OV - - Function sortAtomsById

/beegfs/hackathon/users/eoseret/qaas_runs/170-850-7424/intel/CoMD/build/CoMD/CoMD/src-openmp/haloExchange.c: 655 - 661

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

655:    int bId = ((AtomMsg*) b)->gid;

656:    assert(aId != bId);

657:

658:    if (aId < bId)

659:       return -1;

660:    return 1;

661: }

0x40bf90 MOV	(%RSI),%EAX

0x40bf92 CMP	%EAX,(%RDI)

0x40bf94 JE	40bfa4

0x40bf96 SETGE	%AL

0x40bf99 MOVZX	%AL,%EAX

0x40bf9c LEA	-0x1(,%RAX,2),%EAX

0x40bfa3 RET

0x40bfa4 PUSH	%RBP

0x40bfa5 MOV	%RSP,%RBP

0x40bfa8 MOV	$0x42c671,%EDI

0x40bfad MOV	$0x42c49f,%ESI

0x40bfb2 MOV	$0x42c67c,%ECX

0x40bfb7 MOV	$0x290,%EDX

0x40bfbc CALL	405100 <__assert_fail@plt>

0x40bfc1 NOPW	%CS:(%RAX,%RAX,1)

Path /

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

Source file and lines	haloExchange.c:655-661
Module	exec

The function is defined in /beegfs/hackathon/users/eoseret/qaas_runs/170-850-7424/intel/CoMD/build/CoMD/CoMD/src-openmp/haloExchange.c:655-661.

The structure of this function is probably <if then [else] end>.

The presence of multiple execution paths is typically the main/first bottleneck.
Try to simplify control inside function: ideally, try to remove all conditional expressions, for example by (if applicable):

hoisting them (moving them outside the function)
turning them into conditional moves, MIN or MAX

Ex: if (x<0) x=0 => x = (x<0 ? 0 : x) (or MAX(0,x) after defining the corresponding macro)

gain
potential
hint
expert

Vectorization

Your function is not vectorized. Only 8% of vector register length is used (average across all SSE/AVX instructions).

Details

All SSE/AVX instructions are used in scalar version (process only one data element in vector registers).

No data for this section

Type of elements and instruction set

No instructions are processing arithmetic or math operations on FP elements. This function is probably writing/copying data or processing integer elements.

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 8 bytes.

General properties

nb instructions	8.50
nb uops	9
loop length	27.50
used x86 registers	5
used mmx registers	0
used xmm registers	0
used ymm registers	0
used zmm registers	0
nb stack references	0

Front-end

MACRO FUSION NOT POSSIBLE

micro-operation queue	1.50 cycles
front end	1.50 cycles

Back-end

	ALU0/BRU0	ALU1	ALU2	ALU3	BRU1	AGU0	AGU1	AGU2	FP0	FP1	FP2	FP3	FP4	FP5
uops	1.50	1.38	1.38	1.25	1.00	0.83	0.83	0.83	0.00	0.00	0.00	0.00	0.00	0.00
cycles	1.50	1.38	1.38	1.25	1.00	0.83	0.83	0.83	0.00	0.00	0.00	0.00	0.00	0.00

Execution ports to units layout:

ALU0/BRU0: ALU
ALU1: ALU
ALU2: ALU
ALU3: ALU
BRU1:
AGU0 (256 bits): store address, load
AGU1 (256 bits): store address, load
AGU2 (256 bits): store address, load
FP0 (256 bits): VPU, DIV/SQRT
FP1 (256 bits): VPU, DIV/SQRT
FP2 (256 bits): VPU
FP3 (256 bits): VPU
FP4 (256 bits): FP store data
FP5 (256 bits): FP store data

Cycles executing div or sqrt instructions

NA

Cycles summary

Front-end	1.50
Dispatch	1.50
Overall L1	1.67

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

Vector efficiency ratios

all	8%
load	NA (no load vectorizable/vectorized instructions)
store	NA (no store vectorizable/vectorized instructions)
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	NA (no add-sub vectorizable/vectorized instructions)
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	8%

Cycles and memory resources usage

Assuming all data fit into the L1 cache, each call to the function takes 1.67 cycles. At this rate:

5% of peak load performance is reached (4.85 out of 96.00 bytes loaded per cycle (GB/s @ 1GHz))

Source file and lines	haloExchange.c:655-661
Module	exec

The function is defined in /beegfs/hackathon/users/eoseret/qaas_runs/170-850-7424/intel/CoMD/build/CoMD/CoMD/src-openmp/haloExchange.c:655-661.

The structure of this function is probably <if then [else] end>.

The presence of multiple execution paths is typically the main/first bottleneck.
Try to simplify control inside function: ideally, try to remove all conditional expressions, for example by (if applicable):

hoisting them (moving them outside the function)
turning them into conditional moves, MIN or MAX

Ex: if (x<0) x=0 => x = (x<0 ? 0 : x) (or MAX(0,x) after defining the corresponding macro)

gain
potential
hint
expert

Vectorization

Your function is not vectorized.

Details

No SSE/AVX instruction: only general purpose registers are used and data elements are processed one at a time.

Execution units bottlenecks

Found no such bottlenecks but see expert reports for more complex bottlenecks.

No data for this section

Type of elements and instruction set

No instructions are processing arithmetic or math operations on FP elements. This function is probably writing/copying data or processing integer elements.

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 8 bytes.

General properties

nb instructions	7
nb uops	7
loop length	20
used x86 registers	3
used mmx registers	0
used xmm registers	0
used ymm registers	0
used zmm registers	0
nb stack references	0

Front-end

MACRO FUSION NOT POSSIBLE FIT IN UOP CACHE

micro-operation queue	1.17 cycles
front end	1.17 cycles

Back-end

	ALU0/BRU0	ALU1	ALU2	ALU3	BRU1	AGU0	AGU1	AGU2	FP0	FP1	FP2	FP3	FP4	FP5
uops	1.50	1.25	1.25	1.00	1.00	0.67	0.67	0.67	0.00	0.00	0.00	0.00	0.00	0.00
cycles	1.50	1.25	1.25	1.00	1.00	0.67	0.67	0.67	0.00	0.00	0.00	0.00	0.00	0.00

Execution ports to units layout:

ALU0/BRU0: ALU
ALU1: ALU
ALU2: ALU
ALU3: ALU
BRU1:
AGU0 (256 bits): store address, load
AGU1 (256 bits): store address, load
AGU2 (256 bits): store address, load
FP0 (256 bits): VPU, DIV/SQRT
FP1 (256 bits): VPU, DIV/SQRT
FP2 (256 bits): VPU
FP3 (256 bits): VPU
FP4 (256 bits): FP store data
FP5 (256 bits): FP store data

Cycles executing div or sqrt instructions

NA

Cycles summary

Front-end	1.17
Dispatch	1.50
Overall L1	1.50

Vectorization ratios

No vectorizable/vectorized instructions

Vector efficiency ratios

No vectorizable/vectorized instructions

Cycles and memory resources usage

Assuming all data fit into the L1 cache, each call to the function takes 1.50 cycles. At this rate:

5% of peak load performance is reached (5.33 out of 96.00 bytes loaded per cycle (GB/s @ 1GHz))

Front-end bottlenecks

Found no such bottlenecks.

ASM code

In the binary file, the address of the function is: 40bf90

Instruction	Nb FU	ALU0/BRU0	ALU1	ALU2	ALU3	BRU1	AGU0	AGU1	AGU2	Latency	Recip. throughput
MOV (%RSI),%EAX	1	0	0	0	0	0	0.33	0.33	0.33	3	0.33
CMP %EAX,(%RDI)	1	0.25	0.25	0.25	0.25	0	0.33	0.33	0.33	1	0.33
JE 40bfa4 <sortAtomsById+0x14>	1	0.50	0	0	0	0.50	0	0	0	1	0.50-1
SETGE %AL	1	0.50	0	0	0.50	0	0	0	0	1	0.50
MOVZX %AL,%EAX	1	0.25	0.25	0.25	0.25	0	0	0	0	1	0.25
LEA -0x1(,%RAX,2),%EAX	1	0.25	0.25	0.25	0.25	0	0	0	0	1	0.25
RET	1	0.50	0	0	0	0.50	0	0	0	1	0.50

Source file and lines	haloExchange.c:655-661
Module	exec

The function is defined in /beegfs/hackathon/users/eoseret/qaas_runs/170-850-7424/intel/CoMD/build/CoMD/CoMD/src-openmp/haloExchange.c:655-661.

The structure of this function is probably <if then [else] end>.

The presence of multiple execution paths is typically the main/first bottleneck.
Try to simplify control inside function: ideally, try to remove all conditional expressions, for example by (if applicable):

hoisting them (moving them outside the function)
turning them into conditional moves, MIN or MAX

Ex: if (x<0) x=0 => x = (x<0 ? 0 : x) (or MAX(0,x) after defining the corresponding macro)

gain
potential
hint
expert

Vectorization

Your function is not vectorized. Only 8% of vector register length is used (average across all SSE/AVX instructions). By vectorizing your function, you can lower the cost of an iteration from 1.83 to 0.11 cycles (16.00x speedup).

Details

All SSE/AVX 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): for(i) a[i].x = b[i].x; (slow, non stride 1) => for(i) a.x[i] = b.x[i]; (fast, stride 1)

Execution units bottlenecks

Found no such bottlenecks but see expert reports for more complex bottlenecks.

No data for this section

CALL instructions

Detected function call instructions.

Details

Calling (and then returning from) a function prevents many compiler optimizations (like vectorization), breaks control flow (which reduces pipeline performance) and executes extra instructions to save/restore the registers used inside it, which is very expensive (dozens of cycles). Consider to inline small functions.

__assert_fail@plt: 1 occurrences

Type of elements and instruction set

No instructions are processing arithmetic or math operations on FP elements. This function is probably writing/copying data or processing integer elements.

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 8 bytes.

General properties

nb instructions	10
nb uops	11
loop length	35
used x86 registers	7
used mmx registers	0
used xmm registers	0
used ymm registers	0
used zmm registers	0
nb stack references	0

Front-end

MACRO FUSION NOT POSSIBLE FIT IN UOP CACHE

micro-operation queue	1.83 cycles
front end	1.83 cycles

Back-end

	ALU0/BRU0	ALU1	ALU2	ALU3	BRU1	AGU0	AGU1	AGU2	FP0	FP1	FP2	FP3	FP4	FP5
uops	1.50	1.50	1.50	1.50	1.00	1.00	1.00	1.00	0.00	0.00	0.00	0.00	0.00	0.00
cycles	1.50	1.50	1.50	1.50	1.00	1.00	1.00	1.00	0.00	0.00	0.00	0.00	0.00	0.00

Execution ports to units layout:

ALU0/BRU0: ALU
ALU1: ALU
ALU2: ALU
ALU3: ALU
BRU1:
AGU0 (256 bits): store address, load
AGU1 (256 bits): store address, load
AGU2 (256 bits): store address, load
FP0 (256 bits): VPU, DIV/SQRT
FP1 (256 bits): VPU, DIV/SQRT
FP2 (256 bits): VPU
FP3 (256 bits): VPU
FP4 (256 bits): FP store data
FP5 (256 bits): FP store data

Cycles executing div or sqrt instructions

NA

Cycles summary

Front-end	1.83
Dispatch	1.50
Overall L1	1.83

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

Vector efficiency ratios

all	8%
load	NA (no load vectorizable/vectorized instructions)
store	NA (no store vectorizable/vectorized instructions)
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	NA (no add-sub vectorizable/vectorized instructions)
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	8%

Cycles and memory resources usage

Assuming all data fit into the L1 cache, each call to the function takes 1.83 cycles. At this rate:

4% of peak load performance is reached (4.36 out of 96.00 bytes loaded per cycle (GB/s @ 1GHz))

Front-end bottlenecks

Performance is limited by instruction throughput (loading/decoding program instructions to execution core) (front-end is a bottleneck). By removing all these bottlenecks, you can lower the cost of an iteration from 1.83 to 1.50 cycles (1.22x speedup).

ASM code

In the binary file, the address of the function is: 40bf90

Instruction	Nb FU	ALU0/BRU0	ALU1	ALU2	ALU3	BRU1	AGU0	AGU1	AGU2	Latency	Recip. throughput
MOV (%RSI),%EAX	1	0	0	0	0	0	0.33	0.33	0.33	3	0.33
CMP %EAX,(%RDI)	1	0.25	0.25	0.25	0.25	0	0.33	0.33	0.33	1	0.33
JE 40bfa4 <sortAtomsById+0x14>	1	0.50	0	0	0	0.50	0	0	0	1	0.50-1
PUSH %RBP	1	0	0	0	0	0	0	0	0	1	0.50
MOV %RSP,%RBP	1	0	0	0	0	0	0	0	0	0	0.17
MOV $0x42c671,%EDI	1	0.25	0.25	0.25	0.25	0	0	0	0	1	0.25
MOV $0x42c49f,%ESI	1	0.25	0.25	0.25	0.25	0	0	0	0	1	0.25
MOV $0x42c67c,%ECX	1	0.25	0.25	0.25	0.25	0	0	0	0	1	0.25
MOV $0x290,%EDX	1	0.25	0.25	0.25	0.25	0	0	0	0	1	0.25
CALL 405100 <__assert_fail@plt>	2	0.50	0	0	0	0.50	0.33	0.33	0.33	0	0.50

▶Scalability Runs Description

Run 2x1	Number processes: 1Number nodes: 1Run Command: <executable> -x 100 -y 100 -z 100 --xproc=2 --yproc=1 --zproc=1MPI Command: mpirun -np 2Dataset: Run Directory: /beegfs/hackathon/users/eoseret/qaas_runs/170-850-7424/intel/CoMD/run/oneview_runs/compilers/icx_6/oneview_run_1708509234I_MPI_PIN_DOMAIN: auto:scatterOMP_PLACES: threadsOMP_PROC_BIND: spreadOMP_NUM_THREADS: 1
Run 2x2	OMP_NUM_THREADS: 2I_MPI_PIN_DOMAIN: auto:scatterOMP_PLACES: threadsOMP_PROC_BIND: spread
Run 2x4	OMP_NUM_THREADS: 4I_MPI_PIN_DOMAIN: auto:scatterOMP_PLACES: threadsOMP_PROC_BIND: spread
Run 2x8	OMP_NUM_THREADS: 8I_MPI_PIN_DOMAIN: auto:scatterOMP_PLACES: threadsOMP_PROC_BIND: spread
Run 2x16	OMP_NUM_THREADS: 16I_MPI_PIN_DOMAIN: auto:scatterOMP_PLACES: threadsOMP_PROC_BIND: spread
Run 2x32	OMP_NUM_THREADS: 32I_MPI_PIN_DOMAIN: auto:scatterOMP_PLACES: threadsOMP_PROC_BIND: spread
Run 2x64	OMP_NUM_THREADS: 64I_MPI_PIN_DOMAIN: auto:scatterOMP_PLACES: threadsOMP_PROC_BIND: spread
Run 2x96	OMP_NUM_THREADS: 96I_MPI_PIN_DOMAIN: auto:scatterOMP_PLACES: threadsOMP_PROC_BIND: spread

(2x1) Efficiency	(2x1) Potential Speed-Up (%)	(2x2) Efficiency	(2x2) Potential Speed-Up (%)	(2x4) Efficiency	(2x4) Potential Speed-Up (%)	(2x8) Efficiency	(2x8) Potential Speed-Up (%)	(2x16) Efficiency	(2x16) Potential Speed-Up (%)	(2x32) Efficiency	(2x32) Potential Speed-Up (%)	(2x64) Efficiency	(2x64) Potential Speed-Up (%)	(2x96) Efficiency	(2x96) Potential Speed-Up (%)
1	0	0.95	0.01	0.95	0.01	1.05	0	0.88	0.01	1.31	0	0.66	0.02	0.44	0.03

Speed-Up and Efficiency

▶Detailed Speed-Up and Efficiency

Run	Number of threads	Efficiency (ideal is 1)	Speedup	Ideal Speedup	Time (s)	Coverage (%)
2x1	2	1	1	1	0.43	0.14
2x2	4	0.95	1.91	2	0.24	0.15
2x4	8	0.95	3.82	4	0.16	0.14
2x8	16	1.05	8.4	8	0.07	0.11
2x16	32	0.88	14	16	0.06	0.11
2x32	64	1.31	42	32	0.03	0.08
2x64	123	0.66	42	64	0.03	0.07
2x96	175	0.44	42	96	0.02	0.06

Name	Coverage (%)	Time (s)
○sortAtomsById	0.14	0.42

Report Configuration

Vectorization

Details

Type of elements and instruction set

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

General properties

Front-end

Back-end

Cycles summary

Vectorization ratios

Vector efficiency ratios

Cycles and memory resources usage

Vectorization

Details

Execution units bottlenecks

Type of elements and instruction set

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

General properties

Front-end

Back-end

Cycles summary

Vectorization ratios

Vector efficiency ratios

Cycles and memory resources usage

Front-end bottlenecks

ASM code

Vectorization

Details

Workaround

Execution units bottlenecks

CALL instructions

Details

Type of elements and instruction set

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

General properties

Front-end

Back-end

Cycles summary

Vectorization ratios

Vector efficiency ratios

Cycles and memory resources usage

Front-end bottlenecks

ASM code

▶Scalability Runs Description

▼Colums Filter

Speed-Up and Efficiency

▶Detailed Speed-Up and Efficiency