wg~:ddlmZddlZddlZddlZddlZddlmZddlm Z m Z ddl m Z ddl mZGd d e ZGd d Z dd ZGdde ZdZy)) annotationsN)Dict)do_benchdo_bench_cudagraph)KernelInterface)OutOfResourcesc<eZdZ d ddZdZdZdZdZy) AutotunerNc |stidddg_n|_|D cgc]} |j| c} _i_|_g_|$|D cgc]} |j| c} _g_|$|D cgc]} |j| c} _dd_d_ |r|_n=tj dkDstjdkDr dfd }|_|r|_ n$tjdkDr fd }|_ d_ d _ d_ | rc| jd j_ | jd j_ | jdj_ |_|_t#j$j s;j j_t#j$j s;| _| _ddl}| xr|j,j/_ycc} wcc} wcc} w)a :param prune_configs_by: a dict of functions that are used to prune configs, fields: 'perf_model': performance model used to predicate running time with different configs, returns running time 'top_k': number of configs to bench 'prune_num_stages_by'(optional): a function used to prune num_stages. It takes configs:List[Config] as its input, and returns pruned configs. rr) num_warps num_stagesnum_ctasNcyNr)args reset_onlys ]/home/mcse/projects/flask/flask-venv/lib/python3.12/site-packages/triton/runtime/autotuner.pyz$Autotuner.__init__..6cyrr)r exceptions rrz$Autotuner.__init__..7rrrcjD]}||j|s1jDcgc]}||jc}_yycc}wN) reset_idxzero_ restore_idxclonerestore_copies)rriselfs r _pre_hookz%Autotuner.__init__.._pre_hook<sQ$AGMMO$!DHDTDT*Uq47==?*UD'"*UsActjD]&\}}||jj|(g_yr) enumerater!copy_r#)rrr$jr%s r _post_hookz&Autotuner.__init__.._post_hookHsD%d&6&67:DAqGMM$"5"5a"89:&(#r? perf_modeltop_kearly_config_prune)F)Configconfigsindexkey_idxcache arg_namesrr!pre_hook post_hooklenr- configs_top_kr/getfnbase_fninspect isfunction num_warmupsnum_repstorchcuda is_availableuse_cuda_graph)r%r;r5r1key reset_to_zero restore_valuer6r7prune_configs_bywarmupreprDkr&r+rAs` r__init__zAutotuner.__init__s*"2q1MNDL"DL478q *8  "  $:GHQiooa0HDN  $)B V &DM &DN !! "Q & ) (DN "& .22<QDO!1!5!5gt?Q?Q!RD &6&:&:;OQUQhQh&iD # $$T\\2<.kernel_callms + MM$    NN44N 0 NN41N5E  s)A!! B +A?>B?BBB rmedian)rJ return_mode)g?g?g?)rIrJ quantilesinf)compiler.errorsrNkeyskwargs ValueErrorjoindict all_kwargsnargsrDrArBstreamStreamrr@rr?r float) r%rTrmetarN conflictsrWrA bench_resrUrVs ``` @@r_benchzAutotuner._bench`sOAIIK&--"4"4"66 t3d |j4j6d |j d d|j,d| j(/| j)i|j|| j9|j:j<|i|| j9}d|_|Scc}wcc} w)NTrdtypeFrTrE)rrTRITON_PRINT_AUTOTUNING1zTriton autotuning for function z finished after z.2fzs; best config selected: ;)razipr5rcr8r1appendr3hasattrstrrltupler4 prune_configstimerj bench_timebuiltinsminr:r6configs_timings best_configosgetenvprintr<__name__rbr;rQ)r%rr^used_cached_resultall_args_argsnamer$rEargpruned_configs bench_startrTtimings bench_endrets rrQz Autotuner.runsf#dnnd34 ! t|| q /$**//HE 18#LL$0 1&*\\2582C2 /3(JJs399~. /*C$**$%*"!%!3!3F!;"iik ]klSY6;4;;#NV#Nv#NNll IIK "+k"9"*,,wGKK"H 3 dt 4'.$ZZ_F\\!_F! 99. 5 .s TUrrmr) r1r/rcr-r9 isinstancerfintr8rbsortedr])r%r^rr.rTrs @rrvzAutotuner.prune_configss  " "4T44T\\4::XQWXN ??&&E%'ESLC -56>"U*#1   ODOO** !++- "( (9?V!WX^Y^!_s"6C>c tt|j||_g}|j |D]=}|j |j j|i||j?d|_|Sr) rarqr5rcrvrrr;rIrb)r%rr^rrTs rrIzAutotuner.warmups#dnnd34 ((0 F JJ~tww~~##%     r)NNNdF)rHr)r __module__ __qualname__rLrjrQrvrIrrrr r sA!% NKNK`'gR(T( rr c$eZdZdZddZdZdZy)r0ax An object that represents a possible kernel configuration for the auto-tuner to try. :ivar kwargs: a dictionary of meta-parameters to pass to the kernel as keyword arguments. :type kwargs: dict[Str, Any] :ivar num_warps: the number of warps to use for the kernel when compiled for GPUs. For example, if `num_warps=8`, then each kernel instance will be automatically parallelized to cooperatively execute using `8 * 32 = 256` threads. :type num_warps: int :ivar num_stages: the number of stages that the compiler should use when software-pipelining loops. Mostly useful for matrix multiplication workloads on SM80+ GPUs. :type num_ctas: int :ivar num_ctas: number of blocks in a block cluster. SM90+ only. :type maxnreg: Optional[int] :ivar maxnreg: maximum number of registers one thread can use. Corresponds to ptx .maxnreg directive. Not supported on all platforms. :ivar pre_hook: a function that will be called before the kernel is called. Parameters of this function are args. NcX||_||_||_||_||_||_yr)r^rrrmaxnregr6)r%r^rrrrr6s rrLzConfig.__init__s, "  $   rci|jd|jfd|jfd|jfd|jffDcic] \}}||| c}}Scc}}w)Nrrrr)r^rrrr)r%rKvs rrbzConfig.all_kwargss{  kk  !$..1/!4??3 - Q ] 1  sAcg}|jjD]\}}|j|d||jd|j|jd|j|jd|j |jd|j dj|S)Nz: z num_warps: z num_ctas: z num_stages: z maxnreg: rO)r^itemsrrrrrrr`)r%resrKrs r__str__zConfig.__str__sKK%%' $DAq JJ!Bqc{ # $ [ 012 Z /0 \$//!234 Yt||n-.yy~r)rrrNN)rrr__doc__rLrbrrrrr0r0s(!  rr0c 6  f d} | S)a Decorator for auto-tuning a :code:`triton.jit`'d function. .. highlight:: python .. code-block:: python @triton.autotune(configs=[ triton.Config(kwargs={'BLOCK_SIZE': 128}, num_warps=4), triton.Config(kwargs={'BLOCK_SIZE': 1024}, num_warps=8), ], key=['x_size'] # the two above configs will be evaluated anytime # the value of x_size changes ) @triton.jit def kernel(x_ptr, x_size, **META): BLOCK_SIZE = META['BLOCK_SIZE'] :note: When all the configurations are evaluated, the kernel will run multiple times. This means that whatever value the kernel updates will be updated multiple times. To avoid this undesired behavior, you can use the `reset_to_zero` argument, which resets the value of the provided tensor to `zero` before running any configuration. If the environment variable :code:`TRITON_PRINT_AUTOTUNING` is set to :code:`"1"`, Triton will print a message to stdout after autotuning each kernel, including the time spent autotuning and the best configuration. :param configs: a list of :code:`triton.Config` objects :type configs: list[triton.Config] :param key: a list of argument names whose change in value will trigger the evaluation of all provided configs. :type key: list[str] :param prune_configs_by: a dict of functions that are used to prune configs, fields: 'perf_model': performance model used to predicate running time with different configs, returns running time 'top_k': number of configs to bench 'early_config_prune'(optional): a function used to do early prune (eg, num_stages). It takes configs:List[Config] as its input, and returns pruned configs. :param reset_to_zero: a list of argument names whose value will be reset to zero before evaluating any configs. :type reset_to_zero: list[str] :param restore_value: a list of argument names whose value will be restored after evaluating any configs. :type restore_value: list[str] :param pre_hook: a function that will be called before the kernel is called. This overrides the default pre_hook used for 'reset_to_zero' and 'restore_value'. 'args': a list of arguments passed to the kernel. 'reset_only': a boolean indicating whether the pre_hook is called to reset the values only, without a corresponding post_hook. :type pre_hook: lambda args, reset_only :param post_hook: a function that will be called after the kernel is called. This overrides the default post_hook used for 'restore_value'. 'args': a list of arguments passed to the kernel. 'exception': the exception raised by the kernel in case of a compilation or runtime error. :type post_hook: lambda args, exception :param warmup: Warmup time (in ms) to pass to benchmarking, defaults to 25. :type warmup: int :param rep: Repetition time (in ms) to pass to benchmarking, defaults to 100. :type rep: int cF t||j   S)N)r6r7rHrIrJrD)r r5) r;r1rEr7r6rHrJrFrGrDrIs r decoratorzautotune..decorator@s2R\\7C `h#,?OX^dg(68 8rr) r1rErHrFrGr6r7rIrJrDrs `````````` rautotuner sn88 rceZdZddZdZy) Heuristicsc.||_||_||_yr)r;valuesr5)r%r;r5rs rrLzHeuristics.__init__Js "rc |jjD]1\}}|itt|j||||<3|j j |i|Sr)rrrarqr5r;rQ)r%rr^rheurs rrQzHeuristics.runOsg{{((* LGAtJS%> ?J6JKF1I Ltww{{D+F++rN)returnNone)rrrrLrQrrrrrHs # ,rrcfd}|S)a Decorator for specifying how the values of certain meta-parameters may be computed. This is useful for cases where auto-tuning is prohibitevely expensive, or just not applicable. .. highlight:: python .. code-block:: python @triton.heuristics(values={'BLOCK_SIZE': lambda args: 2 ** int(math.ceil(math.log2(args[1])))}) @triton.jit def kernel(x_ptr, x_size, **META): BLOCK_SIZE = META['BLOCK_SIZE'] # smallest power-of-two >= x_size :param values: a dictionary of meta-parameter names and functions that compute the value of the meta-parameter. each such function takes a list of positional arguments as input. :type values: dict[str, Callable[[list[Any]], Any]] c2t||jSr)rr5)r;rs rrzheuristics..decoratorfs"bllF33rr)rrs` r heuristicsrUs"4 r)NNNNNrrF) __future__rryr}rwr=typingrtestingrrjitr errorsr r r0rrrrrrrs[" 2 "CCL22jtx05<~ , ,r