diff options
author | Juzhe-Zhong <juzhe.zhong@rivai.ai> | 2024-01-10 11:06:50 +0800 |
---|---|---|
committer | Pan Li <pan2.li@intel.com> | 2024-01-11 07:23:59 +0800 |
commit | 2aa83f0ad8a613fc46bf13cdd3403445d5c85b69 (patch) | |
tree | f7a6cef3a429c42ed0bdddcdb3992cdb868ea509 /gcc | |
parent | 8415bceea9d3ca86adc00ae8ad92deaec0457dd1 (diff) | |
download | gcc-2aa83f0ad8a613fc46bf13cdd3403445d5c85b69.zip gcc-2aa83f0ad8a613fc46bf13cdd3403445d5c85b69.tar.gz gcc-2aa83f0ad8a613fc46bf13cdd3403445d5c85b69.tar.bz2 |
RISC-V: Minor tweak dynamic cost model
v2 update: Robostify tests.
While working on cost model, I notice one case that dynamic lmul cost doesn't work well.
Before this patch:
foo:
lui a4,%hi(.LANCHOR0)
li a0,1953
li a1,63
addi a4,a4,%lo(.LANCHOR0)
li a3,64
vsetvli a2,zero,e32,mf2,ta,ma
vmv.v.x v5,a0
vmv.v.x v4,a1
vid.v v3
.L2:
vsetvli a5,a3,e32,mf2,ta,ma
vadd.vi v2,v3,1
vadd.vv v1,v3,v5
mv a2,a5
vmacc.vv v1,v2,v4
slli a1,a5,2
vse32.v v1,0(a4)
sub a3,a3,a5
add a4,a4,a1
vsetvli a5,zero,e32,mf2,ta,ma
vmv.v.x v1,a2
vadd.vv v3,v3,v1
bne a3,zero,.L2
li a0,0
ret
Unexpected: Use scalable vector and LMUL = MF2 which is wasting computation resources.
Ideally, we should use LMUL = M8 VLS modes.
The root cause is the dynamic LMUL heuristic dominates the VLS heuristic.
Adapt the cost model heuristic.
After this patch:
foo:
lui a4,%hi(.LANCHOR0)
addi a4,a4,%lo(.LANCHOR0)
li a3,4096
li a5,32
li a1,2016
addi a2,a4,128
addiw a3,a3,-32
vsetvli zero,a5,e32,m8,ta,ma
li a0,0
vid.v v8
vsll.vi v8,v8,6
vadd.vx v16,v8,a1
vadd.vx v8,v8,a3
vse32.v v16,0(a4)
vse32.v v8,0(a2)
ret
Tested on both RV32/RV64 no regression.
Ok for trunk ?
gcc/ChangeLog:
* config/riscv/riscv-vector-costs.cc (costs::better_main_loop_than_p): Minior tweak.
gcc/testsuite/ChangeLog:
* gcc.dg/vect/costmodel/riscv/rvv/vla_vs_vls-10.c: Fix test.
* gcc.dg/vect/costmodel/riscv/rvv/vla_vs_vls-11.c: Ditto.
* gcc.dg/vect/costmodel/riscv/rvv/vla_vs_vls-12.c: Ditto.
Diffstat (limited to 'gcc')
4 files changed, 9 insertions, 11 deletions
diff --git a/gcc/config/riscv/riscv-vector-costs.cc b/gcc/config/riscv/riscv-vector-costs.cc index f4a1a78..e53f4a1 100644 --- a/gcc/config/riscv/riscv-vector-costs.cc +++ b/gcc/config/riscv/riscv-vector-costs.cc @@ -994,7 +994,8 @@ costs::better_main_loop_than_p (const vector_costs *uncast_other) const vect_vf_for_cost (other_loop_vinfo)); /* Apply the unrolling heuristic described above m_unrolled_vls_niters. */ - if (bool (m_unrolled_vls_stmts) != bool (other->m_unrolled_vls_stmts)) + if (bool (m_unrolled_vls_stmts) != bool (other->m_unrolled_vls_stmts) + && m_cost_type != other->m_cost_type) { bool this_prefer_unrolled = this->prefer_unrolled_loop (); bool other_prefer_unrolled = other->prefer_unrolled_loop (); diff --git a/gcc/testsuite/gcc.dg/vect/costmodel/riscv/rvv/vla_vs_vls-10.c b/gcc/testsuite/gcc.dg/vect/costmodel/riscv/rvv/vla_vs_vls-10.c index 3ddffa3..89a6c67 100644 --- a/gcc/testsuite/gcc.dg/vect/costmodel/riscv/rvv/vla_vs_vls-10.c +++ b/gcc/testsuite/gcc.dg/vect/costmodel/riscv/rvv/vla_vs_vls-10.c @@ -3,7 +3,7 @@ #include <stdint-gcc.h> -#define N 40 +#define N 48 int a[N]; @@ -22,7 +22,6 @@ foo (){ return 0; } -/* { dg-final { scan-assembler-times {vsetivli\s+zero,\s*8,\s*e32,\s*m2,\s*t[au],\s*m[au]} 1 } } */ /* { dg-final { scan-assembler-times {vsetivli\s+zero,\s*16,\s*e32,\s*m4,\s*t[au],\s*m[au]} 1 } } */ -/* { dg-final { scan-assembler-times {vsetivli} 2 } } */ +/* { dg-final { scan-assembler-times {vsetivli} 1 } } */ /* { dg-final { scan-assembler-not {vsetvli} } } */ diff --git a/gcc/testsuite/gcc.dg/vect/costmodel/riscv/rvv/vla_vs_vls-11.c b/gcc/testsuite/gcc.dg/vect/costmodel/riscv/rvv/vla_vs_vls-11.c index 7625ec5..86732ef 100644 --- a/gcc/testsuite/gcc.dg/vect/costmodel/riscv/rvv/vla_vs_vls-11.c +++ b/gcc/testsuite/gcc.dg/vect/costmodel/riscv/rvv/vla_vs_vls-11.c @@ -3,7 +3,7 @@ #include <stdint-gcc.h> -#define N 40 +#define N 64 int a[N]; @@ -22,7 +22,6 @@ foo (){ return 0; } -/* { dg-final { scan-assembler-times {vsetivli\s+zero,\s*8,\s*e32,\s*m2,\s*t[au],\s*m[au]} 1 } } */ /* { dg-final { scan-assembler-times {vsetvli\s+zero,\s*[a-x0-9]+,\s*e32,\s*m8,\s*t[au],\s*m[au]} 1 } } */ -/* { dg-final { scan-assembler-times {vsetivli} 1 } } */ +/* { dg-final { scan-assembler-not {vsetivli} } } */ /* { dg-final { scan-assembler-times {vsetvli} 1 } } */ diff --git a/gcc/testsuite/gcc.dg/vect/costmodel/riscv/rvv/vla_vs_vls-12.c b/gcc/testsuite/gcc.dg/vect/costmodel/riscv/rvv/vla_vs_vls-12.c index 7625ec5..a1fcb3f 100644 --- a/gcc/testsuite/gcc.dg/vect/costmodel/riscv/rvv/vla_vs_vls-12.c +++ b/gcc/testsuite/gcc.dg/vect/costmodel/riscv/rvv/vla_vs_vls-12.c @@ -1,9 +1,9 @@ /* { dg-do compile } */ -/* { dg-options "-march=rv64gcv -mabi=lp64d -O3 --param=riscv-autovec-lmul=m8 -fno-schedule-insns -fno-schedule-insns2" } */ +/* { dg-options "-march=rv64gcv -mabi=lp64d -O3 --param=riscv-autovec-lmul=dynamic -fno-schedule-insns -fno-schedule-insns2" } */ #include <stdint-gcc.h> -#define N 40 +#define N 64 int a[N]; @@ -22,7 +22,6 @@ foo (){ return 0; } -/* { dg-final { scan-assembler-times {vsetivli\s+zero,\s*8,\s*e32,\s*m2,\s*t[au],\s*m[au]} 1 } } */ /* { dg-final { scan-assembler-times {vsetvli\s+zero,\s*[a-x0-9]+,\s*e32,\s*m8,\s*t[au],\s*m[au]} 1 } } */ -/* { dg-final { scan-assembler-times {vsetivli} 1 } } */ +/* { dg-final { scan-assembler-not {vsetivli} } } */ /* { dg-final { scan-assembler-times {vsetvli} 1 } } */ |