From 59e3ed5f879471b79db619ddc7c45b72d18fe4ad Mon Sep 17 00:00:00 2001 From: Albin Chaboissier Date: Sun, 13 Sep 2026 15:48:27 +0200 Subject: [PATCH] Faster coalescing --- shaders/voxel.slang | 67 +++++++++++++++++++++++++++++++++++----- shaders/voxel.spv | Bin 11672 -> 13568 bytes src/main.rs | 34 +++++++++++++------- src/voxel_cache.rs | 35 ++++++++++++++++++++- src/voxel_cache/data.rs | 5 ++- 5 files changed, 119 insertions(+), 22 deletions(-) diff --git a/shaders/voxel.slang b/shaders/voxel.slang index 3ea6db7..6d4c36c 100644 --- a/shaders/voxel.slang +++ b/shaders/voxel.slang @@ -80,7 +80,6 @@ VertexOutput chunk( struct StructurePointer { uint32_t value; - bool subdivided() { return (this.value & 0x80000000) != 0; @@ -129,6 +128,8 @@ struct ByteColor struct StructurePoolElement { + uint32_t occupancy_low; + uint32_t occupancy_high; StructurePointer pointers[64]; } @@ -168,14 +169,23 @@ uint32_t get_children_index(float3 position, uint32_t scale_exp) return cell_position.x + cell_position.y * 4 + cell_position.z * 4 * 4; } +uint64_t get_child_mask(uint32_t low, uint32_t high) +{ + return ((uint64_t)high << 32) | (uint64_t)low; +} + float3 floor_scale(float3 position, uint32_t scale_exp) { uint32_t mask = ~0u << scale_exp; return asfloat(asuint(position) & mask); } -float4 ray_march(float3 ray_direction, float3 ray_origin, uint32_t root_id, float dist_offset) +float4 ray_march(float3 ray_direction, float3 ray_origin, uint32_t root_id, float dist_offset, out float3 hit_pos) { + float fov_deg = 100. / 1920.; + float fov_rad = (float.getPi() * fov_deg) / 180.; + float cone_factor = tan(fov_rad / 2.) * 2; // Horizontal size of pixel + let st_pointer = structure_table_pointer[root_id]; if(!st_pointer.subdivided()) { @@ -216,7 +226,21 @@ float4 ray_march(float3 ray_direction, float3 ray_origin, uint32_t root_id, floa child_index = get_children_index(pos, scale_exp); current_node = structure_pool[current_node_index].pointers[child_index]; - while(current_node.subdivided() && current_node.pointer_valid()) + // Scale computations + //return float4(cone_size * 1000.); + let cone_size = (length(ray_origin - pos) + dist_offset) * cone_factor; + let exponent = + select( + cone_size == 0., + 0, + 23 - (127 - (asuint(cone_size) >> 23)) + ); + + while( + current_node.subdivided() && + current_node.pointer_valid() && + scale_exp - 2 > exponent + ) { scale_exp -= 2; current_node_index = current_node.pointer(); @@ -235,13 +259,22 @@ float4 ray_march(float3 ray_direction, float3 ray_origin, uint32_t root_id, floa if(color_pool[current_node_index].colors[child_index].byte_a != 0) { + hit_pos = pos - float3(1.); return color_pool[current_node_index].colors[child_index].float_color; } + //uint64_t occupancy = get_child_mask(structure_pool[current_node_index].occupancy_low, structure_pool[current_node_index].occupancy_high); + uint64_t occupancy = get_child_mask(structure_pool[current_node_index].occupancy_low, structure_pool[current_node_index].occupancy_high); + uint32_t adv_scale_exp = scale_exp; + if(((occupancy >> (child_index & 0b101010)) & 0x00330033) == 0) + { + adv_scale_exp ++; + } + // Perform dda // Compute correct exponent, and shift it into the exponent part of floatt - let child_scale : float = asfloat((scale_exp - 23 + 127) << 23); - let child_pos : float3 = floor_scale(pos, scale_exp); + let child_scale : float = asfloat((adv_scale_exp - 23 + 127) << 23); + let child_pos : float3 = floor_scale(pos, adv_scale_exp); let child_far : float3 = child_pos + select(ray_direction > 0., float3(child_scale), float3(0.)); // Intersection t let inter_ts : float3 = (child_far - ray_origin) / ray_direction; @@ -250,7 +283,7 @@ float4 ray_march(float3 ray_direction, float3 ray_origin, uint32_t root_id, floa // Perform dda step let neighbor_min : float3 = select(float3(inter_t) == inter_ts, child_pos + copysign(child_scale, ray_direction), child_pos); - let neighbor_max : float3 = asfloat(asint(neighbor_min) + ((1 << scale_exp) - 1)); + let neighbor_max : float3 = asfloat(asint(neighbor_min) + ((1 << adv_scale_exp) - 1)); let previous_pos : float3 = pos; pos = clamp(ray_origin + ray_direction * inter_t, neighbor_min, neighbor_max); @@ -276,17 +309,35 @@ float4 ray_march(float3 ray_direction, float3 ray_origin, uint32_t root_id, floa } + hit_pos = pos - float3(1.); return float4(1., 0., 1., 1.); } +struct FragmentOutput +{ + float depth : SV_Depth; + float4 color : SV_Target<0>; +} + +//[earlydepthstencil] [shader("fragment")] -float4 fragment(VertexOutput vertex_out) : SV_Target<0> +FragmentOutput fragment(VertexOutput vertex_out) { let ray_direction = normalize(vertex_out.world_position - vertex_out.cam_position); let intersection_t = box_intersect(vertex_out.cam_position, ray_direction, vertex_out.chunk_position, vertex_out.chunk_position + float3(1.)); let local_ray_origin = max(intersection_t.x, 0.) * ray_direction + vertex_out.cam_position - vertex_out.chunk_position; // Figure out intersection - return ray_march(ray_direction, local_ray_origin, vertex_out.structure_id, 0.); + var hit_pos : float3; + let color = ray_march(ray_direction, local_ray_origin, vertex_out.structure_id, max(0., intersection_t.x), hit_pos); + let world_hit_pos = hit_pos + vertex_out.chunk_position; + let clip = mul(constants.view_proj, float4(world_hit_pos, 1.)); + let depth = clip.z / clip.w; + + var frag_out : FragmentOutput; + frag_out.depth = depth; + frag_out.color = color; + + return frag_out; } float2 box_intersect(float3 origin, float3 ray_direction, float3 box_min, float3 box_max) diff --git a/shaders/voxel.spv b/shaders/voxel.spv index f51f0ce50f22a765ddb8e73fcb5f7eb877775363..39858e8492d0924c4d2b91346ef84d4300e4f163 100644 GIT binary patch literal 13568 zcmZ{q2b`A05r;oG?&wXbVmS*6f<&cQkaDO26!ZjZ5pa0Cx&hE~;-wULy>w(=%?rLLvkN$cr_HNX1Inv-YEt7&O%YOX!Mu4ZA|k|lM`9p<8j+Lji__AXVF95Zvy z{5j)VS{IL-F!6Btex)*1;p>m(<8+IbwlywC>+Ob=BNa$@qzBRy>4jKN|I(l$rVn@t zRjjCMYy}o$`=ZU-9~poQMD{=iA%l@2$WUY$G8`GeDDyZ^VC^#g=UlF#oJyR#sIg{c zQ%ikoeN&@+drbXeai7wj#MI2KZEdb!J3npDl7^<*)(LAT)EwW`Ft>JDQ!`sJVLUrg z!3g7h7T31cs>@mFUg+w!mZirxHMX?YHnz5~we2GuQ&IShU0q+dre~?*ih#?9j$LK8{^ZoXi?kB+QvoeY8sl>V297%-|1UgfBw?;_oWivcOdOkaW}co z`@_BcK35d?#?HN3+0@j~zBdOjrt>;?u1lX!zODnY=e6%~`!T*{T@@u`e9~ET)OQfv z?&|YSd%Lq62|r_9YaMq@lUH(nZ|L(`h`(l$0j{ZM>0r3dYwE~QEY34C#%Jf+dXx^q zF6OCbd)~6H^1|=XXr%qKS;3C=Nc za}ek9_s#izw7qxp+N@(|ZDmDW{-zo0@031sYWv%$U6yR`zC4%rT%Joia(Tz)xxCXN zm$k?J6l-)o_tW}ZgZ7wCc3E+M#U6Q=`Q+oh!+iH#v*Wc7Kzy{vqVxQRb+XG!-f@v% z?33qrZRR(J`Ly%=iD-pMdtRk&pNj z!QK(ZzY}e|obfZ!#>)?7?pX!zcXqeZY_N7261h$S+vDas8EuT5xx6RMCGYsTXve#L z*K!KF67ezqRCE`lD%sV^HmCm6lI=L>SIzEv(Ey% zzOnrBxMyc0=5U;`YGcIKbPn1zT65%l7dWrs-Pqof;eQX0a$yDl5JlJww@jU%5 z@m%Drb3e44eOZY%mm+VierH&FCE^;7NZ7IagWK(HrT2l2lXKjW$&NhZlkHsAS%tnF z8HOB6F8gQ;5^J~u+!e9r321AInhyZWS;K*7ZRebbwgx$Cos_U~2Z6PXdq3JZIqjp8 z9eIvUcFZvn+xe~K7_@fOaxhrVISxTbj+1B81-j(R-h)?Y6R&Zsqt@|p(cXAA8Z2Z;8_~MIhCwC28JaGosLeSeg?6d2@ zo{@bP@67dJ`!OWOeGraempJYQa21mG=Z)Ayk%-?0?v>d$fp;X^u{R?F5ZC`9w7F&> zld(@l`!01!b~V~uVV?%>o!QI4j(3dXo6y}7r@azfp4qFw6`8Hy`_CK^-y^e~e+ZJ# z?^?swZ_nrU@XTHf?u*zH=J)REm)Vzq`)9WA-GI!tw+Cjn{k2DCyMKd_sNeMuPWEV6 z&-fOk3wSpA)`Y!-ZbLjX_x^U|p_3brwHLgDwj-M}3H}baHJJqeVL1CF#IwB<+jAL> zXy27==Z<+l0`|U*9=sb{E+l&D9ys?U?0do67#Q|_*h`b&_1uqaSigSqBu;-CxGUl_ z9eD-o0kr3F0{TJ39?+hN{%B^;0e>vnQ}9=TA3|K$h1kB24<}ClZ1BesdsbWjClGt~ zGHTI&1aZtfWJ}^6&A3g8`y`yP^O32XpLL#%RN+4d{V7Dwu~FNn6CbsG23u}@%K2Ha zdF9RNTF*ry=VM?w<09wh!1Bhs7w)t1_2|zda>m5^zmRximV>{D*q^5(a`wfS5c#!< zvo{_`(*KMul58uzg?xpLu zX6Fgt&%y4g`(W+P7rtMB-COs=H8`)n2J|lx=QOuDjNQOl`cD1|u`hi$tKsaeUnBAh zkgF5-8?d~&?3G6Jvxw_D1Bo90E%;f)ccL5GapH2cImH!d-+gg+v}43Q(AM!LyLvPd zcT^uZpXrEq)bG&VNv-JLBj$+t{{SA6>{HNxM205Y_5KMNmf3#>4@WMgLj8Y1%wb#~ zDwTT<@tjXY^vB*k4|aSr6661xILBX}a{di0Z=Q(zJJ|7g+&{48jqAgzoco`MbB{#K zV@&kmzrgx^g8Mf(#sv4D#MzhTorV4{V%|}RV~h#k3*hiMhq2*%G4mOB9GrX-;#$Oi#Rq_IVw=w$SZ8@H)7P^Enj#2GSQf9r3*F z&#~w?5q*xcKkSG9A@a`YtZ+@wV^FU-f`B@#lm=Av!lUX5jlHN{05|0F7lRv zoogKudAq^MN8aL}n^RtW6^NX<#IDbEbO-lC$NHY<46WV26V>OPZ%%d8VxBmIp5RyT zU5NOvi}qD7IR4|if!}O^zPl0ML*J>3kl3@{`13vMgDrP4G4AyWbYE=8S%*2ajf*w( z1NX=08ru2t(t2=zc<+!ik(h6Qv50vbr~N8>G7#*seq0Udi_oXFnogTtz*5V_T2<;nX|(fE~Y)7<*@5u#df? zPkTSaacb9PUTuATpKKz&5^Rmt^nmvMVAm7!1oO!s0B5Xzdr$5_u>5wgcYtd;2wUDK zxRHsIj&G}jG34jNtxsbQ!FIg97&{7VKJSzmI~rU5o^0%)*pAm{FC30<40r(AUhw?v zp(DX#!TRh4x$)@3u;q=F^L-zOEpKi)-;sBK;NxV9VbQ_b8lq#YAj*@2vQ?oRm1v$h|)bTkdjj?CH_i^7q2s zM=i!3gKa+7px@p&7Hw~+qpv4}UAz5q95Mw7-&C-Bs&5*S`=(>tKl-Y6@y)hK;`&frDry|ve^Qz5fjJCdw=?qT;yMI?O z_QABqdD!xvTRzV5QOA6+b(q_Db6UsgXdmk^M%%r1UbQi2pd-dTI}x;W+ zF*xob*XVuXT~r6wm)}K8u;uf+=zMH>bH`n@6fAFl#$Dt)B_F?s_2BqDbYAyaZUAC0 z`QCcYYVEiOJb!I_L%%%{XTA(P1IfQRmSfAuyAtEn(W4Dud(_;%nq25wntf{!VScmXF`4jbOR^ z@kO1NVmt06xX85$TRzXV8CyPbT_#5IT<^m+*Uee3%dzG2TwAc^Bi9vRxjfhVvCZZF z#dd5y6*MxfnS_cz|}tHAMoM{nA5SHtVG4`RQs!IqC6 zxHiR{k;PmGC%=mQzY_j>u-udMg!Tu)a$CU>a|5=ypGq+|f{l}p^VkNq|E^-3=Vq>( zu#MGcoZQV|`K!V4=ZX(u%U=U$&)ov{vFG&Nirj{{H){9F7;Sy=O>#SUI^sK*#~2&m z7u!>ed5q7;-+`_FKE}Hrz9S#TmiODo?<0HmPHg#zxeI(Pykp}x<|APF5ZCW-%-wL6 zh<@|UhPwwW{}?gXVc!dW4qbzo&zSq*oHxX8F8TZ6jMslX_6BmuKY(ps{qc@I2$t`O z#Gmay3YK?1e~(<>$FSvnf_o@&(!o8PxctxMABQt$B;puj;?KyR$b8OWY^?7Qu>KIg zPu+`0;q+Ng?CmGP?te(s^C>vj;S=AopN5lgL@)zQJ-h_6*$L7 zk3I=@Y}ikM9T)rkG}tl5>5IO325jHxj~@Ff*mDSZQgYvP@KLE@71oy+lNqZl9Za=~>R$s*Z80=WjDB^yCZLB{3 z{N~u7g88rb|1jux-{gM=zA^dZ`}gN?@>jseo%aiD?>l{)(%4^u<&AT_Rp?(~%RfSl zf4=xNSkB)M@7&*jeY|t^J&VYB$BN_cW#r;)e+%|rG>`LG_fW8W)ciZJoVCQ-eviEk zTc5vGE5LsM`^~N|e$NJhejg0BKDGD#z+{g=d*9nXe+1u(-Vc2-a#8Zl zh5Hluw&dHI`P`pBV{b{mThNYkpZNR_F2&Xt{r)$wWBvO__is7+@7Vr5RD2iS3H}G1yx(%h-G%-qwtRdS zI@iCj;9g0b zba1aG&i5&}*AnNtgL@s$T780hBXQEfy_vW?-~SSq=i342x_n|R$5?RE!F5Sop08`- z@_c1*?txE??UuMaUwPtC9dQ+jGhcj5c87BhUAMKGKYTsF?xpLuX6FfCPq2IHK3Kc+ zg|8Rby>&lagY)WZAWd(ubDG;6#>T(9^a0zO3y6zfRKd?Oib#VO?=i2-> Yu{Q?58LQ7fUk*x~cre&^B>t`C|AUMDSpWb4 literal 11672 zcmZ{p2Y^)76@~}4(0j2X78De*Ar?dyQNaji1zU_dE-=XAvV*&`G&O=??_HBvQf(= zBTEgc!Z!-bXCShCRZr(-XuXX{6Vi+fLIxv4kfF#hWK?Nv5i=aTk}lR%I%~ipcLdt1 zqma?a7-TFm4jGS3Kqex)A*0E#hGR7FKWjVBCPsFwqNnoOYIl3Bz1k^Xk7-{a9$wm= znDXMbT37pqrD-`AZ|u7E%KGx!uIg&Awal~EHHnEM3>e)@plat+kD`axv!_>#JQIE3%mz?~d*&zPr}dv%J>RRVmlk zW>~SNJFlHO-m#{`i~Ze2tf!)1AB}y^OA{x*Z_xEqi8haA6G*+Tm)8fr2w$nY+@3XG0qo;yyVwLtnFF9{V){2vh_o;q1 zJj=a}*4mbJR9FkPs?t^D?!C+0p~XAUKewxLc@IY-y|eZ4_t|BQ=Head>F#~Y2`$dS zBy@RRS6AD{bXdzf`cJc-HPcw7Co)(m~k8IIS$t z|IMqZ@S9T%#jITOgRyr$r>5d=En+#D_4OW7+_nAR=lq`bKI(sWFW^S^j`t2e1kv^m zj`O3fPuqJ>dn(#H$$NewHRFDD4}2DI>v>OlC+Zjbgg-dz$+aNCnIpFdaV&q69N$OV zyEE_2K6-m=DEjhu$XI`0^jTBe-!tuoWP5+*wY;nHTG~;|dn&Kxy%e?VJ?@y8qvK5@ z#{Qgxw(Hw#HxzeJtdaMTPoD2RisvAytAU7*e9>&Z@=|ozS?=c8Q9UUYZhxCHRR&{3C_L_PkA0dZ9izN zQ@0yRlY!Bf*eB2ToZ6R<`PzA3M`HKZYb+hr&ptZY&dVO7R*WGR^@4L8FXO!KF@5ZY zqQ~ev@_UgBwa~XhyozG=9e7;sE_Os_CJI>ZP>^MK;*!`2R_hY5zqQw&! zvjsgFam+)LJrC?Sdt;xBZb9TGW6wu>XUIqVDPZq6<4;AmAacf^hBjV)B4aNoc)xKQ zIVaAuzIRdUbZqY+YrO<*jGVR3KwC@R{QCV@f0vxg;v$ZJ6uPl=CVBwUlI)iz+nUyE zO}2TCuRaTL?qM$hyPnbCQm{|hXM-Ih>~p{aGy7a{12Uao5ZCNH#2V%qt2Rb#Pv@hZ zV+#^BUk1*5xB%PxF8nVCn=_r;(sjNNv99A-*ZS5l&)9iL^bmaXV~qaD6PsfVbHwI2 zwtIIG;xnC9b6qb+63`f`3*#_NmtbKh2>-2>OavpW^-`E;!&aRPilng;bvJ&#a|=1gD-nC> zFHz?zeA~BgpX>OCf$hcd_d#nr{wvUqFQTCv2$m}iPKhxlJR(lcn8pQrQ zo6)Ct;8twwxYto@TV}U_uSL}Er{5Rug`9m(Ld%KwEp)v{zbEWj&bb|tuyK>Y+Ro)V zv~iP@?RSExdVG567hF}2WIwN;Gvm)H+XxpThR9)qmkLvKNIcQ9*}H*bIlFA6+A4n+rjo? zj`^MF#>8oRZ#8B1T5xk_>-RpgM#K-!Y{wsuNN%r}0ns^j>5DcoF(O#Qoihz8~3^N$?NAt;;0% z2jOnZ#B=saY}a}!qW!95JC^&u0R3vjJ1x%DLy41D;+#ATcCKMR0$!DD>pY5V*}Qr7 zEOziDuxH9=F7kJ**Pva~`RLao?xXf;=+|ZT8Q|9^dk+2<@EZ{4aV56z#v2o-e0B>FqUO_JIpd<{+raY1yB3!s#=G8cN92r&`9G6*V^)KoMH-N$ zh@9u(9f{>d1dv=`geIERN#C5QD#|z&Vz^<+9;T#-S-(~18B93WoYZ$wQyFIr4+i z_VFK9^*AK%sNrxvaYua}?48t&{sv-=82_8#@yT9{{uVMJ+0OUd$i&S44tO_YD;?_p zE@BPihSRCs_Yn8_R78KQ-S@%fUyeln4-#ko4XNf2!SdFLxF3PPkK}Pb#+Em3IJ0u> zpCFDs1+k7XaRz@1*6$PC&%lur+|Lu|xwP&A^e+(WPDRWyCVamHhtDyL4d1UapK%u< zx21Xh8f@>z#2WnuEblyH?!N`gyLNwL-+zbp`4@OS`uB*xMe2F%!8G(AkP*mI#69&4 z9*X`WqR%|ffP4NYME+yMX#Y8J>(Eca)zE)Qym|KUSG3PRsiE(0h@5?h|ILbs<)ZH2 z!H%^7iMsy)%SYXRf_<#3?_UM3*A+WI=kaf3Bog!cCHrLW|4F`0XltsY7i-5l{}=4K zUV->u$UHmS5%`6Y1i{j)c=`T8PvAF%bjHzIcuw*13c?!MUO>+>uej&Cw}G}^P^{&|Lu1n&pd zCogvtdVg&B;C$Z?z?QeRah~A=!Se1~oTDk&@~-hgXxCUBXXjwB>*=~rMW!L)I|OWf zeTO2sZ#uSp>pQHUZw9vG>N`C7JUd4uOnXH?M`GKLI_`|4z~*}wJ;l7di;l*YcV3>S znP4B!lfGk+S%|gN&ea%geNp>Zu(iF9qV{pv^48At%y%6e|9J5AXlom9P5U?j?PDLt zXgf#8RU0!K9WnNLA~FXt#`tTPqi55#cQ5V)N58JC{d&j5-vB3J8*5zT&c(h9TVLEo zEnx2=ecmD3){A#(9@sh|@h+VV=ewjo-lh3q`TSiv1zX-H-iK2YCmrv@X&CbPyR-n? ze0`C-5Ny5tT{<0GK7W@yhvw^xcj+bIc$b`Gyh~?*^~HPRyR-;fJ~-c{#n|%Jj(6!y zuzdb5d6wnlK7T1V?sLa=o#mPk&yw$p`>f92KlfkTv!UNJ5qsVWjx!p+nPX+zD;i^5(gQ%fLRap}ysaoNFkKyLJWG+Od}vu-x6` zt|iy+%9Ysi*D}8LrC_;7!O_zyZ1YZsi(2j2@_DV**z!^9GO%1;s{`9w_hhx!V9V#V zIARs0UmACOE$hT(@=D@)5Hhd^5bcamQ@{%ZJ2$wh?XuqThOc3vB|+zsQ)kU|#|D zTm53hdd6G{=eQwHB>pNmW^>xtHJVqZ^qw0o3Z5`&->8%ZNZlJ z3GSN2Ne8zzarxga+u)3uf|z4W{C#q5=5q{VV}93x^@lv0*5Z0NefAS;djr_@4~c$m zgmWG~@$Zsuf|HM!n^O#GN6al4@)2`uig8|1=QcR`=;3y-wawKR`+Emi-vi_ZcPF+z zKM3bp^BeIlY9{kdoN!I>Xt^nS3pVLt#iFV_1(usO!* zi*xfzu;)g9oUvDd-G}g5^LEC3HJm=*gBI+Ez-Oa<5Bxrmdl($NW1IH~oP3CDBL66y zHT64&=lC^X`S?5bwP5$ya}qtj4(yrM7k^)k0Q3Li+f=`OTFbhw|LeiVe3Y@Z-vIU< zl#g|OBUs)ixW^JF{Sk8Ax5qJz)faJ3fX#J}BJN3SWA*v(FU)-k%>Rr3R+@j08TG|)+F-DJ z%ykraBoe>nL%{Z@_D*k3_Gq-::new(200_000, device.clone(), queue.clone()); + let mut voxel_cache = VoxelCache::<4>::new(100_000, device.clone(), queue.clone()); let cache_interface = CacheProducerInterface::new(1024, &device); - let terrain_generator = TerrainGenerator::<4>::new(5, "vxls_height.tif", 0.2, "img.jpg"); + //let terrain_generator = TerrainGenerator::<4>::new(5, "vxls_height.tif", 0.2, "img.jpg"); // let terrain_generator = TerrainGenerator::<4>::new( // 5, // "./pointe_percee/height.tif", // 0.2, // "./pointe_percee/ortho.jpg", // ); - // let terrain_generator = TerrainGenerator::<4>::new( - // 5, - // "/home/albin/Documents/vxls_maps/lapiz/height.tif", - // 0.2, - // "/home/albin/Documents/vxls_maps/lapiz/ortho.jpg", - // ); + let terrain_generator = TerrainGenerator::<4>::new( + 5, + "/home/albin/Documents/vxls_maps/lapiz/height.tif", + 0.2, + "/home/albin/Documents/vxls_maps/lapiz/ortho.jpg", + ); let mut chunk_pos_map = HashMap::new(); let chunk_instances = (0..terrain_generator.chunk_width) @@ -270,7 +272,7 @@ impl State depth_stencil: Some(wgpu::DepthStencilState { format: wgpu::TextureFormat::Depth24PlusStencil8, depth_write_enabled: Some(true), - depth_compare: Some(wgpu::CompareFunction::Less), + depth_compare: Some(wgpu::CompareFunction::LessEqual), stencil: wgpu::StencilState::default(), bias: wgpu::DepthBiasState::default(), }), @@ -645,7 +647,7 @@ impl State let generator = cloned_generator; let gen_test = SineGenerator::<4>::new(5); - let mut structure_nodes = vec![]; + let mut structure_nodes: Vec> = vec![]; let mut color_nodes: Vec<[ColorBytes; 64]> = vec![]; let mut location_nodes = vec![]; @@ -704,7 +706,15 @@ impl State }; } - (node.structure, node.colors, location) + ( + StructurePoolElement { + occupancy_low: 0, + occupancy_high: 0, + pointers: node.structure, + }, + node.colors, + location, + ) }) .collect::>() .into_iter() diff --git a/src/voxel_cache.rs b/src/voxel_cache.rs index 7b1d1c8..d2fa63c 100644 --- a/src/voxel_cache.rs +++ b/src/voxel_cache.rs @@ -346,6 +346,8 @@ where format!(" struct StructurePoolElement {{ + occupancy_low: atomic, + occupancy_high: atomic, pointers: array }} @@ -565,15 +567,37 @@ where {{ // Phase 1 // Copy into cache page + var occupancy_high = u32(0); + var occupancy_low = u32(0); for(var i = 0; i < {children_count}; i++) {{ structure_pool[overwritten_element].pointers[i] = select(u32(0), u32(1)<<31, structure_nodes[index].pointers[i] != 0); + + let turn_on_bit = + select( + u32(0), + u32(u32(1) << u32(i % ({children_count} / 2))), + ((color_nodes[index].colors[i] >> 24) != 0) || + structure_nodes[index].pointers[i] != 0 + ); + + if(i > {children_count} / 2) + {{ + occupancy_high |= turn_on_bit; + }}else + {{ + occupancy_low |= turn_on_bit; + }} }} + atomicStore(&structure_pool[overwritten_element].occupancy_low, occupancy_low); + atomicStore(&structure_pool[overwritten_element].occupancy_high, occupancy_high); + color_pool[overwritten_element] = color_nodes[index]; location_pool[overwritten_element] = locations[index]; // Mark dirty/correct timestamp - usage_buffer[overwritten_element] = parameters.frame_timestamp + 1; + usage_buffer[overwritten_element] = parameters.frame_timestamp + 1; + }} if(parameters.write_pointers != 0 && usage_buffer[overwritten_element] != parameters.frame_timestamp) @@ -588,6 +612,15 @@ where }}else if usage_buffer[requests_wb[index].node_index] != parameters.frame_timestamp + 1 {{ structure_pool[requests_wb[index].node_index].pointers[requests_wb[index].child_index] = new_pointer; + + let turn_on_bit = u32(1 << (requests_wb[index].child_index % ({children_count} / 2))); + if(requests_wb[index].child_index > {children_count} / 2) + {{ + //atomicOr(&structure_pool[requests_wb[index].node_index].occupancy_high, turn_on_bit); + }}else + {{ + //atomicOr(&structure_pool[requests_wb[index].node_index].occupancy_low, turn_on_bit); + }} }} }} diff --git a/src/voxel_cache/data.rs b/src/voxel_cache/data.rs index ba74146..aa94d55 100644 --- a/src/voxel_cache/data.rs +++ b/src/voxel_cache/data.rs @@ -24,11 +24,14 @@ where request_count: [u32; N * N * N], } +#[repr(C)] pub struct StructurePoolElement where [(); N * N * N]:, { - pointers: [StructurePointer; N * N * N], + pub occupancy_low: u32, + pub occupancy_high: u32, + pub pointers: [StructurePointer; N * N * N], } pub struct DestinationElement