatch width.  This means that all of the state
setup code has to re-arrange the data from prog_data into these slots.
This logic has been duplicated 4 times in the GL driver and one more time
in the Vulkan driver.  Let's just put it all in brw_fs.cpp.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
8 years agoi965/gen7_wm: Move where we set the fast clear op
Jason Ekstrand [Thu, 28 Apr 2016 21:40:51 +0000 (14:40 -0700)]
i965/gen7_wm: Move where we set the fast clear op

This better matches gen8 state setup

Acked-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
8 years agoi965/fs: Stop setting dispatch_grf_start_reg from the visitor
Jason Ekstrand [Thu, 28 Apr 2016 21:20:36 +0000 (14:20 -0700)]
i965/fs: Stop setting dispatch_grf_start_reg from the visitor

Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
8 years agoi965/fs: Clean up the logic in compile_fs a bit
Jason Ekstrand [Thu, 28 Apr 2016 19:40:14 +0000 (12:40 -0700)]
i965/fs: Clean up the logic in compile_fs a bit

Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
8 years agoi965/state: Clean up WM/PS state to pull more things out of prog_data
Jason Ekstrand [Tue, 10 May 2016 18:17:27 +0000 (11:17 -0700)]
i965/state: Clean up WM/PS state to pull more things out of prog_data

Now that we have a persample_shading bit in prog_data we can reduce the
amount the state setup code needs to be looking at the GL state.  In
particular, it no longer pulls anything directly out of the
gl_fragment_program and no longer depends on NEW_FRAGMENT_PROGRAM.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
8 years agoi965/fs: Rework the persample shading key/prog_data bits
Jason Ekstrand [Tue, 10 May 2016 00:48:24 +0000 (17:48 -0700)]
i965/fs: Rework the persample shading key/prog_data bits

This commit reworks and simplifies the way we handle persample shading in
the shader key and prog_data.  The previous approach had three different
key bits that had slightly different and hard-to-decern meanings while the
new bits are far more clear.  This commit changes it to two easily
understood bits that communicate everything we need:

 1) key->persample_interp: means that the user has requested persample
    interpolation through the API.  This is equivalent to having
    SAMPLE_SHADING enabled and having MIN_SAMPLE_SHADING_VALUE set high
    enough that you actually get multiple per-sample invocations.

 2) key->multisample_fbo: means that the shader will be running on an
    actual multi-sampled framebuffer.

This commit also adds a new "persample_dispatch" bit to prog_data which
indicates that the shader should be run in persample mode.  This way the
state setup code doesn't have to look at the fragment program or GL state
and can just pull that data out of the prog_data.

In theory, this shuffle could mean more recompiles.  However, in practice,
we were shoving enough state into the key before that we were probably
hitting a recompile on every per-sample shader anyway.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
8 years agonir: Add an info bit for uses_sample_qualifier
Jason Ekstrand [Tue, 10 May 2016 00:15:21 +0000 (17:15 -0700)]
nir: Add an info bit for uses_sample_qualifier

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
8 years agoi965: Fix undefined df bits in brw_reg comparisons.
Kenneth Graunke [Fri, 13 May 2016 23:41:13 +0000 (16:41 -0700)]
i965: Fix undefined df bits in brw_reg comparisons.

Commit 5310bca024f77da40ea6f4c275455f9cb0528f9e added a new "double df"
field to the brw_reg struct, adding an extra 4 bytes of data that isn't
usually initialized (or may contain irrelevant garbage if the struct is
mutated).  This means that it's no longer safe to memcmp().

Instead, add a brw_regs_equal() function which ignores the extra df bits
unless they matter.  To keep the implementation cheap, we wrap the first
set of fields in a union/struct so that we can use a single DWord
comparison.

v2: Drop unnecessary casts (caught by Francisco Jerez).

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
8 years agoi965: disable cull distance temporarily.
Dave Airlie [Sat, 14 May 2016 01:38:38 +0000 (11:38 +1000)]
i965: disable cull distance temporarily.

I'll fix this up on Monday, so leave the docs changes in place.

Signed-off-by: Dave Airlie <airlied@redhat.com>