OV - attention-clang-gnr256 - Loop 59 - attention-clang-gnr256

	Loop Id: 59	Module: attention-clang-gnr256	Source: attention_v2.cpp:52-53	Coverage: 0.09%

0x5eb0 VMOVAPS	%XMM2,0x160(%RSP)    [1]

0x5eb9 VMOVUPS	(%R8,%R13,4),%XMM0    [2]

0x5ebf VSUBPS	0x100(%RSP),%XMM0,%XMM0    [1]

0x5ec8 VMOVAPS	%XMM0,0x60(%RSP)    [1]

0x5ece MOV	%R10,%RBX

0x5ed1 VZEROUPPER

0x5ed4 CALL	1160 <expf@plt>

0x5ed9 VMOVAPS	%XMM0,0xa0(%RSP)    [1]

0x5ee2 VMOVSHDUP	0x60(%RSP),%XMM0    [1]

0x5ee8 CALL	1160 <expf@plt>

0x5eed VMOVAPS	0xa0(%RSP),%XMM1    [1]

0x5ef6 VINSERTPS	$0x10,%XMM0,%XMM1,%XMM0

0x5efc VMOVAPS	%XMM0,0xa0(%RSP)    [1]

0x5f05 VPERMILPD	$0x1,0x60(%RSP),%XMM0    [1]

0x5f0d CALL	1160 <expf@plt>

0x5f12 VMOVAPS	0xa0(%RSP),%XMM1    [1]

0x5f1b VINSERTPS	$0x20,%XMM0,%XMM1,%XMM0

0x5f21 VMOVAPS	%XMM0,0xa0(%RSP)    [1]

0x5f2a VPSHUFD	$-0x1,0x60(%RSP),%XMM0    [1]

0x5f31 CALL	1160 <expf@plt>

0x5f36 VMOVAPS	0x160(%RSP),%XMM2    [1]

0x5f3f MOV	%RBX,%R10

0x5f42 MOV	0x20(%RSP),%R8    [1]

0x5f47 VMOVAPS	0xa0(%RSP),%XMM1    [1]

0x5f50 VINSERTPS	$0x30,%XMM0,%XMM1,%XMM0

0x5f56 VADDPS	%XMM2,%XMM0,%XMM2

0x5f5a ADD	$0x4,%R13

0x5f5e CMP	%R13,%RBX

0x5f61 JNE	5eb0

/home/eoseret/llm-attention/attention_v2.cpp: 52 - 53

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

52:       for (int idx = 0; idx <= row; ++idx) // vectorised

53:          sum += expf(S_row[idx] - max_val);

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	1.40
CQA speedup if FP arith vectorized	1.42
CQA speedup if fully vectorized	6.19
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.30
Bottlenecks	micro-operation queue,
Function	main
Source	attention_v2.cpp:52-53
Source loop unroll info	unrolled by 32
Source loop unroll confidence level	max
Unroll/vectorization loop type	intermediate
Unroll factor	4
CQA cycles	5.83
CQA cycles if no scalar integer	4.17
CQA cycles if FP arith vectorized	4.10
CQA cycles if fully vectorized	0.94
Front-end cycles	5.83
P0 cycles	0.50
P1 cycles	3.00
P2 cycles	3.33
P3 cycles	3.33
P4 cycles	4.50
P5 cycles	4.00
P6 cycles	0.50
P7 cycles	4.50
P8 cycles	4.50
P9 cycles	4.50
P10 cycles	0.00
P11 cycles	3.33
DIV/SQRT cycles	0.00
Inter-iter dependencies cycles	2
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	29.00
Nb uops	35.00
Nb loads	10.00
Nb stores	5.00
Nb stack references	5.00
FLOP/cycle	1.37
Nb FLOP add-sub	8.00
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	39.77
Bytes prefetched	0.00
Bytes loaded	152.00
Bytes stored	80.00
Stride 0	1.00
Stride 1	1.00
Stride n	0.00
Stride unknown	0.00
Stride indirect	0.00
Vectorization ratio all	80.00
Vectorization ratio load	100.00
Vectorization ratio store	100.00
Vectorization ratio mul	NA
Vectorization ratio add_sub	100.00
Vectorization ratio fma	NA
Vectorization ratio div_sqrt	NA
Vectorization ratio other	42.86
Vector-efficiency ratio all	20.94
Vector-efficiency ratio load	23.61
Vector-efficiency ratio store	25.00
Vector-efficiency ratio mul	NA
Vector-efficiency ratio add_sub	25.00
Vector-efficiency ratio fma	NA
Vector-efficiency ratio div_sqrt	NA
Vector-efficiency ratio other	15.18

Metric	Value
CQA speedup if no scalar integer	1.40
CQA speedup if FP arith vectorized	1.42
CQA speedup if fully vectorized	6.19
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.30
Bottlenecks	micro-operation queue,
Function	main
Source	attention_v2.cpp:52-53
Source loop unroll info	unrolled by 32
Source loop unroll confidence level	max
Unroll/vectorization loop type	intermediate
Unroll factor	4
CQA cycles	5.83
CQA cycles if no scalar integer	4.17
CQA cycles if FP arith vectorized	4.10
CQA cycles if fully vectorized	0.94
Front-end cycles	5.83
P0 cycles	0.50
P1 cycles	3.00
P2 cycles	3.33
P3 cycles	3.33
P4 cycles	4.50
P5 cycles	4.00
P6 cycles	0.50
P7 cycles	4.50
P8 cycles	4.50
P9 cycles	4.50
P10 cycles	0.00
P11 cycles	3.33
DIV/SQRT cycles	0.00
Inter-iter dependencies cycles	2
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	29.00
Nb uops	35.00
Nb loads	10.00
Nb stores	5.00
Nb stack references	5.00
FLOP/cycle	1.37
Nb FLOP add-sub	8.00
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	39.77
Bytes prefetched	0.00
Bytes loaded	152.00
Bytes stored	80.00
Stride 0	1.00
Stride 1	1.00
Stride n	0.00
Stride unknown	0.00
Stride indirect	0.00
Vectorization ratio all	80.00
Vectorization ratio load	100.00
Vectorization ratio store	100.00
Vectorization ratio mul	NA
Vectorization ratio add_sub	100.00
Vectorization ratio fma	NA
Vectorization ratio div_sqrt	NA
Vectorization ratio other	42.86
Vector-efficiency ratio all	20.94
Vector-efficiency ratio load	23.61
Vector-efficiency ratio store	25.00
Vector-efficiency ratio mul	NA
Vector-efficiency ratio add_sub	25.00
Vector-efficiency ratio fma	NA
Vector-efficiency ratio div_sqrt	NA
Vector-efficiency ratio other	15.18

Path /

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

Function	main
Source file and lines	attention_v2.cpp:52-53
Module	attention-clang-gnr256

The loop is defined in /home/eoseret/llm-attention/attention_v2.cpp:52-53.

It is intermediate loop of related source loop which is unrolled by 32 (including vectorization).

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 5.83 to 4.17 cycles (1.40x speedup).

Workaround

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

Unrolling/vectorization cost

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

Details

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

Workaround

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

Vectorization

Your loop is partially vectorized. Only 20% of vector register length is used (average across all SSE/AVX instructions).

Details

80% of SSE/AVX instructions are used in vector version (process two or more data elements in vector registers):

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

Execution units bottlenecks

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

No data for this section

CALL instructions

Detected function call instructions.

Details

Calling (and then returning from) a function prevents many compiler optimizations (like vectorization), breaks control flow (which reduces pipeline performance) and executes extra instructions to save/restore the registers used inside it, which is very expensive (dozens of cycles). Consider to inline small functions.

expf@plt: 4 occurrences►
- /home/eoseret/llm-attention/attention_v2.cpp:53
- /home/eoseret/llm-attention/attention_v2.cpp:53
- /home/eoseret/llm-attention/attention_v2.cpp:53
- /home/eoseret/llm-attention/attention_v2.cpp:53

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.

VPERMILPD: 1 occurrences►
- /home/eoseret/llm-attention/attention_v2.cpp:53
VPSHUFD: 1 occurrences►
- /home/eoseret/llm-attention/attention_v2.cpp:53
VZEROUPPER: 1 occurrences►
- /home/eoseret/llm-attention/attention_v2.cpp:53

Type of elements and instruction set

2 SSE or AVX instructions are processing arithmetic or math operations on single precision FP elements in vector mode (four at a time).

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

The binary loop is composed of 8 FP arithmetical operations:

8: addition or subtraction

The binary loop is loading 152 bytes (38 single precision FP elements). The binary loop is storing 80 bytes (20 single precision FP elements).

Arithmetic intensity

Arithmetic intensity is 0.03 FP operations per loaded or stored byte.

General properties

nb instructions	29
nb uops	35
loop length	183
used x86 registers	5
used mmx registers	0
used xmm registers	3
used ymm registers	0
used zmm registers	0
nb stack references	5

Front-end

ASSUMED MACRO FUSION FIT IN UOP CACHE

micro-operation queue	5.83 cycles
front end	5.83 cycles

Back-end

	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11
uops	0.50	3.00	3.33	3.33	4.50	4.00	0.50	4.50	4.50	4.50	0.00	3.33
cycles	0.50	3.00	3.33	3.33	4.50	4.00	0.50	4.50	4.50	4.50	0.00	3.33

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)	2.00

Cycles summary

Front-end	5.83
Dispatch	4.50
Data deps.	2.00
Overall L1	5.83

Vectorization ratios

INT

all	66%
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	NA (no fma vectorizable/vectorized instructions)
other	66%

all	82%
load	100%
store	100%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	100%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	25%

INT+FP

all	80%
load	100%
store	100%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	100%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	42%

Vector efficiency ratios

INT

all	20%
load	25%
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	20%

all	20%
load	23%
store	25%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	25%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	10%

INT+FP

all	20%
load	23%
store	25%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	25%
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 5.83 cycles. At this rate:

13% of peak load performance is reached (26.06 out of 192.00 bytes loaded per cycle (GB/s @ 1GHz))
10% of peak store performance is reached (13.71 out of 128.00 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 5.83 to 4.50 cycles (1.30x speedup).

ASM code

In the binary file, the address of the loop is: 5eb0

Instruction	Nb FU	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11	Latency	Recip. throughput	Vectorization
VMOVAPS %XMM2,0x160(%RSP)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	0-1	0.50	vect (25.0%)
VMOVUPS (%R8,%R13,4),%XMM0	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	0-1	0.33	vect (25.0%)
VSUBPS 0x100(%RSP),%XMM0,%XMM0	1	0	0.50	0.33	0.33	0	0.50	0	0	0	0	0	0.33	3	0.50	vect (25.0%)
VMOVAPS %XMM0,0x60(%RSP)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	0-1	0.50	vect (25.0%)
MOV %R10,%RBX	1	0	0	0	0	0	0	0	0	0	0	0	0	1	0.17	N/A
VZEROUPPER	2	0	0	0	0	0	0	0	0	0	0	0	0	0	1	vect (25.0%)
CALL 1160 <expf@plt>	2	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	0	1	N/A
VMOVAPS %XMM0,0xa0(%RSP)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	0-1	0.50	vect (25.0%)
VMOVSHDUP 0x60(%RSP),%XMM0	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	1	0.33	vect (12.5%)
CALL 1160 <expf@plt>	2	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	0	1	N/A
VMOVAPS 0xa0(%RSP),%XMM1	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	0-1	0.33	vect (25.0%)
VINSERTPS $0x10,%XMM0,%XMM1,%XMM0	1	0	0	0	0	0	1	0	0	0	0	0	0	1	1	scal (6.3%)
VMOVAPS %XMM0,0xa0(%RSP)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	0-1	0.50	vect (25.0%)
VPERMILPD $0x1,0x60(%RSP),%XMM0	2	0	0	0.33	0.33	0	1	0	0	0	0	0	0.33	1	1	vect (25.0%)
CALL 1160 <expf@plt>	2	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	0	1	N/A
VMOVAPS 0xa0(%RSP),%XMM1	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	0-1	0.33	vect (25.0%)
VINSERTPS $0x20,%XMM0,%XMM1,%XMM0	1	0	0	0	0	0	1	0	0	0	0	0	0	1	1	scal (6.3%)
VMOVAPS %XMM0,0xa0(%RSP)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	0-1	0.50	vect (25.0%)
VPSHUFD $-0x1,0x60(%RSP),%XMM0	2	0	0.50	0.33	0.33	0	0.50	0	0	0	0	0	0.33	1	0.50	vect (25.0%)
CALL 1160 <expf@plt>	2	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	0	1	N/A
VMOVAPS 0x160(%RSP),%XMM2	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	0-1	0.33	vect (25.0%)
MOV %RBX,%R10	1	0	0	0	0	0	0	0	0	0	0	0	0	1	0.17	scal (12.5%)
MOV 0x20(%RSP),%R8	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	1	0.33	N/A
VMOVAPS 0xa0(%RSP),%XMM1	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	0-1	0.33	vect (25.0%)
VINSERTPS $0x30,%XMM0,%XMM1,%XMM0	1	0	0	0	0	0	1	0	0	0	0	0	0	1	1	scal (6.3%)
VADDPS %XMM2,%XMM0,%XMM2	1	0	0.50	0	0	0	0.50	0	0	0	0	0	0	3	0.50	vect (25.0%)
ADD $0x4,%R13	1	0	0	0	0	0	0	0	0	0	0	0	0	1	0.17	N/A
CMP %R13,%RBX	1	0.20	0.20	0	0	0	0.20	0.20	0	0	0	0.20	0	1	0.20	N/A
JNE 5eb0 <main+0x31a0>	1	0.50	0	0	0	0	0	0.50	0	0	0	0	0	0	0.50	N/A

Function	main
Source file and lines	attention_v2.cpp:52-53
Module	attention-clang-gnr256

The loop is defined in /home/eoseret/llm-attention/attention_v2.cpp:52-53.

It is intermediate loop of related source loop which is unrolled by 32 (including vectorization).

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 5.83 to 4.17 cycles (1.40x speedup).

Workaround

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

Unrolling/vectorization cost

Details

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

Workaround

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

Vectorization

Your loop is partially vectorized. Only 20% of vector register length is used (average across all SSE/AVX instructions).

Details

80% of SSE/AVX instructions are used in vector version (process two or more data elements in vector registers):

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

Execution units bottlenecks

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

No data for this section

CALL instructions

Detected function call instructions.

Details

expf@plt: 4 occurrences►
- /home/eoseret/llm-attention/attention_v2.cpp:53
- /home/eoseret/llm-attention/attention_v2.cpp:53
- /home/eoseret/llm-attention/attention_v2.cpp:53
- /home/eoseret/llm-attention/attention_v2.cpp:53

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.

VPERMILPD: 1 occurrences►
- /home/eoseret/llm-attention/attention_v2.cpp:53
VPSHUFD: 1 occurrences►
- /home/eoseret/llm-attention/attention_v2.cpp:53
VZEROUPPER: 1 occurrences►
- /home/eoseret/llm-attention/attention_v2.cpp:53

Type of elements and instruction set

2 SSE or AVX instructions are processing arithmetic or math operations on single precision FP elements in vector mode (four at a time).

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

The binary loop is composed of 8 FP arithmetical operations:

8: addition or subtraction

The binary loop is loading 152 bytes (38 single precision FP elements). The binary loop is storing 80 bytes (20 single precision FP elements).

Arithmetic intensity

Arithmetic intensity is 0.03 FP operations per loaded or stored byte.

General properties

nb instructions	29
nb uops	35
loop length	183
used x86 registers	5
used mmx registers	0
used xmm registers	3
used ymm registers	0
used zmm registers	0
nb stack references	5

Front-end

ASSUMED MACRO FUSION FIT IN UOP CACHE

micro-operation queue	5.83 cycles
front end	5.83 cycles

Back-end

	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11
uops	0.50	3.00	3.33	3.33	4.50	4.00	0.50	4.50	4.50	4.50	0.00	3.33
cycles	0.50	3.00	3.33	3.33	4.50	4.00	0.50	4.50	4.50	4.50	0.00	3.33

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)	2.00

Cycles summary

Front-end	5.83
Dispatch	4.50
Data deps.	2.00
Overall L1	5.83

Vectorization ratios

INT

all	66%
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	NA (no fma vectorizable/vectorized instructions)
other	66%

all	82%
load	100%
store	100%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	100%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	25%

INT+FP

all	80%
load	100%
store	100%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	100%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	42%

Vector efficiency ratios

INT

all	20%
load	25%
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	20%

all	20%
load	23%
store	25%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	25%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	10%

INT+FP

all	20%
load	23%
store	25%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	25%
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 5.83 cycles. At this rate:

13% of peak load performance is reached (26.06 out of 192.00 bytes loaded per cycle (GB/s @ 1GHz))
10% of peak store performance is reached (13.71 out of 128.00 bytes stored per cycle (GB/s @ 1GHz))

Front-end bottlenecks

ASM code

In the binary file, the address of the loop is: 5eb0

Instruction	Nb FU	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11	Latency	Recip. throughput	Vectorization
VMOVAPS %XMM2,0x160(%RSP)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	0-1	0.50	vect (25.0%)
VMOVUPS (%R8,%R13,4),%XMM0	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	0-1	0.33	vect (25.0%)
VSUBPS 0x100(%RSP),%XMM0,%XMM0	1	0	0.50	0.33	0.33	0	0.50	0	0	0	0	0	0.33	3	0.50	vect (25.0%)
VMOVAPS %XMM0,0x60(%RSP)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	0-1	0.50	vect (25.0%)
MOV %R10,%RBX	1	0	0	0	0	0	0	0	0	0	0	0	0	1	0.17	N/A
VZEROUPPER	2	0	0	0	0	0	0	0	0	0	0	0	0	0	1	vect (25.0%)
CALL 1160 <expf@plt>	2	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	0	1	N/A
VMOVAPS %XMM0,0xa0(%RSP)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	0-1	0.50	vect (25.0%)
VMOVSHDUP 0x60(%RSP),%XMM0	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	1	0.33	vect (12.5%)
CALL 1160 <expf@plt>	2	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	0	1	N/A
VMOVAPS 0xa0(%RSP),%XMM1	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	0-1	0.33	vect (25.0%)
VINSERTPS $0x10,%XMM0,%XMM1,%XMM0	1	0	0	0	0	0	1	0	0	0	0	0	0	1	1	scal (6.3%)
VMOVAPS %XMM0,0xa0(%RSP)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	0-1	0.50	vect (25.0%)
VPERMILPD $0x1,0x60(%RSP),%XMM0	2	0	0	0.33	0.33	0	1	0	0	0	0	0	0.33	1	1	vect (25.0%)
CALL 1160 <expf@plt>	2	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	0	1	N/A
VMOVAPS 0xa0(%RSP),%XMM1	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	0-1	0.33	vect (25.0%)
VINSERTPS $0x20,%XMM0,%XMM1,%XMM0	1	0	0	0	0	0	1	0	0	0	0	0	0	1	1	scal (6.3%)
VMOVAPS %XMM0,0xa0(%RSP)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	0-1	0.50	vect (25.0%)
VPSHUFD $-0x1,0x60(%RSP),%XMM0	2	0	0.50	0.33	0.33	0	0.50	0	0	0	0	0	0.33	1	0.50	vect (25.0%)
CALL 1160 <expf@plt>	2	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	0	1	N/A
VMOVAPS 0x160(%RSP),%XMM2	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	0-1	0.33	vect (25.0%)
MOV %RBX,%R10	1	0	0	0	0	0	0	0	0	0	0	0	0	1	0.17	scal (12.5%)
MOV 0x20(%RSP),%R8	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	1	0.33	N/A
VMOVAPS 0xa0(%RSP),%XMM1	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	0-1	0.33	vect (25.0%)
VINSERTPS $0x30,%XMM0,%XMM1,%XMM0	1	0	0	0	0	0	1	0	0	0	0	0	0	1	1	scal (6.3%)
VADDPS %XMM2,%XMM0,%XMM2	1	0	0.50	0	0	0	0.50	0	0	0	0	0	0	3	0.50	vect (25.0%)
ADD $0x4,%R13	1	0	0	0	0	0	0	0	0	0	0	0	0	1	0.17	N/A
CMP %R13,%RBX	1	0.20	0.20	0	0	0	0.20	0.20	0	0	0	0.20	0	1	0.20	N/A
JNE 5eb0 <main+0x31a0>	1	0.50	0	0	0	0	0	0.50	0	0	0	0	0	0	0.50	N/A

Report Configuration

Code clean check

Workaround

Unrolling/vectorization cost

Details

Workaround

Vectorization

Details

Execution units bottlenecks

CALL instructions

Details

Complex instructions

Details

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

Code clean check

Workaround

Unrolling/vectorization cost

Details

Workaround

Vectorization

Details

Execution units bottlenecks

CALL instructions

Details

Complex instructions

Details

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