diff mbox

[ARM] implement division using vrecpe/vrecps with -funsafe-math-optimizations

Message ID CAAgBjMkmijb5QdWnRk5RYUFQXXb8HiO4rER+GxOK_DAzhK=rDg@mail.gmail.com
State Superseded
Headers show

Commit Message

Prathamesh Kulkarni July 30, 2015, 9:10 p.m. UTC 
On 29 July 2015 at 16:03, Kyrill Tkachov <kyrylo.tkachov@arm.com> wrote:
> Hi Prathamesh,
>
> This is probably not appropriate for -Os optimisation.
> And for speed optimisation I imagine it can vary a lot on the target the
> code is run.
> Do you have any benchmark results for this patch?
Hi Kyrill,
Thanks for the review. I have attempted to address your comments in
the attached patch.
Does it look OK from correctness perspective ?
Unfortunately I haven't done benchmarking yet.
I ran a test-case (attached) prepared by Charles for target
arm-linux-gnueabihf (on APM Mustang),
and it appeared to run faster with the patch:
Options passed: -O3 -mfpu=neon -mfloat-abi=hard -funsafe-math-optimizations

Before:
t8a, len =       32, took   2593977 ticks
t8a, len =      128, took   2408907 ticks
t8a, len =     1024, took   2354950 ticks
t8a, len =    65536, took   2365041 ticks
t8a, len =  1048576, took   2692928 ticks

After:
t8a, len =       32, took   2027323 ticks
t8a, len =      128, took   1920595 ticks
t8a, len =     1024, took   1827250 ticks
t8a, len =    65536, took   1797924 ticks
t8a, len =  1048576, took   2026274 ticks

I will get back to you soon with benchmarking results.

Thanks,
Prathamesh
>
> Thanks,
> Kyrill
>
>
> On 29/07/15 11:09, Prathamesh Kulkarni wrote:
>>
>> Hi,
>> This patch tries to implement division with multiplication by
>> reciprocal using vrecpe/vrecps
>> with -funsafe-math-optimizations and -freciprocal-math enabled.
>> Tested on arm-none-linux-gnueabihf using qemu.
>> OK for trunk ?
>>
>> Thank you,
>> Prathamesh
>
> +    /* Perform 2 iterations of Newton-Raphson method for better accuracy */
> +    for (int i = 0; i < 2; i++)
> +      {
> +    emit_insn (gen_neon_vrecps<mode> (vrecps_temp, rec, operands[2]));
> +    emit_insn (gen_mul<mode>3 (rec, rec, vrecps_temp));
> +      }
> +
> +    /* We now have reciprocal in rec, perform operands[0] = operands[1] *
> rec */
> +    emit_insn (gen_mul<mode>3 (operands[0], operands[1], rec));
> +    DONE;
> +  }
> +)
> +
>
> Full stop and two spaces at the end of the comments.
>
2015-07-28  Prathamesh Kulkarni  <prathamesh.kulkarni@linaro.org>
	    Charles Baylis  <charles.baylis@linaro.org>

	* config/arm/neon.md (div<mode>3): New pattern.

testsuite/
	* gcc.target/arm/vect-div-1.c: New test-case.
	* gcc.target/arm/vect-div-2.c: Likewise.
diff mbox

Patch

diff --git a/gcc/config/arm/neon.md b/gcc/config/arm/neon.md
index 654d9d5..f2dbcc4 100644
--- a/gcc/config/arm/neon.md
+++ b/gcc/config/arm/neon.md
@@ -548,6 +548,33 @@ 
                     (const_string "neon_mul_<V_elem_ch><q>")))]
 )
 
+(define_expand "div<mode>3"
+  [(set (match_operand:VCVTF 0 "s_register_operand" "=w")
+        (div:VCVTF (match_operand:VCVTF 1 "s_register_operand" "w")
+		  (match_operand:VCVTF 2 "s_register_operand" "w")))]
+  "TARGET_NEON && !optimize_size
+   && flag_unsafe_math_optimizations && flag_reciprocal_math"
+  {
+    rtx rec = gen_reg_rtx (<MODE>mode);
+    rtx vrecps_temp = gen_reg_rtx (<MODE>mode);
+
+    /* Reciprocal estimate.  */
+    emit_insn (gen_neon_vrecpe<mode> (rec, operands[2]));
+
+    /* Perform 2 iterations of newton-raphson method.  */
+    for (int i = 0; i < 2; i++)
+      {
+	emit_insn (gen_neon_vrecps<mode> (vrecps_temp, rec, operands[2]));
+	emit_insn (gen_mul<mode>3 (rec, rec, vrecps_temp));
+      }
+
+    /* We now have reciprocal in rec, perform operands[0] = operands[1] * rec.  */
+    emit_insn (gen_mul<mode>3 (operands[0], operands[1], rec));
+    DONE;
+  }
+)
+
+
 (define_insn "mul<mode>3add<mode>_neon"
   [(set (match_operand:VDQW 0 "s_register_operand" "=w")
         (plus:VDQW (mult:VDQW (match_operand:VDQW 2 "s_register_operand" "w")
diff --git a/gcc/testsuite/gcc.target/arm/vect-div-1.c b/gcc/testsuite/gcc.target/arm/vect-div-1.c
new file mode 100644
index 0000000..e562ef3
--- /dev/null
+++ b/gcc/testsuite/gcc.target/arm/vect-div-1.c
@@ -0,0 +1,14 @@ 
+/* { dg-do compile } */
+/* { dg-require-effective-target arm_v8_neon_ok } */
+/* { dg-options "-O2 -funsafe-math-optimizations -ftree-vectorize -fdump-tree-vect-all" } */
+/* { dg-add-options arm_v8_neon } */
+
+void
+foo (int len, float * __restrict p, float *__restrict x)
+{
+  len = len & ~31;
+  for (int i = 0; i < len; i++)
+    p[i] = p[i] / x[i];
+}
+
+/* { dg-final { scan-tree-dump-times "vectorized 1 loops" 1 "vect" } } */
diff --git a/gcc/testsuite/gcc.target/arm/vect-div-2.c b/gcc/testsuite/gcc.target/arm/vect-div-2.c
new file mode 100644
index 0000000..8e15d0a
--- /dev/null
+++ b/gcc/testsuite/gcc.target/arm/vect-div-2.c
@@ -0,0 +1,14 @@ 
+/* { dg-do compile } */
+/* { dg-require-effective-target arm_v8_neon_ok } */
+/* { dg-options "-O2 -funsafe-math-optimizations -fno-reciprocal-math -ftree-vectorize -fdump-tree-vect-all" } */
+/* { dg-add-options arm_v8_neon } */
+
+void
+foo (int len, float * __restrict p, float *__restrict x)
+{
+  len = len & ~31;
+  for (int i = 0; i < len; i++)
+    p[i] = p[i] / x[i];
+}
+
+/* { dg-final { scan-tree-dump-times "vectorized 0 loops" 1 "vect" } } */