OV - exec - Loop 157

	Loop Id: 157	Module: exec	Source: ljForce.c:191-216 [...]	Coverage: 58.80%

0x413560 ADD	$0x18,%R13

0x413564 DEC	%R11

0x413567 JE	413540

0x413569 MOVUPD	(%RAX),%XMM11    [2]

0x41356e MOVUPD	-0x10(%R13),%XMM12    [4]

0x413574 SUBPD	%XMM12,%XMM11

0x413579 MOVAPD	%XMM11,%XMM12

0x41357e MULPD	%XMM11,%XMM12

0x413583 MOVAPD	%XMM12,%XMM13

0x413588 UNPCKHPD	%XMM12,%XMM13

0x41358d ADDSD	%XMM12,%XMM13

0x413592 MOVSD	0x10(%RAX),%XMM12    [2]

0x413598 SUBSD	(%R13),%XMM12    [4]

0x41359e MOVAPD	%XMM12,%XMM14

0x4135a3 MULSD	%XMM12,%XMM14

0x4135a8 ADDSD	%XMM13,%XMM14

0x4135ad UCOMISD	%XMM2,%XMM14

0x4135b2 JA	413560

0x4135b4 UCOMISD	%XMM4,%XMM14

0x4135b9 JBE	413560

0x4135bb MOVAPD	%XMM6,%XMM13

0x4135c0 DIVSD	%XMM14,%XMM13

0x4135c5 MOVAPD	%XMM13,%XMM14

0x4135ca MULSD	%XMM13,%XMM14

0x4135cf MOVDQA	%XMM1,%XMM15

0x4135d4 MULSD	%XMM13,%XMM15

0x4135d9 MULSD	%XMM14,%XMM15

0x4135de MOVAPD	%XMM15,%XMM14

0x4135e3 ADDSD	%XMM7,%XMM14

0x4135e8 MULSD	%XMM15,%XMM14

0x4135ed SUBSD	0x10(%RBP),%XMM14    [3]

0x4135f3 MULSD	%XMM8,%XMM14

0x4135f8 MOV	0x30(%R14),%R8    [5]

0x4135fc MOVSD	(%R8,%RSI,8),%XMM0    [1]

0x413602 ADDSD	%XMM14,%XMM0

0x413607 MOVSD	%XMM0,(%R8,%RSI,8)    [1]

0x41360d MULSD	%XMM3,%XMM13

0x413612 MULSD	%XMM15,%XMM13

0x413617 MULSD	%XMM9,%XMM15

0x41361c ADDSD	%XMM10,%XMM15

0x413621 MULSD	%XMM15,%XMM13

0x413626 MOV	0x28(%R14),%R8    [5]

0x41362a MOVAPD	%XMM13,%XMM0

0x41362f UNPCKLPD	%XMM13,%XMM0

0x413634 MULPD	%XMM11,%XMM0

0x413639 MOVUPD	(%R8,%RCX,8),%XMM11    [6]

0x41363f ADDPD	%XMM0,%XMM11

0x413644 MOVUPD	%XMM11,(%R8,%RCX,8)    [6]

0x41364a MULSD	%XMM12,%XMM13

0x41364f ADDSD	0x10(%R8,%RCX,8),%XMM13    [6]

0x413656 MOVSD	%XMM13,0x10(%R8,%RCX,8)    [6]

0x41365d ADDSD	%XMM14,%XMM5

0x413662 JMP	413560

/home/eoseret/qaas_runs_ZEN5/174-043-3878/intel/CoMD/build/CoMD/CoMD/src-openmp/ljForce.c: 191 - 216

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

191:             for (int jOff=jBox*MAXATOMS; jOff<(jBox*MAXATOMS+nJBox); jOff++)

[...]

197:                   dr[m] = s->atoms->r[iOff][m]-s->atoms->r[jOff][m];

198:                   r2+=dr[m]*dr[m];

199:                }

200:

201:                if ( r2 <= rCut2 && r2 > 0.0)

202:                {

203:

204:                   // Important note:

205:                   // from this point on r actually refers to 1.0/r

206:                   r2 = 1.0/r2;

207:                   real_t r6 = s6 * (r2*r2*r2);

208:                   real_t eLocal = r6 * (r6 - 1.0) - eShift;

209:                   s->atoms->U[iOff] += 0.5*eLocal;

210:                   ePot += 0.5*eLocal;

211:

212:                   // different formulation to avoid sqrt computation

213:                   real_t fr = - 4.0*epsilon*r6*r2*(12.0*r6 - 6.0);

214:                   for (int m=0; m<3; m++)

215:                   {

216:                      s->atoms->f[iOff][m] -= dr[m]*fr;

Coverage (%)	Name	Source Location	Module
►96.04+	__kmp_invoke_microtask		libiomp5.so
○	__kmp_invoke_task_func		libiomp5.so
○	__kmp_launch_thread		libiomp5.so
○	_INTERNALea951ae6::__kmp_launc[...]		libiomp5.so
○	start_thread		libc.so.6
○	__GI___clone3		libc.so.6
►2.33+	__kmp_invoke_microtask		libiomp5.so
○	__kmp_invoke_task_func		libiomp5.so
○	__kmp_fork_call		libiomp5.so
○	__kmpc_fork_call		libiomp5.so
○	ljForce	ljForce.c:172	exec
○	timestep	timestep.c:51	exec
○	main	CoMD.c:125	exec

min	med	avg	max

Percentile Index	10	20	30	40	50	60	70	80	90	100
Value

min	med	avg	max

Percentile Index	10	20	30	40	50	60	70	80	90	100
Value

Path /

Metric	Value
CQA speedup if no scalar integer	1.00
CQA speedup if FP arith vectorized	2.49
CQA speedup if fully vectorized	6.92
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.06
Bottlenecks
Function	ljForce.extracted
Source	ljForce.c:191-191,ljForce.c:197-216
Source loop unroll info	not unrolled or unrolled with no peel/tail loop
Source loop unroll confidence level	max
Unroll/vectorization loop type	NA
Unroll factor	NA
CQA cycles	4.25
CQA cycles if no scalar integer	4.25
CQA cycles if FP arith vectorized	1.71
CQA cycles if fully vectorized	0.61
Front-end cycles	4.00
P0 cycles	0.67
P1 cycles	0.67
P2 cycles	0.67
P3 cycles	1.00
P4 cycles	1.00
P5 cycles	1.00
P6 cycles	1.75
P7 cycles	1.75
P8 cycles	1.75
P9 cycles	1.75
P10 cycles	3.42
P11 cycles	3.42
P12 cycles	4.25
P13 cycles	4.25
P14 cycles	1.33
P15 cycles	1.33
DIV/SQRT cycles	1.33
Inter-iter dependencies cycles	2
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	30.33
Nb uops	32.00
Nb loads	6.00
Nb stores	1.00
Nb stack references	0.33
FLOP/cycle	3.53
Nb FLOP add-sub	7.67
Nb FLOP mul	7.00
Nb FLOP fma	0.00
Nb FLOP div	0.33
Nb FLOP rcp	0.00
Nb FLOP sqrt	0.00
Nb FLOP rsqrt	0.00
Bytes/cycle	17.99
Bytes prefetched	0.00
Bytes loaded	66.67
Bytes stored	10.67
Stride 0	1.67
Stride 1	1.00
Stride n	0.00
Stride unknown	0.67
Stride indirect	0.00
Vectorization ratio all	44.07
Vectorization ratio load	45.83
Vectorization ratio store	33.33
Vectorization ratio mul	38.46
Vectorization ratio add_sub	22.73
Vectorization ratio fma	NA
Vectorization ratio div_sqrt	0.00
Vectorization ratio other	58.89
Vector-efficiency ratio all	18.01
Vector-efficiency ratio load	18.23
Vector-efficiency ratio store	16.67
Vector-efficiency ratio mul	17.31
Vector-efficiency ratio add_sub	15.34
Vector-efficiency ratio fma	NA
Vector-efficiency ratio div_sqrt	12.50
Vector-efficiency ratio other	19.86

Metric	Value
CQA speedup if no scalar integer	1.00
CQA speedup if FP arith vectorized	2.50
CQA speedup if fully vectorized	6.67
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.05
Bottlenecks	P12, P13,
Function	ljForce.extracted
Source	ljForce.c:191-191,ljForce.c:197-216
Source loop unroll info	not unrolled or unrolled with no peel/tail loop
Source loop unroll confidence level	max
Unroll/vectorization loop type	NA
Unroll factor	NA
CQA cycles	2.50
CQA cycles if no scalar integer	2.50
CQA cycles if FP arith vectorized	1.00
CQA cycles if fully vectorized	0.38
Front-end cycles	2.38
P0 cycles	0.67
P1 cycles	0.67
P2 cycles	0.67
P3 cycles	0.67
P4 cycles	0.67
P5 cycles	0.67
P6 cycles	1.00
P7 cycles	1.00
P8 cycles	1.00
P9 cycles	1.00
P10 cycles	1.50
P11 cycles	1.50
P12 cycles	2.50
P13 cycles	2.50
P14 cycles	0.50
P15 cycles	0.50
DIV/SQRT cycles	0.00
Inter-iter dependencies cycles	2
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	18.00
Nb uops	19.00
Nb loads	4.00
Nb stores	0.00
Nb stack references	0.00
FLOP/cycle	3.20
Nb FLOP add-sub	5.00
Nb FLOP mul	3.00
Nb FLOP fma	0.00
Nb FLOP div	0.00
Nb FLOP rcp	0.00
Nb FLOP sqrt	0.00
Nb FLOP rsqrt	0.00
Bytes/cycle	19.20
Bytes prefetched	0.00
Bytes loaded	48.00
Bytes stored	0.00
Stride 0	1.00
Stride 1	1.00
Stride n	0.00
Stride unknown	0.00
Stride indirect	0.00
Vectorization ratio all	50.00
Vectorization ratio load	50.00
Vectorization ratio store	NA
Vectorization ratio mul	50.00
Vectorization ratio add_sub	25.00
Vectorization ratio fma	NA
Vectorization ratio div_sqrt	NA
Vectorization ratio other	60.00
Vector-efficiency ratio all	18.75
Vector-efficiency ratio load	18.75
Vector-efficiency ratio store	NA
Vector-efficiency ratio mul	18.75
Vector-efficiency ratio add_sub	15.63
Vector-efficiency ratio fma	NA
Vector-efficiency ratio div_sqrt	NA
Vector-efficiency ratio other	20.00

Metric	Value
CQA speedup if no scalar integer	1.00
CQA speedup if FP arith vectorized	2.58
CQA speedup if fully vectorized	7.03
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.05
Bottlenecks	P10, P11, P12, P13,
Function	ljForce.extracted
Source	ljForce.c:191-191,ljForce.c:197-216
Source loop unroll info	not unrolled or unrolled with no peel/tail loop
Source loop unroll confidence level	max
Unroll/vectorization loop type	NA
Unroll factor	NA
CQA cycles	7.25
CQA cycles if no scalar integer	7.25
CQA cycles if FP arith vectorized	2.81
CQA cycles if fully vectorized	1.03
Front-end cycles	6.88
P0 cycles	0.67
P1 cycles	0.67
P2 cycles	0.67
P3 cycles	1.33
P4 cycles	1.33
P5 cycles	1.33
P6 cycles	3.25
P7 cycles	3.25
P8 cycles	3.25
P9 cycles	3.25
P10 cycles	7.25
P11 cycles	7.25
P12 cycles	7.25
P13 cycles	7.25
P14 cycles	2.50
P15 cycles	2.50
DIV/SQRT cycles	4.00
Inter-iter dependencies cycles	2
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	53.00
Nb uops	55.00
Nb loads	10.00
Nb stores	3.00
Nb stack references	1.00
FLOP/cycle	4.00
Nb FLOP add-sub	13.00
Nb FLOP mul	15.00
Nb FLOP fma	0.00
Nb FLOP div	1.00
Nb FLOP rcp	0.00
Nb FLOP sqrt	0.00
Nb FLOP rsqrt	0.00
Bytes/cycle	18.76
Bytes prefetched	0.00
Bytes loaded	104.00
Bytes stored	32.00
Stride 0	3.00
Stride 1	1.00
Stride n	0.00
Stride unknown	2.00
Stride indirect	0.00
Vectorization ratio all	35.56
Vectorization ratio load	37.50
Vectorization ratio store	33.33
Vectorization ratio mul	15.38
Vectorization ratio add_sub	18.18
Vectorization ratio fma	NA
Vectorization ratio div_sqrt	0.00
Vectorization ratio other	66.67
Vector-efficiency ratio all	16.94
Vector-efficiency ratio load	17.19
Vector-efficiency ratio store	16.67
Vector-efficiency ratio mul	14.42
Vector-efficiency ratio add_sub	14.77
Vector-efficiency ratio fma	NA
Vector-efficiency ratio div_sqrt	12.50
Vector-efficiency ratio other	20.83

Metric	Value
CQA speedup if no scalar integer	1.00
CQA speedup if FP arith vectorized	2.29
CQA speedup if fully vectorized	6.86
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.09
Bottlenecks	P12, P13,
Function	ljForce.extracted
Source	ljForce.c:191-191,ljForce.c:197-216
Source loop unroll info	not unrolled or unrolled with no peel/tail loop
Source loop unroll confidence level	max
Unroll/vectorization loop type	NA
Unroll factor	NA
CQA cycles	3.00
CQA cycles if no scalar integer	3.00
CQA cycles if FP arith vectorized	1.31
CQA cycles if fully vectorized	0.44
Front-end cycles	2.75
P0 cycles	0.67
P1 cycles	0.67
P2 cycles	0.67
P3 cycles	1.00
P4 cycles	1.00
P5 cycles	1.00
P6 cycles	1.00
P7 cycles	1.00
P8 cycles	1.00
P9 cycles	1.00
P10 cycles	1.50
P11 cycles	1.50
P12 cycles	3.00
P13 cycles	3.00
P14 cycles	1.00
P15 cycles	1.00
DIV/SQRT cycles	0.00
Inter-iter dependencies cycles	2
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	20.00
Nb uops	22.00
Nb loads	4.00
Nb stores	0.00
Nb stack references	0.00
FLOP/cycle	2.67
Nb FLOP add-sub	5.00
Nb FLOP mul	3.00
Nb FLOP fma	0.00
Nb FLOP div	0.00
Nb FLOP rcp	0.00
Nb FLOP sqrt	0.00
Nb FLOP rsqrt	0.00
Bytes/cycle	16.00
Bytes prefetched	0.00
Bytes loaded	48.00
Bytes stored	0.00
Stride 0	1.00
Stride 1	1.00
Stride n	0.00
Stride unknown	0.00
Stride indirect	0.00
Vectorization ratio all	46.67
Vectorization ratio load	50.00
Vectorization ratio store	NA
Vectorization ratio mul	50.00
Vectorization ratio add_sub	25.00
Vectorization ratio fma	NA
Vectorization ratio div_sqrt	NA
Vectorization ratio other	50.00
Vector-efficiency ratio all	18.33
Vector-efficiency ratio load	18.75
Vector-efficiency ratio store	NA
Vector-efficiency ratio mul	18.75
Vector-efficiency ratio add_sub	15.63
Vector-efficiency ratio fma	NA
Vector-efficiency ratio div_sqrt	NA
Vector-efficiency ratio other	18.75

Path /

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

Function	ljForce.extracted
Source file and lines	ljForce.c:191-216
Module	exec

The loop is defined in /home/eoseret/qaas_runs_ZEN5/174-043-3878/intel/CoMD/build/CoMD/CoMD/src-openmp/ljForce.c:191,197-216.

The related source loop is not unrolled or unrolled with no peel/tail loop.
This loop has 3 execution paths.

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

hoisting them (moving them outside the loop)
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 loop is poorly vectorized. Only 18% of vector register length is used (average across all SSE/AVX instructions). By fully vectorizing your loop, you can lower the cost of an iteration from 4.25 to 0.61 cycles (6.92x speedup).

Details

44% of SSE/AVX instructions are used in vector version (process two or more data elements in vector registers):

45% of SSE/AVX loads are used in vector version.
33% of SSE/AVX stores are used in vector version.
22% of SSE/AVX addition or subtraction instructions are used in vector version.
38% of SSE/AVX multiply instructions are used in vector version.
0% of SSE/AVX divide and square root instructions are used in vector version.
58% of SSE/AVX instructions that are not load, store, addition, subtraction nor multiply instructions are used in vector version.

Since your execution units are vector units, only a fully vectorized loop can use their full power.

Workaround

Try another compiler or update/tune your current one:
- recompile with ffast-math (included in Ofast) to extend loop vectorization to FP reductions.
Remove inter-iterations dependences from your loop and make it unit-stride:
- If your arrays have 2 or more dimensions, check whether elements are accessed contiguously and, otherwise, try to permute loops accordingly: C storage order is row-major: for(i) for(j) a[j][i] = b[j][i]; (slow, non stride 1) => for(i) for(j) a[i][j] = b[i][j]; (fast, stride 1)
- If your loop 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)

FMA

Presence of both ADD/SUB and MUL operations.

Workaround

Pass to your compiler a micro-architecture specialization option:
- use march=native
Try to change order in which elements are evaluated (using parentheses) in arithmetic expressions containing both ADD/SUB and MUL operations to enable your compiler to generate FMA instructions wherever possible. For instance a + b*c is a valid FMA (MUL then ADD). However (a+b)* c cannot be translated into an FMA (ADD then MUL).

Complex instructions

Detected COMPLEX INSTRUCTIONS.

Details

These instructions generate more than one micro-operation and only one of them can be decoded during a cycle and the extra micro-operations increase pressure on execution units.

UCOMISD: 0 occurrences

Slow data structures access

Detected data structures (typically arrays) that cannot be efficiently read/written

Details

Constant unknown stride: 0 occurrence(s)

Non-unit stride (uncontiguous) accesses are not efficiently using data caches

Workaround

Try to reorganize arrays of structures to structures of arrays
Consider to permute loops (see vectorization gain report)

Type of elements and instruction set

11 SSE or AVX instructions are processing arithmetic or math operations on double precision FP elements in scalar mode (one at a time). 2 SSE or AVX instructions are processing arithmetic or math operations on double precision FP elements in vector mode (two at a time).

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

The binary loop is composed of 15 FP arithmetical operations:

7.67: addition or subtraction
7: multiply
0.33: divide

The binary loop is loading 66 bytes (8 double precision FP elements). The binary loop is storing 10 bytes (1 double precision FP elements).

Arithmetic intensity

Arithmetic intensity is 0.19 FP operations per loaded or stored byte.

General properties

nb instructions	30.33
nb uops	32
loop length	146
used x86 registers	4.67
used mmx registers	0
used xmm registers	9
used ymm registers	0
used zmm registers	0
nb stack references	0.33
ADD-SUB / MUL ratio	1.62

Front-end

MACRO FUSION NOT POSSIBLE

micro-operation queue	4.00 cycles
front end	4.00 cycles

Back-end

	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11	P12	P13	P14	P15
uops	0.67	0.67	0.67	1.00	1.00	1.00	1.75	1.75	1.75	1.75	3.42	3.42	4.25	4.25	1.33	1.33
cycles	0.67	0.67	0.67	1.00	1.00	1.00	1.75	1.75	1.75	1.75	3.42	3.42	4.25	4.25	1.33	1.33

Execution ports to units layout:

P0: ALU
P1: ALU
P2: ALU
P3: ALU, BRU
P4: ALU, BRU
P5: ALU, BRU
P6 (512 bits): store address, load
P7 (512 bits): store address, load
P8 (512 bits): store address, load
P9 (512 bits): store address, load
P10 (512 bits): VPU, DIV/SQRT
P11 (512 bits): VPU, DIV/SQRT
P12 (512 bits): VPU
P13 (512 bits): VPU
P14 (512 bits): FP store data
P15 (512 bits): FP store data

Cycles executing div or sqrt instructions	1.33
Longest recurrence chain latency (RecMII)	2.00

Cycles summary

Front-end	4.00
Dispatch	4.25
DIV/SQRT	1.33
Data deps.	2.00
Overall L1	4.25

Vectorization ratios

INT

all	100%
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)
other	100%

all	43%
load	45%
store	33%
mul	38%
add-sub	22%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	0%
other	57%

INT+FP

all	44%
load	45%
store	33%
mul	38%
add-sub	22%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	0%
other	58%

Vector efficiency ratios

INT

all	25%
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)
other	25%

all	17%
load	18%
store	16%
mul	17%
add-sub	15%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	12%
other	19%

INT+FP

all	18%
load	18%
store	16%
mul	17%
add-sub	15%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	12%
other	19%

Cycles and memory resources usage

Assuming all data fit into the L1 cache, each iteration of the binary loop takes 4.25 cycles. At this rate:

6% of peak load performance is reached (16.51 out of 256.00 bytes loaded per cycle (GB/s @ 1GHz))
1% of peak store performance is reached (1.47 out of 128.00 bytes stored per cycle (GB/s @ 1GHz))

Function	ljForce.extracted
Source file and lines	ljForce.c:191-216
Module	exec

hoisting them (moving them outside the loop)
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 loop is partially vectorized. Only 18% of vector register length is used (average across all SSE/AVX instructions). By fully vectorizing your loop, you can lower the cost of an iteration from 2.50 to 0.37 cycles (6.67x speedup).

Details

50% of SSE/AVX instructions are used in vector version (process two or more data elements in vector registers):

50% of SSE/AVX loads are used in vector version.
25% of SSE/AVX addition or subtraction instructions are used in vector version.
50% of SSE/AVX multiply instructions are used in vector version.
60% of SSE/AVX instructions that are not load, store, addition, subtraction nor multiply instructions are used in vector version.

Since your execution units are vector units, only a fully vectorized loop can use their full power.

Workaround

Try another compiler or update/tune your current one:
- recompile with ffast-math (included in Ofast) to extend loop vectorization to FP reductions.
Remove inter-iterations dependences from your loop and make it unit-stride:
- If your arrays have 2 or more dimensions, check whether elements are accessed contiguously and, otherwise, try to permute loops accordingly: C storage order is row-major: for(i) for(j) a[j][i] = b[j][i]; (slow, non stride 1) => for(i) for(j) a[i][j] = b[i][j]; (fast, stride 1)
- If your loop 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

Performance is limited by execution of FP add operations (the FP add unit is a bottleneck). By removing all these bottlenecks, you can lower the cost of an iteration from 2.50 to 2.38 cycles (1.05x speedup).

Workaround

Reduce the number of FP add instructions

FMA

Presence of both ADD/SUB and MUL operations.

Workaround

Pass to your compiler a micro-architecture specialization option:
- use march=native
Try to change order in which elements are evaluated (using parentheses) in arithmetic expressions containing both ADD/SUB and MUL operations to enable your compiler to generate FMA instructions wherever possible. For instance a + b*c is a valid FMA (MUL then ADD). However (a+b)* c cannot be translated into an FMA (ADD then MUL).

Complex instructions

Detected COMPLEX INSTRUCTIONS.

Details

These instructions generate more than one micro-operation and only one of them can be decoded during a cycle and the extra micro-operations increase pressure on execution units.

UCOMISD: 1 occurrences►
- /home/eoseret/qaas_runs_ZEN5/174-043-3878/intel/CoMD/build/CoMD/CoMD/src-openmp/ljForce.c:201

Vector unaligned load/store instructions

Detected 2 suboptimal vector unaligned load/store instructions.

Details

MOVUPD: 2 occurrences►
- /home/eoseret/qaas_runs_ZEN5/174-043-3878/intel/CoMD/build/CoMD/CoMD/src-openmp/ljForce.c:197
- /home/eoseret/qaas_runs_ZEN5/174-043-3878/intel/CoMD/build/CoMD/CoMD/src-openmp/ljForce.c:197

Workaround

Pass to your compiler a micro-architecture specialization option:
- use march=native
Use vector aligned instructions:
1. align your arrays on 64 bytes boundaries: replace { void *p = malloc (size); } with { void *p; posix_memalign (&p, 64, size); }.
2. inform your compiler that your arrays are vector aligned: if array 'foo' is 64 bytes-aligned, define a pointer 'p_foo' as __builtin_assume_aligned (foo, 64) and use it instead of 'foo' in the loop.

Type of elements and instruction set

5 SSE or AVX instructions are processing arithmetic or math operations on double precision FP elements in scalar mode (one at a time). 2 SSE or AVX instructions are processing arithmetic or math operations on double precision FP elements in vector mode (two at a time).

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

The binary loop is composed of 8 FP arithmetical operations:

5: addition or subtraction
3: multiply

The binary loop is loading 48 bytes (6 double precision FP elements).

Arithmetic intensity

Arithmetic intensity is 0.17 FP operations per loaded or stored byte.

General properties

nb instructions	18
nb uops	19
loop length	84
used x86 registers	3
used mmx registers	0
used xmm registers	5
used ymm registers	0
used zmm registers	0
nb stack references	0
ADD-SUB / MUL ratio	2.00

Front-end

MACRO FUSION NOT POSSIBLE FIT IN UOP CACHE

micro-operation queue	2.38 cycles
front end	2.38 cycles

Back-end

	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11	P12	P13	P14	P15
uops	0.67	0.67	0.67	0.67	0.67	0.67	1.00	1.00	1.00	1.00	1.50	1.50	2.50	2.50	0.50	0.50
cycles	0.67	0.67	0.67	0.67	0.67	0.67	1.00	1.00	1.00	1.00	1.50	1.50	2.50	2.50	0.50	0.50

Execution ports to units layout:

P0: ALU
P1: ALU
P2: ALU
P3: ALU, BRU
P4: ALU, BRU
P5: ALU, BRU
P6 (512 bits): store address, load
P7 (512 bits): store address, load
P8 (512 bits): store address, load
P9 (512 bits): store address, load
P10 (512 bits): VPU, DIV/SQRT
P11 (512 bits): VPU, DIV/SQRT
P12 (512 bits): VPU
P13 (512 bits): VPU
P14 (512 bits): FP store data
P15 (512 bits): FP store data

Cycles executing div or sqrt instructions	NA
Longest recurrence chain latency (RecMII)	2.00

Cycles summary

Front-end	2.38
Dispatch	2.50
Data deps.	2.00
Overall L1	2.50

Vectorization ratios

all	50%
load	50%
store	NA (no store vectorizable/vectorized instructions)
mul	50%
add-sub	25%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	60%

Vector efficiency ratios

all	18%
load	18%
store	NA (no store vectorizable/vectorized instructions)
mul	18%
add-sub	15%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	20%

Cycles and memory resources usage

Assuming all data fit into the L1 cache, each iteration of the binary loop takes 2.50 cycles. At this rate:

7% of peak load performance is reached (19.20 out of 256.00 bytes loaded per cycle (GB/s @ 1GHz))

Front-end bottlenecks

Found no such bottlenecks.

ASM code

In the binary file, the address of the loop is: 413569

Instruction	Nb FU	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11	P12	P13	P14	P15	Latency	Recip. throughput	Vectorization
ADD $0x18,%R13	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	N/A
DEC %R11	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	N/A
JE 413540 <ljForce.extracted+0x220>	1	0	0	0	0.33	0.33	0.33	0	0	0	0	0	0	0	0	0	0	1	0.50-1	N/A
MOVUPD (%RAX),%XMM11	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	3	0.50	vect (25.0%)
MOVUPD -0x10(%R13),%XMM12	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	3	0.50	vect (25.0%)
SUBPD %XMM12,%XMM11	1	0	0	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	2	0.50	vect (25.0%)
MOVAPD %XMM11,%XMM12	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	vect (25.0%)
MULPD %XMM11,%XMM12	1	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	0	0	3	0.50	vect (25.0%)
MOVAPD %XMM12,%XMM13	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	vect (25.0%)
UNPCKHPD %XMM12,%XMM13	1	0	0	0	0	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	1	0.25	scal (12.5%)
ADDSD %XMM12,%XMM13	1	0	0	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	2	0.50	scal (12.5%)
MOVSD 0x10(%RAX),%XMM12	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	1	0.50	scal (12.5%)
SUBSD (%R13),%XMM12	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0.50	0.50	0	0	2	0.50	scal (12.5%)
MOVAPD %XMM12,%XMM14	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	vect (25.0%)
MULSD %XMM12,%XMM14	1	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	0	0	3	0.50	scal (12.5%)
ADDSD %XMM13,%XMM14	1	0	0	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	2	0.50	scal (12.5%)
UCOMISD %XMM2,%XMM14	2	0	0	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0.50	0.50	7	1	scal (12.5%)
JA 413560 <ljForce.extracted+0x240>	1	0	0	0	0.33	0.33	0.33	0	0	0	0	0	0	0	0	0	0	1	0.50-1	N/A

Function	ljForce.extracted
Source file and lines	ljForce.c:191-216
Module	exec

hoisting them (moving them outside the loop)
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 loop is poorly vectorized. Only 16% of vector register length is used (average across all SSE/AVX instructions). By fully vectorizing your loop, you can lower the cost of an iteration from 7.25 to 1.03 cycles (7.03x speedup).

Details

35% of SSE/AVX instructions are used in vector version (process two or more data elements in vector registers):

37% of SSE/AVX loads are used in vector version.
33% of SSE/AVX stores are used in vector version.
18% of SSE/AVX addition or subtraction instructions are used in vector version.
15% of SSE/AVX multiply instructions are used in vector version.
0% of SSE/AVX divide and square root instructions are used in vector version.
66% of SSE/AVX instructions that are not load, store, addition, subtraction nor multiply instructions are used in vector version.

Since your execution units are vector units, only a fully vectorized loop can use their full power.

Workaround

Try another compiler or update/tune your current one:
- recompile with ffast-math (included in Ofast) to extend loop vectorization to FP reductions.
Remove inter-iterations dependences from your loop and make it unit-stride:
- If your arrays have 2 or more dimensions, check whether elements are accessed contiguously and, otherwise, try to permute loops accordingly: C storage order is row-major: for(i) for(j) a[j][i] = b[j][i]; (slow, non stride 1) => for(i) for(j) a[i][j] = b[i][j]; (fast, stride 1)
- If your loop 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

Performance is limited by:

execution of FP add operations (the FP add unit is a bottleneck)
execution of FP multiply or FMA (fused multiply-add) operations (the FP multiply/FMA unit is a bottleneck)

By removing all these bottlenecks, you can lower the cost of an iteration from 7.25 to 6.88 cycles (1.05x speedup).

Workaround

Reduce the number of FP add instructions
Reduce the number of FP multiply/FMA instructions

FMA

Presence of both ADD/SUB and MUL operations.

Workaround

Pass to your compiler a micro-architecture specialization option:
- use march=native
Try to change order in which elements are evaluated (using parentheses) in arithmetic expressions containing both ADD/SUB and MUL operations to enable your compiler to generate FMA instructions wherever possible. For instance a + b*c is a valid FMA (MUL then ADD). However (a+b)* c cannot be translated into an FMA (ADD then MUL).

Complex instructions

Detected COMPLEX INSTRUCTIONS.

Details

These instructions generate more than one micro-operation and only one of them can be decoded during a cycle and the extra micro-operations increase pressure on execution units.

UCOMISD: 2 occurrences►
- /home/eoseret/qaas_runs_ZEN5/174-043-3878/intel/CoMD/build/CoMD/CoMD/src-openmp/ljForce.c:201
- /home/eoseret/qaas_runs_ZEN5/174-043-3878/intel/CoMD/build/CoMD/CoMD/src-openmp/ljForce.c:201

Slow data structures access

Detected data structures (typically arrays) that cannot be efficiently read/written

Details

Constant unknown stride: 2 occurrence(s)

Non-unit stride (uncontiguous) accesses are not efficiently using data caches

Workaround

Try to reorganize arrays of structures to structures of arrays
Consider to permute loops (see vectorization gain report)

Vector unaligned load/store instructions

Detected 4 suboptimal vector unaligned load/store instructions.

Details

MOVUPD: 4 occurrences►
- /home/eoseret/qaas_runs_ZEN5/174-043-3878/intel/CoMD/build/CoMD/CoMD/src-openmp/ljForce.c:197
- /home/eoseret/qaas_runs_ZEN5/174-043-3878/intel/CoMD/build/CoMD/CoMD/src-openmp/ljForce.c:197
- /home/eoseret/qaas_runs_ZEN5/174-043-3878/intel/CoMD/build/CoMD/CoMD/src-openmp/ljForce.c:216
- /home/eoseret/qaas_runs_ZEN5/174-043-3878/intel/CoMD/build/CoMD/CoMD/src-openmp/ljForce.c:216

Workaround

Pass to your compiler a micro-architecture specialization option:
- use march=native
Use vector aligned instructions:
1. align your arrays on 64 bytes boundaries: replace { void *p = malloc (size); } with { void *p; posix_memalign (&p, 64, size); }.
2. inform your compiler that your arrays are vector aligned: if array 'foo' is 64 bytes-aligned, define a pointer 'p_foo' as __builtin_assume_aligned (foo, 64) and use it instead of 'foo' in the loop.

Type of elements and instruction set

23 SSE or AVX instructions are processing arithmetic or math operations on double precision FP elements in scalar mode (one at a time). 4 SSE or AVX instructions are processing arithmetic or math operations on double precision FP elements in vector mode (two at a time).

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

The binary loop is composed of 29 FP arithmetical operations:

13: addition or subtraction
15: multiply
1: divide

The binary loop is loading 104 bytes (13 double precision FP elements). The binary loop is storing 32 bytes (4 double precision FP elements).

Arithmetic intensity

Arithmetic intensity is 0.21 FP operations per loaded or stored byte.

General properties

nb instructions	53
nb uops	55
loop length	263
used x86 registers	8
used mmx registers	0
used xmm registers	16
used ymm registers	0
used zmm registers	0
nb stack references	1
ADD-SUB / MUL ratio	0.85

Front-end

MACRO FUSION NOT POSSIBLE FIT IN UOP CACHE

micro-operation queue	6.88 cycles
front end	6.88 cycles

Back-end

	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11	P12	P13	P14	P15
uops	0.67	0.67	0.67	1.33	1.33	1.33	3.25	3.25	3.25	3.25	7.25	7.25	7.25	7.25	2.50	2.50
cycles	0.67	0.67	0.67	1.33	1.33	1.33	3.25	3.25	3.25	3.25	7.25	7.25	7.25	7.25	2.50	2.50

Execution ports to units layout:

P0: ALU
P1: ALU
P2: ALU
P3: ALU, BRU
P4: ALU, BRU
P5: ALU, BRU
P6 (512 bits): store address, load
P7 (512 bits): store address, load
P8 (512 bits): store address, load
P9 (512 bits): store address, load
P10 (512 bits): VPU, DIV/SQRT
P11 (512 bits): VPU, DIV/SQRT
P12 (512 bits): VPU
P13 (512 bits): VPU
P14 (512 bits): FP store data
P15 (512 bits): FP store data

Cycles executing div or sqrt instructions	4.00
Longest recurrence chain latency (RecMII)	2.00

Cycles summary

Front-end	6.88
Dispatch	7.25
DIV/SQRT	4.00
Data deps.	2.00
Overall L1	7.25

Vectorization ratios

INT

all	100%
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)
other	100%

all	34%
load	37%
store	33%
mul	15%
add-sub	18%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	0%
other	63%

INT+FP

all	35%
load	37%
store	33%
mul	15%
add-sub	18%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	0%
other	66%

Vector efficiency ratios

INT

all	25%
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)
other	25%

all	16%
load	17%
store	16%
mul	14%
add-sub	14%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	12%
other	20%

INT+FP

all	16%
load	17%
store	16%
mul	14%
add-sub	14%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	12%
other	20%

Cycles and memory resources usage

Assuming all data fit into the L1 cache, each iteration of the binary loop takes 7.25 cycles. At this rate:

5% of peak load performance is reached (14.34 out of 256.00 bytes loaded per cycle (GB/s @ 1GHz))
3% of peak store performance is reached (4.41 out of 128.00 bytes stored per cycle (GB/s @ 1GHz))

Front-end bottlenecks

Found no such bottlenecks.

ASM code

In the binary file, the address of the loop is: 413569

Instruction	Nb FU	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11	P12	P13	P14	P15	Latency	Recip. throughput	Vectorization
ADD $0x18,%R13	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	N/A
DEC %R11	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	N/A
JE 413540 <ljForce.extracted+0x220>	1	0	0	0	0.33	0.33	0.33	0	0	0	0	0	0	0	0	0	0	1	0.50-1	N/A
MOVUPD (%RAX),%XMM11	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	3	0.50	vect (25.0%)
MOVUPD -0x10(%R13),%XMM12	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	3	0.50	vect (25.0%)
SUBPD %XMM12,%XMM11	1	0	0	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	2	0.50	vect (25.0%)
MOVAPD %XMM11,%XMM12	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	vect (25.0%)
MULPD %XMM11,%XMM12	1	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	0	0	3	0.50	vect (25.0%)
MOVAPD %XMM12,%XMM13	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	vect (25.0%)
UNPCKHPD %XMM12,%XMM13	1	0	0	0	0	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	1	0.25	scal (12.5%)
ADDSD %XMM12,%XMM13	1	0	0	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	2	0.50	scal (12.5%)
MOVSD 0x10(%RAX),%XMM12	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	1	0.50	scal (12.5%)
SUBSD (%R13),%XMM12	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0.50	0.50	0	0	2	0.50	scal (12.5%)
MOVAPD %XMM12,%XMM14	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	vect (25.0%)
MULSD %XMM12,%XMM14	1	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	0	0	3	0.50	scal (12.5%)
ADDSD %XMM13,%XMM14	1	0	0	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	2	0.50	scal (12.5%)
UCOMISD %XMM2,%XMM14	2	0	0	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0.50	0.50	7	1	scal (12.5%)
JA 413560 <ljForce.extracted+0x240>	1	0	0	0	0.33	0.33	0.33	0	0	0	0	0	0	0	0	0	0	1	0.50-1	N/A
UCOMISD %XMM4,%XMM14	2	0	0	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0.50	0.50	7	1	scal (12.5%)
JBE 413560 <ljForce.extracted+0x240>	1	0	0	0	0.33	0.33	0.33	0	0	0	0	0	0	0	0	0	0	1	0.50-1	N/A
MOVAPD %XMM6,%XMM13	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	vect (25.0%)
DIVSD %XMM14,%XMM13	1	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	0	0	13	4	scal (12.5%)
MOVAPD %XMM13,%XMM14	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	vect (25.0%)
MULSD %XMM13,%XMM14	1	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	0	0	3	0.50	scal (12.5%)
MOVDQA %XMM1,%XMM15	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	vect (25.0%)
MULSD %XMM13,%XMM15	1	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	0	0	3	0.50	scal (12.5%)
MULSD %XMM14,%XMM15	1	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	0	0	3	0.50	scal (12.5%)
MOVAPD %XMM15,%XMM14	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	vect (25.0%)
ADDSD %XMM7,%XMM14	1	0	0	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	2	0.50	scal (12.5%)
MULSD %XMM15,%XMM14	1	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	0	0	3	0.50	scal (12.5%)
SUBSD 0x10(%RBP),%XMM14	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0.50	0.50	0	0	2	0.50	scal (12.5%)
MULSD %XMM8,%XMM14	1	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	0	0	3	0.50	scal (12.5%)
MOV 0x30(%R14),%R8	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	3	0.25	N/A
MOVSD (%R8,%RSI,8),%XMM0	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	1	0.50	scal (12.5%)
ADDSD %XMM14,%XMM0	1	0	0	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	2	0.50	scal (12.5%)
MOVSD %XMM0,(%R8,%RSI,8)	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0.50	0.50	1	0.50	scal (12.5%)
MULSD %XMM3,%XMM13	1	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	0	0	3	0.50	scal (12.5%)
MULSD %XMM15,%XMM13	1	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	0	0	3	0.50	scal (12.5%)
MULSD %XMM9,%XMM15	1	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	0	0	3	0.50	scal (12.5%)
ADDSD %XMM10,%XMM15	1	0	0	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	2	0.50	scal (12.5%)
MULSD %XMM15,%XMM13	1	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	0	0	3	0.50	scal (12.5%)
MOV 0x28(%R14),%R8	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	3	0.25	N/A
MOVAPD %XMM13,%XMM0	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	vect (25.0%)
UNPCKLPD %XMM13,%XMM0	1	0	0	0	0	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	1	0.25	scal (12.5%)
MULPD %XMM11,%XMM0	1	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	0	0	3	0.50	vect (25.0%)
MOVUPD (%R8,%RCX,8),%XMM11	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	3	0.50	vect (25.0%)
ADDPD %XMM0,%XMM11	1	0	0	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	2	0.50	vect (25.0%)
MOVUPD %XMM11,(%R8,%RCX,8)	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0.50	0.50	4	0.50	vect (25.0%)
MULSD %XMM12,%XMM13	1	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	0	0	3	0.50	scal (12.5%)
ADDSD 0x10(%R8,%RCX,8),%XMM13	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0.50	0.50	0	0	2	0.50	scal (12.5%)
MOVSD %XMM13,0x10(%R8,%RCX,8)	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0.50	0.50	1	0.50	scal (12.5%)
ADDSD %XMM14,%XMM5	1	0	0	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	2	0.50	scal (12.5%)
JMP 413560 <ljForce.extracted+0x240>	1	0	0	0	0.33	0.33	0.33	0	0	0	0	0	0	0	0	0	0	0	1	N/A

Function	ljForce.extracted
Source file and lines	ljForce.c:191-216
Module	exec

hoisting them (moving them outside the loop)
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

Details

46% of SSE/AVX instructions are used in vector version (process two or more data elements in vector registers):

50% of SSE/AVX loads are used in vector version.
25% of SSE/AVX addition or subtraction instructions are used in vector version.
50% of SSE/AVX multiply instructions are used in vector version.
50% of SSE/AVX instructions that are not load, store, addition, subtraction nor multiply instructions are used in vector version.

Since your execution units are vector units, only a fully vectorized loop can use their full power.

Workaround

Try another compiler or update/tune your current one:
- recompile with ffast-math (included in Ofast) to extend loop vectorization to FP reductions.
Remove inter-iterations dependences from your loop and make it unit-stride:
- If your arrays have 2 or more dimensions, check whether elements are accessed contiguously and, otherwise, try to permute loops accordingly: C storage order is row-major: for(i) for(j) a[j][i] = b[j][i]; (slow, non stride 1) => for(i) for(j) a[i][j] = b[i][j]; (fast, stride 1)
- If your loop 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

Performance is limited by execution of FP add operations (the FP add unit is a bottleneck). By removing all these bottlenecks, you can lower the cost of an iteration from 3.00 to 2.75 cycles (1.09x speedup).

Workaround

Reduce the number of FP add instructions

FMA

Presence of both ADD/SUB and MUL operations.

Workaround

Pass to your compiler a micro-architecture specialization option:
- use march=native
Try to change order in which elements are evaluated (using parentheses) in arithmetic expressions containing both ADD/SUB and MUL operations to enable your compiler to generate FMA instructions wherever possible. For instance a + b*c is a valid FMA (MUL then ADD). However (a+b)* c cannot be translated into an FMA (ADD then MUL).

Complex instructions

Detected COMPLEX INSTRUCTIONS.

Details

These instructions generate more than one micro-operation and only one of them can be decoded during a cycle and the extra micro-operations increase pressure on execution units.

UCOMISD: 2 occurrences►
- /home/eoseret/qaas_runs_ZEN5/174-043-3878/intel/CoMD/build/CoMD/CoMD/src-openmp/ljForce.c:201
- /home/eoseret/qaas_runs_ZEN5/174-043-3878/intel/CoMD/build/CoMD/CoMD/src-openmp/ljForce.c:201

Vector unaligned load/store instructions

Detected 2 suboptimal vector unaligned load/store instructions.

Details

MOVUPD: 2 occurrences►
- /home/eoseret/qaas_runs_ZEN5/174-043-3878/intel/CoMD/build/CoMD/CoMD/src-openmp/ljForce.c:197
- /home/eoseret/qaas_runs_ZEN5/174-043-3878/intel/CoMD/build/CoMD/CoMD/src-openmp/ljForce.c:197

Workaround

Pass to your compiler a micro-architecture specialization option:
- use march=native
Use vector aligned instructions:
1. align your arrays on 64 bytes boundaries: replace { void *p = malloc (size); } with { void *p; posix_memalign (&p, 64, size); }.
2. inform your compiler that your arrays are vector aligned: if array 'foo' is 64 bytes-aligned, define a pointer 'p_foo' as __builtin_assume_aligned (foo, 64) and use it instead of 'foo' in the loop.

Type of elements and instruction set

6 SSE or AVX instructions are processing arithmetic or math operations on double precision FP elements in scalar mode (one at a time). 2 SSE or AVX instructions are processing arithmetic or math operations on double precision FP elements in vector mode (two at a time).

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

The binary loop is composed of 8 FP arithmetical operations:

5: addition or subtraction
3: multiply

The binary loop is loading 48 bytes (6 double precision FP elements).

Arithmetic intensity

Arithmetic intensity is 0.17 FP operations per loaded or stored byte.

General properties

nb instructions	20
nb uops	22
loop length	91
used x86 registers	3
used mmx registers	0
used xmm registers	6
used ymm registers	0
used zmm registers	0
nb stack references	0
ADD-SUB / MUL ratio	2.00

Front-end

MACRO FUSION NOT POSSIBLE FIT IN UOP CACHE

micro-operation queue	2.75 cycles
front end	2.75 cycles

Back-end

	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11	P12	P13	P14	P15
uops	0.67	0.67	0.67	1.00	1.00	1.00	1.00	1.00	1.00	1.00	1.50	1.50	3.00	3.00	1.00	1.00
cycles	0.67	0.67	0.67	1.00	1.00	1.00	1.00	1.00	1.00	1.00	1.50	1.50	3.00	3.00	1.00	1.00

Execution ports to units layout:

P0: ALU
P1: ALU
P2: ALU
P3: ALU, BRU
P4: ALU, BRU
P5: ALU, BRU
P6 (512 bits): store address, load
P7 (512 bits): store address, load
P8 (512 bits): store address, load
P9 (512 bits): store address, load
P10 (512 bits): VPU, DIV/SQRT
P11 (512 bits): VPU, DIV/SQRT
P12 (512 bits): VPU
P13 (512 bits): VPU
P14 (512 bits): FP store data
P15 (512 bits): FP store data

Cycles executing div or sqrt instructions	NA
Longest recurrence chain latency (RecMII)	2.00

Cycles summary

Front-end	2.75
Dispatch	3.00
Data deps.	2.00
Overall L1	3.00

Vectorization ratios

all	46%
load	50%
store	NA (no store vectorizable/vectorized instructions)
mul	50%
add-sub	25%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	50%

Vector efficiency ratios

all	18%
load	18%
store	NA (no store vectorizable/vectorized instructions)
mul	18%
add-sub	15%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	18%

Cycles and memory resources usage

Assuming all data fit into the L1 cache, each iteration of the binary loop takes 3.00 cycles. At this rate:

6% of peak load performance is reached (16.00 out of 256.00 bytes loaded per cycle (GB/s @ 1GHz))

Front-end bottlenecks

Found no such bottlenecks.

ASM code

In the binary file, the address of the loop is: 413569

Instruction	Nb FU	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11	P12	P13	P14	P15	Latency	Recip. throughput	Vectorization
ADD $0x18,%R13	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	N/A
DEC %R11	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	N/A
JE 413540 <ljForce.extracted+0x220>	1	0	0	0	0.33	0.33	0.33	0	0	0	0	0	0	0	0	0	0	1	0.50-1	N/A
MOVUPD (%RAX),%XMM11	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	3	0.50	vect (25.0%)
MOVUPD -0x10(%R13),%XMM12	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	3	0.50	vect (25.0%)
SUBPD %XMM12,%XMM11	1	0	0	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	2	0.50	vect (25.0%)
MOVAPD %XMM11,%XMM12	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	vect (25.0%)
MULPD %XMM11,%XMM12	1	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	0	0	3	0.50	vect (25.0%)
MOVAPD %XMM12,%XMM13	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	vect (25.0%)
UNPCKHPD %XMM12,%XMM13	1	0	0	0	0	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	1	0.25	scal (12.5%)
ADDSD %XMM12,%XMM13	1	0	0	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	2	0.50	scal (12.5%)
MOVSD 0x10(%RAX),%XMM12	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	1	0.50	scal (12.5%)
SUBSD (%R13),%XMM12	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0.50	0.50	0	0	2	0.50	scal (12.5%)
MOVAPD %XMM12,%XMM14	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	vect (25.0%)
MULSD %XMM12,%XMM14	1	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	0	0	3	0.50	scal (12.5%)
ADDSD %XMM13,%XMM14	1	0	0	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	2	0.50	scal (12.5%)
UCOMISD %XMM2,%XMM14	2	0	0	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0.50	0.50	7	1	scal (12.5%)
JA 413560 <ljForce.extracted+0x240>	1	0	0	0	0.33	0.33	0.33	0	0	0	0	0	0	0	0	0	0	1	0.50-1	N/A
UCOMISD %XMM4,%XMM14	2	0	0	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0.50	0.50	7	1	scal (12.5%)
JBE 413560 <ljForce.extracted+0x240>	1	0	0	0	0.33	0.33	0.33	0	0	0	0	0	0	0	0	0	0	1	0.50-1	N/A

Report Configuration

Vectorization

Details

Workaround

FMA

Workaround

Complex instructions

Details

Slow data structures access

Details

Workaround

Type of elements and instruction set

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

Arithmetic intensity

General properties

Front-end

Back-end

Cycles summary

Vectorization ratios

Vector efficiency ratios

Cycles and memory resources usage

Vectorization

Details

Workaround

Execution units bottlenecks

Workaround

FMA

Workaround

Complex instructions

Details

Vector unaligned load/store instructions

Details

Workaround

Type of elements and instruction set

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

Arithmetic intensity

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

Workaround

FMA

Workaround

Complex instructions

Details

Slow data structures access

Details

Workaround

Vector unaligned load/store instructions

Details

Workaround

Type of elements and instruction set

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

Arithmetic intensity

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

Workaround

FMA

Workaround

Complex instructions