OV - exec - Loop 3477

0x49c060 SDIV	X13, X8, X26

0x49c064 MSUB	X14, X26, X13, X8

0x49c068 ADD	X8, X8, #1

0x49c06c MADD	X10, X14, X10, XZR

0x49c070 ADD	X10, X11, X10,LSL #3

0x49c074 LDR	X14, [X10, X13,LSL #3]    [1]

0x49c078 ADD	X14, X14, X12

0x49c07c STR	X14, [X10, X13,LSL #3]    [1]

0x49c080 LDR	X10, [X19]    [2]

0x49c084 MADD	X13, X10, X28, X9

0x49c088 CMP	X8, X13

0x49c08c B.LT	49c060

/home/hbollore/qaas/qaas-runs/174-161-6712/intel/AMG/build/AMG/AMG/seq_mv/csr_matop.c: 380 - 527

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

380:   return idx%dim1*dim2 + idx/dim1;

[...]

524:      for (i = my_thread_num*num_colsA; i < (my_thread_num + 1)*num_colsA - 1; ++i) {

525:        HYPRE_Int transpose_i = transpose_idx(i, num_threads, num_colsA);

526:

527:        bucket[transpose_i] += offset;

Coverage (%)	Name	Source Location	Module
○100.00	__kmp_invoke_microtask		libomp.so

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	1.00
CQA speedup if fully vectorized	2.00
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.25 - 5.00
Bottlenecks	P8, P10,
Function	hypre_CSRMatrixTranspose.omp_outlined
Source	csr_matop.c:380-380,csr_matop.c:524-527
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	5.00 - 20.00
CQA cycles if no scalar integer	5.00 - 20.00
CQA cycles if FP arith vectorized	5.00 - 20.00
CQA cycles if fully vectorized	2.50 - 10.00
Front-end cycles	1.50
P0 cycles	0.50
P1 cycles	0.50
P2 cycles	0.83
P3 cycles	0.75
P4 cycles	0.83
P5 cycles	0.83
P6 cycles	4.00
P7 cycles	0.75
P8 cycles	0.00
P9 cycles	0.00
P10 cycles	0.00
P11 cycles	0.00
P12 cycles	1.17
P13 cycles	0.83
P14 cycles	1.00
P15 cycles	0.50
P16 cycles	0.50
DIV/SQRT cycles	5.00 - 20.00
Inter-iter dependencies cycles	1
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	12.00
Nb uops	12.00
Nb loads	NA
Nb stores	1.00
Nb stack references	0.00
FLOP/cycle	0.00 - 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	1.20 - 4.80
Bytes prefetched	0.00
Bytes loaded	16.00
Bytes stored	8.00
Stride 0	1.00
Stride 1	0.00
Stride n	0.00
Stride unknown	1.00
Stride indirect	0.00
Vectorization ratio all	0.00
Vectorization ratio load	0.00
Vectorization ratio store	0.00
Vectorization ratio mul	NA
Vectorization ratio add_sub	0.00
Vectorization ratio fma	0.00
Vectorization ratio div_sqrt	NA
Vectorization ratio other	NA
Vector-efficiency ratio all	50.00
Vector-efficiency ratio load	50.00
Vector-efficiency ratio store	50.00
Vector-efficiency ratio mul	NA
Vector-efficiency ratio add_sub	50.00
Vector-efficiency ratio fma	50.00
Vector-efficiency ratio div_sqrt	NA
Vector-efficiency ratio other	NA

Metric	Value
CQA speedup if no scalar integer	1.00
CQA speedup if FP arith vectorized	1.00
CQA speedup if fully vectorized	2.00
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.25 - 5.00
Bottlenecks	P8, P10,
Function	hypre_CSRMatrixTranspose.omp_outlined
Source	csr_matop.c:380-380,csr_matop.c:524-527
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	5.00 - 20.00
CQA cycles if no scalar integer	5.00 - 20.00
CQA cycles if FP arith vectorized	5.00 - 20.00
CQA cycles if fully vectorized	2.50 - 10.00
Front-end cycles	1.50
P0 cycles	0.50
P1 cycles	0.50
P2 cycles	0.83
P3 cycles	0.75
P4 cycles	0.83
P5 cycles	0.83
P6 cycles	4.00
P7 cycles	0.75
P8 cycles	0.00
P9 cycles	0.00
P10 cycles	0.00
P11 cycles	0.00
P12 cycles	1.17
P13 cycles	0.83
P14 cycles	1.00
P15 cycles	0.50
P16 cycles	0.50
DIV/SQRT cycles	5.00 - 20.00
Inter-iter dependencies cycles	1
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	12.00
Nb uops	12.00
Nb loads	NA
Nb stores	1.00
Nb stack references	0.00
FLOP/cycle	0.00 - 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	1.20 - 4.80
Bytes prefetched	0.00
Bytes loaded	16.00
Bytes stored	8.00
Stride 0	1.00
Stride 1	0.00
Stride n	0.00
Stride unknown	1.00
Stride indirect	0.00
Vectorization ratio all	0.00
Vectorization ratio load	0.00
Vectorization ratio store	0.00
Vectorization ratio mul	NA
Vectorization ratio add_sub	0.00
Vectorization ratio fma	0.00
Vectorization ratio div_sqrt	NA
Vectorization ratio other	NA
Vector-efficiency ratio all	50.00
Vector-efficiency ratio load	50.00
Vector-efficiency ratio store	50.00
Vector-efficiency ratio mul	NA
Vector-efficiency ratio add_sub	50.00
Vector-efficiency ratio fma	50.00
Vector-efficiency ratio div_sqrt	NA
Vector-efficiency ratio other	NA

Path /

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

Function	hypre_CSRMatrixTranspose.omp_outlined
Source file and lines	csr_matop.c:380-527
Module	exec

The loop is defined in /home/hbollore/qaas/qaas-runs/174-161-6712/intel/AMG/build/AMG/AMG/seq_mv/csr_matop.c:380,524-527.

The related source loop is not unrolled or unrolled with no peel/tail loop.

gain
potential
hint
expert

Vectorization

Your loop is not vectorized. 2 data elements could be processed at once in vector registers. By vectorizing your loop, you can lower the cost of an iteration from 20.00 to 10.00 cycles (2.00x 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
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

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 16 bytes. The binary loop is storing 8 bytes.

General properties

nb instructions	12
loop length	48
nb stack references	0

Front-end

FIT IN UOP CACHE

front end

1.50 cycles

Back-end

	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11	P12	P13	P14	P15	P16
uops	0.50	0.50	0.83	0.75	0.83	0.83	4.00	0.75	0.00	0.00	0.00	0.00	1.17	0.83	1.00	0.50	0.50
cycles	0.50	0.50	0.83	0.75	0.83	0.83	4.00	0.75	0.00	0.00	0.00	0.00	1.17	0.83	1.00	0.50	0.50

Execution ports to units layout:

P0: BRU
P1: BRU
P2: ALU
P3: ALU
P4: ALU
P5: ALU
P6: ALU
P7: ALU
P8 (128 bits): VPU, FP store data, ALU, DIV/SQRT
P9 (128 bits): VPU, ALU, FP store data
P10 (128 bits): VPU, ALU, DIV/SQRT
P11 (128 bits): ALU, VPU
P12 (256 bits): store address, load
P13 (256 bits): store address, load
P14 (256 bits): load
P15 (64 bits): store data
P16 (64 bits): store data

Cycles executing div or sqrt instructions	5.00-20.00
Longest recurrence chain latency (RecMII)	1.00

Cycles summary

Front-end	1.50
Data deps.	1.00
Overall L1	5.00-20.00

Vectorization ratios

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

Cycles and memory resources usage

Assuming all data fit into the L1 cache, each iteration of the binary loop takes 20.00 cycles. At this rate:

0% of peak load performance is reached (0.80 out of 96.00 bytes loaded per cycle (GB/s @ 1GHz))
0% of peak store performance is reached (0.40 out of 48.00 bytes stored per cycle (GB/s @ 1GHz))

ASM code

In the binary file, the address of the loop is: 49c060

Instruction	Nb FU	P0	P1	P2	P3	P4	P5	P6	P7	P12	P13	P14	P15	P16	Latency	Recip. throughput	Vectorization
SDIV X13, X8, X26	1	0	0	0	0	0	0	1	0	0	0	0	0	0	5-20	5-20	N/A
MSUB X14, X26, X13, X8	1	0	0	0	0	0	0	1	0	0	0	0	0	0	2	1	scal (50.0%)
ADD X8, X8, #1	1	0	0	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	1	0.17	N/A
MADD X10, X14, X10, XZR	1	0	0	0	0	0	0	1	0	0	0	0	0	0	2	1	N/A
ADD X10, X11, X10,LSL #3	1	0	0	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	1	0.17	N/A
LDR X14, [X10, X13,LSL #3]	1	0	0	0	0	0	0	0	0	0.33	0.33	0.33	0	0	4	0.33	scal (50.0%)
ADD X14, X14, X12	1	0	0	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	1	0.17	scal (50.0%)
STR X14, [X10, X13,LSL #3]	1	0	0	0	0	0	0	0	0	0.50	0.50	0	0.50	0.50	1	0.50	scal (50.0%)
LDR X10, [X19]	1	0	0	0	0	0	0	0	0	0.33	0.33	0.33	0	0	4	0.33	N/A
MADD X13, X10, X28, X9	1	0	0	0	0	0	0	1	0	0	0	0	0	0	2	1	N/A
CMP X8, X13	1	0	0	0.25	0.25	0	0	0.25	0.25	0	0	0	0	0	1	0.33	N/A
B.LT 49c060 <hypre_CSRMatrixTranspose.omp_outlined+0x310>	1	0.50	0.50	0	0	0	0	0	0	0	0	0	0	0	1	0.50	N/A

Function	hypre_CSRMatrixTranspose.omp_outlined
Source file and lines	csr_matop.c:380-527
Module	exec

The loop is defined in /home/hbollore/qaas/qaas-runs/174-161-6712/intel/AMG/build/AMG/AMG/seq_mv/csr_matop.c:380,524-527.

The related source loop is not unrolled or unrolled with no peel/tail loop.

gain
potential
hint
expert

Vectorization

Your loop is not vectorized. 2 data elements could be processed at once in vector registers. By vectorizing your loop, you can lower the cost of an iteration from 20.00 to 10.00 cycles (2.00x 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
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

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 16 bytes. The binary loop is storing 8 bytes.

General properties

nb instructions	12
loop length	48
nb stack references	0

Front-end

FIT IN UOP CACHE

front end

1.50 cycles

Back-end

	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11	P12	P13	P14	P15	P16
uops	0.50	0.50	0.83	0.75	0.83	0.83	4.00	0.75	0.00	0.00	0.00	0.00	1.17	0.83	1.00	0.50	0.50
cycles	0.50	0.50	0.83	0.75	0.83	0.83	4.00	0.75	0.00	0.00	0.00	0.00	1.17	0.83	1.00	0.50	0.50

Execution ports to units layout:

P0: BRU
P1: BRU
P2: ALU
P3: ALU
P4: ALU
P5: ALU
P6: ALU
P7: ALU
P8 (128 bits): VPU, FP store data, ALU, DIV/SQRT
P9 (128 bits): VPU, ALU, FP store data
P10 (128 bits): VPU, ALU, DIV/SQRT
P11 (128 bits): ALU, VPU
P12 (256 bits): store address, load
P13 (256 bits): store address, load
P14 (256 bits): load
P15 (64 bits): store data
P16 (64 bits): store data

Cycles executing div or sqrt instructions	5.00-20.00
Longest recurrence chain latency (RecMII)	1.00

Cycles summary

Front-end	1.50
Data deps.	1.00
Overall L1	5.00-20.00

Vectorization ratios

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

Cycles and memory resources usage

Assuming all data fit into the L1 cache, each iteration of the binary loop takes 20.00 cycles. At this rate:

0% of peak load performance is reached (0.80 out of 96.00 bytes loaded per cycle (GB/s @ 1GHz))
0% of peak store performance is reached (0.40 out of 48.00 bytes stored per cycle (GB/s @ 1GHz))

ASM code

In the binary file, the address of the loop is: 49c060

Instruction	Nb FU	P0	P1	P2	P3	P4	P5	P6	P7	P12	P13	P14	P15	P16	Latency	Recip. throughput	Vectorization
SDIV X13, X8, X26	1	0	0	0	0	0	0	1	0	0	0	0	0	0	5-20	5-20	N/A
MSUB X14, X26, X13, X8	1	0	0	0	0	0	0	1	0	0	0	0	0	0	2	1	scal (50.0%)
ADD X8, X8, #1	1	0	0	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	1	0.17	N/A
MADD X10, X14, X10, XZR	1	0	0	0	0	0	0	1	0	0	0	0	0	0	2	1	N/A
ADD X10, X11, X10,LSL #3	1	0	0	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	1	0.17	N/A
LDR X14, [X10, X13,LSL #3]	1	0	0	0	0	0	0	0	0	0.33	0.33	0.33	0	0	4	0.33	scal (50.0%)
ADD X14, X14, X12	1	0	0	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	1	0.17	scal (50.0%)
STR X14, [X10, X13,LSL #3]	1	0	0	0	0	0	0	0	0	0.50	0.50	0	0.50	0.50	1	0.50	scal (50.0%)
LDR X10, [X19]	1	0	0	0	0	0	0	0	0	0.33	0.33	0.33	0	0	4	0.33	N/A
MADD X13, X10, X28, X9	1	0	0	0	0	0	0	1	0	0	0	0	0	0	2	1	N/A
CMP X8, X13	1	0	0	0.25	0.25	0	0	0.25	0.25	0	0	0	0	0	1	0.33	N/A
B.LT 49c060 <hypre_CSRMatrixTranspose.omp_outlined+0x310>	1	0.50	0.50	0	0	0	0	0	0	0	0	0	0	0	1	0.50	N/A

Report Configuration

Vectorization

Details

Workaround

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

General properties

Front-end

Back-end

Cycles summary

Vectorization ratios

Cycles and memory resources usage

ASM code

Vectorization

Details

Workaround

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

General properties

Front-end

Back-end

Cycles summary

Vectorization ratios

Cycles and memory resources usage

ASM code