49 KiB

Raw Permalink Blame History

子赛题一、评测数据集生成挑战赛

欢迎参加 评测数据集生成挑战赛 📊！本比赛旨在构建一个标准化、可用于 GPU 性能测试的评测数据集，帮助开发者更高效地比较不同硬件和框架的性能表现。

💻 比赛背景简介

在 AI 模型开发和部署中，GPU 性能评测是一个非常重要的环节。不同 GPU、不同深度学习框架（如 PyTorch、TensorFlow、PaddlePaddle 等）在运行相同任务时，速度、吞吐量、内存占用等表现差异很大。本次挑战赛希望通过社区的力量，构建一个标准化、带权重的评测数据集，让 GPU 性能比较更加科学、公正。

🎯 比赛目标

从 PyTorch、PaddlePaddle、TensorFlow、Jax、MMCV、Transformers 等框架中收集并生成评测样本。
为每个样本提供标准输出和性能指标，确保结果可复现。
最终形成 GPU CodeForces 数据集和评价方法。

初次了解本类比赛的小伙伴可以查看以下两份文档，希望帮助你快速入门和上手：赛题入门、代码解读。

✅ 算子实现状态

该清单由 scripts/update_operator_checklist.py 根据 scripts/operator_targets.txt 和 S1 codes/ 自动生成；新增待实现算子请写入目标清单，新增实现目录后运行 python scripts/update_operator_checklist.py --sync-targets 即可自动勾选。

已实现：578
未实现：0
跟踪总数：578

展开查看算子实现状态

状态	算子	实现目录
[x]	aconc	hli28146_#25
[x]	ActorCriticLoss	uucoco_#74
[x]	adaptive_maxpool1d	ZZZJ_#83
[x]	adaptive_maxpool3d	ZZZJ_#85
[x]	Adaptivepiecewiselinear	gsd123_#32
[x]	AdvantageLoss	uucoco_#75
[x]	AdversarialLoss	uucoco_#76
[x]	Affine+ReLU6 融合	Ljy123_#13
[x]	affine_grid3d	ZZZJ_#173
[x]	affineleakyreluclamp	uucoco_#77
[x]	AHAF	gsd123_#33
[x]	AHerfReLU	hli28146_#86
[x]	alpha_iou	ZZZJ_#16
[x]	alphablend	ZZZJ_#174
[x]	angle	hli28146_#35
[x]	AngularDistance	gsd123_#62
[x]	AngularLoss	uucoco_#57
[x]	AOAF	hli28146_#102
[x]	APALU	hli28146_#112
[x]	AQuLU	hli28146_#105
[x]	arcfaceloss	hli28146_#22
[x]	ArcSinh-Affine-Gate	Ljy123_#116
[x]	ArcSinTanh-Affine-Gate	Ljy123_#109
[x]	Arctan-Affine-Gate	Ljy123_#97
[x]	ARiA2	hli28146_#66
[x]	Atan-Sigmoid-Mix-Gate	LJy123_#50
[x]	balanced_softmax_loss	hli28146_#46
[x]	ball_query	ZZZJ_#175
[x]	batchnorm1d	Icy_Cola10
[x]	bce	wut0n_#8
[x]	bcewithlogitsloss	ZZZJ 14
[x]	BehaviorCloningLoss	uucoco_#79
[x]	BellmanLoss	uucoco_#80
[x]	BetaDivergenceLoss	uucoco_#81
[x]	BhattacharyyaDistance	uucoco_#5
[x]	Bias+GELU（tanh 近似）	Ljy123_#10
[x]	Bias+SiLU（Swish）融合	Ljy123_#11
[x]	bilinear	zizi05_#6
[x]	BilinearGLU	uucoco_#34
[x]	Bipolar	uucoco_#31
[x]	BipolarSigmoid	uucoco_#32
[x]	black_scholes	ZZZJ_#135
[x]	blurpool	ZZZJ_#134
[x]	box_area	ZZZJ_#18
[x]	box_corner_to_center	ZZZJ_#19
[x]	box_iou	ZZZJ_#17
[x]	boxfilter	ZZZJ_#21
[x]	braycurtis	wut0n_#22
[x]	braycurtis_adaptive_triplet	wut0n_#104
[x]	bregman_divergence_gelu	uucoco_#117
[x]	BregmanDivergenceLoss	uucoco_#82
[x]	BReLU	uucoco_#33
[x]	broadcast_tensors	ZZZJ_#180
[x]	bucketize	ZZZJ_#181
[x]	BYOLLoss	gsd123_#105
[x]	canberra	wut0n_#21
[x]	canberra_focalloss	wut0n_#88
[x]	CanberraDistance	gsd123_#26
[x]	cartesian_prod	ZZZJ_#182
[x]	causal_mask	ZZZJ_#183
[x]	CELU-Affine-Gate	Ljy123_#101
[x]	CenterNetLoss	uucoco_#118
[x]	Channel RMS Gate	Ljy123_#35
[x]	Channel-Softmax-Affine gating	Ljy123_#33
[x]	channel_permute	ZZZJ_#130
[x]	channelmeangate	gsd123_#84
[x]	ChannelShuffle	uucoco_#8
[x]	CharbonnierLoss	uucoco_#10
[x]	chebyshev	wut0n_#16
[x]	chebyshev_abs_square	uucoco_#119
[x]	chebyshev_hardswish	wut0n_#105
[x]	chebyshev_leakyrelu	wut0n_#59
[x]	chebyshev_sigmoid	wut0n_#42
[x]	chebyshevaffine	gsd123_#87
[x]	ChebyshevDistance	gsd123_#29
[x]	cholesky	hli28146_#7
[x]	circleloss	gsd123 18 gsd123_#9
[x]	circularpad1d	ZZZJ_#184
[x]	circularpad2d	ZZZJ_#185
[x]	circularpad3d	ZZZJ_#186
[x]	clarksDistance	gsd123_#28
[x]	cmyk_to_rgb	ZZZJ_#50
[x]	Colu	gsd123_#34
[x]	column_stack	ZZZJ_#65
[x]	ComboLoss	uucoco_#58
[x]	complex_abs_angle_polar	uucoco_#83
[x]	complex_conj_mul_div	uucoco_#84
[x]	complex_exp_log_power	uucoco_#85
[x]	complex_mul	ZZZJ_#66
[x]	constantpad3d	ZZZJ_#189
[x]	contrastiveloss	gsd123_#22 wut0n_#11
[x]	conv2d	Icy_cola 11
[x]	Cos-Affine-Gate	Ljy123_#43 Ljy123_#82
[x]	cosfaceloss	hli28146_#21
[x]	cosine	wut0n_#19
[x]	Cosine-Affine-Gate	Ljy123_#113
[x]	Cosine-Square-Affine-Gate	Ljy123_#49
[x]	cosine_swish_gelu	uucoco_#120
[x]	cosinedistance	wut0n_#15
[x]	cosinedistance_softmax	wut0n_#58
[x]	CosineEmbeddingLoss	gsd123 37 gsd123_#15
[x]	cosineloss	ZZZJ_#1
[x]	cosinesimilarity	wwmm_#2
[x]	CoVariance	uucoco_#12
[x]	crop_resize	ZZZJ_#190
[x]	cross	ZZZJ_#164
[x]	cross_layer_norm	ZZZJ_#136
[x]	CrossEntropyDiceLoss	uucoco_#59
[x]	CrossEntropyLoss	gsd123_#6
[x]	CRReLU	hli28146_#115
[x]	cumsum	hli28146_#11
[x]	DecayingSineUnit	uucoco_#29
[x]	deepnorm	hli28146_#32
[x]	depthwise_conv1d	ZZZJ_#74
[x]	dequantize_fp4	ZZZJ_#86
[x]	dequantize_int8	ZZZJ_#87
[x]	dequantize_linear	ZZZJ_#88
[x]	dice	wut0n_#6
[x]	dice_bce	wut0n_#110
[x]	dice_from_2d	wut0n_#108
[x]	DiceLoss	gsd123_#17
[x]	DiceSimilarity	gsd123_#66
[x]	digitization	ZZZJ_#191
[x]	Dilation1d	ZZZJ_#107
[x]	Dilation2d	ZZZJ_#108
[x]	DistillationLoss	gsd123_#108
[x]	dot_mse_tanh	uucoco_#121
[x]	DoubleGLU	uucoco_#44
[x]	dropblock1d	ZZZJ_#192
[x]	DSiLU	hli28146_#107
[x]	DSReLU	hli28146_#131
[x]	Dual-Input Contrastive Gate	Ljy123_#34
[x]	dw_transpose	ZZZJ_#54
[x]	dw_transpose2d	ZZZJ_#55
[x]	dw_transpose3d	ZZZJ_#56
[x]	EIS1	hli28146_#128
[x]	EIS2	hli28146_#129
[x]	EIS3	hli28146_#130
[x]	ELU-Affine-Gate	Ljy123_#89
[x]	ELUGLU	uucoco_#46
[x]	embeddingbag	wwmm_#3
[x]	EMDLoss	gsd123_#109
[x]	Erf-Affine-Gate	Ljy123_#98
[x]	ErfAct	hli28146_#91
[x]	ErfReLU	hli28146_#87
[x]	Erosion1d	ZZZJ_#110
[x]	Erosion2d	ZZZJ_#111
[x]	Esh	hli28146_#84
[x]	euclidean	wut0n_#14
[x]	euclidean_erfc	uucoco_#122
[x]	evonorm	gsd123 13 gsd123_#5
[x]	Exp-Square Affine Gate	Ljy123_#42
[x]	Expm1+Affine	Ljy123_#29
[x]	Expm1-Sigmoid-Affine-Gate	Ljy123_#51
[x]	expnormalizelog	uucoco_#86
[x]	ExponentialLinear-Affine-Gate	Ljy123_#120
[x]	fake_quantize_per_channel_affine	hli28146_#37
[x]	fake_quantize_per_tensor_affine	hli28146_#38
[x]	farthest_point_sampling	ZZZJ_#116
[x]	FastAPLoss	uucoco_#123
[x]	FDivergenceLoss	uucoco_#87
[x]	finite_difference	ZZZJ_#195
[x]	fisherrao_rmsnorm	uucoco_#124
[x]	FlattenT	gsd123_#53
[x]	flip_horizontal	ZZZJ_#146
[x]	fma_activation	wut0n_#36
[x]	focal_eiou	ZZZJ_#15
[x]	FocalLoss	gsd123_#18 wut0n_#2
[x]	focalloss_fused	wut0n_#31
[x]	focalloss_labelsmoothing	wut0n_#109
[x]	focalloss_reduction	wut0n_#40
[x]	focalloss_sigmoid	wut0n_#107
[x]	focaltverskyloss	hli28146_#58 uucoco_#61
[x]	fold	ZZZJ_#196
[x]	fourieraffine	gsd123_#88
[x]	FPFLU	hli28146_#127
[x]	FReLU	gsd123_#45
[x]	fresnel_schlick	ZZZJ_#144
[x]	FTS	gsd123_#55 hli28146_#24
[x]	FunnelActivationforVisualRecognition	gsd123_#46
[x]	fused_adam_step	ZZZJ_#121
[x]	fused_rmsprop_step	ZZZJ_#145
[x]	gamma_correction	ZZZJ_#197
[x]	GammaDivergenceLoss	uucoco_#88
[x]	gateblendnormalize	uucoco_#89
[x]	gather_elements	ZZZJ_#198
[x]	Gaussian+Bias	Ljy123_#22
[x]	Gaussian-Affine-Gate	Ljy123_#100
[x]	gaussian_blur	ZZZJ_#113
[x]	gaussian_filter_2d	ZZZJ_#114
[x]	gaussian_pdf	ZZZJ_#120
[x]	GaussianNLLLoss	hli28146_#113 uucoco 31 uucoco_#1
[x]	GDL	hli28146_#55
[x]	geglu	geglu_sample
[x]	GELU-Affine-Gate	Ljy123_#94
[x]	gelu_dropout	Ljy123_#3
[x]	gempool	hli28146_#12
[x]	GeneratorLoss	gsd123_#19
[x]	giouloss	hli28146_#13
[x]	global_average_pooling	ZZZJ_#118
[x]	global_response_normalization	ZZZJ_#119
[x]	GLU（Gated Linear Unit）融合	Ljy123_#12
[x]	Gompertz-Affine-Gate	Ljy123_#123
[x]	gowerdistance	uucoco_#50
[x]	GradientClip	gsd123_#67
[x]	gridsample1d	ZZZJ_#200
[x]	Group-Softmax-Affine gating	Ljy123_#37
[x]	groupnorm	wut0n_#5 ZZZJ2
[x]	GrowingCosineUnit	uucoco_#35
[x]	GTU（tanh(a) * sigmoid(b)）	Ljy123_#19
[x]	GumbelCDF	hli28146_#82
[x]	GumbelPDF	hli28146_#83
[x]	hamming	wut0n_#20
[x]	hamming_gelu	gsd123_#139
[x]	hamming_relu	gsd123_#140
[x]	hamming_sigmoid	gsd123_#141
[x]	hamming_swish	gsd123_#142
[x]	hamming_xor_and	uucoco_#125
[x]	HammingDistance	uucoco_#4
[x]	hardbootstrappingloss	hli28146_#42
[x]	HardELiSH	uucoco_#36
[x]	hardmish	gsd123_#36 zizi05_#8
[x]	hardshrink	uucoco_#28
[x]	HardShrink-Affine-Gate	Ljy123_#104
[x]	HardSigmoid	uucoco_#27
[x]	Hardsigmoid+Affine	Ljy123_#23
[x]	HardSigmoid-Affine-Gate	Ljy123_#78
[x]	HardSReLUE	hli28146_#110
[x]	HardSwish	uucoco_#26
[x]	Hardswish-Affine-Gate	Ljy123_#48 Ljy123_#95
[x]	hardswishgate	gsd123_#89
[x]	HardTanh	uucoco_#25
[x]	HardTanh-Affine-Gate	Ljy123_#87
[x]	hardtanhgate	gsd123_#90
[x]	haversine_distance	ZZZJ_#117
[x]	hellinger_bhattacharyya	gsd123_#147
[x]	hellinger_gelu	gsd123_#148
[x]	HellingerDistance	uucoco_#6
[x]	hexpo	hli28146_#68
[x]	hingeembeddingloss	ZZZJ 15
[x]	hingeloss	wut0n_#13
[x]	HistogramLoss	uucoco_#90
[x]	Huber-Affine-Gate	Ljy1234_#107
[x]	huber_loss_tukey_biweight	gsd123_#113
[x]	huberloss	wut0n_#12
[x]	HungarianLoss	uucoco_#126
[x]	image_normalize	ZZZJ_#68
[x]	ImitationLearningLoss	uucoco_#91
[x]	infonceloss	gsd123 19 gsd123_#8
[x]	Instancenorm	wut0n_#4
[x]	instancenorm_dropout	wut0n_#34
[x]	instancenorm_relu	wut0n_#35
[x]	IntraClassCorrelation	uucoco_#11
[x]	inverse_lerp	ZZZJ_#148
[x]	InverseReinforcementLearningLoss	uucoco_#92
[x]	InverseSquare-Affine-Gate	Ljy123_#122
[x]	InvMultiquadratic	gsd123_#48
[x]	iou_tanh	uucoco_#127
[x]	IOULoss	uucoco_#17
[x]	IpLU	hli28146_#95
[x]	Isigmoid	hli28146_#124
[x]	ISRLU	gsd123_#49 hli28146_#72
[x]	ISRU	gsd123_#50
[x]	ItakuraSaitoDistanceLoss	uucoco_#93
[x]	jaccard_dice_sqrt	uucoco_#128
[x]	jaccard_legendre	uucoco_#129
[x]	JaccardSimilarity	uucoco_#9
[x]	jaro_winkler_softmax	uucoco_#130
[x]	jensenshannon_groupnorm	uucoco_#131
[x]	jsdivergence	hli28146_#31
[x]	kldiv_jsdiv_swish	uucoco_#132
[x]	kldivloss	gsd123_#24 ZZZJ 29
[x]	KulczynskiIndex	gsd123_#69
[x]	kullbackleibler_layernorm	uucoco_#133
[x]	l1	wut0n_#9
[x]	l1_fused	wut0n_#45
[x]	l1loss	ZZZJ 26
[x]	l2_normalize	ZZZJ_#126
[x]	LaLU	hli28146_#104
[x]	laplacian	ZZZJ_#127
[x]	Laplacian-Affine-Gate	Ljy123_#105
[x]	laplacian_filter	ZZZJ_#128
[x]	layernorm	Ljy123_#1 wut0n_#1
[x]	LDAMLoss	hli28146_#45
[x]	Leaky-Swish+Bias	Ljy123_#31
[x]	LeakyReGLU	uucoco_#37
[x]	LeakyReLU-Affine-Gate	Ljy123_#79
[x]	LeCunTanh	uucoco_#38
[x]	legendreaffine	gsd123_#91
[x]	linear_gelu	HHyy 43
[x]	lisht	hli28146_#19
[x]	localresponsenorm	ZZZJ 8
[x]	Log1p(Square)+Bias	Ljy123_#27
[x]	Log1pAbs-Affine-Gate	LJy123_#115
[x]	logbeta	wut0n_#29
[x]	LogCosh-Affine-Gate	Ljy123_#106 Ljy123_#58
[x]	logcoshdiceloss	hli28146_#56
[x]	LogCoshLoss	Lwh20070813 42 Lwh20070813_#2
[x]	logdet	ZZZJ_#167
[x]	logitsigmoidshift	uucoco_#95
[x]	LogLU	hli28146_#75
[x]	LogMeanExp	uucoco_#24
[x]	LogSigmoid	uucoco_#23
[x]	LogSigmoid Affine Gate	Ljy123_#41
[x]	LogSigmoid-Affine-Gate	Ljy123_#96
[x]	LogSumExp	gsd123_#71 hli28146_#10 wut0n_#28
[x]	LogWeightedSumExp	uucoco_#22
[x]	lp_pool2d	ZZZJ_#149
[x]	lrn_simple	zizi05_#5
[x]	Lsoftmaxloss	hli28146_#53
[x]	MahalanobisDistanceLoss	uucoco_#96
[x]	manhattan	wut0n_#17
[x]	manhattan_distance_matrix	ZZZJ_#80
[x]	manhattan_erf	uucoco_#134
[x]	manhattan_hardswish	wut0n_#103
[x]	manhattan_leakyrelu	wut0n_#60
[x]	manhattan_mse	wut0n_#66
[x]	manhattan_relu	wut0n_#62
[x]	manhattan_sigmoid	wut0n_#61
[x]	manhattan_sqrt	wut0n_#67
[x]	manhattan_swish	wut0n_#64
[x]	manhattan_tanh	wut0n_#65
[x]	marcsinh	gsd123_#51
[x]	marginrankingloss	ZZZJ 16
[x]	matmul	HHyy 25
[x]	maxunpool1d	ZZZJ_#11
[x]	maxunpool2d	ZZZJ_#12
[x]	meanstdnormalizeclip	uucoco_#97
[x]	median_filter_3d	ZZZJ_#106
[x]	MElliott	hli28146_#134
[x]	MeshEdgeLoss	gsd123_#116
[x]	MetaAconC	hli28146_#26
[x]	minkowski	wut0n_#18
[x]	minkowski_contrastiveloss	wut0n_#68
[x]	minkowski_instancenorm	wut0n_#78
[x]	minkowski_relu	wut0n_#74
[x]	MinkowskiDistance	uucoco_#7
[x]	minmax_observer	ZZZJ_#103
[x]	minmaxscaleshift	uucoco_#98
[x]	Mish	gsd123_#31 zizi05_#1
[x]	Mish+Bias	Ljy123_#17
[x]	Mish-Affine-Gate	Ljy123_#93
[x]	MishB	uucoco_#62
[x]	mishglu	hli28146_#59 uucoco_#39
[x]	mixup	ZZZJ_#147
[x]	MMReLU	hli28146_#94
[x]	ModeSeekingLoss	uucoco_#99
[x]	ModReLU	gsd123_#52
[x]	ModSwish	hli28146_#114
[x]	mseloss	wut0n_#10 ZZZJ 28
[x]	MsewithLogitLoss	uucoco_#100
[x]	mulaw_decoding	ZZZJ_#100
[x]	mulaw_encoding	ZZZJ_#101
[x]	multilabelmarginloss	gsd123_#73 hli28146 44
[x]	multimarginloss	hli28146_#2
[x]	Multiquadratic	gsd123_#54
[x]	MutualInformation	uucoco_#51
[x]	NIPUNA	hli28146_#85
[x]	Nish	hli28146_#74
[x]	NISRLU	hli28146_#73
[x]	NLLLoss	uucoco 32 uucoco_#2
[x]	NLReLU	hli28146_#78
[x]	optional_get_element	ZZZJ_#22
[x]	optional_has_element	ZZZJ_#23
[x]	PAA	hli28146_#132
[x]	pairwise_distance	zizi05_#3
[x]	pairwisedistance	ZZZJ 38
[x]	ParametricSigmoid	uucoco_#63
[x]	PATS	hli28146_#88
[x]	PearsonCorrelation	Lwh20070813 41 Lwh20070813_#1
[x]	penalizedtanh	hli28146_#64
[x]	PerceptualLoss	uucoco_#101
[x]	permute	ZZZJ_#143
[x]	PGELU	hli28146_#99
[x]	piecewiseaffine	gsd123_#94
[x]	PiecewiseLinearUnit	gsd123_#39
[x]	poissonnllloss	gsd123 20
[x]	polar_to_cartesian	ZZZJ_#132
[x]	polaraffine	gsd123_#95
[x]	poly1crossentropy	hli28146_#15
[x]	poly1focalloss	hli28146_#16
[x]	PolyLoss	uucoco_#64
[x]	polynomial_eval	ZZZJ_#99
[x]	polynomialaffine	gsd123_#96
[x]	Power-Sigmoid-Affine-Gate	Ljy123_#46
[x]	Power-Sigmoid-Gate	Ljy123_#21
[x]	PPOLoss	uucoco_#102
[x]	PReLU	uucoco_#40
[x]	projectiveaffine	gsd123_#97
[x]	Prototype-Cosine Gate	Ljy123_#36
[x]	Pserf	hli28146_#92
[x]	PseudoHuber-Affine-Gate	Ljy123_#99
[x]	PSGU	hli28146_#101
[x]	PSMish	uucoco_#42
[x]	PTELU	hli28146_#122
[x]	PTLU	uucoco_#65
[x]	python	Ljy123_#67
[x]	QLearningLoss	uucoco_#103
[x]	QReLU	gsd123_#60
[x]	QuantileLoss	hli28146_#69 uucoco_#66
[x]	quantilenormalizeexpand	uucoco_#104
[x]	RadialBasisFunction	uucoco_#43
[x]	Range-Gate Affine	Ljy123_#39
[x]	rangescalegate	uucoco_#68
[x]	rangeshiftgate	uucoco_#69
[x]	rank_normalize_scale	gsd123_#122
[x]	ranknetloss	hli28146_#50
[x]	ransac_normalize_outlier_reject	gsd123_#123
[x]	RationalClip-Affine-Gate	Ljy123_#112
[x]	RationalFunctionApproximator	uucoco_#52
[x]	real_imag_hypot	uucoco_#105
[x]	ReflectionPad3d	gsd123 22
[x]	ReLTanh	hli28146_#125
[x]	ReLU-Affine-Gate	Ljy123_#88
[x]	ReLU6-Affine-Gate	Ljy123_#83
[x]	ReLU^2	Ljy123_#14
[x]	RenyiDivergenceLoss	uucoco_#106
[x]	repeat_interleave	ZZZJ_#102
[x]	replicationpad1d	wwmm_#4
[x]	replicationpad2d	wwmm_#5
[x]	replicationpad3d	wwmm_#6
[x]	RePU	hli28146_#60
[x]	resistance_distance_exp	gsd123_#163
[x]	resize_nearest	ZZZJ_#133
[x]	rgb_to_bayer	ZZZJ_#36
[x]	rgb_to_grayscale	ZZZJ_#41
[x]	rgb_to_xyz	ZZZJ_#39
[x]	rgb_to_yuv	ZZZJ_#47
[x]	RMAF	hli28146_#121
[x]	rmsnorm	Icy_cola9 Ljy123_#8 wut0n_#3
[x]	rmsnorm_residual	wut0n_#32
[x]	rmsnorm_silu	hli28146_#61
[x]	robustscalegate	uucoco_#70
[x]	robustscalehuber	uucoco_#107
[x]	rogers_tanimoto_silu	gsd123_#164
[x]	roiaware_pool1d	ZZZJ_#152
[x]	roll2d	ZZZJ_#52
[x]	roll3d	ZZZJ_#53
[x]	rope	wwmm_#1
[x]	SAAF	hli28146_#93
[x]	SaRa	hli28146_#106
[x]	SbPiPLU	hli28146_#116
[x]	scalenorm	hli28146_#33
[x]	scatter_add	ZZZJ#24
[x]	scatter_div	ZZZJ_25
[x]	scatter_nd	ZZZJ_#30
[x]	sceloss	hli28146_#40
[x]	SCLMish	hli28146_#119
[x]	SCSwish	hli28146_#118
[x]	segment_reduce	ZZZJ_#69
[x]	SELU+Affine	Ljy123_#26
[x]	SELU-Affine-Gate	Ljy123_#90
[x]	selu_clip	zizi05_#4
[x]	Sep	hli28146_#111
[x]	separable_conv2d	ZZZJ_#141
[x]	separable_conv3d	ZZZJ_#142
[x]	serf	hli28146_#65 uucoco_#45
[x]	Serlu	gsd123_#40
[x]	ShiftedSincUnit	uucoco_#47
[x]	Sigmoid+Clamp+Affine	Ljy123_#16
[x]	Sigmoid-Affine-Gate	Ljy123_#85
[x]	sigmoid_derivative	wut0n_#27
[x]	sigmoid_focal_loss	ZZZJ_#131
[x]	SigmoidGLU	uucoco_#53
[x]	SigmoidSlope-Affine-Gate	Ljy123_#110
[x]	SigmoidSquared-Affine-Gate	Ljy123_#117
[x]	signmuladd	uucoco_#108
[x]	SimCLRLoss	gsd123_#133
[x]	Sin-Affine-Gate	Ljy123_#44
[x]	Sinc-Affine-Gate	Ljy123_#92
[x]	sincoshypot	uucoco_#109
[x]	Sine-Affine-Gate	Ljy123_#114
[x]	Sinh-Affine-Gate	Ljy123_#81
[x]	SinLU	hli28146_#89
[x]	SinuGaussian	uucoco_#71
[x]	smelu	hli28146_#27 uucoco_#48
[x]	SmoothAbs	gsd123_#74
[x]	SmoothL1Loss	gsd123 40 gsd123_#13
[x]	SmoothMaximum	gsd123_#75
[x]	SmoothMinimum	gsd123_#76
[x]	SmoothRamp	gsd123_#77
[x]	SmoothStep	gsd123_#78 ZZZJ_#123
[x]	SoftClip	uucoco_#21
[x]	SoftClip+Affine	Ljy123_#30
[x]	SoftExponential	uucoco_#20
[x]	softmarginloss	hli28146 35 hli28146_#3
[x]	softmax	Ljy123_#6
[x]	softmin	uucoco_#14
[x]	softplus	uucoco_#15
[x]	Softplus+Bias	Ljy123_#15
[x]	Softplus^2-Affine-Gate	Ljy123_#80
[x]	Softplus^3-Affine-Gate	Ljy123_#102
[x]	SoftplusGLU	uucoco_#54
[x]	SoftplusSqrt-Affine-Gate	Ljy123_#108
[x]	softshrink	uucoco_#16
[x]	SoftShrink-Affine-Gate	Ljy123_#103
[x]	softsign	uucoco_#18
[x]	Softsign+Affine	Ljy123_#18
[x]	SoftSign-Affine-Gate	Ljy123_#91
[x]	Softsign-Sigmoid-Mix-Gate	Ljy123_#52
[x]	SoftSignSquared-Affine-Gate	Ljy123_#119
[x]	solarize	ZZZJ_#140
[x]	Spatial-Diff Sigmoid Gate	Ljy123_#38
[x]	spherefaceloss	hli28146_#23
[x]	Spherical+Affine	Ljy123_#24
[x]	SQRBF	gsd123_#59
[x]	Sqrt(ReLU)+Affine	Ljy123_#28
[x]	sqrtreciprocalrsqrt	uucoco_#110
[x]	Square-Sigmoid-Affine-Gate	LJy123_#47
[x]	squared_euclidean	wut0n_#23
[x]	SquaredHingeLoss	uucoco_#72
[x]	squareplus	ZZZJ_#124
[x]	SReLU	uucoco_#49
[x]	SRS	uucoco_#73
[x]	std_mean	ZZZJ_#122
[x]	STL	hli28146_#79
[x]	structural_similarity_softplus	uucoco_#135
[x]	SupConLoss	gsd123_#132
[x]	SwAT	hli28146_#109
[x]	swiglu	zizi05_#7 ZZZJ1
[x]	swish	Ljy123_#7
[x]	Swish(SiLU)-Affine-Gate	Ljy123_#86
[x]	Swish^2+Bias	Ljy123_#25
[x]	swish_layernorm	zizi05_10
[x]	switchablenorm	ZZZJ 12
[x]	SørensenDice	gsd123_#27
[x]	Tanh-Affine-Gate	Ljy123_#45 Ljy123_#84
[x]	tanhexp	hli28146_#20
[x]	TanhGLU	uucoco_#55
[x]	TanhLU	hli28146_#90
[x]	tanhshrink	zizi05_#9
[x]	Tanhshrink+Affine	Ljy123_#20
[x]	TanhSlope-Affine-Gate	Ljy123_#111
[x]	TanhSoft1	hli28146_#96
[x]	TanhSoft2	hli28146_#97
[x]	TanimotoCoefficient	uucoco 33 uucoco_#3
[x]	TDLoss	uucoco_#111
[x]	TeLU	hli28146_#80
[x]	Temperature Softmax Channel Affine	Ljy123_#40
[x]	tensor_roll	ZZZJ_#138
[x]	three_interpolate	ZZZJ_#156
[x]	thresholdscalenegate	uucoco_#112
[x]	topk_filtering	ZZZJ_#139
[x]	total_correlation_elu	uucoco_#136
[x]	transpose_scale	ZZZJ_#125
[x]	tripletloss	wut0n_#7
[x]	tripletmarginloss	hli28146_#1 ZZZJ 30
[x]	TripletMarginWithDistanceLoss	gsd123 39 gsd123_#14
[x]	TrustRegionPolicyOptimizationLoss	uucoco_#113
[x]	TsallisDivergenceLoss	uucoco_#114
[x]	tversky_loss	ZZZJ_#72
[x]	TverskyIndex	gsd123_#80
[x]	tverskyloss	hli28146_#57
[x]	TweedieLoss	gsd123_#100
[x]	upsample	ZZZJ 21
[x]	ValueLoss	uucoco_#115
[x]	Variance	uucoco_#13 wut0n_#26
[x]	VariationOfInformation	uucoco_#56
[x]	vecdot	hli28146_#8
[x]	VIDLoss	gsd123_#107
[x]	voxel_hash	ZZZJ_#158
[x]	voxel_mean	ZZZJ_#159
[x]	voxel_to_point	ZZZJ_#160
[x]	wasserstein_energy_gelu	uucoco_#137
[x]	wasserstein_layernorm	gsd123_#166
[x]	waveletaffine	gsd123_#101
[x]	WeightDecay	gsd123_#81
[x]	wingloss	hli28146_#14
[x]	winsorize_scale_normalize	gsd123_#93
[x]	wiou	ZZZJ_#33
[x]	xIELU	hli28146_#77
[x]	xSiLU	hli28146_#76
[x]	xyz_to_rgb	ZZZJ_#38
[x]	yuv_to_rgb	ZZZJ_#49
[x]	zeropad2d	ZZZJ_#9
[x]	zeropad3d	ZZZJ_#10
[x]	zscoresigmoiddenormalize	uucoco_#116
[x]	仿射+ReLU	Ljy123_#9

📥 参赛流程

一句话概括：进入GPUCodeForces赛事首页，完成一份成功合并到仓库内的提交即为参赛成功！时间自由，方法自由，只要有灵感就可以动手开code~

🌰举个栗子

登录or注册自己的Gitlink账号后，进入赛事首页初步查看仓库内的文件内容：

.
├── S1(说明：第一季比赛名称，该目录下文件需选手自建)
│   ├── issue id 1(选手提交算子前创建的issue对应的id)
│   │   ├── cudacode.py(必要提交文件)
│   │   ├── torchcode.py(必要提交文件)
│   │   ├── run_code.py(必要提交文件)
│   │   ├── prompt.txt(必要提交文件)
│   ├── issue id 2
│   ├── issue id ...
├── example(样例提供，供大家上手)
│   ├── 001-example(文件结构与 issue id x 一致)
│   ├── 002-example
├── images(图片文件夹，参赛选手可忽略)
├── FAQ.md(社区收集的问题与解答)
├── LICENSE(证书，参赛选手可忽略)
├── README.md(赛题baseline)
├── how-to-contribute.md(提交指南看这里)

在S1文件夹的文件即为参赛选手需要提交的文件，其余文件夹和文件皆为辅助参赛选手了解比赛、提供思路、排忧解难之用。
我们将当前赛题的项目clone到本地电脑上（建议使用git clone + 链接的方式），在S1文件夹下创建一个以自己issue id命名的文件夹，待提交的参赛代码文件都放在这里面。

 Tips：issue id ≠ issue名称，名称建议为对该算子的概括性描述。创建了issue后，链接末尾处的数字即为issue id。同时，也可以在issue界面的醒目位置查看如"#123"这样的数字标识。

做好了准备工作，就可以开始尽情发挥，去寻找算子亦或是优化算子。在example文件夹中提供了算子样例，如果想简单上手可以查看该文件夹。虽然有效的算子优化也算一次提交，但我们鼓励大家发现新的算子✨真正与其他选手拉开差距。

🔧简单介绍一下样例代码：

example_torchcode.py： 基准模型（Baseline）。示例提供一个简单的PyTorch模型，只包含一个ReLU激活函数。

example_cudacode.py：优化模型。示例使用PyTorch的load_inline功能直接编译和加载CUDA代码，创建了一个新的模型类，使用自定义CUDA实现替代PyTorch的ReLU。

run_code.py：验证和性能测试脚本。验证自定义CUDA实现与原始PyTorch实现的数值精度一致性，比较两种实现的性能，计算加速比。

prompt.txt：提示词文本。提供类似 “融合算子 CUDA 设计” 的 prompt 编写思路。

💡如何将一个example变为自己的一份提交，具体的算子优化思路可参考：GPUCodeForces赛题解读
为了测试代码最终的跑通结果，需要使用规定的GPU。在模力方舟平台上准备了大家此次需要的算力资源，使用免费的算力券购买实例，接着便可以在云端实例上进行代码修改和测试。

相关算力券的领取方式请见算力平台使用说明、算力券兑换发放和兑换。
```
Tips：进行云端服务器的实例选择有库存的即可，点击进入后需要确定配置。其余选项保持默认，只需更改镜像处的选择：基础镜像-->CV-CUDA-->PyTorch2.4.0-->Python任意-->maca 3.0.0.5
```

代码测试运行没有问题后，便可以准备提交了（记得将所有必要文件保存到提交文件夹）。提交流程如下，全程使用git：

# cd到自己克隆到本地的项目路径下，如：C:\Users\Desktop\ODTC AI Infra\GPUCodeForces

git remote -v      # 检查是否链接到自己的Gitlink仓库
git checkout -b dev      # 新建一个分支，dev即分支名字可以任取
git add .      # 将文件的变更操作暂存
git commit -m "输入你本次提交的目的"      # 目的简洁明了最好
git push origin dev      # 提交你的修改到分支上

对git与远程仓库操作的疑问可以点击how-to-contribute.md查看详情。

顺利提交后的代码还只在你自己fork的仓库下，还需要和主仓库合并才能真正让管理员看到你的代码。

fork仓库（你自己的仓库）与主仓库之间关系如下：

ODTC AI Infra/GPUCodeForces(main)   
├── ...
├── folders

your_name/GPUCodeForces(main) # 自己仓库下的main分支内容与主仓库一致
├── ...
├── folders

your_name/GPUCodeForces(dev) # dev分支才是你修改后的代码存放处
├── ...
│   ├── your codes
├── folders

进入自己的仓库，点击上方选项栏：

合并请求(PR)-->+新建合并请求-->源分支选择dev(名称任取)-->填写下方选框-->标题为自己的issue标题-->描述填请简明扼要，描述内关联自己的issue id(如“fixes #001”)

提交后便能看到自己的PR记录了，在对应记录的评论区会有测试结果的告知，请留意查看~

🌳一份完整的提交流程如上，期待各位自由发挥，赛出风采与水平！

⏺如仍有疑问，请点击提交流程演示视频

⭐审核流程

你提交的PR都会得到回复，大概存在的几种情况如下
```
提交PR-->测试通过✔️-->已合并-->有效提交 ヽ(✿ﾟ▽ﾟ)ノ
提交PR-->测试通过✔️-->已关闭-->代码重复/相似-->无效提交 (；′⌒`)
提交PR-->测试失败✖️-->已关闭-->代码不合格-->无效提交 (；′⌒`)
```
也就是说，除了能够自己在算力平台的实例上运行得到算子测算的初步结果外，还可以在这里看到最终的测算结果。这里显示测试通过才能进入后续审核流程，并最终上传至比赛的算子仓库。

✅ 参赛资格

你提交的PR样本中，至少有 1 个样本被评审通过并正式整合到“GPU CodeForces” 数据集，即可算作有效参赛。

⚠️注意事项

请勿抄袭他人代码或成果
请勿恶意提交（如相同算子多次提交、相近优化重复提交）
---相近优化：即指同一份参赛用例在优化后加速比没有提升或更低的用例提交
请遵守提交的格式要求、内容规范

🏅 竞赛排名机制

优先按接受数量从高到低排序作为排名，取前12名。
若接受数量相同：
- 比较总基础评分高者优先
- 若仍相同，比加分项得分高者优先

接受数量 = 提交并被评审通过的样本总数

接受数量相同需要区分排名时如下的评分规则才会生效

📈 评分规则

📊 基础得分

内容	分值
提供标准 GT 输出生成函数（Numpy-CPU / 原始框架实现）	+2 分
CUDA 执行时间评估	+5 分
CUDA 吞吐量评估	+4 分
CUDA 内存带宽评估	+3 分

✨ 加分项

内容	分值
提供 Prompt 让 LLM 生成对应的 CUDA 代码，并同样进行性能评价	额外加分

📚 术语解释

评测数据集：用来测试 GPU 性能的一组标准化样本，包括代码、输入数据和预期结果。
GT（Ground Truth）：标准参考答案或结果，用来验证程序运行是否正确。
吞吐量（Throughput）：每秒钟能处理的数据量，越高表示 GPU 处理能力越强。
内存带宽（Memory Bandwidth）：单位时间内 GPU 内存与计算核心之间的数据传输速度。
Prompt：引导大语言模型（LLM）生成代码或内容的提示词。
LLM：Large Language Model，大语言模型，如 ChatGPT、LLaMA 等。

📬 联系与帮助

如需更多信息或格式说明，请查看官方文档或在本仓库提交想法进行讨论，或直接在交流群内与主办团队进行沟通。

祝各位挑战成功，贡献出高质量的 GPU 评测数据集！🚀

FAQ

第一季FAQ参考

49 KiB Raw Permalink Blame History Unescape Escape