Optimize run_spin_excitation! for GPU

[rkierulf]

This pull request adds a GPU-optimized implementation of run_spin_excitation! to BlochGPU.jl. Compared with the function in BlochSimple.jl, the calculations which can be done beforehand for all time points are stored in preallocated matrices Bz, B, φ, ΔT1, and ΔT2. The sequential calculations are done inside a kernel apply_excitation!, which uses shared memory for all repeated memory accesses, and does the real / imaginary number math in Magnetization.jl directly so that the shared memory arrays store successive 32-bit values, which I think is ideal to avoid bank conflicts (https://developer.nvidia.com/blog/using-shared-memory-cuda-cc/).

The tests pass for Metal on my computer, and I've seen the excitation-heavy MRI Lab benchmark speed up by ~5x. Hopefully the same will be true for the other backends!

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 90.94%. Comparing base (1457a4c) to head (18bcaee).
Report is 1 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #462      +/-   ##
==========================================
+ Coverage   90.91%   90.94%   +0.03%     
==========================================
  Files          53       53              
  Lines        2916     2926      +10     
==========================================
+ Hits         2651     2661      +10     
  Misses        265      265              

Flag	Coverage Δ
base	88.20% <ø> (ø)
core	92.69% <100.00%> (+0.13%)	⬆️
files	93.55% <ø> (ø)
komamri	93.98% <ø> (ø)
plots	89.30% <ø> (ø)

Flags with carried forward coverage won't be shown. Click here to find out more.

Files	Coverage Δ
...RICore/src/simulation/SimMethods/Bloch/BlochGPU.jl	100.00% <100.00%> (ø)
KomaMRICore/src/simulation/SimulatorCore.jl	94.83% <100.00%> (ø)

[github-actions]

KomaMRI Benchmarks

Benchmark suite	Current: 18bcaee	Previous: 1457a4c	Ratio
MRI Lab/Bloch/CPU/2 thread(s)	243027991 ns	227517325.5 ns	1.07
MRI Lab/Bloch/CPU/4 thread(s)	135522120 ns	135033124 ns	1.00
MRI Lab/Bloch/CPU/8 thread(s)	144774394.5 ns	171880824 ns	0.84
MRI Lab/Bloch/CPU/1 thread(s)	408151458 ns	396561930.5 ns	1.03
MRI Lab/Bloch/GPU/CUDA	57005066.5 ns	138134905 ns	0.41
MRI Lab/Bloch/GPU/oneAPI	527703085 ns	14155999496.5 ns	0.037277698768672034
MRI Lab/Bloch/GPU/Metal	543126542 ns	3171338479 ns	0.17
MRI Lab/Bloch/GPU/AMDGPU	36831327 ns	75482754 ns	0.49
Slice Selection 3D/Bloch/CPU/2 thread(s)	1016083368 ns	1168211452 ns	0.87
Slice Selection 3D/Bloch/CPU/4 thread(s)	619287647 ns	612565463 ns	1.01
Slice Selection 3D/Bloch/CPU/8 thread(s)	385912553 ns	495427593 ns	0.78
Slice Selection 3D/Bloch/CPU/1 thread(s)	2252333331 ns	2245843835 ns	1.00
Slice Selection 3D/Bloch/GPU/CUDA	101397562.5 ns	108701927 ns	0.93
Slice Selection 3D/Bloch/GPU/oneAPI	662296008 ns	776956866 ns	0.85
Slice Selection 3D/Bloch/GPU/Metal	564139375 ns	769082459 ns	0.73
Slice Selection 3D/Bloch/GPU/AMDGPU	60677723 ns	64232156 ns	0.94

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