OV - exec - Loop 33

0x24f9a0 VMOVSD	(%R14,%R13,8),%XMM0    [1]

0x24f9a6 VMOVSD	(%RSI,%RAX,8),%XMM1    [2]

0x24f9ab VFMADD213SD	(%RCX,%RAX,8),%XMM0,%XMM1    [3]

0x24f9b1 VMOVSD	%XMM1,(%RCX,%RAX,8)    [3]

0x24f9b6 INC	%RAX

0x24f9b9 CMP	%RAX,%R9

0x24f9bc JNE	24f9a0

/beegfs/hackathon/users/eoseret/qaas_runs/170-850-6313/intel/Kripke/build/Kripke/src/Kripke/Kernel/LTimes.cpp: 62 - 62

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

62:            phi(nm,g,z) += ell(nm, d) * psi(d, g, z);

/beegfs/hackathon/users/eoseret/qaas_runs/170-850-6313/intel/Kripke/build/Kripke/tpl/raja/include/RAJA/policy/loop/forall.hpp: 59 - 59

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

59:   for (decltype(distance_it) i = 0; i < distance_it; ++i) {

Path /

Metric	Value
CQA speedup if no scalar integer	1.00
CQA speedup if FP arith vectorized	1.00
CQA speedup if fully vectorized	4.57
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.33
Bottlenecks	P5, P6, P7,
Function	.omp_outlined.#0x24f610
Source	LTimes.cpp:62-62,forall.hpp:59-59
Source loop unroll info	unrolled by 32
Source loop unroll confidence level	high
Unroll/vectorization loop type	peel/tail
Unroll factor	1
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	0.29
Front-end cycles	1.00
DIV/SQRT cycles	0.50
P0 cycles	0.50
P1 cycles	0.25
P2 cycles	0.25
P3 cycles	0.50
P4 cycles	1.33
P5 cycles	1.33
P6 cycles	1.33
P7 cycles	0.50
P8 cycles	0.50
P9 cycles	0.00
P10 cycles	0.00
P11 cycles	0.50
P12 cycles	0.50
P13 cycles	0.00
Inter-iter dependencies cycles	1
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	7.00
Nb uops	6.00
Nb loads	3.00
Nb stores	1.00
Nb stack references	0.00
FLOP/cycle	1.50
Nb FLOP add-sub	0.00
Nb FLOP mul	0.00
Nb FLOP fma	1.00
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	24.00
Bytes stored	8.00
Stride 0	1.00
Stride 1	2.00
Stride n	0.00
Stride unknown	0.00
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	NA
Vectorization ratio fma	0.00
Vectorization ratio div_sqrt	NA
Vectorization ratio other	NA
Vector-efficiency ratio all	12.50
Vector-efficiency ratio load	12.50
Vector-efficiency ratio store	12.50
Vector-efficiency ratio mul	NA
Vector-efficiency ratio add_sub	NA
Vector-efficiency ratio fma	12.50
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	4.57
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.33
Bottlenecks	P5, P6, P7,
Function	.omp_outlined.#0x24f610
Source	LTimes.cpp:62-62,forall.hpp:59-59
Source loop unroll info	unrolled by 32
Source loop unroll confidence level	high
Unroll/vectorization loop type	peel/tail
Unroll factor	1
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	0.29
Front-end cycles	1.00
DIV/SQRT cycles	0.50
P0 cycles	0.50
P1 cycles	0.25
P2 cycles	0.25
P3 cycles	0.50
P4 cycles	1.33
P5 cycles	1.33
P6 cycles	1.33
P7 cycles	0.50
P8 cycles	0.50
P9 cycles	0.00
P10 cycles	0.00
P11 cycles	0.50
P12 cycles	0.50
P13 cycles	0.00
Inter-iter dependencies cycles	1
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	7.00
Nb uops	6.00
Nb loads	3.00
Nb stores	1.00
Nb stack references	0.00
FLOP/cycle	1.50
Nb FLOP add-sub	0.00
Nb FLOP mul	0.00
Nb FLOP fma	1.00
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	24.00
Bytes stored	8.00
Stride 0	1.00
Stride 1	2.00
Stride n	0.00
Stride unknown	0.00
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	NA
Vectorization ratio fma	0.00
Vectorization ratio div_sqrt	NA
Vectorization ratio other	NA
Vector-efficiency ratio all	12.50
Vector-efficiency ratio load	12.50
Vector-efficiency ratio store	12.50
Vector-efficiency ratio mul	NA
Vector-efficiency ratio add_sub	NA
Vector-efficiency ratio fma	12.50
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	.omp_outlined.#0x24f610
Source file and lines	forall.hpp:59-59
Module	exec

The loop is defined in:

/beegfs/hackathon/users/eoseret/qaas_runs/170-850-6313/intel/Kripke/build/Kripke/src/Kripke/Kernel/LTimes.cpp:62
/beegfs/hackathon/users/eoseret/qaas_runs/170-850-6313/intel/Kripke/build/Kripke/tpl/raja/include/RAJA/policy/loop/forall.hpp:59

It is peel/tail loop of related source loop which is unrolled by 32 (including vectorization).

gain
potential
hint
expert

Unrolling/vectorization cost

This loop is peel/tail of a unrolled/vectorized loop. If its cost is not negligible compared to the main (unrolled/vectorized) loop, unrolling/vectorization is counterproductive due to low trip count.

Details

The more iterations the main loop is processing, the higher the trip count must be to amortize peel/tail overhead.

Workaround

recompile with -fprofile-instr-generate, execute ,merge raw profiles with 'llvm-profdata merge -o default.profdata default.profraw' and recompile with -fprofile-instr-use (profile-guided optimization)
hardcode most frequent values of loop bounds by adding specialized paths.:
- For instance, replace for (i=0; i<n; i++) foo(i) with: switch (n) { case (4): for (i=0; i<4; i++) foo(i); break; case (6): for (i=0; i<6; i++) foo(i); break; default : for (i=0; i<n; i++) foo(i); break; }

Vectorization

Your loop is not vectorized. 8 data elements could be processed at once in vector registers.

Details

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

Execution units bottlenecks

Performance is limited by:

reading data from caches/RAM (load units are a bottleneck)
writing data to caches/RAM (the store unit 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).

Workaround

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

FMA

Detected 1 FMA (fused multiply-add) operations.

Type of elements and instruction set

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

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

The binary loop is composed of 2 FP arithmetical operations:

1: addition or subtraction (all inside FMA instructions)
1: multiply (all inside FMA instructions)

The binary loop is loading 24 bytes (3 double precision FP elements). The binary loop is storing 8 bytes (1 double precision FP elements).

Arithmetic intensity

Arithmetic intensity is 0.06 FP operations per loaded or stored byte.

General properties

nb instructions	7
nb uops	6
loop length	30
used x86 registers	6
used mmx registers	0
used xmm registers	2
used ymm registers	0
used zmm registers	0
nb stack references	0

Front-end

ASSUMED MACRO FUSION FIT IN UOP CACHE

micro-operation queue	1.00 cycles
front end	1.00 cycles

Back-end

	ALU0/BRU0	ALU1	ALU2	ALU3	BRU1	AGU0	AGU1	AGU2	FP0	FP1	FP2	FP3	FP4	FP5
uops	0.50	0.50	0.25	0.25	0.50	1.33	1.33	1.33	0.50	0.50	0.00	0.00	0.50	0.50
cycles	0.50	0.50	0.25	0.25	0.50	1.33	1.33	1.33	0.50	0.50	0.00	0.00	0.50	0.50

Execution ports to units layout:

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

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

Cycles summary

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

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	0%
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	NA (no other vectorizable/vectorized instructions)

Vector efficiency ratios

all	12%
load	12%
store	12%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	NA (no add-sub vectorizable/vectorized instructions)
fma	12%
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 1.33 cycles. At this rate:

18% of peak load performance is reached (18.00 out of 96.00 bytes loaded per cycle (GB/s @ 1GHz))
9% of peak store performance is reached (6.00 out of 64.00 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: 24f9a0

Instruction	Nb FU	ALU0/BRU0	ALU1	ALU2	ALU3	BRU1	AGU0	AGU1	AGU2	FP0	FP1	FP4	FP5	Latency	Recip. throughput
VMOVSD (%R14,%R13,8),%XMM0	1	0	0	0	0	0	0.33	0.33	0.33	0	0	0	0	1	0.50
VMOVSD (%RSI,%RAX,8),%XMM1	1	0	0	0	0	0	0.33	0.33	0.33	0	0	0	0	1	0.50
VFMADD213SD (%RCX,%RAX,8),%XMM0,%XMM1	1	0	0	0	0	0	0.33	0.33	0.33	0.50	0.50	0	0	4	0.50
VMOVSD %XMM1,(%RCX,%RAX,8)	1	0	0	0	0	0	0.33	0.33	0.33	0	0	0.50	0.50	1	1
INC %RAX	1	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	0	1	0.25
CMP %RAX,%R9	1	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	0	1	0.25
JNE 24f9a0 <.omp_outlined.+0x390>	1	0.50	0	0	0	0.50	0	0	0	0	0	0	0	1	0.50-1

Function	.omp_outlined.#0x24f610
Source file and lines	forall.hpp:59-59
Module	exec

The loop is defined in:

/beegfs/hackathon/users/eoseret/qaas_runs/170-850-6313/intel/Kripke/build/Kripke/src/Kripke/Kernel/LTimes.cpp:62
/beegfs/hackathon/users/eoseret/qaas_runs/170-850-6313/intel/Kripke/build/Kripke/tpl/raja/include/RAJA/policy/loop/forall.hpp:59

It is peel/tail loop of related source loop which is unrolled by 32 (including vectorization).

gain
potential
hint
expert

Unrolling/vectorization cost

This loop is peel/tail of a unrolled/vectorized loop. If its cost is not negligible compared to the main (unrolled/vectorized) loop, unrolling/vectorization is counterproductive due to low trip count.

Details

The more iterations the main loop is processing, the higher the trip count must be to amortize peel/tail overhead.

Workaround

recompile with -fprofile-instr-generate, execute ,merge raw profiles with 'llvm-profdata merge -o default.profdata default.profraw' and recompile with -fprofile-instr-use (profile-guided optimization)
hardcode most frequent values of loop bounds by adding specialized paths.:
- For instance, replace for (i=0; i<n; i++) foo(i) with: switch (n) { case (4): for (i=0; i<4; i++) foo(i); break; case (6): for (i=0; i<6; i++) foo(i); break; default : for (i=0; i<n; i++) foo(i); break; }

Vectorization

Your loop is not vectorized. 8 data elements could be processed at once in vector registers.

Details

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

Execution units bottlenecks

Performance is limited by:

reading data from caches/RAM (load units are a bottleneck)
writing data to caches/RAM (the store unit 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).

Workaround

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

FMA

Detected 1 FMA (fused multiply-add) operations.

Type of elements and instruction set

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

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

The binary loop is composed of 2 FP arithmetical operations:

1: addition or subtraction (all inside FMA instructions)
1: multiply (all inside FMA instructions)

The binary loop is loading 24 bytes (3 double precision FP elements). The binary loop is storing 8 bytes (1 double precision FP elements).

Arithmetic intensity

Arithmetic intensity is 0.06 FP operations per loaded or stored byte.

General properties

nb instructions	7
nb uops	6
loop length	30
used x86 registers	6
used mmx registers	0
used xmm registers	2
used ymm registers	0
used zmm registers	0
nb stack references	0

Front-end

ASSUMED MACRO FUSION FIT IN UOP CACHE

micro-operation queue	1.00 cycles
front end	1.00 cycles

Back-end

	ALU0/BRU0	ALU1	ALU2	ALU3	BRU1	AGU0	AGU1	AGU2	FP0	FP1	FP2	FP3	FP4	FP5
uops	0.50	0.50	0.25	0.25	0.50	1.33	1.33	1.33	0.50	0.50	0.00	0.00	0.50	0.50
cycles	0.50	0.50	0.25	0.25	0.50	1.33	1.33	1.33	0.50	0.50	0.00	0.00	0.50	0.50

Execution ports to units layout:

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

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

Cycles summary

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

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	0%
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	NA (no other vectorizable/vectorized instructions)

Vector efficiency ratios

all	12%
load	12%
store	12%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	NA (no add-sub vectorizable/vectorized instructions)
fma	12%
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 1.33 cycles. At this rate:

18% of peak load performance is reached (18.00 out of 96.00 bytes loaded per cycle (GB/s @ 1GHz))
9% of peak store performance is reached (6.00 out of 64.00 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: 24f9a0

Instruction	Nb FU	ALU0/BRU0	ALU1	ALU2	ALU3	BRU1	AGU0	AGU1	AGU2	FP0	FP1	FP4	FP5	Latency	Recip. throughput
VMOVSD (%R14,%R13,8),%XMM0	1	0	0	0	0	0	0.33	0.33	0.33	0	0	0	0	1	0.50
VMOVSD (%RSI,%RAX,8),%XMM1	1	0	0	0	0	0	0.33	0.33	0.33	0	0	0	0	1	0.50
VFMADD213SD (%RCX,%RAX,8),%XMM0,%XMM1	1	0	0	0	0	0	0.33	0.33	0.33	0.50	0.50	0	0	4	0.50
VMOVSD %XMM1,(%RCX,%RAX,8)	1	0	0	0	0	0	0.33	0.33	0.33	0	0	0.50	0.50	1	1
INC %RAX	1	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	0	1	0.25
CMP %RAX,%R9	1	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	0	1	0.25
JNE 24f9a0 <.omp_outlined.+0x390>	1	0.50	0	0	0	0.50	0	0	0	0	0	0	0	1	0.50-1

Report Configuration

Unrolling/vectorization cost

Details

Workaround

Vectorization

Details

Execution units bottlenecks

Workaround

FMA

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

Unrolling/vectorization cost

Details

Workaround

Vectorization

Details

Execution units bottlenecks

Workaround

FMA

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