j+RdZddlZddlZddlZddlZddlZddlZddl m Z m Z ddl Z gdZ d(deddfdZdefd Zd(deddfd Zdefd Zd(deddfd Zdefd ZdeddfdZdefdZdeddfdZdefdZd)dededdfdZdefdZde eeeeffdZde eeffdZdeddfdZ d*de edefdZ!d*de edefdZ"deddfdZ#dede$efd Z%d+d!Z&d"ed#eddfd$Z'd+d%Z(ded&eddfd'Z)dS),a} This module exposes a TunableOp interface. Some operations, such as GEMMs, could be implemented using more than one library or more than one technique. For example, a GEMM could be implemented for CUDA or ROCm using either the blas or blasLt libraries. Further, ROCm's rocblas and hipblaslt libraries allow the user to query for all possible algorithms and then choose one. How does one know which implementation is the fastest and should be chosen? That's what TunableOp provides. Enabling TunableOp and Tuning Separately ======================================== The TunableOp feature is enabled separately from enabling the tuning phase itself. Enabling TunableOp means that PyTorch will replace any standard operators with their Tunable implementations. Any call to a TunableOp first checks whether it has already been tuned for the given operator inputs. If so, it will immediately call the tuned operation; no further tuning will take place even when the tuning setting is enabled. Instead if no tuning result is found, and tuning is enabled, the TunableOp will benchmark every registered implementation of that operator for the given set of inputs and select the fastest. File Input and Output ===================== The first time any TunableOp is invoked, the internal database of tuned operations will be prepared by attempting to read the results from the given file. The default filename is 'tunableop_results.csv'. To support tuning when multiple GPUs are used across multiple processes, the GPU device ordinal is automatically inserted into the filename to avoid multiple processes overwriting the same file. If tuning is enabled and new tunings are discovered during the course of your workload, it will also write out to this same filename with all tunings, both the ones it read in at startup as well as the new ones found at runtime. This can be used, for example, to build up a tunings file across many workloads by reusing the same file. The output file is automatically created when the application terminates. This behavior can be controlled by the C++ and Python APIs but not the environment variables. Assuming you specified a filename, you'll end up with a CSV file with contents like so:: Validator,PT_VERSION,2.2.0 Validator,ROCM_VERSION,6.0.0.0-12969-1544e39 Validator,HIPBLASLT_VERSION,0.6.0-a9c5cc7 Validator,ROCBLAS_VERSION,4.0.0-72e57364-dirty GemmTunableOp_float_NT,nt_25088_4096_64,1219,1.262 GemmTunableOp_float_NT,nt_4096_4096_64,1216,0.033 Note the "Validator" lines. If you change a library version, or ROCm version, or PyTorch version, TunableOp will detect this and reject the tunings file because the prior tunings are likely affected by other software changes. The remaining lines are the tuned solutions for each TunableOp encountered during your execution. Each line consists of 4 comma-separated fields: operator name, operator parameters, solution name, and average execution time. The execution time is an optional field. The CSV file can be edited, but with caution. For example, the solution name (field 3) can be changed to "Default" and it will fall back to the original PyTorch untuned implementation. Or, in the case of ROCm's hipBLAS or hipBLASLt libraries, if you know the specific solution index you can override the solution that TunableOp selected by replacing the value. The operator name and parameters (fields 1 and 2) are internally named and should not be modified. In the case of GemmTunableOp, field 1 indicates the datatype and whether the inputs are transposed (T) or not (N) and field 2 indicates the M, N, K input shapes. There is an option to enable verbose output but it is only recommended for debugging purposes. This will produce a lot of diagnostic messages but may be useful to see if TunableOp is being used at all. Otherwise, TunableOp is completely silent, besides file output, unless there is a warning or error during its use. The verbose option is only available by setting the environment variable PYTORCH_TUNABLEOP_VEROBSE=1. A Note on Tuning Behavior ========================= Tuning an operator consists of iterating through the list or registered implementations and profiling each one. The profile is established by running a single implementation in a loop multiple times and taking the average execution time. By default, each possible solution for a given operator will be run for either 100 iterations or as many iterations that can be run within 30ms, whichever is smaller, and its average execution will be calculated. The fastest solution among all that were successfully profiled will be chosen. A profile might fail if the given solution doesn't achieve the same accuracy as the default implementation or if the solution returns an error code. Current Tunable Operators ========================= TunableGemm for ROCm -------------------- Currently only a TunableGemm for ROCm is implemented. Note that CUDA builds of PyTorch will function correctly when using TunableOp but the only solution available to CUDA builds is the 'Default' implementation i.e. the original cuBLAS default, now called through TunableOp. Any call to at::cuda::blas::gemm() or ::bgemm() will be routed through TunableOp when enabled. Calling gemm() for a given set of input arguments (transa, transb, m, n, k) will attempt to use the fastest available implementation across both rocblas and hipblaslt. Tuning Context ============== The behavior of TunableOp is currently manipulated through environment variables, the C++ interface of at::cuda::tunable::getTuningContext(), or the torch.cuda.tunable python interfaces that wrap the C++ TuningContext. The environment variables take precedence over any setting you manipulate using the C++ or Python APIs. N)OptionalTuple)enable is_enabled tuning_enabletuning_is_enabledrecord_untuned_enablerecord_untuned_is_enabledset_max_tuning_durationget_max_tuning_durationset_max_tuning_iterationsget_max_tuning_iterations set_filename get_filename get_resultsget_validatorswrite_file_on_exit write_file read_filetune_gemm_in_filemgpu_tune_gemm_in_fileTvalreturncDtj|dS)z@This is the big on/off switch for all TunableOp implementations.N)torch_C_cuda_tunableop_enablers CC:\Projects\Historical\venv\Lib\site-packages\torch/cuda/tunable.pyrrs H##C(((((c>tjS)z1Returns whether the TunableOp feature is enabled.)rr_cuda_tunableop_is_enabledr rrrs 8 . . 0 00r cDtj|dS)zEnable tuning of TunableOp implementations. When enabled, if a tuned entry isn't found, run the tuning step and record the entry. N)rr_cuda_tunableop_tuning_enablers rrrs  H**3/////r c>tjS)z7Returns whether TunableOp implementations can be tuned.)rr!_cuda_tunableop_tuning_is_enabledr#r rrrs 8 5 5 7 77r cDtj|dS)zEnable recording untuned of TunableOp perations for offline tuning. When enabled, if a tuned entry isn't found, write it to the untuned file. N)rr_cuda_record_untuned_enablers rr r s  H((-----r c>tjS)zEReturns whether TunableOp operations are recorded for offline tuning.)rr_cuda_record_untuned_is_enabledr#r rr r s 8 3 3 5 55r durationcDtj|dS)zSet max time in milliseconds to spend tuning a given solution. If both max tuning duration and iterations are set, the smaller of the two will be honored. At minimum 1 tuning iteration will always be run. N)rr'_cuda_tunableop_set_max_tuning_duration)r,s rr r s  H44X>>>>>r c>tjS)z.Get max time to spend tuning a given solution.)rr'_cuda_tunableop_get_max_tuning_durationr#r rr r s 8 ; ; = ==r iterationscDtj|dS)zSet max number of iterations to spend tuning a given solution. If both max tuning duration and iterations are set, the smaller of the two will be honored. At minimum 1 tuning iteration will always be run. N)rr)_cuda_tunableop_set_max_tuning_iterations)r1s rr r s  H66zBBBBBr c>tjS)z4Get max iterations to spend tuning a given solution.)rr)_cuda_tunableop_get_max_tuning_iterationsr#r rrrs 8 = = ? ??r Ffilenameinsert_device_ordinalcFtj||dS)a/Set the filename to use for input/output of tuning results. If :attr:`insert_device_ordinal` is ``True`` then the current device ordinal will be added to the given filename automatically. This can be used in a 1-process-per-gpu cenario to ensure all processes write to a separate file. N)rr_cuda_tunableop_set_filename)r6r7s rrrs# H))(4IJJJJJr c>tjS)zGet the results filename.)rr_cuda_tunableop_get_filenamer#r rrrs 8 0 0 2 22r c>tjS)zReturn all TunableOp results.)rr_cuda_tunableop_get_resultsr#r rrrs 8 / / 1 11r c>tjS)z Return the TunableOp validators.)rr_cuda_tunableop_get_validatorsr#r rrrs 8 2 2 4 44r cDtj|dS)aDuring Tuning Context destruction, write file to disk. This is useful as a final flush of your results to disk if your application terminates as result of normal operation or an error. Manual flushing of your results can be achieved by manually calling ``write_file()``.N)rr"_cuda_tunableop_write_file_on_exitrs rrrs  H//44444r c`|t}tj|S)zfWrite results to a CSV file. If :attr:`filename` is not given, ``get_filename()`` is called. )rrr_cuda_tunableop_write_filer6s rrrs) >> 8 . .x 8 88r c`|t}tj|S)zqRead results from a TunableOp CSV file. If :attr:`filename` is not given, ``get_filename()`` is called. )rrr_cuda_tunableop_read_filerDs rrrs) >> 8 - -h 7 77r c(tsJtsJtj}t |5}|D]'}|drt||( ddddS#1swxYwYdS)ztune GEMM in file.Gemm ScaledGemmN)rrrcudacurrent_deviceopen startswith_process_single_offline_gemm)r6deviceidfilelines rrr s <<<    z((**H h=4 = =D566 =,T8<<< ===================s+BB B filename_patternct}tj|D]Y}t|5}|D],}|dr||- dddn #1swxYwYZ|S)zOProcess multiple untuned results file and return a set with duplicates removed.rHN)setglobrMrNadd)rSunique_gemm_entries file_pathrQrRs r&_gather_unique_untuned_gemm_from_filesrZs%%Y/0022 )__ 2 2 2??#9::2'++D111 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 s0A11A5 8A5 c0t}g}t}|F|dkr@|d}|dkr"|dkr|d|dz dz||dz}nSd}nPtjd}||dkrd }n1d |vr|d d}n|dd }d|vsJd }t j|}t|}|D]s} t| 5} | D]D} | d r|s| | /| | E dddn #1swxYwYd}t|dd} t| d5} |D]} | | |D]} | |  dddn #1swxYwYtd|D]:}| dt|}tj| |;dS)zMGather results from multiple tunableop results file and create a single file.N.r?PYTORCH_TUNABLEOP_FILENAMEztunableop_results?.csvz%dz?.F ValidatorT_full0w0)rUrfindosgetenvreplacerVlenrMrNappendrWwriterangestrshutilcopy) gemm_linesvalidator_linesresults_filenamedot_posrSresults_filename_env FirstFilematching_filesnum_matching_filesrYrQrR output_fileout_fileiduplicate_files r_gather_tunableop_resultsr}'sJO$~~ $)9R)?)?"'',, b==Wq[[!7Q;/#58H8RR   "  !y)EFF  '+?2+E+E7   ) ) )3;;D#FF  3;;CFF  " " " " "IY/00N^,,#   )__ ) ) )??;//)%5'..t444NN4((((  ) ) ) ) ) ) ) ) ) ) ) ) ) ) ) ) "**399K k3  !8# ! !D NN4  ! !D NN4  !!!!!!!!!!!!!!!!1( ) )11$,,S#a&&99 K000011s%5AE  E E >5GGGuntuned_gemm_linegpu_idc " dt|z}tjtjtjtjtjtjtjtj tj tj d }| ddd}|dd}d}d}d}d} |dkrM|d d\} } } | ddk} | d dk}|| }n|d d}|d} |d dz|dz}|d dz|d z}|d dz|d z}|dddk} |dd dk}||}||}||} d|d  dd d D\}}}| dkr|r*tj||||ntj||||}| r*tj||||ntj||||}tj||dS| dkrd|d  dd d D\}|rtj|||||ntj|||||}| rtj|||||ntj|||||}|r|d dn|}| r|d dn|}tj||dS| dkrd}d}tjd|}tjd|}|r,tj||f|||ntj||f|||}| r,tj||f|||ntj||f|||}tj||||| dS| dkr| |ksJ|r*tj||||ntj||||}| rtj||||n)tj||||}| rtj|||ntj|||}tjj|||dSt7jd| dS)zProcess a single untuned GEMM.zcuda:) floatdoubleBFloat16Halfzc10::complexzc10::complex Float8_e4m3fn Float8_e5m2Float8_e4m3fnuzFloat8_e5m2fnuz,Nr_Tr_c,g|]}t|Sr#int.0gs r z0_process_single_offline_gemm..sAAAAQAAAr GemmTunableOp)dtypedeviceGemmStridedBatchedTunableOpc,g|]}t|Sr#rrs rrz0_process_single_offline_gemm..s???!s1vv???r ScaledGemmTunableOpg?g?g?)rg?)scale_ascale_b out_dtypeGemmAndBiasTunableOpzerror: unknown op )rnrfloat32float64bfloat16half complex128 complex64 float8_e4m3fn float8_e5m2float8_e4m3fnuzfloat8_e5m2fnuzstripsplitcountgetrandtmm transposebmmtensorfull _scaled_mmnn functionallinearwarningswarn)r~rrP dtype_dict untuned_gemmunderscore_countrdtypeAdtypeBdtypeCop_sig data_typelayouttransAtransBuntuned_gemm_temp data_typeA data_typeB data_typeCnmkmatAmatBbfillAfillBscaleAscaleBXbiass rrOrOfs V$H-N % 0$,( 0 0  J%**,,22377:L#A,,S11 E F F F1&21o&;&;C&@&@#Fc!c!y))(O11#66"1%&q)C/2CA2FF &q)C/2CA2FF &q)C/2CA2FF "1%a(C/"1%a(C/ ++ ++ ++AAa!6!6s!;!;AaC!@AAAIQ1    @EJq!5 : : : < < > > >Aqh???  @EJq!5 : : : < < > > >Aqh??? t 0 0 0??|A44S99!A#>??? CEJq!QeH = = = =Aq!5BBB  CEJq!QeH = = = =Aq!5BBB (.7t~~a###4'-7t~~a###4 $ ( ( (c(333c(333 JEJ1vuF8 D D D F F H H HQFEIII  JEJ1vuF8 D D D F F H H HQFEIII tVVvVVVVVV ) ) ) @EJq!5 : : : < < > > >Aqh???  DEJq!5 : : : :Aqh???AACC  =EJqh 7 7 7 7AU8<<< ""1dD11111 363344444r cFtsJtsJdS)zHelper function for multi-GPU tuning case. Need to check that TunableOp feature is enabled and that tuning is enabled. N)rrr#r r_check_tuning_assertionsrs+ <<<    r num_gpuscJt|}tj}d|cxkr|ksnJt jd}g}g}g}d}t j||5} t|D]1} | t} | | 2t j |D]} | |D];} | t| |} | | |dz|z}>  F MMOOOOx / /A#??:66L   . . . .&.;;MJJ " "L    ! ! ! ! "9"""""""""""""""> Jsrsqqd  """"""""    .))))))) 1D1111 00t0t0000848888 ..t.t....646666 ?c?d????>>>>> C#C$CCCC@3@@@@ KK3KtKKKKK3c3333 2U3S%/02222 5c3h5555 5D5T555599#9$999988 88888 = = = = = = S SX    <1<1<1<1~n5Cn5n5n5n5n5n5b1 S1 C1 D1 1 1 1 1 1 r