From 55e35dd879eb48b083dad2419d26d2c6d5cda1bd Mon Sep 17 00:00:00 2001
From: Raph Levien <raph.levien@gmail.com>
Date: Sat, 25 Apr 2020 10:15:22 -0700
Subject: [PATCH] Dynamic allocation of intermediate buffers

When the initial allocation is exceeded, do an atomic bump allocation.
This is done for both tilegroup instances and per tile command lists.
---
 piet-gpu-hal/src/lib.rs         |   5 ++++
 piet-gpu-types/src/ptcl.rs      |   4 ++++
 piet-gpu-types/src/tilegroup.rs |   4 ++++
 piet-gpu/shader/kernel1.comp    |  14 ++++++++++-
 piet-gpu/shader/kernel1.spv     | Bin 11644 -> 13456 bytes
 piet-gpu/shader/kernel3.comp    |  20 ++++++++++++++++
 piet-gpu/shader/kernel3.spv     | Bin 9964 -> 13176 bytes
 piet-gpu/shader/kernel4.comp    |   4 ++++
 piet-gpu/shader/kernel4.spv     | Bin 5704 -> 6680 bytes
 piet-gpu/shader/ptcl.h          |  39 +++++++++++++++++++++++++++++-
 piet-gpu/shader/setup.h         |   2 +-
 piet-gpu/shader/tilegroup.h     |  39 +++++++++++++++++++++++++++++-
 piet-gpu/src/main.rs            |  41 ++++++++++++++++++++++++++------
 piet-gpu/src/render_ctx.rs      |   8 +++----
 14 files changed, 165 insertions(+), 15 deletions(-)

diff --git a/piet-gpu-hal/src/lib.rs b/piet-gpu-hal/src/lib.rs
index c62678f..77170c0 100644
--- a/piet-gpu-hal/src/lib.rs
+++ b/piet-gpu-hal/src/lib.rs
@@ -71,6 +71,11 @@ pub trait CmdBuf<D: Device> {
 
     unsafe fn memory_barrier(&mut self);
 
+    /// Clear the buffer.
+    ///
+    /// This is readily supported in Vulkan, but for portability it is remarkably
+    /// tricky (unimplemented in gfx-hal right now). Possibly best to write a compute
+    /// kernel, or organize the code not to need it.
     unsafe fn clear_buffer(&self, buffer: &D::Buffer);
 
     unsafe fn copy_buffer(&self, src: &D::Buffer, dst: &D::Buffer);
diff --git a/piet-gpu-types/src/ptcl.rs b/piet-gpu-types/src/ptcl.rs
index b6df77d..ed72e42 100644
--- a/piet-gpu-types/src/ptcl.rs
+++ b/piet-gpu-types/src/ptcl.rs
@@ -33,6 +33,9 @@ piet_gpu! {
         struct CmdSolid {
             rgba_color: u32,
         }
+        struct CmdJump {
+            new_ref: u32,
+        }
         enum Cmd {
             End,
             Circle(CmdCircle),
@@ -42,6 +45,7 @@ piet_gpu! {
             FillEdge(CmdFillEdge),
             DrawFill(CmdDrawFill),
             Solid(CmdSolid),
+            Jump(CmdJump),
             Bail,
         }
     }
diff --git a/piet-gpu-types/src/tilegroup.rs b/piet-gpu-types/src/tilegroup.rs
index 4824178..5912154 100644
--- a/piet-gpu-types/src/tilegroup.rs
+++ b/piet-gpu-types/src/tilegroup.rs
@@ -10,8 +10,12 @@ piet_gpu! {
             // A better type would be Point.
             offset: [f32; 2],
         }
+        struct Jump {
+            new_ref: u32,
+        }
         enum TileGroup {
             Instance(Instance),
+            Jump(Jump),
             End,
         }
     }
diff --git a/piet-gpu/shader/kernel1.comp b/piet-gpu/shader/kernel1.comp
index dbdd492..82ccb8f 100644
--- a/piet-gpu/shader/kernel1.comp
+++ b/piet-gpu/shader/kernel1.comp
@@ -25,6 +25,10 @@ layout(set = 0, binding = 1) buffer TilegroupBuf {
     uint[] tilegroup;
 };
 
+layout(set = 0, binding = 2) buffer AllocBuf {
+    uint alloc;
+};
+
 #include "scene.h"
 #include "tilegroup.h"
 
@@ -43,6 +47,7 @@ void main() {
     uint stack_ix = 0;
     uint tilegroup_ix = gl_GlobalInvocationID.y * WIDTH_IN_TILEGROUPS + gl_GlobalInvocationID.x;
     TileGroupRef tg_ref = TileGroupRef(tilegroup_ix * TILEGROUP_INITIAL_ALLOC);
+    uint tg_limit = tg_ref.offset + TILEGROUP_INITIAL_ALLOC - 2 * TileGroup_size;
     vec2 xy0 = vec2(gl_GlobalInvocationID.xy) * vec2(TILEGROUP_WIDTH_PX, TILEGROUP_HEIGHT_PX);
     PietItemRef root = PietItemRef(0);
     SimpleGroup group = PietItem_Group_read(root);
@@ -62,9 +67,16 @@ void main() {
             if (hit && !is_group) {
                 PietItemRef item_ref = PietItem_index(group.items, tos.index);
                 Instance ins = Instance(item_ref.offset, tos.offset);
+                if (tg_ref.offset > tg_limit) {
+                    // Allocation exceeded; do atomic bump alloc.
+                    uint new_tg = atomicAdd(alloc, TILEGROUP_INITIAL_ALLOC);
+                    Jump jump = Jump(new_tg);
+                    TileGroup_Jump_write(tg_ref, jump);
+                    tg_ref = TileGroupRef(new_tg);
+                    tg_limit = tg_ref.offset + TILEGROUP_INITIAL_ALLOC - 2 * TileGroup_size;
+                }
                 TileGroup_Instance_write(tg_ref, ins);
                 tg_ref.offset += TileGroup_size;
-                // TODO: bump allocate if allocation exceeded
             }
             if (is_group) {
                 PietItemRef item_ref = PietItem_index(group.items, tos.index);
diff --git a/piet-gpu/shader/kernel1.spv b/piet-gpu/shader/kernel1.spv
index 0e9a497e3388a17e9fb102a7409c01a7dae8a420..9ac359394ba6be6c962643b79efacce5218e75ce 100644
GIT binary patch
literal 13456
zcmai)34C8wvBz(lh0;yBQI=8?TA|Q|maUW)XbL4Q6q-^HK{PK-{;h#5B}*4XZCTvK
z^}V8iR%H_r1O*WnESn&rD2n)8pManb6-4&M_x;}go#f=N@8h+H`OR<U%*>f{?#)SS
z)$nOMW!bQ71pk{dJj<VP*>IRF8<o}OvT?=g6?1y~+U6W`=v*DfW>vYLK5xLM23gJM
zY-#V(a3F?V;N`|K^s6gm6?08<2V`|NGHYDX+_-XO<BFxtt?Sn}_x5e>C^dJqca@sk
z+Ivc^eeD}c`i)5Z+Phji`rA^68a&ec>MSZ-vZi~}$)&Y<h3Y=tYuEOc`Z8)Bm5o8y
zr@6hWt+Z)IZ(mP;Yv1f**6jXy^A`5cTZm5$HIG8}5X;!3!TFTjRznZ6tf{@!x4f^^
z8TsUV2J?-{CZdPm(44_CyCUy+=0x5J;K97r*>31oi^$ig?z5)gYUuBt==mCZN-b^w
zYKhog-yX<vo|<ega8q}CSD%aFafju62G>)a)uZQrNOxP^XVYe5j?AWmb9-}|JR;`{
z&2vw~ubjI!+ZVjLy>op>sj;WKe|=to`Q?2E^VeiE&=cRhPC1{h=5}_YcYx!#Wo4$k
z9{oIr_Nvr7$a%5{bBxJmq9-|qtR&&7b8tNeFsHnp(b<9E*xly7mUXcO!3|8r&wZFP
zFn3ILa5^_{$B^{@6vFuj4a~2}4h1jo>g{XkYAxAF`zYr#xW<}nK6?0RtL1-kKK6{P
z1Lr{AUJlM5pDjX<`OO=9m=dcDHda}@VnNfg*&Eg#<V$AlM-azUV&6Nq1zepSg*?8$
zGd~QjLEUGthU#oFx<2yysQYx4Hs<Bm@YPz1uFudKi^%!*74=D#w==Kh80JM?$AZgs
z)n<+0Q`*y$TFx!+Gg#lGYz6v&Z*w$YXe9&of8VaWuH%Urb)5h%*EK#{H6+(RulWDV
zlPo)rzcPv6G!TDmcIuG$OS{_s%`$M$<DG(xdd~v4^W}1XyX9Lp=096%V9r5b-CF7@
zE$JUPcXB?x@<`8!>-90lxkzHnE9+I+RzSHQRoOOh<hr!%SLCYBwwLvRc~_ywyc@wW
z@8+`Kz`War%&X3BFYCU3d2RO~%W<l+hr#PxdRjVjTXpv25WOmU3cZ}CHhTu_{;c!s
zH?aOPk16X58`vH^E6F3S<;9G8Cxg?RyH)7ZD)i|?^qOoYI$w^~vzv`*{Mu|5+(7;T
zenkHBeqBd%V@LOzmX76J8@gLt`r5m@miw-_H}jphuKX2oe;1YYj`q%WeiD~!Z-Cph
zd2a51Jh-R3yRcPfE70?P6!A_(A6#2))`Y&=T9<Z|I!j&oJih_*J~1~q>G(MsoIeW&
z-`A?_WSsiCd-Hlq;K84%)!AC~!Ti-(dxhRPM6b%u$>*2Xc`o?vEZaV81Go1!M|nAS
znOySCwSSkxrTMm3=vNHUYqIPAMT{GU_*7>%4biKzTb$EXUaxmEzngvQthpw;6TP9M
zqq}wBJeE0o+aizj{N0V8k2T*5KCAd1tKW~FuHk_S{oxAzks*3b_9%K~UmhFcQ=L6g
zp}&BhpW&jWS1arj+4Qug-4eZ#n%6V0mf_uU8UH;8>Ek_X8LJH4`-@K_J&^Mb;e6l!
zeHcEW*O%2?e;Pa(zb1e8HK+0X9z;Cf^)#Nk@mN+K&wID`)F}+>bKiW@cybZX{YvAh
z*W*tc@_2g!jJG>u0<3&Jvz;%V#2T%?j^Q&Ei}&m@=dC@<b-6lxcP@18_0%`A(Dl>i
zeDBRMf0x28r~f#lSYChg?~1Hr3}Z}2pNy1q&y43@$<@QS4t4Ws^G+MNrh=_gZdYu3
zBkLH(RI0mX=hY+Y7{=Tum-1!XzQ~BVAJ~|3Q(<QmHrF6;9AoT{)TY1sY@{(<r}{jk
zc6IZ6hl?DCfsG|M6<_aXWt%ltXNMPdd3(43={lUJ&3N9YBHqzp<H>1Tj0{@?Seu-&
zmLQGg9BukHBEx?fSbsVFmm~F;H@DnL31<w~yE@@!qPw4`F?`gG{}#rGL~jGv6}tZZ
ztToqO)U+Py9P9J_RJS)hNPXlM7P{}OT;%Bo`{=(Bsm)knKMy>Np?wR|^?Jsl?hC-~
zoA!?*wNEbeD+|58(61}>I%LG(0XC=e??yV`vk>v`0}o5|2f@CdG5;a3zpsY=IM{c?
zzC4YLzB~g~H_zXY=9IH9>V9jYj#t3e;XC2`F`EL6V_xUaLFZ4-J73*zlbrUs1y1|8
zKqG}2ZwXkN`K;SK$7#>9V?%Gu^(_svI74+9mZ9q-KZ`lOV<&*!zp40m7mzb<T|TO_
zRdB9D+eEURiQJ!6DBsGkmTsi^?U}W&M|Y0A`sPA+PW1R<u)R2-h;<3NXU}hMIe&F_
z7ufa5*&}tozp-Zheu?hcnS%W}uyuOw&HFsqI^<^&)jjZg?RjitC=X-Z#<O-GYkCo;
z25C;eq2BY16S{Z2(0xz6yM?a*qD1$+HYB>WdUp$d&#QO0(6yhK==R@xR@k>-FD1I~
zw|A|uo8P-u=-N9I-M#Uy6?XT=`&O!Z=L+3=F00VJYlYo>+Y?>?D=KvFTH&w#hD3M1
z_pGq{&GViWx_0kcp*Ns=ZwlT1dv6L|`vZxt-Fs8mjqklFbnAJvLVqmL&F`Hl{5@0N
znL>B|s};I;rm&meJCpiRoG#Cj?^^>WY%X{n!}s2^cqHe|Gks9OUx<(T!RT^+mtx)_
z=*}}%%sUiaeoHa$FtBsw&F#6B3;)Bxa*^i<u$<qSCTeLU_Cm0{^F8P4)?i<b1k3qz
zCt@CjE*~`<4VE+J`9)oe!ScR?#`S%XGluWzF$~wXlCCX59$RqoK8p4D)*+XoXupWq
z>dV1$e(Q|wyC*01z4Lo#9=~%Zf#tOM&Z#dT&J1iT!3QJdwfm0B#oX0k=QglIzT>BW
z^?48+{ip+<3a70JKfg_HVmQ}*(`IbnZ~2$_?wp47`=flGI!{NQ!SK=kmcp)l8N1)$
zH#78eo;ddBZD3>H$K2STW_0=R`46ylN6ntM7IgWq(2FyXYZ%Uxe}Q`2kUr+u*2);k
z88^S!*dws70~_0WhWUL5-i|Ksy8UjcM=fW8E9*TQPF{b%Yw8Q=rLj7|W01ztrcWnQ
z-hK3r(FNWUDX-mcr<^&xYxIESV*QihdeP;(sks~3$B=W5d)*JV&RAC++y->{mAo0w
z*~pM{_!J_STZyC3CWf)~wKl)Ya^~^=@D8wCyyO0iJr7+zeBLRp@Oc-y{Ajx5oOgry
zlb>tnYmfO`z~<P2-96HOE4qC6oG)gibzFchzm-~@b0JvnI7Z~X2+W`CjYUj-ZpGg-
z^&W<G>EoGtAJWG&rR}{8InR_hYI#4{dgFU(?hl~LhtCJa6+RzAmyg;$4CYVX6X$D>
z`5ytB<6=hi<fCBu@cEdSk=F5Xbos4CPd)*bYh*;;PlEZA_r#d`L{I(`Y+d^JuE+Y7
z&#{;G?Y|iI$M;Nsd*j*hXSg=;HsmD?Yj@747~1XgrAQz9EdObSoP8F%K7Xct2E2k1
z-wwZ*pGB89$K^<M{ry?+Ik4|U`1>9CJi5I8+mY(NyT<wgcp}m`+Vr^$DR15STnU!b
zb_G)1_uAN3fe%E=Ukm3tuV%=5PMu>P<m_dvQ5k)?25etK|038q*A#ub7A$vl5&uit
z81nY=%Sa!4sqH$3oV^rBFTVn=?B$Jc^456+QayV4Rd8i5zXm6-|4m5s=;h5|dubeP
z`dp8ckC<Nv%f)xj9^Q&BuaDogZ-C94e%Ee8mv6#nK0e<BJ5Sy`w;<&r&$qzy8EKwx
zqs!Z8^L_`sh;d!f*W1DR$lKQ)NFV#E?Yj&)`zkiqM(p>1?`D|mPrM(-{vNu#@1^Hp
z0oXe4Kx&H`?gPtt&smJ$_rY@d-iuV%*E`S;z^+N#{YZ6f&if%)e{ByS)!i3<9A-ZP
zuR=OUTjYKa9J!7EV{|!Vm`6SO`4ezuKOcgVk2}^+!E&xO?pQwq%cplNIp2lor83s|
zb8xKjVX(S3_xu-N_u2X$L8`|Z{~NplALnR`x_<?(tozq+@=^D1z;eH2#Jt~vW1jic
z^|S7Hx0Kd)Ch~WTZ4A%N?~y*98~OiX)E4&=u{r#<Jr1sXAO8R+Z_dY%>QU2^VBd@M
z+xACvdHtV2s>l2IC$R6MakT04C{jLR{x7&P=2LL;(Vstq<;-ggZE_LgFW|};e}$6|
zpQpjDHGLnSL6?vF-QU2@ledq5NBY=DZO<~~eE-C8j$Z`NWJJ9$>BES6waMw<gnXG{
zz1Fr0`3mC&hI3q>x<2l8<WoA|p9imj?W4Z0BGu!*{yNzEy1aJREEjYC0d}sr{nn8p
zuO<D~4fCaC#GQUP*f^DU`YL#BTZvW2lxnbZ<>Q`T12(7MAnn%V+T~r}@4@!d^BH@!
z6Ih@0oQ;5!j~?#~mWv*1lS^Y7TfMEw;oUs;Y!ujiiJzIH(dDeuy&ntqaqqQ_VaQpl
z*tn+`apfcKE@0!v_h%fsT-2#eF7|#r*u9rChUa4fSU!5CuD|c>B(S}SbL@H86;57%
zf9F<@^Dr6gd619uAeZ*V{-~dUbbU(o(+b^nTASbK-I+Td-B|uUuWl|xwg>+?6)FE(
zHjKZw9hU#SlE0()`-zW#k3XNe{$Aw27l?O8j%Jv9EHci)o5A|TTK&0pIzvACeFoUM
z@w@w(U^)NpE6&PW!Oqj>&pGRP8$<qea*Cr?`{_5!XB@-&?EClvTk8adwN64ttu0`E
z(puM`%SWxPVCP1yC9qr@BWhg>cAhqCwVrhhd21C%t^Qu;?{q%?dqwnh53u!3LB`+J
z`ETg1JM``f{oF*~0^VBavA(m>jhn8o16}?O_TRnn|Ij7px5V>dPPtfrA6PEdx&bWb
zJLk7%BXSd?mth|5F?Taqn{lmOJ#6m)r*YqjF6Z4R;=T(Uaka<XcZ0PV*B+>c?R>Dg
z?kd)IAz1DL#(C_G|8{T@L;g;D)ZbHZ`b3=ff{hb%-UpV8IqwI{rE@-jZatCvgJ8Kd
z_lMBsBhH7xa%r54(Tx+iKLVDEIUfbfIj4yeG8y?XhU>5{W9s8M@b3ve!SHOjKlV&r
zf7j}H{3JuWXH^_$wH`d3;j<6JvpN+yjp14KcjGvl+ra5ry|hB#p6FMCuPXF7tCye~
z*BWCFJ_VMKv-)YUT%1*N%7y=BV7WM}mxJZfv-%nII8)kV?q|W;?2omphwXFVH16lo
z<<hhI1@wrkJ?35k)@I!8NcFH?4K|niyf^Y1hFsXc2;R;J`?X*>f1jyCeu?26dHZ-B
z^2-c+Yah+2uAed8&+8f5-D`2|^$hS#hR^;C_j*6%0Sxzg7BcqwMsT{<H&^JlRp_@T
z`rY7r65aFkaG}RpxC!07t|5ByRj_=Vg|C6-;w)H`T=?GtmW#9Sb+BA|7H&n4J<}d@
zzX8@}|6H4T*uDu)<9-WWE<Fq1Mvu7KWA1mr+Kg+D)Wh~&a2odxbh#aj_#WN~j=0)m
z?p<JQ#<kb#Vf!A~HAHXk1>65PAL{yq|M$Vx7C*=C2g^C%pQ%4!IR8F|`PB7^`451t
zA%1565G?2X5y&4goFDt9u20nWV{ql({{&7x_VTA-xrZ3$*$eqIhP*X-p47eLc;<i3
zn9h(lkKdn17}oJ1!#V2uxenijUof=$o`~Z;sRbX%@R`ksyT%-_@5x+btmRi=d3zoG
z`88NR&hc-+a`q$k`M2QMXYDcfcVKPSB#xS__h5$4JVw-X5ZIayLHh3QhkTN;6C-MT
z6s%ub<74RZ(eFQi<sN6)?|qR^Fy!r>G1cS#`bV(e5_xM?k6MR=7cqPmFs$`3<Pilv
z9J!ET-ACrQ{O_1gfz!JGjP9JY?!TbRM{oZMmb16kX$-mO=ik9{vCikfa{Y|x&+}mG
zN%OpbE+6~#BKX<D#~9|4(=T$p1Wt3kj4pQ`!#J-XUuBrfcgq@_yEpPRhBki=c>Z25
zxXAwx^vG{q^ULWM`MK5)$v+HTZU-Xr4@Zyu))o1y;I&2mYOq}7uK`DX<C<Sizw~zs
z_kJnEX9>f-KN@*Vfg6y=GTi&)kn!`TmN@C&?}YB0bWcX;!-&1#87vojZw$HU$4IbT
z?ENUPT)Ow8(XA)VGX_pR);<;-Yd3~#m(#Cu?JJ1ovz+1Dmm!ZYu;=~+hHF0&>DqlS
zCKJcCtLw8{p~v}r1G+KN8h3${k6w%e%h`*ldpvs7t=+oi()CY3uZ%qrPCoiJ2`m@i
zhC|7*E4sWftzX@~&&Req*mY`iE$Z?2aeIIlVRMeQ%Kojw=5rDw`nR&c4ag>j{W}?H
zzdT#hh!g$OXL_MW|Mo;TM%uqAaPrZ=y})v5|E8ix|FpXvIp3Rzy*GMg>^eC4=wCfp
zF8X&A^Y=lQH>UNg+rP!w-UxP`+FXlz^lx9V@0WA5Rrc>xY(A?Q(Z5r`_Uui_sCz%=
f%BMY<0Vf|l*&i&I_T&Kc=!te~l-oh9sOA3vchlCX

literal 11644
zcmai(33yyp6~|wjOejl{rBGyPC$vDJd!dw4wziZ|D6~=pQE*7pX&6Y8m`qB`A}tgY
z6jXK;Q1(T3WM9Nm1k{gB0m1J>1^0agML~bRH}51bBj4vEm-9dWv)psfz3<JNX~WPl
z8zjk)WElUwJ|s!6&6A-pNwRU$l**O`ix*7ltuCFk$Mh*WY??Hre)@a}p8~RxvAng?
zsbLcgo59O%&d{%z$p&JFg4aPdCL1L!3rZ~u7q%>zS8D6-F7;Mdb(BjTmCkZ$X{D#!
zR;{ck>o+XtSLtl)=v$h1DBzKwU!6rIbCz_iJfgfTtx(;oYuU2iay6mmjg#T%`jjf2
zOUo<A^;UcO+Nu+?sEK`3r_Sn|It!lyHE)dUA(y!~0jD9Utw0ZwWKpGDonI|4kNM=h
z2Id={j6e^+!E**nbltpL5EJum2_BfYG1&&)Y7yxg)xDNvT!H>>hn}vnr`)>qgO-Th
z^=*r+%~MEr059sQbXHvqmpd%yHL#w>q!~T+L%Q4QUMp9bbE9M|IJKAZlSk%sq3PUX
z@T<+;l<W*%Tv^`TQEut!>g!G`aDI8Of%yx`IP{!vTBn>>XQ{$&^!9Tcx1`3@)}x>2
z&|Z~$2RKjmz#PMq@#wiZ2CXE=Q|G{Xb|a>?o=uX;;Mm<#wY5FAAh`ZO{M?60{jtN7
z-Se?&I|fbvKPil#)*oL;ri15q_EuXv+sZc5KFWCwtg(>HKo37{P5dY4WzU$k{~XBM
z%YpbUl6}x4zI1XAA+gF}W0l!07BnrJxMJBfZ!&A&n><b>_P%3Vz)i`1$RjKH6H(hQ
zd9Q(5wo2xr_xqM&qQO=5+y8%AX)UwK8FkG8*Xr6dX&E%vyw0WnXBynod<P*T-?88d
zcd*ZuR{7SG(`Re}rVV{@Te-75r?0>Na$ddiNPDI9GDioJ98+t0Lt1OC-iG8{$e8Q=
znqM|oV{&0#+~w#IcU>;-rn<P>2E{cdch_`}VOraR$XcGp<QZ^xYftO))Yh21K1gp!
z-axO_*Oa^kcAwgP3iPkP#$!sVVFTNPpG$Jt0}p6^o*{$uLNW}UyQ1x+lCv4plxzss
zpR=DImVTDCca&N>x|Xzd%<o*$)z(_Abal=@Fq><HaqYD`#5z8VzH-%+w2mFYJzZUV
z+SKYAi=M70(|1B2sIw^<hrZal=5>^pmpjuq?^$`Thz(9UJ`aNPnKp1gG$gy>RPE|b
z>)97P@Y&Fq?1w%ue`7McPM<eOZ%F2+@wI**0Y0HpC2t|=0#|xVQC`Yb<#Nvs_oWXm
zKi|qa{j@=PAvxy*a(rfxPh)c4AiW{Ez?jb3dM_SiYfLU3q!*HF^sTG$+Ce^z$qjY-
zedy`gq0g@A;X3<Eb^0qgy#@az#5FN|Upbin$Njubs@Lv$U^2%4akX%!^l3<OV|uUe
zsDaCT-uum|R@2`E59BYT_mSp&zPHdLpZ85ZpSt-dJI(LAobNu{XZmZI{$cbj`2-z`
zY$lue#9L95^A{Ojqp<kyRpWe<OHy5~i0_7(uDzL#Zj|ZzX*1q;sfgb!v&-r4`=wZ3
zf9KyCS!7UZG7^0xQqG>6&tA(l!?+H0=hNnUPRun5Y@KpjV;hYuGR&z|cg@B%BZ~}k
z+D~=kwe5_IoV$R{DK`psTxN3(^2V8CSEM%m)h8m&;X2iQz-w1`4nNmpjy=KVk{gAu
zpV>8=H8v)DWp;UcI1}kQjMHX5KU*W;{$TUT>3;xHe|i0_XI_pQk8Z!`GrZKzbr{1w
zhkg{enCbeL80Ou9`IjJ#v6eQZy0tGw>LWKR)4d1eB5ym`OaBT|o4La7my{t4?LA1>
z<{8j#9eoV@s(md|yXQsyvP^Hz^s6(yh>ZOAf}K<U2av|QSCRh_@Q|Fo4($CC@lS#M
zjuHB+V9&07c@r6Zc?+!WJU>G^r<{FJ_g<FM{(Gc#cxFAn8&QCHoLBqC==`O;@#^0D
za@vPyIN!&uHBy-Qwgqc*KI?X#QQDK_z|cph`s&#eI73AYMRa}SClKR3Fb3R?9EFdc
zdvfM2ri1$eF6}3B1hegd+?7=*r!%aj1?l|u%-Rn|H%4ClBbjbY^mrNAUhJ0T>dx7&
z0DB&`C-z>jb$OP}^$oDO<tI?7ef1gX8S{KA_hc>h#d^G)=N~Wyq`mgJH7v^$y6+?A
zRo@>uBBy)ie9s8GwfU|Qx@XRJjnK9Gt`WLD^t~W-?*ZQpLiZl^-5_-9o1N3O`)&|+
z`|G<w==Rt5gS_rLLg?00t<!xs2)pz7ZV<Zuz8mCq-wpD*?**Y7?|VV$KBIgu2;F{O
zoYS@YZV-0oyQWU}y&&x7_q`x=&w%d*p&RdeL0<R0Aav*Vy+D0GPJ?^j4xY`4@IKp=
z;l1bC*_ZwI{C_Cp*W#nT8M>U$hKSo7-8gea+!pBatFyQ*!N$rvx93bQ{I>$j#XKXy
za^CBUsKw8KZNT!zdoI<D_v~-Su!e=K&S%2*87J?h=*G8*+yO=V+0>@q43;Z0%<UPM
z6MMdWHaL&Zgq^{1+C1CpGs)v~b{FvONO|qv3vv-V9&GGvI_bSM0j$qD@O*}TMdU;{
zZHw^po}a`p*8XZUxA&3!Ke?YLBfYPcZ<AvxavH-+`|g=t`7ZY9$SDl{j1#+;p07Q@
z=JvC~dff9F=<?yS7udR^W_@O&%il$wJ&=1djFW$ddiOzkIls19jAt0mFE+RL_x@mW
zm%z^NSv>$<-gWzYP>)*Xfa~j>3n#C?&kyyPti)Ugf`=o`qfMWANO||s&*Ou@+acw(
z`;3xv&b9D|faPNSBjFB3m+zuxpZ5zGa>lsVhk>m#)>VW%99@1Poi}D7L(buqiBzs3
zk3NeS=GNEReE!KfkDs-R!E$jgj)prDT|Rt13igcWeU3tx--Iq1b2OO0^jsUSJ>owG
zb`IY++#~&uL6;AokAv&y_ynB%8frD>Sg@R*#WC-3VE&SwvYh%nh`(p*lML(9$1`<2
z(#tcYt;CS?Oo^kGR<QNPbIrMzpv#9(8`ygDK1<Q%qqZ`bzqBXDYmfMLuyZV9L{Clt
z%ZE<|TtCN&aPn)io}2`h8_kG$JHY&<Ju#;~(Uaw1>(a-2J=U-MJ9}y0IvMuI`%Hg(
z<Jma^sZHF4>}FWIF`r^+x6i#uFZ(Rt!;rJjV%O(q&kFEZMm#%wUY?9D?;L$db^U$T
ztO9#agul;`Q_$u0Ux`%r-Zj^$V1LUtk2Za(NO|kl=X9`~w$qU6-q+^-G<Y&n{#H2G
zc?Lt?d%+m{AZIUQjmqfDYOs9?eGS-{o3p;136{Gl%YPQw7<qenHqy&pYFo>YvzOxN
z<!8b5y*w9A-a0>nRF7VM4qV^M&%?><e;!gjdU-zBUYbXnKIb6iBj*>ua`BwAhhIdO
z*T?7Dm%z?j0^5@!@<Md^Mfl7>zX)ucyz^Xul+)kO^-I7bGTX&Sb#2C82G-xWOOfiX
z!S~wB!E=$uXp6eP3^rfB?ysQB$NR{YV7V(85qA|h;+#)iKhHw!kuvuFYOs6nzJ3kq
z<-W>)m9d`TzKWg0d*nKB{k~ogC-0osBGsd&8^QJadJ~+y{x=}iV_$CuyRYWarq4A<
z`N(+-xIX8taPrZguY={hH_f3<E^^!kuFr8hoP7A)0d}qVeYg`{KHd}V0vji9AMZhW
z*+*@6Gvw@}*xq{%?*|WKM7<Ay<)U6~a{7-$KFF|MYugq15aT|EF|JQtANM-uQyTAc
z<q@!b)c0Yedc5~N3iiEEUb}0Si`d7&#yYq6%;R8d$=@^IM3;|u!zaMzsed>87M!*<
z<SLSD9oSg;cwc-H?3_N+v|E$&$-6$!s{LG#jJ<jaT)&^+hLevTKMj_P9&3}!=QOwa
zK1k<y1{r(yEZBXCXYY5wa@OhIe-G^C-fR0VL(V*6^G?ll`N;bm*t}=b6VJx;V7aJM
zn_TSu3t;zN&K%>AFEZq#SL*tE&wd|lZ{i$#9)5r>um4L(^*9eNgFO%OaUSIIeX&34
z(~z!DsXjT=U8lAAe0&w`bIDw<Al04AXW$RPElBzIk|F#qHYEMtzX6=@xnBOZIfGc=
zul-F`ydl!xdYpSxWSoOT!1}~m?dPEk`RMlou(9!7?QpQ1zfHthSqL^xo8RB8XAwjG
zedZKLt@hJrmY2V)TAzL2BE#0YCBs^`LPo8N!TRKDJrZ3$YW*nK*r@eru$;e<M6DkK
z8>h`$t>+kqytRs>R=*!?$MEtu+~}+Sj%|JZULC)s_-|~kJM`8%y(6dlTS#xF$NG*#
zH*bD@pG22mm!0+F!E!!JJRi<^JR{b>3@jIGJpnA&&hS}NL7vDcGn_|z#GVA!W?pMo
z58HBZK5rMgobS((w;LRJwMXoyz}n1f57fg}1v}UJY<(-ha{e}Y8hhjKq$e}vpTtLf
zWya|fc~*hV6EUZN<s#-(uv|XoG<55Uxlaen<>&r1x_sn011y)%vl`tzG4~p<T*RCS
zmNRA%CuAh@EQaf_E_3SRIarH4o8j4Tf9#pM{;t*Ycn(9mXH^_$wHZ8?;k6^fvpNbn
zhT&P=2^nYeTyXxZo?oY5nA0x<U!Li4R?kB>uQkRVd=4xhXZ7=7xj3uNDHr}1faT(>
zegQ0(KdWCvk29q`V!s5|W`C?*J!}_&^LZ~um&>2kOVA^)_K3X{tj)akKs{_%fSt>I
z9*z7mLoV!J0UHzcE5UNU7Z;INF^rM7k5?nV%CNWg(K*%iGpGCcHHLQgS{!>l4m_UW
zwJXED-UYcE!@Zt>jJ>`NoZsu4>h#;{^t*HV1K<a9y65ScOpmj0J-YL{hUmo&VEH%;
zH-hEjELf9V_}>hci?eVGST27SZbdIMoJV`aejTjM{<${wu-y*M=e+}6E`Jv8M321M
zBla$^HuKse^|0Lo&gcCGy4<~tcplydj=b6<_I|K7^V)0mussBJ4bj_&!S>%Z>#wd)
z_&*A^w)h--3@m5-X2{1G#y`R^UR|Gv|0dWP;xqdRu$=M3kl$h$AN!{M1S9Hu5?sIc
z>*3^MFTV|zdy3&aJ0PEC$Xk==N!@oG&-}BDu?%_V@%i%|hIOoC7^AMA>llapE<?Nb
zi8$_)Ch%m2*F;9VYfJ)rpG-l<TAl;T+w17h^I-Wn$1i~8>__bLi{RL2?GgJDSerG8
zqbBR!o#8c=5j9N%Thks$@7-OHuQN7aM2#<l^~=}z3c7ss`-fn;R~hzuXXI-Pd3$G0
z^>}~%5qLa9-dfe8)}i2i7+y0O*19Kh?+ovSoW-#2eN$ZfccwSM`MQ6MZcM)JpP<V}
zZ+{Awi?e19x#;IFz;dz9UxMY@8PT6#fvqP$&#%$tW50d_{(0tO4tpY}U(EGeaDJ}e
zq04nL%=3HX9~jQ{2*VnT9gX}WLz|xip1(h3T+IJx^qAkg&M&85%>Nf~e*V9r%f<7>
z`Txd<`K>GF|9fVO`QJvDi}~LH$Nc7XemVW}-znVtxeTuZ81DW4$k`co@8>Yw`vZ}l
zS<ePvN%}u;^cj-ru}A-aGe>@p{)sLhXW?D2oM$2G{uemv)^6Q$`ThTQ>OVO5dvNk`
zZr%sW#r-iI{2zwAIjvvae$GG+MR%RrT#I`ACe;8=?+>K5`u?>L=QWQJ{W}P3&kjaL
hPa27p&-bJNCm%g&0?XxlvH^PZM7uT0c^;yc{{lh-+SdR8

diff --git a/piet-gpu/shader/kernel3.comp b/piet-gpu/shader/kernel3.comp
index cb344c0..ef3faef 100644
--- a/piet-gpu/shader/kernel3.comp
+++ b/piet-gpu/shader/kernel3.comp
@@ -20,12 +20,26 @@ layout(set = 0, binding = 2) buffer PtclBuf {
     uint[] ptcl;
 };
 
+layout(set = 0, binding = 3) buffer AllocBuf {
+    uint alloc;
+};
+
 #include "scene.h"
 #include "tilegroup.h"
 #include "ptcl.h"
 
 #include "setup.h"
 
+void alloc_cmd(inout CmdRef cmd_ref, inout uint cmd_limit) {
+    if (cmd_ref.offset > cmd_limit) {
+        uint new_cmd = atomicAdd(alloc, PTCL_INITIAL_ALLOC);
+        CmdJump jump = CmdJump(new_cmd);
+        Cmd_Jump_write(cmd_ref, jump);
+        cmd_ref = CmdRef(new_cmd);
+        cmd_limit = new_cmd + PTCL_INITIAL_ALLOC - 2 * Cmd_size;
+    }
+}
+
 void main() {
     uint tile_ix = gl_GlobalInvocationID.y * WIDTH_IN_TILES + gl_GlobalInvocationID.x;
     uint tilegroup_ix = gl_GlobalInvocationID.y * WIDTH_IN_TILEGROUPS
@@ -33,12 +47,17 @@ void main() {
     vec2 xy0 = vec2(gl_GlobalInvocationID.xy) * vec2(TILE_WIDTH_PX, TILE_HEIGHT_PX);
     TileGroupRef tg_ref = TileGroupRef(tilegroup_ix * TILEGROUP_INITIAL_ALLOC);
     CmdRef cmd_ref = CmdRef(tile_ix * PTCL_INITIAL_ALLOC);
+    uint cmd_limit = cmd_ref.offset + PTCL_INITIAL_ALLOC - 2 * Cmd_size;
 
     while (true) {
         uint tg_tag = TileGroup_tag(tg_ref);
         if (tg_tag == TileGroup_End) {
             break;
         }
+        if (tg_tag == TileGroup_Jump) {
+            tg_ref = TileGroupRef(TileGroup_Jump_read(tg_ref).new_ref);
+            continue;
+        }
         // Assume tg_tag is `Instance`, though there will be more cases.
         Instance ins = TileGroup_Instance_read(tg_ref);
         PietItemRef item_ref = PietItemRef(ins.item_ref);
@@ -52,6 +71,7 @@ void main() {
                 && max(center.y - r, xy0.y) < min(center.y + r, xy0.y + float(TILE_HEIGHT_PX)))
             {
                 CmdCircle cmd = CmdCircle(center, r, circle.rgba_color);
+                alloc_cmd(cmd_ref, cmd_limit);
                 Cmd_Circle_write(cmd_ref, cmd);
                 cmd_ref.offset += Cmd_size;
             }
diff --git a/piet-gpu/shader/kernel3.spv b/piet-gpu/shader/kernel3.spv
index 23a7c3efd84eec32a179cc6d5b1047f2ecd6af2e..cd56c48b74ae6db457a49a98e63a7bb323b3dcd5 100644
GIT binary patch
literal 13176
zcmZ{o34m5*xyKL8H-HF;DDEPN5+W#wiipbq>VRUPxR6WZ$IL)O!wk*BQd)rMy;imu
zmX>ANLR~4lX=z!xR<@Xxwr^IpXm)Rz<wDv0{mwbheD8eTj=sGA|NnX3=Xu`cEZ;N?
zojg3thGZl7Z_LmvzxK?AqGZ|b*|5BCUb1}2QN4X_M;&|IOdZB#4Y{8_d*IVZU!kq7
zc64cYErzk^YJ1Wq()6`8@Skb!hg7nWS@V*X=A}!Umz>_xx^7)dZ{McQ_Lk0$uJ)F;
zj-K|`zK-?n`i)5ZI=Wgr``c27Mm*B|ja0;>vSr;JU47@auL9;(s(E#<TGiX$m*G`m
zj-?;H70bOgZqh!Ixt8{pp7v_n)ZV_H{?@)35qU=ctXcE$Ysg09*F&mCqKtJ+M|<DG
zj-J-e_M!k+N!_cKzcCvJulZWTuw2uxXZ6ZzOKW#$cTZ8jzOG?wdskn3qPd<u)wYiQ
zUTYqmjfdA_4lX<`tEk>F4cXqt$m?&+_ASRR?rUFb_13QDRjYkeb|Bn5`nB{`R}ZXC
zsWCQE<T(z&zsNBrJ0u-jEY!;iE%OhGF<~G^V>SuAxU09X+SQt`vURF?)#_}_Cd0!|
zn`^A*)nV-x#pQT)*SI!*R5k@3;|CQPVlq+J@^Ji%c~-Ki;4}NzuEQmdFZZg&uVmBU
z`lwrj+^eg7LtdKaa0XnTLHWzTu|VdYj$e^`Savk{ypGQH=AQ2Uya2fI)xBypjLK%i
z2Yl1b91v@hG|w^k7kT!`<_sDei!!Lv|0hau?v5XbGbWoiD9+1f`)^~MI1po4b~1S3
z+P3u8%jc}_Rhx5Tb}D?p7wp=od8PN5nr9;JG0%@$e%~ecXd`#gY3KtnTQ>A?zT&nR
zkYgpwhsl6GWBsaQW~`bO+etMI*#h(C_hm!25X>!(-k6<H<S9=;QM;N~ZH|>}5u612
z<@1w!)y~$~>}+`H^RkN5BUY<dwJ6gX&tiPkxCC6(Sjm<aahYyWpWLff-|pFRIJsL+
z?`nHl?pP<snCD#jBG0&NMLDi~bkr|&46T)1I1XB3V_q;2vmv_>T+}``yR?k&O#f|R
zwdl14VVrR<PU2oN5O-wO3a)l`cDJ^)t{r%tumDkTfBrtHWL5m|DDVHqY-PziJJxpa
z>dJHQYRoz|iYr+sxTm@ye?NOZd=W3_5VtSrXZ4Hq&tJ~qUUhow%pW>Au4EhNi7|^V
zx4n_RSdWJ68gS&g7F^3!$*wE-Y`T2hjr7I1N_Go)UA3pWHn&x>I|uQG>~46(y$4*2
zJ1n~o?oGei=hwhI3LR6{7dEhSc|eiN^>~QBHkV5FaKU|S<-U*6$J`zVM-5MaYc(`v
zPYw80vR&YGE)6^sYP=yE4v#paz{~ntJ1yJiiri}*n!WXoDK+twWb8a}Ep8<{sg5tI
z<BRL~a=7<5=Z8Fv*?I8!p1FZFEHvxL&th561^9Wneiwn)_|Pc%rSP<-mOB0#xckw<
z0yOfy<~QK#&X(rR?v>Tf#a-*WTdRE?-Cc`)KKhN+SNi9B>!4h3$gV)&xM^ly!&UmM
z&ToO7Ujt9?iEB$<yC*8y4aFF~Mv8gd2v6(&=pf#Z-3ssM>doWt9As<EZqt^0AGnsc
zhn8n3a@;X!TqXMgyfy#!E9`gI*}ts4{DyPyAB6Xm^M6`fzHgQ6nYwY$*6|nW_^vuW
zk{49Wy_r4r9vnvVTX_-x#`9pmoEOgw;J~+0Gqvf{Ky$o@;ZT}a@JR*D`Muy;{KowL
z)jZjMxQurneZ=#AP2<Up$29VIy93gAaSx^O)FPhuQyNb`iC8QH^R1ovbiDUQjQ9Mf
z<K>PIyJy<p2M5ud|7Zfs_owYe7V}KPTYvF*e5`*W&1(V{e}fd7ziP6atL=krH~NXR
zeHk~BUTz%6nMZ9Bwo&wQ$7>r)A2I#CWjwVB%z0dC)8BZG--}*dU&rrDALBg>j#tyy
zxoK0EA58C9xno_Q7<(w#v1$|Wb)B_E-G_nI^mFV~`WSlz*s*Hbrj<5x9toD~qs{Nv
z7&jB_I5mBJd1+IZA4hM_;B&#ok{iqK!iaSW*jQ?=o!@xY?z)9f6IdT@lkk~O?;bg}
ziDt|*=)=DmtiPJJMTu>3X;U-Sne@hTj5huKW{OzLz{XN@Zs*YJukQHs6W&y67bTi8
zCo-;u-Z;i}+-39;N1Nw%75zlo5ZaL?Uk&yOz6R{LgQ+u~E!ONaLvF6_(noDx$=8*d
zajd~-pO^lA+Za<l?3=*Wq5TSaYxg|Mx0T#|3jgikhSGjRY4^-V-8X@a7ydVceI^=j
z2fgv#bLVw0y|p;6`{?ER-A`|wYR*gkK&eICM@p{$qhRBD7UfTtd>XxZ;#p@NpJ{Ua
zM$?-|%{+3SYif~a0el2}Z`7q=WB6RNZeyGe)?dvSa-VH!j1_Q912I>@y%&8Jb%2c(
zYq|!mySnk@J{#4{*I8o6M2+2On-l(e_(Ztz-w4(>>T&J2qA_Crbk3)HdnH`Wb_A6;
z&-c(G{!8G!@eAHau(9yzM7B<!)#mgp*^Vc!`P_Fe=k`2GBfWF-KJc5yJqzBK@NMXR
ztAu?f-0zj(-jjZ>1UH`FD=GI|CAj1LR!O<vD#1NJw<cWwoeB4O?DtCeo6m2R;M(s<
zxc7zMDXHD>l;F<K@08#>;C`nBZ-PIaaOdmyO4!Z+Y{Iqsy%P2&xZf+mwfnsypU*CM
zhSt(NSMJxz?1^WjtJJsQBVPws^L)j)*MS}9b0x<0z|}XG<9fl4RX>H@cdykVMnAYd
z_j)w-$h`rqb~AZm+(vkP?oDXwSC_fp09NxnMm?8<ji=3eJZoy!@A-Wb&A3ZhA@7&X
zrKav>w+>(Ll`W<HD)Py<!PUHnjP1Eqi?z8LY`&P+e}L6|CM+Z7(Zu*qu)5<t=kgf;
zX0YoQ<KF^Si}Aj%-U?TDy!VCNHS?Uj4Xm~r95LSxSC9312UyLRTg!EPCs^HQm2thl
z)L!5|zK-7e+5Yd$VLScxG%xM%F75U&V)t3{E}DLh6I+XOe=pcMo5OolKDVslec;*n
zsB6E0UM=Qv6IkuP@Lx**ewv!^hUf6V8SG^Z+CD&2vj(xXMosDw_e0<`?k#XN?-k>I
z1ngy8Z6Bto8CM*0`xrRp=JQ7G+<Y(X0Q)yN^>fPkejKds6=&lUaCPTzEV*^N-sUrJ
z<ohHz^6do6&F6dZQ{XxD>c+a2UM<ear@?CJIr(q6y0!St`WdjA<8Pyvn_Jsw!D`y>
zpqIx!ehz#fKI+<Ur&o*hz7wqG+J2tiYZrB^%a!ipALnZyIll;woOgre=Jb2yOW;H3
z)y?ViSk3wR{JjV4eEnWAp7-*-aP|2-{q(sH>^Oa5UG%#jP1_b?$sYhaR^41a+ttjq
z1N&FN##sY4m+#=O!qxSEkX|1CUjrK>p0WBr1XqvoUk5wxN^)rb23Q~Ui1SVG6|^+Y
zx8Uldc-A@YVKD#a-y_D*9^=0auCL<}H1(+CJ7DYBQr7WZus-S$=X>D#Ivzz+k8zKI
z)wa^2w(o=aKmVJ*arL>2JWJ_+K(lszJU2h0_wt_9_CuPQXGR>g{21(<;=Sh@`~<EZ
zbNVSbUE80*^-+&FKL<OfG|n&J>QVnM!Tg`EtueI6_{YKZb^HoVJ?i*1I9=P{!1Ym&
zI8T7<>-a62`j&D}eg{_DN{iZl59a@TZH=o>tnDAb)~=7wnpj8szj1C{|0ij#^IbIk
zU1Rs*kMv&dgFa8u)Z7QLbMv106WIC3T>l5`T%DKit^Wmk&%}M__sE~&>dUx4Plf*l
zY#hgG^Im$IUfsC98~+NvrnEgnFV|+gzk#nUZO_un=VI~O_3z+TddFysc+Y{=%w=48
z#CsmBuQub#W8912sN)5&T$}Uw2iUqDw~JmLb^H^&0w2d{i}m>z*m&vsSh#w;*M~R}
z+Dpl}q3{@IKDmDGbL^A7{PXngn|<URhVNDIH1BXU^~gH{tmf~87`GcZ#+g^HU-~Y)
zzRY`FnRh#Vtl#eV#`^g!EB9GXmyP0|<@D<R<W!$P^EZgUW5>|E{JYK^^7>nDG|hSK
zK_9<A{G0vog!_l&ly6A5zmxs@z4n;jYrw`$*QEolem}K%-d;;n^F81im{Tpz&+EWy
zv5vi9HJ{(Ull=YOPg_SbkM<b59<0r{)-Dg*MsOPU^>8(xT@m*U;E1a|#$FEAW?a`m
z9=11u&2>5F*yreGnp)VmfQvDOeJfmT8SCLQ;7ZzNn!4+_js7Z{YwP^XDc8@K?&sAs
z?Ju!^;@E57`QvF`d(qtMJ?Z`3>R#_dAA5Zb*u4vWZ5_WZ;r^DpG2xz{TS^{#|7N&(
z(!GBRTs`*wtzfm-duvdOwR#&^E%yHHV6}Ab-vN(3(jH^q3D)L%ITv}@-UUwMUJqB>
zPK&*NH#p*IkFoCoYcsBEA`jbp!RCrJcpunxihYyo6ZZFm>(}T5XzDTN4}#Thruoh3
z9({=RL7FwV?sD&Y&%%dk-rwrx@xJ^B%{u&MbBtWSsQaT}=Mrc3V_-GM@1XxU&GE5M
za(x^>j{Xxg?VdkzoWEh<18H9S(>#Cs(H}tb{2fFeYq7J`)ARC4H1#+yp8~5{N9@yW
z;MgbaF{e+1wONxmYO>x#XkG`?qNWL8Ynn*!J>Y$EC(ZSa8gB>dm)3X(Ts`*dvtYH@
zud(3I(bQdgW6I;3{4Vf#n!2^hqt?mbDKxLcXx4ft{oyogeFc5ouXls3HTXSs{GmGj
zaKax2KUVUX>lfkXN$2_{xO(izm%(arF04T<_TWLVTCB}iz-m3T*uSrWAE22>dyM@W
zSet9+T;yT<IyjB{4Y=ASTEzV(IO1xLvEKq~Gp=hQ58JoFY25F?)gGZm-0y-TuJ#!F
zJ+L<8y0-GLeIM)`Vr_o_cD<dm{&Ic7|A$~}i~I3MV70LS7;KKP{{*ZS?~tE@9iwi}
zZS+5*ne!%^Ipz8p({uK7ns(2iIL_cS@C=&Qku=ZX5%kk(p24H&<IFt{9-i<g68<Fk
zsglP!{0eRy*CFQiYp{BpvEP8z*3hi`aeB4z|1DT8&e-q3YUvsKJv{bIdyM@9SetRJ
zRUWoKg4Lqdr@_|0nx?;8AN^gUKhd<ie&SfaVc^*`uURzLZzlaQG}rG~`pEI;Qa4A`
z@fWaqtj9B8HP<7~<X^!tKkYI0S+F*15=Tvw!E<O{$I+su<H6Q+0=;|cvt}1<I4x>?
z4y<2V<MVL!*n=0qYVi!51b&gG?%Eks?laI@{z02UQ@2)m)au&Kqj}AxMXe`-t@R{&
zYxPVv3>jSOztHqcYxM`7dek}ut`@aUg${+QTdOhU)@m&ku=j(ywaTN`q3Ea4yynxa
z^%VLAC2pc$NVD$K>Epf`PK<PoN5CDE*1a2=dYpriU^UNybs9r0_I?krTFi4SSS{|0
z9gN!(ZawBvckFKT<IuJFeDVD51y+mmwl{bT&2h#xznXrLe>^zNzYkn(6Jv~H{(a$*
zUp?~ghpsL1?+;ds{0D#|zj4j4reEYg5S->e2(DH<%Zu-tgW=}CDakznO+9iS0#=LM
z6Ty+&nC4c~FLF-;r@0S>t8FJw+((n)=63&L++k?yk?U}9<T8f2)bvY#(|PXBrg@!7
z^V~JlpH<?;^h;=-yI0c3xtl_q;qa*G6>!J6uTj%fH1#-lM}XDh+!;eH{HKA{;@lkx
zRx8e3@pp{raO+9)%s^9*xgQ0Nxf{c|tLaxi_vOU$I)~=mm(ia~bH3-%$NY|Fta_Sv
zCYpN8Zx&e1`9&?qz@rxJ)}rQ|>T6kn=5;<TYPq1qP4pMitm{?u?t{;VdBk!b<W2M^
zl|1U64L3$w_pxZ|QTK6RwY2Wz;Ze7C>sBl7!D27w!0Th5fTkXMI2WuI&&ZRh<wUr;
zF|A+j88{W&$zbQH&AG_qZ#SoaeMUM)TmAZ7ip}dHTCCs2CH5R&LUa9IP49X&(=Q@U
zte-xMOCIZYD%=?9`Zb}c$NJ3&tEKC=5FYEN-TA03phfJ{!1b|DM^lgWI|Hm1>vsmV
zG{e=6Y5j87ZxOaL!Ol~gbCJjTodxz@cZ|0B^=rlEbr~(zuUg_3`js@-uZ=#Q1Ivk%
zuHQuo_g=a*;jV8>$zu(cz^mx#`ksxZ9?ykWg3Y~?_vljaGMYZ>&oR$)>An0LskU=y
zYSt<?S0DBZ!57fX^%Pb6etH#L-E-|dxdQCkIWKKd!^L1Ve@DB{mw?sub<T2qmtuQ0
zxJs|h`O9N}E(7bY&9#un{#3!9VaI5T+$+J6+xV?;HDgpu9=>g*Z=9cYxSI3Y$~rjb
W)nN5iG-J57YX0po*3CZ7$NvE}W$5Gp

literal 9964
zcmaKx2ar|e8HNvRk*0JD2vQWKN)-@6`nrmOfE9IFw&2EPcip{9u^<R)(8QKWF*7Dn
zV~Q!pBpTCaOqoeBGbS-H-SpmLFUj+qbH2TY>r9UO@xR~u)xX_ycdhC_`G_Rxmki*)
zxAjZXXL!;dCP@Y-wdq*5ZqvG1J*DPZ^B2t4;i#l4_0wl4J~fQh$d1Og4h>ZphQZ4X
zM<yWpDpmZ~pYaKl>SRz-x2~aX{rb9fYZ{umx*B>)d)iwX+S@u>8k*a>TbfF3J6rS{
zQ1EN(Xln0mE;`iUQLJA>M@%Z&(Am~e+St+xOuLlx>1=K7X(=UmRWrxjkI|0ReRl8B
zK8U&I_J;13#^zIcO5MFprJ0d=X78LiOYo~oj>NB<QZ-~5W=va4X;oWyQ+rF+fU6|$
zQ?6f=j6g5@+Cy#1>DRq|TVq30XM1ON*1x{4VN**-sinZVp52YjZM{AAd}J~jy`1y#
z#*1xb-5XPtj50^se@${s9$#N-*<ttgF6UG3eMmAE-8%X;lp449txnDn8zs^jWAM*v
z9F?3<jLjD64UOjY56dyWFGo!>5nSKVQ)=vJN>|xF<$TJ0)+CeB!%v%QEa%f^?>5C_
zyu52%jvtatL67*unhdfqQP=V${Ihx1CZ~Wmx3#y_b$9l5;fI>mk@qRqnStKtTU^jS
zvRueI)A7&h3{AY4eQpiW*kRrFnaf$Sy+5@tPi-<6ylO{t@#v=g%KMc2tx4vi_xXZd
zKRKV`8I^NQ<BYoga_PA(obeh?;bQo{oDI9W*`GL8eRTAnU&cOu=FZl+Gh62zc3u`*
zH$B%?$s#aUgf6R-C0QL(Wqqsrl;<}%ITM}a4Qo1@-_S|)V2pdQj4`V-B3Yfs<;S{m
z5)t};+c)RRzMNIbDsa|&RZ<6T+nvqhJbfB>rJu`w?rlKd#NxNC?Cm>iay~usjON~$
zGFgsGz~q=CrEM!?)Vc~BwK~D&TGdG@)8{dyan~}&`W^&#HFh`dNNv^0J%{O4$-U^2
z_dakrZ*B4bx`%bUccyP18ILI`g$-=K4@&a5uMaVn=Te<KoayQD)%P*<=<!K#xyP#H
zDfH;+2jC5*rgqDFuGM|I;2h2MjDW|wuQHZ%RVS}i=w1N(ZlFL7?=#<#+uIxJ+B>&3
zw%2#;>}+Z*wRLvXdv<(>49We|H83dUtCG?1yZ6jaYmU=rd-myN50lWFIIr0`tWIWS
zc0QT29#2It_gIxIL~rZpNpmea%vO^u(N_3u^lY6OT$(Y<x%{xW>SQf?Q~KV^>~$6P
zv$f~%FV}DbdUrmz-P(3Er~T}$h}&18-%-%(nEQHi)FOTlIE#Pd%paTT*}VYveeTy;
z8@&oK-uo~K@d<rmrl$J+;Bx+&^o(kr+~-r3a*@w7S<I(yKH5(6J<V9m7w4;(PcHI#
z){6Pm&DS5XU(b`@U5?N7yL0_G^pSY!H<4`S6OSgw`V$bJ@mTzRlyQFJN>W|!Xlw%+
zCm_cVH;7T)JjPj9ZX&iJjOxZ~8^#zp{Vr%ex$*4sh}@>X`HUaQD6g;ae(#8Q_tJPd
zeVv;&dG&FO#;O}TkuhQ?fsK_LkFV>jE&4tYET^BbzMms@D%e;#>rBgS^6Ik~ja4^x
zE@Q;{;b^R!bN0O#>lZ!?avyCI@i~prwKZ0oIekBc{~2I&%2|IYqyF;7uPEroPrzp_
zqcP@i@6Ss4>^x4-^?Ep;(AR;@JC6EszwE_(qi){wb04`SxxO*y%wxUHh>!jkGMZC9
z>|4P75bX_&_UV~WZ_RbrEc|zZJ;&Ow$?fiY^t~5szVP1<_H#l1LyYDhh&ZqN8STY+
zJ;120-}@QuQ_gv*dxzyB?<2Xc|D#~@dJfF{G-KrT9;@s3Lq>DRnOEJrEa%<o*Ds5U
z`SnNVC*}25_pZyCuR6!Y^*jP@ANJ^T1o{MY^B)b?H|F6wjDaI!#SG4l>pT`+PKg{R
z=))q0exf=Tc@&vufTuEN$C*g@8~~5PE%aN_z56rhz<K%^WpD1kavJsRi!$ls8oUTo
z!|2@C%Hr=jhwgha^j3J^kD*UR_Z=C!_rrH&=;res8M^+yBSUvjeMg3_-S=eZe$M!w
z4BdRbBSY7|yr6qleLoiMz8^!kf8URxA42#282SQq-;tp^AK#OqYu{JUwfmmbzJgWs
ztkfW$AJ=;+Yg)@#oAU?pQ6GRV=Q)bFfnejj%MmvSU4CC4HyCWJ{4&midnFe+js#cM
zJ_=4gY7Yg=?Waz}4MVT2JseK{=DhYuu$<>9`WXc_pEmpPtjO8F=kFNAyz5yZ@5Qk>
zC-39Xg|BzwxZHjd_0*3?m-DWe+jA!uYcn2fy_nYou$=d313A-o30U5E&*?-Y;wOV`
zAmUE~%X#mNUxIlGSl;+ajOs6OUruK9-Z(yw?^MQVh>!N^x!v()?B3Z^5dDl3+lM{Q
z2HTr8yc_C^^EICX_IFl!?cOE1n8RGK+(Y>L*)R{0^X$EV{{pa&J!qSc$k~I~UZW@Z
z=xH%Hdh$N1+moNYr-NrR%G;CoSI)EK{X7FKR}YRpm!iwNu70;z29`74`<?2!Z3TMD
zVOtJXkM&py9!tEuv3?fFMV(b(Ip@BH(dQNVl2?~n%|FiH(K!1#3mi55yim8M@4R}j
ze@l_K=2}KM=j-RmIbi4OJJEce-F4{lE4c0YoC`KipI8_D&V$popIqwe!N$s4%g-t~
zYx!AqKG-~dZ?M)t#*OIm`fp%V5C2VIbHrzg{+rR|BmM%gaR;cO{X(!l@{#8v@M!r$
z-HXxXhj6!zy9CTndZx{xJ>o9~SN5?5PCoj$3~V3!^FA&I>mwg|t^il|(EukOaa+N1
zW02^r5zJ5ecO&!a^Eh?9Guse**T-|y%;@8t(bj~>d1l1XOAFXJ#pk?h(26b}bJ`9r
zuI-iR`p8F~Hn4Ll=6MskeDr@6n4ffQ&7nQw+rgE6?0}PxK03g~we3XLM?Ug&fh+sC
z8cu$Hz9-#ax#HUPp!1Wit$Fo{wJm||T^~PdVjUfy<J`FZy~s9XHKM<3>^|&b^l=~b
z*@?)x4`S!$o!J9+{xR2Uz|Pfq`FV6L*gI1X)~A;7I&}FBX!Fsp2b;%OZQiBbjPk}T
z!RH3>q};ZbQC*wy`@#Afw~tYMF}eM&bt8B+qcPf|uY+Lo75lmgT|VBgH-qI4AQ5*9
zIO43QuAl20>*J`thSBwMjQHEY<B($Ax1h^M-M50}Zbc&Qc5uX5S6x5rOk=c;qx#9Y
zK9w=n?@q9N8+!+%x}UMG(OqEw7AXH3r)d%5cU6B!@&6C__<Qd{Vh1q#{{pNxlrg@i
z)`8=Dn)~kG?TWg8gA3i?fwty)%<pV;^A^|T9CZ07=*4q$E+XgWuk)~`T%4;-V7XYw
z3&3)IZu&0pce9HSe}A)%_K3Y0tj)akt{%2a!Nt6nq09N15qU2MM_%m_dj(jVd0hkb
zur-3MbzPFZOSXaK!rlZnChW~%xecs`cdrF8M&5O7W!#QzMVy~C)%7!{`*|g@6?vWg
z6UScrt{si|j6~dP|3)ziaj%bNjJ>`Jd_+O-tk6pZ-QNkWE$E(~gSj4i-;Qpb;@<B-
zmyf;g0L#VR+k;%JRTo$;_Wo+HTygKa(PNLaM{EyRo9pFV)Wg;bF6P~ZF1Hhjz26Ov
zyxJpn4_KRdT@&@NT?e*StiknQ*D3Z*U7xVu0Ipo4eQ@$I=NnV(`@%C)o$k>Au)ICE
z?&{wA!HhQ{M<Mdo@m}7H*oWVhj8WGw`o0D1T;fc=87yb~A;wz~<71!H^)Y?~<86p`
z&!0HXpZ9w#;&UA0`8$?z4C47ao-x+q_M9)CmpkC(<GkDnma~u8r@O$hPugQncZ0Rr
zlQ?>^-xCm@aY*zu9&Aq&7`+q28Sg>76Vc<_!TJ?@d<VLG?AJTNa<N~-!0$rjU3+t?
z$9Lg(gGVFs_NpGeP6kgwd`?8{brR!Ah`pZ7829&Hu)T(UUxofqh5m3se-!*!uE$*8
zgKnMTT;Gc>AN%n>uw0xAdytDg_yAZg*5-p?x#Io(5c-3Nb+kw9hr!xhGv}fnwvT{|
zc|VFSw*`rF@iB1Z)gG}Q2WvC0YoZ>uPk@VgKZ!2)2oiZe1&+MhBlgo^ZRT}t)x-7~
zuycsDeH`q1J7@jX^$Gvag6%Et<LAI~VgEeX8e#teST5cnUj!Q?Z_QT5FCo_4i&#@#
zKXZD{zKm%142t6n)`DjsKGP7-;8e!xh-YvnW6bM`oOfQ)>sP??ab~^>mh;TSntlx&
zYpOkBzYf-BPvYpwerF>-vykZNRIokGVe}q&SDr+?7t!N4!1@(?{3g15oYQZC<>C%c
z1b-WmckRrn?j5$5?;ukUd3#llUi*U=BR;1g_Bt;eGrR!22(j<e8RPyw1upjeUG#-W
zvG4Dp%g5S&A1vqE+NU|>Vm*HZmWz4*7%aCLaW4)r{sgff>&P2BknyL8HqV@A;%7M*
zd-!woXAtAeYkfKWqW&+y#rn^p%WXlT{x88%Uq0&p3al;a{~9b8^?w77`p<%`FQ;GB
z|1G#!|99wek0728=l^@e`rebM{fC^7+J8isi`stzM{RRjTTZ{I{bz8o_FvHDb|P^$
z{|dIY`xkM41ItIPzk{QeIjkk8U-5gs=k83zXDQ;jJA-i<;#phH7<K=V^X^^L{U=yH
z&ep%ca`qOz{2Ls-Xtx(R&rM}7tKocBAkoXp9D4^=A@;R~(K|4Y@m0h-psvqrxgLH0
z2i+XSzMn&vkG`J=%N6^60X_QGZr^g*Im~{?eG$Df_e*f{aerP0%k4+}ZZ?<k6-3^g
d_OI@m&S!ibah}?oi+cP{``;92bJiB?_djpJ27&+p

diff --git a/piet-gpu/shader/kernel4.comp b/piet-gpu/shader/kernel4.comp
index 6e2392b..cdde198 100644
--- a/piet-gpu/shader/kernel4.comp
+++ b/piet-gpu/shader/kernel4.comp
@@ -44,6 +44,10 @@ void main() {
             vec4 fg_rgba = unpackUnorm4x8(circle.rgba_color);
             // TODO: sRGB
             rgb = mix(rgb, fg_rgba.rgb, alpha * fg_rgba.a);
+            break;
+        case Cmd_Jump:
+            cmd_ref = CmdRef(Cmd_Jump_read(cmd_ref).new_ref);
+            continue;
         }
         cmd_ref.offset += Cmd_size;
     }
diff --git a/piet-gpu/shader/kernel4.spv b/piet-gpu/shader/kernel4.spv
index c2cb755c8ffd098c3a905f3c048bbe5fa95ad2fa..caef463f859ae19392f19ee45eea1b41a0ee71fe 100644
GIT binary patch
literal 6680
zcmZ{o2asK56~}M#c9Tu-y)H2jNYMa+kO*mzB_Wy((V))Y^X2W^<;gZTZ&MtVL=AQ{
z*4VKi8c+f2*o}!THV|W1bYhvo0#PAU{r&E{C%Np5U*^m?|Nkl9`P$t#DGZ)BG)V>}
zP5hfRC`re-WH3yUj7Wy0x@FVWO-pvv+Lm0n>=_1(P6}z9F=LXUNh7)e=_*ybb-WgN
zH!_xd`EjW&cBX{v0y8AZVL(GNJZaffY}vfIWz!YKa&K>OM{Q4MrPx{Rt`ys<eU);p
zy0c<jQ!cLBUGD5}%Lg<Pk?*0Axmj4UuB&ZbwXfV+xw_I0gl$ZD?Vk4b9hF)F)5uzL
zv00p)qg?5(Rr>0=a;~AT)K=}^AvYw=bv=C@+e^iAPiGIYjmb#%DOa=DS1GkE*iq~2
zFV`05yIS18bm>as3&|*J(5C&kS43^>@9IrEw{JPenZ6s6@!%{jU7Og^UD?Hq)*6va
z0%tJ;yU6OF>0u(V^}L2;N<HuN{9?zM`G+Mlz|=0*N*%eyky0<o{DovXG11GMEWUmy
z1KW9O-t1i7ye!X#|9_VIUP#V5omWWC#;)#`Z%7u)mv*IhSp0NwYpvY5w!b~yJF%k|
zPHnBF8Vuu?qsdPx@nv;<e)uo1;|uDzkX(hq6-nz{i;g}@*k}4EBps*Y8j?NOy`{cV
z7uq@Y4Z!=swJHyUV%~AzT1RK`nx4MvTl#wXdpFR8c?aOjoJ)P*8<K<I-Fu4t)$SVp
zru2Q1*V5Uuz0|p(duLC%RIB!Mn~uMbyf2Nf+aJX4-&wbB#pa%;F?V8f?&*AYVdwk)
z8n~Q3ym~)he~Q1590vCx_IeCm>g>JFeGwl=w|5kMPANVCK2y7poCMeRwJ~`dJAHS(
zrSkQaHsTwTF+9)>U8RnS2g~s7$T_O=XxBI{<64+=3EU8*0l5-y+y(EQqlI`mzkJ4f
z3A|^Hz|D1-V$a2Ig?C@S-$U`t43B2+{Cv@0em=RF&tCKMiJPf~O6$+Wu>Le;0$<m`
z=w`Cb*PcWV&Nmlv%*NyQvCjE*O;Rj38Q(DUTx1G)!_i{v%w}(#Cphz)i5ZEOw~n>_
zRz>Y`*w&TPH$L;3HvwC0j6T0HkvA3FJUL@sn?8B*EVOe5ZpOBjcrO0==vbrcKPTnu
zd$usc3*a1q7h=24S@aTj!oECvv331`7$dhb!`=ru>)2!Z5kMQiEc3~Se+BlS9A1r`
z&tDJr?9AVYb{(!&yd}e)LG-;9+j{!9W&ZjZHYDvCc5h;So->j4^Su#mZ!zC%u&pKU
z`o!MnsQ<dmHYe)65qobAzYW_7yg%>6G}iUm-@DMRM?UKNKCEv&`~DEZSP}TcfZGn8
zw_*|Z&vkzUF-G1wZb6$DYxo$pbI4C*=1-%~#jCypiMmIy8LOi1QGlEhF?VA-0^gIO
zy{nw}UTo)Osp;B&hIU?g<6Zath<{g&e*i5P@eg7fFCX<!VJFEA>%`|^4Ps+_SAmCs
z$L8<|uy@`&8{ZZ0tltl@dp;KJ{>!=NV!t7B(dR_469t|DcAiCad?VU#jGW^Ay@<0g
z-?cd0`%^HDXy=%Z-jBFGu{|EhVei4g9QI!OcPHY-{_P3unckYi*1t1{t$$Yz`)>R@
z6#4r7I~3S=a(DnfmcyRq@f<e)L=HcQeKLpb$G=C`U&Ai=PF|1rcHFN^*#po14T$`!
zi4ngk<9tJrcN4aGen%qj&DipLv%I%pn=5}AyYF7hMf_W_<zk+<W6SwATj|B`>N^m5
z^ABX$e9!wmi0^VUEAp+rH{;|TT!S>vGt_6EaqmaudcnTc4`iHkc)ns|_1%of?PpEm
zk0LiCzHj3_f4PYNIJVqoF6u4lPatxN!=9oq?~49MxRZy_-c9vG%zZohlZZqAr_lP<
zkK(@reH&t&ytX~KXP-gbEBo`li7&~{<g?iRCYIOl{ggA;v;G_+cZhhuMW4?&&-r2E
zzle6&gT5~ya`vEYuhEly)cq2+b*=GLwBs?(O<wGtd>Ls%zJgd!9BcXpcC5)aB6dw9
z(BDMnAo8xsHzjATXZ9_`Gu%hK_}dwG6zzT9g8mL--U1}@zMJu{%JRMkHdlTmyEP2`
zeZ=0^BKn=*H!E-LZRj5$*4~O(+q@ql^7bVD5po2PkC-20Tg&;4`3Yj2{7806?x%?J
zFGTd;jogFS$1${i`<RdZIbt8iI)~@|3q;=a+>8DtV!hRfKI<Gu%bVk!{uScg6`S*G
z#F)7AzsWdxhxG;{v98}@+gt3<@375Zk3`JxvE_YtuJb;`Jo#37wU0j__vOSrk^dv&
zoJ$aUwI5?$i+$_&o_g=RXWlpGor3-o68G#xhP`LwvHy(BL*%!>xqp8_<Q?(N`YX1)
z^I1#me7<MrbKW?UzhN7*0I>&a{vDBz@8~}=?tXOS{S!O#oJVY2ejeu&A3{6t?P%w4
zkN$<Y7p}wo{ddNpPILdsIQKH*9?Ce+vz7Pj`9F-9Z@hl*-%Zrj_Xwi@6cTsrQEY1}
z`n%}q0rX>tKG$<H!?E5cz;YKLu}@E8%SX(Av5i@QM2wTjAIZ*Za5k~OS&n900oz*o
zqE-X8wN@ihs}W8<Y7NDf^Y>2FYQnajzNj?}+gj^$wT8pVN3D_Aa{g9|TBES7r_Vik
zoP4>cJ0ANWTz<V1;N;_P>LhGAe>+9qWNh>DdpQM8KJMI9Y;)yf{nN1RVGXx*8G1U{
z-)P=l<Lyz-xF^{7<%qw*M&KWfI7T8D6W4?ug?KN=pyT)0+1NvKc<}(dd;q>|0KPnj
zH)3Cv;aK~*VCT)hxAVaAhqL>z2wQF;;+pJ9F81MkY`NI03$f+g&)A!bu$Lmvqd#&l
z#@1(D*Cr0%CD{48E5UM}Z`6GzcGT4$xtC(=v#xt24&N&5eBCu*xz$M2U5g!c^+)bH
zY<<>suf^fJ0y|%~1uVB7iMm%}M_v7qy8&CDbv-k2_%>m?H!tJ6Fbw@HL@xZ##&(V2
z-;6ES%J~gXbv;MkbKio#8u2c89?mIlMWW_&QvOWM=Yr+T--f;hF@Gy!zSx+^e;&5$
zh<o;YY~$k&iH$LS1o{Pt{wKNr+VNJ#V^2aH6A<6ZIP^rsw=x+WZ{=F-{97sIa65KK
zhNGt!g015oMBhbh`MCQp!j_A^okK3-w`0r2-7jOy<?ntQIL=&u<W{itxqtR54&Qay
za?xuQ+y3Lc#Ksu!9=#aR@BV4W{tdyNfjFii?%!1Obj1CeiH<qCGCsc_-Ei`;A3fM|
z?nm6|UhG(({#e6Hu=UxKcJ$<V&p{lskmzYPwmr>7d!|k3J&0!-J=U;|%lFt1mX9;o
zg)O%e@i&P1yAgT!+M43{z5G(_(TKdgilbNeb{^trMxxiVu<i9}XnP%p-iPFSeL2{;
oe6Ozn%SW#_V9Q0XV^bghn}D`gYwktjH`gn%eGl^XDz?}E0lL_qZU6uP

literal 5704
zcmZ9Od32m*6^B2vbqmGP(y~ZN0c+_71q!qgnzqoiHkt@7M^WZXCzCLCl8lp?LKQW%
zMFj=-qvBTFszngo_k)#MH`KZ-E+|;V9sh7WdVHR5z6&3HbAIRE`@73~-)+9Mx1BL9
zNsdXT^Y^!;HRdF3FiA2i=}Kk)z!d{4CmX{n&tH9x5wnx_G|-sil4(gNvIC42E8{xK
zU<Y^t`SNp8UE7%wvfG&<Ne*CiBr}u#fkOY_VE@2ng;K3nm~8B<mJ8L&c)2iKsh3NQ
z$_@&4(BETmmGM%wIh+sZBqHBKCv&r~WaHTI#!9_ZEni+9!G!HhdG(2rk;!r+f$3ze
zx!5dD&QU6lH_G)^uAFPA7l$j&Nx7~x*Y(s#w-pPeiRuKgok=(Ql&e{&my5$oCL8r;
zsj(v8)r#h-RlUTwC&yZYHeHK*MO50gJ!?B2>v={p51W`mqd1x?OiHU{T5~Qjt(vov
z6SMf%h8*3{<MS3gn%AD3f?m1SHFYGX%NMUn?}7H2*h7s{bwhI`-CJ!(4NmF07OOCf
zUx6flRoLgX>~q6j)3T-0+&(1RoaSDItQG6UvCQ6vY*e@`H0ODH??$VI=S<YM_tz(y
zwJn#n^1PHK_NlcG9Z3`W+MR`FWxPS&^z^Qh*I%93R;+Fr-!V}tHYyY2rl&pZOygVn
zb?8klx9i%2&Y7qBE$EyBdPnkF^n9PUVwciO*y`u)kMp-D@51KFq`CW$#cJ*9VrG97
zIWk)CYNhiW!fx$HXL1-ly>qo<X?uAXe`oR#+?KK8XxRf{cz5I+6?vq+A4!=c>8IZ_
z(7Qkf*o^ng%qQ>YCtl81Z@lN}J#d7*sAZ<MbNEh9&iwusXg@KtA4KNoi~jQS$;Ev3
znx9YG`Fw}Y{|M6hCxN;Ava}%=k!`;E3B)*G4{$8R<8MQY^Oq$_ZMpgQW*~dO0`g`e
zwXL&|y>Xu4%wI%IH&Wg@*7mJNZGZc$D`)KSnNMDOKGIz69&Gn4db4l$tOpzemSy&-
z=#H>^(CzP3=5`OA+de!qZR<V>X^dQNW-rb->zK0yIE-J4G+sXZXQM|S@~eR{;qOIv
z&&^+lbRDiudsAk+chUD|bnEHgn)%)9sCOlL2k;z9Nau|D!|3M9yB=-tUDO-NbaSF!
z1$`=K*U&u}>-(mRbv^dOXCduhKI(gqt#3X1d>LS@4g2L7x9?xNb~*RNbxr|e<eg&|
z(!5y1ZglrZejYQw26+}<<;@`K?nP&;i@LXA?9C7{cVIihzB9F^);aH8=+4^#T-)79
z=ao0!b>9R0o)~`+DHriyK{sAL>feXX$~J0yM}G*s|6$*c&elF^KY;G{$2<BU()!-f
zhmhL#^&6!7FXx_XAIZ4r^I>c!3j42^&a<3~CFCPOPIzB0;4G}^S{&`*pD>+B=kUF4
z1Flcocj)&bZ0~{Ji?BDL`#s3peiOppjlL&mTi<U(_^t2vA#C5ttvOr2--qz~F8n^^
z?R`1>Zgjs9;rFZ#<!t9aoZ0JH^$7YXC<FKBV%FS+yc)>wAx8U!8CO9@-gb2JmVn6f
zBQHOd<@s$kSN;-q+r5>G_z85mnCC_4ay2kSFP^2}LV5E&3vKf~-)n&HaF7-IHu<+C
z<>Vc{PxI1mJi2+ty%@;Vuzi~^$v9&@Pi<rMy%fmZz?!tL&$ub1@t(I_#NU7}H^{}@
zg?t5&6ApWdzPuay|HU1ACDOa7{5x~sh`b3n^uHRZU-@tRHzQvKjFVTl2lwoCz`e3R
z@0a$P>`Y#d?%S5v?|qar*Ry^DklRnZzddivIM4YX#J?Houm^o_0&@1CZm-dkeAIml
zx^=DbHl*V}%ptGsp1c)UM;U8*2YRf<x1jA>W+C4R{2h^ZExsK&b3Lnf1JCYy;<ewC
zaknA8$D5Gv1?Kr3jlB0|{O&C8{n+Npce6_~kRJf{yaDKUe&43NwYMTa2&}ysSlhf0
z0eO4U-VgQy`G~n4-CE9X%!h$-^4;7kxsL$nUk3Ew0qz9$u@9-=K2An{4A@5xa1PJ;
z<3Qf^+=cuEuwEa~XB~e_<jvWI{3P(sYMb*ZU`*WgPiLIG!+LEX*7X^5dyD<~EV}u<
zAYwj;F7JDDou3Eh$q&)1eS87P2iz0+F9PRW3GCH=jCC#ct>63Uee<3vV~@Us?mdfp
z;2s?S^6`82<&3)r8F}}jN1pR&8|OSXA)QC59dG<b?AX(<qTAauvh{uqUCt5Q*U{x%
zTX5gVIQMsm-#^deo4|bI^?M&~p{~Af0sZ%ZIP-6#TT|#C<6I9SzXSBSo<o@(>-{da
z+!Ej#+^6pW`H1;Gx-n;ii1`7!{Jr#J&X0hc|0^8KxF4fiOJCIb3A(jbgQ)dWbor?D
zGjzH8LDc#=y7lx$tzV#9EAE?p{1V7VtzTu_1IVcLYjo@Bb59N<<)ZFy(VHN@-rr%%
z$G7eG=yDH($om7jdHKEkBes0pxj&(sD<A9sGrB#jXD62;{{oHx@2>IoC}-S%Io*qZ
z|NG3so(&w`;6mc2Baa2%)8mlwf6yuD({lFdN7*Zmvd=rpUX!!^zosv<W9<vEoj3p9
zPQ#Wzkllwx=yE;4HQAF~?8B4M<zlbSLYH$tV{ewBF9yz|KXR9&>$9$F(+=NC^nBfO
zu;o18sJjY1>gtc&bJ6u#*FDk>-}&hIx))%}dAFnPh3HXNf8<_-uFtyewRZT{qUY;A
z1zWBcMBR(gqptqQy(HCJbv?6|-FnySvE7^Nc;_>a8-QHoZbWyD;lC7JZiw@niM$M$
zBk#FyLiU3`;CZ+fZR4Wm7IfDaHJ^$uXZ}{?0O$jof%)3TME=vzT}Rxrr=$0Q9$>t-
zF~<8I20<VAU(%N9@m9Lv<^jj?z_&67IT!d=PC&*SmuGzbE?)sBA9wkg=yE+E&g5C>
zaVGj>4bMi`XHV+Ull`6u9P>f+v;f_nPD1*AebZNgX&`!hF1m5~9-oITAMfY+=yF#A
b|Nl1s1wh`tv!-_Z|5ZSr4dm@r+g|?<r0+I@

diff --git a/piet-gpu/shader/ptcl.h b/piet-gpu/shader/ptcl.h
index 583cc10..cc43594 100644
--- a/piet-gpu/shader/ptcl.h
+++ b/piet-gpu/shader/ptcl.h
@@ -28,6 +28,10 @@ struct CmdSolidRef {
     uint offset;
 };
 
+struct CmdJumpRef {
+    uint offset;
+};
+
 struct CmdRef {
     uint offset;
 };
@@ -109,6 +113,16 @@ CmdSolidRef CmdSolid_index(CmdSolidRef ref, uint index) {
     return CmdSolidRef(ref.offset + index * CmdSolid_size);
 }
 
+struct CmdJump {
+    uint new_ref;
+};
+
+#define CmdJump_size 4
+
+CmdJumpRef CmdJump_index(CmdJumpRef ref, uint index) {
+    return CmdJumpRef(ref.offset + index * CmdJump_size);
+}
+
 #define Cmd_End 0
 #define Cmd_Circle 1
 #define Cmd_Line 2
@@ -117,7 +131,8 @@ CmdSolidRef CmdSolid_index(CmdSolidRef ref, uint index) {
 #define Cmd_FillEdge 5
 #define Cmd_DrawFill 6
 #define Cmd_Solid 7
-#define Cmd_Bail 8
+#define Cmd_Jump 8
+#define Cmd_Bail 9
 #define Cmd_size 20
 
 CmdRef Cmd_index(CmdRef ref, uint index) {
@@ -246,6 +261,19 @@ void CmdSolid_write(CmdSolidRef ref, CmdSolid s) {
     ptcl[ix + 0] = s.rgba_color;
 }
 
+CmdJump CmdJump_read(CmdJumpRef ref) {
+    uint ix = ref.offset >> 2;
+    uint raw0 = ptcl[ix + 0];
+    CmdJump s;
+    s.new_ref = raw0;
+    return s;
+}
+
+void CmdJump_write(CmdJumpRef ref, CmdJump s) {
+    uint ix = ref.offset >> 2;
+    ptcl[ix + 0] = s.new_ref;
+}
+
 uint Cmd_tag(CmdRef ref) {
     return ptcl[ref.offset >> 2];
 }
@@ -278,6 +306,10 @@ CmdSolid Cmd_Solid_read(CmdRef ref) {
     return CmdSolid_read(CmdSolidRef(ref.offset + 4));
 }
 
+CmdJump Cmd_Jump_read(CmdRef ref) {
+    return CmdJump_read(CmdJumpRef(ref.offset + 4));
+}
+
 void Cmd_End_write(CmdRef ref) {
     ptcl[ref.offset >> 2] = Cmd_End;
 }
@@ -317,6 +349,11 @@ void Cmd_Solid_write(CmdRef ref, CmdSolid s) {
     CmdSolid_write(CmdSolidRef(ref.offset + 4), s);
 }
 
+void Cmd_Jump_write(CmdRef ref, CmdJump s) {
+    ptcl[ref.offset >> 2] = Cmd_Jump;
+    CmdJump_write(CmdJumpRef(ref.offset + 4), s);
+}
+
 void Cmd_Bail_write(CmdRef ref) {
     ptcl[ref.offset >> 2] = Cmd_Bail;
 }
diff --git a/piet-gpu/shader/setup.h b/piet-gpu/shader/setup.h
index f04462b..9ce2de6 100644
--- a/piet-gpu/shader/setup.h
+++ b/piet-gpu/shader/setup.h
@@ -22,4 +22,4 @@
 #define TILE_WIDTH_PX 16
 #define TILE_HEIGHT_PX 16
 
-#define PTCL_INITIAL_ALLOC 4096
+#define PTCL_INITIAL_ALLOC 1024
diff --git a/piet-gpu/shader/tilegroup.h b/piet-gpu/shader/tilegroup.h
index f1d646f..64b27d3 100644
--- a/piet-gpu/shader/tilegroup.h
+++ b/piet-gpu/shader/tilegroup.h
@@ -4,6 +4,10 @@ struct InstanceRef {
     uint offset;
 };
 
+struct JumpRef {
+    uint offset;
+};
+
 struct TileGroupRef {
     uint offset;
 };
@@ -19,8 +23,19 @@ InstanceRef Instance_index(InstanceRef ref, uint index) {
     return InstanceRef(ref.offset + index * Instance_size);
 }
 
+struct Jump {
+    uint new_ref;
+};
+
+#define Jump_size 4
+
+JumpRef Jump_index(JumpRef ref, uint index) {
+    return JumpRef(ref.offset + index * Jump_size);
+}
+
 #define TileGroup_Instance 0
-#define TileGroup_End 1
+#define TileGroup_Jump 1
+#define TileGroup_End 2
 #define TileGroup_size 16
 
 TileGroupRef TileGroup_index(TileGroupRef ref, uint index) {
@@ -45,6 +60,19 @@ void Instance_write(InstanceRef ref, Instance s) {
     tilegroup[ix + 2] = floatBitsToUint(s.offset.y);
 }
 
+Jump Jump_read(JumpRef ref) {
+    uint ix = ref.offset >> 2;
+    uint raw0 = tilegroup[ix + 0];
+    Jump s;
+    s.new_ref = raw0;
+    return s;
+}
+
+void Jump_write(JumpRef ref, Jump s) {
+    uint ix = ref.offset >> 2;
+    tilegroup[ix + 0] = s.new_ref;
+}
+
 uint TileGroup_tag(TileGroupRef ref) {
     return tilegroup[ref.offset >> 2];
 }
@@ -53,11 +81,20 @@ Instance TileGroup_Instance_read(TileGroupRef ref) {
     return Instance_read(InstanceRef(ref.offset + 4));
 }
 
+Jump TileGroup_Jump_read(TileGroupRef ref) {
+    return Jump_read(JumpRef(ref.offset + 4));
+}
+
 void TileGroup_Instance_write(TileGroupRef ref, Instance s) {
     tilegroup[ref.offset >> 2] = TileGroup_Instance;
     Instance_write(InstanceRef(ref.offset + 4), s);
 }
 
+void TileGroup_Jump_write(TileGroupRef ref, Jump s) {
+    tilegroup[ref.offset >> 2] = TileGroup_Jump;
+    Jump_write(JumpRef(ref.offset + 4), s);
+}
+
 void TileGroup_End_write(TileGroupRef ref) {
     tilegroup[ref.offset >> 2] = TileGroup_End;
 }
diff --git a/piet-gpu/src/main.rs b/piet-gpu/src/main.rs
index 6a243e9..703e156 100644
--- a/piet-gpu/src/main.rs
+++ b/piet-gpu/src/main.rs
@@ -20,7 +20,15 @@ const HEIGHT: usize = 1536;
 const TILE_W: usize = 16;
 const TILE_H: usize = 16;
 
-const N_CIRCLES: usize = 3000;
+const WIDTH_IN_TILEGROUPS: usize = 4;
+const HEIGHT_IN_TILEGROUPS: usize = 96;
+const TILEGROUP_INITIAL_ALLOC: usize = 1024;
+
+const WIDTH_IN_TILES: usize = 124;
+const HEIGHT_IN_TILES: usize = 96;
+const PTCL_INITIAL_ALLOC: usize = 1024;
+
+const N_CIRCLES: usize = 10_000;
 
 fn render_scene(rc: &mut impl RenderContext) {
     let mut rng = rand::thread_rng();
@@ -71,8 +79,7 @@ fn main() {
             .create_buffer(std::mem::size_of_val(&scene[..]) as u64, dev)
             .unwrap();
         device.write_buffer(&scene_buf, &scene).unwrap();
-        // These should only be on the host if we're going to examine them from Rust.
-        let tilegroup_buf = device.create_buffer(384 * 1024, dev).unwrap();
+        let tilegroup_buf = device.create_buffer(4 * 1024 * 1024, dev).unwrap();
         let ptcl_buf = device.create_buffer(12 * 1024 * 4096, dev).unwrap();
         let image_buf = device
             .create_buffer((WIDTH * HEIGHT * 4) as u64, host)
@@ -81,16 +88,34 @@ fn main() {
             .create_buffer((WIDTH * HEIGHT * 4) as u64, dev)
             .unwrap();
 
+        let k1_alloc_buf_host = device.create_buffer(4, host).unwrap();
+        let k1_alloc_buf_dev = device.create_buffer(4, dev).unwrap();
+        let k1_alloc_start = WIDTH_IN_TILEGROUPS * HEIGHT_IN_TILEGROUPS * TILEGROUP_INITIAL_ALLOC;
+        device
+            .write_buffer(&k1_alloc_buf_host, &[k1_alloc_start as u32])
+            .unwrap();
         let k1_code = include_bytes!("../shader/kernel1.spv");
-        let k1_pipeline = device.create_simple_compute_pipeline(k1_code, 2).unwrap();
+        let k1_pipeline = device.create_simple_compute_pipeline(k1_code, 3).unwrap();
         let k1_ds = device
-            .create_descriptor_set(&k1_pipeline, &[&scene_dev, &tilegroup_buf])
+            .create_descriptor_set(
+                &k1_pipeline,
+                &[&scene_dev, &tilegroup_buf, &k1_alloc_buf_dev],
+            )
             .unwrap();
 
+        let k3_alloc_buf_host = device.create_buffer(4, host).unwrap();
+        let k3_alloc_buf_dev = device.create_buffer(4, dev).unwrap();
+        let k3_alloc_start = WIDTH_IN_TILES * HEIGHT_IN_TILES * PTCL_INITIAL_ALLOC;
+        device
+            .write_buffer(&k3_alloc_buf_host, &[k3_alloc_start as u32])
+            .unwrap();
         let k3_code = include_bytes!("../shader/kernel3.spv");
-        let k3_pipeline = device.create_simple_compute_pipeline(k3_code, 3).unwrap();
+        let k3_pipeline = device.create_simple_compute_pipeline(k3_code, 4).unwrap();
         let k3_ds = device
-            .create_descriptor_set(&k3_pipeline, &[&scene_dev, &tilegroup_buf, &ptcl_buf])
+            .create_descriptor_set(
+                &k3_pipeline,
+                &[&scene_dev, &tilegroup_buf, &ptcl_buf, &k3_alloc_buf_dev],
+            )
             .unwrap();
 
         let k4_code = include_bytes!("../shader/kernel4.spv");
@@ -102,6 +127,8 @@ fn main() {
         let mut cmd_buf = device.create_cmd_buf().unwrap();
         cmd_buf.begin();
         cmd_buf.copy_buffer(&scene_buf, &scene_dev);
+        cmd_buf.copy_buffer(&k1_alloc_buf_host, &k1_alloc_buf_dev);
+        cmd_buf.copy_buffer(&k3_alloc_buf_host, &k3_alloc_buf_dev);
         cmd_buf.clear_buffer(&tilegroup_buf);
         cmd_buf.clear_buffer(&ptcl_buf);
         cmd_buf.memory_barrier();
diff --git a/piet-gpu/src/render_ctx.rs b/piet-gpu/src/render_ctx.rs
index 4e9a567..eb67132 100644
--- a/piet-gpu/src/render_ctx.rs
+++ b/piet-gpu/src/render_ctx.rs
@@ -238,7 +238,9 @@ fn flatten_shape(
                 let scene_pt = to_scene_point(p);
                 start_pt = Some(clone_scene_pt(&scene_pt));
                 if !points.is_empty() {
-                    points.push(scene::Point { xy: [std::f32::NAN, std::f32::NAN ]});
+                    points.push(scene::Point {
+                        xy: [std::f32::NAN, std::f32::NAN],
+                    });
                 }
                 last_pt = Some(clone_scene_pt(&scene_pt));
                 points.push(scene_pt);
@@ -350,7 +352,5 @@ fn to_scene_point(point: Point) -> scene::Point {
 
 // TODO: allow #[derive(Clone)] in piet-gpu-derive.
 fn clone_scene_pt(p: &scene::Point) -> scene::Point {
-    scene::Point {
-        xy: p.xy
-    }
+    scene::Point { xy: p.xy }
 }