intel/llvm
1
0
Fork 0
mirror of https://github.com/intel/llvm.git synced 2026-01-14 20:10:50 +08:00
Code Issues Packages Projects Releases Wiki Activity

Do not scale tile loops

Browse Source 
We now generate tile loops as:

 for (int c1 = 0; c1 <= 47; c1 += 1)
   for (int c2 = 0; c2 <= 47; c2 += 1)
     for (int c3 = 0; c3 <= 31; c3 += 1)
       for (int c4 = 0; c4 <= 31; c4 += 4)
         #pragma simd
         for (int c5 = c4; c5 <= c4 + 3; c5 += 1)
           Stmt_for_body3(32 * c1 + c3, 32 * c2 + c5);

instead of

 for (int c1 = 0; c1 <= 1535; c1 += 32)
   for (int c2 = 0; c2 <= 1535; c2 += 32)
     for (int c3 = 0; c3 <= 31; c3 += 1)
       for (int c4 = 0; c4 <= 31; c4 += 4)
         #pragma simd
         for (int c5 = c4; c5 <= c4 + 3; c5 += 1)
           Stmt_for_body3(c1 + c3, c2 + c5);

Run-time performance-wise this makes little difference, but this gives a large
reduction in compile time (10-30% on 17 LNT benchmarks). Apparently the isl
AST generator is not yet very efficient in generating the latter.

llvm-svn: 233675
This commit is contained in:
Tobias Grosser
2015-03-31 07:52:36 +00:00
parent 545b951d3c
commit 4f6bceface
5 changed files with 15 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
polly/lib/Transform/ScheduleOptimizer.cpp
View File

@@ -431,6 +431,7 @@ bool IslScheduleOptimizer::runOnScop(Scop &S) {
isl_options_set_schedule_maximize_band_depth(S.getIslCtx(), IslMaximizeBands);
isl_options_set_schedule_max_constant_term(S.getIslCtx(), MaxConstantTerm);
isl_options_set_schedule_max_coefficient(S.getIslCtx(), MaxCoefficient);
isl_options_set_tile_scale_tile_loops(S.getIslCtx(), 0);
isl_options_set_on_error(S.getIslCtx(), ISL_ON_ERROR_CONTINUE);

4
polly/test/ScheduleOptimizer/line-tiling-2.ll
View File

@@ -1,9 +1,9 @@
; RUN: opt %loadPolly -polly-detect-unprofitable -polly-opt-isl -analyze -polly-no-tiling=0 -polly-ast -polly-tile-sizes=1,64 < %s | FileCheck %s
; CHECK: for (int c0 = 0; c0 <= 1023; c0 += 1)
; CHECK: for (int c1 = 0; c1 <= 511; c1 += 64)
; CHECK: for (int c1 = 0; c1 <= 7; c1 += 1)
; CHECK: for (int c3 = 0; c3 <= 63; c3 += 1)
; CHECK: Stmt_for_body3(c0, c1 + c3);
; CHECK: Stmt_for_body3(c0, 64 * c1 + c3);
target datalayout = "e-m:e-p:32:32-i64:64-v128:64:128-n32-S64"

4
polly/test/ScheduleOptimizer/line-tiling.ll
View File

@@ -1,9 +1,9 @@
; RUN: opt %loadPolly -polly-detect-unprofitable -polly-opt-isl -analyze -polly-no-tiling=0 -polly-ast -polly-tile-sizes=64,1 < %s | FileCheck %s
; CHECK: for (int c0 = 0; c0 <= 1023; c0 += 64)
; CHECK: for (int c0 = 0; c0 <= 15; c0 += 1)
; CHECK: for (int c1 = 0; c1 <= 511; c1 += 1)
; CHECK: for (int c2 = 0; c2 <= 63; c2 += 1)
; CHECK: Stmt_for_body3(c0 + c2, c1);
; CHECK: Stmt_for_body3(64 * c0 + c2, c1);
target datalayout = "e-m:e-p:32:32-i64:64-v128:64:128-n32-S64"

14
polly/test/ScheduleOptimizer/prevectorization.ll
View File

@@ -56,23 +56,23 @@ for.end30: ; preds = %for.inc28
attributes #0 = { nounwind uwtable "less-precise-fpmad"="false" "no-frame-pointer-elim"="true" "no-frame-pointer-elim-non-leaf" "no-infs-fp-math"="false" "no-nans-fp-math"="false" "stack-protector-buffer-size"="8" "unsafe-fp-math"="false" "use-soft-float"="false" }
; CHECK: #pragma known-parallel
; CHECK: for (int c1 = 0; c1 <= 1535; c1 += 32)
; CHECK: for (int c2 = 0; c2 <= 1535; c2 += 32)
; CHECK: for (int c1 = 0; c1 <= 47; c1 += 1)
; CHECK: for (int c2 = 0; c2 <= 47; c2 += 1)
; CHECK: for (int c3 = 0; c3 <= 31; c3 += 1)
; CHECK: for (int c4 = 0; c4 <= 31; c4 += 4)
; CHECK: #pragma simd
; CHECK: for (int c5 = c4; c5 <= c4 + 3; c5 += 1)
; CHECK: Stmt_for_body3(c1 + c3, c2 + c5);
; CHECK: Stmt_for_body3(32 * c1 + c3, 32 * c2 + c5);
; CHECK: #pragma known-parallel
; CHECK: for (int c1 = 0; c1 <= 1535; c1 += 32)
; CHECK: for (int c2 = 0; c2 <= 1535; c2 += 32)
; CHECK: for (int c3 = 0; c3 <= 1535; c3 += 32)
; CHECK: for (int c1 = 0; c1 <= 47; c1 += 1)
; CHECK: for (int c2 = 0; c2 <= 47; c2 += 1)
; CHECK: for (int c3 = 0; c3 <= 47; c3 += 1)
; CHECK: for (int c4 = 0; c4 <= 31; c4 += 1)
; CHECK: for (int c5 = 0; c5 <= 31; c5 += 4)
; CHECK: for (int c6 = 0; c6 <= 31; c6 += 1)
; CHECK: #pragma simd
; CHECK: for (int c7 = c5; c7 <= c5 + 3; c7 += 1)
; CHECK: Stmt_for_body8(c1 + c4, c2 + c7, c3 + c6);
; CHECK: Stmt_for_body8(32 * c1 + c4, 32 * c2 + c7, 32 * c3 + c6);
!llvm.ident = !{!0}

6
polly/test/ScheduleOptimizer/rectangular-tiling.ll
View File

@@ -1,10 +1,10 @@
; RUN: opt %loadPolly -polly-detect-unprofitable -polly-opt-isl -analyze -polly-no-tiling=0 -polly-ast -polly-tile-sizes=256,16 < %s | FileCheck %s
; CHECK: for (int c0 = 0; c0 <= 1023; c0 += 256)
; CHECK: for (int c1 = 0; c1 <= 511; c1 += 16)
; CHECK: for (int c0 = 0; c0 <= 3; c0 += 1)
; CHECK: for (int c1 = 0; c1 <= 31; c1 += 1)
; CHECK: for (int c2 = 0; c2 <= 255; c2 += 1)
; CHECK: for (int c3 = 0; c3 <= 15; c3 += 1)
; CHECK: Stmt_for_body3(c0 + c2, c1 + c3);
; CHECK: Stmt_for_body3(256 * c0 + c2, 16 * c1 + c3);
target datalayout = "e-m:e-p:32:32-i64:64-v128:64:128-n32-S64"