OV - attention-avx512 - Loop 21

0x404ec0 LEA	(%R8,%R14,1),%R15D

0x404ec4 VCVTPS2PD	(%RBX,%R15,4),%ZMM7    [1]

0x404ecb VPBROADCASTD	%R14D,%YMM8

0x404ed1 VPADDD	%YMM2,%YMM8,%YMM8

0x404ed5 VPMULLD	%YMM8,%YMM10,%YMM8

0x404eda VPADDD	%YMM6,%YMM8,%YMM8

0x404ede VPSUBD	%YMM3,%YMM8,%YMM8

0x404ee2 VPCMPEQB	%XMM0,%XMM0,%K1

0x404ee8 VXORPS	%XMM9,%XMM9,%XMM9

0x404eed VGATHERDPS	(%RDI,%YMM8,4),%YMM9{%K1}    [2]

0x404ef4 VCVTPS2PD	%YMM9,%ZMM8

0x404efa VFMADD231PD	%ZMM8,%ZMM7,%ZMM5

0x404f00 ADD	$0x8,%R14

0x404f04 CMP	%RDX,%R14

0x404f07 JB	404ec0

/home/eoseret/llm-attention/attention.cpp: 30 - 31

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

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

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

Coverage (%)	Name	Source Location	Module

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	1.91
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.43
Bottlenecks	P0, P1, P5,
Function	main
Source	attention.cpp:30-31
Source loop unroll info	multi-versionned
Source loop unroll confidence level	max
Unroll/vectorization loop type	NA
Unroll factor	NA
CQA cycles	5.00
CQA cycles if no scalar integer	5.00
CQA cycles if FP arith vectorized	5.00
CQA cycles if fully vectorized	2.62
Front-end cycles	3.50
P0 cycles	5.00
P1 cycles	5.00
P2 cycles	3.00
P3 cycles	3.00
P4 cycles	0.00
P5 cycles	5.00
P6 cycles	1.00
P7 cycles	0.00
P8 cycles	0.00
P9 cycles	0.00
P10 cycles	1.00
P11 cycles	3.00
DIV/SQRT cycles	0.00
Inter-iter dependencies cycles	4
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	15.00
Nb uops	21.00
Nb loads	2.00
Nb stores	0.00
Nb stack references	0.00
FLOP/cycle	3.20
Nb FLOP add-sub	0.00
Nb FLOP mul	0.00
Nb FLOP fma	8.00
Nb FLOP div	0.00
Nb FLOP rcp	0.00
Nb FLOP sqrt	0.00
Nb FLOP rsqrt	0.00
Bytes/cycle	12.80
Bytes prefetched	0.00
Bytes loaded	64.00
Bytes stored	0.00
Stride 0	0.00
Stride 1	1.00
Stride n	0.00
Stride unknown	1.00
Stride indirect	0.00
Vectorization ratio all	90.91
Vectorization ratio load	100.00
Vectorization ratio store	NA
Vectorization ratio mul	100.00
Vectorization ratio add_sub	100.00
Vectorization ratio fma	100.00
Vectorization ratio div_sqrt	NA
Vectorization ratio other	83.33
Vector-efficiency ratio all	46.02
Vector-efficiency ratio load	50.00
Vector-efficiency ratio store	NA
Vector-efficiency ratio mul	50.00
Vector-efficiency ratio add_sub	50.00
Vector-efficiency ratio fma	100.00
Vector-efficiency ratio div_sqrt	NA
Vector-efficiency ratio other	34.38

Metric	Value
CQA speedup if no scalar integer	1.00
CQA speedup if FP arith vectorized	1.00
CQA speedup if fully vectorized	1.91
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.43
Bottlenecks	P0, P1, P5,
Function	main
Source	attention.cpp:30-31
Source loop unroll info	multi-versionned
Source loop unroll confidence level	max
Unroll/vectorization loop type	NA
Unroll factor	NA
CQA cycles	5.00
CQA cycles if no scalar integer	5.00
CQA cycles if FP arith vectorized	5.00
CQA cycles if fully vectorized	2.62
Front-end cycles	3.50
P0 cycles	5.00
P1 cycles	5.00
P2 cycles	3.00
P3 cycles	3.00
P4 cycles	0.00
P5 cycles	5.00
P6 cycles	1.00
P7 cycles	0.00
P8 cycles	0.00
P9 cycles	0.00
P10 cycles	1.00
P11 cycles	3.00
DIV/SQRT cycles	0.00
Inter-iter dependencies cycles	4
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	15.00
Nb uops	21.00
Nb loads	2.00
Nb stores	0.00
Nb stack references	0.00
FLOP/cycle	3.20
Nb FLOP add-sub	0.00
Nb FLOP mul	0.00
Nb FLOP fma	8.00
Nb FLOP div	0.00
Nb FLOP rcp	0.00
Nb FLOP sqrt	0.00
Nb FLOP rsqrt	0.00
Bytes/cycle	12.80
Bytes prefetched	0.00
Bytes loaded	64.00
Bytes stored	0.00
Stride 0	0.00
Stride 1	1.00
Stride n	0.00
Stride unknown	1.00
Stride indirect	0.00
Vectorization ratio all	90.91
Vectorization ratio load	100.00
Vectorization ratio store	NA
Vectorization ratio mul	100.00
Vectorization ratio add_sub	100.00
Vectorization ratio fma	100.00
Vectorization ratio div_sqrt	NA
Vectorization ratio other	83.33
Vector-efficiency ratio all	46.02
Vector-efficiency ratio load	50.00
Vector-efficiency ratio store	NA
Vector-efficiency ratio mul	50.00
Vector-efficiency ratio add_sub	50.00
Vector-efficiency ratio fma	100.00
Vector-efficiency ratio div_sqrt	NA
Vector-efficiency ratio other	34.38

Path /

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

Function	main
Source file and lines	attention.cpp:30-31
Module	attention-avx512

The loop is defined in /home/eoseret/llm-attention/attention.cpp:30-31.

The related source loop is multi-versionned.

gain
potential
hint
expert

Vectorization

Your loop is highly vectorized. Only 46% of vector register length is used (average across all SSE/AVX instructions). By fully vectorizing your loop, you can lower the cost of an iteration from 5.00 to 2.62 cycles (1.91x speedup).

Details

90% of SSE/AVX instructions are used in vector version (process two or more data elements in vector registers):

83% of SSE/AVX instructions that are not load, store, addition, subtraction nor multiply instructions are used in vector version.

Since your execution units are vector units, only a fully vectorized loop can use their full power.

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.

Execution units bottlenecks

Performance is limited by execution of FP multiply or FMA (fused multiply-add) operations (the FP multiply/FMA unit is a bottleneck). By removing all these bottlenecks, you can lower the cost of an iteration from 5.00 to 3.50 cycles (1.43x speedup).

Workaround

Reduce the number of FP multiply/FMA instructions

FMA

Detected 8 FMA (fused multiply-add) operations.

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.

VCVTPS2PD: 2 occurrences►
- /home/eoseret/llm-attention/attention.cpp:31
- /home/eoseret/llm-attention/attention.cpp:31
VGATHERDPS: 1 occurrences►
- /home/eoseret/llm-attention/attention.cpp:31
VPMULLD: 1 occurrences►
- /home/eoseret/llm-attention/attention.cpp:31

Slow data structures access

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

Details

Constant unknown stride: 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)

Gather/scatter instructions

Detected gather/scatter instructions (typically caused by indirect accesses).

Details

VGATHERDPS: 1 occurrences►
- /home/eoseret/llm-attention/attention.cpp:31

Workaround

Try to simplify your code and/or replace indirect accesses with unit-stride ones.

Conversion instructions

Detected expensive conversion instructions.

Details

VCVTPS2PD (FP32 to FP64, SIMD): 2 occurrences►
- /home/eoseret/llm-attention/attention.cpp:31
- /home/eoseret/llm-attention/attention.cpp:31

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

1 AVX-512 instructions are processing arithmetic or math operations on double precision FP elements in vector mode (eight at a time).

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

The binary loop is composed of 16 FP arithmetical operations:

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

The binary loop is loading 64 bytes (8 double precision FP elements).

Arithmetic intensity

Arithmetic intensity is 0.25 FP operations per loaded or stored byte.

Unroll opportunity

Loop is potentially 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	15
nb uops	21
loop length	73
used x86 registers	6
used mmx registers	0
used xmm registers	2
used ymm registers	6
used zmm registers	3
nb stack references	0

Front-end

ASSUMED MACRO FUSION FIT IN UOP CACHE

micro-operation queue	3.50 cycles
front end	3.50 cycles

Back-end

	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11
uops	5.00	5.00	3.00	3.00	0.00	5.00	1.00	0.00	0.00	0.00	1.00	3.00
cycles	5.00	5.00	3.00	3.00	0.00	5.00	1.00	0.00	0.00	0.00	1.00	3.00

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)	4.00

Cycles summary

Front-end	3.50
Dispatch	5.00
Data deps.	4.00
Overall L1	5.00

Vectorization ratios

INT

all	83%
load	NA (no load vectorizable/vectorized instructions)
store	NA (no store vectorizable/vectorized instructions)
mul	100%
add-sub	100%
fma	NA (no fma vectorizable/vectorized instructions)
other	50%

FP

all	100%
load	100%
store	NA (no store vectorizable/vectorized instructions)
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	NA (no add-sub vectorizable/vectorized instructions)
fma	100%
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	100%

INT+FP

all	90%
load	100%
store	NA (no store vectorizable/vectorized instructions)
mul	100%
add-sub	100%
fma	100%
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	83%

Vector efficiency ratios

INT

all	38%
load	NA (no load vectorizable/vectorized instructions)
store	NA (no store vectorizable/vectorized instructions)
mul	50%
add-sub	50%
fma	NA (no fma vectorizable/vectorized instructions)
other	15%

FP

all	55%
load	50%
store	NA (no store vectorizable/vectorized instructions)
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	NA (no add-sub vectorizable/vectorized instructions)
fma	100%
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	43%

INT+FP

all	46%
load	50%
store	NA (no store vectorizable/vectorized instructions)
mul	50%
add-sub	50%
fma	100%
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	34%

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:

6% of peak load performance is reached (12.80 out of 192.00 bytes loaded per cycle (GB/s @ 1GHz))

Front-end bottlenecks

Found no such bottlenecks.

ASM code

In the binary file, the address of the loop is: 404ec0

Instruction	Nb FU	P0	P1	P2	P3	P5	P6	P10	P11	Latency	Recip. throughput	Vectorization
LEA (%R8,%R14,1),%R15D	1	0.20	0.20	0	0	0.20	0.20	0.20	0	1-2	0.20	N/A
VCVTPS2PD (%RBX,%R15,4),%ZMM7	2	1	0	0.33	0.33	1	0	0	0.33	7	1	vect (50.0%)
VPBROADCASTD %R14D,%YMM8	1	0	0	0	0	1	0	0	0	3	1	scal (6.3%)
VPADDD %YMM2,%YMM8,%YMM8	1	0.33	0.33	0	0	0.33	0	0	0	1	0.33	vect (50.0%)
VPMULLD %YMM8,%YMM10,%YMM8	2	1	1	0	0	0	0	0	0	10	1	vect (50.0%)
VPADDD %YMM6,%YMM8,%YMM8	1	0.33	0.33	0	0	0.33	0	0	0	1	0.33	vect (50.0%)
VPSUBD %YMM3,%YMM8,%YMM8	1	0.33	0.33	0	0	0.33	0	0	0	0-1	0.33	vect (50.0%)
VPCMPEQB %XMM0,%XMM0,%K1	1	0	0	0	0	1	0	0	0	3	1	vect (25.0%)
VXORPS %XMM9,%XMM9,%XMM9	1	0	0	0	0	0	0	0	0	0	0.17	vect (25.0%)
VGATHERDPS (%RDI,%YMM8,4),%YMM9{%K1}	5	1	1	2.67	2.67	1	0	0	2.67	29	3	vect (50.0%)
VCVTPS2PD %YMM9,%ZMM8	2	1	0	0	0	1	0	0	0	7	1	vect (50.0%)
VFMADD231PD %ZMM8,%ZMM7,%ZMM5	1	0.50	0	0	0	0.50	0	0	0	4	0.50	vect (100.0%)
ADD $0x8,%R14	1	0	0	0	0	0	0	0	0	1	0.17	N/A
CMP %RDX,%R14	1	0.20	0.20	0	0	0.20	0.20	0.20	0	1	0.20	N/A
JB 404ec0 <main+0x1c00>	1	0.50	0	0	0	0	0.50	0	0	0	0.50	N/A

Function	main
Source file and lines	attention.cpp:30-31
Module	attention-avx512

The loop is defined in /home/eoseret/llm-attention/attention.cpp:30-31.

The related source loop is multi-versionned.

gain
potential
hint
expert

Vectorization

Your loop is highly vectorized. Only 46% of vector register length is used (average across all SSE/AVX instructions). By fully vectorizing your loop, you can lower the cost of an iteration from 5.00 to 2.62 cycles (1.91x speedup).

Details

90% of SSE/AVX instructions are used in vector version (process two or more data elements in vector registers):

83% of SSE/AVX instructions that are not load, store, addition, subtraction nor multiply instructions are used in vector version.

Since your execution units are vector units, only a fully vectorized loop can use their full power.

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.

Execution units bottlenecks

Performance is limited by execution of FP multiply or FMA (fused multiply-add) operations (the FP multiply/FMA unit is a bottleneck). By removing all these bottlenecks, you can lower the cost of an iteration from 5.00 to 3.50 cycles (1.43x speedup).

Workaround

Reduce the number of FP multiply/FMA instructions

FMA

Detected 8 FMA (fused multiply-add) operations.

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.

VCVTPS2PD: 2 occurrences►
- /home/eoseret/llm-attention/attention.cpp:31
- /home/eoseret/llm-attention/attention.cpp:31
VGATHERDPS: 1 occurrences►
- /home/eoseret/llm-attention/attention.cpp:31
VPMULLD: 1 occurrences►
- /home/eoseret/llm-attention/attention.cpp:31

Slow data structures access

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

Details

Constant unknown stride: 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)

Gather/scatter instructions

Detected gather/scatter instructions (typically caused by indirect accesses).

Details

VGATHERDPS: 1 occurrences►
- /home/eoseret/llm-attention/attention.cpp:31

Workaround

Try to simplify your code and/or replace indirect accesses with unit-stride ones.

Conversion instructions

Detected expensive conversion instructions.

Details

VCVTPS2PD (FP32 to FP64, SIMD): 2 occurrences►
- /home/eoseret/llm-attention/attention.cpp:31
- /home/eoseret/llm-attention/attention.cpp:31

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

1 AVX-512 instructions are processing arithmetic or math operations on double precision FP elements in vector mode (eight at a time).

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

The binary loop is composed of 16 FP arithmetical operations:

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

The binary loop is loading 64 bytes (8 double precision FP elements).

Arithmetic intensity

Arithmetic intensity is 0.25 FP operations per loaded or stored byte.

Unroll opportunity

Loop is potentially 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	15
nb uops	21
loop length	73
used x86 registers	6
used mmx registers	0
used xmm registers	2
used ymm registers	6
used zmm registers	3
nb stack references	0

Front-end

ASSUMED MACRO FUSION FIT IN UOP CACHE

micro-operation queue	3.50 cycles
front end	3.50 cycles

Back-end

	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11
uops	5.00	5.00	3.00	3.00	0.00	5.00	1.00	0.00	0.00	0.00	1.00	3.00
cycles	5.00	5.00	3.00	3.00	0.00	5.00	1.00	0.00	0.00	0.00	1.00	3.00

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)	4.00

Cycles summary

Front-end	3.50
Dispatch	5.00
Data deps.	4.00
Overall L1	5.00

Vectorization ratios

INT

all	83%
load	NA (no load vectorizable/vectorized instructions)
store	NA (no store vectorizable/vectorized instructions)
mul	100%
add-sub	100%
fma	NA (no fma vectorizable/vectorized instructions)
other	50%

FP

all	100%
load	100%
store	NA (no store vectorizable/vectorized instructions)
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	NA (no add-sub vectorizable/vectorized instructions)
fma	100%
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	100%

INT+FP

all	90%
load	100%
store	NA (no store vectorizable/vectorized instructions)
mul	100%
add-sub	100%
fma	100%
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	83%

Vector efficiency ratios

INT

all	38%
load	NA (no load vectorizable/vectorized instructions)
store	NA (no store vectorizable/vectorized instructions)
mul	50%
add-sub	50%
fma	NA (no fma vectorizable/vectorized instructions)
other	15%

FP

all	55%
load	50%
store	NA (no store vectorizable/vectorized instructions)
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	NA (no add-sub vectorizable/vectorized instructions)
fma	100%
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	43%

INT+FP

all	46%
load	50%
store	NA (no store vectorizable/vectorized instructions)
mul	50%
add-sub	50%
fma	100%
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	34%

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:

6% of peak load performance is reached (12.80 out of 192.00 bytes loaded per cycle (GB/s @ 1GHz))

Front-end bottlenecks

Found no such bottlenecks.

ASM code

In the binary file, the address of the loop is: 404ec0

Instruction	Nb FU	P0	P1	P2	P3	P5	P6	P10	P11	Latency	Recip. throughput	Vectorization
LEA (%R8,%R14,1),%R15D	1	0.20	0.20	0	0	0.20	0.20	0.20	0	1-2	0.20	N/A
VCVTPS2PD (%RBX,%R15,4),%ZMM7	2	1	0	0.33	0.33	1	0	0	0.33	7	1	vect (50.0%)
VPBROADCASTD %R14D,%YMM8	1	0	0	0	0	1	0	0	0	3	1	scal (6.3%)
VPADDD %YMM2,%YMM8,%YMM8	1	0.33	0.33	0	0	0.33	0	0	0	1	0.33	vect (50.0%)
VPMULLD %YMM8,%YMM10,%YMM8	2	1	1	0	0	0	0	0	0	10	1	vect (50.0%)
VPADDD %YMM6,%YMM8,%YMM8	1	0.33	0.33	0	0	0.33	0	0	0	1	0.33	vect (50.0%)
VPSUBD %YMM3,%YMM8,%YMM8	1	0.33	0.33	0	0	0.33	0	0	0	0-1	0.33	vect (50.0%)
VPCMPEQB %XMM0,%XMM0,%K1	1	0	0	0	0	1	0	0	0	3	1	vect (25.0%)
VXORPS %XMM9,%XMM9,%XMM9	1	0	0	0	0	0	0	0	0	0	0.17	vect (25.0%)
VGATHERDPS (%RDI,%YMM8,4),%YMM9{%K1}	5	1	1	2.67	2.67	1	0	0	2.67	29	3	vect (50.0%)
VCVTPS2PD %YMM9,%ZMM8	2	1	0	0	0	1	0	0	0	7	1	vect (50.0%)
VFMADD231PD %ZMM8,%ZMM7,%ZMM5	1	0.50	0	0	0	0.50	0	0	0	4	0.50	vect (100.0%)
ADD $0x8,%R14	1	0	0	0	0	0	0	0	0	1	0.17	N/A
CMP %RDX,%R14	1	0.20	0.20	0	0	0.20	0.20	0.20	0	1	0.20	N/A
JB 404ec0 <main+0x1c00>	1	0.50	0	0	0	0	0.50	0	0	0	0.50	N/A

Report Configuration

Vectorization

Details

Workaround

Execution units bottlenecks

Workaround

FMA

Complex instructions

Details

Slow data structures access

Details

Workaround

Gather/scatter 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

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

FMA

Complex instructions

Details

Slow data structures access

Details

Workaround

Gather/scatter 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

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