Skip to content

KernelInterface - #1047

Closed
christiangnrd wants to merge 13 commits into
JuliaGPU:mainfrom
christiangnrd:interface
Closed

KernelInterface#1047
christiangnrd wants to merge 13 commits into
JuliaGPU:mainfrom
christiangnrd:interface

Conversation

@christiangnrd

@christiangnrd christiangnrd commented Aug 22, 2026 •
edited
Loading

Copy link
Copy Markdown
Member

Do not merge until KernelInterface has been reviewed and interface fully decided

@christiangnrd
christiangnrd force-pushed the interface branch 5 times, most recently from f95ea6d to 744d030 Compare August 22, 2026 19:57
@christiangnrd
christiangnrd force-pushed the interface branch 3 times, most recently from 8bc556f to c991fed Compare September 5, 2026 22:00
github-actions[bot]
github-actions Bot reviewed Sep 6, 2026
View reviewed changes

@github-actions github-actions Bot left a comment •
edited
Loading

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

AMDGPU.jl Benchmarks

Details
Benchmark suite Current: 21fc97c Previous: 482cd05 Ratio
amdgpu/synchronization/context/device 557.5 ns 545 ns 1.02
amdgpu/synchronization/stream/blocking 227.5 ns 225 ns 1.01
amdgpu/synchronization/stream/nonblocking 305 ns 305 ns 1
applications/bitonic_sort 1102303.5 ns 1190667 ns 0.93
applications/convolution 74421 ns 97316.5 ns 0.76
applications/floyd_warshall 9021578 ns 9117782.75 ns 0.99
applications/histogram 803131.75 ns 812499.25 ns 0.99
applications/prefix_sum 190235.25 ns 186780 ns 1.02
array/accumulate/Float32/1d 81428.5 ns 81956.25 ns 0.99
array/accumulate/Float32/dims=1 267966.5 ns 270074 ns 0.99
array/accumulate/Float32/dims=1L 90609 ns 91604 ns 0.99
array/accumulate/Float32/dims=2 103084 ns 89373.75 ns 1.15
array/accumulate/Float32/dims=2L 3030614.5 ns 3029106.5 ns 1.00
array/accumulate/Int64/1d 83953.5 ns 82571.25 ns 1.02
array/accumulate/Int64/dims=1 249953.75 ns 246918.5 ns 1.01
array/accumulate/Int64/dims=1L 88306.25 ns 87951.25 ns 1.00
array/accumulate/Int64/dims=2 90346.25 ns 76646.25 ns 1.18
array/accumulate/Int64/dims=2L 3321223.75 ns 3388611.75 ns 0.98
array/broadcast 74498.5 ns 73561 ns 1.01
array/construct 2127.5 ns 2072.75 ns 1.03
array/copy 39313 ns 38445.5 ns 1.02
array/copyto!/cpu_to_gpu 89446.25 ns 89271.25 ns 1.00
array/copyto!/gpu_to_cpu 89489 ns 89381.25 ns 1.00
array/copyto!/gpu_to_gpu 33943 ns 33858 ns 1.00
array/iteration/findall/bool 131377 ns 127474.25 ns 1.03
array/iteration/findall/int 147089.5 ns 140999.5 ns 1.04
array/iteration/findfirst/bool 142817 ns 155232.25 ns 0.92
array/iteration/findfirst/int 164550 ns 156597.25 ns 1.05
array/iteration/findmin/1d 103819 ns 103336.5 ns 1.00
array/iteration/findmin/2d 90283.75 ns 90036.25 ns 1.00
array/iteration/logical 226783.25 ns 224733.25 ns 1.01
array/iteration/scalar 293996.75 ns 295471.75 ns 1.00
array/permutedims/2d 69976 ns 69488.5 ns 1.01
array/permutedims/3d 49540.75 ns 55823.25 ns 0.89
array/permutedims/4d 61131 ns 73458.75 ns 0.83
array/random/rand/Float32 45098.25 ns 44653.25 ns 1.01
array/random/rand/Int64 54625.75 ns 54085.75 ns 1.01
array/random/rand!/Float32 41005.5 ns 41163.25 ns 1.00
array/random/rand!/Int64 49220.75 ns 56780.75 ns 0.87
array/random/randn/Float32 70828.5 ns 69911 ns 1.01
array/random/randn!/Float32 56520.75 ns 55493.25 ns 1.02
array/reductions/mapreduce/Float32/1d 81476.25 ns 81456.25 ns 1.00
array/reductions/mapreduce/Float32/dims=1 86468.75 ns 86601.25 ns 1.00
array/reductions/mapreduce/Float32/dims=1L 837757.25 ns 841199.75 ns 1.00
array/reductions/mapreduce/Float32/dims=2 84998.75 ns 81728.5 ns 1.04
array/reductions/mapreduce/Float32/dims=2L 135447 ns 136954.5 ns 0.99
array/reductions/mapreduce/Int64/1d 81181.25 ns 77736 ns 1.04
array/reductions/mapreduce/Int64/dims=1 86361.25 ns 80831.25 ns 1.07
array/reductions/mapreduce/Int64/dims=1L 842372.25 ns 844804.75 ns 1.00
array/reductions/mapreduce/Int64/dims=2 72793.5 ns 75918.5 ns 0.96
array/reductions/mapreduce/Int64/dims=2L 137424.5 ns 137289.5 ns 1.00
array/reductions/reduce/Float32/1d 81683.75 ns 77718.5 ns 1.05
array/reductions/reduce/Float32/dims=1 86538.75 ns 75563.5 ns 1.15
array/reductions/reduce/Float32/dims=1L 850340 ns 841087 ns 1.01
array/reductions/reduce/Float32/dims=2 82871.25 ns 76336 ns 1.09
array/reductions/reduce/Float32/dims=2L 136382 ns 136187 ns 1.00
array/reductions/reduce/Int64/1d 80938.75 ns 81263.5 ns 1.00
array/reductions/reduce/Int64/dims=1 69996 ns 76798.5 ns 0.91
array/reductions/reduce/Int64/dims=1L 841349.75 ns 842952.25 ns 1.00
array/reductions/reduce/Int64/dims=2 73901.25 ns 76308.75 ns 0.97
array/reductions/reduce/Int64/dims=2L 137589.5 ns 137287 ns 1.00
array/reverse/1d 42835.75 ns 35455.5 ns 1.21
array/reverse/1dL 70628.5 ns 70423.5 ns 1.00
array/reverse/1dL_inplace 54688.5 ns 54620.75 ns 1.00
array/reverse/1d_inplace 35825.5 ns 35605.5 ns 1.01
array/reverse/2d 46635.75 ns 46543.25 ns 1.00
array/reverse/2dL 88466.25 ns 88903.75 ns 1.00
array/reverse/2dL_inplace 65548.5 ns 65341 ns 1.00
array/reverse/2d_inplace 31928 ns 36125.5 ns 0.88
array/sorting/1d 327152.25 ns 327314.75 ns 1.00
gemm/tiled 1947713.25 ns 1970010.75 ns 0.99
gemm/tiled_unbounded 2905917 ns 1968730.75 ns 1.48
integration/byval/reference 39401 ns 39211 ns 1.00
integration/byval/slices=1 39810 ns 40181 ns 0.99
integration/byval/slices=2 160742 ns 144092 ns 1.12
integration/byval/slices=3 238863 ns 234554 ns 1.02
integration/volumerhs 4879730 ns 4897830 ns 1.00
kernel/indexing 34285.5 ns 28278 ns 1.21
kernel/indexing_checked 50158.25 ns 33610.5 ns 1.49
kernel/launch 1167.5 ns 1162.5 ns 1.00
kernel/rand 71286 ns 96961.25 ns 0.74
latency/import 1451933250 ns 1444012666 ns 1.01
latency/precompile 22930534917 ns 22844392889 ns 1.00
latency/ttfp 2224689624 ns 2200398894 ns 1.01
stencil/diffusion3d 1603250.75 ns 1621100.75 ns 0.99
stencil/diffusion3d_checked 1658281.5 ns 1656716.25 ns 1.00

This comment was automatically generated by workflow using github-action-benchmark.

@christiangnrd
christiangnrd force-pushed the interface branch 4 times, most recently from 7fe45f5 to 4d75db5 Compare September 12, 2026 15:58
@christiangnrd
christiangnrd force-pushed the interface branch 2 times, most recently from a3c00b8 to 9a165fe Compare September 22, 2026 12:58
@christiangnrd
christiangnrd marked this pull request as ready for review September 22, 2026 20:14
@christiangnrd
christiangnrd force-pushed the interface branch 3 times, most recently from a32063f to db60d30 Compare September 26, 2026 18:13
@maleadt
KernelInterface: report per-dimension launch limits
7f91c42 
Implement KI.max_work_group_dims from the device's maxThreadsDim, cached in
HIPDevice since it is queried on every automatically-sized launch, and
KI.max_num_groups as the number of workgroups that fits the work-item grid
limit for any valid workgroup size. Requires KernelInterface 0.2.3.
@maleadt
Port to KernelInterface 0.3
5d52333 
- subtype `KI.Backend`, and implement `KI.launch` instead of calling the kernel;
- implement the four primitive index queries with `% T`, and let KernelInterface derive
  the global ones;
- typed sub-group queries, `supports_subgroups` and `supports_shuffle`;
- `max_work_group_size(kernel)` is the kernel's limit, `launch_configuration` the
  occupancy-based recommendation;
- report `Float64` and atomics support, and the device of an array;
- `kernel_function` keeps the backend it was given;
- `copyto!` checks the lengths and returns the destination;
- use the generic `zeros` and `ones`;
- `_print` is documented as unsupported.

KernelInterface 0.3 isn't registered yet, so get it from its branch, and develop it from a
clone on Julia 1.10, which ignores `[sources]`.
@maleadt
KernelInterface: fix the sub-group queries for partial and divergent …
e589274 
…wavefronts

- `get_num_sub_groups` counts a partial last wavefront (`cld` instead of `÷`);
- `get_sub_group_size` is the number of work-items in the wavefront, which is smaller than
  the wavefront size for the partial one;
- `get_sub_group_local_id` is the hardware lane (`mbcnt`), not the index among the active
  lanes (`activelane`), which changes in divergent code.
@maleadt
Order memory in sync_wavefront
e651812 
`llvm.amdgcn.wave.barrier` only keeps the compiler from moving code across it, so writes
before it weren't guaranteed to be visible to the other lanes afterwards. Fence it at
wavefront scope, like `sync_workgroup` does at workgroup scope. This makes
`KernelInterface.sub_group_barrier` fence memory, as KernelInterface 0.3 requires.
@maleadt
KernelInterface: guarantee the wavefront size of compiled kernels
20a179e 
`KI.sub_group_size` promises that kernels from `KI.kernel_function` execute with that
sub-group width. `hipfunction` compiles for the device's wavefront size by default, but
`wavefrontsize64` could override it; reject a conflicting value.
@maleadt maleadt mentioned this pull request Sep 28, 2026
Merged
@maleadt
KernelInterface: specialize KI.launch on its arguments
21fc97c 
Julia doesn't specialize a method on `args...` that it only passes through, which
made every launch through KernelInterface's generic launch dispatch dynamically
(+1.2 µs and +1 kB per launch on CUDA).
@maleadt
Accept the problem size in KI.launch_configuration
f915ef1 
KernelInterface 0.3 passes the number of work-items to launch as `nitems`,
separately from the bound on the work-group size.
@maleadt maleadt mentioned this pull request Sep 30, 2026
Open
@maleadt

maleadt commented Sep 30, 2026

Copy link
Copy Markdown
Member

#1117

@maleadt maleadt closed this Sep 30, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@github-actions github-actions[bot] github-actions[bot] left review comments

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

2 participants

@christiangnrd @maleadt