OV - exec - Loop 2214

	Loop Id: 2214	Module: exec	Source: ams.c:3662-3682 [...]	Coverage: 1.31%

0x4c4cc8 ADD	X1, X1, #1

0x4c4ccc CMP	X30, X1

0x4c4cd0 B.EQ	4c4da8

0x4c4cd4 LDR	X5, [X11, X1,LSL #3]

0x4c4cd8 UBFM	X3, X5, #61, #60

0x4c4cdc ADD	X6, X10, X5,LSL #3

0x4c4ce0 LDR	D5, [X10, X3]

0x4c4ce4 FCMP	D31, D5

0x4c4ce8 B.EQ	4c4cc8

0x4c4cec LDR	X2, [X17, X1,LSL #3]

0x4c4cf0 LDR	D0, [X14, X1,LSL #3]

0x4c4cf4 CMP	X5, X2

0x4c4cf8 B.GE	4c4d30

0x4c4cfc MOVZ	X0, #0

0x4c4d00 ADD	X19, X21, X3

0x4c4d04 SUB	X7, X2, X5

0x4c4d08 WHILELO	P7.D, XZR, X7

  (2216) 0x4c4d0c LD1D	{Z4.D}, P7/Z, [X19, X0,LSL #3]

  (2216) 0x4c4d10 LD1D	{Z2.D}, P7/Z, [X6, X0,LSL #3]

  (2216) 0x4c4d14 LD1D	{Z3.D}, P7/Z, [X9, Z4.D,LSL #3]

  (2216) 0x4c4d18 ADD	X0, X0, X4

  (2216) 0x4c4d1c FMUL	Z2.D, P7/M, Z2.D, Z3.D

  (2216) 0x4c4d20 FNEG	Z2.D, P6/M, Z2.D

  (2216) 0x4c4d24 FADDA	D0, P7, D0, Z2.D

  (2216) 0x4c4d28 WHILELO	P7.D, X0, X7

  (2216) 0x4c4d2c B.NE	4c4d0c

0x4c4d30 LDR	X3, [X13, X1,LSL #3]

0x4c4d34 LDR	X6, [X16, X1,LSL #3]

0x4c4d38 CMP	X3, X6

0x4c4d3c B.GE	4c4d84

0x4c4d40 ADD	X5, X18, X3,LSL #3

0x4c4d44 MOVZ	X7, #0

0x4c4d48 SUB	X2, X6, X3

0x4c4d4c ADD	X19, X20, X3,LSL #3

0x4c4d50 WHILELO	P0.D, XZR, X2

0x4c4d54 HINT	#0

0x4c4d58 HINT	#0

0x4c4d5c HINT	#0

  (2215) 0x4c4d60 LD1D	{Z27.D}, P0/Z, [X5, X7,LSL #3]

  (2215) 0x4c4d64 LD1D	{Z1.D}, P0/Z, [X19, X7,LSL #3]

  (2215) 0x4c4d68 LD1D	{Z26.D}, P0/Z, [X8, Z27.D,LSL #3]

  (2215) 0x4c4d6c ADD	X7, X7, X4

  (2215) 0x4c4d70 FMUL	Z1.D, P0/M, Z1.D, Z26.D

  (2215) 0x4c4d74 FNEG	Z1.D, P6/M, Z1.D

  (2215) 0x4c4d78 FADDA	D0, P0, D0, Z1.D

  (2215) 0x4c4d7c WHILELO	P0.D, X7, X2

  (2215) 0x4c4d80 B.NE	4c4d60

0x4c4d84 LDR	D29, [X15, X1,LSL #3]

0x4c4d88 FMUL	D6, D30, D0

0x4c4d8c LDR	D28, [X12, X1,LSL #3]

0x4c4d90 FDIV	D7, D6, D29

0x4c4d94 FADD	D16, D28, D7

0x4c4d98 STR	D16, [X12, X1,LSL #3]

0x4c4d9c ADD	X1, X1, #1

0x4c4da0 CMP	X30, X1

0x4c4da4 B.NE	4c4cd4

/home/eoseret/qaas/qaas_runs/178-188-3659/intel/AMG/build/AMG/AMG/parcsr_ls/ams.c: 3662 - 3682

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

3662: #pragma omp parallel for private(i,ii,jj,res) HYPRE_SMP_SCHEDULE

[...]

3669:          if (A_diag_data[A_diag_i[i]] != zero)

3670:          {

3671:             res = f_data[i];

3672:             for (jj = A_diag_i[i]; jj < A_diag_i[i+1]; jj++)

3673:             {

3674:                ii = A_diag_j[jj];

3675:                res -= A_diag_data[jj] * Vtemp_data[ii];

3676:             }

3677:             for (jj = A_offd_i[i]; jj < A_offd_i[i+1]; jj++)

3678:             {

3679:                ii = A_offd_j[jj];

3680:                res -= A_offd_data[jj] * Vext_data[ii];

3681:             }

3682:             u_data[i] += (relax_weight*res)/l1_norms[i];

Coverage (%)	Name	Source Location	Module
►98.23+	omp_fulfill_event		libgomp.so.1.0.0
○	start_thread		libc.so.6
○	thread_start		libc.so.6
►1.67+	GOMP_parallel		libgomp.so.1.0.0
○	hypre_ParCSRRelaxThreads	ams.c:3662	exec
○	hypre_BoomerAMGCycle	par_cycle.c:394	exec
○	hypre_BoomerAMGSolve	par_amg_solve.c:235	exec
○	hypre_PCGSolve	pcg.c:545	exec
○	main	amg.c:419	exec
○	__libc_start_call_main		libc.so.6
○	__libc_start_main		libc.so.6
○	_start	amg.c:253	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	2.39
CQA speedup if FP arith vectorized	1.54
CQA speedup if fully vectorized	4.00 - 2.50
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.03
Bottlenecks	micro-operation queue,
Function	hypre_ParCSRRelaxThreads._omp_fn.1
Source	ams.c:3662-3662,ams.c:3669-3672,ams.c:3677-3677,ams.c:3682-3682
Source loop unroll info	NA
Source loop unroll confidence level	NA
Unroll/vectorization loop type	NA
Unroll factor	NA
CQA cycles	4.38
CQA cycles if no scalar integer	1.83
CQA cycles if FP arith vectorized	2.83
CQA cycles if fully vectorized	1.09 - 1.75
Front-end cycles	4.38
P0 cycles	2.50
P1 cycles	2.50
P2 cycles	4.25
P3 cycles	4.25
P4 cycles	4.25
P5 cycles	4.25
P6 cycles	1.25
P7 cycles	1.25
P8 cycles	1.25
P9 cycles	1.25
P10 cycles	3.17
P11 cycles	2.83
P12 cycles	3.00
P13 cycles	0.00
P14 cycles	0.00
DIV/SQRT cycles	1.75 - 3.50
Inter-iter dependencies cycles	NA
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	38.00
Nb uops	35.00
Nb loads	NA
Nb stores	1.00
Nb stack references	0.00
FLOP/cycle	0.69
Nb FLOP add-sub	1.00
Nb FLOP mul	1.00
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	0.00
Bytes prefetched	0.00
Bytes loaded	0.00
Bytes stored	0.00
Stride 0	NA
Stride 1	NA
Stride n	NA
Stride unknown	NA
Stride indirect	NA
Vectorization ratio all	0.00
Vectorization ratio load	0.00
Vectorization ratio store	0.00
Vectorization ratio mul	0.00
Vectorization ratio add_sub	0.00
Vectorization ratio fma	NA
Vectorization ratio div_sqrt	0.00
Vectorization ratio other	0.00
Vector-efficiency ratio all	25.00
Vector-efficiency ratio load	25.00
Vector-efficiency ratio store	25.00
Vector-efficiency ratio mul	25.00
Vector-efficiency ratio add_sub	25.00
Vector-efficiency ratio fma	NA
Vector-efficiency ratio div_sqrt	25.00
Vector-efficiency ratio other	25.00

Metric	Value
CQA speedup if no scalar integer	2.39
CQA speedup if FP arith vectorized	1.54
CQA speedup if fully vectorized	4.00 - 2.50
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.03
Bottlenecks	micro-operation queue,
Function	hypre_ParCSRRelaxThreads._omp_fn.1
Source	ams.c:3662-3662,ams.c:3669-3672,ams.c:3677-3677,ams.c:3682-3682
Source loop unroll info	NA
Source loop unroll confidence level	NA
Unroll/vectorization loop type	NA
Unroll factor	NA
CQA cycles	4.38
CQA cycles if no scalar integer	1.83
CQA cycles if FP arith vectorized	2.83
CQA cycles if fully vectorized	1.09 - 1.75
Front-end cycles	4.38
P0 cycles	2.50
P1 cycles	2.50
P2 cycles	4.25
P3 cycles	4.25
P4 cycles	4.25
P5 cycles	4.25
P6 cycles	1.25
P7 cycles	1.25
P8 cycles	1.25
P9 cycles	1.25
P10 cycles	3.17
P11 cycles	2.83
P12 cycles	3.00
P13 cycles	0.00
P14 cycles	0.00
DIV/SQRT cycles	1.75 - 3.50
Inter-iter dependencies cycles	NA
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	38.00
Nb uops	35.00
Nb loads	NA
Nb stores	1.00
Nb stack references	0.00
FLOP/cycle	0.69
Nb FLOP add-sub	1.00
Nb FLOP mul	1.00
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	0.00
Bytes prefetched	0.00
Bytes loaded	0.00
Bytes stored	0.00
Stride 0	NA
Stride 1	NA
Stride n	NA
Stride unknown	NA
Stride indirect	NA
Vectorization ratio all	0.00
Vectorization ratio load	0.00
Vectorization ratio store	0.00
Vectorization ratio mul	0.00
Vectorization ratio add_sub	0.00
Vectorization ratio fma	NA
Vectorization ratio div_sqrt	0.00
Vectorization ratio other	0.00
Vector-efficiency ratio all	25.00
Vector-efficiency ratio load	25.00
Vector-efficiency ratio store	25.00
Vector-efficiency ratio mul	25.00
Vector-efficiency ratio add_sub	25.00
Vector-efficiency ratio fma	NA
Vector-efficiency ratio div_sqrt	25.00
Vector-efficiency ratio other	25.00

Path /

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

Function	hypre_ParCSRRelaxThreads._omp_fn.1
Source file and lines	ams.c:3662-3682
Module	exec

The loop is defined in /home/eoseret/qaas/qaas_runs/178-188-3659/intel/AMG/build/AMG/AMG/parcsr_ls/ams.c:3662,3669-3682.

Analyzed code is defined in /home/eoseret/qaas/qaas_runs/178-188-3659/intel/AMG/build/AMG/AMG/parcsr_ls/ams.c:3662,3669-3672,3677,3682.

Warnings:

Non-innermost loop: analyzing only self part (ignoring child loops).
Ignoring paths for analysis
Too many paths. Rerun with max-paths=5
RecMII not computed since number of paths is unknown or > max_paths
Streams not analyzed since number of paths is unknown or > max_paths

Try to simplify control and/or increase the maximum number of paths per function/loop through the 'max-paths-nb' option.

This loop has 5 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

Code clean check

Detected a slowdown caused by scalar integer instructions (typically used for address computation). By removing them, you can lower the cost of an iteration from 4.38 to 1.83 cycles (2.39x speedup).

Workaround

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

Vectorization

Your loop is not vectorized. 4 data elements could be processed at once in vector registers. By vectorizing your loop, you can lower the cost of an iteration from 4.38 to 1.75 cycles (2.50x 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 loop can use their full power.

Workaround

Try another compiler or update/tune your current one:
- recompile with fassociative-math (included in Ofast or ffast-math) 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

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).

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

The binary loop is composed of 3 FP arithmetical operations:

1: addition or subtraction
1: multiply
1: divide

The binary loop does not load or store any data.

General properties

nb instructions	38
nb uops	35
loop length	152
used w registers	0
used x registers	21
used b registers	0
used h registers	0
used s registers	0
used d registers	9
used q registers	0
used v registers	0
used z registers	0
nb stack references	0
ADD-SUB / MUL ratio	1.00

Front-end

MACRO FUSION NOT POSSIBLE FIT IN UOP CACHE

micro-operation queue	4.38 cycles
front end	4.38 cycles

Back-end

	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11	P12	P13	P14
uops	2.50	2.50	4.25	4.25	4.25	4.25	1.25	1.25	1.25	1.25	3.17	2.83	3.00	0.00	0.00
cycles	2.50	2.50	4.25	4.25	4.25	4.25	1.25	1.25	1.25	1.25	3.17	2.83	3.00	0.00	0.00

Execution ports to units layout:

P0: BRU
P1: BRU
P2: ALU
P3: ALU
P4: ALU
P5: ALU
P6 (128 bits): VPU, FP store data, ALU, DIV/SQRT
P7 (128 bits): VPU, ALU, FP store data
P8 (128 bits): VPU, ALU, DIV/SQRT
P9 (128 bits): ALU, VPU
P10 (256 bits): store address, load
P11 (256 bits): store address, load
P12 (256 bits): load
P13 (64 bits): store data
P14 (64 bits): store data

Cycles executing div or sqrt instructions

1.75-3.50

Cycles summary

Front-end	4.38
Dispatch	4.25
DIV/SQRT	1.75-3.50
Overall L1	4.38

Vectorization ratios

INT

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

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

INT+FP

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

Vector efficiency ratios

INT

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

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

INT+FP

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

Cycles and memory resources usage

Assuming all data fit into the L1 cache, each iteration of the binary loop takes 4.38 cycles.

Front-end bottlenecks

Performance is limited by instruction throughput (loading/decoding program instructions to execution core) (front-end is a bottleneck).

ASM code

In the binary file, the address of the loop is: 4c4cd4

Instruction	Nb FU	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11	P12	P13	P14	Latency	Recip. throughput	Vectorization
ADD X1, X1, #1	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	0	0	1	0.25	N/A
CMP X30, X1	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	0	0	1	0.33	scal (25.0%)
B.EQ 4c4da8 <hypre_ParCSRRelaxThreads._omp_fn.1+0x15c>	1	0.50	0.50	0	0	0	0	0	0	0	0	0	0	0	0	0	1	0.50	N/A
LDR X5, [X11, X1,LSL #3]	1	0	0	0	0	0	0	0	0	0	0	0.33	0.33	0.33	0	0	4	0.33	scal (25.0%)
UBFM X3, X5, #61, #60	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	0	0	1	0.25	N/A
ADD X6, X10, X5,LSL #3	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	0	0	1	0.25	scal (25.0%)
LDR D5, [X10, X3]	1	0	0	0	0	0	0	0	0	0	0	0.33	0.33	0.33	0	0	6	0.33	scal (25.0%)
FCMP D31, D5	1	0	0	0	0	0	0	1	0	0	0	0	0	0	0	0	2	1	scal (25.0%)
B.EQ 4c4cc8 <hypre_ParCSRRelaxThreads._omp_fn.1+0x7c>	1	0.50	0.50	0	0	0	0	0	0	0	0	0	0	0	0	0	1	0.50	N/A
LDR X2, [X17, X1,LSL #3]	1	0	0	0	0	0	0	0	0	0	0	0.33	0.33	0.33	0	0	4	0.33	scal (25.0%)
LDR D0, [X14, X1,LSL #3]	1	0	0	0	0	0	0	0	0	0	0	0.33	0.33	0.33	0	0	6	0.33	scal (25.0%)
CMP X5, X2	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	0	0	1	0.33	scal (25.0%)
B.GE 4c4d30 <hypre_ParCSRRelaxThreads._omp_fn.1+0xe4>	1	0.50	0.50	0	0	0	0	0	0	0	0	0	0	0	0	0	1	0.50	N/A
MOVZ X0, #0	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	0	0	1	0.25	scal (25.0%)
ADD X19, X21, X3	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	0	0	1	0.25	scal (25.0%)
SUB X7, X2, X5	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	0	0	1	0.25	scal (25.0%)
WHILELO P7.D, XZR, X7	1	0	0	0	0	1	0	0	0	0	0	0	0	0	0	0	3	2	N/A
LDR X3, [X13, X1,LSL #3]	1	0	0	0	0	0	0	0	0	0	0	0.33	0.33	0.33	0	0	4	0.33	N/A
LDR X6, [X16, X1,LSL #3]	1	0	0	0	0	0	0	0	0	0	0	0.33	0.33	0.33	0	0	4	0.33	scal (25.0%)
CMP X3, X6	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	0	0	1	0.33	scal (25.0%)
B.GE 4c4d84 <hypre_ParCSRRelaxThreads._omp_fn.1+0x138>	1	0.50	0.50	0	0	0	0	0	0	0	0	0	0	0	0	0	1	0.50	N/A
ADD X5, X18, X3,LSL #3	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	0	0	1	0.25	scal (25.0%)
MOVZ X7, #0	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	0	0	1	0.25	scal (25.0%)
SUB X2, X6, X3	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	0	0	1	0.25	scal (25.0%)
ADD X19, X20, X3,LSL #3	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	0	0	1	0.25	scal (25.0%)
WHILELO P0.D, XZR, X2	1	0	0	0	0	1	0	0	0	0	0	0	0	0	0	0	3	2	N/A
HINT #0																			N/A
HINT #0																			N/A
HINT #0																			N/A
LDR D29, [X15, X1,LSL #3]	1	0	0	0	0	0	0	0	0	0	0	0.33	0.33	0.33	0	0	6	0.33	scal (25.0%)
FMUL D6, D30, D0	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	3	0.25	scal (25.0%)
LDR D28, [X12, X1,LSL #3]	1	0	0	0	0	0	0	0	0	0	0	0.33	0.33	0.33	0	0	6	0.33	scal (25.0%)
FDIV D7, D6, D29	1	0	0	0	0	0	0	0.50	0	0.50	0	0	0	0	0	0	7-15	1.75-3.50	scal (25.0%)
FADD D16, D28, D7	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	2	0.25	scal (25.0%)
STR D16, [X12, X1,LSL #3]	1	0	0	0	0	0	0	0.50	0.50	0	0	0.50	0.50	0	0	0	2	0.50	scal (25.0%)
ADD X1, X1, #1	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	0	0	1	0.25	N/A
CMP X30, X1	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	0	0	1	0.33	N/A
B.NE 4c4cd4 <hypre_ParCSRRelaxThreads._omp_fn.1+0x88>	1	0.50	0.50	0	0	0	0	0	0	0	0	0	0	0	0	0	1	0.50	N/A

Function	hypre_ParCSRRelaxThreads._omp_fn.1
Source file and lines	ams.c:3662-3682
Module	exec

Non-innermost loop: analyzing only self part (ignoring child loops).
Ignoring paths for analysis
Too many paths. Rerun with max-paths=5
RecMII not computed since number of paths is unknown or > max_paths
Streams not analyzed since number of paths is unknown or > max_paths

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

Code clean check

Detected a slowdown caused by scalar integer instructions (typically used for address computation). By removing them, you can lower the cost of an iteration from 4.38 to 1.83 cycles (2.39x speedup).

Workaround

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

Vectorization

Your loop is not vectorized. 4 data elements could be processed at once in vector registers. By vectorizing your loop, you can lower the cost of an iteration from 4.38 to 1.75 cycles (2.50x 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 loop can use their full power.

Workaround

Try another compiler or update/tune your current one:
- recompile with fassociative-math (included in Ofast or ffast-math) 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

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

The binary loop is composed of 3 FP arithmetical operations:

1: addition or subtraction
1: multiply
1: divide

The binary loop does not load or store any data.

General properties

nb instructions	38
nb uops	35
loop length	152
used w registers	0
used x registers	21
used b registers	0
used h registers	0
used s registers	0
used d registers	9
used q registers	0
used v registers	0
used z registers	0
nb stack references	0
ADD-SUB / MUL ratio	1.00

Front-end

MACRO FUSION NOT POSSIBLE FIT IN UOP CACHE

micro-operation queue	4.38 cycles
front end	4.38 cycles

Back-end

	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11	P12	P13	P14
uops	2.50	2.50	4.25	4.25	4.25	4.25	1.25	1.25	1.25	1.25	3.17	2.83	3.00	0.00	0.00
cycles	2.50	2.50	4.25	4.25	4.25	4.25	1.25	1.25	1.25	1.25	3.17	2.83	3.00	0.00	0.00

Execution ports to units layout:

P0: BRU
P1: BRU
P2: ALU
P3: ALU
P4: ALU
P5: ALU
P6 (128 bits): VPU, FP store data, ALU, DIV/SQRT
P7 (128 bits): VPU, ALU, FP store data
P8 (128 bits): VPU, ALU, DIV/SQRT
P9 (128 bits): ALU, VPU
P10 (256 bits): store address, load
P11 (256 bits): store address, load
P12 (256 bits): load
P13 (64 bits): store data
P14 (64 bits): store data

Cycles executing div or sqrt instructions

1.75-3.50

Cycles summary

Front-end	4.38
Dispatch	4.25
DIV/SQRT	1.75-3.50
Overall L1	4.38

Vectorization ratios

INT

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

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

INT+FP

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

Vector efficiency ratios

INT

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

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

INT+FP

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

Cycles and memory resources usage

Assuming all data fit into the L1 cache, each iteration of the binary loop takes 4.38 cycles.

Front-end bottlenecks

Performance is limited by instruction throughput (loading/decoding program instructions to execution core) (front-end is a bottleneck).

ASM code

In the binary file, the address of the loop is: 4c4cd4

Instruction	Nb FU	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11	P12	P13	P14	Latency	Recip. throughput	Vectorization
ADD X1, X1, #1	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	0	0	1	0.25	N/A
CMP X30, X1	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	0	0	1	0.33	scal (25.0%)
B.EQ 4c4da8 <hypre_ParCSRRelaxThreads._omp_fn.1+0x15c>	1	0.50	0.50	0	0	0	0	0	0	0	0	0	0	0	0	0	1	0.50	N/A
LDR X5, [X11, X1,LSL #3]	1	0	0	0	0	0	0	0	0	0	0	0.33	0.33	0.33	0	0	4	0.33	scal (25.0%)
UBFM X3, X5, #61, #60	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	0	0	1	0.25	N/A
ADD X6, X10, X5,LSL #3	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	0	0	1	0.25	scal (25.0%)
LDR D5, [X10, X3]	1	0	0	0	0	0	0	0	0	0	0	0.33	0.33	0.33	0	0	6	0.33	scal (25.0%)
FCMP D31, D5	1	0	0	0	0	0	0	1	0	0	0	0	0	0	0	0	2	1	scal (25.0%)
B.EQ 4c4cc8 <hypre_ParCSRRelaxThreads._omp_fn.1+0x7c>	1	0.50	0.50	0	0	0	0	0	0	0	0	0	0	0	0	0	1	0.50	N/A
LDR X2, [X17, X1,LSL #3]	1	0	0	0	0	0	0	0	0	0	0	0.33	0.33	0.33	0	0	4	0.33	scal (25.0%)
LDR D0, [X14, X1,LSL #3]	1	0	0	0	0	0	0	0	0	0	0	0.33	0.33	0.33	0	0	6	0.33	scal (25.0%)
CMP X5, X2	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	0	0	1	0.33	scal (25.0%)
B.GE 4c4d30 <hypre_ParCSRRelaxThreads._omp_fn.1+0xe4>	1	0.50	0.50	0	0	0	0	0	0	0	0	0	0	0	0	0	1	0.50	N/A
MOVZ X0, #0	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	0	0	1	0.25	scal (25.0%)
ADD X19, X21, X3	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	0	0	1	0.25	scal (25.0%)
SUB X7, X2, X5	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	0	0	1	0.25	scal (25.0%)
WHILELO P7.D, XZR, X7	1	0	0	0	0	1	0	0	0	0	0	0	0	0	0	0	3	2	N/A
LDR X3, [X13, X1,LSL #3]	1	0	0	0	0	0	0	0	0	0	0	0.33	0.33	0.33	0	0	4	0.33	N/A
LDR X6, [X16, X1,LSL #3]	1	0	0	0	0	0	0	0	0	0	0	0.33	0.33	0.33	0	0	4	0.33	scal (25.0%)
CMP X3, X6	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	0	0	1	0.33	scal (25.0%)
B.GE 4c4d84 <hypre_ParCSRRelaxThreads._omp_fn.1+0x138>	1	0.50	0.50	0	0	0	0	0	0	0	0	0	0	0	0	0	1	0.50	N/A
ADD X5, X18, X3,LSL #3	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	0	0	1	0.25	scal (25.0%)
MOVZ X7, #0	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	0	0	1	0.25	scal (25.0%)
SUB X2, X6, X3	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	0	0	1	0.25	scal (25.0%)
ADD X19, X20, X3,LSL #3	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	0	0	1	0.25	scal (25.0%)
WHILELO P0.D, XZR, X2	1	0	0	0	0	1	0	0	0	0	0	0	0	0	0	0	3	2	N/A
HINT #0																			N/A
HINT #0																			N/A
HINT #0																			N/A
LDR D29, [X15, X1,LSL #3]	1	0	0	0	0	0	0	0	0	0	0	0.33	0.33	0.33	0	0	6	0.33	scal (25.0%)
FMUL D6, D30, D0	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	3	0.25	scal (25.0%)
LDR D28, [X12, X1,LSL #3]	1	0	0	0	0	0	0	0	0	0	0	0.33	0.33	0.33	0	0	6	0.33	scal (25.0%)
FDIV D7, D6, D29	1	0	0	0	0	0	0	0.50	0	0.50	0	0	0	0	0	0	7-15	1.75-3.50	scal (25.0%)
FADD D16, D28, D7	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	2	0.25	scal (25.0%)
STR D16, [X12, X1,LSL #3]	1	0	0	0	0	0	0	0.50	0.50	0	0	0.50	0.50	0	0	0	2	0.50	scal (25.0%)
ADD X1, X1, #1	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	0	0	1	0.25	N/A
CMP X30, X1	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	0	0	1	0.33	N/A
B.NE 4c4cd4 <hypre_ParCSRRelaxThreads._omp_fn.1+0x88>	1	0.50	0.50	0	0	0	0	0	0	0	0	0	0	0	0	0	1	0.50	N/A

Report Configuration

Code clean check

Workaround

Vectorization

Details

Workaround

FMA

Workaround

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

General properties

Front-end

Back-end

Cycles summary

Vectorization ratios

Vector efficiency ratios

Cycles and memory resources usage

Front-end bottlenecks

ASM code

Code clean check

Workaround

Vectorization

Details

Workaround

FMA

Workaround

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

General properties

Front-end

Back-end

Cycles summary

Vectorization ratios

Vector efficiency ratios

Cycles and memory resources usage

Front-end bottlenecks

ASM code