[Beignet] [PATCH] libocl: Directly scalarize built-in with vector input.
Zhigang Gong
zhigang.gong at linux.intel.com
Wed Feb 25 00:04:48 PST 2015
This patch LGTM. Actually, after the sin/cos functions have been optimized,
we don't need to fold those vector operations into a for loop.
Will push it latter. Thanks.
On Sun, Feb 15, 2015 at 04:09:07PM +0800, Ruiling Song wrote:
> This revert the following commit:
> "Re-apply "improve the build performance of vector type built-in function.""
> commitId: 06cce8178649759e12a3a353f0550189d371871b.
>
> I finally decide to do this because although below kind of program has less
> instructions and less compile-time, but it will also introduce extra memory access,
> which would cause bad run-time performance if the loop is not unrolled. If the loop
> is unrolled, it would be similar like scalarized version.
>
> OVERLOADABLE float16 func (float16 param0)
> {
> union{
> float va[16];
> float16 vv16;
> }uret;
> union{
> float pa[16];
> float16 pv16;
> }usrc0;
> usrc0.pv16 = param0;
> for(int i =0; i < 16; i++)
> uret.va[i] = func(usrc0.pa[i]);
> return uret.vv16;
> }
>
> I did some experiment on the affected built-in. I fixed the GPU frequency at 1050,
> and increase input data to 862000. The result is like below (obviously the scalarized
> version has better performance):
>
> bultin_asinh_float16:
> loop version: 200ms
> scalarized version: 150ms
>
> builtin_sinh_float16:
> loop version: 250ms
> scalarized version: 160ms
>
> And also this patch would reduce the generation of large integer. Although we support
> large integer legalization, I find sometime it is hard to legalize in very efficient way
> like large integer LE/GT.
> ---
> backend/src/libocl/script/gen_vector.py | 45 +++++--------------------------
> 1 file changed, 6 insertions(+), 39 deletions(-)
>
> diff --git a/backend/src/libocl/script/gen_vector.py b/backend/src/libocl/script/gen_vector.py
> index 291dd87..ffc573a 100755
> --- a/backend/src/libocl/script/gen_vector.py
> +++ b/backend/src/libocl/script/gen_vector.py
> @@ -289,42 +289,9 @@ class builtinProto():
> formatStr += ';'
> self.append(formatStr)
> return formatStr
> - if self.functionName != 'select' and ptypeSeqs[0] == ptypeSeqs[self.paramCount-1] and ptype[1] > 4:
> - formatStr += '\n{ \n union{'
> - formatStr = self.append(formatStr, ' {0} va[{1}];'.format(vtype[0], vtype[1]))
> - formatStr = self.append(formatStr, ' {0}{1} vv{2};'.format(vtype[0], vtype[1], vtype[1]))
> - formatStr += '\n }uret;'
> - formatStr += '\n union{'
> - formatStr = self.append(formatStr, ' {0} pa[{1}];'.format(ptype[0], ptype[1]))
> - formatStr = self.append(formatStr, ' {0}{1} pv{2};'.format(ptype[0], ptype[1], ptype[1]))
> - formatStr += '\n }'
> - for n in range(0, self.paramCount):
> - formatStr += 'usrc{0}'.format(n)
> - if n+1 != self.paramCount:
> - formatStr +=', '
> - formatStr += ';'
> -
> - for n in range(0, self.paramCount):
> - formatStr = self.append(formatStr, ' usrc{0}.pv{1} = param{2};'.format(n, ptype[1], n))
> - formatStr = self.append(formatStr, ' for(int i =0; i < {0}; i++)'.format(ptype[1]))
> - formatStr += '\n uret.va[i] = '
> - if self.prefix == 'relational' and self.functionName != 'bitselect' and self.functionName != 'select':
> - formatStr += '-'
> - formatStr += '{0}('.format(self.functionName)
> -
> - for n in range(0, self.paramCount):
> - formatStr += 'usrc{0}.pa[i]'.format(n)
> - if n+1 != self.paramCount:
> - formatStr +=', '
> - formatStr += ');'
> - formatStr = self.append(formatStr, ' return uret.vv{0};'.format(vtype[1]))
> - formatStr += '\n}'
> - formatStr = self.append(formatStr)
> - return formatStr
> - else:
> - formatStr = self.append(formatStr, '{{return ({0}{1})('.format(vtype[0], vtype[1]))
> - self.indent = len(formatStr)
> - for j in range(0, vtype[1]):
> + formatStr = self.append(formatStr, '{{return ({0}{1})('.format(vtype[0], vtype[1]))
> + self.indent = len(formatStr)
> + for j in range(0, vtype[1]):
> if (j != 0):
> formatStr += ','
> if (j + 1) % 2 == 0:
> @@ -359,10 +326,10 @@ class builtinProto():
>
> formatStr += ')'
>
> - formatStr += '); }\n'
> - self.append(formatStr)
> + formatStr += '); }\n'
> + self.append(formatStr)
>
> - return formatStr
> + return formatStr
>
> def output(self):
> for line in self.outputStr:
> --
> 1.7.10.4
>
> _______________________________________________
> Beignet mailing list
> Beignet at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/beignet
More information about the Beignet
mailing list