1 |
commit: abdf31db4cfb9a43bd7581ab271c03facf933490 |
2 |
Author: Yiyang Wu <xgreenlandforwyy <AT> gmail <DOT> com> |
3 |
AuthorDate: Wed Jul 13 12:18:42 2022 +0000 |
4 |
Commit: Benda XU <heroxbd <AT> gentoo <DOT> org> |
5 |
CommitDate: Thu Sep 22 01:07:37 2022 +0000 |
6 |
URL: https://gitweb.gentoo.org/repo/gentoo.git/commit/?id=abdf31db |
7 |
|
8 |
sci-libs/rocFFT: add 5.1.3, use rocm.eclass |
9 |
|
10 |
profiles/base: mask rocFFT-5.1.3 gfx906 amdgpu target, because |
11 |
intolerable error found on Radeon VII. |
12 |
|
13 |
No need to rebuilt rocFFT when hip upgrade because librocfft is |
14 |
dynamically linked to libhipamd, while llvm SLOT does not matter because |
15 |
it is only used to produce GPU kernels, libclang and libLLVM does not |
16 |
linked in. |
17 |
|
18 |
Signed-off-by: Yiyang Wu <xgreenlandforwyy <AT> gmail.com> |
19 |
Signed-off-by: Benda Xu <heroxbd <AT> gentoo.org> |
20 |
|
21 |
profiles/base/package.use.mask | 8 +- |
22 |
sci-libs/rocFFT/Manifest | 1 + |
23 |
.../files/rocFFT-5.1.3-add-stdexcept-header.patch | 11 ++ |
24 |
.../files/rocFFT-5.1.3-unbundle-sqlite.patch | 67 ++++++++++ |
25 |
sci-libs/rocFFT/rocFFT-5.1.3.ebuild | 144 +++++++++++++++++++++ |
26 |
5 files changed, 230 insertions(+), 1 deletion(-) |
27 |
|
28 |
diff --git a/profiles/base/package.use.mask b/profiles/base/package.use.mask |
29 |
index 39c17ccff6de..13964d29926c 100644 |
30 |
--- a/profiles/base/package.use.mask |
31 |
+++ b/profiles/base/package.use.mask |
32 |
@@ -11,6 +11,12 @@ |
33 |
# Allowing that use flag for >=43.0 would stop users from updating nautilus |
34 |
>=app-text/evince-43.0 nautilus |
35 |
|
36 |
+# Yiyang Wu <xgreenlandforwyy@×××××.com> (2022-09-21) |
37 |
+# Intolerable mathematical errors found on AMD Radeon VII GPU during src_test, |
38 |
+# may lead to incorrect results in computation. |
39 |
+# https://github.com/ROCmSoftwarePlatform/rocFFT/issues/369 |
40 |
+=sci-libs/rocFFT-5.1.3 amdgpu_targets_gfx906 |
41 |
+ |
42 |
# Georgy Yakovlev <gyakovlev@g.o> (2022-09-17) |
43 |
# po4a crashes while translating |
44 |
# mask for now |
45 |
@@ -24,7 +30,7 @@ |
46 |
# bug #870211. |
47 |
dev-lang/rust llvm-libunwind |
48 |
|
49 |
-# Matt Jolly <Matt.Jolly@××××××××.ninja (2022-09-09) |
50 |
+# Matt Jolly <Matt.Jolly@××××××××.ninja> (2022-09-09) |
51 |
# app-antivirus/clamav JIT support requires LLVM <=13 |
52 |
# ebuild support is there but currently failing to link |
53 |
app-antivirus/clamav jit |
54 |
|
55 |
diff --git a/sci-libs/rocFFT/Manifest b/sci-libs/rocFFT/Manifest |
56 |
index a6b401c669d5..62c113ca488d 100644 |
57 |
--- a/sci-libs/rocFFT/Manifest |
58 |
+++ b/sci-libs/rocFFT/Manifest |
59 |
@@ -1,3 +1,4 @@ |
60 |
DIST rocFFT-4.0.0.tar.gz 772176 BLAKE2B f08db9487449e707ba814a66d23992f68c140017a38f72ee45b3b224fc3e46abd40b976aef645f4432601265b38cff8939456981c686360eba135349b5d28fb5 SHA512 db9a6b49162dfb7924aa5f5dc8d42c92a5a741e03c3c0b4ece2cfb72de7506a345f3d28d31a3d7755cfe78078c448c9f952ca671e472335a4301de8702916176 |
61 |
DIST rocFFT-4.3.0.tar.gz 840748 BLAKE2B 236a5faa930751d35e501a9a32ecc8b4075448ef21a0ffb0bb7ec04f4ed1cb67b908aa153b5b424e0823d3e306cf375d7c01a77d687220ed79de6d374037e003 SHA512 765d41e4018a04a53e847c03de80eec830d3e8766187fb8cb493624afa2673175347c16e6eeeebd4ad7d88bb866e58f7e572708f22238723c710b98565a0cd3d |
62 |
DIST rocFFT-5.0.2.tar.gz 837508 BLAKE2B b74d0c6edd7b232f775bffefbca829d88d9d7cd9108c27a0698f45d7fe2760d40463bcda8c710e923874b3bbfb1e83e0fe8622db4dfa28afb29392b3e135e9a5 SHA512 d6810a92a10a27dbcac8a04e8fcd56c7fa198f208da626d8938d5ff28b6ab8b89a058c7a56b5666464165e8f6bd7c4c031a94315625a3613eb02a9b8eccd9eb9 |
63 |
+DIST rocFFT-5.1.3.tar.gz 845470 BLAKE2B 86060ce4f8612869fedc5bc92ed6a5bb7fbc539b26c7f5ab0ded5406ab6c340f64116782285d040824e54557af87d9ddb6f907fd1550b80c3da903d3c1d5e5c0 SHA512 175402856bd4753e74ce14326569be99221e21d0ce675e007900f6982c53e1566ad0b060eec545e0f17ff4dc4811688cc1003dc986226205ee95ee6f22d1ca43 |
64 |
|
65 |
diff --git a/sci-libs/rocFFT/files/rocFFT-5.1.3-add-stdexcept-header.patch b/sci-libs/rocFFT/files/rocFFT-5.1.3-add-stdexcept-header.patch |
66 |
new file mode 100644 |
67 |
index 000000000000..edfd89ea4b70 |
68 |
--- /dev/null |
69 |
+++ b/sci-libs/rocFFT/files/rocFFT-5.1.3-add-stdexcept-header.patch |
70 |
@@ -0,0 +1,11 @@ |
71 |
+Contributed-by: perestoronin https://bugs.gentoo.org/693200#c25 |
72 |
+--- a/library/src/include/twiddles.h |
73 |
++++ b/library/src/include/twiddles.h |
74 |
+@@ -14,6 +14,7 @@ |
75 |
+ #include <numeric> |
76 |
+ #include <tuple> |
77 |
+ #include <vector> |
78 |
++#include <stdexcept> |
79 |
+ |
80 |
+ static const size_t LTWD_BASE_DEFAULT = 8; |
81 |
+ static const size_t LARGE_TWIDDLE_THRESHOLD = 4096; |
82 |
|
83 |
diff --git a/sci-libs/rocFFT/files/rocFFT-5.1.3-unbundle-sqlite.patch b/sci-libs/rocFFT/files/rocFFT-5.1.3-unbundle-sqlite.patch |
84 |
new file mode 100644 |
85 |
index 000000000000..b18b55eb8048 |
86 |
--- /dev/null |
87 |
+++ b/sci-libs/rocFFT/files/rocFFT-5.1.3-unbundle-sqlite.patch |
88 |
@@ -0,0 +1,67 @@ |
89 |
+Disable rocFFT fetching SQLite3 and link to system libsqlite.so |
90 |
+Index: rocFFT-rocm-5.1.3/library/src/CMakeLists.txt |
91 |
+=================================================================== |
92 |
+--- rocFFT-rocm-5.1.3.orig/library/src/CMakeLists.txt |
93 |
++++ rocFFT-rocm-5.1.3/library/src/CMakeLists.txt |
94 |
+@@ -134,49 +134,22 @@ set( rocfft_source |
95 |
+ rtcsubprocess.cpp |
96 |
+ ) |
97 |
+ |
98 |
+-# SQLite 3.36.0 enabled the backup API by default, which we need |
99 |
+-# for cache serialization. We also want to use a static SQLite, |
100 |
+-# and distro static libraries aren't typically built |
101 |
+-# position-independent. |
102 |
+-include( FetchContent ) |
103 |
+- |
104 |
+-set(SQLITE_SRC_URL https://sqlite.org/2021/sqlite-amalgamation-3360000.zip CACHE STRING "Location of SQLite source code") |
105 |
+-set(SQLITE_SRC_SHA256 999826fe4c871f18919fdb8ed7ec9dd8217180854dd1fe21eea96aed36186729 CACHE STRING "SHA256 hash of SQLite source code") |
106 |
+- |
107 |
+-# embed SQLite |
108 |
+-FetchContent_Declare(sqlite_local |
109 |
+- URL ${SQLITE_SRC_URL} |
110 |
+- URL_HASH SHA256=${SQLITE_SRC_SHA256} |
111 |
+-) |
112 |
+-FetchContent_MakeAvailable(sqlite_local) |
113 |
+-add_library( sqlite3 OBJECT ${sqlite_local_SOURCE_DIR}/sqlite3.c ) |
114 |
+-set_target_properties( sqlite3 PROPERTIES |
115 |
+- C_VISIBILITY_PRESET "hidden" |
116 |
+- VISIBILITY_INLINES_HIDDEN ON |
117 |
+- POSITION_INDEPENDENT_CODE ON |
118 |
+- ) |
119 |
+- |
120 |
+-# we don't need extensions, and omitting them from SQLite removes the |
121 |
+-# need for dlopen/dlclose from within rocFFT |
122 |
+-target_compile_options( |
123 |
+- sqlite3 |
124 |
+- PRIVATE -DSQLITE_OMIT_LOAD_EXTENSION |
125 |
+-) |
126 |
+- |
127 |
+ set_property( |
128 |
+ SOURCE rtc.cpp |
129 |
+ PROPERTY OBJECT_DEPENDS ${kgen_embed_h} |
130 |
+ ) |
131 |
+ |
132 |
++find_package( SQLite3 REQUIRED ) |
133 |
++ |
134 |
+ prepend_path( ".." rocfft_headers_public relative_rocfft_headers_public ) |
135 |
+ |
136 |
+ add_library( rocfft |
137 |
+ ${rocfft_source} |
138 |
+ ${relative_rocfft_headers_public} |
139 |
+- $<TARGET_OBJECTS:sqlite3> |
140 |
+ ) |
141 |
+ add_library( roc::rocfft ALIAS rocfft ) |
142 |
+ |
143 |
++target_link_libraries( rocfft PRIVATE ${SQLite3_LIBRARIES} ) |
144 |
+ if( NOT WIN32 ) |
145 |
+ target_link_libraries( rocfft PRIVATE -lstdc++fs ) |
146 |
+ endif() |
147 |
+@@ -209,7 +182,7 @@ endif() |
148 |
+ target_include_directories( rocfft |
149 |
+ PRIVATE $<BUILD_INTERFACE:${PROJECT_SOURCE_DIR}/library/src/include> |
150 |
+ $<BUILD_INTERFACE:${PROJECT_BINARY_DIR}/library/src/device> |
151 |
+- ${sqlite_local_SOURCE_DIR} |
152 |
++ ${SQLite3_LIBRARIES} |
153 |
+ PUBLIC $<BUILD_INTERFACE:${PROJECT_SOURCE_DIR}/library/include> |
154 |
+ $<BUILD_INTERFACE:${PROJECT_BINARY_DIR}/include> |
155 |
+ $<INSTALL_INTERFACE:include> |
156 |
|
157 |
diff --git a/sci-libs/rocFFT/rocFFT-5.1.3.ebuild b/sci-libs/rocFFT/rocFFT-5.1.3.ebuild |
158 |
new file mode 100644 |
159 |
index 000000000000..67d33f28847e |
160 |
--- /dev/null |
161 |
+++ b/sci-libs/rocFFT/rocFFT-5.1.3.ebuild |
162 |
@@ -0,0 +1,144 @@ |
163 |
+# Copyright 1999-2022 Gentoo Authors |
164 |
+# Distributed under the terms of the GNU General Public License v2 |
165 |
+ |
166 |
+EAPI=8 |
167 |
+ |
168 |
+PYTHON_COMPAT=( python3_{8..10} ) |
169 |
+ROCM_VERSION=${PV} |
170 |
+ |
171 |
+inherit cmake check-reqs edo multiprocessing python-r1 rocm |
172 |
+ |
173 |
+DESCRIPTION="Next generation FFT implementation for ROCm" |
174 |
+HOMEPAGE="https://github.com/ROCmSoftwarePlatform/rocFFT" |
175 |
+SRC_URI="https://github.com/ROCmSoftwarePlatform/rocFFT/archive/rocm-${PV}.tar.gz -> rocFFT-${PV}.tar.gz" |
176 |
+ |
177 |
+LICENSE="MIT" |
178 |
+KEYWORDS="~amd64" |
179 |
+SLOT="0/$(ver_cut 1-2)" |
180 |
+ |
181 |
+# RDEPEND: perfscripts? dev-python/plotly[${PYTHON_USEDEP}] # currently masked by arch/amd64/x32/package.mask |
182 |
+RDEPEND=" |
183 |
+perfscripts? ( |
184 |
+ >=media-gfx/asymptote-2.61 |
185 |
+ dev-texlive/texlive-latex |
186 |
+ dev-tex/latexmk |
187 |
+ sys-apps/texinfo |
188 |
+ dev-python/sympy[${PYTHON_USEDEP}] |
189 |
+ dev-python/numpy[${PYTHON_USEDEP}] |
190 |
+ dev-python/scipy[${PYTHON_USEDEP}] |
191 |
+ dev-python/pandas[${PYTHON_USEDEP}] ) |
192 |
+${PYTHON_DEPS}" |
193 |
+ |
194 |
+DEPEND="dev-util/hip |
195 |
+ ${PYTHON_DEPS}" |
196 |
+ |
197 |
+BDEPEND=" |
198 |
+ test? ( dev-cpp/gtest dev-libs/boost |
199 |
+ >=sci-libs/fftw-3 sys-libs/libomp ) |
200 |
+ >=dev-util/cmake-3.22 |
201 |
+ dev-util/rocm-cmake:${SLOT} |
202 |
+" |
203 |
+ |
204 |
+CHECKREQS_DISK_BUILD="7G" |
205 |
+ |
206 |
+IUSE="benchmark perfscripts test" |
207 |
+REQUIRED_USE="perfscripts? ( benchmark ) ${PYTHON_REQUIRED_USE} ${ROCM_REQUIRED_USE}" |
208 |
+ |
209 |
+RESTRICT="!test? ( test )" |
210 |
+ |
211 |
+S="${WORKDIR}/rocFFT-rocm-${PV}" |
212 |
+ |
213 |
+PATCHES=( |
214 |
+ "${FILESDIR}/${PN}-4.2.0-add-functional-header.patch" |
215 |
+ "${FILESDIR}/${PN}-5.1.3-unbundle-sqlite.patch" |
216 |
+ "${FILESDIR}/${PN}-5.0.2-add-math-header.patch" |
217 |
+ "${FILESDIR}/${PN}-5.1.3-add-stdexcept-header.patch" ) |
218 |
+ |
219 |
+required_mem() { |
220 |
+ if use test; then |
221 |
+ echo "52G" |
222 |
+ else |
223 |
+ if [[ -n "${AMDGPU_TARGETS}" ]]; then |
224 |
+ local NARCH=$(($(awk -F";" '{print NF-1}' <<< "${AMDGPU_TARGETS}" || die)+1)) # count how many archs user specified in ${AMDGPU_TARGETS} |
225 |
+ else |
226 |
+ local NARCH=7 # The default number of AMDGPU_TARGETS for rocFFT-4.3.0. May change in the future. |
227 |
+ fi |
228 |
+ echo "$(($(makeopts_jobs)*${NARCH}*25+2200))M" # A linear function estimating how much memory required |
229 |
+ fi |
230 |
+} |
231 |
+ |
232 |
+pkg_pretend() { |
233 |
+ return # leave the disk space check to pkg_setup phase |
234 |
+} |
235 |
+ |
236 |
+pkg_setup() { |
237 |
+ export CHECKREQS_MEMORY=$(required_mem) |
238 |
+ check-reqs_pkg_setup |
239 |
+ python_setup |
240 |
+} |
241 |
+ |
242 |
+src_prepare() { |
243 |
+ sed -e "s/PREFIX rocfft//" \ |
244 |
+ -e "/rocm_install_symlink_subdir/d" \ |
245 |
+ -e "/<INSTALL_INTERFACE/s,include,include/rocFFT," \ |
246 |
+ -i library/src/CMakeLists.txt || die |
247 |
+ |
248 |
+ sed -e "/rocm_install_symlink_subdir/d" \ |
249 |
+ -e "$!N;s:PREFIX\n[ ]*rocfft:# PREFIX rocfft\n:;P;D" \ |
250 |
+ -i library/src/device/CMakeLists.txt || die |
251 |
+ |
252 |
+ if use perfscripts; then |
253 |
+ pushd scripts/perf || die |
254 |
+ sed -e "/\/opt\/rocm/d" -e "/rocmversion/s,rocm_info.strip(),\"${PV}\"," -i perflib/specs.py || dir |
255 |
+ sed -e "/^top/,+1d" -i rocfft-perf suites.py || die |
256 |
+ sed -e "s,perflib,${PN}_perflib,g" -i rocfft-perf suites.py perflib/*.py || die |
257 |
+ sed -e "/^top = /s,__file__).*$,\"${EPREFIX}/usr/share/${PN}-perflib\")," -i perflib/pdf.py perflib/generators.py || die |
258 |
+ popd |
259 |
+ fi |
260 |
+ |
261 |
+ cmake_src_prepare |
262 |
+} |
263 |
+ |
264 |
+src_configure() { |
265 |
+ addpredict /dev/kfd |
266 |
+ addpredict /dev/dri/ |
267 |
+ |
268 |
+ local mycmakeargs=( |
269 |
+ -DCMAKE_SKIP_RPATH=On |
270 |
+ -DAMDGPU_TARGETS="$(get_amdgpu_flags)" |
271 |
+ -Wno-dev |
272 |
+ -DCMAKE_INSTALL_INCLUDEDIR="include/rocFFT/" |
273 |
+ -DBUILD_CLIENTS_TESTS=$(usex test ON OFF) |
274 |
+ -DBUILD_CLIENTS_SELFTEST=$(usex test ON OFF) |
275 |
+ -DPYTHON3_EXE=${EPYTHON} |
276 |
+ -DBUILD_CLIENTS_RIDER=$(usex benchmark ON OFF) |
277 |
+ ) |
278 |
+ |
279 |
+ CXX=hipcc cmake_src_configure |
280 |
+} |
281 |
+ |
282 |
+src_test() { |
283 |
+ check_amdgpu |
284 |
+ cd "${BUILD_DIR}/clients/staging" || die |
285 |
+ export LD_LIBRARY_PATH=${BUILD_DIR}/library/src/:${BUILD_DIR}/library/src/device |
286 |
+ edob ./${PN,,}-test |
287 |
+ edob ./${PN,,}-selftest |
288 |
+} |
289 |
+ |
290 |
+src_install() { |
291 |
+ cmake_src_install |
292 |
+ |
293 |
+ if use benchmark; then |
294 |
+ cd "${BUILD_DIR}"/clients/staging || die |
295 |
+ dobin *rider |
296 |
+ fi |
297 |
+ |
298 |
+ if use perfscripts; then |
299 |
+ cd "${S}"/scripts/perf || die |
300 |
+ python_foreach_impl python_doexe rocfft-perf |
301 |
+ python_moduleinto ${PN}_perflib |
302 |
+ python_foreach_impl python_domodule perflib/*.py |
303 |
+ insinto /usr/share/${PN}-perflib |
304 |
+ doins *.asy suites.py |
305 |
+ fi |
306 |
+} |