Skip to content

Commit 0fad81f

Browse files
author
Raghuveer Devulapalli
authored
Merge pull request #83 from sterrettm2/perfwork
Various performance improvements
2 parents a0eef89 + 68e5393 commit 0fad81f

16 files changed

+1266
-395
lines changed

‎.github/workflows/c-cpp.yml‎

Lines changed: 31 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -125,3 +125,34 @@ jobs:
125125
126126
- name: Run test suite on SPR
127127
run: sde -spr -- ./builddir/testexe
128+
129+
SPR-gcc13-min-networksort:
130+
131+
runs-on: intel-ubuntu-latest
132+
133+
steps:
134+
- uses: actions/checkout@v3
135+
136+
- name: Install dependencies
137+
run: |
138+
sudo apt update
139+
sudo apt -y install g++-13 libgtest-dev meson curl git cmake
140+
141+
- name: Install Intel SDE
142+
run: |
143+
curl -o /tmp/sde.tar.xz https://downloadmirror.intel.com/784319/sde-external-9.24.0-2023-07-13-lin.tar.xz
144+
mkdir /tmp/sde && tar -xvf /tmp/sde.tar.xz -C /tmp/sde/
145+
sudo mv /tmp/sde/* /opt/sde && sudo ln -s /opt/sde/sde64 /usr/bin/sde
146+
147+
- name: Build
148+
env:
149+
CXX: g++-13
150+
CXXFLAGS: -DXSS_MINIMAL_NETWORK_SORT
151+
run: |
152+
make clean
153+
meson setup --warnlevel 2 --werror --buildtype release builddir
154+
cd builddir
155+
ninja
156+
157+
- name: Run test suite on SPR
158+
run: sde -spr -- ./builddir/testexe

‎Makefile‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -1,6 +1,6 @@
11
# When unset, discover g++. Prioritise the latest version on the path.
22
ifeq (, $(and $(strip $(CXX)), $(filter-out default undefined, $(origin CXX))))
3-
override CXX := $(shell which g++-12 g++-11 g++-10 g++-9 g++-8 g++ 2>/dev/null | head -n 1)
3+
override CXX := $(shell which g++-13 g++-12 g++-11 g++-10 g++-9 g++-8 g++ 2>/dev/null | head -n 1)
44
ifeq (, $(strip $(CXX)))
55
$(error Could not locate the g++ compiler. Please manually specify its path using the CXX variable)
66
endif

‎README.md‎

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -170,3 +170,4 @@ Skylake https://arxiv.org/pdf/1704.08579.pdf
170170

171171
* [4] http://mitp-content-server.mit.edu:18180/books/content/sectbyfn?collid=books_pres_0&fn=Chapter%2027.pdf&id=8030
172172

173+
* [5] https://bertdobbelaere.github.io/sorting_networks.html

‎src/avx512-16bit-common.h‎

Lines changed: 216 additions & 26 deletions
Original file line numberDiff line numberDiff line change
@@ -8,7 +8,6 @@
88
#define AVX512_16BIT_COMMON
99

1010
#include "avx512-common-qsort.h"
11-
#include "xss-network-qsort.hpp"
1211

1312
/*
1413
* Constants used in sorting 32 elements in a ZMM registers. Based on Bitonic
@@ -93,30 +92,221 @@ X86_SIMD_SORT_INLINE reg_t sort_zmm_16bit(reg_t zmm)
9392
return zmm;
9493
}
9594

96-
// Assumes zmm is bitonic and performs a recursive half cleaner
97-
template <typename vtype, typename reg_t = typename vtype::reg_t>
98-
X86_SIMD_SORT_INLINE reg_t bitonic_merge_zmm_16bit(reg_t zmm)
99-
{
100-
// 1) half_cleaner[32]: compare 1-17, 2-18, 3-19 etc ..
101-
zmm = cmp_merge<vtype>(
102-
zmm, vtype::permutexvar(vtype::get_network(6), zmm), 0xFFFF0000);
103-
// 2) half_cleaner[16]: compare 1-9, 2-10, 3-11 etc ..
104-
zmm = cmp_merge<vtype>(
105-
zmm, vtype::permutexvar(vtype::get_network(5), zmm), 0xFF00FF00);
106-
// 3) half_cleaner[8]
107-
zmm = cmp_merge<vtype>(
108-
zmm, vtype::permutexvar(vtype::get_network(3), zmm), 0xF0F0F0F0);
109-
// 3) half_cleaner[4]
110-
zmm = cmp_merge<vtype>(
111-
zmm,
112-
vtype::template shuffle<SHUFFLE_MASK(1, 0, 3, 2)>(zmm),
113-
0xCCCCCCCC);
114-
// 3) half_cleaner[2]
115-
zmm = cmp_merge<vtype>(
116-
zmm,
117-
vtype::template shuffle<SHUFFLE_MASK(2, 3, 0, 1)>(zmm),
118-
0xAAAAAAAA);
119-
return zmm;
120-
}
95+
struct avx512_16bit_swizzle_ops {
96+
template <typename vtype, int scale>
97+
X86_SIMD_SORT_INLINE typename vtype::reg_t swap_n(typename vtype::reg_t reg)
98+
{
99+
__m512i v = vtype::cast_to(reg);
100+
101+
if constexpr (scale == 2) {
102+
__m512i mask = _mm512_set_epi16(30,
103+
31,
104+
28,
105+
29,
106+
26,
107+
27,
108+
24,
109+
25,
110+
22,
111+
23,
112+
20,
113+
21,
114+
18,
115+
19,
116+
16,
117+
17,
118+
14,
119+
15,
120+
12,
121+
13,
122+
10,
123+
11,
124+
8,
125+
9,
126+
6,
127+
7,
128+
4,
129+
5,
130+
2,
131+
3,
132+
0,
133+
1);
134+
v = _mm512_permutexvar_epi16(mask, v);
135+
}
136+
else if constexpr (scale == 4) {
137+
v = _mm512_shuffle_epi32(v, (_MM_PERM_ENUM)0b10110001);
138+
}
139+
else if constexpr (scale == 8) {
140+
v = _mm512_shuffle_epi32(v, (_MM_PERM_ENUM)0b01001110);
141+
}
142+
else if constexpr (scale == 16) {
143+
v = _mm512_shuffle_i64x2(v, v, 0b10110001);
144+
}
145+
else if constexpr (scale == 32) {
146+
v = _mm512_shuffle_i64x2(v, v, 0b01001110);
147+
}
148+
else {
149+
static_assert(scale == -1, "should not be reached");
150+
}
151+
152+
return vtype::cast_from(v);
153+
}
154+
155+
template <typename vtype, int scale>
156+
X86_SIMD_SORT_INLINE typename vtype::reg_t
157+
reverse_n(typename vtype::reg_t reg)
158+
{
159+
__m512i v = vtype::cast_to(reg);
160+
161+
if constexpr (scale == 2) { return swap_n<vtype, 2>(reg); }
162+
else if constexpr (scale == 4) {
163+
__m512i mask = _mm512_set_epi16(28,
164+
29,
165+
30,
166+
31,
167+
24,
168+
25,
169+
26,
170+
27,
171+
20,
172+
21,
173+
22,
174+
23,
175+
16,
176+
17,
177+
18,
178+
19,
179+
12,
180+
13,
181+
14,
182+
15,
183+
8,
184+
9,
185+
10,
186+
11,
187+
4,
188+
5,
189+
6,
190+
7,
191+
0,
192+
1,
193+
2,
194+
3);
195+
v = _mm512_permutexvar_epi16(mask, v);
196+
}
197+
else if constexpr (scale == 8) {
198+
__m512i mask = _mm512_set_epi16(24,
199+
25,
200+
26,
201+
27,
202+
28,
203+
29,
204+
30,
205+
31,
206+
16,
207+
17,
208+
18,
209+
19,
210+
20,
211+
21,
212+
22,
213+
23,
214+
8,
215+
9,
216+
10,
217+
11,
218+
12,
219+
13,
220+
14,
221+
15,
222+
0,
223+
1,
224+
2,
225+
3,
226+
4,
227+
5,
228+
6,
229+
7);
230+
v = _mm512_permutexvar_epi16(mask, v);
231+
}
232+
else if constexpr (scale == 16) {
233+
__m512i mask = _mm512_set_epi16(16,
234+
17,
235+
18,
236+
19,
237+
20,
238+
21,
239+
22,
240+
23,
241+
24,
242+
25,
243+
26,
244+
27,
245+
28,
246+
29,
247+
30,
248+
31,
249+
0,
250+
1,
251+
2,
252+
3,
253+
4,
254+
5,
255+
6,
256+
7,
257+
8,
258+
9,
259+
10,
260+
11,
261+
12,
262+
13,
263+
14,
264+
15);
265+
v = _mm512_permutexvar_epi16(mask, v);
266+
}
267+
else if constexpr (scale == 32) {
268+
return vtype::reverse(reg);
269+
}
270+
else {
271+
static_assert(scale == -1, "should not be reached");
272+
}
273+
274+
return vtype::cast_from(v);
275+
}
276+
277+
template <typename vtype, int scale>
278+
X86_SIMD_SORT_INLINE typename vtype::reg_t
279+
merge_n(typename vtype::reg_t reg, typename vtype::reg_t other)
280+
{
281+
__m512i v1 = vtype::cast_to(reg);
282+
__m512i v2 = vtype::cast_to(other);
283+
284+
if constexpr (scale == 2) {
285+
v1 = _mm512_mask_blend_epi16(
286+
0b01010101010101010101010101010101, v1, v2);
287+
}
288+
else if constexpr (scale == 4) {
289+
v1 = _mm512_mask_blend_epi16(
290+
0b00110011001100110011001100110011, v1, v2);
291+
}
292+
else if constexpr (scale == 8) {
293+
v1 = _mm512_mask_blend_epi16(
294+
0b00001111000011110000111100001111, v1, v2);
295+
}
296+
else if constexpr (scale == 16) {
297+
v1 = _mm512_mask_blend_epi16(
298+
0b00000000111111110000000011111111, v1, v2);
299+
}
300+
else if constexpr (scale == 32) {
301+
v1 = _mm512_mask_blend_epi16(
302+
0b00000000000000001111111111111111, v1, v2);
303+
}
304+
else {
305+
static_assert(scale == -1, "should not be reached");
306+
}
307+
308+
return vtype::cast_from(v1);
309+
}
310+
};
121311

122312
#endif // AVX512_16BIT_COMMON

0 commit comments

Comments
 (0)