OV - attention-native - Loop 57

	Loop Id: 57	Module: attention-native	Source: attention.cpp:30-31	Coverage: 7.40%

0x125c0 SUB	W17, W16, #1

0x125c4 LDR	S3, [X22, W15,UXTW #2]    [1]

0x125c8 ADD	X14, X14, #2

0x125cc LDR	S2, [X20, W17,UXTW #2]    [2]

0x125d0 ADD	W17, W25, W15

0x125d4 ADD	X15, X15, X19

0x125d8 CMP	X9, X14

0x125dc FCVT	D3, S3

0x125e0 FCVT	D2, S2

0x125e4 FMADD	D1, D2, D3, D1

0x125e8 LDR	S2, [X20, W16,UXTW #2]    [2]

0x125ec LDR	S3, [X22, W17,UXTW #2]    [1]

0x125f0 ADD	W16, W16, #2

0x125f4 FCVT	D2, S2

0x125f8 FCVT	D3, S3

0x125fc FMADD	D1, D2, D3, D1

0x12600 B.NE	125c0

/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	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	1.00
CQA speedup if FP arith vectorized	1.45
CQA speedup if fully vectorized	6.48
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.88
Bottlenecks
Function	main
Source	attention.cpp:30-31
Source loop unroll info	unrolled by 4
Source loop unroll confidence level	medium
Unroll/vectorization loop type	NA
Unroll factor	1
CQA cycles	4.00
CQA cycles if no scalar integer	4.00
CQA cycles if FP arith vectorized	2.76
CQA cycles if fully vectorized	0.62
Front-end cycles	2.13
P0 cycles	0.50
P1 cycles	0.50
P2 cycles	1.50
P3 cycles	1.50
P4 cycles	1.50
P5 cycles	1.50
P6 cycles	2.00
P7 cycles	1.00
P8 cycles	2.00
P9 cycles	1.00
P10 cycles	1.33
P11 cycles	1.33
P12 cycles	1.33
P13 cycles	0.00
P14 cycles	0.00
DIV/SQRT cycles	0.00
Inter-iter dependencies cycles	4
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	17.00
Nb uops	17.00
Nb loads	NA
Nb stores	0.00
Nb stack references	0.00
FLOP/cycle	1.00
Nb FLOP add-sub	0.00
Nb FLOP mul	0.00
Nb FLOP fma	2.00
Nb FLOP div	0.00
Nb FLOP rcp	0.00
Nb FLOP sqrt	0.00
Nb FLOP rsqrt	0.00
Bytes/cycle	3.00
Bytes prefetched	0.00
Bytes loaded	12.00
Bytes stored	0.00
Stride 0	0.00
Stride 1	0.00
Stride n	0.00
Stride unknown	2.00
Stride indirect	0.00
Vectorization ratio all	0.00
Vectorization ratio load	0.00
Vectorization ratio store	NA
Vectorization ratio mul	NA
Vectorization ratio add_sub	0.00
Vectorization ratio fma	0.00
Vectorization ratio div_sqrt	NA
Vectorization ratio other	0.00
Vector-efficiency ratio all	16.67
Vector-efficiency ratio load	12.50
Vector-efficiency ratio store	NA
Vector-efficiency ratio mul	NA
Vector-efficiency ratio add_sub	25.00
Vector-efficiency ratio fma	25.00
Vector-efficiency ratio div_sqrt	NA
Vector-efficiency ratio other	15.00

Metric	Value
CQA speedup if no scalar integer	1.00
CQA speedup if FP arith vectorized	1.45
CQA speedup if fully vectorized	6.48
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.88
Bottlenecks
Function	main
Source	attention.cpp:30-31
Source loop unroll info	unrolled by 4
Source loop unroll confidence level	medium
Unroll/vectorization loop type	NA
Unroll factor	1
CQA cycles	4.00
CQA cycles if no scalar integer	4.00
CQA cycles if FP arith vectorized	2.76
CQA cycles if fully vectorized	0.62
Front-end cycles	2.13
P0 cycles	0.50
P1 cycles	0.50
P2 cycles	1.50
P3 cycles	1.50
P4 cycles	1.50
P5 cycles	1.50
P6 cycles	2.00
P7 cycles	1.00
P8 cycles	2.00
P9 cycles	1.00
P10 cycles	1.33
P11 cycles	1.33
P12 cycles	1.33
P13 cycles	0.00
P14 cycles	0.00
DIV/SQRT cycles	0.00
Inter-iter dependencies cycles	4
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	17.00
Nb uops	17.00
Nb loads	NA
Nb stores	0.00
Nb stack references	0.00
FLOP/cycle	1.00
Nb FLOP add-sub	0.00
Nb FLOP mul	0.00
Nb FLOP fma	2.00
Nb FLOP div	0.00
Nb FLOP rcp	0.00
Nb FLOP sqrt	0.00
Nb FLOP rsqrt	0.00
Bytes/cycle	3.00
Bytes prefetched	0.00
Bytes loaded	12.00
Bytes stored	0.00
Stride 0	0.00
Stride 1	0.00
Stride n	0.00
Stride unknown	2.00
Stride indirect	0.00
Vectorization ratio all	0.00
Vectorization ratio load	0.00
Vectorization ratio store	NA
Vectorization ratio mul	NA
Vectorization ratio add_sub	0.00
Vectorization ratio fma	0.00
Vectorization ratio div_sqrt	NA
Vectorization ratio other	0.00
Vector-efficiency ratio all	16.67
Vector-efficiency ratio load	12.50
Vector-efficiency ratio store	NA
Vector-efficiency ratio mul	NA
Vector-efficiency ratio add_sub	25.00
Vector-efficiency ratio fma	25.00
Vector-efficiency ratio div_sqrt	NA
Vector-efficiency ratio other	15.00

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-native

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

The related source loop is unrolled by 4 (including vectorization).

gain
potential
hint
expert

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 not vectorized. Only 16% of vector register length is used (average across all VPU instructions).

Details

All VPU instructions are used in scalar version (process only one data element in vector registers).

FMA

Detected 2 FMA (fused multiply-add) operations.

Slow data structures access

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

Details

Constant unknown stride: 2 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)

Conversion instructions

Detected expensive conversion instructions.

Details

FCVT: 4 occurrences►
- /home/eoseret/llm-attention/attention.cpp:31
- /home/eoseret/llm-attention/attention.cpp:31
- /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.

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

The binary loop is composed of 4 FP arithmetical operations:

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

The binary loop is loading 12 bytes.

Arithmetic intensity

Arithmetic intensity is 0.33 FP operations per loaded or stored byte.

General properties

nb instructions	17
nb uops	17
loop length	68
used w registers	4
used x registers	6
used b registers	0
used h registers	0
used s registers	2
used d registers	3
used q registers	0
used v registers	0
used z registers	0
nb stack references	0

Front-end

MACRO FUSION NOT POSSIBLE FIT IN UOP CACHE

micro-operation queue	2.13 cycles
front end	2.13 cycles

Back-end

	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11	P12	P13	P14
uops	0.50	0.50	1.50	1.50	1.50	1.50	2.00	1.00	2.00	1.00	1.33	1.33	1.33	0.00	0.00
cycles	0.50	0.50	1.50	1.50	1.50	1.50	2.00	1.00	2.00	1.00	1.33	1.33	1.33	0.00	0.00

Execution ports to units layout:

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

Cycles executing div or sqrt instructions	NA
Longest recurrence chain latency (RecMII)	4.00

Cycles summary

Front-end	2.13
Dispatch	2.00
Data deps.	4.00
Overall L1	4.00

Vectorization ratios

INT

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

INT+FP

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

Vector efficiency ratios

INT

all	16%
load	12%
store	NA (no store vectorizable/vectorized instructions)
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	25%
fma	NA (no fma vectorizable/vectorized instructions)
other	25%

all	16%
load	NA (no load vectorizable/vectorized instructions)
store	NA (no store vectorizable/vectorized instructions)
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	NA (no add-sub vectorizable/vectorized instructions)
fma	25%
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	12%

INT+FP

all	16%
load	12%
store	NA (no store vectorizable/vectorized instructions)
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	25%
fma	25%
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 4.00 cycles. At this rate:

3% of peak load performance is reached (3.00 out of 96.00 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: 125c0

Instruction	Nb FU	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11	P12	Latency	Recip. throughput	Vectorization
SUB W17, W16, #1	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	1	0.25	N/A
LDR S3, [X22, W15,UXTW #2]	1	0	0	0	0	0	0	0	0	0	0	0.33	0.33	0.33	6	0.33	scal (12.5%)
ADD X14, X14, #2	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	1	0.25	scal (25.0%)
LDR S2, [X20, W17,UXTW #2]	1	0	0	0	0	0	0	0	0	0	0	0.33	0.33	0.33	6	0.33	scal (12.5%)
ADD W17, W25, W15	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	1	0.25	N/A
ADD X15, X15, X19	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	1	0.25	N/A
CMP X9, X14	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	1	0.33	scal (25.0%)
FCVT D3, S3	1	0	0	0	0	0	0	0.50	0	0.50	0	0	0	0	3	0.50	scal (12.5%)
FCVT D2, S2	1	0	0	0	0	0	0	0.50	0	0.50	0	0	0	0	3	0.50	scal (12.5%)
FMADD D1, D2, D3, D1	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	4	0.25	scal (25.0%)
LDR S2, [X20, W16,UXTW #2]	1	0	0	0	0	0	0	0	0	0	0	0.33	0.33	0.33	6	0.33	scal (12.5%)
LDR S3, [X22, W17,UXTW #2]	1	0	0	0	0	0	0	0	0	0	0	0.33	0.33	0.33	6	0.33	scal (12.5%)
ADD W16, W16, #2	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	1	0.25	N/A
FCVT D2, S2	1	0	0	0	0	0	0	0.50	0	0.50	0	0	0	0	3	0.50	scal (12.5%)
FCVT D3, S3	1	0	0	0	0	0	0	0.50	0	0.50	0	0	0	0	3	0.50	scal (12.5%)
FMADD D1, D2, D3, D1	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	4	0.25	scal (25.0%)
B.NE 125c0 <main+0x15c0>	1	0.50	0.50	0	0	0	0	0	0	0	0	0	0	0	1	0.50	N/A

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

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

The related source loop is unrolled by 4 (including vectorization).

gain
potential
hint
expert

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 not vectorized. Only 16% of vector register length is used (average across all VPU instructions).

Details

All VPU instructions are used in scalar version (process only one data element in vector registers).

FMA

Detected 2 FMA (fused multiply-add) operations.

Slow data structures access

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

Details

Constant unknown stride: 2 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)

Conversion instructions

Detected expensive conversion instructions.

Details

FCVT: 4 occurrences►
- /home/eoseret/llm-attention/attention.cpp:31
- /home/eoseret/llm-attention/attention.cpp:31
- /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.

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

The binary loop is composed of 4 FP arithmetical operations:

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

The binary loop is loading 12 bytes.

Arithmetic intensity

Arithmetic intensity is 0.33 FP operations per loaded or stored byte.

General properties

nb instructions	17
nb uops	17
loop length	68
used w registers	4
used x registers	6
used b registers	0
used h registers	0
used s registers	2
used d registers	3
used q registers	0
used v registers	0
used z registers	0
nb stack references	0

Front-end

MACRO FUSION NOT POSSIBLE FIT IN UOP CACHE

micro-operation queue	2.13 cycles
front end	2.13 cycles

Back-end

	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11	P12	P13	P14
uops	0.50	0.50	1.50	1.50	1.50	1.50	2.00	1.00	2.00	1.00	1.33	1.33	1.33	0.00	0.00
cycles	0.50	0.50	1.50	1.50	1.50	1.50	2.00	1.00	2.00	1.00	1.33	1.33	1.33	0.00	0.00

Execution ports to units layout:

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

Cycles executing div or sqrt instructions	NA
Longest recurrence chain latency (RecMII)	4.00

Cycles summary

Front-end	2.13
Dispatch	2.00
Data deps.	4.00
Overall L1	4.00

Vectorization ratios

INT

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

INT+FP

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

Vector efficiency ratios

INT

all	16%
load	12%
store	NA (no store vectorizable/vectorized instructions)
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	25%
fma	NA (no fma vectorizable/vectorized instructions)
other	25%

all	16%
load	NA (no load vectorizable/vectorized instructions)
store	NA (no store vectorizable/vectorized instructions)
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	NA (no add-sub vectorizable/vectorized instructions)
fma	25%
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	12%

INT+FP

all	16%
load	12%
store	NA (no store vectorizable/vectorized instructions)
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	25%
fma	25%
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 4.00 cycles. At this rate:

3% of peak load performance is reached (3.00 out of 96.00 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: 125c0

Instruction	Nb FU	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11	P12	Latency	Recip. throughput	Vectorization
SUB W17, W16, #1	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	1	0.25	N/A
LDR S3, [X22, W15,UXTW #2]	1	0	0	0	0	0	0	0	0	0	0	0.33	0.33	0.33	6	0.33	scal (12.5%)
ADD X14, X14, #2	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	1	0.25	scal (25.0%)
LDR S2, [X20, W17,UXTW #2]	1	0	0	0	0	0	0	0	0	0	0	0.33	0.33	0.33	6	0.33	scal (12.5%)
ADD W17, W25, W15	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	1	0.25	N/A
ADD X15, X15, X19	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	1	0.25	N/A
CMP X9, X14	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	1	0.33	scal (25.0%)
FCVT D3, S3	1	0	0	0	0	0	0	0.50	0	0.50	0	0	0	0	3	0.50	scal (12.5%)
FCVT D2, S2	1	0	0	0	0	0	0	0.50	0	0.50	0	0	0	0	3	0.50	scal (12.5%)
FMADD D1, D2, D3, D1	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	4	0.25	scal (25.0%)
LDR S2, [X20, W16,UXTW #2]	1	0	0	0	0	0	0	0	0	0	0	0.33	0.33	0.33	6	0.33	scal (12.5%)
LDR S3, [X22, W17,UXTW #2]	1	0	0	0	0	0	0	0	0	0	0	0.33	0.33	0.33	6	0.33	scal (12.5%)
ADD W16, W16, #2	1	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	1	0.25	N/A
FCVT D2, S2	1	0	0	0	0	0	0	0.50	0	0.50	0	0	0	0	3	0.50	scal (12.5%)
FCVT D3, S3	1	0	0	0	0	0	0	0.50	0	0.50	0	0	0	0	3	0.50	scal (12.5%)
FMADD D1, D2, D3, D1	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	4	0.25	scal (25.0%)
B.NE 125c0 <main+0x15c0>	1	0.50	0.50	0	0	0	0	0	0	0	0	0	0	0	1	0.50	N/A

Report Configuration

Unrolling/vectorization cost

Details

Workaround

Vectorization

Details

FMA

Slow data structures access

Details

Workaround

Conversion instructions

Details

Workaround

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

Unrolling/vectorization cost

Details

Workaround

Vectorization

Details

FMA

Slow data structures access

Details

Workaround

Conversion instructions

Details

Workaround

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