I'd be lying if I said I understood very much of this. But I'm interested and I'll keep watching your videos. I would find it useful if you would frequently refer your code to a chart to show what you are doing. Like a chart that shows Command Queues, Command Buffers, drawables, Render Pass Descriptors, etc. "This code is our buffer; this is how it relates to the chart." Lotta work, I know. Your enthusiasm is infectious.
Bro this is an awesome video, thank you so much for taking the time. What a great introduction to Metal! Would love to see you back with more new tutorials now that Metal3 is around
This is absolutely the best Metal introductory tutorial ever! I was looking for tuts to start getting my head around Metal and all of them went over my head. This didn't. Thanks. :)
the quality of your content become so much better. And i know that, because i was watching your videos right from the first one (and i've seen your backflip)
Hey!! Very good explanation of the MTL basics! I like the hurry-through-the-code way of presenting. If I don’t understand something, I can stop the video at any point 😊 looking forward to the next episodes 😀
Hi, I’m new to this area. I only had experience in cuda. Does M1 Pro/Max work the same way as Radeon does on MBP 2019-? I’m confused with the merged vRAM of silicon platform and does the 10 GPU cores of M-processors accelerate parallel computing like cuda does with its thousands of cuda cores?
Wow, thanks for the tutorial! That seems to be a pretty precise way to control a gpu. Was feeling more nerdy than I prefer to while watching though. Haha.
i went over your videos months ago, but this time, i'm not going to give up to confusion. to those not in a graphics rendering or game dev background, this stuff seems so damned foreign..but stay strong, everyone!
Hi. How do I set up metal to draw over the previous frame? For example, 1000 fixed triangles of different colors are drawn. And in each new frame, need to change the color of only 10 random triangle. I want to draw 10 triangle with the desired color and draw it on top of the previous frame.
Hi, thanks for these awesome videos, hopefully you're still active! I plan on using the shader to update an array using two other arrays similar to what you've described. I can't quite figure out how to get the full updated array back from the buffer after processing, to use on the CPU. trying to print the array after binding to memory just brings up the following. Hope that makes sense I'm fairly new to this!
Thank you! Very very nice video! I strongly believe that this is the only video / tutorial that explains loud and clear how to use GPU and metal to performs calulations instead of rendering! I have some issues - however - undestanding how to bind back complex data structures as results in Swift; in C with pointer is quite easy but in Swift is a bit tricky; can you point me to some examples? Thanks!
MTLValidateFeatureSupport:6508: failed assertion `Dispatch Threads with Non-Uniform Threadgroup Size is not supported on this device' When I am trying to run it on the simulator iPhone 15. What could be wrong?
Did you try using the new m1 pro/max/ultra with Metal? I would really like to know how the unified memory architecture comes into play when it comes to sending data back and forth between cpu and gpu.
@@2etime Supposedly Apple uses a new unified architecture so that GPU and CPU both share memory. Does that affect how you would program for Metal? Especially when it comes to copying data / results. How does the performance compare to your previous machine? Do you feel the benchmarks are representative of how slow it is compared to NVIDIA?
Am I right in thinking that for processing a 2d texture, you would want the largest number of threads that will fit 'square' inside of each max thread group. Is this done to ensure that the uint2 thread coords align with texture's x,y coords? Or is it just as convenient/efficient to create thread groups that might be 'retangular'?
Copy/paste from apple documentation : "Always use a multiple of a type’s stride instead of its size when allocating memory or accounting for the distance between instances in memory."
@@2etime I think @platin21 is talking about the SIMD matrix computation from the Accelerate framework - developer.apple.com/documentation/accelerate/simd
When I am trying to run main.swift, it turns out to show "main/main.swift:31: Fatal error: Unexpectedly found nil while unwrapping an Optional value", do you have any idea why this happens?
i would make it a 1D array for sure. The information will be better if it is contiguous. If you want to access a value in the buffer / array you can use this formula where width is the column count of the 2d version.
I am new to MSL ios, I have a windows laptop, but I need to work on MSL , can anyone help me know which IDE should i be using now to start coding in metal?
@@2etime gr8 i found the plugin for msl for vsc. But , may be it is a dumb question but i don't know how do i run my project, any specific cli for vsc ?
I would suggest learning OpenGL on windows! My videos will give you a high level understanding of graphics as a whole. OpenGL and metal are very similar regarding the pipeline of the gpu
The 3+ seconds for the CPU based componentwise addition of two arrays of 3 million numbers is some orders of magnitude too long, why is that? I get 0.02 seconds in JavaScript (not even a statically typed language, and time includes some interpreted execution and warmed up JIT compilation phases too): a = [...new Array(3e6)].map(Math.random); b = [...new Array(3e6)].map(Math.random); (t0 = performance.now(), result=new Array(3e6).fill(0), (() => {for(let i = 0; i < 3e6; i++) result[i]=a[i]+b[i]})(), console.log((performance.now() - t0)/1000), result);
interesting... My discord would love this question lol. I tested it out as well and seem to be getting pretty quick numbers in js too. computers are awesome! I think my main point in the video is running multiple functions at the same time, as opposed to linearly or single instruction multiple data (simd). www.heavy.ai/technical-glossary/cpu-vs-gpu
Awesome! Please don't stop making these, your videos are so underrated!
I'd be lying if I said I understood very much of this. But I'm interested and I'll keep watching your videos. I would find it useful if you would frequently refer your code to a chart to show what you are doing. Like a chart that shows Command Queues, Command Buffers, drawables, Render Pass Descriptors, etc. "This code is our buffer; this is how it relates to the chart." Lotta work, I know. Your enthusiasm is infectious.
Ah ok will do!
This is for you guys and if a chart helps I can do that!
Bro this is an awesome video, thank you so much for taking the time. What a great introduction to Metal! Would love to see you back with more new tutorials now that Metal3 is around
This is absolutely the best Metal introductory tutorial ever! I was looking for tuts to start getting my head around Metal and all of them went over my head.
This didn't.
Thanks. :)
the quality of your content become so much better. And i know that, because i was watching your videos right from the first one
(and i've seen your backflip)
You, sir, are a gentleman and a scholar. Thank you.
Hey!! Very good explanation of the MTL basics! I like the hurry-through-the-code way of presenting. If I don’t understand something, I can stop the video at any point 😊 looking forward to the next episodes 😀
Thank you very much! This is a really clear and good explanation about how we use GPU in Metal. Keep going! You're the best!
Thank you so much! The last time I saw such a precious video like this was about 10 years ago.
What was that video about? Lol
You are awesome. Thank you for creating such a wonderful intro.
Thank u so much! Metal is suqch lowkey framework and u explained the compute functions very well!
this is just a brilliant tutorial. thanks man!
Very nice. Love the by example approach.
you fucking god. this is the best metal channel on YT.
Top 5 favorite comments
Hi, I’m new to this area. I only had experience in cuda. Does M1 Pro/Max work the same way as Radeon does on MBP 2019-? I’m confused with the merged vRAM of silicon platform and does the 10 GPU cores of M-processors accelerate parallel computing like cuda does with its thousands of cuda cores?
This is fantastic! Thanks for sharing
Watching you teach is making my wave function collapse.
Is that a good thing?
Wow, thanks for the tutorial! That seems to be a pretty precise way to control a gpu. Was feeling more nerdy than I prefer to while watching though. Haha.
i went over your videos months ago, but this time, i'm not going to give up to confusion.
to those not in a graphics rendering or game dev background, this stuff seems so damned foreign..but stay strong, everyone!
Feel free to reach out if you need some help :)
Hi. How do I set up metal to draw over the previous frame? For example, 1000 fixed triangles of different colors are drawn. And in each new frame, need to change the color of only 10 random triangle. I want to draw 10 triangle with the desired color and draw it on top of the previous frame.
Hi, thanks for these awesome videos, hopefully you're still active! I plan on using the shader to update an array using two other arrays similar to what you've described. I can't quite figure out how to get the full updated array back from the buffer after processing, to use on the CPU. trying to print the array after binding to memory just brings up the following. Hope that makes sense I'm fairly new to this!
Brilliant video!
Hi, I love the videos. However, I specialize in cpp. Can you release a video of this but as a cpp (or I guess objective cpp) rather than a swift file?
Could you also point me in the direction of resources that would help?
well produced video!
Thank you! Very very nice video! I strongly believe that this is the only video / tutorial that explains loud and clear how to use GPU and metal to performs calulations instead of rendering! I have some issues - however - undestanding how to bind back complex data structures as results in Swift; in C with pointer is quite easy but in Swift is a bit tricky; can you point me to some examples? Thanks!
do you mean how to access the data into a Type on the swift side?
MTLValidateFeatureSupport:6508: failed assertion `Dispatch Threads with Non-Uniform Threadgroup Size is not supported on this device'
When I am trying to run it on the simulator iPhone 15. What could be wrong?
Thank you so much.
Did you try using the new m1 pro/max/ultra with Metal? I would really like to know how the unified memory architecture comes into play when it comes to sending data back and forth between cpu and gpu.
I’m getting one in a week and plan on checking that out asap :)
Ok… It works really well! My computer runs metal so much faster. However it could just be the upgrade! What specs do you want? :)
@@2etime Supposedly Apple uses a new unified architecture so that GPU and CPU both share memory. Does that affect how you would program for Metal? Especially when it comes to copying data / results. How does the performance compare to your previous machine? Do you feel the benchmarks are representative of how slow it is compared to NVIDIA?
Am I right in thinking that for processing a 2d texture, you would want the largest number of threads that will fit 'square' inside of each max thread group. Is this done to ensure that the uint2 thread coords align with texture's x,y coords? Or is it just as convenient/efficient to create thread groups that might be 'retangular'?
Simply Awesome !!!!!!!!!!!
Awesome :D
Copy/paste from apple documentation : "Always use a multiple of a type’s stride instead of its size when allocating memory or accounting for the distance between instances in memory."
What did I say?
@@2etime when you allocate MTLBuffer your code use size instead of stride
with simple type, size and stride are probably the same anyway. But still :)
Ah! Yes I know this now. I hate the permanence of TH-cam :(
@@2etime hehehe :D
This is gold
Your metal shading language playlist has only video :(
Hey! Just a comment: the bindMemory(to:capacity:) expects capacity as the number of T elements, not bytes. (at 20:10, line 81)
Oh nice! Thank you!
So ideally we would also add a simd way just for good measure (by directly using the instrinsics).
I’m not sure how to do that. Can you elaborate a little bit for me?
@@2etime I think @platin21 is talking about the SIMD matrix computation from the Accelerate framework - developer.apple.com/documentation/accelerate/simd
Can I use it to render image frames into video ? I never use metal before
When I am trying to run main.swift, it turns out to show "main/main.swift:31: Fatal error: Unexpectedly found nil while unwrapping an Optional value", do you have any idea why this happens?
How should I make a buffer for 2-D array? Do I have to concatenate it into 1D array?
i would make it a 1D array for sure. The information will be better if it is contiguous. If you want to access a value in the buffer / array you can use this formula where width is the column count of the 2d version.
@@2etime Thank you.
I am new to MSL ios, I have a windows laptop, but I need to work on MSL , can anyone help me know which IDE should i be using now to start coding in metal?
You could prob use visual studio and use c++!
@@2etime gr8 i found the plugin for msl for vsc. But , may be it is a dumb question but i don't know how do i run my project, any specific cli for vsc ?
I do not think you can run a project with a Windows machine. Hopefully somebody who knows more about that will post.
@@2etime oh ok .. I thought so. Thank you so much for the reply
I would suggest learning OpenGL on windows! My videos will give you a high level understanding of graphics as a whole. OpenGL and metal are very similar regarding the pipeline of the gpu
GPU is executing multiple tasks all at once
CPU is executing multiple tasks one at a time
The 3+ seconds for the CPU based componentwise addition of two arrays of 3 million numbers is some orders of magnitude too long, why is that? I get 0.02 seconds in JavaScript (not even a statically typed language, and time includes some interpreted execution and warmed up JIT compilation phases too):
a = [...new Array(3e6)].map(Math.random);
b = [...new Array(3e6)].map(Math.random);
(t0 = performance.now(), result=new Array(3e6).fill(0), (() => {for(let i = 0; i < 3e6; i++) result[i]=a[i]+b[i]})(), console.log((performance.now() - t0)/1000), result);
interesting... My discord would love this question lol. I tested it out as well and seem to be getting pretty quick numbers in js too. computers are awesome! I think my main point in the video is running multiple functions at the same time, as opposed to linearly or single instruction multiple data (simd). www.heavy.ai/technical-glossary/cpu-vs-gpu
I'm pretty sure you don't need to explain that we can do general purpose computation on GPU in 2020+ ;)
it's still a good tutorial of course. Thank you