|
12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576777879808182838485868788899091929394959697989910010110210310410510610710810911011111211311411511611711811912012112212312412512612712812913013113213313413513613713813914014114214314414514614714814915015115215315415515615715815916016116216316416516616716816917017117217317417517617717817918018118218318418518618718818919019119219319419519619719819920020120220320420520620720820921021121221321421521621721821922022122222322422522622722822923023123223323423523623723823924024124224324424524624724824925025125225325425525625725825926026126226326426526626726826927027127227327427527627727827928028128228328428528628728828929029129229329429529629729829930030130230330430530630730830931031131231331431531631731831932032132232332432532632732832933033133233333433533633733833934034134234334434534634734834935035135235335435535635735835936036136236336436536636736836937037137237337437537637737837938038138238338438538638738838939039139239339439539639739839940040140240340440540640740840941041141241341441541641741841942042142242342442542642742842943043143243343443543643743843944044144244344444544644744844945045145245345445545645745845946046146246346446546646746846947047147247347447547647747847948048148248348448548648748848949049149249349449549649749849950050150250350450550650750850951051151251351451551651751851952052152252352452552652752852953053153253353453553653753853954054154254354454554654754854955055155255355455555655755855956056156256356456556656756856957057157257357457557657757857958058158258358458558658758858959059159259359459559659759859960060160260360460560660760860961061161261361461561661761861962062162262362462562662762862963063163263363463563663763863964064164264364464564664764864965065165265365465565665765865966066166266366466566666766866967067167267367467567667767867968068168268368468568668768868969069169269369469569669769869970070170270370470570670770870971071171271371471571671771871972072172272372472572672772872973073173273373473573673773873974074174274374474574674774874975075175275375475575675775875976076176276376476576676776876977077177277377477577677777877978078178278378478578678778878979079179279379479579679779879980080180280380480580680780880981081181281381481581681781881982082182282382482582682782882983083183283383483583683783883984084184284384484584684784884985085185285385485585685785885986086186286386486586686786886987087187287387487587687787887988088188288388488588688788888989089189289389489589689789889990090190290390490590690790890991091191291391491591691791891992092192292392492592692792892993093193293393493593693793893994094194294394494594694794894995095195295395495595695795895996096196296396496596696796896997097197297397497597697797897998098198298398498598698798898999099199299399499599699799899910001001100210031004100510061007100810091010101110121013101410151016101710181019102010211022102310241025102610271028102910301031103210331034103510361037103810391040104110421043104410451046104710481049105010511052105310541055105610571058105910601061106210631064106510661067106810691070107110721073107410751076107710781079108010811082108310841085108610871088108910901091109210931094109510961097109810991100110111021103110411051106110711081109111011111112111311141115111611171118111911201121112211231124112511261127112811291130113111321133113411351136113711381139114011411142114311441145114611471148114911501151115211531154115511561157115811591160116111621163116411651166116711681169117011711172117311741175117611771178117911801181118211831184118511861187118811891190119111921193119411951196119711981199120012011202120312041205120612071208120912101211121212131214121512161217121812191220122112221223122412251226122712281229123012311232123312341235123612371238123912401241124212431244124512461247124812491250125112521253125412551256125712581259126012611262126312641265126612671268126912701271127212731274127512761277127812791280128112821283128412851286128712881289129012911292129312941295129612971298129913001301130213031304130513061307130813091310131113121313131413151316131713181319132013211322132313241325132613271328132913301331133213331334133513361337133813391340134113421343134413451346134713481349135013511352135313541355135613571358135913601361136213631364136513661367136813691370137113721373137413751376137713781379138013811382138313841385 |
- # Generated by dnn/scripts/cutlass_generator/gen_list.py
-
- cutlass_gen_list = [
- "cutlass_simt_sgemm_8x32_8x2_nn_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_8x32_8x2_nn_align1.cu",
- "cutlass_simt_sgemm_16x32_8x2_nn_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_16x32_8x2_nn_align1.cu",
- "cutlass_simt_sgemm_16x64_8x2_nn_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_16x64_8x2_nn_align1.cu",
- "cutlass_simt_sgemm_32x32_8x2_nn_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_32x32_8x2_nn_align1.cu",
- "cutlass_simt_sgemm_32x64_8x2_nn_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_32x64_8x2_nn_align1.cu",
- "cutlass_simt_sgemm_64x32_8x2_nn_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_64x32_8x2_nn_align1.cu",
- "cutlass_simt_sgemm_16x128_8x2_nn_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_16x128_8x2_nn_align1.cu",
- "cutlass_simt_sgemm_32x128_8x2_nn_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_32x128_8x2_nn_align1.cu",
- "cutlass_simt_sgemm_64x64_8x2_nn_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_64x64_8x2_nn_align1.cu",
- "cutlass_simt_sgemm_128x32_8x2_nn_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_128x32_8x2_nn_align1.cu",
- "cutlass_simt_sgemm_64x128_8x2_nn_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_64x128_8x2_nn_align1.cu",
- "cutlass_simt_sgemm_128x64_8x2_nn_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_128x64_8x2_nn_align1.cu",
- "cutlass_simt_sgemm_32x256_8x2_nn_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_32x256_8x2_nn_align1.cu",
- "cutlass_simt_sgemm_64x256_8x2_nn_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_64x256_8x2_nn_align1.cu",
- "cutlass_simt_sgemm_128x128_8x2_nn_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_128x128_8x2_nn_align1.cu",
- "cutlass_simt_sgemm_256x32_8x2_nn_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_256x32_8x2_nn_align1.cu",
- "cutlass_simt_sgemm_256x64_8x2_nn_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_256x64_8x2_nn_align1.cu",
- "cutlass_simt_sgemm_8x32_8x2_nt_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_8x32_8x2_nt_align1.cu",
- "cutlass_simt_sgemm_16x32_8x2_nt_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_16x32_8x2_nt_align1.cu",
- "cutlass_simt_sgemm_16x64_8x2_nt_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_16x64_8x2_nt_align1.cu",
- "cutlass_simt_sgemm_32x32_8x2_nt_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_32x32_8x2_nt_align1.cu",
- "cutlass_simt_sgemm_32x64_8x2_nt_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_32x64_8x2_nt_align1.cu",
- "cutlass_simt_sgemm_64x32_8x2_nt_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_64x32_8x2_nt_align1.cu",
- "cutlass_simt_sgemm_16x128_8x2_nt_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_16x128_8x2_nt_align1.cu",
- "cutlass_simt_sgemm_32x128_8x2_nt_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_32x128_8x2_nt_align1.cu",
- "cutlass_simt_sgemm_64x64_8x2_nt_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_64x64_8x2_nt_align1.cu",
- "cutlass_simt_sgemm_128x32_8x2_nt_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_128x32_8x2_nt_align1.cu",
- "cutlass_simt_sgemm_64x128_8x2_nt_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_64x128_8x2_nt_align1.cu",
- "cutlass_simt_sgemm_128x64_8x2_nt_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_128x64_8x2_nt_align1.cu",
- "cutlass_simt_sgemm_32x256_8x2_nt_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_32x256_8x2_nt_align1.cu",
- "cutlass_simt_sgemm_64x256_8x2_nt_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_64x256_8x2_nt_align1.cu",
- "cutlass_simt_sgemm_128x128_8x2_nt_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_128x128_8x2_nt_align1.cu",
- "cutlass_simt_sgemm_256x32_8x2_nt_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_256x32_8x2_nt_align1.cu",
- "cutlass_simt_sgemm_256x64_8x2_nt_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_256x64_8x2_nt_align1.cu",
- "cutlass_simt_sgemm_8x32_8x2_tn_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_8x32_8x2_tn_align1.cu",
- "cutlass_simt_sgemm_16x32_8x2_tn_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_16x32_8x2_tn_align1.cu",
- "cutlass_simt_sgemm_16x64_8x2_tn_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_16x64_8x2_tn_align1.cu",
- "cutlass_simt_sgemm_32x32_8x2_tn_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_32x32_8x2_tn_align1.cu",
- "cutlass_simt_sgemm_32x64_8x2_tn_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_32x64_8x2_tn_align1.cu",
- "cutlass_simt_sgemm_64x32_8x2_tn_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_64x32_8x2_tn_align1.cu",
- "cutlass_simt_sgemm_16x128_8x2_tn_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_16x128_8x2_tn_align1.cu",
- "cutlass_simt_sgemm_32x128_8x2_tn_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_32x128_8x2_tn_align1.cu",
- "cutlass_simt_sgemm_64x64_8x2_tn_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_64x64_8x2_tn_align1.cu",
- "cutlass_simt_sgemm_128x32_8x2_tn_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_128x32_8x2_tn_align1.cu",
- "cutlass_simt_sgemm_64x128_8x2_tn_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_64x128_8x2_tn_align1.cu",
- "cutlass_simt_sgemm_128x64_8x2_tn_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_128x64_8x2_tn_align1.cu",
- "cutlass_simt_sgemm_32x256_8x2_tn_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_32x256_8x2_tn_align1.cu",
- "cutlass_simt_sgemm_64x256_8x2_tn_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_64x256_8x2_tn_align1.cu",
- "cutlass_simt_sgemm_128x128_8x2_tn_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_128x128_8x2_tn_align1.cu",
- "cutlass_simt_sgemm_256x32_8x2_tn_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_256x32_8x2_tn_align1.cu",
- "cutlass_simt_sgemm_256x64_8x2_tn_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_256x64_8x2_tn_align1.cu",
- "cutlass_simt_sgemm_8x32_8x2_tt_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_8x32_8x2_tt_align1.cu",
- "cutlass_simt_sgemm_16x32_8x2_tt_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_16x32_8x2_tt_align1.cu",
- "cutlass_simt_sgemm_16x64_8x2_tt_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_16x64_8x2_tt_align1.cu",
- "cutlass_simt_sgemm_32x32_8x2_tt_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_32x32_8x2_tt_align1.cu",
- "cutlass_simt_sgemm_32x64_8x2_tt_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_32x64_8x2_tt_align1.cu",
- "cutlass_simt_sgemm_64x32_8x2_tt_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_64x32_8x2_tt_align1.cu",
- "cutlass_simt_sgemm_16x128_8x2_tt_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_16x128_8x2_tt_align1.cu",
- "cutlass_simt_sgemm_32x128_8x2_tt_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_32x128_8x2_tt_align1.cu",
- "cutlass_simt_sgemm_64x64_8x2_tt_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_64x64_8x2_tt_align1.cu",
- "cutlass_simt_sgemm_128x32_8x2_tt_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_128x32_8x2_tt_align1.cu",
- "cutlass_simt_sgemm_64x128_8x2_tt_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_64x128_8x2_tt_align1.cu",
- "cutlass_simt_sgemm_128x64_8x2_tt_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_128x64_8x2_tt_align1.cu",
- "cutlass_simt_sgemm_32x256_8x2_tt_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_32x256_8x2_tt_align1.cu",
- "cutlass_simt_sgemm_64x256_8x2_tt_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_64x256_8x2_tt_align1.cu",
- "cutlass_simt_sgemm_128x128_8x2_tt_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_128x128_8x2_tt_align1.cu",
- "cutlass_simt_sgemm_256x32_8x2_tt_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_256x32_8x2_tt_align1.cu",
- "cutlass_simt_sgemm_256x64_8x2_tt_align1.cu",
- "cutlass_simt_sgemm_split_k_parallel_256x64_8x2_tt_align1.cu",
- "all_gemm_simt_operations.cu",
- "cutlass_tensorop_f16_s1688gemm_f16_256x128_32x2_nn_align8.cu",
- "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_256x128_32x2_nn_align8.cu",
- "cutlass_tensorop_f16_s1688gemm_f16_128x256_32x2_nn_align8.cu",
- "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_128x256_32x2_nn_align8.cu",
- "cutlass_tensorop_f16_s1688gemm_f16_128x128_32x2_nn_align8.cu",
- "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_128x128_32x2_nn_align8.cu",
- "cutlass_tensorop_f16_s1688gemm_f16_256x128_32x2_nn_align4.cu",
- "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_256x128_32x2_nn_align4.cu",
- "cutlass_tensorop_f16_s1688gemm_f16_128x256_32x2_nn_align4.cu",
- "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_128x256_32x2_nn_align4.cu",
- "cutlass_tensorop_f16_s1688gemm_f16_128x128_32x2_nn_align4.cu",
- "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_128x128_32x2_nn_align4.cu",
- "cutlass_tensorop_f16_s1688gemm_f16_256x128_32x2_nn_align2.cu",
- "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_256x128_32x2_nn_align2.cu",
- "cutlass_tensorop_f16_s1688gemm_f16_128x256_32x2_nn_align2.cu",
- "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_128x256_32x2_nn_align2.cu",
- "cutlass_tensorop_f16_s1688gemm_f16_128x128_32x2_nn_align2.cu",
- "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_128x128_32x2_nn_align2.cu",
- "cutlass_tensorop_f16_s1688gemm_f16_256x128_32x2_nt_align8.cu",
- "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_256x128_32x2_nt_align8.cu",
- "cutlass_tensorop_f16_s1688gemm_f16_128x256_32x2_nt_align8.cu",
- "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_128x256_32x2_nt_align8.cu",
- "cutlass_tensorop_f16_s1688gemm_f16_128x128_32x2_nt_align8.cu",
- "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_128x128_32x2_nt_align8.cu",
- "cutlass_tensorop_f16_s1688gemm_f16_256x128_32x2_nt_align4.cu",
- "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_256x128_32x2_nt_align4.cu",
- "cutlass_tensorop_f16_s1688gemm_f16_128x256_32x2_nt_align4.cu",
- "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_128x256_32x2_nt_align4.cu",
- "cutlass_tensorop_f16_s1688gemm_f16_128x128_32x2_nt_align4.cu",
- "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_128x128_32x2_nt_align4.cu",
- "cutlass_tensorop_f16_s1688gemm_f16_256x128_32x2_nt_align2.cu",
- "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_256x128_32x2_nt_align2.cu",
- "cutlass_tensorop_f16_s1688gemm_f16_128x256_32x2_nt_align2.cu",
- "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_128x256_32x2_nt_align2.cu",
- "cutlass_tensorop_f16_s1688gemm_f16_128x128_32x2_nt_align2.cu",
- "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_128x128_32x2_nt_align2.cu",
- "cutlass_tensorop_f16_s1688gemm_f16_256x128_32x2_tn_align8.cu",
- "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_256x128_32x2_tn_align8.cu",
- "cutlass_tensorop_f16_s1688gemm_f16_128x256_32x2_tn_align8.cu",
- "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_128x256_32x2_tn_align8.cu",
- "cutlass_tensorop_f16_s1688gemm_f16_128x128_32x2_tn_align8.cu",
- "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_128x128_32x2_tn_align8.cu",
- "cutlass_tensorop_f16_s1688gemm_f16_256x128_32x2_tn_align4.cu",
- "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_256x128_32x2_tn_align4.cu",
- "cutlass_tensorop_f16_s1688gemm_f16_128x256_32x2_tn_align4.cu",
- "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_128x256_32x2_tn_align4.cu",
- "cutlass_tensorop_f16_s1688gemm_f16_128x128_32x2_tn_align4.cu",
- "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_128x128_32x2_tn_align4.cu",
- "cutlass_tensorop_f16_s1688gemm_f16_256x128_32x2_tn_align2.cu",
- "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_256x128_32x2_tn_align2.cu",
- "cutlass_tensorop_f16_s1688gemm_f16_128x256_32x2_tn_align2.cu",
- "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_128x256_32x2_tn_align2.cu",
- "cutlass_tensorop_f16_s1688gemm_f16_128x128_32x2_tn_align2.cu",
- "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_128x128_32x2_tn_align2.cu",
- "cutlass_tensorop_f16_s1688gemm_f16_256x128_32x2_tt_align8.cu",
- "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_256x128_32x2_tt_align8.cu",
- "cutlass_tensorop_f16_s1688gemm_f16_128x256_32x2_tt_align8.cu",
- "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_128x256_32x2_tt_align8.cu",
- "cutlass_tensorop_f16_s1688gemm_f16_128x128_32x2_tt_align8.cu",
- "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_128x128_32x2_tt_align8.cu",
- "cutlass_tensorop_f16_s1688gemm_f16_256x128_32x2_tt_align4.cu",
- "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_256x128_32x2_tt_align4.cu",
- "cutlass_tensorop_f16_s1688gemm_f16_128x256_32x2_tt_align4.cu",
- "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_128x256_32x2_tt_align4.cu",
- "cutlass_tensorop_f16_s1688gemm_f16_128x128_32x2_tt_align4.cu",
- "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_128x128_32x2_tt_align4.cu",
- "cutlass_tensorop_f16_s1688gemm_f16_256x128_32x2_tt_align2.cu",
- "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_256x128_32x2_tt_align2.cu",
- "cutlass_tensorop_f16_s1688gemm_f16_128x256_32x2_tt_align2.cu",
- "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_128x256_32x2_tt_align2.cu",
- "cutlass_tensorop_f16_s1688gemm_f16_128x128_32x2_tt_align2.cu",
- "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_128x128_32x2_tt_align2.cu",
- "cutlass_tensorop_h1688gemm_256x128_32x2_nn_align8.cu",
- "cutlass_tensorop_h1688gemm_split_k_parallel_256x128_32x2_nn_align8.cu",
- "cutlass_tensorop_h1688gemm_128x256_32x2_nn_align8.cu",
- "cutlass_tensorop_h1688gemm_split_k_parallel_128x256_32x2_nn_align8.cu",
- "cutlass_tensorop_h1688gemm_128x128_32x2_nn_align8.cu",
- "cutlass_tensorop_h1688gemm_split_k_parallel_128x128_32x2_nn_align8.cu",
- "cutlass_tensorop_h1688gemm_256x128_32x2_nn_align4.cu",
- "cutlass_tensorop_h1688gemm_split_k_parallel_256x128_32x2_nn_align4.cu",
- "cutlass_tensorop_h1688gemm_128x256_32x2_nn_align4.cu",
- "cutlass_tensorop_h1688gemm_split_k_parallel_128x256_32x2_nn_align4.cu",
- "cutlass_tensorop_h1688gemm_128x128_32x2_nn_align4.cu",
- "cutlass_tensorop_h1688gemm_split_k_parallel_128x128_32x2_nn_align4.cu",
- "cutlass_tensorop_h1688gemm_256x128_32x2_nn_align2.cu",
- "cutlass_tensorop_h1688gemm_split_k_parallel_256x128_32x2_nn_align2.cu",
- "cutlass_tensorop_h1688gemm_128x256_32x2_nn_align2.cu",
- "cutlass_tensorop_h1688gemm_split_k_parallel_128x256_32x2_nn_align2.cu",
- "cutlass_tensorop_h1688gemm_128x128_32x2_nn_align2.cu",
- "cutlass_tensorop_h1688gemm_split_k_parallel_128x128_32x2_nn_align2.cu",
- "cutlass_tensorop_h1688gemm_256x128_32x2_nt_align8.cu",
- "cutlass_tensorop_h1688gemm_split_k_parallel_256x128_32x2_nt_align8.cu",
- "cutlass_tensorop_h1688gemm_128x256_32x2_nt_align8.cu",
- "cutlass_tensorop_h1688gemm_split_k_parallel_128x256_32x2_nt_align8.cu",
- "cutlass_tensorop_h1688gemm_128x128_32x2_nt_align8.cu",
- "cutlass_tensorop_h1688gemm_split_k_parallel_128x128_32x2_nt_align8.cu",
- "cutlass_tensorop_h1688gemm_256x128_32x2_nt_align4.cu",
- "cutlass_tensorop_h1688gemm_split_k_parallel_256x128_32x2_nt_align4.cu",
- "cutlass_tensorop_h1688gemm_128x256_32x2_nt_align4.cu",
- "cutlass_tensorop_h1688gemm_split_k_parallel_128x256_32x2_nt_align4.cu",
- "cutlass_tensorop_h1688gemm_128x128_32x2_nt_align4.cu",
- "cutlass_tensorop_h1688gemm_split_k_parallel_128x128_32x2_nt_align4.cu",
- "cutlass_tensorop_h1688gemm_256x128_32x2_nt_align2.cu",
- "cutlass_tensorop_h1688gemm_split_k_parallel_256x128_32x2_nt_align2.cu",
- "cutlass_tensorop_h1688gemm_128x256_32x2_nt_align2.cu",
- "cutlass_tensorop_h1688gemm_split_k_parallel_128x256_32x2_nt_align2.cu",
- "cutlass_tensorop_h1688gemm_128x128_32x2_nt_align2.cu",
- "cutlass_tensorop_h1688gemm_split_k_parallel_128x128_32x2_nt_align2.cu",
- "cutlass_tensorop_h1688gemm_256x128_32x2_tn_align8.cu",
- "cutlass_tensorop_h1688gemm_split_k_parallel_256x128_32x2_tn_align8.cu",
- "cutlass_tensorop_h1688gemm_128x256_32x2_tn_align8.cu",
- "cutlass_tensorop_h1688gemm_split_k_parallel_128x256_32x2_tn_align8.cu",
- "cutlass_tensorop_h1688gemm_128x128_32x2_tn_align8.cu",
- "cutlass_tensorop_h1688gemm_split_k_parallel_128x128_32x2_tn_align8.cu",
- "cutlass_tensorop_h1688gemm_256x128_32x2_tn_align4.cu",
- "cutlass_tensorop_h1688gemm_split_k_parallel_256x128_32x2_tn_align4.cu",
- "cutlass_tensorop_h1688gemm_128x256_32x2_tn_align4.cu",
- "cutlass_tensorop_h1688gemm_split_k_parallel_128x256_32x2_tn_align4.cu",
- "cutlass_tensorop_h1688gemm_128x128_32x2_tn_align4.cu",
- "cutlass_tensorop_h1688gemm_split_k_parallel_128x128_32x2_tn_align4.cu",
- "cutlass_tensorop_h1688gemm_256x128_32x2_tn_align2.cu",
- "cutlass_tensorop_h1688gemm_split_k_parallel_256x128_32x2_tn_align2.cu",
- "cutlass_tensorop_h1688gemm_128x256_32x2_tn_align2.cu",
- "cutlass_tensorop_h1688gemm_split_k_parallel_128x256_32x2_tn_align2.cu",
- "cutlass_tensorop_h1688gemm_128x128_32x2_tn_align2.cu",
- "cutlass_tensorop_h1688gemm_split_k_parallel_128x128_32x2_tn_align2.cu",
- "cutlass_tensorop_h1688gemm_256x128_32x2_tt_align8.cu",
- "cutlass_tensorop_h1688gemm_split_k_parallel_256x128_32x2_tt_align8.cu",
- "cutlass_tensorop_h1688gemm_128x256_32x2_tt_align8.cu",
- "cutlass_tensorop_h1688gemm_split_k_parallel_128x256_32x2_tt_align8.cu",
- "cutlass_tensorop_h1688gemm_128x128_32x2_tt_align8.cu",
- "cutlass_tensorop_h1688gemm_split_k_parallel_128x128_32x2_tt_align8.cu",
- "cutlass_tensorop_h1688gemm_256x128_32x2_tt_align4.cu",
- "cutlass_tensorop_h1688gemm_split_k_parallel_256x128_32x2_tt_align4.cu",
- "cutlass_tensorop_h1688gemm_128x256_32x2_tt_align4.cu",
- "cutlass_tensorop_h1688gemm_split_k_parallel_128x256_32x2_tt_align4.cu",
- "cutlass_tensorop_h1688gemm_128x128_32x2_tt_align4.cu",
- "cutlass_tensorop_h1688gemm_split_k_parallel_128x128_32x2_tt_align4.cu",
- "cutlass_tensorop_h1688gemm_256x128_32x2_tt_align2.cu",
- "cutlass_tensorop_h1688gemm_split_k_parallel_256x128_32x2_tt_align2.cu",
- "cutlass_tensorop_h1688gemm_128x256_32x2_tt_align2.cu",
- "cutlass_tensorop_h1688gemm_split_k_parallel_128x256_32x2_tt_align2.cu",
- "cutlass_tensorop_h1688gemm_128x128_32x2_tt_align2.cu",
- "cutlass_tensorop_h1688gemm_split_k_parallel_128x128_32x2_tt_align2.cu",
- "all_gemm_tensorop1688_operations.cu",
- "cutlass_tensorop_f16_s884gemm_f16_256x128_32x2_nn_align8.cu",
- "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_256x128_32x2_nn_align8.cu",
- "cutlass_tensorop_f16_s884gemm_f16_128x256_32x2_nn_align8.cu",
- "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_128x256_32x2_nn_align8.cu",
- "cutlass_tensorop_f16_s884gemm_f16_128x128_32x2_nn_align8.cu",
- "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_128x128_32x2_nn_align8.cu",
- "cutlass_tensorop_f16_s884gemm_f16_256x128_32x2_nn_align4.cu",
- "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_256x128_32x2_nn_align4.cu",
- "cutlass_tensorop_f16_s884gemm_f16_128x256_32x2_nn_align4.cu",
- "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_128x256_32x2_nn_align4.cu",
- "cutlass_tensorop_f16_s884gemm_f16_128x128_32x2_nn_align4.cu",
- "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_128x128_32x2_nn_align4.cu",
- "cutlass_tensorop_f16_s884gemm_f16_256x128_32x2_nn_align2.cu",
- "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_256x128_32x2_nn_align2.cu",
- "cutlass_tensorop_f16_s884gemm_f16_128x256_32x2_nn_align2.cu",
- "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_128x256_32x2_nn_align2.cu",
- "cutlass_tensorop_f16_s884gemm_f16_128x128_32x2_nn_align2.cu",
- "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_128x128_32x2_nn_align2.cu",
- "cutlass_tensorop_f16_s884gemm_f16_256x128_32x2_nt_align8.cu",
- "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_256x128_32x2_nt_align8.cu",
- "cutlass_tensorop_f16_s884gemm_f16_128x256_32x2_nt_align8.cu",
- "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_128x256_32x2_nt_align8.cu",
- "cutlass_tensorop_f16_s884gemm_f16_128x128_32x2_nt_align8.cu",
- "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_128x128_32x2_nt_align8.cu",
- "cutlass_tensorop_f16_s884gemm_f16_256x128_32x2_nt_align4.cu",
- "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_256x128_32x2_nt_align4.cu",
- "cutlass_tensorop_f16_s884gemm_f16_128x256_32x2_nt_align4.cu",
- "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_128x256_32x2_nt_align4.cu",
- "cutlass_tensorop_f16_s884gemm_f16_128x128_32x2_nt_align4.cu",
- "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_128x128_32x2_nt_align4.cu",
- "cutlass_tensorop_f16_s884gemm_f16_256x128_32x2_nt_align2.cu",
- "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_256x128_32x2_nt_align2.cu",
- "cutlass_tensorop_f16_s884gemm_f16_128x256_32x2_nt_align2.cu",
- "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_128x256_32x2_nt_align2.cu",
- "cutlass_tensorop_f16_s884gemm_f16_128x128_32x2_nt_align2.cu",
- "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_128x128_32x2_nt_align2.cu",
- "cutlass_tensorop_f16_s884gemm_f16_256x128_32x2_tn_align8.cu",
- "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_256x128_32x2_tn_align8.cu",
- "cutlass_tensorop_f16_s884gemm_f16_128x256_32x2_tn_align8.cu",
- "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_128x256_32x2_tn_align8.cu",
- "cutlass_tensorop_f16_s884gemm_f16_128x128_32x2_tn_align8.cu",
- "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_128x128_32x2_tn_align8.cu",
- "cutlass_tensorop_f16_s884gemm_f16_256x128_32x2_tn_align4.cu",
- "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_256x128_32x2_tn_align4.cu",
- "cutlass_tensorop_f16_s884gemm_f16_128x256_32x2_tn_align4.cu",
- "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_128x256_32x2_tn_align4.cu",
- "cutlass_tensorop_f16_s884gemm_f16_128x128_32x2_tn_align4.cu",
- "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_128x128_32x2_tn_align4.cu",
- "cutlass_tensorop_f16_s884gemm_f16_256x128_32x2_tn_align2.cu",
- "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_256x128_32x2_tn_align2.cu",
- "cutlass_tensorop_f16_s884gemm_f16_128x256_32x2_tn_align2.cu",
- "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_128x256_32x2_tn_align2.cu",
- "cutlass_tensorop_f16_s884gemm_f16_128x128_32x2_tn_align2.cu",
- "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_128x128_32x2_tn_align2.cu",
- "cutlass_tensorop_f16_s884gemm_f16_256x128_32x2_tt_align8.cu",
- "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_256x128_32x2_tt_align8.cu",
- "cutlass_tensorop_f16_s884gemm_f16_128x256_32x2_tt_align8.cu",
- "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_128x256_32x2_tt_align8.cu",
- "cutlass_tensorop_f16_s884gemm_f16_128x128_32x2_tt_align8.cu",
- "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_128x128_32x2_tt_align8.cu",
- "cutlass_tensorop_f16_s884gemm_f16_256x128_32x2_tt_align4.cu",
- "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_256x128_32x2_tt_align4.cu",
- "cutlass_tensorop_f16_s884gemm_f16_128x256_32x2_tt_align4.cu",
- "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_128x256_32x2_tt_align4.cu",
- "cutlass_tensorop_f16_s884gemm_f16_128x128_32x2_tt_align4.cu",
- "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_128x128_32x2_tt_align4.cu",
- "cutlass_tensorop_f16_s884gemm_f16_256x128_32x2_tt_align2.cu",
- "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_256x128_32x2_tt_align2.cu",
- "cutlass_tensorop_f16_s884gemm_f16_128x256_32x2_tt_align2.cu",
- "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_128x256_32x2_tt_align2.cu",
- "cutlass_tensorop_f16_s884gemm_f16_128x128_32x2_tt_align2.cu",
- "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_128x128_32x2_tt_align2.cu",
- "cutlass_tensorop_h884gemm_256x128_32x2_nn_align8.cu",
- "cutlass_tensorop_h884gemm_split_k_parallel_256x128_32x2_nn_align8.cu",
- "cutlass_tensorop_h884gemm_128x256_32x2_nn_align8.cu",
- "cutlass_tensorop_h884gemm_split_k_parallel_128x256_32x2_nn_align8.cu",
- "cutlass_tensorop_h884gemm_128x128_32x2_nn_align8.cu",
- "cutlass_tensorop_h884gemm_split_k_parallel_128x128_32x2_nn_align8.cu",
- "cutlass_tensorop_h884gemm_256x128_32x2_nn_align4.cu",
- "cutlass_tensorop_h884gemm_split_k_parallel_256x128_32x2_nn_align4.cu",
- "cutlass_tensorop_h884gemm_128x256_32x2_nn_align4.cu",
- "cutlass_tensorop_h884gemm_split_k_parallel_128x256_32x2_nn_align4.cu",
- "cutlass_tensorop_h884gemm_128x128_32x2_nn_align4.cu",
- "cutlass_tensorop_h884gemm_split_k_parallel_128x128_32x2_nn_align4.cu",
- "cutlass_tensorop_h884gemm_256x128_32x2_nn_align2.cu",
- "cutlass_tensorop_h884gemm_split_k_parallel_256x128_32x2_nn_align2.cu",
- "cutlass_tensorop_h884gemm_128x256_32x2_nn_align2.cu",
- "cutlass_tensorop_h884gemm_split_k_parallel_128x256_32x2_nn_align2.cu",
- "cutlass_tensorop_h884gemm_128x128_32x2_nn_align2.cu",
- "cutlass_tensorop_h884gemm_split_k_parallel_128x128_32x2_nn_align2.cu",
- "cutlass_tensorop_h884gemm_256x128_32x2_nt_align8.cu",
- "cutlass_tensorop_h884gemm_split_k_parallel_256x128_32x2_nt_align8.cu",
- "cutlass_tensorop_h884gemm_128x256_32x2_nt_align8.cu",
- "cutlass_tensorop_h884gemm_split_k_parallel_128x256_32x2_nt_align8.cu",
- "cutlass_tensorop_h884gemm_128x128_32x2_nt_align8.cu",
- "cutlass_tensorop_h884gemm_split_k_parallel_128x128_32x2_nt_align8.cu",
- "cutlass_tensorop_h884gemm_256x128_32x2_nt_align4.cu",
- "cutlass_tensorop_h884gemm_split_k_parallel_256x128_32x2_nt_align4.cu",
- "cutlass_tensorop_h884gemm_128x256_32x2_nt_align4.cu",
- "cutlass_tensorop_h884gemm_split_k_parallel_128x256_32x2_nt_align4.cu",
- "cutlass_tensorop_h884gemm_128x128_32x2_nt_align4.cu",
- "cutlass_tensorop_h884gemm_split_k_parallel_128x128_32x2_nt_align4.cu",
- "cutlass_tensorop_h884gemm_256x128_32x2_nt_align2.cu",
- "cutlass_tensorop_h884gemm_split_k_parallel_256x128_32x2_nt_align2.cu",
- "cutlass_tensorop_h884gemm_128x256_32x2_nt_align2.cu",
- "cutlass_tensorop_h884gemm_split_k_parallel_128x256_32x2_nt_align2.cu",
- "cutlass_tensorop_h884gemm_128x128_32x2_nt_align2.cu",
- "cutlass_tensorop_h884gemm_split_k_parallel_128x128_32x2_nt_align2.cu",
- "cutlass_tensorop_h884gemm_256x128_32x2_tn_align8.cu",
- "cutlass_tensorop_h884gemm_split_k_parallel_256x128_32x2_tn_align8.cu",
- "cutlass_tensorop_h884gemm_128x256_32x2_tn_align8.cu",
- "cutlass_tensorop_h884gemm_split_k_parallel_128x256_32x2_tn_align8.cu",
- "cutlass_tensorop_h884gemm_128x128_32x2_tn_align8.cu",
- "cutlass_tensorop_h884gemm_split_k_parallel_128x128_32x2_tn_align8.cu",
- "cutlass_tensorop_h884gemm_256x128_32x2_tn_align4.cu",
- "cutlass_tensorop_h884gemm_split_k_parallel_256x128_32x2_tn_align4.cu",
- "cutlass_tensorop_h884gemm_128x256_32x2_tn_align4.cu",
- "cutlass_tensorop_h884gemm_split_k_parallel_128x256_32x2_tn_align4.cu",
- "cutlass_tensorop_h884gemm_128x128_32x2_tn_align4.cu",
- "cutlass_tensorop_h884gemm_split_k_parallel_128x128_32x2_tn_align4.cu",
- "cutlass_tensorop_h884gemm_256x128_32x2_tn_align2.cu",
- "cutlass_tensorop_h884gemm_split_k_parallel_256x128_32x2_tn_align2.cu",
- "cutlass_tensorop_h884gemm_128x256_32x2_tn_align2.cu",
- "cutlass_tensorop_h884gemm_split_k_parallel_128x256_32x2_tn_align2.cu",
- "cutlass_tensorop_h884gemm_128x128_32x2_tn_align2.cu",
- "cutlass_tensorop_h884gemm_split_k_parallel_128x128_32x2_tn_align2.cu",
- "cutlass_tensorop_h884gemm_256x128_32x2_tt_align8.cu",
- "cutlass_tensorop_h884gemm_split_k_parallel_256x128_32x2_tt_align8.cu",
- "cutlass_tensorop_h884gemm_128x256_32x2_tt_align8.cu",
- "cutlass_tensorop_h884gemm_split_k_parallel_128x256_32x2_tt_align8.cu",
- "cutlass_tensorop_h884gemm_128x128_32x2_tt_align8.cu",
- "cutlass_tensorop_h884gemm_split_k_parallel_128x128_32x2_tt_align8.cu",
- "cutlass_tensorop_h884gemm_256x128_32x2_tt_align4.cu",
- "cutlass_tensorop_h884gemm_split_k_parallel_256x128_32x2_tt_align4.cu",
- "cutlass_tensorop_h884gemm_128x256_32x2_tt_align4.cu",
- "cutlass_tensorop_h884gemm_split_k_parallel_128x256_32x2_tt_align4.cu",
- "cutlass_tensorop_h884gemm_128x128_32x2_tt_align4.cu",
- "cutlass_tensorop_h884gemm_split_k_parallel_128x128_32x2_tt_align4.cu",
- "cutlass_tensorop_h884gemm_256x128_32x2_tt_align2.cu",
- "cutlass_tensorop_h884gemm_split_k_parallel_256x128_32x2_tt_align2.cu",
- "cutlass_tensorop_h884gemm_128x256_32x2_tt_align2.cu",
- "cutlass_tensorop_h884gemm_split_k_parallel_128x256_32x2_tt_align2.cu",
- "cutlass_tensorop_h884gemm_128x128_32x2_tt_align2.cu",
- "cutlass_tensorop_h884gemm_split_k_parallel_128x128_32x2_tt_align2.cu",
- "all_gemm_tensorop884_operations.cu",
- "cutlass_simt_sgemv_batched_strided_1x128_32_tt_align4x4.cu",
- "cutlass_simt_sgemv_batched_strided_1x128_16_tt_align4x2.cu",
- "cutlass_simt_sgemv_batched_strided_1x128_8_tt_align4x1.cu",
- "cutlass_simt_sgemv_batched_strided_1x128_16_tt_align2x4.cu",
- "cutlass_simt_sgemv_batched_strided_1x128_8_tt_align2x2.cu",
- "cutlass_simt_sgemv_batched_strided_1x128_4_tt_align2x1.cu",
- "cutlass_simt_sgemv_batched_strided_1x128_8_tt_align1x4.cu",
- "cutlass_simt_sgemv_batched_strided_1x128_4_tt_align1x2.cu",
- "cutlass_simt_sgemv_batched_strided_1x128_2_tt_align1x1.cu",
- "cutlass_simt_sgemv_batched_strided_1x64_64_tt_align4x4.cu",
- "cutlass_simt_sgemv_batched_strided_1x64_32_tt_align4x2.cu",
- "cutlass_simt_sgemv_batched_strided_1x64_16_tt_align4x1.cu",
- "cutlass_simt_sgemv_batched_strided_1x64_32_tt_align2x4.cu",
- "cutlass_simt_sgemv_batched_strided_1x64_16_tt_align2x2.cu",
- "cutlass_simt_sgemv_batched_strided_1x64_8_tt_align2x1.cu",
- "cutlass_simt_sgemv_batched_strided_1x64_16_tt_align1x4.cu",
- "cutlass_simt_sgemv_batched_strided_1x64_8_tt_align1x2.cu",
- "cutlass_simt_sgemv_batched_strided_1x64_4_tt_align1x1.cu",
- "cutlass_simt_sgemv_batched_strided_1x32_128_tt_align4x4.cu",
- "cutlass_simt_sgemv_batched_strided_1x32_64_tt_align4x2.cu",
- "cutlass_simt_sgemv_batched_strided_1x32_32_tt_align4x1.cu",
- "cutlass_simt_sgemv_batched_strided_1x32_64_tt_align2x4.cu",
- "cutlass_simt_sgemv_batched_strided_1x32_32_tt_align2x2.cu",
- "cutlass_simt_sgemv_batched_strided_1x32_16_tt_align2x1.cu",
- "cutlass_simt_sgemv_batched_strided_1x32_32_tt_align1x4.cu",
- "cutlass_simt_sgemv_batched_strided_1x32_16_tt_align1x2.cu",
- "cutlass_simt_sgemv_batched_strided_1x32_8_tt_align1x1.cu",
- "cutlass_simt_s8_idgrad_id_s8_32x128x32_32x64x32_2_nc4hw4_k4rsc4_align4x16.cu",
- "cutlass_simt_s8_idgrad_s2_id_s8_32x128x32_32x64x32_2_nc4hw4_k4rsc4_align4x16.cu",
- "cutlass_simt_s8_idgrad_id_s8_16x128x16_16x64x16_2_nc4hw4_k4rsc4_align4x4.cu",
- "cutlass_simt_s8_idgrad_s2_id_s8_16x128x16_16x64x16_2_nc4hw4_k4rsc4_align4x4.cu",
- "cutlass_simt_s8_idgrad_id_s8_16x128x16_16x128x16_1_nc4hw4_k4rsc4_align4x8.cu",
- "cutlass_simt_s8_idgrad_s2_id_s8_16x128x16_16x128x16_1_nc4hw4_k4rsc4_align4x8.cu",
- "cutlass_simt_s8_idgrad_id_s8_16x64x8_16x64x8_2_nc4hw4_k4rsc4_align4x4.cu",
- "cutlass_simt_s8_idgrad_s2_id_s8_16x64x8_16x64x8_2_nc4hw4_k4rsc4_align4x4.cu",
- "all_deconv_simt_operations.cu",
- "cutlass_tensorop_s8_i8816dgrad_id_s8_128x32x32_64x32x32_1_nhwc_ck4rs4_align4x4.cu",
- "cutlass_tensorop_s8_i8816dgrad_s2_id_s8_128x32x32_64x32x32_1_nhwc_ck4rs4_align4x4.cu",
- "cutlass_tensorop_s8_i8816dgrad_id_s8_64x16x32_64x16x32_2_nhwc_ck4rs4_align4x4.cu",
- "cutlass_tensorop_s8_i8816dgrad_s2_id_s8_64x16x32_64x16x32_2_nhwc_ck4rs4_align4x4.cu",
- "cutlass_tensorop_s8_i8816dgrad_id_s8_128x32x32_64x32x32_1_nhwc_ck8rs8_align8x8.cu",
- "cutlass_tensorop_s8_i8816dgrad_s2_id_s8_128x32x32_64x32x32_1_nhwc_ck8rs8_align8x8.cu",
- "cutlass_tensorop_s8_i8816dgrad_id_s8_64x16x32_64x16x32_2_nhwc_ck8rs8_align8x8.cu",
- "cutlass_tensorop_s8_i8816dgrad_s2_id_s8_64x16x32_64x16x32_2_nhwc_ck8rs8_align8x8.cu",
- "cutlass_tensorop_s8_i8816dgrad_id_s8_128x32x32_64x32x32_1_nhwc_ck16rs16_align16x16.cu",
- "cutlass_tensorop_s8_i8816dgrad_s2_id_s8_128x32x32_64x32x32_1_nhwc_ck16rs16_align16x16.cu",
- "cutlass_tensorop_s8_i8816dgrad_id_s8_64x16x32_64x16x32_2_nhwc_ck16rs16_align16x16.cu",
- "cutlass_tensorop_s8_i8816dgrad_s2_id_s8_64x16x32_64x16x32_2_nhwc_ck16rs16_align16x16.cu",
- "all_deconv_tensorop8816_operations.cu",
- "cutlass_simt_s8_ifprop_id_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
- "cutlass_simt_s8_ifprop_1x1_id_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
- "cutlass_simt_s8_ifprop_relu_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
- "cutlass_simt_s8_ifprop_1x1_relu_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
- "cutlass_simt_s8_ifprop_hswish_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
- "cutlass_simt_s8_ifprop_1x1_hswish_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
- "cutlass_simt_s8_ifprop_id_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
- "cutlass_simt_s8_ifprop_1x1_id_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
- "cutlass_simt_s8_ifprop_relu_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
- "cutlass_simt_s8_ifprop_1x1_relu_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
- "cutlass_simt_s8_ifprop_hswish_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
- "cutlass_simt_s8_ifprop_1x1_hswish_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
- "cutlass_simt_s8_ifprop_id_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
- "cutlass_simt_s8_ifprop_1x1_id_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
- "cutlass_simt_s8_ifprop_relu_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
- "cutlass_simt_s8_ifprop_1x1_relu_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
- "cutlass_simt_s8_ifprop_hswish_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
- "cutlass_simt_s8_ifprop_1x1_hswish_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
- "cutlass_simt_s8_ifprop_id_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
- "cutlass_simt_s8_ifprop_1x1_id_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
- "cutlass_simt_s8_ifprop_relu_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
- "cutlass_simt_s8_ifprop_1x1_relu_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
- "cutlass_simt_s8_ifprop_hswish_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
- "cutlass_simt_s8_ifprop_1x1_hswish_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
- "cutlass_simt_s8_ifprop_id_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_align4x16.cu",
- "cutlass_simt_s8_ifprop_1x1_id_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_align4x16.cu",
- "cutlass_simt_s8_ifprop_relu_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_align4x16.cu",
- "cutlass_simt_s8_ifprop_1x1_relu_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_align4x16.cu",
- "cutlass_simt_s8_ifprop_hswish_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_align4x16.cu",
- "cutlass_simt_s8_ifprop_1x1_hswish_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_align4x16.cu",
- "cutlass_simt_s8_ifprop_id_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_align4x16.cu",
- "cutlass_simt_s8_ifprop_1x1_id_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_align4x16.cu",
- "cutlass_simt_s8_ifprop_relu_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_align4x16.cu",
- "cutlass_simt_s8_ifprop_1x1_relu_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_align4x16.cu",
- "cutlass_simt_s8_ifprop_hswish_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_align4x16.cu",
- "cutlass_simt_s8_ifprop_1x1_hswish_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_align4x16.cu",
- "cutlass_simt_s8_ifprop_id_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
- "cutlass_simt_s8_ifprop_1x1_id_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
- "cutlass_simt_s8_ifprop_relu_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
- "cutlass_simt_s8_ifprop_1x1_relu_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
- "cutlass_simt_s8_ifprop_hswish_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
- "cutlass_simt_s8_ifprop_1x1_hswish_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
- "cutlass_simt_s8_ifprop_id_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_align4x8.cu",
- "cutlass_simt_s8_ifprop_1x1_id_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_align4x8.cu",
- "cutlass_simt_s8_ifprop_relu_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_align4x8.cu",
- "cutlass_simt_s8_ifprop_1x1_relu_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_align4x8.cu",
- "cutlass_simt_s8_ifprop_hswish_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_align4x8.cu",
- "cutlass_simt_s8_ifprop_1x1_hswish_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_align4x8.cu",
- "cutlass_simt_s8_ifprop_id_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_align4x4.cu",
- "cutlass_simt_s8_ifprop_1x1_id_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_align4x4.cu",
- "cutlass_simt_s8_ifprop_relu_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_align4x4.cu",
- "cutlass_simt_s8_ifprop_1x1_relu_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_align4x4.cu",
- "cutlass_simt_s8_ifprop_hswish_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_align4x4.cu",
- "cutlass_simt_s8_ifprop_1x1_hswish_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_align4x4.cu",
- "cutlass_simt_s8_ifprop_id_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32_align4x16.cu",
- "cutlass_simt_s8_ifprop_1x1_id_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32_align4x16.cu",
- "cutlass_simt_s8_ifprop_relu_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32_align4x16.cu",
- "cutlass_simt_s8_ifprop_1x1_relu_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32_align4x16.cu",
- "cutlass_simt_s8_ifprop_hswish_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32_align4x16.cu",
- "cutlass_simt_s8_ifprop_1x1_hswish_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32_align4x16.cu",
- "cutlass_simt_s8_ifprop_id_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32_align4x16.cu",
- "cutlass_simt_s8_ifprop_1x1_id_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32_align4x16.cu",
- "cutlass_simt_s8_ifprop_relu_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32_align4x16.cu",
- "cutlass_simt_s8_ifprop_1x1_relu_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32_align4x16.cu",
- "cutlass_simt_s8_ifprop_hswish_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32_align4x16.cu",
- "cutlass_simt_s8_ifprop_1x1_hswish_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32_align4x16.cu",
- "cutlass_simt_u4_ifprop_id_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nhwc_align4x8.cu",
- "cutlass_simt_u4_ifprop_relu_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nhwc_align4x8.cu",
- "cutlass_simt_u4_ifprop_hswish_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nhwc_align4x8.cu",
- "cutlass_simt_u4_ifprop_id_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nhwc_align4x4.cu",
- "cutlass_simt_u4_ifprop_relu_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nhwc_align4x4.cu",
- "cutlass_simt_u4_ifprop_hswish_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nhwc_align4x4.cu",
- "cutlass_simt_s4_ifprop_id_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nhwc_align4x8.cu",
- "cutlass_simt_s4_ifprop_relu_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nhwc_align4x8.cu",
- "cutlass_simt_s4_ifprop_hswish_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nhwc_align4x8.cu",
- "cutlass_simt_s4_ifprop_id_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nhwc_align4x4.cu",
- "cutlass_simt_s4_ifprop_relu_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nhwc_align4x4.cu",
- "cutlass_simt_s4_ifprop_hswish_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nhwc_align4x4.cu",
- "cutlass_simt_f32_ifprop_id_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nchw_align4x8.cu",
- "cutlass_simt_f32_ifprop_1x1_id_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nchw_align4x8.cu",
- "cutlass_simt_f32_ifprop_relu_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nchw_align4x8.cu",
- "cutlass_simt_f32_ifprop_1x1_relu_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nchw_align4x8.cu",
- "cutlass_simt_f32_ifprop_hswish_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nchw_align4x8.cu",
- "cutlass_simt_f32_ifprop_1x1_hswish_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nchw_align4x8.cu",
- "cutlass_simt_f32_ifprop_id_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nchw_align4x4.cu",
- "cutlass_simt_f32_ifprop_1x1_id_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nchw_align4x4.cu",
- "cutlass_simt_f32_ifprop_relu_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nchw_align4x4.cu",
- "cutlass_simt_f32_ifprop_1x1_relu_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nchw_align4x4.cu",
- "cutlass_simt_f32_ifprop_hswish_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nchw_align4x4.cu",
- "cutlass_simt_f32_ifprop_1x1_hswish_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nchw_align4x4.cu",
- "all_conv2d_simt_operations.cu",
- "cutlass_tensorop_s8_i8816fprop_roc_id_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_roc_id_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_roc_relu_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_roc_relu_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_roc_hswish_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_roc_hswish_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_roc_id_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_roc_id_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_roc_relu_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_roc_relu_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_roc_hswish_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_roc_hswish_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_roc_id_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_roc_id_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_roc_relu_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_roc_relu_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_roc_hswish_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_roc_hswish_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_roc_id_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_roc_id_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_roc_relu_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_roc_relu_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_roc_hswish_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_roc_hswish_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_roc_id_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_roc_id_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_roc_relu_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_roc_relu_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_roc_hswish_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_roc_hswish_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_roc_id_s8_128x64x32_64x32x32_1_nc32hw32_c32rsk32_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_roc_id_s8_128x64x32_64x32x32_1_nc32hw32_c32rsk32_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_roc_relu_s8_128x64x32_64x32x32_1_nc32hw32_c32rsk32_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_roc_relu_s8_128x64x32_64x32x32_1_nc32hw32_c32rsk32_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_roc_hswish_s8_128x64x32_64x32x32_1_nc32hw32_c32rsk32_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_roc_hswish_s8_128x64x32_64x32x32_1_nc32hw32_c32rsk32_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_roc_id_s8_128x32x32_64x32x32_1_nc32hw32_c32rsk32_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_roc_id_s8_128x32x32_64x32x32_1_nc32hw32_c32rsk32_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_roc_relu_s8_128x32x32_64x32x32_1_nc32hw32_c32rsk32_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_roc_relu_s8_128x32x32_64x32x32_1_nc32hw32_c32rsk32_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_roc_hswish_s8_128x32x32_64x32x32_1_nc32hw32_c32rsk32_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_roc_hswish_s8_128x32x32_64x32x32_1_nc32hw32_c32rsk32_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_id_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32_nc4hw4_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_id_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32_nc4hw4_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_relu_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32_nc4hw4_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32_nc4hw4_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_hswish_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32_nc4hw4_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32_nc4hw4_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_id_s8_32x128x32_32x64x32_1_nc32hw32_c32rsk32_nc4hw4_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_id_s8_32x128x32_32x64x32_1_nc32hw32_c32rsk32_nc4hw4_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_relu_s8_32x128x32_32x64x32_1_nc32hw32_c32rsk32_nc4hw4_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_32x128x32_32x64x32_1_nc32hw32_c32rsk32_nc4hw4_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_hswish_s8_32x128x32_32x64x32_1_nc32hw32_c32rsk32_nc4hw4_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_32x128x32_32x64x32_1_nc32hw32_c32rsk32_nc4hw4_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_id_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_id_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_s8_i8816fprop_relu_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_s8_i8816fprop_hswish_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_s8_i8816fprop_roc_id_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_roc_id_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_s8_i8816fprop_roc_relu_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_roc_relu_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_s8_i8816fprop_roc_hswish_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_roc_hswish_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_s8_i8816fprop_id_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_id_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_s8_i8816fprop_relu_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_s8_i8816fprop_hswish_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_s8_i8816fprop_roc_id_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_roc_id_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_s8_i8816fprop_roc_relu_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_roc_relu_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_s8_i8816fprop_roc_hswish_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_roc_hswish_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_s8_i8816fprop_id_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_id_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8816fprop_relu_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8816fprop_hswish_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8816fprop_roc_id_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_roc_id_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8816fprop_roc_relu_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_roc_relu_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8816fprop_roc_hswish_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_roc_hswish_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8816fprop_id_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_id_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8816fprop_relu_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8816fprop_hswish_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8816fprop_roc_id_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_roc_id_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8816fprop_roc_relu_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_roc_relu_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8816fprop_roc_hswish_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_roc_hswish_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8816fprop_id_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_id_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_relu_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_hswish_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_roc_id_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_roc_id_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_roc_relu_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_roc_relu_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_roc_hswish_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_roc_hswish_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_id_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_id_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_relu_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_hswish_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_roc_id_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_roc_id_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_roc_relu_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_roc_relu_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_roc_hswish_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8816fprop_1x1_roc_hswish_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s4_i8816fprop_id_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_s4_i8816fprop_1x1_id_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_s4_i8816fprop_relu_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_s4_i8816fprop_1x1_relu_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_s4_i8816fprop_hswish_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_s4_i8816fprop_1x1_hswish_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_s4_i8816fprop_roc_id_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_s4_i8816fprop_1x1_roc_id_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_s4_i8816fprop_roc_relu_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_s4_i8816fprop_1x1_roc_relu_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_s4_i8816fprop_roc_hswish_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_s4_i8816fprop_1x1_roc_hswish_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_s4_i8816fprop_id_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_s4_i8816fprop_1x1_id_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_s4_i8816fprop_relu_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_s4_i8816fprop_1x1_relu_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_s4_i8816fprop_hswish_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_s4_i8816fprop_1x1_hswish_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_s4_i8816fprop_roc_id_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_s4_i8816fprop_1x1_roc_id_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_s4_i8816fprop_roc_relu_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_s4_i8816fprop_1x1_roc_relu_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_s4_i8816fprop_roc_hswish_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_s4_i8816fprop_1x1_roc_hswish_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_u4_i8816fprop_id_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_u4_i8816fprop_1x1_id_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_u4_i8816fprop_relu_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_u4_i8816fprop_1x1_relu_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_u4_i8816fprop_hswish_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_u4_i8816fprop_1x1_hswish_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_u4_i8816fprop_roc_id_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_u4_i8816fprop_1x1_roc_id_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_u4_i8816fprop_roc_relu_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_u4_i8816fprop_1x1_roc_relu_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_u4_i8816fprop_roc_hswish_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_u4_i8816fprop_1x1_roc_hswish_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_u4_i8816fprop_id_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_u4_i8816fprop_1x1_id_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_u4_i8816fprop_relu_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_u4_i8816fprop_1x1_relu_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_u4_i8816fprop_hswish_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_u4_i8816fprop_1x1_hswish_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_u4_i8816fprop_roc_id_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_u4_i8816fprop_1x1_roc_id_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_u4_i8816fprop_roc_relu_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_u4_i8816fprop_1x1_roc_relu_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_u4_i8816fprop_roc_hswish_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_u4_i8816fprop_1x1_roc_hswish_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_f32_i8816fprop_id_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_f32_i8816fprop_1x1_id_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_f32_i8816fprop_relu_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_f32_i8816fprop_1x1_relu_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_f32_i8816fprop_hswish_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_f32_i8816fprop_1x1_hswish_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_f32_i8816fprop_id_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_f32_i8816fprop_1x1_id_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_f32_i8816fprop_relu_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_f32_i8816fprop_1x1_relu_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_f32_i8816fprop_hswish_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_f32_i8816fprop_1x1_hswish_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_f32_i8816fprop_roc_id_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_f32_i8816fprop_1x1_roc_id_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_f32_i8816fprop_roc_relu_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_f32_i8816fprop_1x1_roc_relu_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_f32_i8816fprop_roc_hswish_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_f32_i8816fprop_1x1_roc_hswish_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
- "cutlass_tensorop_s4_i8816fprop_id_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s4_i8816fprop_1x1_id_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s4_i8816fprop_relu_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s4_i8816fprop_1x1_relu_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s4_i8816fprop_hswish_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s4_i8816fprop_1x1_hswish_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s4_i8816fprop_roc_id_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s4_i8816fprop_1x1_roc_id_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s4_i8816fprop_roc_relu_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s4_i8816fprop_1x1_roc_relu_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s4_i8816fprop_roc_hswish_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s4_i8816fprop_1x1_roc_hswish_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s4_i8816fprop_id_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s4_i8816fprop_1x1_id_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s4_i8816fprop_relu_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s4_i8816fprop_1x1_relu_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s4_i8816fprop_hswish_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s4_i8816fprop_1x1_hswish_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s4_i8816fprop_roc_id_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s4_i8816fprop_1x1_roc_id_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s4_i8816fprop_roc_relu_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s4_i8816fprop_1x1_roc_relu_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s4_i8816fprop_roc_hswish_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s4_i8816fprop_1x1_roc_hswish_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_u4_i8816fprop_id_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_u4_i8816fprop_1x1_id_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_u4_i8816fprop_relu_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_u4_i8816fprop_1x1_relu_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_u4_i8816fprop_hswish_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_u4_i8816fprop_1x1_hswish_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_u4_i8816fprop_roc_id_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_u4_i8816fprop_1x1_roc_id_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_u4_i8816fprop_roc_relu_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_u4_i8816fprop_1x1_roc_relu_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_u4_i8816fprop_roc_hswish_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_u4_i8816fprop_1x1_roc_hswish_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_u4_i8816fprop_id_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_u4_i8816fprop_1x1_id_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_u4_i8816fprop_relu_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_u4_i8816fprop_1x1_relu_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_u4_i8816fprop_hswish_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_u4_i8816fprop_1x1_hswish_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_u4_i8816fprop_roc_id_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_u4_i8816fprop_1x1_roc_id_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_u4_i8816fprop_roc_relu_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_u4_i8816fprop_1x1_roc_relu_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_u4_i8816fprop_roc_hswish_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_u4_i8816fprop_1x1_roc_hswish_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_f32_i8816fprop_id_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_f32_i8816fprop_1x1_id_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_f32_i8816fprop_relu_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_f32_i8816fprop_1x1_relu_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_f32_i8816fprop_hswish_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_f32_i8816fprop_1x1_hswish_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_f32_i8816fprop_id_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_f32_i8816fprop_1x1_id_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_f32_i8816fprop_relu_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_f32_i8816fprop_1x1_relu_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_f32_i8816fprop_hswish_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_f32_i8816fprop_1x1_hswish_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_f32_i8816fprop_roc_id_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_f32_i8816fprop_1x1_roc_id_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_f32_i8816fprop_roc_relu_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_f32_i8816fprop_1x1_roc_relu_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_f32_i8816fprop_roc_hswish_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_f32_i8816fprop_1x1_roc_hswish_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s4_i8816fprop_id_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s4_i8816fprop_1x1_id_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s4_i8816fprop_relu_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s4_i8816fprop_1x1_relu_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s4_i8816fprop_hswish_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s4_i8816fprop_1x1_hswish_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s4_i8816fprop_roc_id_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s4_i8816fprop_1x1_roc_id_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s4_i8816fprop_roc_relu_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s4_i8816fprop_1x1_roc_relu_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s4_i8816fprop_roc_hswish_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s4_i8816fprop_1x1_roc_hswish_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s4_i8816fprop_id_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s4_i8816fprop_1x1_id_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s4_i8816fprop_relu_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s4_i8816fprop_1x1_relu_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s4_i8816fprop_hswish_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s4_i8816fprop_1x1_hswish_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s4_i8816fprop_roc_id_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s4_i8816fprop_1x1_roc_id_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s4_i8816fprop_roc_relu_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s4_i8816fprop_1x1_roc_relu_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s4_i8816fprop_roc_hswish_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s4_i8816fprop_1x1_roc_hswish_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_u4_i8816fprop_id_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_u4_i8816fprop_1x1_id_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_u4_i8816fprop_relu_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_u4_i8816fprop_1x1_relu_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_u4_i8816fprop_hswish_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_u4_i8816fprop_1x1_hswish_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_u4_i8816fprop_roc_id_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_u4_i8816fprop_1x1_roc_id_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_u4_i8816fprop_roc_relu_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_u4_i8816fprop_1x1_roc_relu_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_u4_i8816fprop_roc_hswish_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_u4_i8816fprop_1x1_roc_hswish_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_u4_i8816fprop_id_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_u4_i8816fprop_1x1_id_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_u4_i8816fprop_relu_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_u4_i8816fprop_1x1_relu_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_u4_i8816fprop_hswish_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_u4_i8816fprop_1x1_hswish_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_u4_i8816fprop_roc_id_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_u4_i8816fprop_1x1_roc_id_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_u4_i8816fprop_roc_relu_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_u4_i8816fprop_1x1_roc_relu_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_u4_i8816fprop_roc_hswish_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_u4_i8816fprop_1x1_roc_hswish_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_f32_i8816fprop_id_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_f32_i8816fprop_1x1_id_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_f32_i8816fprop_relu_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_f32_i8816fprop_1x1_relu_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_f32_i8816fprop_hswish_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_f32_i8816fprop_1x1_hswish_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_f32_i8816fprop_id_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_f32_i8816fprop_1x1_id_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_f32_i8816fprop_relu_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_f32_i8816fprop_1x1_relu_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_f32_i8816fprop_hswish_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_f32_i8816fprop_1x1_hswish_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_f32_i8816fprop_roc_id_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_f32_i8816fprop_1x1_roc_id_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_f32_i8816fprop_roc_relu_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_f32_i8816fprop_1x1_roc_relu_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_f32_i8816fprop_roc_hswish_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_f32_i8816fprop_1x1_roc_hswish_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
- "all_conv2d_tensorop8816_operations.cu",
- "cutlass_tensorop_s4_i8832fprop_roc_id_s4_128x256x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
- "cutlass_tensorop_s4_i8832fprop_1x1_roc_id_s4_128x256x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
- "cutlass_tensorop_s4_i8832fprop_roc_relu_s4_128x256x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
- "cutlass_tensorop_s4_i8832fprop_1x1_roc_relu_s4_128x256x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
- "cutlass_tensorop_s4_i8832fprop_roc_hswish_s4_128x256x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
- "cutlass_tensorop_s4_i8832fprop_1x1_roc_hswish_s4_128x256x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
- "cutlass_tensorop_s4_i8832fprop_roc_id_s4_128x128x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
- "cutlass_tensorop_s4_i8832fprop_1x1_roc_id_s4_128x128x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
- "cutlass_tensorop_s4_i8832fprop_roc_relu_s4_128x128x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
- "cutlass_tensorop_s4_i8832fprop_1x1_roc_relu_s4_128x128x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
- "cutlass_tensorop_s4_i8832fprop_roc_hswish_s4_128x128x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
- "cutlass_tensorop_s4_i8832fprop_1x1_roc_hswish_s4_128x128x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
- "cutlass_tensorop_s4_i8832fprop_roc_id_s4_128x64x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
- "cutlass_tensorop_s4_i8832fprop_1x1_roc_id_s4_128x64x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
- "cutlass_tensorop_s4_i8832fprop_roc_relu_s4_128x64x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
- "cutlass_tensorop_s4_i8832fprop_1x1_roc_relu_s4_128x64x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
- "cutlass_tensorop_s4_i8832fprop_roc_hswish_s4_128x64x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
- "cutlass_tensorop_s4_i8832fprop_1x1_roc_hswish_s4_128x64x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
- "cutlass_tensorop_s4_i8832fprop_roc_id_s4_128x64x64_64x64x64_1_nc64hw64_c64rsk64_align32x32.cu",
- "cutlass_tensorop_s4_i8832fprop_1x1_roc_id_s4_128x64x64_64x64x64_1_nc64hw64_c64rsk64_align32x32.cu",
- "cutlass_tensorop_s4_i8832fprop_roc_relu_s4_128x64x64_64x64x64_1_nc64hw64_c64rsk64_align32x32.cu",
- "cutlass_tensorop_s4_i8832fprop_1x1_roc_relu_s4_128x64x64_64x64x64_1_nc64hw64_c64rsk64_align32x32.cu",
- "cutlass_tensorop_s4_i8832fprop_roc_hswish_s4_128x64x64_64x64x64_1_nc64hw64_c64rsk64_align32x32.cu",
- "cutlass_tensorop_s4_i8832fprop_1x1_roc_hswish_s4_128x64x64_64x64x64_1_nc64hw64_c64rsk64_align32x32.cu",
- "cutlass_tensorop_u4_i8832fprop_roc_id_u4_s4_128x256x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
- "cutlass_tensorop_u4_i8832fprop_1x1_roc_id_u4_s4_128x256x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
- "cutlass_tensorop_u4_i8832fprop_roc_relu_u4_s4_128x256x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
- "cutlass_tensorop_u4_i8832fprop_1x1_roc_relu_u4_s4_128x256x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
- "cutlass_tensorop_u4_i8832fprop_roc_id_u4_s4_128x128x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
- "cutlass_tensorop_u4_i8832fprop_1x1_roc_id_u4_s4_128x128x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
- "cutlass_tensorop_u4_i8832fprop_roc_relu_u4_s4_128x128x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
- "cutlass_tensorop_u4_i8832fprop_1x1_roc_relu_u4_s4_128x128x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
- "cutlass_tensorop_u4_i8832fprop_roc_id_u4_s4_128x64x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
- "cutlass_tensorop_u4_i8832fprop_1x1_roc_id_u4_s4_128x64x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
- "cutlass_tensorop_u4_i8832fprop_roc_relu_u4_s4_128x64x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
- "cutlass_tensorop_u4_i8832fprop_1x1_roc_relu_u4_s4_128x64x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
- "cutlass_tensorop_u4_i8832fprop_roc_id_u4_s4_128x64x64_64x64x64_1_nc64hw64_c64rsk64_align32x32.cu",
- "cutlass_tensorop_u4_i8832fprop_1x1_roc_id_u4_s4_128x64x64_64x64x64_1_nc64hw64_c64rsk64_align32x32.cu",
- "cutlass_tensorop_u4_i8832fprop_roc_relu_u4_s4_128x64x64_64x64x64_1_nc64hw64_c64rsk64_align32x32.cu",
- "cutlass_tensorop_u4_i8832fprop_1x1_roc_relu_u4_s4_128x64x64_64x64x64_1_nc64hw64_c64rsk64_align32x32.cu",
- "cutlass_tensorop_s4_i8832fprop_id_s4_128x16x64_128x16x64_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s4_i8832fprop_1x1_id_s4_128x16x64_128x16x64_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s4_i8832fprop_relu_s4_128x16x64_128x16x64_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x16x64_128x16x64_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s4_i8832fprop_hswish_s4_128x16x64_128x16x64_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x16x64_128x16x64_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s4_i8832fprop_id_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s4_i8832fprop_1x1_id_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s4_i8832fprop_relu_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s4_i8832fprop_hswish_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s4_i8832fprop_roc_id_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s4_i8832fprop_1x1_roc_id_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s4_i8832fprop_roc_relu_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s4_i8832fprop_1x1_roc_relu_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s4_i8832fprop_roc_hswish_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s4_i8832fprop_1x1_roc_hswish_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s4_i8832fprop_id_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s4_i8832fprop_1x1_id_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s4_i8832fprop_relu_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s4_i8832fprop_hswish_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s4_i8832fprop_roc_id_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s4_i8832fprop_1x1_roc_id_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s4_i8832fprop_roc_relu_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s4_i8832fprop_1x1_roc_relu_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s4_i8832fprop_roc_hswish_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s4_i8832fprop_1x1_roc_hswish_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s4_i8832fprop_id_s4_128x16x64_128x16x64_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s4_i8832fprop_1x1_id_s4_128x16x64_128x16x64_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s4_i8832fprop_relu_s4_128x16x64_128x16x64_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x16x64_128x16x64_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s4_i8832fprop_hswish_s4_128x16x64_128x16x64_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x16x64_128x16x64_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s4_i8832fprop_id_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s4_i8832fprop_1x1_id_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s4_i8832fprop_relu_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s4_i8832fprop_hswish_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s4_i8832fprop_roc_id_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s4_i8832fprop_1x1_roc_id_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s4_i8832fprop_roc_relu_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s4_i8832fprop_1x1_roc_relu_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s4_i8832fprop_roc_hswish_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s4_i8832fprop_1x1_roc_hswish_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s4_i8832fprop_id_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s4_i8832fprop_1x1_id_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s4_i8832fprop_relu_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s4_i8832fprop_hswish_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s4_i8832fprop_roc_id_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s4_i8832fprop_1x1_roc_id_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s4_i8832fprop_roc_relu_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s4_i8832fprop_1x1_roc_relu_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s4_i8832fprop_roc_hswish_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s4_i8832fprop_1x1_roc_hswish_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s4_i8832fprop_id_s4_128x16x64_128x16x64_2_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s4_i8832fprop_1x1_id_s4_128x16x64_128x16x64_2_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s4_i8832fprop_relu_s4_128x16x64_128x16x64_2_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x16x64_128x16x64_2_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s4_i8832fprop_hswish_s4_128x16x64_128x16x64_2_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x16x64_128x16x64_2_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s4_i8832fprop_id_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s4_i8832fprop_1x1_id_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s4_i8832fprop_relu_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s4_i8832fprop_hswish_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s4_i8832fprop_roc_id_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s4_i8832fprop_1x1_roc_id_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s4_i8832fprop_roc_relu_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s4_i8832fprop_1x1_roc_relu_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s4_i8832fprop_roc_hswish_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s4_i8832fprop_1x1_roc_hswish_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s4_i8832fprop_id_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s4_i8832fprop_1x1_id_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s4_i8832fprop_relu_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s4_i8832fprop_hswish_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s4_i8832fprop_roc_id_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s4_i8832fprop_1x1_roc_id_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s4_i8832fprop_roc_relu_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s4_i8832fprop_1x1_roc_relu_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s4_i8832fprop_roc_hswish_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s4_i8832fprop_1x1_roc_hswish_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_u4_i8832fprop_id_u4_s4_128x16x64_128x16x64_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_u4_i8832fprop_1x1_id_u4_s4_128x16x64_128x16x64_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x16x64_128x16x64_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x16x64_128x16x64_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_u4_i8832fprop_id_u4_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_u4_i8832fprop_1x1_id_u4_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_u4_i8832fprop_roc_id_u4_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_u4_i8832fprop_1x1_roc_id_u4_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_u4_i8832fprop_roc_relu_u4_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_u4_i8832fprop_1x1_roc_relu_u4_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_u4_i8832fprop_id_u4_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_u4_i8832fprop_1x1_id_u4_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_u4_i8832fprop_roc_id_u4_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_u4_i8832fprop_1x1_roc_id_u4_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_u4_i8832fprop_roc_relu_u4_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_u4_i8832fprop_1x1_roc_relu_u4_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_u4_i8832fprop_id_u4_s4_128x16x64_128x16x64_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_u4_i8832fprop_1x1_id_u4_s4_128x16x64_128x16x64_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x16x64_128x16x64_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x16x64_128x16x64_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_u4_i8832fprop_id_u4_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_u4_i8832fprop_1x1_id_u4_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_u4_i8832fprop_roc_id_u4_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_u4_i8832fprop_1x1_roc_id_u4_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_u4_i8832fprop_roc_relu_u4_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_u4_i8832fprop_1x1_roc_relu_u4_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_u4_i8832fprop_id_u4_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_u4_i8832fprop_1x1_id_u4_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_u4_i8832fprop_roc_id_u4_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_u4_i8832fprop_1x1_roc_id_u4_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_u4_i8832fprop_roc_relu_u4_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_u4_i8832fprop_1x1_roc_relu_u4_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_u4_i8832fprop_id_u4_s4_128x16x64_128x16x64_2_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_u4_i8832fprop_1x1_id_u4_s4_128x16x64_128x16x64_2_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x16x64_128x16x64_2_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x16x64_128x16x64_2_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_u4_i8832fprop_id_u4_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_u4_i8832fprop_1x1_id_u4_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_u4_i8832fprop_roc_id_u4_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_u4_i8832fprop_1x1_roc_id_u4_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_u4_i8832fprop_roc_relu_u4_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_u4_i8832fprop_1x1_roc_relu_u4_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_u4_i8832fprop_id_u4_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_u4_i8832fprop_1x1_id_u4_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_u4_i8832fprop_roc_id_u4_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_u4_i8832fprop_1x1_roc_id_u4_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_u4_i8832fprop_roc_relu_u4_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_u4_i8832fprop_1x1_roc_relu_u4_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s8_i8832fprop_id_s4_128x16x64_128x16x64_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_id_s4_128x16x64_128x16x64_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8832fprop_relu_s4_128x16x64_128x16x64_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_relu_s4_128x16x64_128x16x64_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8832fprop_hswish_s4_128x16x64_128x16x64_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_hswish_s4_128x16x64_128x16x64_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8832fprop_id_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_id_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8832fprop_relu_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_relu_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8832fprop_hswish_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_hswish_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8832fprop_roc_id_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_roc_id_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8832fprop_roc_relu_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_roc_relu_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8832fprop_roc_hswish_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_roc_hswish_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8832fprop_id_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_id_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8832fprop_relu_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_relu_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8832fprop_hswish_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_hswish_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8832fprop_roc_id_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_roc_id_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8832fprop_roc_relu_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_roc_relu_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8832fprop_roc_hswish_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_roc_hswish_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8832fprop_id_s4_128x16x64_128x16x64_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_id_s4_128x16x64_128x16x64_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8832fprop_relu_s4_128x16x64_128x16x64_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_relu_s4_128x16x64_128x16x64_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8832fprop_hswish_s4_128x16x64_128x16x64_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_hswish_s4_128x16x64_128x16x64_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8832fprop_id_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_id_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8832fprop_relu_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_relu_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8832fprop_hswish_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_hswish_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8832fprop_roc_id_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_roc_id_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8832fprop_roc_relu_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_roc_relu_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8832fprop_roc_hswish_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_roc_hswish_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8832fprop_id_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_id_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8832fprop_relu_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_relu_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8832fprop_hswish_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_hswish_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8832fprop_roc_id_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_roc_id_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8832fprop_roc_relu_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_roc_relu_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8832fprop_roc_hswish_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_roc_hswish_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8832fprop_id_s4_128x16x64_128x16x64_2_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_id_s4_128x16x64_128x16x64_2_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s8_i8832fprop_relu_s4_128x16x64_128x16x64_2_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_relu_s4_128x16x64_128x16x64_2_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s8_i8832fprop_hswish_s4_128x16x64_128x16x64_2_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_hswish_s4_128x16x64_128x16x64_2_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s8_i8832fprop_id_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_id_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s8_i8832fprop_relu_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_relu_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s8_i8832fprop_hswish_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_hswish_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s8_i8832fprop_roc_id_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_roc_id_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s8_i8832fprop_roc_relu_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_roc_relu_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s8_i8832fprop_roc_hswish_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_roc_hswish_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s8_i8832fprop_id_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_id_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s8_i8832fprop_relu_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_relu_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s8_i8832fprop_hswish_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_hswish_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s8_i8832fprop_roc_id_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_roc_id_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s8_i8832fprop_roc_relu_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_roc_relu_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s8_i8832fprop_roc_hswish_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_roc_hswish_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s8_i8832fprop_id_u4_s4_128x16x64_128x16x64_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_id_u4_s4_128x16x64_128x16x64_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8832fprop_relu_u4_s4_128x16x64_128x16x64_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_relu_u4_s4_128x16x64_128x16x64_2_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8832fprop_id_u4_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_id_u4_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8832fprop_relu_u4_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_relu_u4_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8832fprop_roc_id_u4_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_roc_id_u4_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8832fprop_roc_relu_u4_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_roc_relu_u4_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8832fprop_id_u4_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_id_u4_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8832fprop_relu_u4_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_relu_u4_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8832fprop_roc_id_u4_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_roc_id_u4_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8832fprop_roc_relu_u4_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_roc_relu_u4_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
- "cutlass_tensorop_s8_i8832fprop_id_u4_s4_128x16x64_128x16x64_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_id_u4_s4_128x16x64_128x16x64_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8832fprop_relu_u4_s4_128x16x64_128x16x64_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_relu_u4_s4_128x16x64_128x16x64_2_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8832fprop_id_u4_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_id_u4_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8832fprop_relu_u4_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_relu_u4_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8832fprop_roc_id_u4_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_roc_id_u4_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8832fprop_roc_relu_u4_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_roc_relu_u4_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8832fprop_id_u4_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_id_u4_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8832fprop_relu_u4_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_relu_u4_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8832fprop_roc_id_u4_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_roc_id_u4_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8832fprop_roc_relu_u4_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_roc_relu_u4_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
- "cutlass_tensorop_s8_i8832fprop_id_u4_s4_128x16x64_128x16x64_2_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_id_u4_s4_128x16x64_128x16x64_2_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s8_i8832fprop_relu_u4_s4_128x16x64_128x16x64_2_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_relu_u4_s4_128x16x64_128x16x64_2_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s8_i8832fprop_id_u4_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_id_u4_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s8_i8832fprop_relu_u4_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_relu_u4_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s8_i8832fprop_roc_id_u4_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_roc_id_u4_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s8_i8832fprop_roc_relu_u4_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_roc_relu_u4_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s8_i8832fprop_id_u4_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_id_u4_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s8_i8832fprop_relu_u4_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_relu_u4_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s8_i8832fprop_roc_id_u4_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_roc_id_u4_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s8_i8832fprop_roc_relu_u4_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
- "cutlass_tensorop_s8_i8832fprop_1x1_roc_relu_u4_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
- "all_conv2d_tensorop8832_operations.cu",
- "cutlass_simt_sdwfprop_id_f32_32x32x8_32x32x8_2_nchw_nchw_align4x1.cu",
- "cutlass_simt_sdwfprop_relu_f32_32x32x8_32x32x8_2_nchw_nchw_align4x1.cu",
- "cutlass_simt_sdwfprop_id_f32_32x32x8_32x32x8_2_nchw_nchw_align1x1.cu",
- "cutlass_simt_sdwfprop_relu_f32_32x32x8_32x32x8_2_nchw_nchw_align1x1.cu",
- "cutlass_simt_sdwfprop_id_f32_32x64x8_32x64x8_2_nchw_nchw_align4x1.cu",
- "cutlass_simt_sdwfprop_relu_f32_32x64x8_32x64x8_2_nchw_nchw_align4x1.cu",
- "cutlass_simt_sdwfprop_id_f32_32x64x8_32x64x8_2_nchw_nchw_align1x1.cu",
- "cutlass_simt_sdwfprop_relu_f32_32x64x8_32x64x8_2_nchw_nchw_align1x1.cu",
- "cutlass_simt_sdwfprop_id_f32_64x32x8_64x32x8_2_nchw_nchw_align4x1.cu",
- "cutlass_simt_sdwfprop_relu_f32_64x32x8_64x32x8_2_nchw_nchw_align4x1.cu",
- "cutlass_simt_sdwfprop_id_f32_64x32x8_64x32x8_2_nchw_nchw_align1x1.cu",
- "cutlass_simt_sdwfprop_relu_f32_64x32x8_64x32x8_2_nchw_nchw_align1x1.cu",
- "cutlass_simt_sdwfprop_id_f32_32x128x8_32x64x8_2_nchw_nchw_align4x1.cu",
- "cutlass_simt_sdwfprop_relu_f32_32x128x8_32x64x8_2_nchw_nchw_align4x1.cu",
- "cutlass_simt_sdwfprop_id_f32_32x128x8_32x64x8_2_nchw_nchw_align1x1.cu",
- "cutlass_simt_sdwfprop_relu_f32_32x128x8_32x64x8_2_nchw_nchw_align1x1.cu",
- "cutlass_simt_sdwfprop_id_f32_64x64x8_32x64x8_2_nchw_nchw_align4x1.cu",
- "cutlass_simt_sdwfprop_relu_f32_64x64x8_32x64x8_2_nchw_nchw_align4x1.cu",
- "cutlass_simt_sdwfprop_id_f32_64x64x8_32x64x8_2_nchw_nchw_align1x1.cu",
- "cutlass_simt_sdwfprop_relu_f32_64x64x8_32x64x8_2_nchw_nchw_align1x1.cu",
- "cutlass_simt_sdwfprop_id_f32_128x32x8_64x32x8_2_nchw_nchw_align4x1.cu",
- "cutlass_simt_sdwfprop_relu_f32_128x32x8_64x32x8_2_nchw_nchw_align4x1.cu",
- "cutlass_simt_sdwfprop_id_f32_128x32x8_64x32x8_2_nchw_nchw_align1x1.cu",
- "cutlass_simt_sdwfprop_relu_f32_128x32x8_64x32x8_2_nchw_nchw_align1x1.cu",
- "cutlass_simt_sdwfprop_id_f32_64x128x8_32x64x8_2_nchw_nchw_align4x1.cu",
- "cutlass_simt_sdwfprop_relu_f32_64x128x8_32x64x8_2_nchw_nchw_align4x1.cu",
- "cutlass_simt_sdwfprop_id_f32_64x128x8_32x64x8_2_nchw_nchw_align1x1.cu",
- "cutlass_simt_sdwfprop_relu_f32_64x128x8_32x64x8_2_nchw_nchw_align1x1.cu",
- "cutlass_simt_sdwfprop_id_f32_128x64x8_64x32x8_2_nchw_nchw_align4x1.cu",
- "cutlass_simt_sdwfprop_relu_f32_128x64x8_64x32x8_2_nchw_nchw_align4x1.cu",
- "cutlass_simt_sdwfprop_id_f32_128x64x8_64x32x8_2_nchw_nchw_align1x1.cu",
- "cutlass_simt_sdwfprop_relu_f32_128x64x8_64x32x8_2_nchw_nchw_align1x1.cu",
- "cutlass_simt_sdwfprop_id_f32_128x128x8_32x64x8_2_nchw_nchw_align4x1.cu",
- "cutlass_simt_sdwfprop_relu_f32_128x128x8_32x64x8_2_nchw_nchw_align4x1.cu",
- "cutlass_simt_sdwfprop_id_f32_128x128x8_32x64x8_2_nchw_nchw_align1x1.cu",
- "cutlass_simt_sdwfprop_relu_f32_128x128x8_32x64x8_2_nchw_nchw_align1x1.cu",
- "all_dwconv2d_fprop_simt_operations.cu",
- "cutlass_tensorop_f16_s884dwfprop_id_f16_128x256x32_64x64x32_2_nchw_nchw_align8x1.cu",
- "cutlass_tensorop_f16_s884dwfprop_relu_f16_128x256x32_64x64x32_2_nchw_nchw_align8x1.cu",
- "cutlass_tensorop_f16_s884dwfprop_id_f16_128x128x32_32x32x32_2_nchw_nchw_align8x1.cu",
- "cutlass_tensorop_f16_s884dwfprop_relu_f16_128x128x32_32x32x32_2_nchw_nchw_align8x1.cu",
- "cutlass_tensorop_f16_s884dwfprop_id_f16_64x128x32_32x32x32_2_nchw_nchw_align8x1.cu",
- "cutlass_tensorop_f16_s884dwfprop_relu_f16_64x128x32_32x32x32_2_nchw_nchw_align8x1.cu",
- "cutlass_tensorop_f16_s884dwfprop_id_f16_128x64x32_32x32x32_2_nchw_nchw_align8x1.cu",
- "cutlass_tensorop_f16_s884dwfprop_relu_f16_128x64x32_32x32x32_2_nchw_nchw_align8x1.cu",
- "cutlass_tensorop_f16_s884dwfprop_id_f16_64x64x32_32x32x32_2_nchw_nchw_align8x1.cu",
- "cutlass_tensorop_f16_s884dwfprop_relu_f16_64x64x32_32x32x32_2_nchw_nchw_align8x1.cu",
- "cutlass_tensorop_f16_s884dwfprop_id_f16_128x256x32_64x64x32_2_nchw_nchw_align2x1.cu",
- "cutlass_tensorop_f16_s884dwfprop_relu_f16_128x256x32_64x64x32_2_nchw_nchw_align2x1.cu",
- "cutlass_tensorop_f16_s884dwfprop_id_f16_128x128x32_32x32x32_2_nchw_nchw_align2x1.cu",
- "cutlass_tensorop_f16_s884dwfprop_relu_f16_128x128x32_32x32x32_2_nchw_nchw_align2x1.cu",
- "cutlass_tensorop_f16_s884dwfprop_id_f16_64x128x32_32x32x32_2_nchw_nchw_align2x1.cu",
- "cutlass_tensorop_f16_s884dwfprop_relu_f16_64x128x32_32x32x32_2_nchw_nchw_align2x1.cu",
- "cutlass_tensorop_f16_s884dwfprop_id_f16_128x64x32_32x32x32_2_nchw_nchw_align2x1.cu",
- "cutlass_tensorop_f16_s884dwfprop_relu_f16_128x64x32_32x32x32_2_nchw_nchw_align2x1.cu",
- "cutlass_tensorop_f16_s884dwfprop_id_f16_64x64x32_32x32x32_2_nchw_nchw_align2x1.cu",
- "cutlass_tensorop_f16_s884dwfprop_relu_f16_64x64x32_32x32x32_2_nchw_nchw_align2x1.cu",
- "cutlass_tensorop_f16_s884dwfprop_id_f16_128x256x32_64x64x32_2_nchw_nchw_align1x1.cu",
- "cutlass_tensorop_f16_s884dwfprop_relu_f16_128x256x32_64x64x32_2_nchw_nchw_align1x1.cu",
- "cutlass_tensorop_f16_s884dwfprop_id_f16_128x128x32_32x32x32_2_nchw_nchw_align1x1.cu",
- "cutlass_tensorop_f16_s884dwfprop_relu_f16_128x128x32_32x32x32_2_nchw_nchw_align1x1.cu",
- "cutlass_tensorop_f16_s884dwfprop_id_f16_64x128x32_32x32x32_2_nchw_nchw_align1x1.cu",
- "cutlass_tensorop_f16_s884dwfprop_relu_f16_64x128x32_32x32x32_2_nchw_nchw_align1x1.cu",
- "cutlass_tensorop_f16_s884dwfprop_id_f16_128x64x32_32x32x32_2_nchw_nchw_align1x1.cu",
- "cutlass_tensorop_f16_s884dwfprop_relu_f16_128x64x32_32x32x32_2_nchw_nchw_align1x1.cu",
- "cutlass_tensorop_f16_s884dwfprop_id_f16_64x64x32_32x32x32_2_nchw_nchw_align1x1.cu",
- "cutlass_tensorop_f16_s884dwfprop_relu_f16_64x64x32_32x32x32_2_nchw_nchw_align1x1.cu",
- "cutlass_tensorop_h884dwfprop_id_f16_128x256x32_64x64x32_2_nchw_nchw_align8x1.cu",
- "cutlass_tensorop_h884dwfprop_relu_f16_128x256x32_64x64x32_2_nchw_nchw_align8x1.cu",
- "cutlass_tensorop_h884dwfprop_id_f16_128x128x32_32x32x32_2_nchw_nchw_align8x1.cu",
- "cutlass_tensorop_h884dwfprop_relu_f16_128x128x32_32x32x32_2_nchw_nchw_align8x1.cu",
- "cutlass_tensorop_h884dwfprop_id_f16_64x128x32_32x32x32_2_nchw_nchw_align8x1.cu",
- "cutlass_tensorop_h884dwfprop_relu_f16_64x128x32_32x32x32_2_nchw_nchw_align8x1.cu",
- "cutlass_tensorop_h884dwfprop_id_f16_128x64x32_32x32x32_2_nchw_nchw_align8x1.cu",
- "cutlass_tensorop_h884dwfprop_relu_f16_128x64x32_32x32x32_2_nchw_nchw_align8x1.cu",
- "cutlass_tensorop_h884dwfprop_id_f16_64x64x32_32x32x32_2_nchw_nchw_align8x1.cu",
- "cutlass_tensorop_h884dwfprop_relu_f16_64x64x32_32x32x32_2_nchw_nchw_align8x1.cu",
- "cutlass_tensorop_h884dwfprop_id_f16_128x256x32_64x64x32_2_nchw_nchw_align2x1.cu",
- "cutlass_tensorop_h884dwfprop_relu_f16_128x256x32_64x64x32_2_nchw_nchw_align2x1.cu",
- "cutlass_tensorop_h884dwfprop_id_f16_128x128x32_32x32x32_2_nchw_nchw_align2x1.cu",
- "cutlass_tensorop_h884dwfprop_relu_f16_128x128x32_32x32x32_2_nchw_nchw_align2x1.cu",
- "cutlass_tensorop_h884dwfprop_id_f16_64x128x32_32x32x32_2_nchw_nchw_align2x1.cu",
- "cutlass_tensorop_h884dwfprop_relu_f16_64x128x32_32x32x32_2_nchw_nchw_align2x1.cu",
- "cutlass_tensorop_h884dwfprop_id_f16_128x64x32_32x32x32_2_nchw_nchw_align2x1.cu",
- "cutlass_tensorop_h884dwfprop_relu_f16_128x64x32_32x32x32_2_nchw_nchw_align2x1.cu",
- "cutlass_tensorop_h884dwfprop_id_f16_64x64x32_32x32x32_2_nchw_nchw_align2x1.cu",
- "cutlass_tensorop_h884dwfprop_relu_f16_64x64x32_32x32x32_2_nchw_nchw_align2x1.cu",
- "cutlass_tensorop_h884dwfprop_id_f16_128x256x32_64x64x32_2_nchw_nchw_align1x1.cu",
- "cutlass_tensorop_h884dwfprop_relu_f16_128x256x32_64x64x32_2_nchw_nchw_align1x1.cu",
- "cutlass_tensorop_h884dwfprop_id_f16_128x128x32_32x32x32_2_nchw_nchw_align1x1.cu",
- "cutlass_tensorop_h884dwfprop_relu_f16_128x128x32_32x32x32_2_nchw_nchw_align1x1.cu",
- "cutlass_tensorop_h884dwfprop_id_f16_64x128x32_32x32x32_2_nchw_nchw_align1x1.cu",
- "cutlass_tensorop_h884dwfprop_relu_f16_64x128x32_32x32x32_2_nchw_nchw_align1x1.cu",
- "cutlass_tensorop_h884dwfprop_id_f16_128x64x32_32x32x32_2_nchw_nchw_align1x1.cu",
- "cutlass_tensorop_h884dwfprop_relu_f16_128x64x32_32x32x32_2_nchw_nchw_align1x1.cu",
- "cutlass_tensorop_h884dwfprop_id_f16_64x64x32_32x32x32_2_nchw_nchw_align1x1.cu",
- "cutlass_tensorop_h884dwfprop_relu_f16_64x64x32_32x32x32_2_nchw_nchw_align1x1.cu",
- "all_dwconv2d_fprop_tensorop884_operations.cu",
- "cutlass_simt_sdwdgrad_id_f32_32x32x8_32x32x8_2_nchw_nchw_align4x1.cu",
- "cutlass_simt_sdwdgrad_id_f32_32x32x8_32x32x8_2_nchw_nchw_align1x1.cu",
- "cutlass_simt_sdwdgrad_id_f32_32x64x8_32x64x8_2_nchw_nchw_align4x1.cu",
- "cutlass_simt_sdwdgrad_id_f32_32x64x8_32x64x8_2_nchw_nchw_align1x1.cu",
- "cutlass_simt_sdwdgrad_id_f32_64x32x8_64x32x8_2_nchw_nchw_align4x1.cu",
- "cutlass_simt_sdwdgrad_id_f32_64x32x8_64x32x8_2_nchw_nchw_align1x1.cu",
- "cutlass_simt_sdwdgrad_id_f32_32x128x8_32x64x8_2_nchw_nchw_align4x1.cu",
- "cutlass_simt_sdwdgrad_id_f32_32x128x8_32x64x8_2_nchw_nchw_align1x1.cu",
- "cutlass_simt_sdwdgrad_id_f32_64x64x8_32x64x8_2_nchw_nchw_align4x1.cu",
- "cutlass_simt_sdwdgrad_id_f32_64x64x8_32x64x8_2_nchw_nchw_align1x1.cu",
- "cutlass_simt_sdwdgrad_id_f32_128x32x8_64x32x8_2_nchw_nchw_align4x1.cu",
- "cutlass_simt_sdwdgrad_id_f32_128x32x8_64x32x8_2_nchw_nchw_align1x1.cu",
- "cutlass_simt_sdwdgrad_id_f32_64x128x8_32x64x8_2_nchw_nchw_align4x1.cu",
- "cutlass_simt_sdwdgrad_id_f32_64x128x8_32x64x8_2_nchw_nchw_align1x1.cu",
- "cutlass_simt_sdwdgrad_id_f32_128x64x8_64x32x8_2_nchw_nchw_align4x1.cu",
- "cutlass_simt_sdwdgrad_id_f32_128x64x8_64x32x8_2_nchw_nchw_align1x1.cu",
- "cutlass_simt_sdwdgrad_id_f32_128x128x8_32x64x8_2_nchw_nchw_align4x1.cu",
- "cutlass_simt_sdwdgrad_id_f32_128x128x8_32x64x8_2_nchw_nchw_align1x1.cu",
- "all_dwconv2d_dgrad_simt_operations.cu",
- "cutlass_tensorop_f16_s884dwdgrad_id_f16_128x256x32_64x64x32_2_nchw_nchw_align8x1.cu",
- "cutlass_tensorop_f16_s884dwdgrad_id_f16_128x128x32_32x32x32_2_nchw_nchw_align8x1.cu",
- "cutlass_tensorop_f16_s884dwdgrad_id_f16_64x128x32_32x32x32_2_nchw_nchw_align8x1.cu",
- "cutlass_tensorop_f16_s884dwdgrad_id_f16_128x64x32_32x32x32_2_nchw_nchw_align8x1.cu",
- "cutlass_tensorop_f16_s884dwdgrad_id_f16_64x64x32_32x32x32_2_nchw_nchw_align8x1.cu",
- "cutlass_tensorop_f16_s884dwdgrad_id_f16_128x256x32_64x64x32_2_nchw_nchw_align2x1.cu",
- "cutlass_tensorop_f16_s884dwdgrad_id_f16_128x128x32_32x32x32_2_nchw_nchw_align2x1.cu",
- "cutlass_tensorop_f16_s884dwdgrad_id_f16_64x128x32_32x32x32_2_nchw_nchw_align2x1.cu",
- "cutlass_tensorop_f16_s884dwdgrad_id_f16_128x64x32_32x32x32_2_nchw_nchw_align2x1.cu",
- "cutlass_tensorop_f16_s884dwdgrad_id_f16_64x64x32_32x32x32_2_nchw_nchw_align2x1.cu",
- "cutlass_tensorop_f16_s884dwdgrad_id_f16_128x256x32_64x64x32_2_nchw_nchw_align1x1.cu",
- "cutlass_tensorop_f16_s884dwdgrad_id_f16_128x128x32_32x32x32_2_nchw_nchw_align1x1.cu",
- "cutlass_tensorop_f16_s884dwdgrad_id_f16_64x128x32_32x32x32_2_nchw_nchw_align1x1.cu",
- "cutlass_tensorop_f16_s884dwdgrad_id_f16_128x64x32_32x32x32_2_nchw_nchw_align1x1.cu",
- "cutlass_tensorop_f16_s884dwdgrad_id_f16_64x64x32_32x32x32_2_nchw_nchw_align1x1.cu",
- "cutlass_tensorop_h884dwdgrad_id_f16_128x256x32_64x64x32_2_nchw_nchw_align8x1.cu",
- "cutlass_tensorop_h884dwdgrad_id_f16_128x128x32_32x32x32_2_nchw_nchw_align8x1.cu",
- "cutlass_tensorop_h884dwdgrad_id_f16_64x128x32_32x32x32_2_nchw_nchw_align8x1.cu",
- "cutlass_tensorop_h884dwdgrad_id_f16_128x64x32_32x32x32_2_nchw_nchw_align8x1.cu",
- "cutlass_tensorop_h884dwdgrad_id_f16_64x64x32_32x32x32_2_nchw_nchw_align8x1.cu",
- "cutlass_tensorop_h884dwdgrad_id_f16_128x256x32_64x64x32_2_nchw_nchw_align2x1.cu",
- "cutlass_tensorop_h884dwdgrad_id_f16_128x128x32_32x32x32_2_nchw_nchw_align2x1.cu",
- "cutlass_tensorop_h884dwdgrad_id_f16_64x128x32_32x32x32_2_nchw_nchw_align2x1.cu",
- "cutlass_tensorop_h884dwdgrad_id_f16_128x64x32_32x32x32_2_nchw_nchw_align2x1.cu",
- "cutlass_tensorop_h884dwdgrad_id_f16_64x64x32_32x32x32_2_nchw_nchw_align2x1.cu",
- "cutlass_tensorop_h884dwdgrad_id_f16_128x256x32_64x64x32_2_nchw_nchw_align1x1.cu",
- "cutlass_tensorop_h884dwdgrad_id_f16_128x128x32_32x32x32_2_nchw_nchw_align1x1.cu",
- "cutlass_tensorop_h884dwdgrad_id_f16_64x128x32_32x32x32_2_nchw_nchw_align1x1.cu",
- "cutlass_tensorop_h884dwdgrad_id_f16_128x64x32_32x32x32_2_nchw_nchw_align1x1.cu",
- "cutlass_tensorop_h884dwdgrad_id_f16_64x64x32_32x32x32_2_nchw_nchw_align1x1.cu",
- "all_dwconv2d_dgrad_tensorop884_operations.cu",
- ]
|