OV - attention-gcc-native - Loop 2 - attention-gcc-native

	Loop Id: 2	Module: attention-gcc-native	Source: random.tcc:333-3371 [...]	Coverage: 0.52%

0x4107c8 MOVZ	X7, #5400

0x4107cc ADD	X3, X1, #1

0x4107d0 ADD	X0, SP, X7,UXTX

0x4107d4 LDR	X0, [X0, X1,LSL #3]    [3]

0x4107d8 STR	X3, [SP, #10392]    [1]

0x4107dc UBFM	X1, X0, #11, #42

0x4107e0 EOR	X0, X0, X1

0x4107e4 MOVZ	X1, #22144

0x4107e8 MOVK	X1, #40236

0x4107ec AND	X1, X1, X0,LSL #7

0x4107f0 EOR	X0, X0, X1

0x4107f4 AND	X1, X10, X0,LSL #15

0x4107f8 EOR	X0, X0, X1

0x4107fc EOR	X0, X0, X0,LSR #18

0x410800 UCVTF	S31, X0

0x410804 FMUL	S31, S31, S15

0x410808 FCMPE	S31, S13

0x41080c B.GE	410840

0x410810 STR	S31, [X19], #4    [2]

0x410814 CMP	X19, X20

0x410818 B.EQ	410850

0x41081c ORR	X1, XZR, X3

0x410820 CMP	X1, #623

0x410824 B.LS	4107c8

0x410840 MOVN	W0, #49280

0x410844 STR	W0, [X19], #4    [2]

0x410848 CMP	X20, X19

0x41084c B.NE	41081c

/home/eoseret/llm-attention/attention_v2.cpp: 163 - 163

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

163:       for (size_t i = 0; i < elemsX; ++i) h_X[i] = dist(rng);

/usr/include/c++/14/bits/random.tcc: 333 - 3371

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

333:       for (size_t __i = 1; __i < state_size; ++__i)

[...]

458:       if (_M_p >= state_size)

459: 	_M_gen_rand();

460:

461:       // Calculate o(x(i)).

462:       result_type __z = _M_x[_M_p++];

463:       __z ^= (__z >> __u) & __d;

464:       __z ^= (__z << __s) & __b;

465:       __z ^= (__z << __t) & __c;

466:       __z ^= (__z >> __l);

[...]

3367: 	  __sum += _RealType(__urng() - __urng.min()) * __tmp;

3368: 	  __tmp *= __r;

3369: 	}

3370:       __ret = __sum / __tmp;

3371:       if (__builtin_expect(__ret >= _RealType(1), 0))

Coverage (%)	Name	Source Location	Module
►100.00+	__libc_start_call_main		libc.so.6
○	__libc_start_main		libc.so.6
○	_start	stl_vector.h:1920	attention-gcc-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	1.35
CQA speedup if FP arith vectorized	2.31
CQA speedup if fully vectorized	3.44
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.05
Bottlenecks
Function	main
Source	attention_v2.cpp:163-163,random.tcc:333-333,random.tcc:458-466,random.tcc:3367-3371
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	3.06
CQA cycles if no scalar integer	2.27
CQA cycles if FP arith vectorized	1.33
CQA cycles if fully vectorized	0.89
Front-end cycles	3.06
P0 cycles	1.50
P1 cycles	1.50
P2 cycles	2.92
P3 cycles	2.92
P4 cycles	2.92
P5 cycles	2.92
P6 cycles	2.92
P7 cycles	2.92
P8 cycles	1.00
P9 cycles	0.75
P10 cycles	0.38
P11 cycles	0.38
P12 cycles	1.00
P13 cycles	1.00
P14 cycles	1.00
P15 cycles	0.75
P16 cycles	0.75
DIV/SQRT cycles	0.00
Inter-iter dependencies cycles	0
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	24.50
Nb uops	24.50
Nb loads	NA
Nb stores	2.00
Nb stack references	1.00
FLOP/cycle	0.33
Nb FLOP add-sub	0.00
Nb FLOP mul	1.00
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	0.98
Bytes prefetched	0.00
Bytes loaded	0.00
Bytes stored	3.00
Stride 0	2.00
Stride 1	0.00
Stride n	0.00
Stride unknown	0.00
Stride indirect	1.00
Vectorization ratio all	0.00
Vectorization ratio load	NA
Vectorization ratio store	0.00
Vectorization ratio mul	0.00
Vectorization ratio add_sub	0.00
Vectorization ratio fma	NA
Vectorization ratio div_sqrt	NA
Vectorization ratio other	0.00
Vector-efficiency ratio all	40.63
Vector-efficiency ratio load	NA
Vector-efficiency ratio store	37.50
Vector-efficiency ratio mul	25.00
Vector-efficiency ratio add_sub	50.00
Vector-efficiency ratio fma	NA
Vector-efficiency ratio div_sqrt	NA
Vector-efficiency ratio other	43.75

Metric	Value
CQA speedup if no scalar integer	2.12
CQA speedup if FP arith vectorized	2.29
CQA speedup if fully vectorized	3.43
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.06
Bottlenecks	micro-operation queue,
Function	main
Source	attention_v2.cpp:163-163,random.tcc:333-333,random.tcc:458-466,random.tcc:3367-3371
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	3.00
CQA cycles if no scalar integer	1.42
CQA cycles if FP arith vectorized	1.31
CQA cycles if fully vectorized	0.88
Front-end cycles	3.00
P0 cycles	1.50
P1 cycles	1.50
P2 cycles	2.83
P3 cycles	2.83
P4 cycles	2.83
P5 cycles	2.83
P6 cycles	2.83
P7 cycles	2.83
P8 cycles	1.00
P9 cycles	1.00
P10 cycles	0.50
P11 cycles	0.50
P12 cycles	1.00
P13 cycles	1.00
P14 cycles	1.00
P15 cycles	0.50
P16 cycles	0.50
DIV/SQRT cycles	0.00
Inter-iter dependencies cycles	0
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	24.00
Nb uops	24.00
Nb loads	NA
Nb stores	2.00
Nb stack references	1.00
FLOP/cycle	0.33
Nb FLOP add-sub	0.00
Nb FLOP mul	1.00
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	1.00
Bytes prefetched	0.00
Bytes loaded	0.00
Bytes stored	3.00
Stride 0	2.00
Stride 1	0.00
Stride n	0.00
Stride unknown	0.00
Stride indirect	1.00
Vectorization ratio all	0.00
Vectorization ratio load	NA
Vectorization ratio store	0.00
Vectorization ratio mul	0.00
Vectorization ratio add_sub	0.00
Vectorization ratio fma	NA
Vectorization ratio div_sqrt	NA
Vectorization ratio other	0.00
Vector-efficiency ratio all	40.63
Vector-efficiency ratio load	NA
Vector-efficiency ratio store	37.50
Vector-efficiency ratio mul	25.00
Vector-efficiency ratio add_sub	50.00
Vector-efficiency ratio fma	NA
Vector-efficiency ratio div_sqrt	NA
Vector-efficiency ratio other	43.75

Metric	Value
CQA speedup if no scalar integer	1.00
CQA speedup if FP arith vectorized	2.33
CQA speedup if fully vectorized	3.45
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.04
Bottlenecks	micro-operation queue,
Function	main
Source	attention_v2.cpp:163-163,random.tcc:333-333,random.tcc:458-466,random.tcc:3367-3371
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	3.13
CQA cycles if no scalar integer	3.13
CQA cycles if FP arith vectorized	1.34
CQA cycles if fully vectorized	0.91
Front-end cycles	3.13
P0 cycles	1.50
P1 cycles	1.50
P2 cycles	3.00
P3 cycles	3.00
P4 cycles	3.00
P5 cycles	3.00
P6 cycles	3.00
P7 cycles	3.00
P8 cycles	1.00
P9 cycles	0.50
P10 cycles	0.25
P11 cycles	0.25
P12 cycles	1.00
P13 cycles	1.00
P14 cycles	1.00
P15 cycles	1.00
P16 cycles	1.00
DIV/SQRT cycles	0.00
Inter-iter dependencies cycles	0
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	25.00
Nb uops	25.00
Nb loads	NA
Nb stores	2.00
Nb stack references	1.00
FLOP/cycle	0.32
Nb FLOP add-sub	0.00
Nb FLOP mul	1.00
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	0.96
Bytes prefetched	0.00
Bytes loaded	0.00
Bytes stored	3.00
Stride 0	2.00
Stride 1	0.00
Stride n	0.00
Stride unknown	0.00
Stride indirect	1.00
Vectorization ratio all	0.00
Vectorization ratio load	NA
Vectorization ratio store	0.00
Vectorization ratio mul	0.00
Vectorization ratio add_sub	0.00
Vectorization ratio fma	NA
Vectorization ratio div_sqrt	NA
Vectorization ratio other	0.00
Vector-efficiency ratio all	40.63
Vector-efficiency ratio load	NA
Vector-efficiency ratio store	37.50
Vector-efficiency ratio mul	25.00
Vector-efficiency ratio add_sub	50.00
Vector-efficiency ratio fma	NA
Vector-efficiency ratio div_sqrt	NA
Vector-efficiency ratio other	43.75

Path /

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

Function	main
Source file and lines	random.tcc:333-3371
Module	attention-gcc-native

The loop is defined in:

/home/eoseret/llm-attention/attention_v2.cpp:163
/usr/include/c++/14/bits/random.tcc:333,458-466,3367-3371

The related source loop is not unrolled or unrolled with no peel/tail loop.
The structure of this loop is probably <if then [else] end>.

The presence of multiple execution paths is typically the main/first bottleneck.
Try to simplify control inside loop: ideally, try to remove all conditional expressions, for example by (if applicable):

hoisting them (moving them outside the loop)
turning them into conditional moves, MIN or MAX

Ex: if (x<0) x=0 => x = (x<0 ? 0 : x) (or MAX(0,x) after defining the corresponding macro)

gain
potential
hint
expert

Vectorization

Your loop is not vectorized. Only 40% of vector register length is used (average across all VPU instructions). By vectorizing your loop, you can lower the cost of an iteration from 3.06 to 0.89 cycles (3.44x speedup).

Details

All VPU instructions are used in scalar version (process only one data element in vector registers). Since your execution units are vector units, only a vectorized loop can use their full power.

Workaround

Try another compiler or update/tune your current one
Remove inter-iterations dependences from your loop and make it unit-stride:
- If your arrays have 2 or more dimensions, check whether elements are accessed contiguously and, otherwise, try to permute loops accordingly: C storage order is row-major: for(i) for(j) a[j][i] = b[j][i]; (slow, non stride 1) => for(i) for(j) a[i][j] = b[i][j]; (fast, stride 1)
- If your loop streams arrays of structures (AoS), try to use structures of arrays instead (SoA): for(i) a[i].x = b[i].x; (slow, non stride 1) => for(i) a.x[i] = b.x[i]; (fast, stride 1)

No data for this section

Slow data structures access

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

Details

Irregular (variable stride) or indirect: 1 occurrence(s)

Non-unit stride (uncontiguous) accesses are not efficiently using data caches

Workaround

Try to remove indirect accesses. If applicable, precompute elements out of the innermost loop.

Conversion instructions

Detected expensive conversion instructions.

Details

UCVTF: 0 occurrences

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 1 FP arithmetical operations:

1: multiply

The binary loop is storing 3 bytes.

Arithmetic intensity

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

General properties

nb instructions	24.50
nb uops	24.50
loop length	98
used w registers	0.50
used x registers	8
used b registers	0
used h registers	0
used s registers	4
used d registers	0
used q registers	0
used v registers	0
used z registers	0
nb stack references	1

Front-end

MACRO FUSION NOT POSSIBLE

micro-operation queue	3.06 cycles
front end	3.06 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	2.92	2.92	2.92	2.92	2.92	2.92	1.00	0.75	0.38	0.38	1.00	1.00	1.00	0.75	0.75
cycles	1.50	1.50	2.92	2.92	2.92	2.92	2.92	2.92	1.00	0.75	0.38	0.38	1.00	1.00	1.00	0.75	0.75

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	NA
Longest recurrence chain latency (RecMII)	0.00

Cycles summary

Front-end	3.06
Dispatch	2.92
Data deps.	0.00
Overall L1	3.06

Vectorization ratios

INT

all	0%
load	NA (no load vectorizable/vectorized instructions)
store	0%
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	0%
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	0%

INT+FP

all	0%
load	NA (no load vectorizable/vectorized instructions)
store	0%
mul	0%
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	45%
load	NA (no load vectorizable/vectorized instructions)
store	37%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	50%
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	25%
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	25%

INT+FP

all	40%
load	NA (no load vectorizable/vectorized instructions)
store	37%
mul	25%
add-sub	50%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	43%

Cycles and memory resources usage

Assuming all data fit into the L1 cache, each iteration of the binary loop takes 3.06 cycles. At this rate:

2% of peak store performance is reached (0.98 out of 48.00 bytes stored per cycle (GB/s @ 1GHz))

Function	main
Source file and lines	random.tcc:333-3371
Module	attention-gcc-native

The loop is defined in:

/home/eoseret/llm-attention/attention_v2.cpp:163
/usr/include/c++/14/bits/random.tcc:333,458-466,3367-3371

hoisting them (moving them outside the loop)
turning them into conditional moves, MIN or MAX

Ex: if (x<0) x=0 => x = (x<0 ? 0 : x) (or MAX(0,x) after defining the corresponding macro)

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 3.00 to 1.42 cycles (2.12x speedup).

Workaround

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

Vectorization

Details

All VPU instructions are used in scalar version (process only one data element in vector registers). Since your execution units are vector units, only a vectorized loop can use their full power.

Workaround

Try another compiler or update/tune your current one
Remove inter-iterations dependences from your loop and make it unit-stride:
- If your arrays have 2 or more dimensions, check whether elements are accessed contiguously and, otherwise, try to permute loops accordingly: C storage order is row-major: for(i) for(j) a[j][i] = b[j][i]; (slow, non stride 1) => for(i) for(j) a[i][j] = b[i][j]; (fast, stride 1)
- If your loop streams arrays of structures (AoS), try to use structures of arrays instead (SoA): for(i) a[i].x = b[i].x; (slow, non stride 1) => for(i) a.x[i] = b.x[i]; (fast, stride 1)

No data for this section

Slow data structures access

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

Details

Irregular (variable stride) or indirect: 1 occurrence(s)

Non-unit stride (uncontiguous) accesses are not efficiently using data caches

Workaround

Try to remove indirect accesses. If applicable, precompute elements out of the innermost loop.

Conversion instructions

Detected expensive conversion instructions.

Details

UCVTF: 1 occurrences►
- /usr/include/c++/14/bits/random.tcc:3367

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 1 FP arithmetical operations:

1: multiply

The binary loop is storing 3 bytes.

Arithmetic intensity

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

General properties

nb instructions	24
nb uops	24
loop length	96
used w registers	0
used x registers	8
used b registers	0
used h registers	0
used s registers	4
used d registers	0
used q registers	0
used v registers	0
used z registers	0
nb stack references	1

Front-end

MACRO FUSION NOT POSSIBLE FIT IN UOP CACHE

micro-operation queue	3.00 cycles
front end	3.00 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	2.83	2.83	2.83	2.83	2.83	2.83	1.00	1.00	0.50	0.50	1.00	1.00	1.00	0.50	0.50
cycles	1.50	1.50	2.83	2.83	2.83	2.83	2.83	2.83	1.00	1.00	0.50	0.50	1.00	1.00	1.00	0.50	0.50

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	NA
Longest recurrence chain latency (RecMII)	0.00

Cycles summary

Front-end	3.00
Dispatch	2.83
Data deps.	0.00
Overall L1	3.00

Vectorization ratios

INT

all	0%
load	NA (no load vectorizable/vectorized instructions)
store	0%
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	0%
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	0%

INT+FP

all	0%
load	NA (no load vectorizable/vectorized instructions)
store	0%
mul	0%
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	45%
load	NA (no load vectorizable/vectorized instructions)
store	37%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	50%
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	25%
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	25%

INT+FP

all	40%
load	NA (no load vectorizable/vectorized instructions)
store	37%
mul	25%
add-sub	50%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	43%

Cycles and memory resources usage

Assuming all data fit into the L1 cache, each iteration of the binary loop takes 3.00 cycles. At this rate:

2% of peak store performance is reached (1.00 out of 48.00 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 3.00 to 2.83 cycles (1.06x speedup).

ASM code

In the binary file, the address of the loop is: 4107c8

Instruction	Nb FU	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11	P12	P13	P14	P15	P16	Latency	Recip. throughput	Vectorization
MOVZ X7, #5400	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	scal (50.0%)
ADD X3, X1, #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	scal (50.0%)
ADD X0, SP, X7,UXTX	1	0	0	0	0	0	0	0.50	0.50	0	0	0	0	0	0	0	0	0	2	0.50	N/A
LDR X0, [X0, X1,LSL #3]	1	0	0	0	0	0	0	0	0	0	0	0	0	0.33	0.33	0.33	0	0	4	0.33	N/A
STR X3, [SP, #10392]	1	0	0	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0.50	0.50	1	0.50	scal (50.0%)
UBFM X1, X0, #11, #42	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
EOR X0, X0, X1	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
MOVZ X1, #22144	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
MOVK X1, #40236	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
AND X1, X1, X0,LSL #7	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
EOR X0, X0, X1	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
AND X1, X10, X0,LSL #15	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
EOR X0, X0, X1	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
EOR X0, X0, X0,LSR #18	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
UCVTF S31, X0	1	0	0	0	0	0	0	1	0	0	0	0	0	0	0	0	0	0	3	1	scal (50.0%)
FMUL S31, S31, S15	1	0	0	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	3	0.25	scal (25.0%)
FCMPE S31, S13	1	0	0	0	0	0	0	0	0	1	0	0	0	0	0	0	0	0	2	1	scal (25.0%)
B.GE 410840 <main+0x680>	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
STR S31, [X19], #4	1	0	0	0.17	0.17	0.17	0.17	0.17	0.17	0.50	0.50	0	0	0.50	0.50	0	0	0	2	0.50	scal (25.0%)
CMP X19, X20	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	scal (50.0%)
B.EQ 410850 <main+0x690>	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
ORR X1, XZR, X3	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
CMP X1, #623	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.LS 4107c8 <main+0x608>	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	random.tcc:333-3371
Module	attention-gcc-native

The loop is defined in:

/home/eoseret/llm-attention/attention_v2.cpp:163
/usr/include/c++/14/bits/random.tcc:333,458-466,3367-3371

hoisting them (moving them outside the loop)
turning them into conditional moves, MIN or MAX

Ex: if (x<0) x=0 => x = (x<0 ? 0 : x) (or MAX(0,x) after defining the corresponding macro)

gain
potential
hint
expert

Vectorization

Details

All VPU instructions are used in scalar version (process only one data element in vector registers). Since your execution units are vector units, only a vectorized loop can use their full power.

Workaround

Try another compiler or update/tune your current one
Remove inter-iterations dependences from your loop and make it unit-stride:
- If your arrays have 2 or more dimensions, check whether elements are accessed contiguously and, otherwise, try to permute loops accordingly: C storage order is row-major: for(i) for(j) a[j][i] = b[j][i]; (slow, non stride 1) => for(i) for(j) a[i][j] = b[i][j]; (fast, stride 1)
- If your loop streams arrays of structures (AoS), try to use structures of arrays instead (SoA): for(i) a[i].x = b[i].x; (slow, non stride 1) => for(i) a.x[i] = b.x[i]; (fast, stride 1)

No data for this section

Slow data structures access

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

Details

Irregular (variable stride) or indirect: 1 occurrence(s)

Non-unit stride (uncontiguous) accesses are not efficiently using data caches

Workaround

Try to remove indirect accesses. If applicable, precompute elements out of the innermost loop.

Conversion instructions

Detected expensive conversion instructions.

Details

UCVTF: 1 occurrences►
- /usr/include/c++/14/bits/random.tcc:3367

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 1 FP arithmetical operations:

1: multiply

The binary loop is storing 3 bytes.

Arithmetic intensity

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

General properties

nb instructions	25
nb uops	25
loop length	100
used w registers	1
used x registers	8
used b registers	0
used h registers	0
used s registers	4
used d registers	0
used q registers	0
used v registers	0
used z registers	0
nb stack references	1

Front-end

MACRO FUSION NOT POSSIBLE FIT IN UOP CACHE

micro-operation queue	3.13 cycles
front end	3.13 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	3.00	3.00	3.00	3.00	3.00	3.00	1.00	0.50	0.25	0.25	1.00	1.00	1.00	1.00	1.00
cycles	1.50	1.50	3.00	3.00	3.00	3.00	3.00	3.00	1.00	0.50	0.25	0.25	1.00	1.00	1.00	1.00	1.00

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	NA
Longest recurrence chain latency (RecMII)	0.00

Cycles summary

Front-end	3.13
Dispatch	3.00
Data deps.	0.00
Overall L1	3.13

Vectorization ratios

INT

all	0%
load	NA (no load vectorizable/vectorized instructions)
store	0%
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	0%
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	0%

INT+FP

all	0%
load	NA (no load vectorizable/vectorized instructions)
store	0%
mul	0%
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	45%
load	NA (no load vectorizable/vectorized instructions)
store	37%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	50%
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	25%
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	25%

INT+FP

all	40%
load	NA (no load vectorizable/vectorized instructions)
store	37%
mul	25%
add-sub	50%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	43%

Cycles and memory resources usage

Assuming all data fit into the L1 cache, each iteration of the binary loop takes 3.13 cycles. At this rate:

2% of peak store performance is reached (0.96 out of 48.00 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).

ASM code

In the binary file, the address of the loop is: 4107c8

Instruction	Nb FU	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11	P12	P13	P14	P15	P16	Latency	Recip. throughput	Vectorization
MOVZ X7, #5400	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	scal (50.0%)
ADD X3, X1, #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	scal (50.0%)
ADD X0, SP, X7,UXTX	1	0	0	0	0	0	0	0.50	0.50	0	0	0	0	0	0	0	0	0	2	0.50	N/A
LDR X0, [X0, X1,LSL #3]	1	0	0	0	0	0	0	0	0	0	0	0	0	0.33	0.33	0.33	0	0	4	0.33	N/A
STR X3, [SP, #10392]	1	0	0	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0.50	0.50	1	0.50	scal (50.0%)
UBFM X1, X0, #11, #42	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
EOR X0, X0, X1	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
MOVZ X1, #22144	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
MOVK X1, #40236	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
AND X1, X1, X0,LSL #7	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
EOR X0, X0, X1	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
AND X1, X10, X0,LSL #15	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
EOR X0, X0, X1	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
EOR X0, X0, X0,LSR #18	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
UCVTF S31, X0	1	0	0	0	0	0	0	1	0	0	0	0	0	0	0	0	0	0	3	1	scal (50.0%)
FMUL S31, S31, S15	1	0	0	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	3	0.25	scal (25.0%)
FCMPE S31, S13	1	0	0	0	0	0	0	0	0	1	0	0	0	0	0	0	0	0	2	1	scal (25.0%)
B.GE 410840 <main+0x680>	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
ORR X1, XZR, X3	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
CMP X1, #623	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	scal (50.0%)
B.LS 4107c8 <main+0x608>	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
MOVN W0, #49280	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
STR W0, [X19], #4	1	0	0	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0.50	0.50	0	0.50	0.50	1	0.50	scal (25.0%)
CMP X20, X19	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 41081c <main+0x65c>	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

Vectorization

Details

Workaround

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

Code clean check

Workaround

Vectorization

Details

Workaround

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

Vectorization

Details

Workaround

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