Merged
fix(bn3d_training_reduce): preserve FP16 full-R slot merge precision #13622
fix(bn3d_training_reduce): preserve FP16 full-R slot merge precision #13622
Merged
raoliang_sacAdd on 18 days ago
Total: 2 files changed+4-3
@@ -344,9 +344,9 @@ private:
344 } else {344 } else {
345 AccumulateChannelFull(c, accUb);345 AccumulateChannelFull(c, accUb);
346 }346 }
347- // FP16 sub-R 的多个搬运块使用独立累加槽,槽间合并也必须保留舍入低位。347+ // FP16 sub-R 与普通 full-R 多 tile 的槽间合并均需保留新产生的舍入低位。
348 if constexpr (IsSameType<T_X, half>::value) {348 if constexpr (IsSameType<T_X, half>::value) {
349- if (isSubR_ != 0) {349+ if (isSubR_ != 0 || !C0_PACKED) {
350 FoldAccSlots<false>(accUb);350 FoldAccSlots<false>(accUb);
351 } else {351 } else {
352 FoldAccSlots(accUb);352 FoldAccSlots(accUb);
@@ -1092,7 +1092,7 @@ private:
1092 // value = acc - comp;槽间主值相加也会产生新舍入。BF16 / FP32 路径用1092 // value = acc - comp;槽间主值相加也会产生新舍入。BF16 / FP32 路径用
1093 // TwoSum 把该误差同步并入 comp;旧实现只做 Σacc - Σcomp,会在大输出的临界1093 // TwoSum 把该误差同步并入 comp;旧实现只做 Σacc - Σcomp,会在大输出的临界
1094 // 舍入点丢 1 ULP。FP16 sub-R 配合 K=2 的短块同样保留槽间舍入误差;1094 // 舍入点丢 1 ULP。FP16 sub-R 配合 K=2 的短块同样保留槽间舍入误差;
1095- // FP16 full-R 和 split-reduce 保持原有折叠策略。1095+ // 普通 FP16 full-R 同样保留该误差;C0 full-R 和 split-reduce 保持原策略。
1096 // activeSlots 每趟减半。1096 // activeSlots 每趟减半。
1097 // 趟数由 Host 算好(foldPasses_ = log2(numAccSlots_))传进来,__VEC_SCOPE__ 内1097 // 趟数由 Host 算好(foldPasses_ = log2(numAccSlots_))传进来,__VEC_SCOPE__ 内
1098 // 没有任何数据依赖分支;activeSlots 逐趟减半是纯标量运算,同族 BinaryAddVF 同样写法。1098 // 没有任何数据依赖分支;activeSlots 逐趟减半是纯标量运算,同族 BinaryAddVF 同样写法。
@@ -14,3 +14,4 @@ bn3d_ncdhw_unknown_rank_ndhwc_outputs,bn3d_training_reduce,"((2,5,2,3,4),)","('f
14bn3d_fp16_short_r0_one_square_sum,bn3d_training_reduce,"((53,5),)","('float16',)","('NCDHW',)","((53,5),)","('NCDHW',)","((5,),(5,))","('float32','float32')","('NCDHW','NCDHW')","((-1000,-10),)","((53,-1),)",{},FP16 square-sum compensation across short rows14bn3d_fp16_short_r0_one_square_sum,bn3d_training_reduce,"((53,5),)","('float16',)","('NCDHW',)","((53,5),)","('NCDHW',)","((5,),(5,))","('float32','float32')","('NCDHW','NCDHW')","((-1000,-10),)","((53,-1),)",{},FP16 square-sum compensation across short rows
15BN3DTrainingReduce_L1_531,bn3d_training_reduce,"((12, 17, 12, 20, 10),)","('float16',)","('NDHWC',)","((12, 17, 12, 20, 10),)","('NDHWC',)","((10,),(10,),)","('float32','float32',)","('NDHWC', 'NDHWC')","((-1,1,),)","((12, -1, 12, -1, 10),)",{},FP16 fused long-channel square-sum block rounding; run with --seed 4215BN3DTrainingReduce_L1_531,bn3d_training_reduce,"((12, 17, 12, 20, 10),)","('float16',)","('NDHWC',)","((12, 17, 12, 20, 10),)","('NDHWC',)","((10,),(10,),)","('float32','float32',)","('NDHWC', 'NDHWC')","((-1,1,),)","((12, -1, 12, -1, 10),)",{},FP16 fused long-channel square-sum block rounding; run with --seed 42
16bn3d_fp16_fused_long_ncdhw_square_sum,bn3d_training_reduce,"((12, 10, 17, 12, 20),)","('float16',)","('NCDHW',)","((12, 10, 17, 12, 20),)","('NCDHW',)","((10,),(10,),)","('float32','float32',)","('NCDHW', 'NCDHW')","((-1,1,),)","((12, 10, -1, 12, -1),)",{},FP16 fused long-channel square-sum block rounding; run with --seed 4216bn3d_fp16_fused_long_ncdhw_square_sum,bn3d_training_reduce,"((12, 10, 17, 12, 20),)","('float16',)","('NCDHW',)","((12, 10, 17, 12, 20),)","('NCDHW',)","((10,),(10,),)","('float32','float32',)","('NCDHW', 'NCDHW')","((-1,1,),)","((12, 10, -1, 12, -1),)",{},FP16 fused long-channel square-sum block rounding; run with --seed 42
17+BN3DTrainingReduce_L1_149,bn3d_training_reduce,"((83,6,46,34,),)","('float16',)","('NCDHW',)","((83,6,46,34,),)","('NCDHW',)","((6,),(6,),)","('float32','float32',)","('NCDHW','NCDHW',)","((0.01,1,),)","((-1,-1,-1,-1,),)",{},FP16 full-R multi-slot sum rounding regression; run with --seed 42