| Commit message (Collapse) | Author | Age | Files | Lines |
|---|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Many shaders use a pattern such as:
for (int i = 0; i < NUM_LIGHTS; i++) {
...access a uniform array, or shader input/output array...
}
where NUM_LIGHTS is a small constant (such as 2, 4, or 8).
The expectation is that the compiler will unroll those loops, turning
the array access into constant indexing, which is more efficient, and
which may enable array splitting and other optimizations.
In many cases, our heuristic fails - either there's another tiny nested
loop inside, or the estimated number of instructions is just barely
beyond the threshold. So, we fail to unroll the loop, leaving the
variable indexing in place.
Drivers which don't support the particular flavor of variable indexing
will call lower_variable_index_to_cond_assign(), which generates piles
and piles of immensely inefficient code. We'd like to avoid generating
that.
This patch detects unsupported forms of variable-indexing in loops, where
the array index is a loop induction variable. In that case, it bypasses
the loop-too-large heuristic and forces unrolling.
Improves performance in various microbenchmarks: Gl32PSBump8 by 47%,
Gl32ShMapVsm by 80%, and Gl32ShMapPcf by 27%. No changes in shader-db.
v2: Check ir->array for being an array or matrix, rather than the
ir_dereference_array itself.
v3: Fix and expand statistics in commit message.
Signed-off-by: Kenneth Graunke <[email protected]>
Reviewed-by: Eric Anholt <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
| |
The "fail" flag is set if loop_unroll_count encounters a nested loop;
calling the flag "nested_loop" is a bit clearer.
The original reasoning was that count is inaccurate (too small) if there
are nested loops, as we don't do any sort of analysis on the inner loop.
Signed-off-by: Kenneth Graunke <[email protected]>
Reviewed-by: Eric Anholt <[email protected]>
|
| |
|
|
|
|
|
| |
Loop unrolling will need to know a few more options in the future.
Signed-off-by: Kenneth Graunke <[email protected]>
Reviewed-by: Eric Anholt <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
../../src/glsl/builtin_functions.cpp:72:1: warning: unused parameter 'state' [-Wunused-parameter]
../../src/glsl/ir_clone.cpp:31:1: warning: unused parameter 'ht' [-Wunused-parameter]
../../src/glsl/ir_equals.cpp:44:1: warning: unused parameter 'ir' [-Wunused-parameter]
../../src/glsl/ir_equals.cpp:50:1: warning: unused parameter 'ignore' [-Wunused-parameter]
../../src/glsl/ir_equals.cpp:68:1: warning: unused parameter 'ignore' [-Wunused-parameter]
../../src/glsl/ir_print_visitor.cpp:149:6: warning: unused parameter 'ir' [-Wunused-parameter]
../../src/glsl/ir_print_visitor.cpp:556:1: warning: unused parameter 'ir' [-Wunused-parameter]
../../src/glsl/ir_print_visitor.cpp:562:1: warning: unused parameter 'ir' [-Wunused-parameter]
../../src/glsl/link_uniforms.cpp:213:1: warning: unused parameter 'record_type' [-Wunused-parameter]
../../src/glsl/loop_analysis.cpp:225:1: warning: unused parameter 'ir' [-Wunused-parameter]
../../src/glsl/loop_unroll.cpp:73:30: warning: unused parameter 'ir' [-Wunused-parameter]
../../src/glsl/loop_unroll.cpp:79:30: warning: unused parameter 'ir' [-Wunused-parameter]
../../src/glsl/loop_unroll.cpp:85:30: warning: unused parameter 'ir' [-Wunused-parameter]
../../src/glsl/opt_copy_propagation_elements.cpp:189:1: warning: unused parameter 'ir' [-Wunused-parameter]
../../src/glsl/opt_cse.cpp:402:1: warning: unused parameter 'ir' [-Wunused-parameter]
../../src/glsl/opt_dead_code_local.cpp:117:30: warning: unused parameter 'ir' [-Wunused-parameter]
../../src/glsl/opt_redundant_jumps.cpp:53:1: warning: unused parameter 'ir' [-Wunused-parameter]
../../src/glsl/opt_vectorize.cpp:301:1: warning: unused parameter 'ir' [-Wunused-parameter]
Signed-off-by: Ian Romanick <[email protected]>
Reviewed-by: Kenneth Graunke <[email protected]>
|
| |
|
|
|
|
|
|
| |
Now that loop_controls no longer creates normatively bound loops,
there is no need for ir_loop::normative_bound or the
lower_bounded_loops pass.
Reviewed-by: Ian Romanick <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Previously, when loop_controls analyzed a loop and found that it had a
fixed bound (known at compile time), it would remove all of the loop
terminators and instead set the loop's normative_bound field to force
the loop to execute the correct number of times.
This made loop unrolling easy, but it had a serious disadvantage.
Since most GPU's don't have a native mechanism for executing a loop a
fixed number of times, in order to implement the normative bound, the
back-ends would have to synthesize a new loop induction variable. As
a result, many loops wound up having two induction variables instead
of one. This caused extra register pressure and unnecessary
instructions.
This patch modifies loop_controls so that it doesn't set the loop's
normative_bound anymore. Instead it leaves one of the terminators in
the loop (the limiting terminator), so the back-end doesn't have to go
to any extra work to ensure the loop terminates at the right time.
This complicates loop unrolling slightly: when deciding whether a loop
can be unrolled, we have to account for the presence of the limiting
terminator. And when we do unroll the loop, we have to remove the
limiting terminator first.
For an example of how this results in more efficient back end code,
consider the loop:
for (int i = 0; i < 100; i++) {
total += i;
}
Previous to this patch, on i965, this loop would compile down to this
(vec4) native code:
mov(8) g4<1>.xD 0D
mov(8) g8<1>.xD 0D
loop:
cmp.ge.f0(8) null g8<4;4,1>.xD 100D
(+f0) if(8)
break(8)
endif(8)
add(8) g5<1>.xD g5<4;4,1>.xD g4<4;4,1>.xD
add(8) g8<1>.xD g8<4;4,1>.xD 1D
add(8) g4<1>.xD g4<4;4,1>.xD 1D
while(8) loop
(notice that both g8 and g4 are loop induction variables; one is used
to terminate the loop, and the other is used to accumulate the total).
After this patch, the same loop compiles to:
mov(8) g4<1>.xD 0D
loop:
cmp.ge.f0(8) null g4<4;4,1>.xD 100D
(+f0) if(8)
break(8)
endif(8)
add(8) g5<1>.xD g5<4;4,1>.xD g4<4;4,1>.xD
add(8) g4<1>.xD g4<4;4,1>.xD 1D
while(8) loop
Reviewed-by: Ian Romanick <[email protected]>
|
| |
|
|
|
|
|
|
| |
This value is now redundant with
loop_variable_state::limiting_terminator->iterations and
ir_loop::normative_bound.
Reviewed-by: Ian Romanick <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The old logic of loop_unroll_visitor::visit_leave(ir_loop *) was:
heuristics to skip unrolling in various circumstances;
if (loop contains more than one jump)
return;
else if (loop contains one jump) {
if (the jump is an unconditional "break" at the end of the loop) {
remove the break and set iteration count to 1;
fall through to simple loop unrolling code;
} else {
for (each "if" statement in the loop body)
see if the jump is a "break" at the end of one of its forks;
if (the "break" wasn't found)
return;
splice the remainder of the loop into the other fork of the "if";
remove the "break";
complex loop unrolling code;
return;
}
}
simple loop unrolling code;
return;
These tasks have been moved to their own functions:
- splice the remainder of the loop into the other fork of the "if"
- simple loop unrolling code
- complex loop unrolling code
And the logic has been flattened to:
heuristics to skip unrolling in various circumstances;
if (loop contains more than one jump)
return;
if (loop contains no jumps) {
simple loop unroll;
return;
}
if (the jump is an unconditional "break" at the end of the loop) {
remove the break;
simple loop unroll with iteration count of 1;
return;
}
for (each "if" statement in the loop body) {
if (the jump is a "break" at the end of one of its forks) {
splice the remainder of the loop into the other fork of the "if";
remove the "break";
complex loop unroll;
return;
}
}
This will make it easier to modify the loop unrolling algorithm in a
future patch.
Reviewed-by: Ian Romanick <[email protected]>
|
| |
|
|
|
|
|
|
| |
This gives the compiler the chance to inline and not export class symbols
even in the absence of LTO. Saves about 60kb on disk.
Reviewed-by: Kenneth Graunke <[email protected]>
Reviewed-by: Ian Romanick <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
| |
Before, we were only counting top-level instructions. But if we have
an assignment of a giant expression tree (such as the ones eventually
generated by glsl-fs-unroll), we were counting the same as an
assignment of a variable deref.
glsl-fs-unroll-explosion now fails in a reasonable amount of time on
i965 because the unrolling didn't go ridiculously far.
Reviewed-by: Kenneth Graunke <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Avoid unrollong loops that are either nested loops or
where the loop body times the unroll count is huge.
The change is far from being perfect but it extends the
loop unrolling decision heuristic by some additional
safeguard. In particular this cuts down compilation of
a shader precomputing atmospheric scattering integral
tables containing two nesting levels in a loop from
something way beyond some minutes (I never waited for
it to finish) to some fractions of a second.
This fixes piglit tests glsl-fs-unroll-explosion and
glsl-vs-unroll-explosion on r600g.
Reviewed-by: Eric Anholt <[email protected]>
Signed-off-by: Mathias Fröhlich <[email protected]>
|
| | |
|
| |
|
|
|
|
|
|
|
|
|
| |
Currently we only unroll loops with conditional breaks at the end, which is
the form that lower_jumps generates.
However, if breaks are not lowered, they tend to appear at the beginning, so
add support for a conditional break anywhere.
Signed-off-by: Luca Barbieri <[email protected]>
Signed-off-by: Kenneth Graunke <[email protected]>
|
| |
|
|
|
|
|
| |
Found this bug by code inspection. Based off the comments just before
this code, the intent is to find whether the break exists in the "then"
branch or the "else" branch. However, the code actually looked at the
last instruction in the "then" branch twice.
|
| | |
|
| | |
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
If the loop ends with an if with one break or in a single break unroll
it. Loops that end with a continue will have that continue removed by
the redundant jump optimizer. Likewise loops that end with an
if-statement with a break at the end of both branches will have the
break pulled out after the if-statement.
Loops of the form
for (...) {
do_something1();
if (cond) {
do_something2();
break;
} else {
do_something3();
}
}
will be unrolled as
do_something1();
if (cond) {
do_something2();
} else {
do_something3();
do_something1();
if (cond) {
do_something2();
} else {
do_something3();
/* Repeat inserting iterations here.*/
}
}
ir_lower_jumps can guarantee that all loops are put in this form
and thus all loops are now potentially unrollable if an upper bound
on the number of iterations can be found.
Signed-off-by: Ian Romanick <[email protected]>
|
| |
|
|
|
|
|
|
|
| |
This increases the chance that GLSL programs will actually work.
Note that continues and returns are not yet lowered, so linking
will just fail if not supported.
Signed-off-by: Ian Romanick <[email protected]>
|
| |
|
