Skip to content

[MIOpen] fix bn fwd training edge case #2193

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
bghimireamd wants to merge 9 commits into from
Closed

[MIOpen] fix bn fwd training edge case #2193

Show file tree
Hide file tree
Changes from 6 commits
Commits
Show all changes
9 commits
Select commit Hold shift + click to select a range
03f1067
fix bn fwd training edge case
bghimireamd Oct 21, 2025
e10a0ce
fix minor comment
bghimireamd Oct 21, 2025
b52a7e7
handle other variant too
bghimireamd Oct 21, 2025
eaa63bd
clang format
bghimireamd Oct 21, 2025
7a08adb
fix in hip variant 1
bghimireamd Oct 22, 2025
d015b5b
Merge branch 'develop' into users/bghimire/fix_bn_variance_issue
bghimireamd Nov 19, 2025
07825bf
addresss review comments
bghimireamd Nov 19, 2025
cecaa85
Merge branch 'users/bghimire/fix_bn_variance_issue' of github.com:ROC…
bghimireamd Nov 19, 2025
ecfade1
Merge branch 'develop' into users/bghimire/fix_bn_variance_issue
bghimireamd Nov 21, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
46 changes: 40 additions & 6 deletions projects/miopen/src/kernels/MIOpenBatchNormFwdTrainSpatial.cl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -141,7 +141,14 @@ MIOpenBatchNormFwdTrainSpatial(const __global _FLOAT* __restrict in,
gcn_reduce2(&mean, &variance, (_FLOAT_ACCUM)INHW, lcl_data_x, lcl_data_y, lid);
#endif

variance = mad(-mean, mean, variance);
if(MIO_BN_NHW == 1)
{
variance = 0;
}
else
{
variance = mad(-mean, mean, variance);
}
if(variance < 0)
{
variance = 0;
Expand Down Expand Up @@ -353,7 +360,14 @@ MIOpenBatchNormFwdTrainSpatial(const __global _FLOAT* __restrict in,
#endif

// REDUCTION COMPLETE ---------------------------
variance = mad(-mean, mean, variance);
if(MIO_BN_NHW == 1)
{
variance = 0;
}
else
{
variance = mad(-mean, mean, variance);
}
if(variance < 0)
{
variance = 0;
Expand Down Expand Up @@ -629,7 +643,14 @@ MIOpenBatchNormFwdTrainSpatialFinalMeanVariance(
gcn_reduce2(&mean, &variance, INHW, lcl_data_x, lcl_data_y, ylid + zlid * ygrp_sz);
#endif

variance = mad(-mean, mean, variance);
if(MIO_BN_NHW == 1)
{
variance = 0;
}
else
{
variance = mad(-mean, mean, variance);
}
variance = max(variance, (_FLOAT_PREC_C)0.);
invVariance = rsqrt(variance + (_FLOAT_PREC_C)epsilon);

Expand Down Expand Up @@ -842,8 +863,14 @@ MIOpenBatchNormFwdTrainSpatial(const __global _FLOAT* __restrict in,
local _FLOAT_ACCUM lcl_data_y[MIO_BN_LDSGCN_SIZE];
gcn_reduce2(&mean, &variance, (_FLOAT_ACCUM)INHW, lcl_data_x, lcl_data_y, lid);
#endif

variance = mad(-mean, mean, variance);
if(MIO_BN_NHW == 1)
{
variance = 0;
}
else
{
variance = mad(-mean, mean, variance);
}
if(variance < 0)
{
variance = 0;
Expand Down Expand Up @@ -965,7 +992,14 @@ __kernel void MIOpenBatchNormFwdTrainSpatial(const __global _FLOAT* __restrict i
gcn_reduce2(&mean, &variance, (_FLOAT_ACCUM)INHW, lcl_data_x, lcl_data_y, lid);
#endif

variance = mad(-mean, mean, variance);
if(MIO_BN_NHW == 1)
{
variance = 0;
}
else
{
variance = mad(-mean, mean, variance);
}
variance = variance > 0. ? variance : 0.;
invVariance = rsqrt(variance + (_FLOAT_PREC)epsilon);
pvscale = lcl_scale;
Expand Down
9 changes: 8 additions & 1 deletion projects/miopen/src/kernels/MIOpenBatchNormFwdTrainSpatialHIP.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -233,7 +233,14 @@ struct MIOpenBatchNormFwdTrainSpatialHIPImpl<1, FpType, FpPrecType, FpAccumType>
}

// REDUCTION COMPLETE ---------------------------
variance = fma(-mean, mean, variance);
if(mio_bn_config::nhw == 1)
Comment thread
EnricoDeg marked this conversation as resolved.
Outdated
{
variance = FpPrecType{0};
}
else
{
variance = fma(-mean, mean, variance);
}
if(variance < FpPrecType{0})
{
variance = FpPrecType{0};
Expand Down
9 changes: 4 additions & 5 deletions projects/miopen/src/kernels/batchnorm_functions.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -180,8 +180,7 @@
#if defined(__AMDGCN__) && \
!((defined(MIO_BN_GFX103X) && MIO_BN_GFX103X) || \
(defined(MIO_BN_GFX110X) && MIO_BN_GFX110X) || \
(defined(MIO_BN_GFX120X) && MIO_BN_GFX120X) || \
(defined(MIO_BN_GFX115X) && MIO_BN_GFX115X))
(defined(MIO_BN_GFX120X) && MIO_BN_GFX120X) || (defined(MIO_BN_GFX115X) && MIO_BN_GFX115X))
#define MIOPEN_USE_AMDGCN 1
#else
#define MIOPEN_USE_AMDGCN 0
Expand Down Expand Up @@ -559,7 +558,7 @@ static inline void running_stash(global _FLOAT_PREC_C* resultRunningMean,
mean, (_FLOAT_ACCUM)expAvgFactor, pvt_newRunMean); // newMean*factor + tmp
const _FLOAT_ACCUM_C adjust =
(_FLOAT_ACCUM_C)((MIO_BN_NHW == 1)
? variance
? (_FLOAT_ACCUM_C)0.0
: variance * ((_FLOAT_ACCUM)MIO_BN_NHW /
((_FLOAT_ACCUM)MIO_BN_NHW - (_FLOAT_ACCUM)1.0)));
resultRunningVariance[channel] =
Expand All @@ -585,7 +584,7 @@ static inline void running_stash_pa(global _FLOAT_PREC* resultRunningMean,
(_FLOAT_PREC)expAvgFactor,
pvt_newRunMean); // newMean*factor + tmp

const _FLOAT_PREC adjust = (MIO_BN_N == 1) ? variance : variance * (N / (N - 1.0));
const _FLOAT_PREC adjust = (MIO_BN_N == 1) ? (_FLOAT_PREC)0.0 : variance * (N / (N - 1.0));
resultRunningVariance[index] =
(1 - (_FLOAT_PREC)expAvgFactor) * *(resultRunningVariance + index) +
(_FLOAT_PREC)expAvgFactor * adjust;
Expand All @@ -607,7 +606,7 @@ static inline void running_stash_dyn(global _FLOAT_PREC* resultRunningMean,
resultRunningMean[channel] =
(_FLOAT_PREC)mad(mean, (_FLOAT_ACCUM)expAvgFactor, pvt_newRunMean); // newMean*factor + tmp
const _FLOAT_ACCUM adjust =
(_FLOAT_ACCUM)((inhw == 1) ? variance : variance * (1. / (1. - inhw)));
(_FLOAT_ACCUM)((inhw == 1) ? (_FLOAT_ACCUM)0.0 : variance * (1. / (1. - inhw)));
resultRunningVariance[channel] =
(_FLOAT_PREC)((1 - (_FLOAT_ACCUM)expAvgFactor) *
(_FLOAT_ACCUM)(*(resultRunningVariance + channel)) +
Expand Down
Loading