| 30 Nov 2021 |
 eris | id personally put 0,0 at the top left since thats where most people look first | 14:45:31 |
| eris | and also because | 14:45:56 |
 loke | Whoops. You're right. | 14:46:11 |
| eris | 
Download image.png | 14:48:00 |
| eris | yea | 14:48:03 |
| loke | The generation of the SVG got it wrong. Easy to fix. | 14:48:25 |
| loke | OK, fixed the 2-dimensional diagram | 14:58:32 |
 Adám#2156 | 'aeiou' are already normalised, so you don't need ⎕C to apply to the left argument of ∊ | 15:02:55 |
|  maybe William changed their display name from maybe William to maybe William#2431. | 18:23:12 |
| maybe William changed their display name from maybe William#2431 to maybe William. | 18:23:13 |
| maybe William | if i dont recall there exsists an NGN in here, assuming this NGN is the same as http://archive.vector.org.uk/art10501160 i were wondering, how does your Strided representation of arrays stack up when it comes to SIMD instructions and cache misses [of the L1 cache]? | 18:23:13 |
 Marshall | ngn/apl ended up dropping that strided representation and using shape only. NumPy uses strides though. | 18:28:59 |
| Marshall | Depends on the operation, but a ravel-ordered array is easier to handle in a cache-friendly way than a strided one. In particular, if the strided array is a scattered selection from a larger array then accessing its elements inherently breaks cache locality. | 18:32:28 |
| Marshall | The advantage of a strided representation is that you can skip many kinds of operations, that is, things like reverse, drop, and transpose can be done only by manipulating metadata. The problem is that later operations have to have those transformations built in somehow. | 18:34:10 |
| Marshall | Transpose can be done in blocks to make optimal use of the cache; fusing a blocked transpose with another operation would be pretty hard and I doubt NumPy does it. I have no idea whether they have block transpose code in the first place. Dyalog only does that for booleans. | 18:38:06 |
 ngn | In reply to @_discord_346616332726304768:t2bot.io
if i dont recall there exsists an NGN in here, assuming this NGN is the same as http://archive.vector.org.uk/art10501160 i were wondering, how does your Strided representation of arrays stack up when it comes to SIMD instructions and cache misses [of the L1 cache]? me exists :) | 18:42:19 |
| ngn | js ÷ l2 cache ≈ snail ÷ aerodynamics | 18:43:11 |
| eris | snail :) | 18:44:38 |
| maybe William | awesome thanks a lot, this cleares up a lot, i guess ill just end up implementing a system like classes in haskell or traits in rust where i can have multiple representations work in different ways. i guess that would be the best solution, thanks for the information, this was really interesting | 18:47:46 |
| eris | 
Download screenshot_20211130-184849_fennec.png | 18:49:01 |
| eris | this is cool :) | 18:49:06 |
| eris | so like, with
1 2 3
4 5 6
7 8 9 | 18:49:55 |
| ngn | eris: an idea stolen from numpy. marshall described its drawbacks well. | 18:50:03 |
| eris | what would the stride be | 18:50:05 |
| ngn | eris: 1 3 | 18:50:23 |
| ngn | or 3 1 | 18:50:33 |
| eris | huh | 18:50:41 |
| eris | ok | 18:50:43 |
| eris | i get the meaning of axis now :) | 18:50:51 |
| eris | took a while......... | 18:51:32 |
