From b33217d8f04ddf8d7b22de0af239a84dbf3a5e24 Mon Sep 17 00:00:00 2001
From: Megvii Engine Team <megengine@megvii.com>
Date: Thu, 27 May 2021 19:32:24 +0800
Subject: [PATCH] refactor(dnn/cuda): refactor cutlass kernel generator for
 deconv operation

GitOrigin-RevId: 88e962a912dc0f31b8b31787a9ab7aa1b38d1264
---
 ...> implicit_gemm_deconv_cutlass_wrapper.cuinl} |   2 +-
 ...d_identity_s8_16x128x16_16x128x16_1_nc4hw4.cu | Bin 0 -> 1716 bytes
 ...ad_identity_s8_16x128x16_16x64x16_2_nc4hw4.cu | Bin 0 -> 1714 bytes
 ...dgrad_identity_s8_16x64x8_16x64x8_2_nc4hw4.cu | Bin 0 -> 1708 bytes
 ...ad_identity_s8_32x128x32_32x64x32_2_nc4hw4.cu | Bin 0 -> 1715 bytes
 ...ad_identity_s8_64x128x32_64x32x32_2_nc4hw4.cu | Bin 0 -> 1715 bytes
 ...gemm_dp4a_ncdiv4hw4_16x128x16_16x128x16_id.cu | Bin 1829 -> 0 bytes
 ..._gemm_dp4a_ncdiv4hw4_16x128x16_16x64x16_id.cu | Bin 1828 -> 0 bytes
 ...cit_gemm_dp4a_ncdiv4hw4_16x64x8_16x64x8_id.cu | Bin 1825 -> 0 bytes
 ..._gemm_dp4a_ncdiv4hw4_32x128x32_32x64x32_id.cu | Bin 1829 -> 0 bytes
 ..._gemm_dp4a_ncdiv4hw4_64x128x32_64x32x32_id.cu | Bin 1829 -> 0 bytes
 11 files changed, 1 insertion(+), 1 deletion(-)
 rename dnn/src/cuda/convolution/backward_data/{int8/deconv_int8_implicit_gemm_cutlass_wrapper.cuinl => implicit_gemm_deconv_cutlass_wrapper.cuinl} (97%)
 create mode 100644 dnn/src/cuda/convolution/backward_data/int8/kimpl/cutlass_simt_s8_idgrad_identity_s8_16x128x16_16x128x16_1_nc4hw4.cu
 create mode 100644 dnn/src/cuda/convolution/backward_data/int8/kimpl/cutlass_simt_s8_idgrad_identity_s8_16x128x16_16x64x16_2_nc4hw4.cu
 create mode 100644 dnn/src/cuda/convolution/backward_data/int8/kimpl/cutlass_simt_s8_idgrad_identity_s8_16x64x8_16x64x8_2_nc4hw4.cu
 create mode 100644 dnn/src/cuda/convolution/backward_data/int8/kimpl/cutlass_simt_s8_idgrad_identity_s8_32x128x32_32x64x32_2_nc4hw4.cu
 create mode 100644 dnn/src/cuda/convolution/backward_data/int8/kimpl/cutlass_simt_s8_idgrad_identity_s8_64x128x32_64x32x32_2_nc4hw4.cu
 delete mode 100644 dnn/src/cuda/convolution/backward_data/int8/kimpl/deconv_int8_implicit_gemm_dp4a_ncdiv4hw4_16x128x16_16x128x16_id.cu
 delete mode 100644 dnn/src/cuda/convolution/backward_data/int8/kimpl/deconv_int8_implicit_gemm_dp4a_ncdiv4hw4_16x128x16_16x64x16_id.cu
 delete mode 100644 dnn/src/cuda/convolution/backward_data/int8/kimpl/deconv_int8_implicit_gemm_dp4a_ncdiv4hw4_16x64x8_16x64x8_id.cu
 delete mode 100644 dnn/src/cuda/convolution/backward_data/int8/kimpl/deconv_int8_implicit_gemm_dp4a_ncdiv4hw4_32x128x32_32x64x32_id.cu
 delete mode 100644 dnn/src/cuda/convolution/backward_data/int8/kimpl/deconv_int8_implicit_gemm_dp4a_ncdiv4hw4_64x128x32_64x32x32_id.cu

diff --git a/dnn/src/cuda/convolution/backward_data/int8/deconv_int8_implicit_gemm_cutlass_wrapper.cuinl b/dnn/src/cuda/convolution/backward_data/implicit_gemm_deconv_cutlass_wrapper.cuinl
similarity index 97%
rename from dnn/src/cuda/convolution/backward_data/int8/deconv_int8_implicit_gemm_cutlass_wrapper.cuinl
rename to dnn/src/cuda/convolution/backward_data/implicit_gemm_deconv_cutlass_wrapper.cuinl
index f5382f919..349b392ae 100644
--- a/dnn/src/cuda/convolution/backward_data/int8/deconv_int8_implicit_gemm_cutlass_wrapper.cuinl
+++ b/dnn/src/cuda/convolution/backward_data/implicit_gemm_deconv_cutlass_wrapper.cuinl
@@ -1,6 +1,6 @@
 /**
  * \file
- * dnn/src/cuda/conv_bias/int8/conv_bias_int8_implicit_gemm_cutlass_wrapper.cuinl
+ * dnn/src/cuda/convolution/backward_data/implicit_gemm_deconv_cutlass_wrapper.cuinl
  * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
  *
  * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
diff --git a/dnn/src/cuda/convolution/backward_data/int8/kimpl/cutlass_simt_s8_idgrad_identity_s8_16x128x16_16x128x16_1_nc4hw4.cu b/dnn/src/cuda/convolution/backward_data/int8/kimpl/cutlass_simt_s8_idgrad_identity_s8_16x128x16_16x128x16_1_nc4hw4.cu
new file mode 100644
index 0000000000000000000000000000000000000000..63d8b005d12b85d227605f38e1ce943ebd54a88e
GIT binary patch
literal 1716
zcmbtV+invv5PjdTFsX!~N{IqdRGCnzBqd1D7IiC?x0Pc%8!KMh@+FskJ>&Im(klw$
z#dhpDvvX$VOtK>}M{h1Krjx6y;`U;Cb6))JXT2U0rP7w8HL*&l66raj!As&?c4RCm
zD}tuuF=B#%o%4bs6Wrppx8#hW&dFU+!EttCh$R*G+;$GYIBx~@CqxS3AQ*rMMP*<d
zb)2O=8W`zOtyWqFFSP2-2wg(ptYE~Go~VozRQRIgRaGz^EsKUsv9`n*Za>f<l<Z_#
zq`u@<afyU-o+t_*X^owzeBn+D!AeUQY+QNaxACe!-1I-5Zu-NAr$W)eVm;`fk}GbB
z=M2rZZ64c8ZL<(rp-CNC!vTFk8A5(*xPm;j7>r}^7&E>Slw<q|G6DdCsQqc-yAh4{
zFiEyLc#Lnka@t;vH{b6zqrqACV91ZnO+NlN$n&p!`gnL`i2$&92}CSiV0>-Hu_1G)
zN+QzK3FGvhSHu33hu8ubPhrg$#PCsn*hNq`ScXk9c0T^c6#R8jo8FyAZiLijfV%mG
z<%D7U6&m`Sv7bV5V#m6g2}NQD1Fy=YdhRM(mkOalVpLy0NX_%#bE%21K~I^b9!fFH
zq^3)Z$9wfmVsEF~-PU2&Ciz<2-%H+&U6o4Hq;}A>8sxD4n709}>f?@hTdvC>y)beM
zl+Qkc)-umQo1hD{(t@FimrN<>anN&I>v7X6bbpZ)+~}lz9zBx6**gdTXga@_&?5tk
zFC?!(zPY8xh!rqK>Yxw8d=nA|SA@;$)8WWi?}(ehemH94d<5Osh}}qdnNMTOTp9tO
zpJTMvcIgbEsV=WMc+djB+TomX_%?2<q8{0QkD+Ps>)=iB*9PO-lWX0y0!Y-M%DD#}
ZRE0+{;o<N;HpOdyFWO{BTrn}v{s5NlI+6eY

literal 0
HcmV?d00001

diff --git a/dnn/src/cuda/convolution/backward_data/int8/kimpl/cutlass_simt_s8_idgrad_identity_s8_16x128x16_16x64x16_2_nc4hw4.cu b/dnn/src/cuda/convolution/backward_data/int8/kimpl/cutlass_simt_s8_idgrad_identity_s8_16x128x16_16x64x16_2_nc4hw4.cu
new file mode 100644
index 0000000000000000000000000000000000000000..e704b90291a1d3a4cd68b038346c6854f499cb4a
GIT binary patch
literal 1714
zcmb7F+in^$5PjdTaEVm45h;Qq6uC4~3MrM^Bvm$2`nEE*SuDJ^<qNm3?|98Zt`&W;
zjXh^}&di)4I}tPV?)qvnzPTyxuO@ew#h-50=^#-mZ8=&ItAr|%o*^2%B+g|g#*(rk
zXfhfhCJ5L$FDNp>&EI-U&KPQ)Jp>gTXJ>|3QgP31>j;eVR#1ONq#zE00f<mk2F6j#
zS=yn2kq*^rsb%m&tIm|r1q99tMm*_=%1A+lFG^lj1>@1O*pVq#mKekBCmMv3tt^Yw
z7u+f?kx<SPMd2fju@jXq+<75bX$gakD=+*eUUdiSZtr~E9mLb19}m4k(f)kZZ=sSa
zZi(j%O*c&%n@er85J{nN9oUWn`ie4y{KjwvX=*SS$JjAud?_f$_z7eP00cq%^TM|y
z8XaJgY;^D#-*e@(y&0{4JgkTPi}ulwpX<AP^nZ|-U-{fSKC(mr*t`ZJmd-K0HKW*%
zITR%k*_8?7^qyCP?z4y30T@qU&F94MVRz6*P&QZwyJYNryk<J+A7u$fZPa9U?~ofI
zbs3;+er`Eo7=MFyzGUo|P@LG2uBJkf*uOxlGO3*Vg4UHnsGkVc*N-yu^7l+?;%m@z
zVyT5v3{$D;0^`v^JLA~dsdV>sn6*j165Flh?buYQGQAWCB&7aGvwIuBs=nG2Z>n`2
zq!&hRfbzu`P+I0WC=+CXmRc}W@scS8Ee>jqYdvfleI727es?lyo`+AAaPa{G0GiGp
z6!gRZ<15K45N~ejDPjc-kviz3FyDoQu@zyn`gA-p);r>+upf^aJ0C&Y8nK@Uuk+d2
zGM6R*$mbNTv|TtuXsXLw4j#3@?`AZo9KMg+vZx34zhl@n_-*he_*;W<?a4K68UZBg
eQ03f%4ywW<nD21-7@Oj)KNoGX6Rwz;Wq$!$ZaOXi

literal 0
HcmV?d00001

diff --git a/dnn/src/cuda/convolution/backward_data/int8/kimpl/cutlass_simt_s8_idgrad_identity_s8_16x64x8_16x64x8_2_nc4hw4.cu b/dnn/src/cuda/convolution/backward_data/int8/kimpl/cutlass_simt_s8_idgrad_identity_s8_16x64x8_16x64x8_2_nc4hw4.cu
new file mode 100644
index 0000000000000000000000000000000000000000..a55ed79e7c9a07690745992978225ef42e145a7b
GIT binary patch
literal 1708
zcmbVN+in^$5PjdTm=dXKBT__)P~_4`Eu>UxlT;y6`nEE*SuDI>%NH*B`i|Eu<XX|b
ztQUKZcg~zSLv|!*@aFPjJi5BdZ!g9-U-IAGtkVHm=*kLMQLCjcpqc>-UQy?=BV%b%
zQWy`1z$L{v=OqIZ-2An(bQy=%$z9OF34UUzrKRw~whq8JZzc05R7vVE8H7m9RNw+y
z&aw^*oOW2Hmz4@$R=P8#Y=MdMoKsIbvNTFE>GMLAWzI#6%(rCnm8HfA`+)_iRV&LP
z^@XrnD3IEDsu_Ny33jsdxjV}xFD&J_2<@fc#I0_B-S4gU>uIi8Z@%iapb%PE>IH}C
zrm11GS=lT^Mlh<g+U5_RAp`ImBQ(~i!4MJ?CxnZoWP*@KkO2Y!Gxle>Z$~ujVTx|5
z;0d`E+Ew;yxc+{(9`w%J2Sa|WZzjWkgFOGr$J4_zO9YV3OCW05oRDiXjLkSfMG}#%
zMuepEN!jl{If(KTGREg*PK_9J`)%m=+Hkfl#qNgx4{SrX4NYlx_X#6aRRq+`&#j=G
zkgwRtUpW6MwV-xbl~bu{>|Eqjnxy4U%~~daUSd;kKPb!d-!oNFUxS|VO8t{^n5v2`
z2pR5OGm1T(bi1w7tV1R%d4I1&J9bo(<uoONg{wc(tlkEUs<(E;8>udX^3tdcQa=BL
z9?QHyPogJaSxFA1D7e<x-{|A`tjA2_%>71k?$#sC@!*jS&fj4INYnYfeI5l6a-l?t
zww+jZ3_Qp2Q73&6=9^SFup(?$za5T@dPm$8*TYdG=OgI8M(oDH%Y3%6+$H;?caC9I
z*@ZKdC0$;-;6V%gYKC&E;oG<@^LkYOJ%??BUk7i3zcz%_H@W7T#sG;rsyy-NgEIGk
VlO12)$EJAgZ>2KXk<eVuvOn9sIS>E<

literal 0
HcmV?d00001

diff --git a/dnn/src/cuda/convolution/backward_data/int8/kimpl/cutlass_simt_s8_idgrad_identity_s8_32x128x32_32x64x32_2_nc4hw4.cu b/dnn/src/cuda/convolution/backward_data/int8/kimpl/cutlass_simt_s8_idgrad_identity_s8_32x128x32_32x64x32_2_nc4hw4.cu
new file mode 100644
index 0000000000000000000000000000000000000000..0cc081ec86f7ad01bed385d0fd59cac309eed32e
GIT binary patch
literal 1715
zcmb7F+in^$5PjdTaEVm45h>y#6uC4~3MrM^BvpvizO9UH77MR!`NAb%-|>1Ea;@mg
z+Sqf(b7tlY*@;-7ch^_b$<0lEe>J_k%>Q(=P6vrXY0J@uSS3_}^a9b~C2=l0F_si1
zLDTUVF+sr2c|nm0Zu!<-a>h{W>~~PXadu{iB_;RVwvNC!Zw2*dL<-^{7=Q>xWndh&
zoTVKa80k>0)>;NHwCc<WT|wY1XT+0^D2)_U_`Kj{nKK?O^O{V)vBVf|KhYqRY-L%b
zzT#GKiG*^VC<>1>#!i$zcjviag(VCYuDtNOc+~H0yS?*mzZZAILEQ9mMF-2xpoI#q
zxFw!5G~YF8Y!0=}LL`MIRbVv*^aW)I`JLek($ruuj<I9R_*zho@e{}h00@Hi=eci3
zG&;Z}+3DahzURtmdo$jC|GgazF4{*!er)e%<Nt%ad<AlHbYzJDuz3wcEL~!JYsRr3
zXHb+xq%IT2>3&uYyU#9S2Vgvf?`%m7AN6}}1Z8WZey>i(-p6aE!@)6@UGUjPO?LMV
zxe-zq0m|l=mJ^2YS7_%;#(oOLi5=^5E)<FV3$#j;%DFFSRVjo9iBNU?C^IjAFQg{E
z0zD^|S}4UZmzu6H9v`$biJhHFcVC5Bnanog;X(3tY^qe5UJ3*fQa#cg)*EjFm{n(c
z=1sY-gY?449dN$*3{uNH2Wf&X&{_+IN?tIfpvOVZ@mo!sMxcj-B;Z;n&GqPs6fQnM
z0Kn7vgM^+KV0<Nc3F@6$dWu*MQ=|&|D9kq@VQxj(qS_sgj0KOlIjqN{CeBCDLxtFn
zgxC4hw#+4=fqqWWM%$G$gr>T@<={~d{Ay-%%HjLCtn+GO|2u}d!*7E(!QUE;D^IR*
i(<mTOhbqrJ2%*e9f&mYkkFhD<`ghSLJK>6nMfMl%J~~kV

literal 0
HcmV?d00001

diff --git a/dnn/src/cuda/convolution/backward_data/int8/kimpl/cutlass_simt_s8_idgrad_identity_s8_64x128x32_64x32x32_2_nc4hw4.cu b/dnn/src/cuda/convolution/backward_data/int8/kimpl/cutlass_simt_s8_idgrad_identity_s8_64x128x32_64x32x32_2_nc4hw4.cu
new file mode 100644
index 0000000000000000000000000000000000000000..5eaa5d469f5c63686b53d679be09d077adcb0370
GIT binary patch
literal 1715
zcmb7F+in^$5PjdTaEVm45h>y#6uC4~3MrM^BvpvizO9UH77MR!`NAb%-|>1Ea;@mg
z+Sqf(b7tlY*@;-7ch^_b$<0lEe>J_k%>Q(=P6vrXY0J@uSS3_}^a9b~C2=l0F_si1
zLDTUVF+sr2c|nm0Zu!<-a>h{W>~~PXadu{iB_;RVwvNC!Zw2*dL<-^{7=Q>xWndh&
zoTVKa80k>0)>;NHwCc<WT|wY1XT+0^D2)_U_`Kj{nKK?O^O{V)vBVf|KhYqRY-L%b
zzT#GKiG*^VC<>1>#!i$zcjviag(VCYuDtNOcr+YryS?*mzZZA?UflF@MF-2xpoI#q
zxFw!5G~YF8Y!0=}LL`MIRbVv*^aW)I`JLek($ruuj<I9R_*zho@e{}h00@Hi=eci3
zG&;Z}+3DahzURtmdo$jC|GgazF4{*!er)e%<Nt%ad<AlHbYzJDuz3wcEL~!JYsRr3
zXHb+xq%IT2>3&uYyU#9S2Vgvf?`%m79}Nd>1Z8WZey>i(-p6aE{oXN_UGUjPO?LMV
zxe-zq0m|l=mJ^2YS7_%;#(oOLi5=^5E)<FV3$#j;%DFFSRVjo9iBNU?C^IjAFQg{E
z0zD^|S}4UZmzu6H9v`$biJhHFcVC5Bnanog;X(3tY^qe5UJ3;AS3S}l)*EjFm{n(c
z=1sY-gY?449dN$*3{uNH2Wf&X&{_+IN?tIfpvOVZ@mo!sMxcj-B;Z;n&GqPs6fQnM
z0Kn7vgM^+KV0<Nc3F@6$dWu*MQ=|&|D9kq@VQxj(qS_sgj0KOlIjqN{CeBCDLxtFn
zgxC4hw#+4=fqqWWM%$G$gr>T@<={~d{Ay-%%HjLCtn+GO|2u}d!*7E(!QUE;D^IR*
i(<mTOhbqrJ2%*e9f&mYkkFhD<`ghSLJK>6nMfMjP3_4o?

literal 0
HcmV?d00001

diff --git a/dnn/src/cuda/convolution/backward_data/int8/kimpl/deconv_int8_implicit_gemm_dp4a_ncdiv4hw4_16x128x16_16x128x16_id.cu b/dnn/src/cuda/convolution/backward_data/int8/kimpl/deconv_int8_implicit_gemm_dp4a_ncdiv4hw4_16x128x16_16x128x16_id.cu
deleted file mode 100644
index d3d701ea716156562346f89547b09461fbb44aa0..0000000000000000000000000000000000000000
GIT binary patch
literal 0
HcmV?d00001

literal 1829
zcmbVN*-jfV6n)QExIwB~6o_amgiMMANt#w=sTd^sHgardV#VXJyd;FL@3m*i1c-p<
z!EtP#yPb1y_QVW*xEfDJ*VpOYcyfE5{ypmS`zYgz8{#-a({=cfQqM?A3$;wAg4lGy
zjY>tXr9IT^01%lf3`Z+slu#KeW{7$xiM5?QZAg|AG#L&N69nE_Cn(b1&Uf9F#u(}z
z+<WCMX9t=Xl5@vR_YD~9jG*p-NQe}I0f<mkdd5-L8rl!3^g}9z^iCA2KP7YlfwPnm
zNBTm!lRo1y*RYRi3MyQh@jQn@#I(|tAzE{DNWD<9+v)h=3SE+Q;ay@V`d$|RV;K)I
zzT?Uk=6d+>^ZsG*?d)_5^Fv5TZ#V4s!)-Er4jWmw8}rQ7-P~}(E@VL$$((53hKB4g
zp1>Nf21my|boAu}mg6SWdt&r|0e>kV#nVvk*YV)8hc@*YYaxrw^Bdg`msh74yO<Ql
z7eMzpW0yj4VunRN6^aChLrHQy=%5!qkzWVSdI+{(zof8lHZxfe*Q7p*il`*h4kQQ#
zW5$<)a*Vg0l}>9XjW(i>li;g7DyiR`#!$;Eptlv#OK0(MB}$Bmp>vFHG~{e;B62`K
zxl8h6sBT0(RYMOo{PxgO#(HSOLxpcFmnA59m<DlSJlyoxC^LkWQY<~jC&@}YK1yDh
z3WWC%hz(RypEc&|q^`VlLd!Lf474~7;|E3*S{8z#oM%iaj6-vWb*o7?2W4MZEY>|d
z*dp!hQzI3~3);)AHCHfw(AG1I$CBq<g=XJJEQQHZ<q~xECS=Kb0MBCczZV|_AJAbU
zzZVO_3ed+E?v*?%1Ph96Dg@fzN2|gttR}Q3E)-ywv6e3W)HA)T*2nO*Oyi_)dDvyB
uQKNU-yBm#hv}bE$^?YcV3ff2<Ou#&K2<8!NKA-OTKhcHm>~Y1!tn&{T&u_i}

diff --git a/dnn/src/cuda/convolution/backward_data/int8/kimpl/deconv_int8_implicit_gemm_dp4a_ncdiv4hw4_16x128x16_16x64x16_id.cu b/dnn/src/cuda/convolution/backward_data/int8/kimpl/deconv_int8_implicit_gemm_dp4a_ncdiv4hw4_16x128x16_16x64x16_id.cu
deleted file mode 100644
index 934d3264512000e1a6ee531a5dab297aa59e94b7..0000000000000000000000000000000000000000
GIT binary patch
literal 0
HcmV?d00001

literal 1828
zcmb7FTTdb}6n^JdbTQfN8kFc7Vk&E*z~W}TY{bNUo0PT;4V_NY3y8mdPdgU|m8d+F
z()K&I?>mP*F+*Q&M&se_ZSpW0-(MvEj#}L=O1a{OIL^>?9eyO#Gm_9;Et9DrHd%0^
z5|L?X5A`|#M5=Pb(TW%)REqK$qTWejZEH^(l4b;r2Lr?efp^vkinO=$U3Y~shS~>@
zUU|#efhLAz+;P+X0LD5as5>AMB86Z8A{3RLan!bkc0(%NkV-DS6S?Y630*+oEMdfv
zu2Al@%XrK+>|>gO3YVlj%b*Z3tz>10*4!LYFO+PzT0Xc!*JPc0H!&3btO|g!jE5LM
zaAk9IJ9xT!eCnT^pKW1&2?^=#hW&ZEpA25ZhSu%IymIw0H=M9bnbXB&PBd>qLv|RC
zVU1V)qh1Fc9iPI|t3!PxMt=y{JK4hhR=|m;q1x}m0og&D@{F{QdFuI{Zbr+?Q;c2A
z3FAv(`+~7+p*S&vJevwdg2RC%neMmH8=pw817{rs+aJdX?CZ@;=ET*R52GMT#WVv6
zI>DInrJx++t!IVP%1JGa=;Ji_Dvk=~H>WYw;tJSpMD*5Kd|YZ0V`AtW<2wyGTRRaM
zpkLfgvK~}7qF#!jgKBy^=p|zvwBez|*OrSC6g*6WwlE%S+H05^!U`o89^;eAN<2SH
zUYa_N9Gr9zsH8fp%~wg?c<F?eYakhDaU7-(OenO>1w$E6nNk>s-VW<lk8Tc%wk}z$
zTDZSO+WEIyDv%d+ms@MDVECY|XBdwp&$tS`zK>V}gQd(R=<Gwtg7*NP#oqrcJ_tUb
z!$AHl7K9a`&kfu=d6ozk6j@gYw7rj3xmj3EXhmEoz%FAAUHq*^dQq*f;cJ=1LErGO
w%TTRG@3eP!8slisR>rFN&@dIWF>x>ev&12oN3i*Rw(IXi=eo7W6%(`8e^L`~iU0rr

diff --git a/dnn/src/cuda/convolution/backward_data/int8/kimpl/deconv_int8_implicit_gemm_dp4a_ncdiv4hw4_16x64x8_16x64x8_id.cu b/dnn/src/cuda/convolution/backward_data/int8/kimpl/deconv_int8_implicit_gemm_dp4a_ncdiv4hw4_16x64x8_16x64x8_id.cu
deleted file mode 100644
index 28512b4914d68cb0b3c52d81fc3fbb82c91de502..0000000000000000000000000000000000000000
GIT binary patch
literal 0
HcmV?d00001

literal 1825
zcmbVN+fExX5Pi>AxIwB~v=BiNLMBCmB&AikR0N5>tsL9gSn+x-UlPLCckI1n14Kac
zuyJO6=623ZcEl8Yx*lB)Z*J22(dFHF`uCvI>!FM*ZiwRyO;+JYN<AYfEz}~N2x8MY
zH!2mmmUdsS0zhP{FdQw3Q9@;?m?G+(B-VCzv>{ne(B)u&m>_V^Izf^4cDC(Z>5QT7
z-h)@(a<-?5Avt&4bl-uo&Is!Eh=fcb8Gr~yrDq&<t)ab8N-vaBNbf|UdJ{tDkT^>j
zaik}dJLxeVYYpdErl7*58P9WQL@X;^8lp8f`_v01yPb{?q0kjs72b^vMc?ZNU@W5%
z<9n`bVQvPGKOY|ZM`x!Sm>)tydD~&XAMeJ4m$0FA+c7Uf-Omgs>_QfFKAsWH+t5%Q
z#+UGnulfha2|7MX&`Fc%12OtPV80X);%R90+vo%xBxv26kruMZJipcLRC#5^*u{b{
zz5uGv8M_jS6Ei6CiBKd297vMueh0nsiQGB}mLNEOJ51r+9Hz1$t|@#N)lkW$ok$P~
z#*8ln<rr^VD}&ZS8ePN~Cm~j8DrsL^W2)sVpthCKYqof~5+ugN&>6<J8fvz7EOJ0U
zdpFLHp}P_FTnPzk*iF!L#uBvVp~5$w%NCS8OoFg59;|0;m>I%K8I}>_mGM$MJxN}<
zI*c3~B?wef+Zy+EQP*BNq2&rl23j14)dTAZEegR<&NHSI#$m9-b7Mf)hH|DW7VG)#
zZ;*EOrI8BM1*7Fwnk(2mXyY2jBgu2F!l3UWmcmx4Y6&`f7qa9%fTwZrKZ*~659qLs
zKZ*rm1?W=?_ePx+f(1o39Rh9dqGe&`Rufti7aFk5SxXmx>U~~z>vQ;8q;bo)B5ZTi
usL?y^-L1wr`m?pOdOx&W1#OHSY`;8p2=);izMgLTFVThW>~O`zwDS)*L2mm1

diff --git a/dnn/src/cuda/convolution/backward_data/int8/kimpl/deconv_int8_implicit_gemm_dp4a_ncdiv4hw4_32x128x32_32x64x32_id.cu b/dnn/src/cuda/convolution/backward_data/int8/kimpl/deconv_int8_implicit_gemm_dp4a_ncdiv4hw4_32x128x32_32x64x32_id.cu
deleted file mode 100644
index 3fdf28326b729d102a2fddf01efdb7fdf3724e96..0000000000000000000000000000000000000000
GIT binary patch
literal 0
HcmV?d00001

literal 1829
zcmb7FTTk0C6n^io@QO5TC{R%lLZ(B4w(F*GnFtbli(K1HtvZfnJ1ync@7TFC1uCEq
zse9t@+`jLe?ui-tax<QcZf}!^@#Ow8`FGUm^-#(cH^gy<rfdHrp#>ue&DAoQ3SyH5
zH!2aCmiADuJwT)?Hyo{qQ9`9CpCMW}Nv!ScX+zSCpviEEm>_V^Izf>x?0na|(iubD
zgU3P@ma_v*49U3TruzYmbw*HkKqO=e$pAzsDhtL@*BaXMrSyC$xh$N>Rc}h@0upBl
zBaZZha%VlpL#_TC$`n+%B;{EKjR<8WD?_yA=8zUb$!@1p_)zGYtn<RfhN2&J12C4s
zi17ngHaEA!r(cgxgOiK%EzD0Lp}gI&KTr4Z@HK2?-EPdQP!Ds%3A>UxUBq*uc^m4h
z!*~ME@M>_}kI+&7EJDZqCe=q`^oM|_CtJAR3OFgK@AkX!BM2LzO?$>#$UH6hoo;8#
zD=Wq>)P(UBuzktcwNRXxVV+HeB0k_yl1vXe=#5XH*Fmrd!ST4Cz_~fhWKLXD`6vjY
zQcOD$uM><JUkb`G-nv!>t%Eeu2r<rltkP66zp;i?%U8f|E2Fn;;c}%(jESLhjPEqm
zY;7zuK)-kwXQ$BJfO;v02sQLZ=p|zj+VD`}8_#76N*<<OTNn>F<26bRVWkqwh~Y}S
z63@?)SFS*6CqVmYgg_>>vGHG5byG+uv|Izrz>DLseqcqRWiA-Xc*>N**bldVZVl<i
zP|kHlV?D!zE!Hl+HClnfV7%N~a|PQ6ZC%57EP2M2ANGC364)$NF<xjNLYBmP@GK1f
zXYpR~9_=^sXR)9x4}EUo-l?-fFfYlbLm=*bw93uGYC>z`d;@klYYF3Tz0=EXef3|<
zBy9Rtgk6prITlVA?oMMI{Mp)By&zhyyf$J78!$^8f`tT!@8`SzPjs$3dt5Ow>-+~j
Cq;ItV

diff --git a/dnn/src/cuda/convolution/backward_data/int8/kimpl/deconv_int8_implicit_gemm_dp4a_ncdiv4hw4_64x128x32_64x32x32_id.cu b/dnn/src/cuda/convolution/backward_data/int8/kimpl/deconv_int8_implicit_gemm_dp4a_ncdiv4hw4_64x128x32_64x32x32_id.cu
deleted file mode 100644
index 20ba90c374eb433b189273f757dc8c466d972947..0000000000000000000000000000000000000000
GIT binary patch
literal 0
HcmV?d00001

literal 1829
zcmb7FTTk0C6n^io@QO5TC{R%lLZ(B4w(F*GnFtbli(K1HtvZfnJ1ync@7TFC1uCEq
zse9t@+`jLe?ui-tax<QcZf}!^@#Ow8`FGUm^-#(cH^gy<rfdHrp#>ue&DAoQ3SyH5
zH!2aCmiADuJwT)?Hyo{qQ9`9CpCMW}Nv!ScX+zSCpviEEm>_V^Izf>x?0na|(iubD
zgU3P@ma_v*49U3TruzYmbw*HkKqO=e$pAzsDhtL@*BaXMrSyC$xh$N>Rc}h@0upBl
zBaZZha%VlpL#_TC$`n+%B;{EKjR<8WD?_yA=8zUb$!@1p_)zGYtn<RfhN2&J12C4s
zi17ngHaEA!r(cgxgOiK%EzD0Lp}gI&KTr4Z@HK2?-EPdQP!Ds%3A>UxUBq*uc^m4h
z!*~ME@M>^+5}~91S%i-JO{$N?=nnyp`&+o*3OFgK@AkX!BM2LzO?$>#$UH6hoo;8#
zD=Wq>)P(UBuzktcwNRXxVV+HeB0k_yl1vXe=#5XH*Fmrd!ST4Cz_~fhWKLXD`6vjY
zQcOD$uM><JUkb`G-nv!>t%Eeu2r<rltkP66zp;i?%U8f|E2Fn;;c}%(jESLhjPEqm
zY;7zuK)-kwXQ$BJfO;v02sQLZ=p|zj+VD`}8_#76N*<<OTNn>F<26bRVWkqwh~Y}S
z63@?)SFS)UCqVmYgg_>>vGHG5byG+uv|Izrz>DLseqcqRWiA-Xc*>N**bldVZVl<i
zP|kHlV?D!zE!Hl+HClnfV7%N~a|PQ6ZC%57EP2M2ANGC364)$NF<xjNLYBmP@GK1f
zXYpR~9_=^sXR)9x4}EUo-l?-fFfYlbLm=*bw93uGYC>z`d;@klYYF3Tz0=EXef3|<
zBy9Rtgk6prITlVA?oMMI{Mp)By&zhyyf$J78!$^8f`tT!@8`SzPjs$3dt5Ow>-+~+
C=x@3J

-- 
GitLab