OV - attention-icx-skl256 - Loop 51

0x404470 VMOVDQU	0x2e0(%RSP,%RDX,1),%YMM0    [1]

0x404479 VPSRLQ	$0x1,%YMM0,%YMM1

0x40447e VPAND	%YMM3,%YMM1,%YMM1

0x404482 VPBROADCASTQ	%RAX,%YMM2

0x404488 VPBLENDD	$-0x40,%YMM2,%YMM0,%YMM2

0x40448e VPSRLQ	$0x1,%YMM2,%YMM2

0x404493 VPERMQ	$-0x6d,%YMM2,%YMM2

0x404499 VPAND	%YMM4,%YMM2,%YMM2

0x40449d VPTERNLOGQ	$0x56,0xf40(%RSP,%RDX,1),%YMM1,%YMM2    [1]

0x4044a6 VPTESTMQ	%YMM5,%YMM0,%K1

0x4044ac VPXORQ	%YMM6,%YMM2,%YMM2{%K1}

0x4044b2 VMOVDQU	%YMM2,0x2d8(%RSP,%RDX,1)    [1]

0x4044bb MOV	0x2f8(%RSP,%RDX,1),%RAX    [1]

0x4044c3 ADD	$0x20,%RDX

0x4044c7 CMP	$0x700,%RDX

0x4044ce JNE	404470

/usr/lib64/gcc/x86_64-pc-linux-gnu/16.1.1/../../../../include/c++/16.1.1/bits/random.tcc: 404 - 409

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

404:       for (size_t __k = 0; __k < (__n - __m); ++__k)

405:         {

406: 	  _UIntType __y = ((_M_x[__k] & __upper_mask)

407: 			   | (_M_x[__k + 1] & __lower_mask));

408: 	  _M_x[__k] = (_M_x[__k + __m] ^ (__y >> 1)

409: 		       ^ ((__y & 0x01) ? __a : 0));

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	1.07
CQA speedup if FP arith vectorized	1.00
CQA speedup if fully vectorized	2.74
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.20
Bottlenecks	micro-operation queue,
Function	main
Source	random.tcc:404-409
Source loop unroll info	multi-versionned
Source loop unroll confidence level	max
Unroll/vectorization loop type	NA
Unroll factor	NA
CQA cycles	4.00
CQA cycles if no scalar integer	3.75
CQA cycles if FP arith vectorized	4.00
CQA cycles if fully vectorized	1.46
Front-end cycles	4.00
P0 cycles	3.33
P1 cycles	3.33
P2 cycles	1.50
P3 cycles	1.50
P4 cycles	1.00
P5 cycles	3.33
P6 cycles	2.00
P7 cycles	1.00
DIV/SQRT cycles	0.00
Inter-iter dependencies cycles	1
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	16.00
Nb uops	16.00
Nb loads	3.00
Nb stores	1.00
Nb stack references	0.00
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	26.00
Bytes prefetched	0.00
Bytes loaded	72.00
Bytes stored	32.00
Stride 0	0.00
Stride 1	0.00
Stride n	0.00
Stride unknown	0.00
Stride indirect	0.00
Vectorization ratio all	84.62
Vectorization ratio load	66.67
Vectorization ratio store	100.00
Vectorization ratio mul	NA
Vectorization ratio add_sub	NA
Vectorization ratio fma	NA
Vectorization ratio div_sqrt	NA
Vectorization ratio other	90.00
Vector-efficiency ratio all	44.23
Vector-efficiency ratio load	37.50
Vector-efficiency ratio store	50.00
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	46.25

Metric	Value
CQA speedup if no scalar integer	1.07
CQA speedup if FP arith vectorized	1.00
CQA speedup if fully vectorized	2.74
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.20
Bottlenecks	micro-operation queue,
Function	main
Source	random.tcc:404-409
Source loop unroll info	multi-versionned
Source loop unroll confidence level	max
Unroll/vectorization loop type	NA
Unroll factor	NA
CQA cycles	4.00
CQA cycles if no scalar integer	3.75
CQA cycles if FP arith vectorized	4.00
CQA cycles if fully vectorized	1.46
Front-end cycles	4.00
P0 cycles	3.33
P1 cycles	3.33
P2 cycles	1.50
P3 cycles	1.50
P4 cycles	1.00
P5 cycles	3.33
P6 cycles	2.00
P7 cycles	1.00
DIV/SQRT cycles	0.00
Inter-iter dependencies cycles	1
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	16.00
Nb uops	16.00
Nb loads	3.00
Nb stores	1.00
Nb stack references	0.00
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	26.00
Bytes prefetched	0.00
Bytes loaded	72.00
Bytes stored	32.00
Stride 0	0.00
Stride 1	0.00
Stride n	0.00
Stride unknown	0.00
Stride indirect	0.00
Vectorization ratio all	84.62
Vectorization ratio load	66.67
Vectorization ratio store	100.00
Vectorization ratio mul	NA
Vectorization ratio add_sub	NA
Vectorization ratio fma	NA
Vectorization ratio div_sqrt	NA
Vectorization ratio other	90.00
Vector-efficiency ratio all	44.23
Vector-efficiency ratio load	37.50
Vector-efficiency ratio store	50.00
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	46.25

Path /

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

Function	main
Source file and lines	random.tcc:404-409
Module	attention-icx-skl256

The loop is defined in /usr/lib64/gcc/x86_64-pc-linux-gnu/16.1.1/../../../../include/c++/16.1.1/bits/random.tcc:404-409.

The related source loop is multi-versionned.

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 4.00 to 3.75 cycles (1.07x speedup).

Workaround

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

Vectorization

Your loop is partially vectorized. Only 44% of vector register length is used (average across all SSE/AVX instructions).

Details

84% of SSE/AVX instructions are used in vector version (process two or more data elements in vector registers):

66% of SSE/AVX loads are used in vector version.
90% 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

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

Masked instructions

Detected masked instructions.

Details

Vector registers are partially exploited, which is expected if your loop is irregular or mixes elements of different sizes.

Workaround

If your loop is irregular, try to remove or hoist conditional structures out of your loop. If it mixes elements of different sizes, try to uniformize them.

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.

VPTERNLOGQ: 1 occurrences►
- /usr/lib64/gcc/x86_64-pc-linux-gnu/16.1.1/../../../../include/c++/16.1.1/bits/random.tcc:408

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 72 bytes. The binary loop is storing 32 bytes.

General properties

nb instructions	16
nb uops	16
loop length	96
used x86 registers	3
used mmx registers	0
used xmm registers	0
used ymm registers	7
used zmm registers	0
nb stack references	0

Front-end

ASSUMED MACRO FUSION FIT IN UOP CACHE

micro-operation queue	4.00 cycles
front end	4.00 cycles

Back-end

	P0	P1	P2	P3	P4	P5	P6	P7
uops	3.33	3.33	1.50	1.50	1.00	3.33	2.00	1.00
cycles	3.33	3.33	1.50	1.50	1.00	3.33	2.00	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)	1.00

Cycles summary

Front-end	4.00
Dispatch	3.33
Data deps.	1.00
Overall L1	4.00

Vectorization ratios

all	84%
load	66%
store	100%
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	90%

Vector efficiency ratios

all	44%
load	37%
store	50%
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	46%

Cycles and memory resources usage

Detected masked instructions: assuming all mask elements are active. Assuming all data fit into the L1 cache, each iteration of the binary loop takes 4.00 cycles. At this rate:

14% of peak load performance is reached (18.00 out of 128.00 bytes loaded per cycle (GB/s @ 1GHz))
12% of peak store performance is reached (8.00 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 4.00 to 3.33 cycles (1.20x speedup).

ASM code

In the binary file, the address of the loop is: 404470

Instruction	Nb FU	P0	P1	P2	P3	P4	P5	P6	P7	Latency	Recip. throughput	Vectorization
VMOVDQU 0x2e0(%RSP,%RDX,1),%YMM0	1	0	0	0.50	0.50	0	0	0	0	5	0.50	vect (50.0%)
VPSRLQ $0x1,%YMM0,%YMM1	1	0.50	0.50	0	0	0	0	0	0	1	0.50	vect (50.0%)
VPAND %YMM3,%YMM1,%YMM1	1	0.33	0.33	0	0	0	0.33	0	0	1	0.33	vect (50.0%)
VPBROADCASTQ %RAX,%YMM2	1	0	0	0	0	0	1	0	0	1	1	scal (12.5%)
VPBLENDD $-0x40,%YMM2,%YMM0,%YMM2	1	0.33	0.33	0	0	0	0.33	0	0	1	0.33	vect (50.0%)
VPSRLQ $0x1,%YMM2,%YMM2	1	0.50	0.50	0	0	0	0	0	0	1	0.50	vect (50.0%)
VPERMQ $-0x6d,%YMM2,%YMM2	1	0	0	0	0	0	1	0	0	3	1	vect (50.0%)
VPAND %YMM4,%YMM2,%YMM2	1	0.33	0.33	0	0	0	0.33	0	0	1	0.33	vect (50.0%)
VPTERNLOGQ $0x56,0xf40(%RSP,%RDX,1),%YMM1,%YMM2	2	0.33	0.33	0.50	0.50	0	0.33	0	0	1	0.50	vect (50.0%)
VPTESTMQ %YMM5,%YMM0,%K1	1	0	0	0	0	0	1	0	0	3	1	vect (50.0%)
VPXORQ %YMM6,%YMM2,%YMM2{%K1}	1	0.33	0.33	0	0	0	0.33	0	0	1	0.33	vect (50.0%)
VMOVDQU %YMM2,0x2d8(%RSP,%RDX,1)	1	0	0	0.33	0.33	1	0	0	0.33	4	1	vect (50.0%)
MOV 0x2f8(%RSP,%RDX,1),%RAX	1	0	0	0.50	0.50	0	0	0	0	4-5	0.50	scal (12.5%)
ADD $0x20,%RDX	1	0.25	0.25	0	0	0	0.25	0.25	0	1	0.25	N/A
CMP $0x700,%RDX	1	0.25	0.25	0	0	0	0.25	0.25	0	1	0.25	N/A
JNE 404470 <main+0x1350>	1	0.50	0	0	0	0	0	0.50	0	0	0.50-1	N/A

Function	main
Source file and lines	random.tcc:404-409
Module	attention-icx-skl256

The loop is defined in /usr/lib64/gcc/x86_64-pc-linux-gnu/16.1.1/../../../../include/c++/16.1.1/bits/random.tcc:404-409.

The related source loop is multi-versionned.

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 4.00 to 3.75 cycles (1.07x speedup).

Workaround

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

Vectorization

Your loop is partially vectorized. Only 44% of vector register length is used (average across all SSE/AVX instructions).

Details

84% of SSE/AVX instructions are used in vector version (process two or more data elements in vector registers):

66% of SSE/AVX loads are used in vector version.
90% 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

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

Masked instructions

Detected masked instructions.

Details

Vector registers are partially exploited, which is expected if your loop is irregular or mixes elements of different sizes.

Workaround

If your loop is irregular, try to remove or hoist conditional structures out of your loop. If it mixes elements of different sizes, try to uniformize them.

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.

VPTERNLOGQ: 1 occurrences►
- /usr/lib64/gcc/x86_64-pc-linux-gnu/16.1.1/../../../../include/c++/16.1.1/bits/random.tcc:408

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 72 bytes. The binary loop is storing 32 bytes.

General properties

nb instructions	16
nb uops	16
loop length	96
used x86 registers	3
used mmx registers	0
used xmm registers	0
used ymm registers	7
used zmm registers	0
nb stack references	0

Front-end

ASSUMED MACRO FUSION FIT IN UOP CACHE

micro-operation queue	4.00 cycles
front end	4.00 cycles

Back-end

	P0	P1	P2	P3	P4	P5	P6	P7
uops	3.33	3.33	1.50	1.50	1.00	3.33	2.00	1.00
cycles	3.33	3.33	1.50	1.50	1.00	3.33	2.00	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)	1.00

Cycles summary

Front-end	4.00
Dispatch	3.33
Data deps.	1.00
Overall L1	4.00

Vectorization ratios

all	84%
load	66%
store	100%
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	90%

Vector efficiency ratios

all	44%
load	37%
store	50%
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	46%

Cycles and memory resources usage

Detected masked instructions: assuming all mask elements are active. Assuming all data fit into the L1 cache, each iteration of the binary loop takes 4.00 cycles. At this rate:

14% of peak load performance is reached (18.00 out of 128.00 bytes loaded per cycle (GB/s @ 1GHz))
12% of peak store performance is reached (8.00 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 4.00 to 3.33 cycles (1.20x speedup).

ASM code

In the binary file, the address of the loop is: 404470

Instruction	Nb FU	P0	P1	P2	P3	P4	P5	P6	P7	Latency	Recip. throughput	Vectorization
VMOVDQU 0x2e0(%RSP,%RDX,1),%YMM0	1	0	0	0.50	0.50	0	0	0	0	5	0.50	vect (50.0%)
VPSRLQ $0x1,%YMM0,%YMM1	1	0.50	0.50	0	0	0	0	0	0	1	0.50	vect (50.0%)
VPAND %YMM3,%YMM1,%YMM1	1	0.33	0.33	0	0	0	0.33	0	0	1	0.33	vect (50.0%)
VPBROADCASTQ %RAX,%YMM2	1	0	0	0	0	0	1	0	0	1	1	scal (12.5%)
VPBLENDD $-0x40,%YMM2,%YMM0,%YMM2	1	0.33	0.33	0	0	0	0.33	0	0	1	0.33	vect (50.0%)
VPSRLQ $0x1,%YMM2,%YMM2	1	0.50	0.50	0	0	0	0	0	0	1	0.50	vect (50.0%)
VPERMQ $-0x6d,%YMM2,%YMM2	1	0	0	0	0	0	1	0	0	3	1	vect (50.0%)
VPAND %YMM4,%YMM2,%YMM2	1	0.33	0.33	0	0	0	0.33	0	0	1	0.33	vect (50.0%)
VPTERNLOGQ $0x56,0xf40(%RSP,%RDX,1),%YMM1,%YMM2	2	0.33	0.33	0.50	0.50	0	0.33	0	0	1	0.50	vect (50.0%)
VPTESTMQ %YMM5,%YMM0,%K1	1	0	0	0	0	0	1	0	0	3	1	vect (50.0%)
VPXORQ %YMM6,%YMM2,%YMM2{%K1}	1	0.33	0.33	0	0	0	0.33	0	0	1	0.33	vect (50.0%)
VMOVDQU %YMM2,0x2d8(%RSP,%RDX,1)	1	0	0	0.33	0.33	1	0	0	0.33	4	1	vect (50.0%)
MOV 0x2f8(%RSP,%RDX,1),%RAX	1	0	0	0.50	0.50	0	0	0	0	4-5	0.50	scal (12.5%)
ADD $0x20,%RDX	1	0.25	0.25	0	0	0	0.25	0.25	0	1	0.25	N/A
CMP $0x700,%RDX	1	0.25	0.25	0	0	0	0.25	0.25	0	1	0.25	N/A
JNE 404470 <main+0x1350>	1	0.50	0	0	0	0	0	0.50	0	0	0.50-1	N/A

Report Configuration

Code clean check

Workaround

Vectorization

Details

Workaround

Execution units bottlenecks

512-bits vectorization

Workaround

Masked instructions

Details

Workaround

Complex instructions

Details

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

Masked instructions

Details

Workaround

Complex instructions

Details

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