Add register-tiled GEMM and 3D diffusion stencil benchmarks

[gbaraldi]

Adds two real-workload benchmarks distilled from an ISA-level analysis of Oceananigans/ParallelStencil/Trixi kernels on MI300A, chosen because each pins a codegen behavior the existing suite doesn't cover:

• gemm/tiled and gemm/tiled_unbounded: register-tiled FP32 GEMM (16×16 workgroup, 8×8 accumulator microtile, LDS-staged). Register-bound: the pair tracks launch-bounds support (Support launch bounds (@roc maxthreads) and set them for static KernelAbstractions workgroups #1060) — on MI300A the bound is worth +76% (15.96 → 28.11 TFLOPS at 4096³) because the default 1,1024 flat-workgroup-size caps the register budget at 128 VGPRs and forces spills.
• stencil/diffusion3d and stencil/diffusion3d_checked: 7-point Float64 diffusion step, 256³. Short and bandwidth-bound, so it is dominated by per-wave entry costs — it regresses ~3× if workgroup dimension reads fall back to uncached dispatch-packet loads (the Read workgroup/grid dimensions from hidden kernargs instead of the dispatch packet (2x on short kernels) #1058 sentinel), and the checked variant tracks the bounds-check control-flow overhead.

Kernel correctness and timings were validated on MI300A (gfx942, ROCm 7.2.4).

🤖 Generated with Claude Code

https://claude.ai/code/session_01FJL9JQhcpXUVWkhuuiA7ad

[gbaraldi]

Added the rocm-examples Applications ports as applications/: histogram (LDS byte counters), prefix_sum (multi-launch scan sweep), convolution (kernarg mask), bitonic_sort (210 launches) and floyd_warshall (1024 dependent launches — the sharpest launch-overhead sentinel). Full suite smoke-tested on MI300A: gemm 1.00/1.57 ms (bounded/unbounded), stencil 0.44/0.49 ms, histogram 1.00 ms, convolution 0.054 ms, prefix_sum 0.155 ms, bitonic 3.7 ms, floyd 14.7 ms.

[github-actions]

AMDGPU.jl Benchmarks

Details

Benchmark suite	Current: 047b3a5	Previous: 7cf1826	Ratio
amdgpu/synchronization/context/device	560 ns	535 ns	1.05
amdgpu/synchronization/stream/blocking	235 ns	230 ns	1.02
amdgpu/synchronization/stream/nonblocking	317.5 ns	307.5 ns	1.03
applications/bitonic_sort	1651043.75 ns
applications/convolution	107264 ns
applications/floyd_warshall	11770948.75 ns
applications/histogram	791381.25 ns
applications/prefix_sum	251868.75 ns
array/accumulate/Float32/1d	76623.5 ns	77873.75 ns	0.98
array/accumulate/Float32/dims=1	248063.5 ns	272711.5 ns	0.91
array/accumulate/Float32/dims=1L	80976.25 ns	91106.5 ns	0.89
array/accumulate/Float32/dims=2	71406 ns	87163.75 ns	0.82
array/accumulate/Float32/dims=2L	2762639.75 ns	2756483.25 ns	1.00
array/accumulate/Int64/1d	77811.25 ns	81851.25 ns	0.95
array/accumulate/Int64/dims=1	242673.5 ns	245181.25 ns	0.99
array/accumulate/Int64/dims=1L	83988.75 ns	83678.75 ns	1.00
array/accumulate/Int64/dims=2	85726 ns	90186.5 ns	0.95
array/accumulate/Int64/dims=2L	2897179.25 ns	2898395.25 ns	1.00
array/broadcast	73043.75 ns	73666.25 ns	0.99
array/construct	2235 ns	2057.5 ns	1.09
array/copy	33183 ns	37383 ns	0.89
array/copyto!/cpu_to_gpu	111814 ns	111264 ns	1.00
array/copyto!/gpu_to_cpu	111966.75 ns	111711.75 ns	1.00
array/copyto!/gpu_to_gpu	53695.75 ns	59595.75 ns	0.90
array/iteration/findall/bool	135384.25 ns	139419.75 ns	0.97
array/iteration/findall/int	137344.5 ns	149934.75 ns	0.92
array/iteration/findfirst/bool	184492.5 ns	146172.25 ns	1.26
array/iteration/findfirst/int	162022.5 ns	145079.75 ns	1.12
array/iteration/findmin/1d	118321.5 ns	123587 ns	0.96
array/iteration/findmin/2d	108454.25 ns	113761.75 ns	0.95
array/iteration/logical	245131 ns	237901.25 ns	1.03
array/iteration/scalar	287591.75 ns	298014.5 ns	0.97
array/permutedims/2d	71481 ns	71631 ns	1.00
array/permutedims/3d	70581 ns	70726 ns	1.00
array/permutedims/4d	60888.5 ns	73603.75 ns	0.83
array/random/rand/Float32	45635.75 ns	45208 ns	1.01
array/random/rand/Int64	54348.25 ns	54288.25 ns	1.00
array/random/rand!/Float32	64160.75 ns	65161 ns	0.98
array/random/rand!/Int64	71556 ns	66998.5 ns	1.07
array/random/randn/Float32	79106 ns	79741.25 ns	0.99
array/random/randn!/Float32	80946.25 ns	81228.75 ns	1.00
array/reductions/mapreduce/Float32/1d	97088.75 ns	87228.75 ns	1.11
array/reductions/mapreduce/Float32/dims=1	83813.75 ns	75168.75 ns	1.12
array/reductions/mapreduce/Float32/dims=1L	834459.5 ns	831937.25 ns	1.00
array/reductions/mapreduce/Float32/dims=2	83818.75 ns	88813.75 ns	0.94
array/reductions/mapreduce/Float32/dims=2L	144232 ns	142234.5 ns	1.01
array/reductions/mapreduce/Int64/1d	96833.75 ns	101484 ns	0.95
array/reductions/mapreduce/Int64/dims=1	83708.5 ns	91746.5 ns	0.91
array/reductions/mapreduce/Int64/dims=1L	832542 ns	832082.5 ns	1.00
array/reductions/mapreduce/Int64/dims=2	83903.75 ns	90866.25 ns	0.92
array/reductions/mapreduce/Int64/dims=2L	144754.5 ns	217318.25 ns	0.67
array/reductions/reduce/Float32/1d	96846.25 ns	101954 ns	0.95
array/reductions/reduce/Float32/dims=1	84021.25 ns	75456.25 ns	1.11
array/reductions/reduce/Float32/dims=1L	832932 ns	829999.75 ns	1.00
array/reductions/reduce/Float32/dims=2	84116.25 ns	89351.25 ns	0.94
array/reductions/reduce/Float32/dims=2L	143159.75 ns	143432 ns	1.00
array/reductions/reduce/Int64/1d	89141.25 ns	101771.5 ns	0.88
array/reductions/reduce/Int64/dims=1	93796.25 ns	91171.25 ns	1.03
array/reductions/reduce/Int64/dims=1L	829429.5 ns	835642.5 ns	0.99
array/reductions/reduce/Int64/dims=2	83151.25 ns	92804 ns	0.90
array/reductions/reduce/Int64/dims=2L	144569.75 ns	144482.25 ns	1.00
array/reverse/1d	44985.75 ns	45710.5 ns	0.98
array/reverse/1dL	73921 ns	72093.5 ns	1.03
array/reverse/1dL_inplace	65543.5 ns	80361.25 ns	0.82
array/reverse/1d_inplace	37133 ns	61453.5 ns	0.60
array/reverse/2d	49930.75 ns	43585.75 ns	1.15
array/reverse/2dL	83538.75 ns	87671.25 ns	0.95
array/reverse/2dL_inplace	91774 ns	92566.5 ns	0.99
array/reverse/2d_inplace	57886 ns	63991 ns	0.90
array/sorting/1d	334265 ns	333142.5 ns	1.00
gemm/tiled	1870684.5 ns
gemm/tiled_unbounded	1868396.75 ns
integration/byval/reference	38761 ns	38921 ns	1.00
integration/byval/slices=1	39160 ns	41091 ns	0.95
integration/byval/slices=2	116571 ns	159003 ns	0.73
integration/byval/slices=3	233593 ns	231453 ns	1.01
integration/volumerhs	4887940 ns	4910793 ns	1.00
kernel/indexing	57155.75 ns	57765.75 ns	0.99
kernel/indexing_checked	39290.5 ns	51668.25 ns	0.76
kernel/launch	1425 ns	1402.5 ns	1.02
kernel/rand	80776 ns	70716 ns	1.14
latency/import	1720025947 ns	1713216916 ns	1.00
latency/precompile	39753959217 ns	39864575175 ns	1.00
latency/ttfp	2325086865 ns	2323506210 ns	1.00
stencil/diffusion3d	1627126.25 ns
stencil/diffusion3d_checked	1658231.75 ns

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

[luraess]

Thanks!