GPGPU: Use a tile size of 32 by default

The tile size was previously uninitialized. As a result, it was often zero (aka.
no tiling), which is not what we want in general. More importantly, there was
the risk for arbitrary tile sizes to be choosen, which we did not observe, but
which still is highly problematic.

llvm-svn: 275418

diff --git a/polly/lib/CodeGen/PPCGCodeGeneration.cpp b/polly/lib/CodeGen/PPCGCodeGeneration.cpp
index deb80db..d0921db 100644 (file)
--- a/polly/lib/CodeGen/PPCGCodeGeneration.cpp
+++ b/polly/lib/CodeGen/PPCGCodeGeneration.cpp
@@ -75,6 +75,8 @@ public:
     Options->ctx = nullptr;
     Options->sizes = nullptr;
 
+    Options->tile_size = 32;
+
     Options->use_private_memory = false;
     Options->use_shared_memory = false;
     Options->max_shared_memory = 0;

diff --git a/polly/test/GPGPU/double-parallel-loop.ll b/polly/test/GPGPU/double-parallel-loop.ll
index 7ae5010..963a411 100644 (file)
--- a/polly/test/GPGPU/double-parallel-loop.ll
+++ b/polly/test/GPGPU/double-parallel-loop.ll
@@ -29,22 +29,21 @@
 ; SCHED:           child:
 ; SCHED:             mark: "kernel"
 ; SCHED:             child:
-; SCHED:               context: "[b0, b1, t0, t1] -> { [] : 0 <= b0 <= 255 and 0 <= b1 <= 255 and 0 <= t0 <= 3 and 0 <= t1 <= 3 }"
+; SCHED:               context: "[b0, b1, t0, t1] -> { [] : 0 <= b0 <= 31 and 0 <= b1 <= 31 and 0 <= t0 <= 31 and 0 <= t1 <= 15 }"
 ; SCHED:               child:
-; SCHED:                 filter: "[b0, b1] -> { Stmt_bb5[i0, i1] : -3 - 4b0 + i0 <= 1024*floor((i0)/1024) <= -4b0 + i0 and -3 - 4b1 + i1 <= 1024*floor((i1)/1024) <= -4b1 + i1 }"
+; SCHED:                 filter: "[b0, b1] -> { Stmt_bb5[i0, i1] : -31 - 32b0 + i0 <= 8192*floor((i0)/8192) <= -32b0 + i0 and -31 - 32b1 + i1 <= 8192*floor((i1)/8192) <= -32b1 + i1 }"
 ; SCHED:                 child:
-; SCHED:                   schedule: "[{ Stmt_bb5[i0, i1] -> [(floor((i0)/1024))] }, { Stmt_bb5[i0, i1] -> [(floor((i1)/1024))] }]"
+; SCHED:                   schedule: "[{ Stmt_bb5[i0, i1] -> [(floor((i0)/8192))] }, { Stmt_bb5[i0, i1] -> [(floor((i1)/8192))] }]"
 ; SCHED:                   permutable: 1
 ; SCHED:                   coincident: [ 1, 1 ]
 ; SCHED:                   child:
-; SCHED:                     filter: "[t0, t1] -> { Stmt_bb5[i0, i1] : 4*floor((-t0 + i0)/4) = -t0 + i0 and 4*floor((-t1 + i1)/4) = -t1 + i1 and 0 <= t0 <= 3 and 0 <= t1 <= 3 }"
+; SCHED:                     filter: "[t0, t1] -> { Stmt_bb5[i0, i1] : 32*floor((-t0 + i0)/32) = -t0 + i0 and 16*floor((-t1 + i1)/16) = -t1 + i1 and 0 <= t0 <= 31 and 0 <= t1 <= 15 }"
 ; SCHED:                     child:
-; SCHED:                       schedule: "[{ Stmt_bb5[i0, i1] -> [(0)] }, { Stmt_bb5[i0, i1] -> [(0)] }]"
+; SCHED:                       schedule: "[{ Stmt_bb5[i0, i1] -> [(0)] }, { Stmt_bb5[i0, i1] -> [(floor((i1)/16) - 2*floor((i1)/32))] }]"
 ; SCHED:                       permutable: 1
 ; SCHED:                       coincident: [ 1, 1 ]
 ; SCHED:       - filter: "{  }"
 
-
 ;    void double_parallel_loop(float A[][1024]) {
 ;      for (long i = 0; i < 1024; i++)
 ;        for (long j = 0; j < 1024; j++)