OV - attention-gcc-skl256 - Loop 41

0x33d0 VPANDQ	0x8(%RAX),%YMM8,%YMM1    [2]

0x33da VMOVDQA	%YMM9,%YMM0

0x33de ADD	$0x20,%RAX

0x33e2 VPTERNLOGQ	$-0x14,-0x20(%RAX),%YMM1,%YMM0    [1]

0x33ea VPSRLQ	$0x1,%YMM0,%YMM1

0x33ef VPANDQ	%YMM7,%YMM0,%YMM0

0x33f5 VPMULLQ	%YMM6,%YMM0,%YMM0

0x33fb VPTERNLOGQ	$-0x6a,0xc48(%RAX),%YMM0,%YMM1    [1]

0x3406 VMOVDQU	%YMM1,-0x20(%RAX)    [1]

0x340b CMP	%RAX,%RDX

0x340e JNE	33d0

/usr/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	Source Location	Module
►100.00+	__libc_init_first		libc.so.6
○	__libc_start_main		libc.so.6
○	main	new_allocator.h:183	attention-gcc-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	1.00
CQA speedup if FP arith vectorized	1.00
CQA speedup if fully vectorized	2.00
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.31
Bottlenecks	micro-operation queue,
Function	std::mersenne_twister_engine::_M_gen_rand()
Source	random.tcc:404-409
Source loop unroll info	not unrolled or unrolled with no peel/tail loop
Source loop unroll confidence level	max
Unroll/vectorization loop type	NA
Unroll factor	NA
CQA cycles	3.50
CQA cycles if no scalar integer	3.50
CQA cycles if FP arith vectorized	3.50
CQA cycles if fully vectorized	1.75
Front-end cycles	3.50
P0 cycles	2.67
P1 cycles	2.67
P2 cycles	1.50
P3 cycles	1.50
P4 cycles	1.00
P5 cycles	2.67
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	11.00
Nb uops	14.00
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	36.57
Bytes prefetched	0.00
Bytes loaded	96.00
Bytes stored	32.00
Stride 0	0.00
Stride 1	1.00
Stride n	0.00
Stride unknown	0.00
Stride indirect	0.00
Vectorization ratio all	100.00
Vectorization ratio load	100.00
Vectorization ratio store	100.00
Vectorization ratio mul	100.00
Vectorization ratio add_sub	NA
Vectorization ratio fma	NA
Vectorization ratio div_sqrt	NA
Vectorization ratio other	100.00
Vector-efficiency ratio all	50.00
Vector-efficiency ratio load	50.00
Vector-efficiency ratio store	50.00
Vector-efficiency ratio mul	50.00
Vector-efficiency ratio add_sub	NA
Vector-efficiency ratio fma	NA
Vector-efficiency ratio div_sqrt	NA
Vector-efficiency ratio other	50.00

Metric	Value
CQA speedup if no scalar integer	1.00
CQA speedup if FP arith vectorized	1.00
CQA speedup if fully vectorized	2.00
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.31
Bottlenecks	micro-operation queue,
Function	std::mersenne_twister_engine::_M_gen_rand()
Source	random.tcc:404-409
Source loop unroll info	not unrolled or unrolled with no peel/tail loop
Source loop unroll confidence level	max
Unroll/vectorization loop type	NA
Unroll factor	NA
CQA cycles	3.50
CQA cycles if no scalar integer	3.50
CQA cycles if FP arith vectorized	3.50
CQA cycles if fully vectorized	1.75
Front-end cycles	3.50
P0 cycles	2.67
P1 cycles	2.67
P2 cycles	1.50
P3 cycles	1.50
P4 cycles	1.00
P5 cycles	2.67
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	11.00
Nb uops	14.00
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	36.57
Bytes prefetched	0.00
Bytes loaded	96.00
Bytes stored	32.00
Stride 0	0.00
Stride 1	1.00
Stride n	0.00
Stride unknown	0.00
Stride indirect	0.00
Vectorization ratio all	100.00
Vectorization ratio load	100.00
Vectorization ratio store	100.00
Vectorization ratio mul	100.00
Vectorization ratio add_sub	NA
Vectorization ratio fma	NA
Vectorization ratio div_sqrt	NA
Vectorization ratio other	100.00
Vector-efficiency ratio all	50.00
Vector-efficiency ratio load	50.00
Vector-efficiency ratio store	50.00
Vector-efficiency ratio mul	50.00
Vector-efficiency ratio add_sub	NA
Vector-efficiency ratio fma	NA
Vector-efficiency ratio div_sqrt	NA
Vector-efficiency ratio other	50.00

Path /

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

Function	std::mersenne_twister_engine::_M_gen_rand()
Source file and lines	random.tcc:404-409
Module	attention-gcc-skl256

The loop is defined in /usr/include/c++/16.1.1/bits/random.tcc:404-409.

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

gain
potential
hint
expert

Vectorization

Your loop is vectorized, but using only 256 out of 512 bits (AVX/AVX2 instructions on AVX-512 processors).

Details

All SSE/AVX instructions are used in vector version (process two or more data elements in vector registers).

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

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.

VPMULLQ: 1 occurrences►
- /usr/include/c++/16.1.1/bits/random.tcc:408
VPTERNLOGQ: 2 occurrences►
- /usr/include/c++/16.1.1/bits/random.tcc:406
- /usr/include/c++/16.1.1/bits/random.tcc:408

Type of elements and instruction set

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 96 bytes. The binary loop is storing 32 bytes.

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 by recompiling with -funroll-loops and/or -floop-unroll-and-jam. Or with the unroll (resp. unroll_and_jam) directive on top of the inner (resp. surrounding) loop. You can enforce an unroll factor: #pragma GCC unroll N

General properties

nb instructions	11
nb uops	14
loop length	64
used x86 registers	2
used mmx registers	0
used xmm registers	0
used ymm registers	6
used zmm registers	0
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
uops	2.67	2.67	1.50	1.50	1.00	2.67	2.00	1.00
cycles	2.67	2.67	1.50	1.50	1.00	2.67	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	3.50
Dispatch	2.67
Data deps.	1.00
Overall L1	3.50

Vectorization ratios

all	100%
load	100%
store	100%
mul	100%
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	100%

Vector efficiency ratios

all	50%
load	50%
store	50%
mul	50%
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%

Cycles and memory resources usage

Assuming all data fit into the L1 cache, each iteration of the binary loop takes 3.50 cycles. At this rate:

21% of peak load performance is reached (27.43 out of 128.00 bytes loaded per cycle (GB/s @ 1GHz))
14% of peak store performance is reached (9.14 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 3.50 to 2.67 cycles (1.31x speedup).

ASM code

In the binary file, the address of the loop is: 33d0

Instruction	Nb FU	P0	P1	P2	P3	P4	P5	P6	P7	Latency	Recip. throughput	Vectorization
VPANDQ 0x8(%RAX),%YMM8,%YMM1	1	0.33	0.33	0.50	0.50	0	0.33	0	0	1	0.50	vect (50.0%)
VMOVDQA %YMM9,%YMM0	1	0	0	0	0	0	0	0	0	0	0.25	vect (50.0%)
ADD $0x20,%RAX	1	0.25	0.25	0	0	0	0.25	0.25	0	1	0.25	N/A
VPTERNLOGQ $-0x14,-0x20(%RAX),%YMM1,%YMM0	2	0.33	0.33	0.50	0.50	0	0.33	0	0	1	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%)
VPANDQ %YMM7,%YMM0,%YMM0	1	0.33	0.33	0	0	0	0.33	0	0	1	0.33	vect (50.0%)
VPMULLQ %YMM6,%YMM0,%YMM0	3	1.50	1.50	0	0	0	0	0	0	15	1.50	vect (50.0%)
VPTERNLOGQ $-0x6a,0xc48(%RAX),%YMM0,%YMM1	2	0.33	0.33	0.50	0.50	0	0.33	0	0	1	0.50	vect (50.0%)
VMOVDQU %YMM1,-0x20(%RAX)	1	0	0	0.33	0.33	1	0	0	0.33	4	1	vect (50.0%)
CMP %RAX,%RDX	1	0.25	0.25	0	0	0	0.25	0.25	0	1	0.25	N/A
JNE 33d0 <_ZNSt23mersenne_twister_engineImLm32ELm624ELm397ELm31ELm2567483615ELm11ELm4294967295ELm7ELm2636928640ELm15ELm4022730752ELm18ELm1812433253EE11_M_gen_randEv+0x50>	1	0.50	0	0	0	0	0	0.50	0	0	0.50-1	N/A

Function	std::mersenne_twister_engine::_M_gen_rand()
Source file and lines	random.tcc:404-409
Module	attention-gcc-skl256

The loop is defined in /usr/include/c++/16.1.1/bits/random.tcc:404-409.

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

gain
potential
hint
expert

Vectorization

Your loop is vectorized, but using only 256 out of 512 bits (AVX/AVX2 instructions on AVX-512 processors).

Details

All SSE/AVX instructions are used in vector version (process two or more data elements in vector registers).

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

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.

VPMULLQ: 1 occurrences►
- /usr/include/c++/16.1.1/bits/random.tcc:408
VPTERNLOGQ: 2 occurrences►
- /usr/include/c++/16.1.1/bits/random.tcc:406
- /usr/include/c++/16.1.1/bits/random.tcc:408

Type of elements and instruction set

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 96 bytes. The binary loop is storing 32 bytes.

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 by recompiling with -funroll-loops and/or -floop-unroll-and-jam. Or with the unroll (resp. unroll_and_jam) directive on top of the inner (resp. surrounding) loop. You can enforce an unroll factor: #pragma GCC unroll N

General properties

nb instructions	11
nb uops	14
loop length	64
used x86 registers	2
used mmx registers	0
used xmm registers	0
used ymm registers	6
used zmm registers	0
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
uops	2.67	2.67	1.50	1.50	1.00	2.67	2.00	1.00
cycles	2.67	2.67	1.50	1.50	1.00	2.67	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	3.50
Dispatch	2.67
Data deps.	1.00
Overall L1	3.50

Vectorization ratios

all	100%
load	100%
store	100%
mul	100%
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	100%

Vector efficiency ratios

all	50%
load	50%
store	50%
mul	50%
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%

Cycles and memory resources usage

Assuming all data fit into the L1 cache, each iteration of the binary loop takes 3.50 cycles. At this rate:

21% of peak load performance is reached (27.43 out of 128.00 bytes loaded per cycle (GB/s @ 1GHz))
14% of peak store performance is reached (9.14 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 3.50 to 2.67 cycles (1.31x speedup).

ASM code

In the binary file, the address of the loop is: 33d0

Instruction	Nb FU	P0	P1	P2	P3	P4	P5	P6	P7	Latency	Recip. throughput	Vectorization
VPANDQ 0x8(%RAX),%YMM8,%YMM1	1	0.33	0.33	0.50	0.50	0	0.33	0	0	1	0.50	vect (50.0%)
VMOVDQA %YMM9,%YMM0	1	0	0	0	0	0	0	0	0	0	0.25	vect (50.0%)
ADD $0x20,%RAX	1	0.25	0.25	0	0	0	0.25	0.25	0	1	0.25	N/A
VPTERNLOGQ $-0x14,-0x20(%RAX),%YMM1,%YMM0	2	0.33	0.33	0.50	0.50	0	0.33	0	0	1	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%)
VPANDQ %YMM7,%YMM0,%YMM0	1	0.33	0.33	0	0	0	0.33	0	0	1	0.33	vect (50.0%)
VPMULLQ %YMM6,%YMM0,%YMM0	3	1.50	1.50	0	0	0	0	0	0	15	1.50	vect (50.0%)
VPTERNLOGQ $-0x6a,0xc48(%RAX),%YMM0,%YMM1	2	0.33	0.33	0.50	0.50	0	0.33	0	0	1	0.50	vect (50.0%)
VMOVDQU %YMM1,-0x20(%RAX)	1	0	0	0.33	0.33	1	0	0	0.33	4	1	vect (50.0%)
CMP %RAX,%RDX	1	0.25	0.25	0	0	0	0.25	0.25	0	1	0.25	N/A
JNE 33d0 <_ZNSt23mersenne_twister_engineImLm32ELm624ELm397ELm31ELm2567483615ELm11ELm4294967295ELm7ELm2636928640ELm15ELm4022730752ELm18ELm1812433253EE11_M_gen_randEv+0x50>	1	0.50	0	0	0	0	0	0.50	0	0	0.50-1	N/A

Report Configuration

Vectorization

Details

Workaround

Execution units bottlenecks

512-bits vectorization

Workaround

Complex instructions

Details

Type of elements and instruction set

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

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

512-bits vectorization

Workaround

Complex instructions

Details

Type of elements and instruction set

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

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