Abstracting the Work Queue — Handmade Hero — Episode Guide

0:07We are absolute control freaks here, people^α

1:25Recap and set the stage for today

3:46win32_handmade.cpp: Introduce DoWorkerWork

5:43Let our normal thread do work

6:55Run and see what the threads are doing

7:37Follow the compression oriented programming approach

9:10handmade_render_group.cpp: Figure out a way to do TiledRenderGroupToOutput on multiple threads

10:38handmade_platform.h: Consider pulling in work_queue_entry

11:54win32_handmade.cpp: Rewrite PushString as AddWorkQueueEntry

15:18Note the necessity of _mm_sfence

16:13Pull work_queue_entry down into the test code

17:18Split DoWorkerWork in two

20:34Think

21:18Put while(EntryCount != EntryCompletionCount) into QueueWorkStillInProgress

22:49Rename and finish writing these functions

30:48Compile and run and see what the threads are doing

31:14Discuss our options

34:31Rename GetNextWorkQueueEntry to CompleteAndGetNextWorkQueueEntry and make it take work_queue_entry Completed

35:29Rearrange ThreadProc slightly

37:24Massage DoWorkerWork

37:51Tweak the QueueWorkStillInProgress loop

38:55Compile and consider removing one more call

40:12Go for it and make the work_queue two separate things

44:22Run this again

45:19handmade_platform.h: Hoist these functions in

46:01Think about this a little bit more

47:11handmade_render_group.cpp: Write the usage code first

50:06Compile and express hate for const^β

50:40Finish writing TiledRenderGroupToOutput

54:58Compile and run and crash^γ

55:22Moment of realisation: Gotta increment by the correct value

55:32Recap and glimpse into the multithreaded future

56:09Q&A

🗩

56:09Q&A

🗩

56:09Q&A

🗩

57:35BrainCruser Will you start new threads for every queue that you make?

🗪

57:35BrainCruser Will you start new threads for every queue that you make?

🗪

57:35BrainCruser Will you start new threads for every queue that you make?

🗪

59:29niegrfiegr0 Still don't understand the use of volatile and memory barrier

🗪

59:29niegrfiegr0 Still don't understand the use of volatile and memory barrier

🗪

59:29niegrfiegr0 Still don't understand the use of volatile and memory barrier

🗪

59:53Blackboard: Memory and Code Fences

1:07:46kelimion Can Entry.IsValid be removed and replaced with a test to see if Entry.Data != NULL?

🗪

1:07:46kelimion Can Entry.IsValid be removed and replaced with a test to see if Entry.Data != NULL?

🗪

1:07:46kelimion Can Entry.IsValid be removed and replaced with a test to see if Entry.Data != NULL?

🗪

1:08:04kil4h What is your take on Naughty Dog's approach using fibers (+ manual management) and thread affinity to core instead of using classic worker / job approach for multithreaded gameplay?

🗪

1:08:04kil4h What is your take on Naughty Dog's approach using fibers (+ manual management) and thread affinity to core instead of using classic worker / job approach for multithreaded gameplay?

🗪

1:08:04kil4h What is your take on Naughty Dog's approach using fibers (+ manual management) and thread affinity to core instead of using classic worker / job approach for multithreaded gameplay?

🗪

1:08:23robrobby The work queue will take any function to do it multithreaded? Does the function need to be special so that this will work?

🗪

1:08:23robrobby The work queue will take any function to do it multithreaded? Does the function need to be special so that this will work?

🗪

1:08:23robrobby The work queue will take any function to do it multithreaded? Does the function need to be special so that this will work?

🗪

1:08:53waterlimon Please write a lock free queue, even though I don't know what those are and if you used one

🗪

1:08:53waterlimon Please write a lock free queue, even though I don't know what those are and if you used one

🗪

1:08:53waterlimon Please write a lock free queue, even though I don't know what those are and if you used one

🗪

1:09:33boogie0815 How many CPU cycles does spawning a thread cost? Or better: what's the minimum amount of cycles to work in 2 threads to gain speed?

🗪

1:09:33boogie0815 How many CPU cycles does spawning a thread cost? Or better: what's the minimum amount of cycles to work in 2 threads to gain speed?

🗪

1:09:33boogie0815 How many CPU cycles does spawning a thread cost? Or better: what's the minimum amount of cycles to work in 2 threads to gain speed?

🗪

1:10:45gasto5 I don't understand why you call it a queue if it is done potentially simultaneously

🗪

1:10:45gasto5 I don't understand why you call it a queue if it is done potentially simultaneously

🗪

1:10:45gasto5 I don't understand why you call it a queue if it is done potentially simultaneously

🗪

1:11:46flyingwafflenyc Wasn't there already a bit of thread-related code in the win32 file?

🗪

1:11:46flyingwafflenyc Wasn't there already a bit of thread-related code in the win32 file?

🗪

1:11:46flyingwafflenyc Wasn't there already a bit of thread-related code in the win32 file?

🗪

1:12:04waterlimon Will you add a cool graph over time that shows what task (e.g. from which subsystem) each thread is working on at each moment?

🗪

1:12:04waterlimon Will you add a cool graph over time that shows what task (e.g. from which subsystem) each thread is working on at each moment?

🗪

1:12:04waterlimon Will you add a cool graph over time that shows what task (e.g. from which subsystem) each thread is working on at each moment?

🗪

1:14:19zuurr_ Is false sharing between the entries in the work queue potentially problematic (from a performance standpoint)?

🗪

1:14:19zuurr_ Is false sharing between the entries in the work queue potentially problematic (from a performance standpoint)?

🗪

1:14:19zuurr_ Is false sharing between the entries in the work queue potentially problematic (from a performance standpoint)?

🗪

1:14:34popcorn0x90 Does volatile clear the assembly registers by pushing them into the stack and then restore by popping?

🗪

1:14:34popcorn0x90 Does volatile clear the assembly registers by pushing them into the stack and then restore by popping?

🗪

1:14:34popcorn0x90 Does volatile clear the assembly registers by pushing them into the stack and then restore by popping?

🗪

1:14:56grubuck Why would you want a compiler fence and not a process fence, and vice versa?

🗪

1:14:56grubuck Why would you want a compiler fence and not a process fence, and vice versa?

🗪

1:14:56grubuck Why would you want a compiler fence and not a process fence, and vice versa?

🗪

1:15:41Pseudonym73 Shouldn't _mm_fence() imply a compiler fence? Surely there's no point otherwise...

🗪

1:15:41Pseudonym73 Shouldn't _mm_fence() imply a compiler fence? Surely there's no point otherwise...

🗪

1:15:41Pseudonym73 Shouldn't _mm_fence() imply a compiler fence? Surely there's no point otherwise...

🗪

1:16:44thordura Have you implemented friction?

🗪

1:16:44thordura Have you implemented friction?

🗪

1:16:44thordura Have you implemented friction?

🗪

1:16:54jameswidman So, thread management is a bit like memory management (in that you want to set it up ahead of time rather than allocating them on-demand)

🗪

1:16:54jameswidman So, thread management is a bit like memory management (in that you want to set it up ahead of time rather than allocating them on-demand)

🗪

1:16:54jameswidman So, thread management is a bit like memory management (in that you want to set it up ahead of time rather than allocating them on-demand)

🗪

1:17:57robrobby The code to ask how many threads will be done simultaneously by the processor is to be added?

🗪

1:17:57robrobby The code to ask how many threads will be done simultaneously by the processor is to be added?

🗪

1:17:57robrobby The code to ask how many threads will be done simultaneously by the processor is to be added?

🗪

1:18:35alephant Is it possible that a work queue entry spawn another work queue entry?

🗪

1:18:35alephant Is it possible that a work queue entry spawn another work queue entry?

🗪

1:18:35alephant Is it possible that a work queue entry spawn another work queue entry?

🗪

1:19:26zuurr_ (Not an expert at all, which was part of why I asked the question) False sharing causes the processor to skip the cache when different threads access stuff on the same cache line

🗪

1:19:26zuurr_ (Not an expert at all, which was part of why I asked the question) False sharing causes the processor to skip the cache when different threads access stuff on the same cache line

🗪

1:19:26zuurr_ (Not an expert at all, which was part of why I asked the question) False sharing causes the processor to skip the cache when different threads access stuff on the same cache line

🗪

1:21:41Pseudonym73 Actually, I can think of one use case for a compiler fence without a memory fence: writing to CPU special registers like control registers or MSRs

🗪

1:21:41Pseudonym73 Actually, I can think of one use case for a compiler fence without a memory fence: writing to CPU special registers like control registers or MSRs

🗪

1:21:41Pseudonym73 Actually, I can think of one use case for a compiler fence without a memory fence: writing to CPU special registers like control registers or MSRs

🗪

1:22:34kemosabe76 Getting back into C/C++ coding after many years. Don't know why you are mixing C-style structs and C++ structs?

🗪

1:22:34kemosabe76 Getting back into C/C++ coding after many years. Don't know why you are mixing C-style structs and C++ structs?

🗪

1:22:34kemosabe76 Getting back into C/C++ coding after many years. Don't know why you are mixing C-style structs and C++ structs?

🗪

1:22:52Close things down

🗩

1:22:52Close things down

🗩

1:22:52Close things down

🗩

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