OV - exec - Loop 1327

0x454630 INC	%RDX

0x454633 CMP	%RSI,%RDX

0x454636 JGE	454698

0x454638 MOV	-0x160(%RBP),%RDI    [5]

0x45463f MOV	(%RDI,%RDX,8),%RDI    [9]

0x454643 MOV	-0x100(%RBP),%R8    [5]

0x45464a CMPQ	$0,(%R8,%RDI,8)    [1]

0x45464f JS	454630

0x454651 CMP	%R9,(%RBX,%RDI,8)    [10]

0x454655 JGE	454630

0x454657 MOV	-0x38(%RBP),%R8    [5]

0x45465b MOV	%R8,(%RBX,%RDI,8)    [10]

0x45465f MOV	(%R12,%RDI,8),%RSI    [2]

0x454663 MOV	-0x150(%RBP),%RDI    [5]

0x45466a MOV	(%RDI),%RDI    [8]

0x45466d MOV	%RSI,(%RDI,%R8,8)    [7]

0x454671 MOV	-0x90(%RBP),%RSI    [5]

0x454678 MOV	(%RSI),%RSI    [6]

0x45467b MOVQ	$0,(%RSI,%R8,8)    [4]

0x454683 INC	%R8

0x454686 MOV	%R8,-0x38(%RBP)    [5]

0x45468a MOV	-0xe8(%RBP),%RSI    [5]

0x454691 MOV	0x8(%RSI,%RCX,8),%RSI    [3]

0x454696 JMP	454630

/home/eoseret/qaas_runs_GNR/173-927-0874/intel/AMG/build/AMG/AMG/parcsr_ls/par_lr_interp.c: 1516 - 1526

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

1516:                for (kk = S_diag_i[i1]; kk < S_diag_i[i1+1]; kk++)

1517:                {

1518:                  k1 = S_diag_j[kk];

1519:                  if (CF_marker[k1] >= 0)

1520:                  {

1521:                    if(P_marker[k1] < jj_begin_row)

1522:                    {

1523:                      P_marker[k1] = jj_counter;

1524:                      P_diag_j[jj_counter] = fine_to_coarse[k1];

1525:                      P_diag_data[jj_counter] = zero;

1526:                      jj_counter++;

Coverage (%)	Name	Source Location	Module
►100.00+	__kmp_invoke_microtask		libiomp5.so
○	__kmp_invoke_task_func		libiomp5.so

Path /

Metric	Value
CQA speedup if no scalar integer	1.00
CQA speedup if FP arith vectorized	1.00
CQA speedup if fully vectorized	10.47
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.10
Bottlenecks
Function	hypre_BoomerAMGBuildExtPIInterp.extracted
Source	par_lr_interp.c:1516-1526
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	2.44
CQA cycles if no scalar integer	2.44
CQA cycles if FP arith vectorized	2.44
CQA cycles if fully vectorized	0.23
Front-end cycles	2.39
P0 cycles	1.33
P1 cycles	0.87
P2 cycles	2.44
P3 cycles	2.44
P4 cycles	0.67
P5 cycles	0.73
P6 cycles	1.33
P7 cycles	0.67
P8 cycles	0.67
P9 cycles	0.67
P10 cycles	0.73
P11 cycles	2.44
DIV/SQRT cycles	0.00
Inter-iter dependencies cycles	1
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	14.00
Nb uops	13.67
Nb loads	7.33
Nb stores	1.33
Nb stack references	3.33
FLOP/cycle	0.00
Nb FLOP add-sub	0.00
Nb FLOP mul	0.00
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	26.46
Bytes prefetched	0.00
Bytes loaded	58.67
Bytes stored	10.67
Stride 0	1.00
Stride 1	1.00
Stride n	0.00
Stride unknown	1.00
Stride indirect	2.67
Vectorization ratio all	0.00
Vectorization ratio load	0.00
Vectorization ratio store	0.00
Vectorization ratio mul	NA
Vectorization ratio add_sub	NA
Vectorization ratio fma	NA
Vectorization ratio div_sqrt	NA
Vectorization ratio other	0.00
Vector-efficiency ratio all	12.20
Vector-efficiency ratio load	12.50
Vector-efficiency ratio store	10.94
Vector-efficiency ratio mul	NA
Vector-efficiency ratio add_sub	NA
Vector-efficiency ratio fma	NA
Vector-efficiency ratio div_sqrt	NA
Vector-efficiency ratio other	12.50

Metric	Value
CQA speedup if no scalar integer	1.00
CQA speedup if FP arith vectorized	1.00
CQA speedup if fully vectorized	8.00
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.33
Bottlenecks	micro-operation queue, P2, P3, P11,
Function	hypre_BoomerAMGBuildExtPIInterp.extracted
Source	par_lr_interp.c:1516-1526
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	1.33
CQA cycles if no scalar integer	1.33
CQA cycles if FP arith vectorized	1.33
CQA cycles if fully vectorized	0.17
Front-end cycles	1.33
P0 cycles	1.00
P1 cycles	0.80
P2 cycles	1.33
P3 cycles	1.33
P4 cycles	0.00
P5 cycles	0.60
P6 cycles	1.00
P7 cycles	0.00
P8 cycles	0.00
P9 cycles	0.00
P10 cycles	0.60
P11 cycles	1.33
DIV/SQRT cycles	0.00
Inter-iter dependencies cycles	1
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	8.00
Nb uops	8.00
Nb loads	4.00
Nb stores	0.00
Nb stack references	2.00
FLOP/cycle	0.00
Nb FLOP add-sub	0.00
Nb FLOP mul	0.00
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	24.00
Bytes prefetched	0.00
Bytes loaded	32.00
Bytes stored	0.00
Stride 0	1.00
Stride 1	1.00
Stride n	0.00
Stride unknown	0.00
Stride indirect	1.00
Vectorization ratio all	0.00
Vectorization ratio load	NA
Vectorization ratio store	NA
Vectorization ratio mul	NA
Vectorization ratio add_sub	NA
Vectorization ratio fma	NA
Vectorization ratio div_sqrt	NA
Vectorization ratio other	0.00
Vector-efficiency ratio all	12.50
Vector-efficiency ratio load	NA
Vector-efficiency ratio store	NA
Vector-efficiency ratio mul	NA
Vector-efficiency ratio add_sub	NA
Vector-efficiency ratio fma	NA
Vector-efficiency ratio div_sqrt	NA
Vector-efficiency ratio other	12.50

Metric	Value
CQA speedup if no scalar integer	1.00
CQA speedup if FP arith vectorized	1.00
CQA speedup if fully vectorized	13.31
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.04
Bottlenecks	P2, P3, P11,
Function	hypre_BoomerAMGBuildExtPIInterp.extracted
Source	par_lr_interp.c:1516-1526
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	4.33
CQA cycles if no scalar integer	4.33
CQA cycles if FP arith vectorized	4.33
CQA cycles if fully vectorized	0.33
Front-end cycles	4.17
P0 cycles	1.50
P1 cycles	1.00
P2 cycles	4.33
P3 cycles	4.33
P4 cycles	2.00
P5 cycles	1.00
P6 cycles	1.50
P7 cycles	2.00
P8 cycles	2.00
P9 cycles	2.00
P10 cycles	1.00
P11 cycles	4.33
DIV/SQRT cycles	0.00
Inter-iter dependencies cycles	1
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	24.00
Nb uops	24.00
Nb loads	13.00
Nb stores	4.00
Nb stack references	6.00
FLOP/cycle	0.00
Nb FLOP add-sub	0.00
Nb FLOP mul	0.00
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	31.38
Bytes prefetched	0.00
Bytes loaded	104.00
Bytes stored	32.00
Stride 0	1.00
Stride 1	1.00
Stride n	0.00
Stride unknown	3.00
Stride indirect	5.00
Vectorization ratio all	0.00
Vectorization ratio load	0.00
Vectorization ratio store	0.00
Vectorization ratio mul	NA
Vectorization ratio add_sub	NA
Vectorization ratio fma	NA
Vectorization ratio div_sqrt	NA
Vectorization ratio other	0.00
Vector-efficiency ratio all	11.61
Vector-efficiency ratio load	12.50
Vector-efficiency ratio store	10.94
Vector-efficiency ratio mul	NA
Vector-efficiency ratio add_sub	NA
Vector-efficiency ratio fma	NA
Vector-efficiency ratio div_sqrt	NA
Vector-efficiency ratio other	12.50

Metric	Value
CQA speedup if no scalar integer	1.00
CQA speedup if FP arith vectorized	1.00
CQA speedup if fully vectorized	8.00
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.11
Bottlenecks	micro-operation queue,
Function	hypre_BoomerAMGBuildExtPIInterp.extracted
Source	par_lr_interp.c:1516-1526
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	1.67
CQA cycles if no scalar integer	1.67
CQA cycles if FP arith vectorized	1.67
CQA cycles if fully vectorized	0.21
Front-end cycles	1.67
P0 cycles	1.50
P1 cycles	0.80
P2 cycles	1.67
P3 cycles	1.67
P4 cycles	0.00
P5 cycles	0.60
P6 cycles	1.50
P7 cycles	0.00
P8 cycles	0.00
P9 cycles	0.00
P10 cycles	0.60
P11 cycles	1.67
DIV/SQRT cycles	0.00
Inter-iter dependencies cycles	1
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	10.00
Nb uops	9.00
Nb loads	5.00
Nb stores	0.00
Nb stack references	2.00
FLOP/cycle	0.00
Nb FLOP add-sub	0.00
Nb FLOP mul	0.00
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	24.00
Bytes prefetched	0.00
Bytes loaded	40.00
Bytes stored	0.00
Stride 0	1.00
Stride 1	1.00
Stride n	0.00
Stride unknown	0.00
Stride indirect	2.00
Vectorization ratio all	0.00
Vectorization ratio load	0.00
Vectorization ratio store	NA
Vectorization ratio mul	NA
Vectorization ratio add_sub	NA
Vectorization ratio fma	NA
Vectorization ratio div_sqrt	NA
Vectorization ratio other	0.00
Vector-efficiency ratio all	12.50
Vector-efficiency ratio load	12.50
Vector-efficiency ratio store	NA
Vector-efficiency ratio mul	NA
Vector-efficiency ratio add_sub	NA
Vector-efficiency ratio fma	NA
Vector-efficiency ratio div_sqrt	NA
Vector-efficiency ratio other	12.50

Path /

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

Function	hypre_BoomerAMGBuildExtPIInterp.extracted
Source file and lines	par_lr_interp.c:1516-1526
Module	exec

The loop is defined in /home/eoseret/qaas_runs_GNR/173-927-0874/intel/AMG/build/AMG/AMG/parcsr_ls/par_lr_interp.c:1516-1526.

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 not vectorized. Only 12% of vector register length is used (average across all SSE/AVX instructions). By vectorizing your loop, you can lower the cost of an iteration from 2.44 to 0.23 cycles (10.47x 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 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)

No data for this section

Slow data structures access

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

Details

Constant unknown stride: 1 occurrence(s)
Irregular (variable stride) or indirect: 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)
Try to remove indirect accesses. If applicable, precompute elements out of the innermost loop.

Type of elements and instruction set

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

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

The binary loop does not contain any FP arithmetical operations. The binary loop is loading 58 bytes. The binary loop is storing 10 bytes.

Unroll opportunity

Loop is data access bound.

Workaround

Unroll your loop if trip count is significantly higher than target unroll factor and if some data references are common to consecutive iterations. This can be done manually. Or with the unroll (resp. unroll_and_jam) directive on top of the inner (resp. surrounding) loop. You can enforce an unroll factor: #pragma unroll_and_jam N, unroll_and_jam(N), unroll N or unroll(N)

General properties

nb instructions	14
nb uops	13.67
loop length	58.67
used x86 registers	7
used mmx registers	0
used xmm registers	0
used ymm registers	0
used zmm registers	0
nb stack references	3.33

Front-end

MACRO FUSION NOT POSSIBLE

micro-operation queue	2.39 cycles
front end	2.39 cycles

Back-end

	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11
uops	1.33	0.87	2.44	2.44	0.67	0.73	1.33	0.67	0.67	0.67	0.73	2.44
cycles	1.33	0.87	2.44	2.44	0.67	0.73	1.33	0.67	0.67	0.67	0.73	2.44

Execution ports to units layout:

P0 (256 bits): VPU, BRU, ALU, DIV/SQRT
P1 (256 bits): ALU, VPU
P2 (512 bits): load
P3 (512 bits): load
P4 (512 bits): store data
P5 (512 bits): ALU, VPU
P6: ALU, BRU
P7: store address
P8: store address
P9 (512 bits): store data
P10: ALU
P11 (512 bits): load

Cycles executing div or sqrt instructions	NA
Longest recurrence chain latency (RecMII)	1.00

Cycles summary

Front-end	2.39
Dispatch	2.44
Data deps.	1.00
Overall L1	2.44

Vectorization ratios

all	0%
load	0%
store	0%
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	12%
load	12%
store	10%
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	12%

Cycles and memory resources usage

Assuming all data fit into the L1 cache, each iteration of the binary loop takes 2.44 cycles. At this rate:

12% of peak load performance is reached (24.00 out of 192.00 bytes loaded per cycle (GB/s @ 1GHz))
1% of peak store performance is reached (2.46 out of 128.00 bytes stored per cycle (GB/s @ 1GHz))

Function	hypre_BoomerAMGBuildExtPIInterp.extracted
Source file and lines	par_lr_interp.c:1516-1526
Module	exec

The loop is defined in /home/eoseret/qaas_runs_GNR/173-927-0874/intel/AMG/build/AMG/AMG/parcsr_ls/par_lr_interp.c:1516-1526.

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 not vectorized. 8 data elements could be processed at once in vector registers. By vectorizing your loop, you can lower the cost of an iteration from 1.33 to 0.17 cycles (8.00x 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 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 reading data from caches/RAM (load units are a bottleneck). By removing all these bottlenecks, you can lower the cost of an iteration from 1.33 to 1.00 cycles (1.33x speedup).

Workaround

Read less array elements
Provide more information to your compiler:
- hardcode the bounds of the corresponding 'for' loop
- use the 'restrict' C99 keyword

No data for this section

Slow data structures access

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

Details

Irregular (variable stride) or indirect: 1 occurrence(s)

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

Workaround

Try to remove indirect accesses. If applicable, precompute elements out of the innermost loop.

Type of elements and instruction set

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

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

The binary loop does not contain any FP arithmetical operations. The binary loop is loading 32 bytes.

Unroll opportunity

Loop is data access bound.

Workaround

Unroll your loop if trip count is significantly higher than target unroll factor and if some data references are common to consecutive iterations. This can be done manually. Or with the unroll (resp. unroll_and_jam) directive on top of the inner (resp. surrounding) loop. You can enforce an unroll factor: #pragma unroll_and_jam N, unroll_and_jam(N), unroll N or unroll(N)

General properties

nb instructions	8
nb uops	8
loop length	33
used x86 registers	5
used mmx registers	0
used xmm registers	0
used ymm registers	0
used zmm registers	0
nb stack references	2

Front-end

MACRO FUSION NOT POSSIBLE FIT IN UOP CACHE

micro-operation queue	1.33 cycles
front end	1.33 cycles

Back-end

	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11
uops	1.00	0.80	1.33	1.33	0.00	0.60	1.00	0.00	0.00	0.00	0.60	1.33
cycles	1.00	0.80	1.33	1.33	0.00	0.60	1.00	0.00	0.00	0.00	0.60	1.33

Execution ports to units layout:

P0 (256 bits): VPU, BRU, ALU, DIV/SQRT
P1 (256 bits): ALU, VPU
P2 (512 bits): load
P3 (512 bits): load
P4 (512 bits): store data
P5 (512 bits): ALU, VPU
P6: ALU, BRU
P7: store address
P8: store address
P9 (512 bits): store data
P10: ALU
P11 (512 bits): load

Cycles executing div or sqrt instructions	NA
Longest recurrence chain latency (RecMII)	1.00

Cycles summary

Front-end	1.33
Dispatch	1.33
Data deps.	1.00
Overall L1	1.33

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	12%
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	12%

Cycles and memory resources usage

Assuming all data fit into the L1 cache, each iteration of the binary loop takes 1.33 cycles. At this rate:

12% of peak load performance is reached (24.00 out of 192.00 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 1.33 to 1.00 cycles (1.33x speedup).

ASM code

In the binary file, the address of the loop is: 454633

Instruction	Nb FU	P0	P1	P2	P3	P5	P6	P10	P11	Latency	Recip. throughput	Vectorization
INC %RDX	1	0	0	0	0	0	0	0	0	1	0.17	N/A
CMP %RSI,%RDX	1	0.20	0.20	0	0	0.20	0.20	0.20	0	1	0.20	scal (12.5%)
JGE 454698 <hypre_BoomerAMGBuildExtPIInterp.extracted+0x12a8>	1	0.50	0	0	0	0	0.50	0	0	0	0.50	N/A
MOV -0x160(%RBP),%RDI	1	0	0	0.33	0.33	0	0	0	0.33	1	0.33	N/A
MOV (%RDI,%RDX,8),%RDI	1	0	0	0.33	0.33	0	0	0	0.33	1	0.33	N/A
MOV -0x100(%RBP),%R8	1	0	0	0.33	0.33	0	0	0	0.33	1	0.33	N/A
CMPQ $0,(%R8,%RDI,8)	1	0.20	0.20	0.33	0.33	0.20	0.20	0.20	0.33	1	0.33	N/A
JS 454630 <hypre_BoomerAMGBuildExtPIInterp.extracted+0x1240>	1	0.50	0	0	0	0	0.50	0	0	0	0.50	N/A

Function	hypre_BoomerAMGBuildExtPIInterp.extracted
Source file and lines	par_lr_interp.c:1516-1526
Module	exec

The loop is defined in /home/eoseret/qaas_runs_GNR/173-927-0874/intel/AMG/build/AMG/AMG/parcsr_ls/par_lr_interp.c:1516-1526.

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 not vectorized. Only 11% of vector register length is used (average across all SSE/AVX instructions). By vectorizing your loop, you can lower the cost of an iteration from 4.33 to 0.33 cycles (13.31x 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 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 reading data from caches/RAM (load units are a bottleneck).

Workaround

Read less array elements
Provide more information to your compiler:
- hardcode the bounds of the corresponding 'for' loop
- use the 'restrict' C99 keyword

No data for this section

Slow data structures access

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

Details

Constant unknown stride: 3 occurrence(s)
Irregular (variable stride) or indirect: 5 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)
Try to remove indirect accesses. If applicable, precompute elements out of the innermost loop.

Type of elements and instruction set

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

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

The binary loop does not contain any FP arithmetical operations. The binary loop is loading 104 bytes. The binary loop is storing 32 bytes.

Unroll opportunity

Loop is data access bound.

Workaround

Unroll your loop if trip count is significantly higher than target unroll factor and if some data references are common to consecutive iterations. This can be done manually. Or with the unroll (resp. unroll_and_jam) directive on top of the inner (resp. surrounding) loop. You can enforce an unroll factor: #pragma unroll_and_jam N, unroll_and_jam(N), unroll N or unroll(N)

General properties

nb instructions	24
nb uops	24
loop length	104
used x86 registers	9
used mmx registers	0
used xmm registers	0
used ymm registers	0
used zmm registers	0
nb stack references	6

Front-end

MACRO FUSION NOT POSSIBLE FIT IN UOP CACHE

micro-operation queue	4.17 cycles
front end	4.17 cycles

Back-end

	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11
uops	1.50	1.00	4.33	4.33	2.00	1.00	1.50	2.00	2.00	2.00	1.00	4.33
cycles	1.50	1.00	4.33	4.33	2.00	1.00	1.50	2.00	2.00	2.00	1.00	4.33

Execution ports to units layout:

P0 (256 bits): VPU, BRU, ALU, DIV/SQRT
P1 (256 bits): ALU, VPU
P2 (512 bits): load
P3 (512 bits): load
P4 (512 bits): store data
P5 (512 bits): ALU, VPU
P6: ALU, BRU
P7: store address
P8: store address
P9 (512 bits): store data
P10: ALU
P11 (512 bits): load

Cycles executing div or sqrt instructions	NA
Longest recurrence chain latency (RecMII)	1.00

Cycles summary

Front-end	4.17
Dispatch	4.33
Data deps.	1.00
Overall L1	4.33

Vectorization ratios

all	0%
load	0%
store	0%
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	11%
load	12%
store	10%
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	12%

Cycles and memory resources usage

Assuming all data fit into the L1 cache, each iteration of the binary loop takes 4.33 cycles. At this rate:

12% of peak load performance is reached (24.00 out of 192.00 bytes loaded per cycle (GB/s @ 1GHz))
5% of peak store performance is reached (7.38 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: 454633

Instruction	Nb FU	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11	Latency	Recip. throughput	Vectorization
INC %RDX	1	0	0	0	0	0	0	0	0	0	0	0	0	1	0.17	N/A
CMP %RSI,%RDX	1	0.20	0.20	0	0	0	0.20	0.20	0	0	0	0.20	0	1	0.20	scal (12.5%)
JGE 454698 <hypre_BoomerAMGBuildExtPIInterp.extracted+0x12a8>	1	0.50	0	0	0	0	0	0.50	0	0	0	0	0	0	0.50	N/A
MOV -0x160(%RBP),%RDI	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	1	0.33	N/A
MOV (%RDI,%RDX,8),%RDI	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	1	0.33	N/A
MOV -0x100(%RBP),%R8	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	1	0.33	N/A
CMPQ $0,(%R8,%RDI,8)	1	0.20	0.20	0.33	0.33	0	0.20	0.20	0	0	0	0.20	0.33	1	0.33	scal (12.5%)
JS 454630 <hypre_BoomerAMGBuildExtPIInterp.extracted+0x1240>	1	0.50	0	0	0	0	0	0.50	0	0	0	0	0	0	0.50	N/A
CMP %R9,(%RBX,%RDI,8)	1	0.20	0.20	0.33	0.33	0	0.20	0.20	0	0	0	0.20	0.33	1	0.33	scal (12.5%)
JGE 454630 <hypre_BoomerAMGBuildExtPIInterp.extracted+0x1240>	1	0.50	0	0	0	0	0	0.50	0	0	0	0	0	0	0.50	N/A
MOV -0x38(%RBP),%R8	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	1	0.33	N/A
MOV %R8,(%RBX,%RDI,8)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	1	0.50	scal (12.5%)
MOV (%R12,%RDI,8),%RSI	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	1	0.33	N/A
MOV -0x150(%RBP),%RDI	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	1	0.33	N/A
MOV (%RDI),%RDI	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	1	0.33	N/A
MOV %RSI,(%RDI,%R8,8)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	1	0.50	scal (12.5%)
MOV -0x90(%RBP),%RSI	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	1	0.33	N/A
MOV (%RSI),%RSI	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	1	0.33	N/A
MOVQ $0,(%RSI,%R8,8)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	1	0.50	scal (6.3%)
INC %R8	1	0	0	0	0	0	0	0	0	0	0	0	0	1	0.17	N/A
MOV %R8,-0x38(%RBP)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	1	0.50	scal (12.5%)
MOV -0xe8(%RBP),%RSI	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	1	0.33	N/A
MOV 0x8(%RSI,%RCX,8),%RSI	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	1	0.33	N/A
JMP 454630 <hypre_BoomerAMGBuildExtPIInterp.extracted+0x1240>	1	0	0	0	0	0	0	0	0	0	0	0	0	0	5.84	N/A

Function	hypre_BoomerAMGBuildExtPIInterp.extracted
Source file and lines	par_lr_interp.c:1516-1526
Module	exec

The loop is defined in /home/eoseret/qaas_runs_GNR/173-927-0874/intel/AMG/build/AMG/AMG/parcsr_ls/par_lr_interp.c:1516-1526.

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 not vectorized. 8 data elements could be processed at once in vector registers. By vectorizing your loop, you can lower the cost of an iteration from 1.67 to 0.21 cycles (8.00x 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 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

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

No data for this section

Slow data structures access

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

Details

Irregular (variable stride) or indirect: 2 occurrence(s)

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

Workaround

Try to remove indirect accesses. If applicable, precompute elements out of the innermost loop.

Type of elements and instruction set

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

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

The binary loop does not contain any FP arithmetical operations. The binary loop is loading 40 bytes.

Unroll opportunity

Loop is data access bound.

Workaround

Unroll your loop if trip count is significantly higher than target unroll factor and if some data references are common to consecutive iterations. This can be done manually. Or with the unroll (resp. unroll_and_jam) directive on top of the inner (resp. surrounding) loop. You can enforce an unroll factor: #pragma unroll_and_jam N, unroll_and_jam(N), unroll N or unroll(N)

General properties

nb instructions	10
nb uops	9
loop length	39
used x86 registers	7
used mmx registers	0
used xmm registers	0
used ymm registers	0
used zmm registers	0
nb stack references	2

Front-end

ASSUMED MACRO FUSION FIT IN UOP CACHE

micro-operation queue	1.67 cycles
front end	1.67 cycles

Back-end

	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11
uops	1.50	0.80	1.67	1.67	0.00	0.60	1.50	0.00	0.00	0.00	0.60	1.67
cycles	1.50	0.80	1.67	1.67	0.00	0.60	1.50	0.00	0.00	0.00	0.60	1.67

Execution ports to units layout:

P0 (256 bits): VPU, BRU, ALU, DIV/SQRT
P1 (256 bits): ALU, VPU
P2 (512 bits): load
P3 (512 bits): load
P4 (512 bits): store data
P5 (512 bits): ALU, VPU
P6: ALU, BRU
P7: store address
P8: store address
P9 (512 bits): store data
P10: ALU
P11 (512 bits): load

Cycles executing div or sqrt instructions	NA
Longest recurrence chain latency (RecMII)	1.00

Cycles summary

Front-end	1.67
Dispatch	1.67
Data deps.	1.00
Overall L1	1.67

Vectorization ratios

all	0%
load	0%
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	12%
load	12%
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	12%

Cycles and memory resources usage

Assuming all data fit into the L1 cache, each iteration of the binary loop takes 1.67 cycles. At this rate:

12% of peak load performance is reached (24.00 out of 192.00 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 1.67 to 1.50 cycles (1.11x speedup).

ASM code

In the binary file, the address of the loop is: 454633

Instruction	Nb FU	P0	P1	P2	P3	P5	P6	P10	P11	Latency	Recip. throughput	Vectorization
INC %RDX	1	0	0	0	0	0	0	0	0	1	0.17	N/A
CMP %RSI,%RDX	1	0.20	0.20	0	0	0.20	0.20	0.20	0	1	0.20	scal (12.5%)
JGE 454698 <hypre_BoomerAMGBuildExtPIInterp.extracted+0x12a8>	1	0.50	0	0	0	0	0.50	0	0	0	0.50	N/A
MOV -0x160(%RBP),%RDI	1	0	0	0.33	0.33	0	0	0	0.33	1	0.33	N/A
MOV (%RDI,%RDX,8),%RDI	1	0	0	0.33	0.33	0	0	0	0.33	1	0.33	N/A
MOV -0x100(%RBP),%R8	1	0	0	0.33	0.33	0	0	0	0.33	1	0.33	N/A
CMPQ $0,(%R8,%RDI,8)	1	0.20	0.20	0.33	0.33	0.20	0.20	0.20	0.33	1	0.33	scal (12.5%)
JS 454630 <hypre_BoomerAMGBuildExtPIInterp.extracted+0x1240>	1	0.50	0	0	0	0	0.50	0	0	0	0.50	N/A
CMP %R9,(%RBX,%RDI,8)	1	0.20	0.20	0.33	0.33	0.20	0.20	0.20	0.33	1	0.33	N/A
JGE 454630 <hypre_BoomerAMGBuildExtPIInterp.extracted+0x1240>	1	0.50	0	0	0	0	0.50	0	0	0	0.50	N/A

Report Configuration

Vectorization

Details

Workaround

Slow data structures access

Details

Workaround

Type of elements and instruction set

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

Unroll opportunity

Workaround

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

Slow data structures access

Details

Workaround

Type of elements and instruction set

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

Unroll opportunity

Workaround

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

Slow data structures access

Details

Workaround

Type of elements and instruction set

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

Unroll opportunity

Workaround

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

Slow data structures access

Details

Workaround

Type of elements and instruction set

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

Unroll opportunity

Workaround

General properties

Front-end

Back-end

Cycles summary

Vectorization ratios

Vector efficiency ratios

Cycles and memory resources usage

Front-end bottlenecks

ASM code