Skip to content

Commit 5e6b4c2

Browse files
charleskawczynskicharleskawczynski
committed
Add field-last benchmark script
wip wip
1 parent 9ab3c3f commit 5e6b4c2

File tree

2 files changed

+343
-0
lines changed

2 files changed

+343
-0
lines changed

‎.buildkite/pipeline.yml

+10
Original file line numberDiff line numberDiff line change
@@ -1294,6 +1294,16 @@ steps:
12941294
agents:
12951295
slurm_gpus: 1
12961296

1297+
- label: "Perf: benchmark scripts benchmark_field_last"
1298+
key: benchmark_field_last
1299+
command:
1300+
- "julia --project=.buildkite -e 'using CUDA; CUDA.versioninfo()'"
1301+
- "julia --color=yes --project=.buildkite benchmarks/scripts/benchmark_field_last.jl"
1302+
env:
1303+
CLIMACOMMS_DEVICE: "CUDA"
1304+
agents:
1305+
slurm_gpus: 1
1306+
12971307
- group: "Perf: Operators"
12981308
steps:
12991309

‎benchmarks/scripts/benchmark_field_last.jl

+333
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,333 @@
1+
#=
2+
julia --project=.buildkite
3+
using Revise; include(joinpath("benchmarks", "scripts", "benchmark_field_last.jl"))
4+
5+
# Info
6+
7+
# Benchmark results:
8+
9+
Clima A100:
10+
```
11+
Kernel `add3(x1, x2, x3) = x1+x2+x3` and `n_reads_writes=4`:
12+
[ Info: ArrayType = CuArray
13+
Problem size: (63, 4, 4, 5400, 1), float_type = Float32, device_bandwidth_GBs=2039
14+
┌─────────────────────────────────────────────────────────────────────┬──────────────────────────────────┬─────────┬─────────────┬────────────────┬────────┐
15+
│ funcs │ time per call │ bw % │ achieved bw │ n-reads/writes │ n-reps │
16+
├─────────────────────────────────────────────────────────────────────┼──────────────────────────────────┼─────────┼─────────────┼────────────────┼────────┤
17+
│ FLD.aos_cart_offset!(X_aos_ref, Y_aos_ref, us; bm, nreps = 100) │ 72 microseconds, 899 nanoseconds │ 54.568 │ 1112.64 │ 4 │ 100 │
18+
│ FLD.aos_lin_offset!(X_aos, Y_aos, us; bm, nreps = 100) │ 56 microseconds, 259 nanoseconds │ 70.708 │ 1441.74 │ 4 │ 100 │
19+
│ FLD.soa_linear_index!(X_soa, Y_soa, us; bm, nreps = 100) │ 56 microseconds, 515 nanoseconds │ 70.3877 │ 1435.21 │ 4 │ 100 │
20+
│ FLD.soa_cart_index!(X_soa, Y_soa, us; bm, nreps = 100) │ 67 microseconds, 462 nanoseconds │ 58.9663 │ 1202.32 │ 4 │ 100 │
21+
└─────────────────────────────────────────────────────────────────────┴──────────────────────────────────┴─────────┴─────────────┴────────────────┴────────┘
22+
23+
Kernel `add3(x1, x2, x3) = x1+x2+x3` and `n_reads_writes=4`:
24+
[ Info: ArrayType = CuArray
25+
Problem size: (63, 4, 4, 5400, 1), float_type = Float64, device_bandwidth_GBs=2039
26+
┌─────────────────────────────────────────────────────────────────────┬───────────────────────────────────┬─────────┬─────────────┬────────────────┬────────┐
27+
│ funcs │ time per call │ bw % │ achieved bw │ n-reads/writes │ n-reps │
28+
├─────────────────────────────────────────────────────────────────────┼───────────────────────────────────┼─────────┼─────────────┼────────────────┼────────┤
29+
│ FLD.aos_cart_offset!(X_aos_ref, Y_aos_ref, us; bm, nreps = 100) │ 106 microseconds, 783 nanoseconds │ 74.5051 │ 1519.16 │ 4 │ 100 │
30+
│ FLD.aos_lin_offset!(X_aos, Y_aos, us; bm, nreps = 100) │ 102 microseconds, 472 nanoseconds │ 77.6396 │ 1583.07 │ 4 │ 100 │
31+
│ FLD.soa_linear_index!(X_soa, Y_soa, us; bm, nreps = 100) │ 102 microseconds, 523 nanoseconds │ 77.6008 │ 1582.28 │ 4 │ 100 │
32+
│ FLD.soa_cart_index!(X_soa, Y_soa, us; bm, nreps = 100) │ 106 microseconds, 834 nanoseconds │ 74.4694 │ 1518.43 │ 4 │ 100 │
33+
└─────────────────────────────────────────────────────────────────────┴───────────────────────────────────┴─────────┴─────────────┴────────────────┴────────┘
34+
35+
Kernel `add3(x1, x2, x3) = x1` and `n_reads_writes=2`:
36+
[ Info: ArrayType = CuArray
37+
Problem size: (63, 4, 4, 5400, 1), float_type = Float32, device_bandwidth_GBs=2039
38+
┌─────────────────────────────────────────────────────────────────────┬──────────────────────────────────┬─────────┬─────────────┬────────────────┬────────┐
39+
│ funcs │ time per call │ bw % │ achieved bw │ n-reads/writes │ n-reps │
40+
├─────────────────────────────────────────────────────────────────────┼──────────────────────────────────┼─────────┼─────────────┼────────────────┼────────┤
41+
│ FLD.aos_cart_offset!(X_aos_ref, Y_aos_ref, us; bm, nreps = 100) │ 61 microseconds, 185 nanoseconds │ 32.5079 │ 662.837 │ 2 │ 100 │
42+
│ FLD.aos_lin_offset!(X_aos, Y_aos, us; bm, nreps = 100) │ 31 microseconds, 376 nanoseconds │ 63.3926 │ 1292.57 │ 2 │ 100 │
43+
│ FLD.soa_linear_index!(X_soa, Y_soa, us; bm, nreps = 100) │ 31 microseconds, 120 nanoseconds │ 63.9141 │ 1303.21 │ 2 │ 100 │
44+
│ FLD.soa_cart_index!(X_soa, Y_soa, us; bm, nreps = 100) │ 44 microseconds, 53 nanoseconds │ 45.1499 │ 920.607 │ 2 │ 100 │
45+
└─────────────────────────────────────────────────────────────────────┴──────────────────────────────────┴─────────┴─────────────┴────────────────┴────────┘
46+
```
47+
48+
# CPU (Mac M1)
49+
```
50+
[ Info: ArrayType = identity
51+
Problem size: (63, 4, 4, 5400, 1), float_type = Float32, device_bandwidth_GBs=2039
52+
┌─────────────────────────────────────────────────────────────────────┬───────────────────────────────────┬──────────┬─────────────┬────────────────┬────────┐
53+
│ funcs │ time per call (CPU) │ bw % │ achieved bw │ n-reads/writes │ n-reps │
54+
├─────────────────────────────────────────────────────────────────────┼───────────────────────────────────┼──────────┼─────────────┼────────────────┼────────┤
55+
│ FLD.aos_cart_offset!(X_aos_ref, Y_aos_ref, us; bm, nreps = 100) │ 16 milliseconds, 494 microseconds │ 0.241171 │ 4.91747 │ 4 │ 100 │
56+
│ FLD.aos_lin_offset!(X_aos, Y_aos, us; bm, nreps = 100) │ 783 microseconds, 256 nanoseconds │ 5.07871 │ 103.555 │ 4 │ 100 │
57+
│ FLD.soa_linear_index!(X_soa, Y_soa, us; bm, nreps = 100) │ 790 microseconds, 894 nanoseconds │ 5.02966 │ 102.555 │ 4 │ 100 │
58+
│ FLD.soa_cart_index!(X_soa, Y_soa, us; bm, nreps = 100) │ 12 milliseconds, 522 microseconds │ 0.317663 │ 6.47714 │ 4 │ 100 │
59+
└─────────────────────────────────────────────────────────────────────┴───────────────────────────────────┴──────────┴─────────────┴────────────────┴────────┘
60+
```
61+
62+
=#
63+
64+
#! format: off
65+
module BenchmarkFieldLastIndex
66+
67+
include("benchmark_utils.jl")
68+
69+
@inline function const_linear_index(us::UniversalSizesStatic, I, field_index)
70+
n = (get_Nv(us), get_Nij(us), get_Nij(us), get_Nh(us), 1)
71+
i = I + prod(n)*field_index
72+
return i
73+
end
74+
75+
@inline function const_linear_index_reference(us::UniversalSizesStatic, I, field_index)
76+
CI = CartesianIndices((get_Nv(us), get_Nij(us), get_Nij(us), get_Nh(us), 1))
77+
LI = LinearIndices((get_Nv(us), get_Nij(us), get_Nij(us), get_Nh(us), field_index+1))
78+
return LI[CI[I] + CartesianIndex((0, 0, 0, 0, field_index))]
79+
end
80+
81+
# add3(x1, x2, x3) = x1 + x2 + x3
82+
add3(x1, x2, x3) = x1
83+
84+
function aos_cart_offset!(X, Y, us; nreps = 1, bm=nothing, n_trials = 30)
85+
if Y isa Array
86+
e = Inf
87+
CI = CartesianIndices((get_Nv(us), get_Nij(us), get_Nij(us), get_Nh(us), 1))
88+
for t in 1:n_trials
89+
et = Base.@elapsed begin
90+
for i in 1:nreps
91+
@inbounds @simd for I in 1:get_N(us)
92+
CI1 = CI[I]
93+
CI2 = CI1 + CartesianIndex((0, 0, 0, 0, 1))
94+
CI3 = CI1 + CartesianIndex((0, 0, 0, 0, 2))
95+
Y[CI1] = add3(X[CI1], X[CI2], X[CI3])
96+
end
97+
end
98+
end
99+
e = min(e, et)
100+
end
101+
else
102+
e = Inf
103+
kernel = CUDA.@cuda always_inline = true launch = false aos_cart_offset_kernel!(X,Y,us)
104+
config = CUDA.launch_configuration(kernel.fun)
105+
threads = min(get_N(us), config.threads)
106+
blocks = cld(get_N(us), threads)
107+
for t in 1:n_trials
108+
et = CUDA.@elapsed begin
109+
for i in 1:nreps # reduce variance / impact of launch latency
110+
kernel(X,Y,us; threads, blocks)
111+
end
112+
end
113+
e = min(e, et)
114+
end
115+
end
116+
push_info(bm; e, nreps, caller = @caller_name(@__FILE__),n_reads_writes=4)
117+
return nothing
118+
end;
119+
function aos_cart_offset_kernel!(X, Y, us)
120+
@inbounds begin
121+
I = (CUDA.blockIdx().x - Int32(1)) * CUDA.blockDim().x + CUDA.threadIdx().x
122+
if I ≤ get_N(us)
123+
n = (get_Nv(us), get_Nij(us), get_Nij(us), get_Nh(us), 1)
124+
CI1 = CartesianIndices(map(x -> Base.OneTo(x), n))[I]
125+
CI2 = CI1 + CartesianIndex((0, 0, 0, 0, 1))
126+
CI3 = CI1 + CartesianIndex((0, 0, 0, 0, 2))
127+
Y[CI1] = add3(X[CI1], X[CI2], X[CI3])
128+
end
129+
end
130+
return nothing
131+
end;
132+
133+
function aos_lin_offset!(X, Y, us; nreps = 1, bm=nothing, n_trials = 30)
134+
if Y isa Array
135+
e = Inf
136+
for t in 1:n_trials
137+
et = Base.@elapsed begin
138+
for i in 1:nreps
139+
@inbounds @simd for I in 1:get_N(us)
140+
LY1 = const_linear_index(us, I, 0)
141+
LX1 = const_linear_index(us, I, 0)
142+
LX2 = const_linear_index(us, I, 1)
143+
LX3 = const_linear_index(us, I, 2)
144+
Y[LY1] = add3(X[LX1], X[LX2], X[LX3])
145+
end
146+
end
147+
end
148+
e = min(e, et)
149+
end
150+
else
151+
e = Inf
152+
kernel = CUDA.@cuda always_inline = true launch = false aos_lin_offset_kernel!(X,Y,us)
153+
config = CUDA.launch_configuration(kernel.fun)
154+
threads = min(get_N(us), config.threads)
155+
blocks = cld(get_N(us), threads)
156+
for t in 1:n_trials
157+
et = CUDA.@elapsed begin
158+
for i in 1:nreps
159+
kernel(X,Y,us; threads, blocks)
160+
end
161+
end
162+
e = min(e, et)
163+
end
164+
end
165+
push_info(bm; e, nreps, caller = @caller_name(@__FILE__),n_reads_writes=4)
166+
return nothing
167+
end;
168+
function aos_lin_offset_kernel!(X, Y, us)
169+
@inbounds begin
170+
I = (CUDA.blockIdx().x - Int32(1)) * CUDA.blockDim().x + CUDA.threadIdx().x
171+
if I ≤ get_N(us)
172+
LY1 = const_linear_index(us, I, 0)
173+
LX1 = const_linear_index(us, I, 0)
174+
LX2 = const_linear_index(us, I, 1)
175+
LX3 = const_linear_index(us, I, 2)
176+
Y[LY1] = add3(X[LX1], X[LX2], X[LX3])
177+
end
178+
end
179+
return nothing
180+
end;
181+
182+
function soa_cart_index!(X, Y, us; nreps = 1, bm=nothing, n_trials = 30)
183+
e = Inf
184+
if first(Y) isa Array
185+
CI = CartesianIndices((get_Nv(us), get_Nij(us), get_Nij(us), get_Nh(us)))
186+
for t in 1:n_trials
187+
et = Base.@elapsed begin
188+
for i in 1:nreps
189+
(y1,) = Y
190+
(x1, x2, x3) = X
191+
@inbounds @simd for I in 1:get_N(us)
192+
y1[CI[I]] = add3(x1[CI[I]], x2[CI[I]], x3[CI[I]])
193+
end
194+
end
195+
end
196+
e = min(e, et)
197+
end
198+
else
199+
kernel = CUDA.@cuda always_inline = true launch = false soa_cart_index_kernel!(X,Y,us)
200+
config = CUDA.launch_configuration(kernel.fun)
201+
threads = min(get_N(us), config.threads)
202+
blocks = cld(get_N(us), threads)
203+
for t in 1:n_trials
204+
et = CUDA.@elapsed begin
205+
for i in 1:nreps # reduce variance / impact of launch latency
206+
kernel(X,Y,us; threads, blocks)
207+
end
208+
end
209+
e = min(e, et)
210+
end
211+
end
212+
push_info(bm; e, nreps, caller = @caller_name(@__FILE__),n_reads_writes=4)
213+
return nothing
214+
end;
215+
function soa_cart_index_kernel!(X, Y, us)
216+
@inbounds begin
217+
I = (CUDA.blockIdx().x - Int32(1)) * CUDA.blockDim().x + CUDA.threadIdx().x
218+
if I ≤ get_N(us)
219+
CI = CartesianIndices((get_Nv(us), get_Nij(us), get_Nij(us), get_Nh(us)))
220+
(y1,) = Y
221+
(x1, x2, x3) = X
222+
y1[CI[I]] = add3(x1[CI[I]], x2[CI[I]], x3[CI[I]])
223+
end
224+
end
225+
return nothing
226+
end;
227+
228+
function soa_linear_index!(X, Y, us; nreps = 1, bm=nothing, n_trials = 30)
229+
e = Inf
230+
if first(Y) isa Array
231+
for t in 1:n_trials
232+
et = Base.@elapsed begin
233+
for i in 1:nreps
234+
(y1,) = Y
235+
(x1, x2, x3) = X
236+
@inbounds @simd for I in 1:get_N(us)
237+
y1[I] = add3(x1[I], x2[I], x3[I])
238+
end
239+
end
240+
end
241+
e = min(e, et)
242+
end
243+
else
244+
kernel = CUDA.@cuda always_inline = true launch = false soa_linear_index_kernel!(X,Y,us)
245+
config = CUDA.launch_configuration(kernel.fun)
246+
threads = min(get_N(us), config.threads)
247+
blocks = cld(get_N(us), threads)
248+
for t in 1:n_trials
249+
et = CUDA.@elapsed begin
250+
for i in 1:nreps # reduce variance / impact of launch latency
251+
kernel(X,Y,us; threads, blocks)
252+
end
253+
end
254+
e = min(e, et)
255+
end
256+
end
257+
push_info(bm; e, nreps, caller = @caller_name(@__FILE__),n_reads_writes=4)
258+
return nothing
259+
end;
260+
function soa_linear_index_kernel!(X, Y, us)
261+
@inbounds begin
262+
I = (CUDA.blockIdx().x - Int32(1)) * CUDA.blockDim().x + CUDA.threadIdx().x
263+
if I ≤ get_N(us)
264+
(y1,) = Y
265+
(x1, x2, x3) = X
266+
y1[I] = add3(x1[I], x2[I], x3[I])
267+
end
268+
end
269+
return nothing
270+
end;
271+
272+
end # module
273+
274+
import .BenchmarkFieldLastIndex as FLD
275+
276+
function fill_with_rand!(arr)
277+
FT = eltype(arr)
278+
T = typeof(arr)
279+
s = size(arr)
280+
arr .= T(rand(FT, s))
281+
end
282+
283+
using CUDA
284+
using Test
285+
@testset "Field last dim benchmark" begin
286+
bm = FLD.Benchmark(;problem_size=(63,4,4,5400,1), float_type=Float32) # size(problem_size, 4) == 1 to avoid double counting reads/writes
287+
ArrayType = CUDA.CuArray;
288+
# ArrayType = Base.identity;
289+
arr(float_type, problem_size, T) = T(zeros(float_type, problem_size...))
290+
291+
s = (63,4,4,5400,3);
292+
sY = (63,4,4,5400,1);
293+
st = (63,4,4,5400);
294+
ndofs = prod(st);
295+
us = FLD.UniversalSizesStatic(s[1], s[2], s[end-1]);
296+
297+
X_aos = arr(bm.float_type, s, ArrayType);
298+
Y_aos = arr(bm.float_type, sY, ArrayType);
299+
X_aos_ref = arr(bm.float_type, s, ArrayType);
300+
Y_aos_ref = arr(bm.float_type, sY, ArrayType);
301+
X_soa = ntuple(_ -> arr(bm.float_type, st, ArrayType), 3);
302+
Y_soa = ntuple(_ -> arr(bm.float_type, st, ArrayType), 1);
303+
fill_with_rand!(X_aos)
304+
fill_with_rand!(Y_aos)
305+
X_aos_ref .= X_aos
306+
Y_aos_ref .= Y_aos
307+
for i in 1:3; X_soa[i] .= X_aos[:,:,:,:, i]; end
308+
for i in 1:1; Y_soa[i] .= Y_aos[:,:,:,:, i]; end
309+
@info "ArrayType = $ArrayType"
310+
311+
FLD.aos_cart_offset!(X_aos_ref, Y_aos_ref, us; n_trials = 1, nreps = 1)
312+
FLD.aos_lin_offset!(X_aos, Y_aos, us; n_trials = 1, nreps = 1)
313+
FLD.soa_linear_index!(X_soa, Y_soa, us; n_trials = 1, nreps = 1)
314+
315+
@test all(X_aos .== X_aos_ref)
316+
@test all(Y_aos .== Y_aos_ref)
317+
for i in 1:3; @test all(X_soa[i] .== X_aos_ref[:,:,:,:,i]); end
318+
for i in 1:1; @test all(Y_soa[i] .== Y_aos_ref[:,:,:,:,i]); end
319+
320+
FLD.soa_cart_index!(X_soa, Y_soa, us; n_trials = 1, nreps = 1)
321+
322+
for i in 1:3; @test all(X_soa[i] .== X_aos_ref[:,:,:,:,i]); end
323+
for i in 1:1; @test all(Y_soa[i] .== Y_aos_ref[:,:,:,:,i]); end
324+
325+
FLD.aos_cart_offset!(X_aos_ref, Y_aos_ref, us; bm, nreps = 100)
326+
FLD.aos_lin_offset!(X_aos, Y_aos, us; bm, nreps = 100)
327+
FLD.soa_linear_index!(X_soa, Y_soa, us; bm, nreps = 100)
328+
FLD.soa_cart_index!(X_soa, Y_soa, us; bm, nreps = 100)
329+
330+
FLD.tabulate_benchmark(bm)
331+
end
332+
333+
# #! format: on

0 commit comments

Comments
 (0)