OV - gmx_mpi - Function clearRVecs(gmx::ArrayRef<gmx::BasicVector<float> >, bool) [clone ._omp

	Function: clearRVecs(gmx::ArrayRef<gmx::BasicVector<float> >, bool) [clone ._omp_fn.0]	Module: libgromacs_mpi.so.7	Source: sim_util.cpp:473-474 [...]	Coverage: 0.05%

/home/eoseret/GROMACS/gromacs-2022.4/api/legacy/include/gromacs/math/vec.h: 292 - 294

--------------------------------------------------------------------------------

292:     a[XX] = 0.0_real;

293:     a[YY] = 0.0_real;

294:     a[ZZ] = 0.0_real;

/home/eoseret/GROMACS/gromacs-2022.4/src/gromacs/mdlib/sim_util.cpp: 473 - 474

--------------------------------------------------------------------------------

473: #pragma omp parallel for num_threads(nth) schedule(static)

474:         for (gmx::index i = 0; i < v.ssize(); i++)

/home/eoseret/GROMACS/gromacs-2022.4/api/legacy/include/gromacs/utility/arrayref.h: 85 - 85

--------------------------------------------------------------------------------

85:     constexpr auto operator-(ArrayRefIter other) const noexcept { return it_ - other.it_; }

0x97de00 STP	X29, X30, [SP, #976]!

0x97de04 ORR	X1, XZR, #3841

0x97de08 MOVK	X1, #43691

0x97de0c ADD	X29, SP, #0

0x97de10 LDR	X0, [X0]

0x97de14 STP	X19, X20, [SP, #16]

0x97de18 LDR	X19, [X0, #8]

0x97de1c STR	X21, [SP, #32]

0x97de20 LDR	X21, [X0]

0x97de24 SUB	X2, X19, X21

0x97de28 SBFM	X3, X2, #2, #63

0x97de2c MADD	X19, X3, X1, XZR

0x97de30 BL	199830

0x97de34 SBFM	X20, X0, #0, #31

0x97de38 BL	18ded0

0x97de3c SBFM	X4, X0, #0, #31

0x97de40 SDIV	X5, X19, X20

0x97de44 MSUB	X6, X5, X20, X19

0x97de48 CMP	X4, X6

0x97de4c B.LT	97dfd0

(17753) 0x97de50 MADD	X7, X5, X4, X6

(17753) 0x97de54 ADD	X8, X5, X7

(17753) 0x97de58 CMP	X7, X8

(17753) 0x97de5c B.GE	97dfc0

(17753) 0x97de60 ADD	X12, X5, X5,LSL #1

(17753) 0x97de64 ADD	X9, X7, X7,LSL #1

(17753) 0x97de68 CMP	X12, #16

(17753) 0x97de6c MOVZ	X10, #16

(17753) 0x97de70 CSEL	X11, X12, X10, #2

(17753) 0x97de74 MOVZ	X14, #8

(17753) 0x97de78 SUB	X13, X11, #16

(17753) 0x97de7c ADD	X0, X21, X9,LSL #2

(17753) 0x97de80 WHILELO	P6.S, XZR, X13

(17753) 0x97de84 MOVZ	X2, #0

(17753) 0x97de88 CMP	X12, X14

(17753) 0x97de8c DUP	Z0.S, #0

(17753) 0x97de90 CSEL	X15, X12, X14, #2

(17753) 0x97de94 SUB	X16, X15, #8

(17753) 0x97de98 WHILELO	P7.S, XZR, X12

(17753) 0x97de9c WHILELO	P0.S, XZR, X16

  (17754) 0x97dea0 ADD	X17, X2, #24

  (17754) 0x97dea4 WHILELO	P3.S, X17, X13

  (17754) 0x97dea8 WHILELO	P4.S, X17, X16

  (17754) 0x97deac ST1W	{Z0.S}, P7, [X0, MUL VL]

  (17754) 0x97deb0 ST1W	{Z0.S}, P0, [X0, #1, MUL VL]

  (17754) 0x97deb4 ST1W	{Z0.S}, P6, [X0, #2, MUL VL]

  (17754) 0x97deb8 ADD	X18, X2, #48

  (17754) 0x97debc ADD	X30, X2, #72

  (17754) 0x97dec0 ADD	X1, X2, #96

  (17754) 0x97dec4 ADD	X21, X2, #120

  (17754) 0x97dec8 ADD	X3, X2, #144

  (17754) 0x97decc ADD	X19, X0, #96

  (17754) 0x97ded0 WHILELO	P0.S, X17, X12

  (17754) 0x97ded4 B.EQ	97dfc0

  (17754) 0x97ded8 WHILELO	P1.S, X18, X13

  (17754) 0x97dedc WHILELO	P2.S, X18, X16

  (17754) 0x97dee0 ST1W	{Z0.S}, P0, [X0, #3, MUL VL]

  (17754) 0x97dee4 ST1W	{Z0.S}, P4, [X19, #1, MUL VL]

  (17754) 0x97dee8 ST1W	{Z0.S}, P3, [X19, #2, MUL VL]

  (17754) 0x97deec ADD	X20, X2, #168

  (17754) 0x97def0 ADD	X4, X0, #192

  (17754) 0x97def4 WHILELO	P5.S, X18, X12

  (17754) 0x97def8 B.EQ	97dfc0

  (17754) 0x97defc WHILELO	P6.S, X30, X13

  (17754) 0x97df00 WHILELO	P7.S, X30, X16

  (17754) 0x97df04 ST1W	{Z0.S}, P5, [X0, #6, MUL VL]

  (17754) 0x97df08 ST1W	{Z0.S}, P2, [X4, #1, MUL VL]

  (17754) 0x97df0c ST1W	{Z0.S}, P1, [X4, #2, MUL VL]

  (17754) 0x97df10 ADD	X2, X2, #192

  (17754) 0x97df14 ADD	X5, X0, #288

  (17754) 0x97df18 WHILELO	P3.S, X30, X12

  (17754) 0x97df1c B.EQ	97dfc0

  (17754) 0x97df20 WHILELO	P1.S, X1, X13

  (17754) 0x97df24 WHILELO	P2.S, X1, X16

  (17754) 0x97df28 ST1W	{Z0.S}, P3, [X5, MUL VL]

  (17754) 0x97df2c ST1W	{Z0.S}, P7, [X5, #1, MUL VL]

  (17754) 0x97df30 ST1W	{Z0.S}, P6, [X5, #2, MUL VL]

  (17754) 0x97df34 ADD	X6, X0, #384

  (17754) 0x97df38 WHILELO	P0.S, X1, X12

  (17754) 0x97df3c B.EQ	97dfc0

  (17754) 0x97df40 WHILELO	P5.S, X21, X13

  (17754) 0x97df44 WHILELO	P4.S, X21, X16

  (17754) 0x97df48 ST1W	{Z0.S}, P0, [X6, MUL VL]

  (17754) 0x97df4c ST1W	{Z0.S}, P2, [X6, #1, MUL VL]

  (17754) 0x97df50 ST1W	{Z0.S}, P1, [X6, #2, MUL VL]

  (17754) 0x97df54 ADD	X7, X0, #480

  (17754) 0x97df58 WHILELO	P6.S, X21, X12

  (17754) 0x97df5c B.EQ	97dfc0

  (17754) 0x97df60 WHILELO	P7.S, X3, X13

  (17754) 0x97df64 WHILELO	P1.S, X3, X16

  (17754) 0x97df68 ST1W	{Z0.S}, P6, [X7, MUL VL]

  (17754) 0x97df6c ST1W	{Z0.S}, P4, [X7, #1, MUL VL]

  (17754) 0x97df70 ST1W	{Z0.S}, P5, [X7, #2, MUL VL]

  (17754) 0x97df74 ADD	X8, X0, #576

  (17754) 0x97df78 WHILELO	P2.S, X3, X12

  (17754) 0x97df7c B.EQ	97dfc0

  (17754) 0x97df80 WHILELO	P3.S, X20, X13

  (17754) 0x97df84 WHILELO	P5.S, X20, X16

  (17754) 0x97df88 ST1W	{Z0.S}, P2, [X8, MUL VL]

  (17754) 0x97df8c ST1W	{Z0.S}, P1, [X8, #1, MUL VL]

  (17754) 0x97df90 ST1W	{Z0.S}, P7, [X8, #2, MUL VL]

  (17754) 0x97df94 ADD	X9, X0, #672

  (17754) 0x97df98 WHILELO	P4.S, X20, X12

  (17754) 0x97df9c B.EQ	97dfc0

  (17754) 0x97dfa0 WHILELO	P6.S, X2, X13

  (17754) 0x97dfa4 WHILELO	P0.S, X2, X16

  (17754) 0x97dfa8 ST1W	{Z0.S}, P4, [X9, MUL VL]

  (17754) 0x97dfac ST1W	{Z0.S}, P5, [X9, #1, MUL VL]

  (17754) 0x97dfb0 ST1W	{Z0.S}, P3, [X9, #2, MUL VL]

  (17754) 0x97dfb4 ADD	X0, X0, #768

  (17754) 0x97dfb8 WHILELO	P7.S, X2, X12

  (17754) 0x97dfbc B.NE	97dea0

(17753) 0x97dfc0 LDP	X19, X20, [SP, #16]

(17753) 0x97dfc4 LDR	X21, [SP, #32]

(17753) 0x97dfc8 LDP	X29, X30, [SP], #48

(17753) 0x97dfcc RET

(17753) 0x97dfd0 ADD	X5, X5, #1

(17753) 0x97dfd4 MOVZ	X6, #0

(17753) 0x97dfd8 B	97de50

0x97dfdc HINT	#0

Coverage (%)	Name	Module
►96.19+	__kmp_GOMP_microtask_wrapper(i[...]	libomp.so
○	__kmp_invoke_microtask	libomp.so
○2.86	__kmp_invoke_microtask	libomp.so

Path /

Average path: Display a virtual path defined by average values of all real paths

Source file and lines	sim_util.cpp:473-474
Module	libgromacs_mpi.so.7

The function is defined in:

/home/eoseret/GROMACS/gromacs-2022.4/api/legacy/include/gromacs/math/vec.h:292-294
/home/eoseret/GROMACS/gromacs-2022.4/src/gromacs/mdlib/sim_util.cpp:473-474
/home/eoseret/GROMACS/gromacs-2022.4/api/legacy/include/gromacs/utility/arrayref.h:85

Warnings:
Ignoring paths for analysis

gain
potential
hint
expert

Vectorization

Your function is not vectorized. 4 data elements could be processed at once in vector registers. By vectorizing your function, you can lower the cost of an iteration from 9.08 to 0.87 cycles (10.38x speedup).

Details

All VPU instructions are used in scalar version (process only one data element in vector registers). Since your execution units are vector units, only a vectorized function can use their full power.

Workaround

Try another compiler or update/tune your current one
Make array accesses unit-stride:
- If your function streams arrays of structures (AoS), try to use structures of arrays instead (SoA)

Expensive FP math instructions/calls

Detected performance impact from expensive FP math instructions/calls. By removing/reexpressing them, you can lower the cost of an iteration from 9.08 to 1.00 cycles (9.08x speedup).

Matching between your function (in the source code) and the binary function

The binary function does not contain any FP arithmetical operations. The binary function is loading 24 bytes. The binary function is storing 40 bytes.

General properties

nb instructions	21
loop length	84
nb stack references	0

Front-end

FIT IN UOP CACHE

front end

2.50 cycles

Back-end

	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11	P12	P13	P14
uops	1.50	1.50	3.50	3.50	3.50	3.50	0.00	0.00	0.00	0.00	2.17	1.83	2.00	1.50	1.50
cycles	1.50	1.50	3.50	3.50	3.50	3.50	0.00	0.00	0.00	0.00	2.17	1.83	2.00	1.50	1.50

Cycles executing div or sqrt instructions

1.00-0.50

Cycles summary

Front-end	2.50
Overall L1	9.08

Vectorization ratios

all	0%
load	NA (no load vectorizable/vectorized instructions)
store	NA (no store vectorizable/vectorized instructions)
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	0%
fma	0%
div/sqrt	0%
other	0%

Cycles and memory resources usage

Assuming all data fit into the L1 cache, each call to the function takes 9.08 cycles. At this rate:

2% of peak load performance is reached (2.64 out of 96.00 bytes loaded per cycle (GB/s @ 1GHz))
9% of peak store performance is reached (4.40 out of 48.00 bytes stored per cycle (GB/s @ 1GHz))

ASM code

In the binary file, the address of the function is: 97de00

Instruction	Nb FU	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11	P12	P13	P14	Latency	Recip. throughput
STP X29, X30, [SP, #976]!	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0.50	0.50	0	0.50	0.50	1	0.50
ORR X1, XZR, #3841	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	0	0	1	0.25
MOVK X1, #43691	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	0	0	1	0.25
ADD X29, SP, #0	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	0	0	1	0.25
LDR X0, [X0]	1	0	0	0	0	0	0	0	0	0	0	0.33	0.33	0.33	0	0	4	0.33
STP X19, X20, [SP, #16]	1	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0.50	0.50	1	0.50
LDR X19, [X0, #8]	1	0	0	0	0	0	0	0	0	0	0	0.33	0.33	0.33	0	0	4	0.33
STR X21, [SP, #32]	1	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0.50	0.50	1	0.50
LDR X21, [X0]	1	0	0	0	0	0	0	0	0	0	0	0.33	0.33	0.33	0	0	4	0.33
SUB X2, X19, X21	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	0	0	1	0.25
SBFM X3, X2, #2, #63	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	0	0	1	0.25
MADD X19, X3, X1, XZR	1	0	0	0	0	1	0	0	0	0	0	0	0	0	0	0	2	1
BL 199830 <@plt_start@+0xe910>	1	0.50	0.50	0.50	0.50	0	0	0	0	0	0	0	0	0	0	0	1	0.50
SBFM X20, X0, #0, #31	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	0	0	1	0.25
BL 18ded0 <@plt_start@+0x2fb0>	1	0.50	0.50	0.50	0.50	0	0	0	0	0	0	0	0	0	0	0	1	0.50
SBFM X4, X0, #0, #31	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	0	0	1	0.25
SDIV X5, X19, X20	1	0	0	0	0	1	0	0	0	0	0	0	0	0	0	0	5-20	1-0.50
MSUB X6, X5, X20, X19	1	0	0	0	0	1	0	0	0	0	0	0	0	0	0	0	2	1
CMP X4, X6	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	0	0	1	0.33
B.LT 97dfd0 <_ZL10clearRVecsN3gmx8ArrayRefINS_11BasicVectorIfEEEEb._omp_fn.0+0x1d0>	1	0.50	0.50	0	0	0	0	0	0	0	0	0	0	0	0	0	1	0.50
HINT #0

Source file and lines	sim_util.cpp:473-474
Module	libgromacs_mpi.so.7

The function is defined in:

/home/eoseret/GROMACS/gromacs-2022.4/api/legacy/include/gromacs/math/vec.h:292-294
/home/eoseret/GROMACS/gromacs-2022.4/src/gromacs/mdlib/sim_util.cpp:473-474
/home/eoseret/GROMACS/gromacs-2022.4/api/legacy/include/gromacs/utility/arrayref.h:85

Warnings:
Ignoring paths for analysis

gain
potential
hint
expert

Vectorization

Details

All VPU instructions are used in scalar version (process only one data element in vector registers). Since your execution units are vector units, only a vectorized function can use their full power.

Workaround

Try another compiler or update/tune your current one
Make array accesses unit-stride:
- If your function streams arrays of structures (AoS), try to use structures of arrays instead (SoA)

Expensive FP math instructions/calls

Detected performance impact from expensive FP math instructions/calls. By removing/reexpressing them, you can lower the cost of an iteration from 9.08 to 1.00 cycles (9.08x speedup).

Matching between your function (in the source code) and the binary function

The binary function does not contain any FP arithmetical operations. The binary function is loading 24 bytes. The binary function is storing 40 bytes.

General properties

nb instructions	21
loop length	84
nb stack references	0

Front-end

FIT IN UOP CACHE

front end

2.50 cycles

Back-end

	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11	P12	P13	P14
uops	1.50	1.50	3.50	3.50	3.50	3.50	0.00	0.00	0.00	0.00	2.17	1.83	2.00	1.50	1.50
cycles	1.50	1.50	3.50	3.50	3.50	3.50	0.00	0.00	0.00	0.00	2.17	1.83	2.00	1.50	1.50

Cycles executing div or sqrt instructions

1.00-0.50

Cycles summary

Front-end	2.50
Overall L1	9.08

Vectorization ratios

all	0%
load	NA (no load vectorizable/vectorized instructions)
store	NA (no store vectorizable/vectorized instructions)
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	0%
fma	0%
div/sqrt	0%
other	0%

Cycles and memory resources usage

Assuming all data fit into the L1 cache, each call to the function takes 9.08 cycles. At this rate:

2% of peak load performance is reached (2.64 out of 96.00 bytes loaded per cycle (GB/s @ 1GHz))
9% of peak store performance is reached (4.40 out of 48.00 bytes stored per cycle (GB/s @ 1GHz))

ASM code

In the binary file, the address of the function is: 97de00

Instruction	Nb FU	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11	P12	P13	P14	Latency	Recip. throughput
STP X29, X30, [SP, #976]!	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0.50	0.50	0	0.50	0.50	1	0.50
ORR X1, XZR, #3841	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	0	0	1	0.25
MOVK X1, #43691	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	0	0	1	0.25
ADD X29, SP, #0	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	0	0	1	0.25
LDR X0, [X0]	1	0	0	0	0	0	0	0	0	0	0	0.33	0.33	0.33	0	0	4	0.33
STP X19, X20, [SP, #16]	1	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0.50	0.50	1	0.50
LDR X19, [X0, #8]	1	0	0	0	0	0	0	0	0	0	0	0.33	0.33	0.33	0	0	4	0.33
STR X21, [SP, #32]	1	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0.50	0.50	1	0.50
LDR X21, [X0]	1	0	0	0	0	0	0	0	0	0	0	0.33	0.33	0.33	0	0	4	0.33
SUB X2, X19, X21	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	0	0	1	0.25
SBFM X3, X2, #2, #63	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	0	0	1	0.25
MADD X19, X3, X1, XZR	1	0	0	0	0	1	0	0	0	0	0	0	0	0	0	0	2	1
BL 199830 <@plt_start@+0xe910>	1	0.50	0.50	0.50	0.50	0	0	0	0	0	0	0	0	0	0	0	1	0.50
SBFM X20, X0, #0, #31	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	0	0	1	0.25
BL 18ded0 <@plt_start@+0x2fb0>	1	0.50	0.50	0.50	0.50	0	0	0	0	0	0	0	0	0	0	0	1	0.50
SBFM X4, X0, #0, #31	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	0	0	1	0.25
SDIV X5, X19, X20	1	0	0	0	0	1	0	0	0	0	0	0	0	0	0	0	5-20	1-0.50
MSUB X6, X5, X20, X19	1	0	0	0	0	1	0	0	0	0	0	0	0	0	0	0	2	1
CMP X4, X6	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	0	0	1	0.33
B.LT 97dfd0 <_ZL10clearRVecsN3gmx8ArrayRefINS_11BasicVectorIfEEEEb._omp_fn.0+0x1d0>	1	0.50	0.50	0	0	0	0	0	0	0	0	0	0	0	0	0	1	0.50
HINT #0

Name	Coverage (%)	Time (s)
▼clearRVecs(gmx::ArrayRef >, bool) [clone ._omp_fn.0]–	0.05	0.02
▼Loop 17753 - arrayref.h:85-85 - libgromacs_mpi.so.7–	0	0
○Loop 17754 - vec.h:292-294 - libgromacs_mpi.so.7	0.04	0.02

Report Configuration

Vectorization

Details

Workaround

Expensive FP math instructions/calls

Matching between your function (in the source code) and the binary function

General properties

Front-end

Back-end

Cycles summary

Vectorization ratios

Cycles and memory resources usage

ASM code

Vectorization

Details

Workaround

Expensive FP math instructions/calls

Matching between your function (in the source code) and the binary function

General properties

Front-end

Back-end

Cycles summary

Vectorization ratios

Cycles and memory resources usage

ASM code