From ddf36431f5e4b31acaaa58ebf80f44c3e3c5b660 Mon Sep 17 00:00:00 2001
From: LegendWasTaken <caiodcw@gmail.com>
Date: Mon, 10 May 2021 01:12:30 +0800
Subject: [PATCH] Add resolution scaling

---
 assets/shaders/compute_pass.comp       |  68 +++++++++++++------------
 assets/shaders/compute_pass.comp.spv   | Bin 65264 -> 65612 bytes
 assets/shaders/debug_vis.frag.spv      | Bin 556 -> 556 bytes
 assets/shaders/debug_vis.vert.spv      | Bin 1340 -> 1340 bytes
 assets/shaders/fullscreen_tri.vert.spv | Bin 1312 -> 1312 bytes
 assets/shaders/tex_sample.frag.spv     | Bin 656 -> 656 bytes
 src/rvpt/bvh_builder.h                 |   2 +-
 src/rvpt/main.cpp                      |   6 +--
 src/rvpt/rvpt.cpp                      |  10 ++--
 src/rvpt/rvpt.h                        |   5 +-
 10 files changed, 49 insertions(+), 42 deletions(-)

diff --git a/assets/shaders/compute_pass.comp b/assets/shaders/compute_pass.comp
index be1b23f..9f1990d 100644
--- a/assets/shaders/compute_pass.comp
+++ b/assets/shaders/compute_pass.comp
@@ -126,38 +126,42 @@ void main()
 		5: Kajiya
 		
 	*/
-    int integrator_idx = render_settings.top_left_render_mode;
-    vec2 pixel_split = vec2(gl_GlobalInvocationID) / dim;
-    if (pixel_split.y > render_settings.split_ratio.y)
-    {
-        if (pixel_split.x < render_settings.split_ratio.x)
-            integrator_idx = render_settings.bottom_left_render_mode;
-        else
-            integrator_idx = render_settings.bottom_right_render_mode;
-    }
-    else if (pixel_split.x > render_settings.split_ratio.x)
-        integrator_idx = render_settings.top_right_render_mode;
 
-    vec3 temporal_accumulation_sample =
-        (imageLoad(temporal_image, ivec2(gl_GlobalInvocationID.xy))).xyz *
-        min(render_settings.current_frame, 1);
-
-    vec3 sampled = vec3(0);
-    for (int i = 0; i < render_settings.aa; i++)
+    if (gl_GlobalInvocationID.y < dim.y && gl_GlobalInvocationID.x < dim.x)
     {
-        vec2 coord = (vec2(gl_GlobalInvocationID.xy) + vec2(rand(), rand())) / dim;
-		coord.y = 1.0-coord.y; /* flip image vertically */
-        
-		Ray ray = get_camera_ray(render_settings.camera_mode, coord.x, coord.y);
-		sampled += eval_integrator(integrator_idx, ray);
-	}
-
-    float current_frame = float(render_settings.current_frame);
-
-    sampled /= render_settings.aa;
-    sampled = temporal_accumulation_sample * current_frame / (current_frame + 1) +
-              sampled / (current_frame + 1);
-
-    imageStore(temporal_image, ivec2(gl_GlobalInvocationID.xy), vec4(sampled, 0));
-    imageStore(result_image, ivec2(gl_GlobalInvocationID.xy), vec4(sampled, 0));
+        int integrator_idx = render_settings.top_left_render_mode;
+        vec2 pixel_split = vec2(gl_GlobalInvocationID) / dim;
+        if (pixel_split.y > render_settings.split_ratio.y)
+        {
+            if (pixel_split.x < render_settings.split_ratio.x)
+                integrator_idx = render_settings.bottom_left_render_mode;
+            else
+                integrator_idx = render_settings.bottom_right_render_mode;
+        }
+        else if (pixel_split.x > render_settings.split_ratio.x)
+            integrator_idx = render_settings.top_right_render_mode;
+
+        vec3 temporal_accumulation_sample =
+            (imageLoad(temporal_image, ivec2(gl_GlobalInvocationID.xy))).xyz *
+            min(render_settings.current_frame, 1);
+
+        vec3 sampled = vec3(0);
+        for (int i = 0; i < render_settings.aa; i++)
+        {
+            vec2 coord = (vec2(gl_GlobalInvocationID.xy) + vec2(rand(), rand())) / dim;
+            coord.y = 1.0-coord.y; /* flip image vertically */
+
+            Ray ray = get_camera_ray(render_settings.camera_mode, coord.x, coord.y);
+            sampled += eval_integrator(integrator_idx, ray);
+        }
+
+        float current_frame = float(render_settings.current_frame);
+
+        sampled /= render_settings.aa;
+        sampled = temporal_accumulation_sample * current_frame / (current_frame + 1) +
+                  sampled / (current_frame + 1);
+
+        imageStore(temporal_image, ivec2(gl_GlobalInvocationID.xy), vec4(sampled, 0));
+        imageStore(result_image, ivec2(gl_GlobalInvocationID.xy), vec4(sampled, 0));
+    }
 }
diff --git a/assets/shaders/compute_pass.comp.spv b/assets/shaders/compute_pass.comp.spv
index 460c936872ad5fdbab846bc4308ffa7ed6990717..69f3aadae6e3323813299b2cfc034112981e4ff7 100644
GIT binary patch
delta 3751
zcmZve$#Ydz5XMgihC~qX2@wTkP|*ZLP!WV6!2!epqN0KWK9Y!0Nx}n^(HnH*$|~#q
z1+H8wMM+ug+`Dn-gg7Cfs0<>54E}yEx8p0yn>uyA@9RE&`t<F4UXEX^`@Fw?WL@pF
zVJX$5#xy+j?-{wHdHm6al)8I1cdaiKH}{s-c6V+~X$XJE${#crYQ~m7X`Ws)u6&@m
zDVx@q(x&2uzSp}tvDc-W;N3~@`gA+Eqqn!zS>bi56<#WCtJoHS`--LFhTOO93jV+q
zeAyLz`2cUs0}M^=<y|dJGj}0BQS9mLCGOBPEXjRJ6>aUN-Yun$F1V>?dp~PgUzd%;
zWopGKrKMyop8~IC$alc60*}x6PWVu8J!2VxVvA{TpQVIWtE=Gpl!2?`%Z;zSkZSjS
zIP0s%><k4Tj?hBbxy<$XCc%fpc}Hmh_M^-u>C*(iF84KI{Y`E=+Gyq(unWu{Ld{WI
zlhQ=C_aLLNO-glBD;X<HCcs-dVek}gD9_4M7A(l@i%4TT9>=Qq4ER_)!!EZwg71Ud
z9l>31hXh{;Z>{3`+a+Nyq_ia2CRIYRO@h0?HVJM(+ax#vQv2SC3qPvKE(5<L_e;!{
zqctzvS7_>Oyw@S<UxN=<`S@zSK~vwEulW|N_AzT(u^#{%NBs<CKFI8I9)E4$HKeQ}
zo~6i;%4R0?J=lcKGa{iM(9|QLAHgORJ|^@NntCMkGgvJWIs`V3`Z@MI*zjaZZf<6$
zDUVtI0yeuLU>BH}*&ku{F?)RuGt^u!HnI1ae=UE}TF8DAF&44XUXgngTkv<V$LoV%
z%lrpit(^$A;h$hNpU{ry8m}U)#BV9{2{e7R8T%00Nib!}QZIr|rw|n1%@_RzcA>h-
z*|=(^Z(p1StJUVsU_yVx)!X=n)}WsSpFsCpt@b+y*S0_1KhOc^AuiTn4KFbJSVL|9
zFs`C8t)V!QxClO0E-r4$V#~`Yv1Pn+Q{$>m+CC+H-(C&;5?tLK>zIAqQClrz7{eXK
z@z#cbjqi11Ep4Zvjnq*?yhMab-x$;Y28sX;U;}tRMSw=IdaRe*cd_1CZdr_HDWU`;
zz{eW0{~FxWjzrWDRk<4ce^nWUu5Eu_mC<0=s#}$7n0>5@wlNGflMu(D3SQA}FgNc8
z%W^GRff1*Ydx{9)Tzx7}XGM;%wh5azw%ku)JlIcR1s|Wc_w@`lhfjH4+lZMr<c4*e
z=1#;3a098Y;K#ymXd+mBLV06bQ#J`?+9w%q<qyG)XzC-_N<JCv*K|v6tA+SQTIVU)
zw7<uWMxP3H2X&9_?XD)C#{39BVPbFfO*LtG<@`58EU5F}1fI@(E5pb7+?Ly&qhp%E
z28kEYQf^;T$Y!9lF!toHb0%0%bzh<U7KS^xqc|!&3tU~{JK$;#A4}k|y(?y;X`i2$
zVGekVmq`eB4hVxc&|Pqy)OVAdLGA_{NSjG`E6oLSD-B*s_rTTSQkn;Low{}O*HCTX
zW8fp04^i|b8Pu0_FI@c{HkK~{cVSl_SeVj6aI7=1T)(LAePA19a~{|JUW}%$zdyT+
zZOjO`1Z=>{9aJtpSjgI7_g8s5n0r*G`VgABA^oA2&!-QsXS=1~pP1dsSladTKLBmf
z8V`g0CDG=OyBu3q3G#@9p>a90d_HY%^F9hb!0aMz22_6xZ1e8UoBVOGn*PSore?3&
zw|16uB(ws2QzI{@p8ytaC4%B=ifI5FS<L|PG1#Kk@)Z70fUo>YpF~rSFYzg`n(^A1
z<?-@XflriIEg6+P4U+kzXDd7dR*R!P3-&lG^OCFv>l4F1v?OcL^wl0EP>VRvfyd_D
zeVzxCRW^80ynvvS1{dk{;($}7&hoWzHG4G@dI_!`BN8g28B=>Cu@3BEB7TSOz}omC
z-JlcP!KmJ#3$7j`Hh3A$fZC%G)`Q(xTU0^K4ZXx)0gubMmw2}yNcA1_D!kiKp50c+
zUIVE_{;z}0KmN7Z09Nw}ttZzg(V*6wYyLyNl8?uqf<83Y>#x0^1s$YmitdYj&S2gf
z5q@QniI{f@UY+(PxLT=__GXYuq`d`Po%S1WHJ?cP&0MQaduy&$r@al$_4;d%w9U|c
Uv5%^JHt`SUk%ql*FYDR&FD{U*uK)l5

delta 3454
zcmZ9O$#Ydj6vjJwFM$LiL(oVx$e^Ghh>AvH00SZcQ3*H!D)>kOMj^omRyo#%a^uP>
z>iq?-Tv}90%5tY`H||8l2^m!cQ6>TX{cdhkm)um--}jx<r%!jE?#I`Aif67C8Vc33
z#)ME6#_~6`zv121so&OxFfh2ge_Oe<d#Jp5pzo~^YWO=;`M9;Qs#y7~bxu`H<#20L
zGPyp4U8Nm6xA*s9FN8+;K+wAwTEM+SL*>4d7eX7nTzWgT%>(Z&l}kG^-}^@K1*7=G
zqj<*%ug?P1hOWx{bDQREM_yMN>>DC(Z5R{ee&tl#yld#qa&JG}R0oH@oV%@%)Zj8R
zbqb-ItmQM{)in8Q@NwX&8GjvK3og=D5-3hF3+~rVXthF$i(wF)k1sdA_Qp^>{K@=-
z^~o_dd@RCT!nV_&gF8v&7vP*x=)``K-X#5+;A6o0ny~)j0knzqZLll!E<(*#T@^wL
z)4Pxf*rtWT%rs+#=>*uJ69!k|gz~}4jLy!)x`=7KgU6v9Z-W={jCQ%z5%~eQ)e*Vl
zt&qqU!P|0Nf2$<g8$(zYER&oNER)C`V3|a2K+7a@0)(#NmPMadCHu(xAnhzt3)``n
zRl4opW%?4f$F<6}xrN_@uhH*>PoX~yS6fRBxQ#!6)%>FNW2P}h=)td>{-?T-q|O=*
zeiZR%@NxPfxK6*o)jrGy9RV{i)lJTgS6hnRb^Quf^K6Z2LPz21OR0g4=*Ph>_*h+X
zqvHt-8owaeVjNC_9jtzV$$q2vyG&qhr)ZaH7pW3)OyYO2NnFM^Ch-ScJtlD)e3tpi
z;zq|m;Tk{Bk~jl)u)0b7Mek=4+RoC>(@a8a5+3W{G*{^jG6|3MpG=R(nkZaipg|2_
zpcvpH*Z^L+7~qnAv=~61A8#DFJ7qj~C+^C>;8DBM9Iha0h}&`%oSsePjjpC79>jmB
zdS0Ws2US$FpL?J!z}1{m9G6fHHd=eGv<qnJaS3u$j^pjKB_wG;*W&t55U3WTPmbKX
zFc$1xSi|M=`KhC++5D{c_wco}>ofZn4rl><9Nc*7Yk2g$OXI=nwh$^`Et_1`P}#Sl
zF}Vq$fi|9reAgy`4L+Nuts3HmwB9GewSUA!2Au>pu)3@EaZ(feFs<eV7W*jeB95z{
z0(Lj;&xc}rxCPKe^V8wh%)vfZ$5gOE;`TRJz9IG7P?~A`vx2@Itf#sMD8HHJ3{EMI
z`#TMs-`_jnYBoQ&z-9YXOh?n+k!`~a@T9;S6NQ;0!u1De7F;Lwy(DLlJHZChW)dFU
zUEutiG#joK-=sNU$EmxIKIUp8h0h(o8=~kT8Pr3%2d@4B6U$q{{n*t<2J#~Yj^hk0
z*Dvnxy<p3GcNW(xJ|9h8e}6URC;Srw-VZil{$Y9muI>PTRpinCL2yhUCjJnby8iy^
z$UCUYW+rb3AE9?*V`)!x^!H4oLCkU?*xwax{#41e8RQYL{@VQ6l6O$_mU1WfFuh~6
zIYxaE*izn`mHc9`8va|;KpGV-UCY=)vyTZa0b9Cz2w;Mbg4H&#wI-rXEq^yGSa<3i
z{!6{3^OmX5;!S)EtY(BRdbuNfpO>eX<8hSbG=J|bgB4)4xUg=px%XrnuoA3KY}TL~
zunJ9I?at;HwHW6Ku=if>JgdQElw5x#o<z_|Lv(s-#3|iX2R#i}vrc0|YvAg!#e{m$
zjHx{)u@;;jM3%(5dM=0i;;}iwGvIZ!`~=U!)nkhjJcnjL?eW;32RpI0xC?3~;Pd?g
zxF+L1-|M{``S)c5e0|oir|(OAzc-?5kE?kR>}ulwab5ze`9<yJOk=~YYnw97|5+wI
z+#G)lHX|Cq0XpnqL@!yJu@h^Lt9S)`iZOCmQG(|eu?4PX^K<gJh+Z`9*&>qA2g$FZ
wAFk#XSFttI@~haEX|5tJ;#D-q>#sd7!i1ez`&H6eNNeZFYs2vARhtg{52m|%5C8xG

diff --git a/assets/shaders/debug_vis.frag.spv b/assets/shaders/debug_vis.frag.spv
index 056dc866bc2d7f48a975acedb3a654e48d2ad1d7..050cd02d9d038162cfaf3754fa3a10b5a13acc9c 100644
GIT binary patch
delta 18
ZcmZ3(vWA6|nMs+Qfq{{MYa^#J6964f0)qem

delta 18
ZcmZ3(vWA6|nMs+Qfq{{MV<V?B6964V0)YSk

diff --git a/assets/shaders/debug_vis.vert.spv b/assets/shaders/debug_vis.vert.spv
index e8a35e3b9f9c3acf7509c37e45fb6c5bd842ae64..c9847eeb5db8053a03c9fc914f6f3f5e691edb08 100644
GIT binary patch
delta 18
ZcmdnPwTFw7nMs+Qfq{{MYa^#ID*zx-0>S_Q

delta 18
ZcmdnPwTFw7nMs+Qfq{{MV<V?AD*zxz0>A(O

diff --git a/assets/shaders/fullscreen_tri.vert.spv b/assets/shaders/fullscreen_tri.vert.spv
index 72f4f7f569aaef1d93d3ec380913744f228101b0..fb6c58167750317b5c3455297a50e275e0893f75 100644
GIT binary patch
delta 18
ZcmZ3$wSbG0nMs+Qfq{{MYa^#PD*zl}0&V~R

delta 18
ZcmZ3$wSbG0nMs+Qfq{{MV<V?HD*zl<0&D;P

diff --git a/assets/shaders/tex_sample.frag.spv b/assets/shaders/tex_sample.frag.spv
index 49caa24eac34218c4545d48788784761d503c3b8..4abdbf2568ec54cf0b89aa59445f6dba0641392b 100644
GIT binary patch
delta 18
ZcmbQhI)RmwnMs+Qfq{{MYa?en695{10<i!9

delta 18
ZcmbQhI)RmwnMs+Qfq{{MV<Trf695`?0<Qo7

diff --git a/src/rvpt/bvh_builder.h b/src/rvpt/bvh_builder.h
index 1a350e2..69b356d 100644
--- a/src/rvpt/bvh_builder.h
+++ b/src/rvpt/bvh_builder.h
@@ -43,7 +43,7 @@ class BinnedBvhBuilder : public BvhBuilder
 
 private:
     // Threshold below which nodes are no longer split
-    static constexpr size_t min_primitives_per_leaf = 2;
+    static constexpr size_t min_primitives_per_leaf = 4;
     // Threshold above which nodes that are not split with binning are still split with the fallback strategy.
     static constexpr size_t max_primitives_per_leaf = 8;
     // Number of bins used to approximate the SAH. Higher = More accuracy, but slower.
diff --git a/src/rvpt/main.cpp b/src/rvpt/main.cpp
index 0c88b1f..3008d44 100644
--- a/src/rvpt/main.cpp
+++ b/src/rvpt/main.cpp
@@ -96,13 +96,13 @@ int main()
 
 
     Window::Settings settings;
-    settings.width = 1024;
-    settings.height = 512;
+    settings.width = 3330;
+    settings.height = 1340;
     Window window(settings);
 
     RVPT rvpt(window);
 
-    load_model(rvpt, "models/rabbit.obj", 1);
+    load_model(rvpt, "models/tridel-interior-test.obj", 1);
 
     // Setup Demo Scene
     rvpt.add_material(
diff --git a/src/rvpt/rvpt.cpp b/src/rvpt/rvpt.cpp
index 525a69f..46257c1 100644
--- a/src/rvpt/rvpt.cpp
+++ b/src/rvpt/rvpt.cpp
@@ -799,11 +799,11 @@ void RVPT::add_per_frame_data(int index)
     auto output_image = VK::Image(vk_device, memory_allocator, *graphics_queue,
                                   "raytrace_output_image_" + std::to_string(index),
                                   VK_FORMAT_R8G8B8A8_UNORM, VK_IMAGE_TILING_OPTIMAL,
-                                  window_ref.get_settings().width, window_ref.get_settings().height,
+                                  window_ref.get_settings().width / resolution_scale, window_ref.get_settings().height / resolution_scale,
                                   VK_IMAGE_USAGE_SAMPLED_BIT | VK_IMAGE_USAGE_STORAGE_BIT,
                                   VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_ASPECT_COLOR_BIT,
-                                  static_cast<VkDeviceSize>(window_ref.get_settings().width *
-                                                            window_ref.get_settings().height * 4),
+                                  static_cast<VkDeviceSize>((window_ref.get_settings().width *
+                                                            window_ref.get_settings().height * 4) / resolution_scale),
                                   VK::MemoryUsage::gpu);
     auto random_buffer =
         VK::Buffer(vk_device, memory_allocator, "random_data_uniform_" + std::to_string(index),
@@ -1028,8 +1028,8 @@ void RVPT::record_compute_command_buffer()
         cmd_buf, VK_PIPELINE_BIND_POINT_COMPUTE, rendering_resources->raytrace_pipeline_layout, 0,
         1, &per_frame_data[current_frame_index].raytracing_descriptor_sets.set, 0, 0);
 
-    vkCmdDispatch(cmd_buf, per_frame_data[current_frame_index].output_image.width / 16,
-                  per_frame_data[current_frame_index].output_image.height / 16, 1);
+    vkCmdDispatch(cmd_buf, glm::ceil(per_frame_data[current_frame_index].output_image.width / 16.f),
+                  glm::ceil(per_frame_data[current_frame_index].output_image.height / 16.f), 1);
 
     command_buffer.end();
 }
diff --git a/src/rvpt/rvpt.h b/src/rvpt/rvpt.h
index 0376d9f..e78d8d0 100644
--- a/src/rvpt/rvpt.h
+++ b/src/rvpt/rvpt.h
@@ -4,6 +4,7 @@
 #include <vector>
 #include <optional>
 #include <random>
+#include <atomic>
 
 #include <vulkan/vulkan.h>
 
@@ -85,7 +86,6 @@ class RVPT
         int bottom_left_render_mode = 9;
         int bottom_right_render_mode = 9;
         glm::vec2 split_ratio = glm::vec2(0.5, 0.5);
-
     } render_settings;
 
 private:
@@ -100,6 +100,9 @@ class RVPT
 
     bool debug_bvh_enabled = false;
 
+    // Currently hardcoded, should be able to change in the future.
+    float resolution_scale = 3.0f;
+
     Window& window_ref;
     std::string source_folder = "";