OV - attention-icx-skl256 - Loop 41

	Loop Id: 41	Module: attention-icx-skl256	Source: attention_v2.cpp:27-33	Coverage: 0.56%

0x404a20 ADD	%EDX,%EDI

0x404a22 VCVTSD2SS	%XMM0,%XMM0,%XMM0

0x404a26 MOV	0x108(%RSP),%R8

0x404a2e VMOVSS	%XMM0,(%R8,%RDI,4)

0x404a34 LEA	0x1(%RSI),%RDI

0x404a38 CMP	0x158(%RSP),%RSI

0x404a40 MOV	%RDI,%RSI

0x404a43 JE	404a00

0x404a45 TEST	%R15,%R15

0x404a48 JE	404af0

0x404a4e MOV	%ESI,%EDI

0x404a50 VPBROADCASTD	%ESI,%XMM1

0x404a56 VXORPS	%XMM0,%XMM0,%XMM0

0x404a5a XOR	%R8D,%R8D

0x404a5d NOPL	(%RAX)

  (42) 0x404a60 LEA	(%RAX,%R8,1),%R9D

  (42) 0x404a64 VPBROADCASTD	%R9D,%XMM2

  (42) 0x404a6a LEA	(%RBX,%R12,1),%R9

  (42) 0x404a6e VXORPD	%XMM3,%XMM3,%XMM3

  (42) 0x404a72 VPCMPEQB	%XMM0,%XMM0,%K1

  (42) 0x404a78 VPBROADCASTD	%R8D,%XMM4

  (42) 0x404a7e VPADDD	%XMM7,%XMM2,%XMM2

  (42) 0x404a82 VPADDD	%XMM4,%XMM8,%XMM4

  (42) 0x404a86 VPMULLD	%XMM4,%XMM6,%XMM4

  (42) 0x404a8b VPADDD	%XMM4,%XMM1,%XMM4

  (42) 0x404a8f LEA	(%R11,%R12,1),%R10

  (42) 0x404a93 VPSUBD	%XMM9,%XMM4,%XMM4

  (42) 0x404a98 VXORPS	%XMM5,%XMM5,%XMM5

  (42) 0x404a9c VGATHERDPS	(%R9,%XMM2,4),%XMM3{%K1}

  (42) 0x404aa3 VPCMPEQB	%XMM0,%XMM0,%K1

  (42) 0x404aa9 VGATHERDPS	(%R10,%XMM4,4),%XMM5{%K1}

  (42) 0x404ab0 VCVTPS2PD	%XMM3,%YMM2

  (42) 0x404ab4 VCVTPS2PD	%XMM5,%YMM3

  (42) 0x404ab8 VFMADD231PD	%YMM3,%YMM2,%YMM0

  (42) 0x404abd ADD	$0x4,%R8

  (42) 0x404ac1 CMP	%R15,%R8

  (42) 0x404ac4 JB	404a60

0x404ac6 VEXTRACTF128	$0x1,%YMM0,%XMM1

0x404acc VADDPD	%XMM1,%XMM0,%XMM0

0x404ad0 VSHUFPD	$0x1,%XMM0,%XMM0,%XMM1

0x404ad5 VADDSD	%XMM1,%XMM0,%XMM0

0x404ad9 MOV	%R15,%R8

0x404adc TESTB	$0x1,0xe4(%RSP)

0x404ae4 JNE	404a20

0x404aea JMP	404af7

0x404af0 VXORPS	%XMM0,%XMM0,%XMM0

0x404af4 XOR	%R8D,%R8D

0x404af7 MOV	%R13,%RDI

0x404afa IMUL	%R8,%RDI

0x404afe ADD	%RSI,%RDI

0x404b01 MOV	0xf0(%RSP),%R10

0x404b09 NOPL	(%RAX)

  (22) 0x404b10 LEA	(%RAX,%R8,1),%R9D

  (22) 0x404b14 VMOVSS	(%RBX,%R9,4),%XMM1

  (22) 0x404b1a VCVTSS2SD	%XMM1,%XMM1,%XMM1

  (22) 0x404b1e MOV	%EDI,%R9D

  (22) 0x404b21 VMOVSS	(%R10,%R9,4),%XMM2

  (22) 0x404b27 VCVTSS2SD	%XMM2,%XMM2,%XMM2

  (22) 0x404b2b VFMADD231SD	%XMM2,%XMM1,%XMM0

  (22) 0x404b30 INC	%R8

  (22) 0x404b33 ADD	%R13,%RDI

  (22) 0x404b36 CMP	%R8,%R14

  (22) 0x404b39 JNE	404b10

0x404b3b MOV	%ESI,%EDI

0x404b3d JMP	404a20

/home/eoseret/Applications/llm-attention/attention_v2.cpp: 27 - 33

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

27:       for (unsigned int j = 0; j < N; ++j) {  //vectorized

28:          double sum = 0.0;

29: #pragma clang loop vectorize(enable)

30:          for (unsigned int k = 0; k < K; ++k) { //vectorized

31:             sum += (double)A[i * K + k] * (double)B[k * N + j];

32:          }

33:          C[i * N + j] = alpha * static_cast<float>(sum);

Coverage (%)	Name	Module
►100.00+	__libc_init_first	libc.so.6
○	__libc_start_main	libc.so.6
○	_start	attention-icx-skl256

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	1.68
CQA speedup if FP arith vectorized	1.42
CQA speedup if fully vectorized	10.55
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.61
Bottlenecks
Function	main
Source	attention_v2.cpp:27-27,attention_v2.cpp:30-33
Source loop unroll info	NA
Source loop unroll confidence level	NA
Unroll/vectorization loop type	NA
Unroll factor	NA
CQA cycles	6.17
CQA cycles if no scalar integer	3.67
CQA cycles if FP arith vectorized	4.33
CQA cycles if fully vectorized	0.58
Front-end cycles	6.17
P0 cycles	3.75
P1 cycles	3.75
P2 cycles	1.67
P3 cycles	1.67
P4 cycles	1.00
P5 cycles	3.83
P6 cycles	3.67
P7 cycles	1.00
DIV/SQRT cycles	0.00
Inter-iter dependencies cycles	0
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	23.67
Nb uops	24.67
Nb loads	3.33
Nb stores	1.00
Nb stack references	3.33
FLOP/cycle	0.32
Nb FLOP add-sub	2.00
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	4.33
Bytes prefetched	0.00
Bytes loaded	22.00
Bytes stored	4.00
Stride 0	0.00
Stride 1	0.00
Stride n	0.00
Stride unknown	1.33
Stride indirect	1.33
Vectorization ratio all	30.00
Vectorization ratio load	0.00
Vectorization ratio store	0.00
Vectorization ratio mul	NA
Vectorization ratio add_sub	50.00
Vectorization ratio fma	NA
Vectorization ratio div_sqrt	NA
Vectorization ratio other	35.12
Vector-efficiency ratio all	14.84
Vector-efficiency ratio load	11.28
Vector-efficiency ratio store	6.25
Vector-efficiency ratio mul	NA
Vector-efficiency ratio add_sub	18.75
Vector-efficiency ratio fma	NA
Vector-efficiency ratio div_sqrt	NA
Vector-efficiency ratio other	15.88

Metric	Value
CQA speedup if no scalar integer	1.82
CQA speedup if FP arith vectorized	1.00
CQA speedup if fully vectorized	13.22
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.82
Bottlenecks	micro-operation queue,
Function	main
Source	attention_v2.cpp:27-27,attention_v2.cpp:30-33
Source loop unroll info	NA
Source loop unroll confidence level	NA
Unroll/vectorization loop type	NA
Unroll factor	NA
CQA cycles	5.00
CQA cycles if no scalar integer	2.75
CQA cycles if FP arith vectorized	5.00
CQA cycles if fully vectorized	0.38
Front-end cycles	5.00
P0 cycles	2.75
P1 cycles	2.75
P2 cycles	1.50
P3 cycles	1.50
P4 cycles	1.00
P5 cycles	2.75
P6 cycles	2.75
P7 cycles	1.00
DIV/SQRT cycles	0.00
Inter-iter dependencies cycles	0
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	19.00
Nb uops	20.00
Nb loads	3.00
Nb stores	1.00
Nb stack references	3.00
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	5.60
Bytes prefetched	0.00
Bytes loaded	24.00
Bytes stored	4.00
Stride 0	0.00
Stride 1	0.00
Stride n	0.00
Stride unknown	2.00
Stride indirect	0.00
Vectorization ratio all	16.67
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	25.00
Vector-efficiency ratio all	13.54
Vector-efficiency ratio load	12.50
Vector-efficiency ratio store	6.25
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	15.63

Metric	Value
CQA speedup if no scalar integer	1.82
CQA speedup if FP arith vectorized	1.94
CQA speedup if fully vectorized	11.27
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.63
Bottlenecks	micro-operation queue,
Function	main
Source	attention_v2.cpp:27-27,attention_v2.cpp:30-33
Source loop unroll info	NA
Source loop unroll confidence level	NA
Unroll/vectorization loop type	NA
Unroll factor	NA
CQA cycles	7.75
CQA cycles if no scalar integer	4.25
CQA cycles if FP arith vectorized	4.00
CQA cycles if fully vectorized	0.69
Front-end cycles	7.75
P0 cycles	4.75
P1 cycles	4.75
P2 cycles	2.00
P3 cycles	2.00
P4 cycles	1.00
P5 cycles	4.75
P6 cycles	4.75
P7 cycles	1.00
DIV/SQRT cycles	0.00
Inter-iter dependencies cycles	0
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	30.00
Nb uops	31.00
Nb loads	4.00
Nb stores	1.00
Nb stack references	4.00
FLOP/cycle	0.39
Nb FLOP add-sub	3.00
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	3.74
Bytes prefetched	0.00
Bytes loaded	25.00
Bytes stored	4.00
Stride 0	0.00
Stride 1	0.00
Stride n	0.00
Stride unknown	2.00
Stride indirect	2.00
Vectorization ratio all	33.33
Vectorization ratio load	0.00
Vectorization ratio store	0.00
Vectorization ratio mul	NA
Vectorization ratio add_sub	50.00
Vectorization ratio fma	NA
Vectorization ratio div_sqrt	NA
Vectorization ratio other	37.50
Vector-efficiency ratio all	14.71
Vector-efficiency ratio load	8.85
Vector-efficiency ratio store	6.25
Vector-efficiency ratio mul	NA
Vector-efficiency ratio add_sub	18.75
Vector-efficiency ratio fma	NA
Vector-efficiency ratio div_sqrt	NA
Vector-efficiency ratio other	15.04

Metric	Value
CQA speedup if no scalar integer	1.44
CQA speedup if FP arith vectorized	1.44
CQA speedup if fully vectorized	8.36
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.44
Bottlenecks	micro-operation queue,
Function	main
Source	attention_v2.cpp:27-27,attention_v2.cpp:30-33
Source loop unroll info	NA
Source loop unroll confidence level	NA
Unroll/vectorization loop type	NA
Unroll factor	NA
CQA cycles	5.75
CQA cycles if no scalar integer	4.00
CQA cycles if FP arith vectorized	4.00
CQA cycles if fully vectorized	0.69
Front-end cycles	5.75
P0 cycles	3.75
P1 cycles	3.75
P2 cycles	1.50
P3 cycles	1.50
P4 cycles	1.00
P5 cycles	4.00
P6 cycles	3.50
P7 cycles	1.00
DIV/SQRT cycles	0.00
Inter-iter dependencies cycles	0
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	22.00
Nb uops	23.00
Nb loads	3.00
Nb stores	1.00
Nb stack references	3.00
FLOP/cycle	0.52
Nb FLOP add-sub	3.00
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	3.65
Bytes prefetched	0.00
Bytes loaded	17.00
Bytes stored	4.00
Stride 0	0.00
Stride 1	0.00
Stride n	0.00
Stride unknown	0.00
Stride indirect	2.00
Vectorization ratio all	40.00
Vectorization ratio load	0.00
Vectorization ratio store	0.00
Vectorization ratio mul	NA
Vectorization ratio add_sub	50.00
Vectorization ratio fma	NA
Vectorization ratio div_sqrt	NA
Vectorization ratio other	42.86
Vector-efficiency ratio all	16.25
Vector-efficiency ratio load	12.50
Vector-efficiency ratio store	6.25
Vector-efficiency ratio mul	NA
Vector-efficiency ratio add_sub	18.75
Vector-efficiency ratio fma	NA
Vector-efficiency ratio div_sqrt	NA
Vector-efficiency ratio other	16.96

Path /

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

Function	main
Source file and lines	attention_v2.cpp:27-33
Module	attention-icx-skl256

The loop is defined in /home/eoseret/Applications/llm-attention/attention_v2.cpp:27-33.

Analyzed code is defined in /home/eoseret/Applications/llm-attention/attention_v2.cpp:27,30-33.

Warnings:
Non-innermost loop: analyzing only self part (ignoring child loops).
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 14% of vector register length is used (average across all SSE/AVX instructions).

Details

30% of SSE/AVX instructions are used in vector version (process two or more data elements in vector registers):

0% of SSE/AVX loads are used in vector version.
0% of SSE/AVX stores are used in vector version.
50% of SSE/AVX addition or subtraction instructions are used in vector version.
35% of SSE/AVX instructions that are not load, store, addition, subtraction nor multiply instructions are used in vector version.

Workaround

Read the "512-bits vectorization" report at "Potential" confidence level.

512-bits vectorization

On some x86 processors supporting 512-bits vectorization, compilers are often too conservative and limit vectorization to 256 bits. Performance can then be improved by enforcing 512-bits vectorization, especially with many vectorized and high trip count loops. 512-bits vectorization performance overhead (compared to 256-bits) is generally lower on newer processors.

Workaround

Recompile with -mprefer-vector-width=512

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.

VCVTSD2SS: 0 occurrences

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: 1 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.

Conversion instructions

Detected expensive conversion instructions.

Details

VCVTSD2SS (FP64 to FP32, scalar): 0 occurrences

Workaround

Avoid mixing data with different types. In particular, check if the type of constants is the same as array elements. Use double instead of single precision only when/where needed by numerical stability and avoid mixing precision. In C/C++, FP constants are double precision by default and must be suffixed by 'f' to make them single precision.

Type of elements and instruction set

0 SSE or AVX instructions are processing arithmetic or math operations on double precision FP elements in scalar mode (one at a time). 0 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 2 FP arithmetical operations:

2: addition or subtraction

The binary loop is loading 22 bytes (2 double precision FP elements). The binary loop is storing 4 bytes (0 double precision FP elements).

Arithmetic intensity

Arithmetic intensity is 0.08 FP operations per loaded or stored byte.

General properties

nb instructions	23.67
nb uops	24.67
loop length	106.33
used x86 registers	8.33
used mmx registers	0
used xmm registers	1.67
used ymm registers	0.67
used zmm registers	0
nb stack references	3.33

Front-end

MACRO FUSION NOT POSSIBLE

micro-operation queue	6.17 cycles
front end	6.17 cycles

Back-end

	P0	P1	P2	P3	P4	P5	P6	P7
uops	3.75	3.75	1.67	1.67	1.00	3.83	3.67	1.00
cycles	3.75	3.75	1.67	1.67	1.00	3.83	3.67	1.00

Execution ports to units layout:

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

Cycles executing div or sqrt instructions	NA
Longest recurrence chain latency (RecMII)	0.00

Cycles summary

Front-end	6.17
Dispatch	3.83
Data deps.	0.00
Overall L1	6.17

Vectorization ratios

INT

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)
other	0%

all	49%
load	NA (no load vectorizable/vectorized instructions)
store	0%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	50%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	66%

INT+FP

all	30%
load	0%
store	0%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	50%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	35%

Vector efficiency ratios

INT

all	10%
load	11%
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	10%

all	17%
load	NA (no load vectorizable/vectorized instructions)
store	6%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	18%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	20%

INT+FP

all	14%
load	11%
store	6%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	18%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	15%

Cycles and memory resources usage

Assuming all data fit into the L1 cache, each iteration of the binary loop takes 6.17 cycles. At this rate:

2% of peak load performance is reached (3.66 out of 128.00 bytes loaded per cycle (GB/s @ 1GHz))
1% of peak store performance is reached (0.67 out of 64.00 bytes stored per cycle (GB/s @ 1GHz))

Function	main
Source file and lines	attention_v2.cpp:27-33
Module	attention-icx-skl256

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 5.00 to 2.75 cycles (1.82x speedup).

Workaround

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

Vectorization

Your loop is probably not vectorized. Only 13% 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 5.00 to 0.38 cycles (13.22x speedup).

Details

Store and arithmetical 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

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.

VCVTSD2SS: 1 occurrences►
- /home/eoseret/Applications/llm-attention/attention_v2.cpp:33

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)

Conversion instructions

Detected expensive conversion instructions.

Details

VCVTSD2SS (FP64 to FP32, scalar): 1 occurrences►
- /home/eoseret/Applications/llm-attention/attention_v2.cpp:33

Workaround

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 24 bytes. The binary loop is storing 4 bytes.

General properties

nb instructions	19
nb uops	20
loop length	85
used x86 registers	9
used mmx registers	0
used xmm registers	1
used ymm registers	0
used zmm registers	0
nb stack references	3

Front-end

MACRO FUSION NOT POSSIBLE FIT IN UOP CACHE

micro-operation queue	5.00 cycles
front end	5.00 cycles

Back-end

	P0	P1	P2	P3	P4	P5	P6	P7
uops	2.75	2.75	1.50	1.50	1.00	2.75	2.75	1.00
cycles	2.75	2.75	1.50	1.50	1.00	2.75	2.75	1.00

Execution ports to units layout:

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

Cycles executing div or sqrt instructions	NA
Longest recurrence chain latency (RecMII)	0.00

Cycles summary

Front-end	5.00
Dispatch	2.75
Data deps.	0.00
Overall L1	5.00

Vectorization ratios

INT

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)
other	0%

all	33%
load	NA (no load vectorizable/vectorized instructions)
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	50%

INT+FP

all	16%
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	25%

Vector efficiency ratios

INT

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)
other	12%

all	14%
load	NA (no load vectorizable/vectorized instructions)
store	6%
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	18%

INT+FP

all	13%
load	12%
store	6%
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	15%

Cycles and memory resources usage

Assuming all data fit into the L1 cache, each iteration of the binary loop takes 5.00 cycles. At this rate:

3% of peak load performance is reached (4.80 out of 128.00 bytes loaded per cycle (GB/s @ 1GHz))
1% of peak store performance is reached (0.80 out of 64.00 bytes stored 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 5.00 to 2.75 cycles (1.82x speedup).

ASM code

In the binary file, the address of the loop is: 404a45

Instruction	Nb FU	P0	P1	P2	P3	P4	P5	P6	P7	Latency	Recip. throughput	Vectorization
ADD %EDX,%EDI	1	0.25	0.25	0	0	0	0.25	0.25	0	1	0.25	N/A
VCVTSD2SS %XMM0,%XMM0,%XMM0	2	0.50	0.50	0	0	0	1	0	0	5	1	scal (12.5%)
MOV 0x108(%RSP),%R8	1	0	0	0.50	0.50	0	0	0	0	4-5	0.50	N/A
VMOVSS %XMM0,(%R8,%RDI,4)	1	0	0	0.33	0.33	1	0	0	0.33	3	1	scal (6.3%)
LEA 0x1(%RSI),%RDI	1	0	0.50	0	0	0	0.50	0	0	1	0.50	N/A
CMP 0x158(%RSP),%RSI	1	0.25	0.25	0.50	0.50	0	0.25	0.25	0	1	0.50	scal (12.5%)
MOV %RDI,%RSI	1	0	0	0	0	0	0	0	0	0	0.25	N/A
JE 404a00 <main+0x18e0>	1	0.50	0	0	0	0	0	0.50	0	0	0.50-1	N/A
TEST %R15,%R15	1	0.25	0.25	0	0	0	0.25	0.25	0	1	0.25	scal (12.5%)
JE 404af0 <main+0x19d0>	1	0.50	0	0	0	0	0	0.50	0	0	0.50-1	N/A
VXORPS %XMM0,%XMM0,%XMM0	1	0	0	0	0	0	0	0	0	0	0.25	vect (25.0%)
XOR %R8D,%R8D	1	0	0	0	0	0	0	0	0	0	0.25	N/A
MOV %R13,%RDI	1	0	0	0	0	0	0	0	0	0	0.25	N/A
IMUL %R8,%RDI	1	0	1	0	0	0	0	0	0	3	1	N/A
ADD %RSI,%RDI	1	0.25	0.25	0	0	0	0.25	0.25	0	1	0.25	N/A
MOV 0xf0(%RSP),%R10	1	0	0	0.50	0.50	0	0	0	0	4-5	0.50	scal (12.5%)
NOPL (%RAX)	1	0	0	0	0	0	0	0	0	0	0.25	N/A
MOV %ESI,%EDI	1	0	0	0	0	0	0	0	0	0	0.25	N/A
JMP 404a20 <main+0x1900>	1	0	0	0	0	0	0	1	0	0	1-2	N/A

Function	main
Source file and lines	attention_v2.cpp:27-33
Module	attention-icx-skl256

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 7.75 to 4.25 cycles (1.82x speedup).

Workaround

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

Vectorization

Your loop is poorly vectorized. Only 14% of vector register length is used (average across all SSE/AVX instructions).

Details

33% of SSE/AVX instructions are used in vector version (process two or more data elements in vector registers):

0% of SSE/AVX loads are used in vector version.
0% of SSE/AVX stores are used in vector version.
50% of SSE/AVX addition or subtraction instructions are used in vector version.
37% of SSE/AVX instructions that are not load, store, addition, subtraction nor multiply instructions are used in vector version.

Workaround

Read the "512-bits vectorization" report at "Potential" confidence level.

Execution units bottlenecks

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

512-bits vectorization

Workaround

Recompile with -mprefer-vector-width=512

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.

VCVTSD2SS: 1 occurrences►
- /home/eoseret/Applications/llm-attention/attention_v2.cpp:33

Slow data structures access

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

Details

Constant unknown stride: 2 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.

Vector unaligned load/store instructions

Detected 1 suboptimal vector unaligned load/store instructions.

Details

VEXTRACTF128: 1 occurrences►
- /home/eoseret/Applications/llm-attention/attention_v2.cpp:31

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.

Conversion instructions

Detected expensive conversion instructions.

Details

VCVTSD2SS (FP64 to FP32, scalar): 1 occurrences►
- /home/eoseret/Applications/llm-attention/attention_v2.cpp:33

Workaround

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). 1 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 3 FP arithmetical operations:

3: addition or subtraction

The binary loop is loading 25 bytes (3 double precision FP elements). The binary loop is storing 4 bytes (0 double precision FP elements).

Arithmetic intensity

Arithmetic intensity is 0.10 FP operations per loaded or stored byte.

General properties

nb instructions	30
nb uops	31
loop length	134
used x86 registers	9
used mmx registers	0
used xmm registers	2
used ymm registers	1
used zmm registers	0
nb stack references	4

Front-end

MACRO FUSION NOT POSSIBLE FIT IN UOP CACHE

micro-operation queue	7.75 cycles
front end	7.75 cycles

Back-end

	P0	P1	P2	P3	P4	P5	P6	P7
uops	4.75	4.75	2.00	2.00	1.00	4.75	4.75	1.00
cycles	4.75	4.75	2.00	2.00	1.00	4.75	4.75	1.00

Execution ports to units layout:

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

Cycles executing div or sqrt instructions	NA
Longest recurrence chain latency (RecMII)	0.00

Cycles summary

Front-end	7.75
Dispatch	4.75
Data deps.	0.00
Overall L1	7.75

Vectorization ratios

INT

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)
other	0%

all	57%
load	NA (no load vectorizable/vectorized instructions)
store	0%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	50%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	75%

INT+FP

all	33%
load	0%
store	0%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	50%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	37%

Vector efficiency ratios

INT

all	9%
load	8%
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	8%

all	18%
load	NA (no load vectorizable/vectorized instructions)
store	6%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	18%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	21%

INT+FP

all	14%
load	8%
store	6%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	18%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	15%

Cycles and memory resources usage

Assuming all data fit into the L1 cache, each iteration of the binary loop takes 7.75 cycles. At this rate:

2% of peak load performance is reached (3.23 out of 128.00 bytes loaded per cycle (GB/s @ 1GHz))
0% of peak store performance is reached (0.52 out of 64.00 bytes stored per cycle (GB/s @ 1GHz))

Front-end bottlenecks

ASM code

In the binary file, the address of the loop is: 404a45

Instruction	Nb FU	P0	P1	P2	P3	P4	P5	P6	P7	Latency	Recip. throughput	Vectorization
ADD %EDX,%EDI	1	0.25	0.25	0	0	0	0.25	0.25	0	1	0.25	N/A
VCVTSD2SS %XMM0,%XMM0,%XMM0	2	0.50	0.50	0	0	0	1	0	0	5	1	scal (12.5%)
MOV 0x108(%RSP),%R8	1	0	0	0.50	0.50	0	0	0	0	4-5	0.50	N/A
VMOVSS %XMM0,(%R8,%RDI,4)	1	0	0	0.33	0.33	1	0	0	0.33	3	1	scal (6.3%)
LEA 0x1(%RSI),%RDI	1	0	0.50	0	0	0	0.50	0	0	1	0.50	N/A
CMP 0x158(%RSP),%RSI	1	0.25	0.25	0.50	0.50	0	0.25	0.25	0	1	0.50	scal (12.5%)
MOV %RDI,%RSI	1	0	0	0	0	0	0	0	0	0	0.25	N/A
JE 404a00 <main+0x18e0>	1	0.50	0	0	0	0	0	0.50	0	0	0.50-1	N/A
TEST %R15,%R15	1	0.25	0.25	0	0	0	0.25	0.25	0	1	0.25	scal (12.5%)
JE 404af0 <main+0x19d0>	1	0.50	0	0	0	0	0	0.50	0	0	0.50-1	N/A
MOV %ESI,%EDI	1	0	0	0	0	0	0	0	0	0	0.25	N/A
VPBROADCASTD %ESI,%XMM1	1	0	0	0	0	0	1	0	0	1	1	scal (6.3%)
VXORPS %XMM0,%XMM0,%XMM0	1	0	0	0	0	0	0	0	0	0	0.25	vect (25.0%)
XOR %R8D,%R8D	1	0	0	0	0	0	0	0	0	0	0.25	N/A
NOPL (%RAX)	1	0	0	0	0	0	0	0	0	0	0.25	N/A
VEXTRACTF128 $0x1,%YMM0,%XMM1	1	0	0	0	0	0	1	0	0	3	1	vect (25.0%)
VADDPD %XMM1,%XMM0,%XMM0	1	0.50	0.50	0	0	0	0	0	0	4	0.50	vect (25.0%)
VSHUFPD $0x1,%XMM0,%XMM0,%XMM1	1	0	0	0	0	0	1	0	0	1	1	vect (25.0%)
VADDSD %XMM1,%XMM0,%XMM0	1	0.50	0.50	0	0	0	0	0	0	4	0.50	scal (12.5%)
MOV %R15,%R8	1	0	0	0	0	0	0	0	0	0	0.25	N/A
TESTB $0x1,0xe4(%RSP)	1	0.25	0.25	0.50	0.50	0	0.25	0.25	0	1	0.50	scal (1.6%)
JNE 404a20 <main+0x1900>	1	0.50	0	0	0	0	0	0.50	0	0	0.50-1	N/A
JMP 404af7 <main+0x19d7>	1	0	0	0	0	0	0	1	0	0	1-2	N/A
MOV %R13,%RDI	1	0	0	0	0	0	0	0	0	0	0.25	N/A
IMUL %R8,%RDI	1	0	1	0	0	0	0	0	0	3	1	N/A
ADD %RSI,%RDI	1	0.25	0.25	0	0	0	0.25	0.25	0	1	0.25	N/A
MOV 0xf0(%RSP),%R10	1	0	0	0.50	0.50	0	0	0	0	4-5	0.50	scal (12.5%)
NOPL (%RAX)	1	0	0	0	0	0	0	0	0	0	0.25	N/A
MOV %ESI,%EDI	1	0	0	0	0	0	0	0	0	0	0.25	N/A
JMP 404a20 <main+0x1900>	1	0	0	0	0	0	0	1	0	0	1-2	N/A

Function	main
Source file and lines	attention_v2.cpp:27-33
Module	attention-icx-skl256

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 5.75 to 4.00 cycles (1.44x speedup).

Workaround

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

Vectorization

Your loop is poorly vectorized. Only 16% of vector register length is used (average across all SSE/AVX instructions).

Details

40% of SSE/AVX instructions are used in vector version (process two or more data elements in vector registers):

0% of SSE/AVX loads are used in vector version.
0% of SSE/AVX stores are used in vector version.
50% of SSE/AVX addition or subtraction instructions are used in vector version.
42% of SSE/AVX instructions that are not load, store, addition, subtraction nor multiply instructions are used in vector version.

Workaround

Read the "512-bits vectorization" report at "Potential" confidence level.

Execution units bottlenecks

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

512-bits vectorization

Workaround

Recompile with -mprefer-vector-width=512

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.

VCVTSD2SS: 1 occurrences►
- /home/eoseret/Applications/llm-attention/attention_v2.cpp:33

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.

Vector unaligned load/store instructions

Detected 1 suboptimal vector unaligned load/store instructions.

Details

VEXTRACTF128: 1 occurrences►
- /home/eoseret/Applications/llm-attention/attention_v2.cpp:31

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.

Conversion instructions

Detected expensive conversion instructions.

Details

VCVTSD2SS (FP64 to FP32, scalar): 1 occurrences►
- /home/eoseret/Applications/llm-attention/attention_v2.cpp:33

Workaround

Type of elements and instruction set

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

The binary loop is composed of 3 FP arithmetical operations:

3: addition or subtraction

The binary loop is loading 17 bytes (2 double precision FP elements). The binary loop is storing 4 bytes (0 double precision FP elements).

Arithmetic intensity

Arithmetic intensity is 0.14 FP operations per loaded or stored byte.

General properties

nb instructions	22
nb uops	23
loop length	100
used x86 registers	7
used mmx registers	0
used xmm registers	2
used ymm registers	1
used zmm registers	0
nb stack references	3

Front-end

MACRO FUSION NOT POSSIBLE FIT IN UOP CACHE

micro-operation queue	5.75 cycles
front end	5.75 cycles

Back-end

	P0	P1	P2	P3	P4	P5	P6	P7
uops	3.75	3.75	1.50	1.50	1.00	4.00	3.50	1.00
cycles	3.75	3.75	1.50	1.50	1.00	4.00	3.50	1.00

Execution ports to units layout:

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

Cycles executing div or sqrt instructions	NA
Longest recurrence chain latency (RecMII)	0.00

Cycles summary

Front-end	5.75
Dispatch	4.00
Data deps.	0.00
Overall L1	5.75

Vectorization ratios

INT

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)
other	0%

all	57%
load	NA (no load vectorizable/vectorized instructions)
store	0%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	50%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	75%

INT+FP

all	40%
load	0%
store	0%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	50%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	42%

Vector efficiency ratios

INT

all	10%
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)
other	10%

all	18%
load	NA (no load vectorizable/vectorized instructions)
store	6%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	18%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	21%

INT+FP

all	16%
load	12%
store	6%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	18%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	16%

Cycles and memory resources usage

Assuming all data fit into the L1 cache, each iteration of the binary loop takes 5.75 cycles. At this rate:

2% of peak load performance is reached (2.96 out of 128.00 bytes loaded per cycle (GB/s @ 1GHz))
1% of peak store performance is reached (0.70 out of 64.00 bytes stored per cycle (GB/s @ 1GHz))

Front-end bottlenecks

ASM code

In the binary file, the address of the loop is: 404a45

Instruction	Nb FU	P0	P1	P2	P3	P4	P5	P6	P7	Latency	Recip. throughput	Vectorization
ADD %EDX,%EDI	1	0.25	0.25	0	0	0	0.25	0.25	0	1	0.25	N/A
VCVTSD2SS %XMM0,%XMM0,%XMM0	2	0.50	0.50	0	0	0	1	0	0	5	1	scal (12.5%)
MOV 0x108(%RSP),%R8	1	0	0	0.50	0.50	0	0	0	0	4-5	0.50	N/A
VMOVSS %XMM0,(%R8,%RDI,4)	1	0	0	0.33	0.33	1	0	0	0.33	3	1	scal (6.3%)
LEA 0x1(%RSI),%RDI	1	0	0.50	0	0	0	0.50	0	0	1	0.50	N/A
CMP 0x158(%RSP),%RSI	1	0.25	0.25	0.50	0.50	0	0.25	0.25	0	1	0.50	scal (12.5%)
MOV %RDI,%RSI	1	0	0	0	0	0	0	0	0	0	0.25	N/A
JE 404a00 <main+0x18e0>	1	0.50	0	0	0	0	0	0.50	0	0	0.50-1	N/A
TEST %R15,%R15	1	0.25	0.25	0	0	0	0.25	0.25	0	1	0.25	scal (12.5%)
JE 404af0 <main+0x19d0>	1	0.50	0	0	0	0	0	0.50	0	0	0.50-1	N/A
MOV %ESI,%EDI	1	0	0	0	0	0	0	0	0	0	0.25	N/A
VPBROADCASTD %ESI,%XMM1	1	0	0	0	0	0	1	0	0	1	1	scal (6.3%)
VXORPS %XMM0,%XMM0,%XMM0	1	0	0	0	0	0	0	0	0	0	0.25	vect (25.0%)
XOR %R8D,%R8D	1	0	0	0	0	0	0	0	0	0	0.25	N/A
NOPL (%RAX)	1	0	0	0	0	0	0	0	0	0	0.25	N/A
VEXTRACTF128 $0x1,%YMM0,%XMM1	1	0	0	0	0	0	1	0	0	3	1	vect (25.0%)
VADDPD %XMM1,%XMM0,%XMM0	1	0.50	0.50	0	0	0	0	0	0	4	0.50	vect (25.0%)
VSHUFPD $0x1,%XMM0,%XMM0,%XMM1	1	0	0	0	0	0	1	0	0	1	1	vect (25.0%)
VADDSD %XMM1,%XMM0,%XMM0	1	0.50	0.50	0	0	0	0	0	0	4	0.50	scal (12.5%)
MOV %R15,%R8	1	0	0	0	0	0	0	0	0	0	0.25	N/A
TESTB $0x1,0xe4(%RSP)	1	0.25	0.25	0.50	0.50	0	0.25	0.25	0	1	0.50	N/A
JNE 404a20 <main+0x1900>	1	0.50	0	0	0	0	0	0.50	0	0	0.50-1	N/A

Report Configuration

Vectorization

Details

Workaround

512-bits vectorization

Workaround

Complex instructions

Details

Slow data structures access

Details

Workaround

Conversion 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

Code clean check

Workaround

Vectorization

Details

Workaround

Execution units bottlenecks

Complex instructions

Details

Slow data structures access

Details

Workaround

Conversion instructions

Details

Workaround

Type of elements and instruction set

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

Execution units bottlenecks

512-bits vectorization

Workaround

Complex instructions

Details

Slow data structures access

Details

Workaround

Vector unaligned load/store instructions

Details

Workaround

Conversion 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