1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
|
# Copyright 2013-2023 Lawrence Livermore National Security, LLC and other
# Spack Project Developers. See the top-level COPYRIGHT file for details.
#
# SPDX-License-Identifier: (Apache-2.0 OR MIT)
import re # To get the variant name after (+)
from spack.package import *
def find_model_flag(str):
res = re.findall(r"\+(\w+)", str)
if not res:
return ""
return res
class Babelstream(CMakePackage, CudaPackage, ROCmPackage):
"""Measure memory transfer rates to/from global device memory on GPUs.
This benchmark is similar in spirit, and based on, the STREAM benchmark for CPUs."""
homepage = "https://github.com/UoB-HPC/BabelStream"
url = "https://github.com/UoB-HPC/BabelStream/archive/refs/tags/v4.0.tar.gz"
git = "https://github.com/UoB-HPC/BabelStream.git"
version("4.0", sha256="a9cd39277fb15d977d468435eb9b894f79f468233f0131509aa540ffda4f5953")
version("main", branch="main")
version("develop", branch="develop")
maintainers("tomdeakin", "kaanolgu", "tom91136", "robj0nes")
# Languages
# Also supported variants are cuda and rocm (for HIP)
variant("sycl", default=False, description="Enable SYCL support")
variant("sycl2020", default=False, description="Enable SYCL support")
variant("omp", default=False, description="Enable OpenMP support")
variant("ocl", default=False, description="Enable OpenCL support")
variant("tbb", default=False, description="Enable TBB support")
variant("acc", default=False, description="Enable OpenACC support")
variant("thrust", default=False, description="Enable THRUST support")
variant("raja", default=False, description="Enable RAJA support")
variant("stddata", default=False, description="Enable STD-data support")
variant("stdindices", default=False, description="Enable STD-indices support")
variant("stdranges", default=False, description="Enable STD-ranges support")
# Some models need to have the programming model abstraction downloaded -
# this variant enables a path to be provided.
variant("dir", values=str, default="none", description="Enable Directory support")
# Kokkos conflict and variant
conflicts(
"dir=none", when="+kokkos", msg="KOKKKOS requires architecture to be specfied by dir="
)
variant("kokkos", default=False, description="Enable KOKKOS support")
# ACC conflict
variant("cpu_arch", values=str, default="none", description="Enable CPU Target for ACC")
variant("target", values=str, default="none", description="Enable CPU Target for ACC")
# STD conflicts
conflicts("+stddata", when="%gcc@:10.1.0", msg="STD-data requires newer version of GCC")
conflicts("+stdindices", when="%gcc@:10.1.0", msg="STD-indices requires newer version of GCC")
conflicts("+stdranges", when="%gcc@:10.1.0", msg="STD-ranges requires newer version of GCC")
# CUDA conflict
conflicts(
"cuda_arch=none",
when="+cuda",
msg="CUDA requires architecture to be specfied by cuda_arch=",
)
variant("mem", values=str, default="DEFAULT", description="Enable MEM Target for CUDA")
# Raja Conflict
variant(
"offload", values=str, default="none", description="Enable RAJA Target [CPU or NVIDIA]"
)
conflicts(
"offload=none",
when="+raja",
msg="RAJA requires architecture to be specfied by target=[CPU,NVIDIA]",
)
# download raja from https://github.com/LLNL/RAJA
conflicts(
"dir=none",
when="+raja",
msg="RAJA implementation requires architecture to be specfied by dir=",
)
# Thrust Conflict
# conflicts("~cuda", when="+thrust", msg="Thrust requires +cuda variant")
depends_on("thrust", when="+thrust")
depends_on("rocthrust", when="+thrust implementation=rocm")
# TBB Dependency
depends_on("intel-oneapi-tbb", when="+tbb")
partitioner_vals = ["auto", "affinity", "static", "simple"]
variant(
"partitioner",
values=partitioner_vals,
default="auto",
description="Partitioner specifies how a loop template should partition its work among threads.\
Possible values are:\
AUTO - Optimize range subdivision based on work-stealing events.\
AFFINITY - Proportional splitting that optimizes for cache affinity.\
STATIC - Distribute work uniformly with no additional load balancing.\
SIMPLE - Recursively split its range until it cannot be further subdivided.\
See https://spec.oneapi.com/versions/latest/elements/oneTBB/source/algorithms.html#partitioners for more details.",
)
# Kokkos Dependency
depends_on("kokkos@3.7.1", when="+kokkos")
# OpenCL Dependency
backends = {
"ocl": [
("amd", "rocm-opencl", "enable ROCM backend"),
("cuda", "cuda", "enable Cuda backend"),
("intel", "intel-oneapi-compilers", "enable Intel backend"),
("pocl", "pocl@1.5", "enable POCL backend"),
],
"kokkos": [
("cuda", "cuda", "enable Cuda backend"),
("omp", "none", "enable Cuda backend"),
],
}
backend_vals = ["none"]
for lang in backends:
for item in backends[lang]:
backend, dpdncy, descr = item
backend_vals.append(backend.lower())
variant("backend", values=backend_vals, default="none", description="Enable backend support")
for lang in backends:
for item in backends[lang]:
backend, dpdncy, descr = item
if dpdncy.lower() != "none":
depends_on("%s" % dpdncy.lower(), when="backend=%s" % backend.lower())
# this flag could be used in all required languages
variant("flags", values=str, default="none", description="Additional CXX flags to be provided")
# comp_impl_vals=["ONEAPI-DPCPP","DPCPP","HIPSYCL","COMPUTECPP"]
variant(
"implementation",
values=str,
default="none",
description="Compile using the specified SYCL compiler option",
)
conflicts(
"implementation=none",
when="+sycl",
msg="SYCL requires compiler implementation to be specified by option=",
)
conflicts(
"implementation=none",
when="+thrust",
msg="Which Thrust implementation to use, supported options include:\
- CUDA (via https://github.com/NVIDIA/thrust)\
- ROCM (via https://github.com/ROCmSoftwarePlatform/rocThrust)",
)
# This applies to all
depends_on("cmake@3.14.0:", type="build")
depends_on("opencl-c-headers", when="+ocl")
def cmake_args(self):
# convert spec to string to work on it
spec_string = str(self.spec)
# take only the first portion of the spec until space
spec_string_truncate = spec_string.split(" ", 1)[0]
model_list = find_model_flag(spec_string_truncate) # Prints out ['cuda', 'thrust']
if len(model_list) > 1:
ignore_list = ["cuda"] # if +acc is provided ignore the cuda model
model = list(set(model_list) - set(ignore_list))
# We choose 'thrust' from the list of ['cuda', 'thrust']
args = ["-DMODEL=" + model[0]]
else:
# if it is +stddata,indices etc. we need to pass it
# as std-data to the CMake compiler
# do some alterations here
if "std" in model_list[0]:
args = ["-DMODEL=" + "std-" + model_list[0].split("d", 1)[1]]
else:
args = ["-DMODEL=" + model_list[0]]
# ===================================
# ACC
# ===================================
if ("+acc" in self.spec) and ("~cuda" in self.spec):
args.append("-DCMAKE_CXX_COMPILER=" + self.compiler.cxx)
if "cuda_arch" in self.spec.variants:
cuda_arch_list = self.spec.variants["cuda_arch"].value
# the architecture value is only number so append sm_ to the name
cuda_arch = "cc" + cuda_arch_list[0]
args.append("-DTARGET_DEVICE=gpu")
args.append("-DCUDA_ARCH=" + cuda_arch)
elif "cpu_arch" in self.spec.variants:
cpu_arch_list = self.spec.variants["cpu_arch"].value
# the architecture value is only number so append sm_ to the name
cpu_arch = cpu_arch_list[0]
args.append("-DTARGET_DEVICE=multicore")
args.append("-DTARGET_PROCESSOR=" + cpu_arch)
# ===================================
# STDdata,STDindices,STDranges
# ===================================
std_list = ["+stddata", "+stdindices", "+stdranges"]
if spec_string.startswith(tuple(std_list)):
args.append("-DCMAKE_CXX_COMPILER=" + self.compiler.cxx)
# ===================================
# CUDA
# ===================================
if ("+cuda" in self.spec) and ("~kokkos" in self.spec) and ("~acc" in self.spec):
# Set up the cuda macros needed by the build
cuda_arch_list = self.spec.variants["cuda_arch"].value
# the architecture value is only number so append sm_ to the name
cuda_arch = "sm_" + cuda_arch_list[0]
args.append("-DCUDA_ARCH=" + cuda_arch)
cuda_dir = self.spec["cuda"].prefix
cuda_comp = cuda_dir + "/bin/nvcc"
args.append("-DCMAKE_CUDA_COMPILER=" + cuda_comp)
args.append("-DMEM=" + self.spec.variants["mem"].value)
if self.spec.variants["flags"].value != "none":
args.append("-DCUDA_EXTRA_FLAGS=" + self.spec.variants["flags"].value)
# ===================================
# OMP
# ===================================
# `~kokkos` option is there to prevent +kokkos +omp setting to use omp directly from here
# Same applies for raja
if ("+omp" in self.spec) and ("~kokkos" in self.spec) and ("~raja" in self.spec):
args.append("-DCMAKE_CXX_COMPILER=" + self.compiler.cxx)
if "cuda_arch" in self.spec.variants:
cuda_arch_list = self.spec.variants["cuda_arch"].value
# the architecture value is only number so append sm_ to the name
cuda_arch = "sm_" + cuda_arch_list[0]
args.append("-DOFFLOAD= " + "NVIDIA:" + cuda_arch)
elif "amdgpu_target" in self.spec.variants:
rocm_arch = self.spec.variants["amdgpu_target"].value
# the architecture value is only number so append sm_ to the name
args.append("-DOFFLOAD=" + " AMD:" + rocm_arch)
else:
args.append("-DOFFLOAD=" + "INTEL")
# ===================================
# SYCL
# ===================================
if "+sycl" in self.spec:
args.append("-DSYCL_COMPILER=" + self.spec.variants["implementation"].value.upper())
if self.spec.variants["implementation"].value.upper() != "ONEAPI-DPCPP":
args.append(
"-DSYCL_COMPILER_DIR=" + self.spec.variants["implementation"].value.upper()
)
if self.spec.variants["implementation"].value.upper() == "COMPUTE-CPP":
args.append("-DOpenCL_LIBRARY=")
# ===================================
# SYCL 2020
# ===================================
if "+sycl2020" in self.spec:
if self.spec.satisfies("%oneapi"):
# -fsycl flag is required for setting up sycl/sycl.hpp seems like
# it doesn't get it from the CMake file
args.append("-DCXX_EXTRA_FLAGS= -fsycl -O3")
# this is required to enable -DCMAKE_CXX_COMPILER=icpx flag from CMake
args.append("-DSYCL_COMPILER=ONEAPI-ICPX")
else:
args.append(
"-DSYCL_COMPILER=" + self.spec.variants["implementation"].value.upper()
)
if self.spec.variants["implementation"].value.upper() != "ONEAPI-DPCPP":
args.append(
"-DSYCL_COMPILER_DIR=" + self.spec.variants["implementation"].value.upper()
)
if self.spec.variants["implementation"].value.upper() == "COMPUTE-CPP":
args.append("-DOpenCL_LIBRARY=")
# ===================================
# HIP(ROCM)
# ===================================
if "+rocm" in self.spec:
hip_comp = self.spec["rocm"].prefix + "/bin/hipcc"
args.append("-DCMAKE_CXX_COMPILER=" + hip_comp)
args.append(
"-DCXX_EXTRA_FLAGS= --offload-arch="
+ self.spec.variants["amdgpu_target"].value
+ " "
+ self.spec.variants["flags"].value
+ " -O3"
)
# ===================================
# TBB
# ===================================
if "+tbb" in self.spec:
args.append("-DONE_TBB_DIR=" + self.spec["tbb"].prefix + "/tbb/latest/")
args.append("-DPARTITIONER=" + self.spec.variants["partitioner"].value.upper())
# ===================================
# OpenCL (ocl)
# ===================================
if "+ocl" in self.spec:
if "backend" in self.spec.variants:
if "cuda" in self.spec.variants["backend"].value:
cuda_dir = self.spec["cuda"].prefix
args.append("-DOpenCL_LIBRARY=" + cuda_dir + "/lib64/libOpenCL.so")
elif "amd" in self.spec.variants["backend"].value:
rocm_dir = self.spec["rocm-opencl"].prefix
args.append("-DOpenCL_LIBRARY=" + rocm_dir + "/lib64/libOpenCL.so")
elif "intel" in self.spec.variants["backend"].value:
intel_lib = (
self.spec["intel-oneapi-compilers"].prefix
+ "/compiler/2023.0.0/linux/lib/libOpenCL.so"
)
args.append("-DOpenCL_LIBRARY=" + intel_lib)
elif "pocl" in self.spec.variants["backend"].value:
args.append("-DCMAKE_CXX_COMPILER=" + self.compiler.cxx)
pocl_lib = self.spec["pocl"].prefix + "/lib64/libOpenCL.so"
args.append("-DOpenCL_LIBRARY=" + pocl_lib)
args.append("-DCMAKE_CXX_COMPILER=" + self.compiler.cxx)
# ===================================
# RAJA
# ===================================
if "+raja" in self.spec:
args.append("-DCMAKE_CXX_COMPILER=" + self.compiler.cxx)
args.append("-DRAJA_IN_TREE=" + self.spec.variants["dir"].value)
if "offload" in self.spec.variants:
if "nvidia" in self.spec.variants["offload"].value:
cuda_dir = self.spec["cuda"].prefix
cuda_comp = cuda_dir + "/bin/nvcc"
args.append("-DCMAKE_CUDA_COMPILER=" + cuda_comp)
args.append("-DTARGET=NVIDIA")
cuda_arch_list = self.spec.variants["cuda_arch"].value
int_cuda_arch = int(cuda_arch_list[0])
cuda_arch = "sm_" + cuda_arch_list[0]
args.append("-DCUDA_ARCH=" + cuda_arch)
args.append("DCUDA_TOOLKIT_ROOT_DIR=" + self.spec["cuda"].prefix)
if self.spec.variants["flags"].value != "none":
args.append("-DCUDA_EXTRA_FLAGS=" + self.spec.variants["flags"].value)
# if("cpu" in self.spec.variants['offload'].value):
if "omp" in self.spec.variants["backend"].value:
args.append("-DENABLE_OPENMP=ON")
if "cuda" in self.spec.variants["backend"].value:
args.append("-DENABLE_CUDA=ON")
# ===================================
# THRUST
# ===================================
if "+thrust" in self.spec:
if "cuda" in self.spec.variants["implementation"].value:
args.append("-DTHRUST_IMPL=" + self.spec.variants["implementation"].value.upper())
args.append("-SDK_DIR=" + self.spec["thrust"].prefix + "/include")
cuda_arch_list = self.spec.variants["cuda_arch"].value
# the architecture value is only number so append sm_ to the name
cuda_arch = "sm_" + cuda_arch_list[0]
args.append("-DCUDA_ARCH=" + cuda_arch)
cuda_dir = self.spec["cuda"].prefix
cuda_comp = cuda_dir + "/bin/nvcc"
args.append("-DCMAKE_CUDA_COMPILER=" + cuda_comp)
args.append("-DBACKEND=" + self.spec.variants["backend"].value.upper())
if self.spec.variants["flags"].value != "none":
args.append("-DCUDA_EXTRA_FLAGS=" + self.spec.variants["flags"].value)
if "rocm" in self.spec.variants["implementation"].value:
args.append("-DTHRUST_IMPL=" + self.spec.variants["implementation"].value.upper())
args.append("-SDK_DIR=" + self.spec["rocthrust"].prefix)
args.append("-DBACKEND=" + self.spec.variants["backend"].value.upper())
# ===================================
# kokkos
# ===================================
# kokkos implementation is versatile and it could use cuda or omp architectures as backend
# The usage should be spack install babelstream +kokkos +cuda [or +omp]
if "+kokkos" in self.spec:
args.append("-DCMAKE_CXX_COMPILER=" + self.compiler.cxx)
args.append("-DKOKKOS_IN_TREE=" + self.spec.variants["dir"].value)
# args.append("-DKOKKOS_IN_PACKAGE=" + self.spec["kokkos"].prefix)
if "backend" in self.spec.variants:
if "cuda" in self.spec.variants["backend"].value:
args.append("-DKokkos_ENABLE_CUDA=ON")
cuda_arch_list = self.spec.variants["cuda_arch"].value
int_cuda_arch = int(cuda_arch_list[0])
# arhitecture kepler optimisations
if int_cuda_arch in (30, 32, 35, 37):
args.append("-D" + "Kokkos_ARCH_KEPLER" + str(int_cuda_arch) + "=ON")
# arhitecture maxwell optimisations
if int_cuda_arch in (50, 52, 53):
args.append("-D" + "Kokkos_ARCH_MAXWELL" + str(int_cuda_arch) + "=ON")
# arhitecture pascal optimisations
if int_cuda_arch in (60, 61):
args.append("-D" + "Kokkos_ARCH_PASCAL" + str(int_cuda_arch) + "=ON")
# architecture volta optimisations
if int_cuda_arch in (70, 72):
args.append("-D" + "Kokkos_ARCH_VOLTA" + str(int_cuda_arch) + "=ON")
if int_cuda_arch == 75:
args.append("-DKokkos_ARCH_TURING75=ON")
if "omp" in self.spec.variants["backend"].value:
args.append("-DKokkos_ENABLE_OPENMP=ON")
# not in ["kokkos", "raja", "acc", "hip"] then compiler forced true
if set(model_list).intersection(["kokkos", "raja", "acc", "hip"]) is True:
args.append("-DCMAKE_CXX_COMPILER_FORCED=True")
return args
|
