OV - attention-native - Loop 39

	Loop Id: 39	Module: attention-native	Source: attention.cpp:52-53	Coverage: 0.13%

0x12f44 MOVI	D22, #0

0x12f48 ORR	X22, XZR, XZR

0x12f4c B	13080

0x13074 HINT	#0

0x13078 HINT	#0

0x1307c HINT	#0

0x13080 LDR	S0, [X28, X22,LSL #2]    [1]

0x13084 STR	Q22, [SP, #320]    [2]

0x13088 FSUB	S0, S0, S9

0x1308c BL	101c0

0x13090 LDR	Q22, [SP, #320]    [2]

0x13094 LDR	Q9, [SP, #288]    [2]

0x13098 ADD	X22, X22, #1

0x1309c FADD	S22, S0, S22

0x130a0 CMP	X26, X22

0x130a4 B.NE	13080

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

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

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

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

Coverage (%)	Name	Module
►100.00+	__libc_start_call_main	libc.so.6
○	__libc_start_main	libc.so.6
○	_start	attention-native

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	1.08
CQA speedup if FP arith vectorized	2.60
CQA speedup if fully vectorized	1.08
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.08
Bottlenecks	micro-operation queue,
Function	main
Source	attention.cpp:52-53
Source loop unroll info	unrolled by 16
Source loop unroll confidence level	max
Unroll/vectorization loop type	NA
Unroll factor	1
CQA cycles	1.63
CQA cycles if no scalar integer	1.50
CQA cycles if FP arith vectorized	0.63
CQA cycles if fully vectorized	1.50
Front-end cycles	1.63
P0 cycles	1.50
P1 cycles	1.50
P2 cycles	0.67
P3 cycles	0.67
P4 cycles	0.67
P5 cycles	0.67
P6 cycles	0.67
P7 cycles	0.67
P8 cycles	1.00
P9 cycles	1.00
P10 cycles	1.00
P11 cycles	1.00
P12 cycles	1.50
P13 cycles	1.17
P14 cycles	1.33
P15 cycles	0.00
P16 cycles	0.00
DIV/SQRT cycles	0.00
Inter-iter dependencies cycles	NA
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	16.00
Nb uops	13.00
Nb loads	NA
Nb stores	1.00
Nb stack references	3.00
FLOP/cycle	1.23
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	1.85
Bytes prefetched	0.00
Bytes loaded	3.00
Bytes stored	0.00
Stride 0	NA
Stride 1	NA
Stride n	NA
Stride unknown	NA
Stride indirect	NA
Vectorization ratio all	42.86
Vectorization ratio load	66.67
Vectorization ratio store	100.00
Vectorization ratio mul	NA
Vectorization ratio add_sub	0.00
Vectorization ratio fma	NA
Vectorization ratio div_sqrt	NA
Vectorization ratio other	0.00
Vector-efficiency ratio all	60.71
Vector-efficiency ratio load	75.00
Vector-efficiency ratio store	100.00
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	50.00

Metric	Value
CQA speedup if no scalar integer	1.08
CQA speedup if FP arith vectorized	2.60
CQA speedup if fully vectorized	1.08
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.08
Bottlenecks	micro-operation queue,
Function	main
Source	attention.cpp:52-53
Source loop unroll info	unrolled by 16
Source loop unroll confidence level	max
Unroll/vectorization loop type	NA
Unroll factor	1
CQA cycles	1.63
CQA cycles if no scalar integer	1.50
CQA cycles if FP arith vectorized	0.63
CQA cycles if fully vectorized	1.50
Front-end cycles	1.63
P0 cycles	1.50
P1 cycles	1.50
P2 cycles	0.67
P3 cycles	0.67
P4 cycles	0.67
P5 cycles	0.67
P6 cycles	0.67
P7 cycles	0.67
P8 cycles	1.00
P9 cycles	1.00
P10 cycles	1.00
P11 cycles	1.00
P12 cycles	1.50
P13 cycles	1.17
P14 cycles	1.33
P15 cycles	0.00
P16 cycles	0.00
DIV/SQRT cycles	0.00
Inter-iter dependencies cycles	NA
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	16.00
Nb uops	13.00
Nb loads	NA
Nb stores	1.00
Nb stack references	3.00
FLOP/cycle	1.23
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	1.85
Bytes prefetched	0.00
Bytes loaded	3.00
Bytes stored	0.00
Stride 0	NA
Stride 1	NA
Stride n	NA
Stride unknown	NA
Stride indirect	NA
Vectorization ratio all	42.86
Vectorization ratio load	66.67
Vectorization ratio store	100.00
Vectorization ratio mul	NA
Vectorization ratio add_sub	0.00
Vectorization ratio fma	NA
Vectorization ratio div_sqrt	NA
Vectorization ratio other	0.00
Vector-efficiency ratio all	60.71
Vector-efficiency ratio load	75.00
Vector-efficiency ratio store	100.00
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	50.00

Path /

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

Function	main
Source file and lines	attention.cpp:52-53
Module	attention-native

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

The related source loop is unrolled by 16 (including vectorization).
Warnings:

Ignoring paths for analysis
Failed to get the number of paths
RecMII not computed since number of paths is unknown or > max_paths
Streams not analyzed since number of paths is unknown or > max_paths

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 1.63 to 1.50 cycles (1.08x 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 poorly vectorized. Only 60% of vector register length is used (average across all VPU instructions).

Details

42% of VPU instructions are used in vector version (process two or more data elements in vector registers):

66% of VPU loads are used in vector version.
0% of VPU addition or subtraction instructions are used in vector version.
0% of VPU instructions that are not load, store, addition, subtraction nor multiply instructions are used in vector version.

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.

unknown: 1 occurrences►
- /home/eoseret/llm-attention/attention.cpp:53

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 3 bytes.

Arithmetic intensity

Arithmetic intensity is 0.67 FP operations per loaded or stored byte.

General properties

nb instructions	16
nb uops	13
loop length	64
used w registers	0
used x registers	4
used b registers	0
used h registers	0
used s registers	4
used d registers	1
used q registers	2
used v registers	0
used z registers	0
nb stack references	3

Front-end

MACRO FUSION NOT POSSIBLE FIT IN UOP CACHE

micro-operation queue	1.63 cycles
front end	1.63 cycles

Back-end

	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11	P12	P13	P14	P15	P16
uops	1.50	1.50	0.67	0.67	0.67	0.67	0.67	0.67	1.00	1.00	1.00	1.00	1.50	1.17	1.33	0.00	0.00
cycles	1.50	1.50	0.67	0.67	0.67	0.67	0.67	0.67	1.00	1.00	1.00	1.00	1.50	1.17	1.33	0.00	0.00

Execution ports to units layout:

P0: BRU
P1: BRU
P2: ALU
P3: ALU
P4: ALU
P5: ALU
P6: ALU
P7: ALU
P8 (128 bits): VPU, FP store data, ALU, DIV/SQRT
P9 (128 bits): VPU, ALU, FP store data
P10 (128 bits): VPU, ALU, DIV/SQRT
P11 (128 bits): ALU, VPU
P12 (256 bits): store address, load
P13 (256 bits): store address, load
P14 (256 bits): load
P15 (64 bits): store data
P16 (64 bits): store data

Cycles executing div or sqrt instructions

Cycles summary

Front-end	1.63
Dispatch	1.50
Overall L1	1.63

Vectorization ratios

INT

all	60%
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)
other	0%

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

INT+FP

all	42%
load	66%
store	100%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	0%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	0%

Vector efficiency ratios

INT

all	75%
load	75%
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)
other	50%

all	25%
load	NA (no load vectorizable/vectorized instructions)
store	NA (no store vectorizable/vectorized instructions)
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	NA (no other vectorizable/vectorized instructions)

INT+FP

all	60%
load	75%
store	100%
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	50%

Cycles and memory resources usage

Assuming all data fit into the L1 cache, each iteration of the binary loop takes 1.63 cycles. At this rate:

1% of peak load performance is reached (1.85 out of 96.00 bytes loaded 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 1.63 to 1.50 cycles (1.08x speedup).

ASM code

In the binary file, the address of the loop is: 13074

Instruction	Nb FU	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11	P12	P13	P14	P15	P16	Latency	Recip. throughput	Vectorization
MOVI D22, #0	1	0	0	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	2	0.25	scal (50.0%)
ORR X22, XZR, XZR	1	0	0	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	1	0.17	N/A
B 13080 <main+0x1e60>	1	0.50	0.50	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	1	0.50	N/A
HINT #0																					N/A
HINT #0																					N/A
HINT #0																					N/A
LDR S0, [X28, X22,LSL #2]	1	0	0	0	0	0	0	0	0	0	0	0	0	0.33	0.33	0.33	0	0	6	0.33	scal (25.0%)
STR Q22, [SP, #320]	1	0	0	0	0	0	0	0	0	0.50	0.50	0	0	0.50	0.50	0	0	0	2	0.50	vect (100.0%)
FSUB S0, S0, S9	1	0	0	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	2	0.25	scal (25.0%)
BL 101c0 <@plt_start@+0x1a0>	1	0.50	0.50	0.50	0.50	0	0	0	0	0	0	0	0	0	0	0	0	0	1	0.50	N/A
LDR Q22, [SP, #320]	1	0	0	0	0	0	0	0	0	0	0	0	0	0.33	0.33	0.33	0	0	6	0.33	vect (100.0%)
LDR Q9, [SP, #288]	1	0	0	0	0	0	0	0	0	0	0	0	0	0.33	0.33	0.33	0	0	6	0.33	vect (100.0%)
ADD X22, X22, #1	1	0	0	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	1	0.17	N/A
FADD S22, S0, S22	1	0	0	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	2	0.25	scal (25.0%)
CMP X26, X22	1	0	0	0.25	0.25	0	0	0.25	0.25	0	0	0	0	0	0	0	0	0	1	0.33	N/A
B.NE 13080 <main+0x1e60>	1	0.50	0.50	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	1	0.50	N/A

Function	main
Source file and lines	attention.cpp:52-53
Module	attention-native

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

The related source loop is unrolled by 16 (including vectorization).
Warnings:

Ignoring paths for analysis
Failed to get the number of paths
RecMII not computed since number of paths is unknown or > max_paths
Streams not analyzed since number of paths is unknown or > max_paths

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 1.63 to 1.50 cycles (1.08x 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 poorly vectorized. Only 60% of vector register length is used (average across all VPU instructions).

Details

42% of VPU instructions are used in vector version (process two or more data elements in vector registers):

66% of VPU loads are used in vector version.
0% of VPU addition or subtraction instructions are used in vector version.
0% of VPU instructions that are not load, store, addition, subtraction nor multiply instructions are used in vector version.

No data for this section

CALL instructions

Detected function call instructions.

Details

unknown: 1 occurrences►
- /home/eoseret/llm-attention/attention.cpp:53

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 3 bytes.

Arithmetic intensity

Arithmetic intensity is 0.67 FP operations per loaded or stored byte.

General properties

nb instructions	16
nb uops	13
loop length	64
used w registers	0
used x registers	4
used b registers	0
used h registers	0
used s registers	4
used d registers	1
used q registers	2
used v registers	0
used z registers	0
nb stack references	3

Front-end

MACRO FUSION NOT POSSIBLE FIT IN UOP CACHE

micro-operation queue	1.63 cycles
front end	1.63 cycles

Back-end

	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11	P12	P13	P14	P15	P16
uops	1.50	1.50	0.67	0.67	0.67	0.67	0.67	0.67	1.00	1.00	1.00	1.00	1.50	1.17	1.33	0.00	0.00
cycles	1.50	1.50	0.67	0.67	0.67	0.67	0.67	0.67	1.00	1.00	1.00	1.00	1.50	1.17	1.33	0.00	0.00

Execution ports to units layout:

P0: BRU
P1: BRU
P2: ALU
P3: ALU
P4: ALU
P5: ALU
P6: ALU
P7: ALU
P8 (128 bits): VPU, FP store data, ALU, DIV/SQRT
P9 (128 bits): VPU, ALU, FP store data
P10 (128 bits): VPU, ALU, DIV/SQRT
P11 (128 bits): ALU, VPU
P12 (256 bits): store address, load
P13 (256 bits): store address, load
P14 (256 bits): load
P15 (64 bits): store data
P16 (64 bits): store data

Cycles executing div or sqrt instructions

Cycles summary

Front-end	1.63
Dispatch	1.50
Overall L1	1.63

Vectorization ratios

INT

all	60%
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)
other	0%

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

INT+FP

all	42%
load	66%
store	100%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	0%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	0%

Vector efficiency ratios

INT

all	75%
load	75%
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)
other	50%

all	25%
load	NA (no load vectorizable/vectorized instructions)
store	NA (no store vectorizable/vectorized instructions)
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	NA (no other vectorizable/vectorized instructions)

INT+FP

all	60%
load	75%
store	100%
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	50%

Cycles and memory resources usage

Assuming all data fit into the L1 cache, each iteration of the binary loop takes 1.63 cycles. At this rate:

1% of peak load performance is reached (1.85 out of 96.00 bytes loaded per cycle (GB/s @ 1GHz))

Front-end bottlenecks

ASM code

In the binary file, the address of the loop is: 13074

Instruction	Nb FU	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11	P12	P13	P14	P15	P16	Latency	Recip. throughput	Vectorization
MOVI D22, #0	1	0	0	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	2	0.25	scal (50.0%)
ORR X22, XZR, XZR	1	0	0	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	1	0.17	N/A
B 13080 <main+0x1e60>	1	0.50	0.50	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	1	0.50	N/A
HINT #0																					N/A
HINT #0																					N/A
HINT #0																					N/A
LDR S0, [X28, X22,LSL #2]	1	0	0	0	0	0	0	0	0	0	0	0	0	0.33	0.33	0.33	0	0	6	0.33	scal (25.0%)
STR Q22, [SP, #320]	1	0	0	0	0	0	0	0	0	0.50	0.50	0	0	0.50	0.50	0	0	0	2	0.50	vect (100.0%)
FSUB S0, S0, S9	1	0	0	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	2	0.25	scal (25.0%)
BL 101c0 <@plt_start@+0x1a0>	1	0.50	0.50	0.50	0.50	0	0	0	0	0	0	0	0	0	0	0	0	0	1	0.50	N/A
LDR Q22, [SP, #320]	1	0	0	0	0	0	0	0	0	0	0	0	0	0.33	0.33	0.33	0	0	6	0.33	vect (100.0%)
LDR Q9, [SP, #288]	1	0	0	0	0	0	0	0	0	0	0	0	0	0.33	0.33	0.33	0	0	6	0.33	vect (100.0%)
ADD X22, X22, #1	1	0	0	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	1	0.17	N/A
FADD S22, S0, S22	1	0	0	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	2	0.25	scal (25.0%)
CMP X26, X22	1	0	0	0.25	0.25	0	0	0.25	0.25	0	0	0	0	0	0	0	0	0	1	0.33	N/A
B.NE 13080 <main+0x1e60>	1	0.50	0.50	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	1	0.50	N/A

Report Configuration

Code clean check

Workaround

Unrolling/vectorization cost

Details

Workaround

Vectorization

Details

CALL instructions

Details

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

CALL instructions

Details

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