Merged
fix(l2_normalize): 恢复普通 FP32 开方精度并补齐 FP16 边界回归 #13500
fix(l2_normalize): 恢复普通 FP32 开方精度并补齐 FP16 边界回归 #13500
Merged
raoliang_sacAdd on 18 days ago
Total: 3 files changed+16-3
@@ -72,7 +72,8 @@ constexpr AscendC::Reg::CastTrait CAST_TRAIT_FROM_FP32_FP16{
72 AscendC::RoundMode::CAST_RINT};72 AscendC::RoundMode::CAST_RINT};
73 73 
74// 高精度模式(设计文档 §7「精度处理」认可的备选:ST 实测超差时切 0-ULP,不换数据流):74// 高精度模式(设计文档 §7「精度处理」认可的备选:ST 实测超差时切 0-ULP,不换数据流):
75-// FP16 uses 0-ULP FTZ_FALSE. FP32 uses the scaled norm helpers below.75+// FP16 and ordinary FP32 norms use 0-ULP FTZ_FALSE. FP32 high/low retries
76+// use the scaled 1-ULP FTZ_FALSE helpers below; FP32 Div remains 1-ULP FTZ_FALSE.
76// 默认 INTRINSIC Sqrt 对 subnormal 输入 FTZ77// 默认 INTRINSIC Sqrt 对 subnormal 输入 FTZ
77// 刷 0,eps<=0 且 x² 落 fp32 subnormal 窗口(|x|<1.08e-19)时 denom 被刷成 0 →78// 刷 0,eps<=0 且 x² 落 fp32 subnormal 窗口(|x|<1.08e-19)时 denom 被刷成 0 →
78// y=±inf,与 golden(numpy IEEE 保留 subnormal)NaN/Inf mismatch(FR-P1-001 L1_094)。79// y=±inf,与 golden(numpy IEEE 保留 subnormal)NaN/Inf mismatch(FR-P1-001 L1_094)。
@@ -439,7 +440,7 @@ __simd_vf__ inline void FinalizeSumVfImpl(__ubuf__ float* sPtr, __ubuf__ float*
439}440}
440 441 
441// S6+S7 PostElewise:Maxs(eps) + Sqrt 同一 VF(全 fp32 域;eps 钳在平方和上,§9.6)442// S6+S7 PostElewise:Maxs(eps) + Sqrt 同一 VF(全 fp32 域;eps 钳在平方和上,§9.6)
442-// Sqrt 恒 0-ULP FTZ_FALSE(IEEE 正确舍入,subnormal 输入不刷零,见 SQRT_MODE_0ULP 注释)。443+// FP16 input path only: Sqrt uses 0-ULP FTZ_FALSE on the FP32 intermediate.
443__simd_vf__ inline void PostElewiseVfImpl(__ubuf__ float* rootPtr, __ubuf__ float* denomPtr, uint32_t laneN, float eps,444__simd_vf__ inline void PostElewiseVfImpl(__ubuf__ float* rootPtr, __ubuf__ float* denomPtr, uint32_t laneN, float eps,
444 uint16_t repeatTime)445 uint16_t repeatTime)
445{446{
@@ -31,6 +31,12 @@ constexpr float SCALE_LOW_MARKER = -2.0f;
31constexpr uint32_t SCALE_LANES = 64;31constexpr uint32_t SCALE_LANES = 64;
32constexpr AscendC::Reg::DivSpecificMode SCALE_DIV_MODE{AscendC::Reg::MaskMergeMode::ZEROING, false,32constexpr AscendC::Reg::DivSpecificMode SCALE_DIV_MODE{AscendC::Reg::MaskMergeMode::ZEROING, false,
33 AscendC::DivAlgo::PRECISION_1ULP_FTZ_FALSE};33 AscendC::DivAlgo::PRECISION_1ULP_FTZ_FALSE};
34+// Retained finite norms in the first pass have sums in [2^-120, 2^124),
35+// below the fast-inverse sqrt limit 2^126. Keep their original 0-ULP mode.
36+// High/low lanes become retry markers; rescaled passes retain the full-range
37+// 1-ULP FTZ_FALSE mode independently of the ordinary-input precision choice.
38+constexpr AscendC::Reg::SqrtSpecificMode NORMAL_SQRT_MODE_0ULP{AscendC::Reg::MaskMergeMode::ZEROING, true,
39+ AscendC::SqrtAlgo::PRECISION_0ULP_FTZ_FALSE};
34constexpr AscendC::Reg::SqrtSpecificMode SCALE_SQRT_MODE{AscendC::Reg::MaskMergeMode::ZEROING, false,40constexpr AscendC::Reg::SqrtSpecificMode SCALE_SQRT_MODE{AscendC::Reg::MaskMergeMode::ZEROING, false,
35 AscendC::SqrtAlgo::PRECISION_1ULP_FTZ_FALSE};41 AscendC::SqrtAlgo::PRECISION_1ULP_FTZ_FALSE};
36 42 
@@ -50,7 +56,7 @@ __simd_vf__ inline void PrepareScaledNormVf(__ubuf__ float* sums, __ubuf__ float
50 Reg::Maxs(sum, sum, eps, mask);56 Reg::Maxs(sum, sum, eps, mask);
51 Reg::Compares<float, CMPMODE::GE>(high, sum, SCALE_SUM_HIGH, mask);57 Reg::Compares<float, CMPMODE::GE>(high, sum, SCALE_SUM_HIGH, mask);
52 Reg::Compares<float, CMPMODE::LT>(low, sum, SCALE_SUM_LOW, mask);58 Reg::Compares<float, CMPMODE::LT>(low, sum, SCALE_SUM_LOW, mask);
53- Reg::Sqrt<float, &SCALE_SQRT_MODE>(norm, sum, mask);59+ Reg::Sqrt<float, &NORMAL_SQRT_MODE_0ULP>(norm, sum, mask);
54 Reg::Select(norm, highMarker, norm, high);60 Reg::Select(norm, highMarker, norm, high);
55 Reg::Select(norm, lowMarker, norm, low);61 Reg::Select(norm, lowMarker, norm, low);
56 Reg::StoreAlign(denom + off, norm, mask);62 Reg::StoreAlign(denom + off, norm, mask);
@@ -27,3 +27,9 @@ l2_fp16_group_taila,,l2_normalize,"((65537, 2),)","('float16',)","('ND',)","((65
27l2_fp32_eps_negative,,l2_normalize,"((3,),)","('float32',)","('ND',)","((3,),)","('float32',)","('ND',)","((3,),)","('ND',)","((3,),)","('ND',)","{'axis': [0], 'eps': -1.0}","((1, 2),)",,1e-08,,,True,Small numerical stability/Group regression,,0,,(),(),"((3,),)"27l2_fp32_eps_negative,,l2_normalize,"((3,),)","('float32',)","('ND',)","((3,),)","('float32',)","('ND',)","((3,),)","('ND',)","((3,),)","('ND',)","{'axis': [0], 'eps': -1.0}","((1, 2),)",,1e-08,,,True,Small numerical stability/Group regression,,0,,(),(),"((3,),)"
28l2_fp32_eps_max,,l2_normalize,"((3,),)","('float32',)","('ND',)","((3,),)","('float32',)","('ND',)","((3,),)","('ND',)","((3,),)","('ND',)","{'axis': [0], 'eps': 3.4028234663852886e+38}","((1e+18, 1e+18),)",,1e-08,,,True,Small numerical stability/Group regression,,0,,(),(),"((3,),)"28l2_fp32_eps_max,,l2_normalize,"((3,),)","('float32',)","('ND',)","((3,),)","('float32',)","('ND',)","((3,),)","('ND',)","((3,),)","('ND',)","{'axis': [0], 'eps': 3.4028234663852886e+38}","((1e+18, 1e+18),)",,1e-08,,,True,Small numerical stability/Group regression,,0,,(),(),"((3,),)"
29l2_fp32_subnormal,,l2_normalize,"((3,),)","('float32',)","('ND',)","((3,),)","('float32',)","('ND',)","((3,),)","('ND',)","((3,),)","('ND',)","{'axis': [0], 'eps': 0.0}","((1.401298464324817e-45, 1.401298464324817e-45),)",,1e-08,,,True,Small numerical stability/Group regression,,0,,(),(),"((3,),)"29l2_fp32_subnormal,,l2_normalize,"((3,),)","('float32',)","('ND',)","((3,),)","('float32',)","('ND',)","((3,),)","('ND',)","((3,),)","('ND',)","{'axis': [0], 'eps': 0.0}","((1.401298464324817e-45, 1.401298464324817e-45),)",,1e-08,,,True,Small numerical stability/Group regression,,0,,(),(),"((3,),)"
30+l2_fp16_finite_extremes,UNKNOWN,l2_normalize,"((2, 4),)","('float16',)","('ND',)","((2, 4),)","('float16',)","('ND',)","((2, 4),)","('ND',)","((2, 4),)","('ND',)","{'axis': [1], 'eps': 0.0, 'x': [[65504.0,-65504.0,0.0,2**-24],[2**-24,-2**-24,2**-14,-2**-14]]}","((-1,1),)",,,(),(),True,FP16 special values and FP32 normal sqrt precision regression,,0,,(),(),"((2, 4),)"
31+l2_fp16_mixed_nonfinite,UNKNOWN,l2_normalize,"((3, 4),)","('float16',)","('ND',)","((3, 4),)","('float16',)","('ND',)","((3, 4),)","('ND',)","((3, 4),)","('ND',)","{'axis': [1], 'eps': 0.0, 'x': [[1.0,float('inf'),-2.0,0.0],[-float('inf'),1.0,-1.0,0.0],[1.0,float('nan'),0.0,-1.0]]}","((-1,1),)",,,(),(),True,FP16 special values and FP32 normal sqrt precision regression,,0,,(),(),"((3, 4),)"
32+l2_fp16_empty_nonfinite,UNKNOWN,l2_normalize,"((8,),)","('float16',)","('ND',)","((8,),)","('float16',)","('ND',)","((8,),)","('ND',)","((8,),)","('ND',)","{'axis': [], 'eps': 0.0001, 'x': [float('inf'),-float('inf'),float('nan'),0.0,-0.0,2**-24,65504.0,-65504.0]}","((-1,1),)",,,(),(),True,FP16 special values and FP32 normal sqrt precision regression,,0,,(),(),"((8,),)"
33+l2_fp32_identity_normal,UNKNOWN,l2_normalize,"((8,),)","('float32',)","('ND',)","((8,),)","('float32',)","('ND',)","((8,),)","('ND',)","((8,),)","('ND',)","{'axis': [], 'eps': 0.0, 'x': [-841.525634765625, -460.3163757324219, -0.10450954735279083, -0.003607991384342313, -817.0941162109375, -0.23848676681518555, -0.7906009554862976, 13.625178337097168]}","((-1,1),)",,,(),(),True,FP16 special values and FP32 normal sqrt precision regression,,0,,(),(),"((8,),)"
34+l2_fp16_group_inf,UNKNOWN,l2_normalize,"((65537, 2),)","('float16',)","('ND',)","((65537, 2),)","('float16',)","('ND',)","((65537, 2),)","('ND',)","((65537, 2),)","('ND',)","{'axis': [0], 'eps': 0.0, 'x': [[float('inf'),1.0]]}","((-1,1),)",,,(),(),True,FP16 special values and FP32 normal sqrt precision regression,,0,,(),(),"((65537, 2),)"
35+l2_fp16_group_nan,UNKNOWN,l2_normalize,"((65537, 2),)","('float16',)","('ND',)","((65537, 2),)","('float16',)","('ND',)","((65537, 2),)","('ND',)","((65537, 2),)","('ND',)","{'axis': [0], 'eps': 0.0, 'x': [[float('nan'),1.0]]}","((-1,1),)",,,(),(),True,FP16 special values and FP32 normal sqrt precision regression,,0,,(),(),"((65537, 2),)"
atomgit-bot
atomgit-botatomgit-bot18 days ago

🟡 Medium Priority

改动行:CSV 第 30–35 行新增 6 条用例,第 16 列 precision_tolerances 与第 17 列 absolute_precision 全为空(,,,(),() 段),而同文件其余 33 条用例(第 2–11 行 fp16 填 1.52587890625e-5,第 12–29 行填 1e-08)均显式给出 absolute_precision。本文件所有用例的 precision_tolerances 都是空的,也就是说 absolute_precision 是这套 CSV 唯一的容差来源。

影响契约:第 33 行 l2_fp32_identity_normal(axis=[]、eps=0、8 个冻结元素)正是本 PR 用来锁定“普通 FP32 首遍 Sqrt 由 1-ULP 改回 0-ULP”的回归用例,其输出期望为精确的 ±1。

失效模式:若容差为空、由 TTK 按 dtype 默认(fp32 通常为 1e-4 量级相对误差)判定,则把 PrepareScaledNormVf 的 sqrt 模式改回 SCALE_SQRT_MODE(1-ULP)后,输出相对精确 ±1 的偏差仅 1.1920928955078125e-7(issue #8024 实测最大绝对差),仍会判 PASS——即这条“回归”用例恰恰检不出它要防的回归。反之若按同文件惯例填 absolute_precision=1e-08,1.19e-7 > 1e-8,回归会被正确捕获。同理第 30/32 行 FP16 用例(含 2^-24 次正规、65504 边界)也缺少 fp16 侧 1.52587890625e-5 的量级约束。

建议:为新增 6 条用例补第 17 列 absolute_precision:第 33 行(fp32)填 1e-08;第 30、31、32、34、35 行(fp16)填 1.52587890625e-5,与第 2–29 行既有写法一致,确保新回归门限严于 1 ULP。

likedislike
Incorrect?