Metal support

[lazarusA]

just guessing now, but I think the current failure is due to the random number generator.

possible related issues:

• Add support for rand(rng, n) JuliaGPU/GPUArrays.jl#578
• Replace GPUArrays-based RNG with native one JuliaGPU/Metal.jl#691

doing

using GPUArrays
using Metal
GArray = MtlArray; # for Apple hardware
rng = GPUArrays.default_rng(GArray);
rand(rng,1)

fails:

ERROR: MethodError: no method matching rng_native_52(::GPUArrays.RNG)
The function `rng_native_52` exists, but no method is defined for this combination of argument types.

Closest candidates are:
  rng_native_52(::Random.MersenneTwister)
   @ Random ~/.julia/juliaup/julia-1.12.1+0.aarch64.apple.darwin14/share/julia/stdlib/v1.12/Random/src/RNGs.jl:444
  rng_native_52(::Random.RandomDevice)
   @ Random ~/.julia/juliaup/julia-1.12.1+0.aarch64.apple.darwin14/share/julia/stdlib/v1.12/Random/src/RNGs.jl:36
  rng_native_52(::Random.TaskLocalRNG)
   @ Random ~/.julia/juliaup/julia-1.12.1+0.aarch64.apple.darwin14/share/julia/stdlib/v1.12/Random/src/Xoshiro.jl:230

and doing directly Metal.rand

using GLMakie
Metal.rand(Vec2f)

ERROR: MethodError: no method matching Random.Sampler(::Type{GPUArrays.RNG}, ::Random.SamplerType{UInt32}, ::Val{1})
This error has been manually thrown, explicitly, so the method may exist but be intentionally marked as unimplemented.

Closest candidates are:
  Random.Sampler(::Type{<:Random.AbstractRNG}, ::Random.Sampler, ::Union{Val{1}, Val{Inf}})
   @ Random ~/.julia/juliaup/julia-1.12.1+0.aarch64.apple.darwin14/share/julia/stdlib/v1.12/Random/src/Random.jl:147
  Random.Sampler(::Type{<:Random.AbstractRNG}, ::Any, ::Union{Val{1}, Val{Inf}})
   @ Random ~/.julia/juliaup/julia-1.12.1+0.aarch64.apple.darwin14/share/julia/stdlib/v1.12/Random/src/Random.jl:183
  Random.Sampler(::Type{<:Random.AbstractRNG}, ::BitSet, ::Union{Val{1}, Val{Inf}})
   @ Random ~/.julia/juliaup/julia-1.12.1+0.aarch64.apple.darwin14/share/julia/stdlib/v1.12/Random/src/generation.jl:488
  ...

Stacktrace:
  [1] Random.Sampler(T::Type{GPUArrays.RNG}, sp::Random.SamplerType{UInt32}, r::Val{1})
    @ Random ~/.julia/juliaup/julia-1.12.1+0.aarch64.apple.darwin14/share/julia/stdlib/v1.12/Random/src/Random.jl:147
  [2] Random.Sampler(rng::GPUArrays.RNG, x::Random.SamplerType{UInt32}, r::Val{1})
    @ Random ~/.julia/juliaup/julia-1.12.1+0.aarch64.apple.darwin14/share/julia/stdlib/v1.12/Random/src/Random.jl:141
  [3] rand(rng::GPUArrays.RNG, X::Random.SamplerType{UInt32})
    @ Random ~/.julia/juliaup/julia-1.12.1+0.aarch64.apple.darwin14/share/julia/stdlib/v1.12/Random/src/Random.jl:255
  [4] rand(rng::GPUArrays.RNG, ::Type{UInt32})
    @ Random ~/.julia/juliaup/julia-1.12.1+0.aarch64.apple.darwin14/share/julia/stdlib/v1.12/Random/src/Random.jl:258
  [5] rand(r::GPUArrays.RNG, ::Random.SamplerTrivial{Random.UInt23Raw{UInt32}, UInt32})
    @ Random ~/.julia/juliaup/julia-1.12.1+0.aarch64.apple.darwin14/share/julia/stdlib/v1.12/Random/src/generation.jl:112
  [6] rand(rng::GPUArrays.RNG, X::Random.UInt23Raw{UInt32})
    @ Random ~/.julia/juliaup/julia-1.12.1+0.aarch64.apple.darwin14/share/julia/stdlib/v1.12/Random/src/Random.jl:255
  [7] rand(r::GPUArrays.RNG, ::Random.SamplerTrivial{Random.UInt23{UInt32}, UInt32})
    @ Random ~/.julia/juliaup/julia-1.12.1+0.aarch64.apple.darwin14/share/julia/stdlib/v1.12/Random/src/generation.jl:124
  [8] rand(rng::GPUArrays.RNG, X::Random.UInt23{UInt32})
    @ Random ~/.julia/juliaup/julia-1.12.1+0.aarch64.apple.darwin14/share/julia/stdlib/v1.12/Random/src/Random.jl:255
  [9] rand(r::GPUArrays.RNG, ::Random.SamplerTrivial{Random.CloseOpen01{Float32}, Float32})
    @ Random ~/.julia/juliaup/julia-1.12.1+0.aarch64.apple.darwin14/share/julia/stdlib/v1.12/Random/src/generation.jl:29
 [10] rand(rng::GPUArrays.RNG, ::Type{Float32})
    @ Random ~/.julia/juliaup/julia-1.12.1+0.aarch64.apple.darwin14/share/julia/stdlib/v1.12/Random/src/Random.jl:258
 [11] macro expansion
    @ ~/.julia/packages/StaticArrays/DsPgf/src/arraymath.jl:77 [inlined]
 [12] _rand
    @ ~/.julia/packages/StaticArrays/DsPgf/src/arraymath.jl:62 [inlined]
 [13] rand
    @ ~/.julia/packages/StaticArrays/DsPgf/src/arraymath.jl:60 [inlined]
 [14] rand!
    @ ~/.julia/juliaup/julia-1.12.1+0.aarch64.apple.darwin14/share/julia/stdlib/v1.12/Random/src/Random.jl:273 [inlined]
 [15] rand!
    @ ~/.julia/juliaup/julia-1.12.1+0.aarch64.apple.darwin14/share/julia/stdlib/v1.12/Random/src/Random.jl:269 [inlined]
 [16] rand
    @ ~/.julia/packages/StaticArrays/DsPgf/src/arraymath.jl:59 [inlined]
 [17] rand
    @ ~/.julia/juliaup/julia-1.12.1+0.aarch64.apple.darwin14/share/julia/stdlib/v1.12/Random/src/Random.jl:293 [inlined]
 [18] rand!(rng::GPUArrays.RNG, A::MtlVector{Vec{2, Float32}, Metal.SharedStorage})
    @ GPUArrays ~/.julia/packages/GPUArrays/w335n/src/host/random.jl:115
 [19] rand(::Type, ::Int64; storage::Type)
    @ Metal ~/.julia/packages/Metal/av3If/src/random.jl:40
 [20] rand(T::Type; storage::Type)
    @ Metal ~/.julia/packages/Metal/av3If/src/random.jl:52
 [21] top-level scope
    @ REPL[4]:1

which is / could be used for example in

function sample_light(bvh, ctx, width, height, camera_pos, focal_length, aspect, x, y, sky_color)
    jitter = rand(Vec2f)
....
end

[lazarusA]

well... things are not ok even before going there, but this only if I want to print the output, doing bvh_gpu = to_gpu(GArray, bvh); seems to be ok

bvh_gpu = to_gpu(GArray, bvh)

[91451] signal 10 (1): Bus error: 10
in expression starting at REPL[31]:1
macro expansion at /Users/lalonso/.julia/packages/LLVM/iza6e/src/interop/base.jl:39 [inlined]
macro expansion at ./none:0 [inlined]
pointerref at ./none:0 [inlined]
unsafe_load at /Users/lalonso/.julia/packages/LLVM/iza6e/src/interop/pointer.jl:85 [inlined]
arrayref at /Users/lalonso/.julia/packages/Metal/av3If/src/device/array.jl:82 [inlined]
getindex at /Users/lalonso/.julia/packages/Metal/av3If/src/device/array.jl:103 [inlined]
iterate at /Users/lalonso/.julia/packages/Metal/av3If/src/device/array.jl:159 [inlined]
iterate at /Users/lalonso/.julia/packages/Metal/av3If/src/device/array.jl:158 [inlined]
_foldl_impl at ./reduce.jl:48
foldl_impl at ./reduce.jl:40 [inlined]
mapfoldl_impl at ./reduce.jl:36 [inlined]
_mapreduce_dim at ./reducedim.jl:331 [inlined]
#mapreduce#726 at ./reducedim.jl:326 [inlined]
mapreduce at ./reducedim.jl:326 [inlined]
#_sum#736 at ./reducedim.jl:984 [inlined]
_sum at ./reducedim.jl:984 [inlined]
#sum#734 at ./reducedim.jl:980 [inlined]
sum at ./reducedim.jl:980 [inlined]
_simple_count at ./reduce.jl:1084 [inlined]
_count at ./reducedim.jl:411 [inlined]
#count#730 at ./reducedim.jl:409 [inlined]
count at ./reducedim.jl:409 [inlined]
show at /Users/lalonso/Documents/Raycore.jl/src/bvh.jl:706
unknown function (ip: 0x381bdc11b) at (unknown file)
show_repl at /Users/julia/.julia/scratchspaces/a66863c6-20e8-4ff4-8a62-49f30b1f605e/agent-cache/default-honeycrisp-XG3Q6T6R70.0/build/default-honeycrisp-XG3Q6T6R70-0/julialang/julia-release-1-dot-12/usr/share/julia/stdlib/v1.12/REPL/src/REPL.jl:548
show_limited at /Users/julia/.julia/scratchspaces/a66863c6-20e8-4ff4-8a62-49f30b1f605e/agent-cache/default-honeycrisp-XG3Q6T6R70.0/build/default-honeycrisp-XG3Q6T6R70-0/julialang/julia-release-1-dot-12/usr/share/julia/stdlib/v1.12/REPL/src/REPL.jl:517
#display##0 at /Users/julia/.julia/scratchspaces/a66863c6-20e8-4ff4-8a62-49f30b1f605e/agent-cache/default-honeycrisp-XG3Q6T6R70.0/build/default-honeycrisp-XG3Q6T6R70-0/julialang/julia-release-1-dot-12/usr/share/julia/stdlib/v1.12/REPL/src/REPL.jl:540
jfptr_YY.displayYY.YY.0_18283.1 at /Users/lalonso/.julia/juliaup/julia-1.12.1+0.aarch64.apple.darwin14/share/julia/compiled/v1.12/REPL/u0gqU_vMi9E.dylib (unknown line)
with_repl_linfo at /Users/julia/.julia/scratchspaces/a66863c6-20e8-4ff4-8a62-49f30b1f605e/agent-cache/default-honeycrisp-XG3Q6T6R70.0/build/default-honeycrisp-XG3Q6T6R70-0/julialang/julia-release-1-dot-12/usr/share/julia/stdlib/v1.12/REPL/src/REPL.jl:833
jfptr_with_repl_linfo_17935.1 at /Users/lalonso/.julia/juliaup/julia-1.12.1+0.aarch64.apple.darwin14/share/julia/compiled/v1.12/REPL/u0gqU_vMi9E.dylib (unknown line)
display at /Users/julia/.julia/scratchspaces/a66863c6-20e8-4ff4-8a62-49f30b1f605e/agent-cache/default-honeycrisp-XG3Q6T6R70.0/build/default-honeycrisp-XG3Q6T6R70-0/julialang/julia-release-1-dot-12/usr/share/julia/stdlib/v1.12/REPL/src/REPL.jl:526
display at /Users/julia/.julia/scratchspaces/a66863c6-20e8-4ff4-8a62-49f30b1f605e/agent-cache/default-honeycrisp-XG3Q6T6R70.0/build/default-honeycrisp-XG3Q6T6R70-0/julialang/julia-release-1-dot-12/usr/share/julia/stdlib/v1.12/REPL/src/REPL.jl:546
unknown function (ip: 0x1225fc083) at (unknown file)
display at ./multimedia.jl:340
unknown function (ip: 0x1225f42c7) at (unknown file)
jl_apply at /Users/julia/.julia/scratchspaces/a66863c6-20e8-4ff4-8a62-49f30b1f605e/agent-cache/default-honeycrisp-XG3Q6T6R70.0/build/default-honeycrisp-XG3Q6T6R70-0/julialang/julia-release-1-dot-12/src/./julia.h:2391 [inlined]
jl_f_invokelatest at /Users/julia/.julia/scratchspaces/a66863c6-20e8-4ff4-8a62-49f30b1f605e/agent-cache/default-honeycrisp-XG3Q6T6R70.0/build/default-honeycrisp-XG3Q6T6R70-0/julialang/julia-release-1-dot-12/src/builtins.c:881
#print_response##2 at /Users/julia/.julia/scratchspaces/a66863c6-20e8-4ff4-8a62-49f30b1f605e/agent-cache/default-honeycrisp-XG3Q6T6R70.0/build/default-honeycrisp-XG3Q6T6R70-0/julialang/julia-release-1-dot-12/usr/share/julia/stdlib/v1.12/REPL/src/REPL.jl:590
jfptr_YY.print_responseYY.YY.2_18764.1 at /Users/lalonso/.julia/juliaup/julia-1.12.1+0.aarch64.apple.darwin14/share/julia/compiled/v1.12/REPL/u0gqU_vMi9E.dylib (unknown line)
repl_backend_loop at /Users/julia/.julia/scratchspaces/a66863c6-20e8-4ff4-8a62-49f30b1f605e/agent-cache/default-honeycrisp-XG3Q6T6R70.0/build/default-honeycrisp-XG3Q6T6R70-0/julialang/julia-release-1-dot-12/usr/share/julia/stdlib/v1.12/REPL/src/REPL.jl:445
#start_repl_backend#41 at /Users/julia/.julia/scratchspaces/a66863c6-20e8-4ff4-8a62-49f30b1f605e/agent-cache/default-honeycrisp-XG3Q6T6R70.0/build/default-honeycrisp-XG3Q6T6R70-0/julialang/julia-release-1-dot-12/usr/share/julia/stdlib/v1.12/REPL/src/REPL.jl:427
start_repl_backend at /Users/julia/.julia/scratchspaces/a66863c6-20e8-4ff4-8a62-49f30b1f605e/agent-cache/default-honeycrisp-XG3Q6T6R70.0/build/default-honeycrisp-XG3Q6T6R70-0/julialang/julia-release-1-dot-12/usr/share/julia/stdlib/v1.12/REPL/src/REPL.jl:424 [inlined]
#run_repl#50 at /Users/julia/.julia/scratchspaces/a66863c6-20e8-4ff4-8a62-49f30b1f605e/agent-cache/default-honeycrisp-XG3Q6T6R70.0/build/default-honeycrisp-XG3Q6T6R70-0/julialang/julia-release-1-dot-12/usr/share/julia/stdlib/v1.12/REPL/src/REPL.jl:653
run_repl at /Users/julia/.julia/scratchspaces/a66863c6-20e8-4ff4-8a62-49f30b1f605e/agent-cache/default-honeycrisp-XG3Q6T6R70.0/build/default-honeycrisp-XG3Q6T6R70-0/julialang/julia-release-1-dot-12/usr/share/julia/stdlib/v1.12/REPL/src/REPL.jl:639
jfptr_run_repl_18811.1 at /Users/lalonso/.julia/juliaup/julia-1.12.1+0.aarch64.apple.darwin14/share/julia/compiled/v1.12/REPL/u0gqU_vMi9E.dylib (unknown line)
run_std_repl at ./client.jl:478
jfptr_run_std_repl_53553.1 at /Users/lalonso/.julia/juliaup/julia-1.12.1+0.aarch64.apple.darwin14/lib/julia/sys.dylib (unknown line)
jl_apply at /Users/julia/.julia/scratchspaces/a66863c6-20e8-4ff4-8a62-49f30b1f605e/agent-cache/default-honeycrisp-XG3Q6T6R70.0/build/default-honeycrisp-XG3Q6T6R70-0/julialang/julia-release-1-dot-12/src/./julia.h:2391 [inlined]
jl_f_invokelatest at /Users/julia/.julia/scratchspaces/a66863c6-20e8-4ff4-8a62-49f30b1f605e/agent-cache/default-honeycrisp-XG3Q6T6R70.0/build/default-honeycrisp-XG3Q6T6R70-0/julialang/julia-release-1-dot-12/src/builtins.c:881
run_main_repl at ./client.jl:499
repl_main at ./client.jl:586 [inlined]
_start at ./client.jl:561
jfptr__start_66711.1 at /Users/lalonso/.julia/juliaup/julia-1.12.1+0.aarch64.apple.darwin14/lib/julia/sys.dylib (unknown line)
jl_apply at /Users/julia/.julia/scratchspaces/a66863c6-20e8-4ff4-8a62-49f30b1f605e/agent-cache/default-honeycrisp-XG3Q6T6R70.0/build/default-honeycrisp-XG3Q6T6R70-0/julialang/julia-release-1-dot-12/src/./julia.h:2391 [inlined]
true_main at /Users/julia/.julia/scratchspaces/a66863c6-20e8-4ff4-8a62-49f30b1f605e/agent-cache/default-honeycrisp-XG3Q6T6R70.0/build/default-honeycrisp-XG3Q6T6R70-0/julialang/julia-release-1-dot-12/src/jlapi.c:971
jl_repl_entrypoint at /Users/julia/.julia/scratchspaces/a66863c6-20e8-4ff4-8a62-49f30b1f605e/agent-cache/default-honeycrisp-XG3Q6T6R70.0/build/default-honeycrisp-XG3Q6T6R70-0/julialang/julia-release-1-dot-12/src/jlapi.c:1139
Allocations: 162685491 (Pool: 162682191; Big: 3300); GC: 40
zsh: bus error  Julia

[lazarusA]

ok, ok... also a mwe using to_gpu and Metal fails

# MWE for debugging Metal GPU arrays
using Metal
using Raycore: to_gpu
GArray = MtlArray;
# Small test
A = rand(Float32, 10)
gpuA = to_gpu(GArray, A)  # issue here?

# Try accessing it
cpuA = Array(gpuA)  # should copy back safely but errors out!

note that

gpuA
10-element device array at Core.LLVMPtr{Float32, 1}(0x0000001500000000)

mwe_to_gpu.mov

using Metal directly is fine

metalA = Metal.rand(Float32, 10)
cpuA = Array(metalA)  # this works fine

note that

MetalA
10-element MtlVector{Float32, Metal.PrivateStorage}:
 0.8307109
 0.95224166
 0.09675121
 0.6214751
 0.38179517
 0.34523797
 0.64462495
 0.70480347
 0.16316044
 0.747296

[lazarusA]

last, what is the expected behaviour for argconvert and mtlconvert, note that in the following none of the output an MtlMatrix

# Check backend detection and KernelAbstractions functionality
using KernelAbstractions
import KernelAbstractions as KA
using Metal
metalA = Metal.rand(Float32, 2, 2)
cpuA = Array(metalA)

backend = KA.get_backend(metalA)
output = KA.ones(backend, Float32, 2, 2)

_die_kernel = raytrace_kernel_v1!(backend)

# non of the following go into a `MtlMatrix`! they should?  right?

KA.argconvert(_die_kernel, cpuA)  # ????
2×2 Matrix{Float32}:
 0.831668  0.946562
 0.560413  0.212343

Metal.mtlconvert(cpuA) # ???
2×2 Matrix{Float32}:
 0.831668  0.946562
 0.560413  0.212343

[lazarusA]

from Simon's version this works. (unroll for loops for GPU)

using Raycore, Colors
import KernelAbstractions as KA
using KernelAbstractions, GLMakie
using Raycore: to_gpu
include("raytracing-core.jl")

# Compute interpolated normal at hit point
function compute_normal(triangle, bary_coords)
    v0, v1, v2 = Raycore.normals(triangle)
    u, v, w = bary_coords[1], bary_coords[2], bary_coords[3]
    return Vec3f(normalize(v0 * u + v1 * v + v2 * w))
end

# Generate camera ray for a pixel with optional jitter
function camera_ray(x, y, width, height, camera_pos, focal_length, aspect)
    ndc_x = (2.0f0 * (Float32(x) - 0.5f0) / Float32(width) - 1.0f0) * aspect
    ndc_y = 1.0f0 - 2.0f0 * (Float32(y) - 0.5f0) / Float32(height)
    direction = normalize(Vec3f(ndc_x, ndc_y, focal_length))
    return Raycore.Ray(o=camera_pos, d=direction)
end

# Convert between color representations
to_vec3f(c::RGB) = Vec3f(c.r, c.g, c.b)
to_rgb(v::Vec3f) = RGB{Float32}(v...)

function compute_light(
    bvh, point, normal, light_pos, light_intensity, light_color; shadow_samples=1)

    light_vec = light_pos - point
    light_dist = norm(light_vec)
    light_dir = light_vec / light_dist

    diffuse = max(0.0f0, dot(normal, light_dir))

    # Shadow testing with optional soft shadows
    shadow_factor = 0.0f0
    light_radius = 0.2f0  # Size of area light for soft shadows

    for _ in 1:shadow_samples
        # For shadow_samples=1, this is just the light position (hard shadow)
        # For shadow_samples>1, we sample random points on a disk (soft shadow)
        if shadow_samples > 1
            # Random point on disk perpendicular to light direction
            offset = (rand(Vec3f) .* 2.0f0 .- 1.0f0) * light_radius
            offset = offset - light_dir * dot(offset, light_dir)
            shadow_target = light_pos + offset
        else
            shadow_target = light_pos
        end

        shadow_vec = shadow_target - point
        shadow_dist = norm(shadow_vec)
        shadow_dir = normalize(shadow_vec)

        shadow_ray = Raycore.Ray(o=point + normal * 0.001f0, d=shadow_dir)
        shadow_hit, _, hit_dist, _ = Raycore.any_hit(bvh, shadow_ray)

        if !shadow_hit || hit_dist >= shadow_dist
            shadow_factor += 1.0f0
        end
    end
    shadow_factor /= shadow_samples

    # Compute final light contribution
    attenuation = light_intensity / (light_dist * light_dist)
    return to_vec3f(light_color) * (diffuse * attenuation * shadow_factor)
end

function shadow_kernel(bvh, tri, dist, bary, ray; shadow_samples=1)
    hit_point = ray.o + ray.d * dist
    normal = compute_normal(tri, bary)
    # Single point light
    light_pos = Point3f(3, 4, -2)
    light_intensity = 50.0f0
    light_color = RGB{Float32}(1.0f0, 0.9f0, 0.8f0)
    # Hard shadows (shadow_samples=1)
    light_contrib = compute_light(
        bvh, hit_point, normal, light_pos, light_intensity, light_color;
        shadow_samples=shadow_samples
    )
    ambient = 0.1f0

    brightness = ambient .+ light_contrib
    return brightness
end
# function compute_multi_light(bvh, ctx, point, normal, mat; shadow_samples=1)
#     base_color = to_vec3f(mat.base_color)
#     total_color = base_color * ctx.ambient

#     for light in ctx.lights
#         light_contrib = compute_light(bvh, point, normal, light.position, light.intensity, light.color, shadow_samples=shadow_samples)
#         total_color += base_color .* light_contrib
#     end

#     return total_color
# end
function compute_multi_light(bvh, ctx, point, normal, mat; shadow_samples=1)
    base_color = to_vec3f(mat.base_color)
    total_color = base_color * ctx.ambient
    light = ctx.lights[1]
    light_contrib = compute_light(bvh, point, normal, light.position, light.intensity, light.color, shadow_samples=shadow_samples)
    total_color += base_color .* light_contrib
    # 2
    light = ctx.lights[2]
    light_contrib = compute_light(bvh, point, normal, light.position, light.intensity, light.color, shadow_samples=shadow_samples)
    total_color += base_color .* light_contrib
    # 3
    light = ctx.lights[3]
    light_contrib = compute_light(bvh, point, normal, light.position, light.intensity, light.color, shadow_samples=shadow_samples)
    total_color += base_color .* light_contrib
    return total_color
end
function material_kernel(bvh, ctx, tri, dist, bary, ray)
    hit_point = ray.o + ray.d * dist
    normal = compute_normal(tri, bary)
    mat = ctx.materials[tri.material_idx]

    color = compute_multi_light(bvh, ctx, hit_point, normal, mat, shadow_samples=2)
    return color
end
# Basic kernel: one thread per pixel, straightforward implementation
@kernel function raytrace_kernel_v1!(
    kernel, img, @Const(bvh), ctx, camera_pos, focal_length, aspect
)
    # Get pixel coordinates
    idx = @index(Global, Cartesian)
    height, width = size(img)
    # Convert linear index to 2D coordinates
    if checkbounds(Bool, img, idx)
        x, y = Tuple(idx)
        # Generate camera ray and calculate a simple light model
        ray = camera_ray(y, x, width, height, camera_pos, focal_length, aspect)
        hit_found, triangle, dist, bary_coords = Raycore.closest_hit(bvh, ray)
        if hit_found
            @inbounds img[x, y] = kernel(bvh, ctx, triangle, dist, bary_coords, ray)
        end
    end
end
function trace_gpu(kernel, img, bvh, ctx;
        camera_pos=Point3f(0, -0.9, -2.5), fov=45.0f0,
        ndrange=size(img), tilesize=nothing
    )
    height, width = size(img)
    aspect = Float32(width / height)
    focal_length = 1.0f0 / tan(deg2rad(fov / 2))
    # KernelAbstractions automatically detects the backend (CPU/GPU) from the array type
    backend = KA.get_backend(img)
    # Create the kernel with or without tilesize (for workgroup configuration)
    kernel! = raytrace_kernel_v1!(backend)

    kernel!(kernel, img, bvh, ctx, camera_pos, focal_length, aspect, ndrange=ndrange)
    # Ensure GPU computation completes before returning
    KA.synchronize(backend)
    return img
end

bvh, ctx = example_scene()

img = zeros(Vec3f, 512, 1024)
trace_gpu(material_kernel, img, bvh, ctx)
image(rotr90(to_rgb.(Array(img))); axis =(; aspect=DataAspect()))

# Metal
using Metal
GArray = MtlArray;
img = zeros(Vec3f, 512, 1024)

img_gpu = GArray(img);
bvh_gpu = Raycore.to_gpu(GArray, bvh);
ctx_gpu = Raycore.to_gpu(GArray, ctx);
@time trace_gpu(material_kernel, img_gpu, bvh_gpu, ctx_gpu);

to_cpu = to_rgb.(Array(img_gpu))
image(rotr90(to_cpu); axis =(; aspect=DataAspect()))