Optimizing Ray vs. AABB Intersections — Handmade Hero — Episode Guide

0:02Recap and set the stage for the day refining the lighting

🗩

0:02Recap and set the stage for the day refining the lighting

🗩

0:02Recap and set the stage for the day refining the lighting

🗩

0:31Take Visual Studio's quick feedback survey

🗹

💢

0:31Take Visual Studio's quick feedback survey

🗹

💢

0:31Take Visual Studio's quick feedback survey

🗹

💢

8:15Run the game to see our current lighting situation

🏃

8:15Run the game to see our current lighting situation

🏃

8:15Run the game to see our current lighting situation

🏃

11:59Consider the performance of ComputeLightPropagation()

🏃

11:59Consider the performance of ComputeLightPropagation()

🏃

11:59Consider the performance of ComputeLightPropagation()

🏃

14:18Track TotalPartitionLeavesUsed in ComputeLightPropagation()

17:57Consult the TotalPartitionLeavesUsed performance figures

18:22Determine to change SplitBox() from doing k-d to quad tree partitioning

20:15k-d tree vs Quad tree

🖌

20:15k-d tree vs Quad tree

🖌

20:15k-d tree vs Quad tree

🖌

21:00Allow SplitBox() to have 8 leaves per child, increased from 4

21:32Run the game to see that that doesn't appreciably change our runtime

🏃

21:32Run the game to see that that doesn't appreciably change our runtime

🏃

21:32Run the game to see that that doesn't appreciably change our runtime

🏃

21:44Track PartitionsPerLeaf in ComputeLightPropagation()

22:37Run the game and compare the PartitionsPerLeaf with 8 and 4 leaves per child

🏃

22:37Run the game and compare the PartitionsPerLeaf with 8 and 4 leaves per child

🏃

22:37Run the game and compare the PartitionsPerLeaf with 8 and 4 leaves per child

🏃

23:09Toggle off ComputeLightPropagation()

23:41Run the game to determine that our shader is the bottleneck

🏃

23:41Run the game to determine that our shader is the bottleneck

🏃

23:41Run the game to determine that our shader is the bottleneck

🏃

26:00Toggle off wglSwapIntervalExt in Win32InitOpenGL()

26:25Run the game and watch the threads view with the determination to investigate the time delay

🏃

26:25Run the game and watch the threads view with the determination to investigate the time delay

🏃

26:25Run the game and watch the threads view with the determination to investigate the time delay

🏃

28:29There's a knock at the door

🗹

28:29There's a knock at the door

🗹

28:29There's a knock at the door

🗹

28:50afk

29:08Return and consider the threads view to be misleading

🏃

29:08Return and consider the threads view to be misleading

🏃

29:08Return and consider the threads view to be misleading

🏃

31:33Introduce HUD_TIMED_FUNCTION() called in LightingTest() to enable more direct textual profiling of the lighting rendering

36:53Run the game to see that it does nothing different

🏃

36:53Run the game to see that it does nothing different

🏃

36:53Run the game to see that it does nothing different

🏃

37:21Enable DEBUGEnd() and DEBUGInit() to handle HUD_TIMED_FUNCTION(), renaming AddTooltip() to AddLine() and DrawTooltips() to DrawLineBuffer()

57:53Crash in DEBUGDrawElement() and inspect what's going on

🏃

57:53Crash in DEBUGDrawElement() and inspect what's going on

🏃

57:53Crash in DEBUGDrawElement() and inspect what's going on

🏃

1:02:39Enable DrawTreeLink() to handle the case when a tree has no element and no children, and change HasChildren() to CanHaveChildren()

1:04:52Run the game to see that we do not crash, and see our new HUD element

🏃

1:04:52Run the game to see that we do not crash, and see our new HUD element

🏃

1:04:52Run the game to see that we do not crash, and see our new HUD element

🏃

1:06:24Make DEBUGEnd() display only the function name of our HUD_TIMED_FUNCTION(), expandable to contain textual profiling information for its child functions

1:17:27Step in to DEBUGEnd() and inspect the HUD element

🏃

1:17:27Step in to DEBUGEnd() and inspect the HUD element

🏃

1:17:27Step in to DEBUGEnd() and inspect the HUD element

🏃

1:22:41Enable DEBUGEnd() to gather profiling information for our HUD element

1:31:08Run the game and consult our LightingTest() HUD to see that our average value is busted

🏃

1:31:08Run the game and consult our LightingTest() HUD to see that our average value is busted

🏃

1:31:08Run the game and consult our LightingTest() HUD to see that our average value is busted

🏃

1:32:24Fix DEBUGEnd() to correctly compute the average cycle count

1:33:55Run the game and consult the performance of LightingTest() in the new HUD

🏃

1:33:55Run the game and consult the performance of LightingTest() in the new HUD

🏃

1:33:55Run the game and consult the performance of LightingTest() in the new HUD

🏃

1:35:37Tweak the leaves per child value in SplitBox(), comparing their performance in the HUD, and consider that our spatial partitioning is not effective enough

🏃

1:35:37Tweak the leaves per child value in SplitBox(), comparing their performance in the HUD, and consider that our spatial partitioning is not effective enough

🏃

1:35:37Tweak the leaves per child value in SplitBox(), comparing their performance in the HUD, and consider that our spatial partitioning is not effective enough

🏃

1:39:35Consider making the RayCast() more performant with AABB testing

📖

1:39:35Consider making the RayCast() more performant with AABB testing

📖

1:39:35Consider making the RayCast() more performant with AABB testing

📖

1:41:49Read 'Fast, Branchless Ray / Bounding Box Intersections'¹ and 'Fast Ray / Axis-Aligned Bounding Box Overlap Tests using Ray Slopes'²

📖

1:41:49Read 'Fast, Branchless Ray / Bounding Box Intersections'¹ and 'Fast Ray / Axis-Aligned Bounding Box Overlap Tests using Ray Slopes'²

📖

1:41:49Read 'Fast, Branchless Ray / Bounding Box Intersections'¹ and 'Fast Ray / Axis-Aligned Bounding Box Overlap Tests using Ray Slopes'²

📖

1:45:36Ray AABB testing

🖌

1:45:36Ray AABB testing

🖌

1:45:36Ray AABB testing

🖌

1:47:18Continue to read these papers on ray AABB testing³^,4

📖

1:47:18Continue to read these papers on ray AABB testing³^,4

📖

1:47:18Continue to read these papers on ray AABB testing³^,4

📖

1:56:23Change RayCast() to perform AABB testing⁵

1:59:15Computing X, Y and Z intersections in T

🖌

1:59:15Computing X, Y and Z intersections in T

🖌

1:59:15Computing X, Y and Z intersections in T

🖌

2:06:21Doing this without caring which direction the normal is facing

🖌

2:06:21Doing this without caring which direction the normal is facing

🖌

2:06:21Doing this without caring which direction the normal is facing

🖌

2:09:31Continue to enable RayCast() to perform fast AABB testing

2:13:05Determining the tMin and tMax intersection points

🖌

2:13:05Determining the tMin and tMax intersection points

🖌

2:13:05Determining the tMin and tMax intersection points

🖌

2:14:12Enable RayCast() to compute those tMin and tMax

2:15:50The Maximum-Minimum and Minimum-Maximum

🖌

2:15:50The Maximum-Minimum and Minimum-Maximum

🖌

2:15:50The Maximum-Minimum and Minimum-Maximum

🖌

2:16:29Set tMin and tMax and continue enabling RayCast() to perform fast AABB testing

2:27:31Consider how to determine which side of the box we hit

📖

2:27:31Consider how to determine which side of the box we hit

📖

2:27:31Consider how to determine which side of the box we hit

📖

2:30:24Enable RayCast() to select the correct BoxSurfaceIndex

📖

2:30:24Enable RayCast() to select the correct BoxSurfaceIndex

📖

2:30:24Enable RayCast() to select the correct BoxSurfaceIndex

📖

2:35:27Introduce f32_4x and v3_4x versions of Min() and Max(), and the v3_4x / operator

2:38:03Run the game to see that it looks pretty similar, and is faster

🏃

2:38:03Run the game to see that it looks pretty similar, and is faster

🏃

2:38:03Run the game to see that it looks pretty similar, and is faster

🏃

2:39:51Introduce v3_4x * operator and make RayCast() precompute the RayD

2:40:53Compare the performance of this with the previous way

🏃

2:40:53Compare the performance of this with the previous way

🏃

2:40:53Compare the performance of this with the previous way

🏃

2:41:29Consider breaking down our profiling of LightingTest()

📖

2:41:29Consider breaking down our profiling of LightingTest()

📖

2:41:29Consider breaking down our profiling of LightingTest()

📖

2:43:25Consult the threads view to see that the performance is more reasonable

2:43:52Try to stress the system

🏃

🖮

2:43:52Try to stress the system

🏃

🖮

2:43:52Try to stress the system

🏃

🖮

2:45:17Q&A

🗩

2:45:17Q&A

🗩

2:45:17Q&A

🗩

2:46:01@0b0000000000000 SurfaceIndexLookupTable[movemask(tmin == tBoxMin)], and just choose the first set bit in the return from the movemask for which index

🗪

2:46:01@0b0000000000000 SurfaceIndexLookupTable[movemask(tmin == tBoxMin)], and just choose the first set bit in the return from the movemask for which index

🗪

2:46:01@0b0000000000000 SurfaceIndexLookupTable[movemask(tmin == tBoxMin)], and just choose the first set bit in the return from the movemask for which index

🗪

2:47:15@vaualbus Q: Would we get an improvement if we switch all the v3 to a v3_4x so we have not to load the value into the __m128 each frame?

🗪

2:47:15@vaualbus Q: Would we get an improvement if we switch all the v3 to a v3_4x so we have not to load the value into the __m128 each frame?

🗪

2:47:15@vaualbus Q: Would we get an improvement if we switch all the v3 to a v3_4x so we have not to load the value into the __m128 each frame?

🗪

2:47:40@0b0000000000000 It would only be 256 entries

🗪

2:47:40@0b0000000000000 It would only be 256 entries

🗪

2:47:40@0b0000000000000 It would only be 256 entries

🗪

2:48:05@alexkelbo Q: Could you recap how we retain the state of the debug UI between frames?

🗪

2:48:05@alexkelbo Q: Could you recap how we retain the state of the debug UI between frames?

🗪

2:48:05@alexkelbo Q: Could you recap how we retain the state of the debug UI between frames?

🗪

2:48:31Consider the movemask instruction

🗩

2:48:31Consider the movemask instruction

🗩

2:48:31Consider the movemask instruction

🗩

2:49:46@0b0000000000000 You can collapse the 4 bytes into 2 bytes

🗪

2:49:46@0b0000000000000 You can collapse the 4 bytes into 2 bytes

🗪

2:49:46@0b0000000000000 You can collapse the 4 bytes into 2 bytes

🗪

2:52:19@roam00010011 Q: tMin3 will always be the closest in distance to RayOrigin, since RayD that hits BoxMax will result in tBoxMax < tBoxMin, no?

🗪

2:52:19@roam00010011 Q: tMin3 will always be the closest in distance to RayOrigin, since RayD that hits BoxMax will result in tBoxMax < tBoxMin, no?

🗪

2:52:19@roam00010011 Q: tMin3 will always be the closest in distance to RayOrigin, since RayD that hits BoxMax will result in tBoxMax < tBoxMin, no?

🗪

2:53:00Determining the closest hit when rays can be cast in all directions

🖌

2:53:00Determining the closest hit when rays can be cast in all directions

🖌

2:53:00Determining the closest hit when rays can be cast in all directions

🖌

2:53:33Sketch out RayCast() computing the RayPosition from the ray direction

2:56:08Glimpse into the future optimising the spatial hierarchy stuff

🗩

2:56:08Glimpse into the future optimising the spatial hierarchy stuff

🗩

2:56:08Glimpse into the future optimising the spatial hierarchy stuff

🗩

2:57:00Consult the profiler in terms of multithreading

🏃

2:57:00Consult the profiler in terms of multithreading

🏃

2:57:00Consult the profiler in terms of multithreading

🏃

2:57:41Try quadrupling PointsPerWork in ComputeLightPropagation()

2:57:54Consult the threads view to see that there is more empty space now

🏃

2:57:54Consult the threads view to see that there is more empty space now

🏃

2:57:54Consult the threads view to see that there is more empty space now

🏃

2:59:40Optimising lane usage when threading

🖌

2:59:40Optimising lane usage when threading

🖌

2:59:40Optimising lane usage when threading

🖌

3:01:54Note that we are not measuring dead computation time, and just enjoy our lighting

🏃

3:01:54Note that we are not measuring dead computation time, and just enjoy our lighting

🏃

3:01:54Note that we are not measuring dead computation time, and just enjoy our lighting

🏃

3:03:05Close down, with one last look at 'Fast, Branchless Ray / Bounding Box Intersections'⁶

📖

3:03:05Close down, with one last look at 'Fast, Branchless Ray / Bounding Box Intersections'⁶

📖

3:03:05Close down, with one last look at 'Fast, Branchless Ray / Bounding Box Intersections'⁶

📖

3:04:08Glimpse into the future, either continuing with optimisation or investigating the lighting flicker

🗩

3:04:08Glimpse into the future, either continuing with optimisation or investigating the lighting flicker

🗩

3:04:08Glimpse into the future, either continuing with optimisation or investigating the lighting flicker

🗩

Keyboard Navigation

Global Keys

Menu toggling

In-Menu and Index Controls

Quotes and References Menus and Index

Quotes, References and Credits Menus

Filter Menu

Filter and Link Menus

Credits Menu