OV - attention-gcc-native - Loop 31

	Loop Id: 31	Module: attention-gcc-native	Source: attention_v2.cpp:55-56	Coverage: 0.46%

0x411b20 LDR	S0, [X20, X25,LSL #2]    [1]

0x411b24 FSUB	S0, S0, S15

0x411b28 BL	410140

0x411b2c FDIV	S0, S0, S14

0x411b30 STR	S0, [X21, X25,LSL #2]    [2]

0x411b34 ADD	X25, X25, #1

0x411b38 CMP	X19, X25

0x411b3c B.NE	411b20

/home/eoseret/llm-attention/attention_v2.cpp: 55 - 56

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

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

56:          P[row * N + idx] = expf(S_row[idx] - max_val) / sum;

Coverage (%)	Name	Source Location	Module
►100.00+	main	attention_v2.cpp:283	attention-gcc-native
○	__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	1.00 - 1.13
CQA speedup if FP arith vectorized	1.35
CQA speedup if fully vectorized	4.00
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.13
Bottlenecks	P8, P10,
Function	softmax(float const, float, float*, int)
Source	attention_v2.cpp:55-56
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	1.00 - 1.13
CQA cycles if no scalar integer	1.00
CQA cycles if FP arith vectorized	0.83
CQA cycles if fully vectorized	0.25 - 0.28
Front-end cycles	1.00
P0 cycles	1.00
P1 cycles	1.00
P2 cycles	0.50
P3 cycles	0.50
P4 cycles	0.50
P5 cycles	0.50
P6 cycles	0.50
P7 cycles	0.50
P8 cycles	1.00
P9 cycles	0.75
P10 cycles	0.75
P11 cycles	0.50
P12 cycles	0.83
P13 cycles	0.50
P14 cycles	0.67
P15 cycles	0.00
P16 cycles	0.00
DIV/SQRT cycles	0.87 - 1.13
Inter-iter dependencies cycles	1
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	8.00
Nb uops	8.00
Nb loads	NA
Nb stores	1.00
Nb stack references	0.00
FLOP/cycle	2.00 - 1.78
Nb FLOP add-sub	1.00
Nb FLOP mul	0.00
Nb FLOP fma	0.00
Nb FLOP div	1.00
Nb FLOP rcp	0.00
Nb FLOP sqrt	0.00
Nb FLOP rsqrt	0.00
Bytes/cycle	5.33 - 6.00
Bytes prefetched	0.00
Bytes loaded	3.00
Bytes stored	3.00
Stride 0	0.00
Stride 1	0.00
Stride n	0.00
Stride unknown	0.00
Stride indirect	0.00
Vectorization ratio all	0.00
Vectorization ratio load	0.00
Vectorization ratio store	0.00
Vectorization ratio mul	NA
Vectorization ratio add_sub	0.00
Vectorization ratio fma	NA
Vectorization ratio div_sqrt	0.00
Vectorization ratio other	NA
Vector-efficiency ratio all	25.00
Vector-efficiency ratio load	25.00
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	25.00
Vector-efficiency ratio other	NA

Metric	Value
CQA speedup if no scalar integer	1.00 - 1.13
CQA speedup if FP arith vectorized	1.35
CQA speedup if fully vectorized	4.00
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.13
Bottlenecks	P8, P10,
Function	softmax(float const, float, float*, int)
Source	attention_v2.cpp:55-56
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	1.00 - 1.13
CQA cycles if no scalar integer	1.00
CQA cycles if FP arith vectorized	0.83
CQA cycles if fully vectorized	0.25 - 0.28
Front-end cycles	1.00
P0 cycles	1.00
P1 cycles	1.00
P2 cycles	0.50
P3 cycles	0.50
P4 cycles	0.50
P5 cycles	0.50
P6 cycles	0.50
P7 cycles	0.50
P8 cycles	1.00
P9 cycles	0.75
P10 cycles	0.75
P11 cycles	0.50
P12 cycles	0.83
P13 cycles	0.50
P14 cycles	0.67
P15 cycles	0.00
P16 cycles	0.00
DIV/SQRT cycles	0.87 - 1.13
Inter-iter dependencies cycles	1
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	8.00
Nb uops	8.00
Nb loads	NA
Nb stores	1.00
Nb stack references	0.00
FLOP/cycle	2.00 - 1.78
Nb FLOP add-sub	1.00
Nb FLOP mul	0.00
Nb FLOP fma	0.00
Nb FLOP div	1.00
Nb FLOP rcp	0.00
Nb FLOP sqrt	0.00
Nb FLOP rsqrt	0.00
Bytes/cycle	5.33 - 6.00
Bytes prefetched	0.00
Bytes loaded	3.00
Bytes stored	3.00
Stride 0	0.00
Stride 1	0.00
Stride n	0.00
Stride unknown	0.00
Stride indirect	0.00
Vectorization ratio all	0.00
Vectorization ratio load	0.00
Vectorization ratio store	0.00
Vectorization ratio mul	NA
Vectorization ratio add_sub	0.00
Vectorization ratio fma	NA
Vectorization ratio div_sqrt	0.00
Vectorization ratio other	NA
Vector-efficiency ratio all	25.00
Vector-efficiency ratio load	25.00
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	25.00
Vector-efficiency ratio other	NA

Path /

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

Function	softmax(float const, float, float*, int)
Source file and lines	attention_v2.cpp:55-56
Module	attention-gcc-native

The loop is defined in /home/eoseret/llm-attention/attention_v2.cpp:55-56.

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

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 1.13 to 1.00 cycles (1.13x speedup).

Workaround

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

Vectorization

Your loop is not vectorized. 4 data elements could be processed at once in vector registers. By vectorizing your loop, you can lower the cost of an iteration from 1.13 to 0.28 cycles (4.00x 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)

Expensive FP math instructions/calls

Detected performance impact from expensive FP math instructions/calls. By removing/reexpressing them, you can lower the cost of an iteration from 1.13 to 1.00 cycles (1.13x speedup).

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_v2.cpp:56

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

The binary loop is composed of 2 FP arithmetical operations:

1: addition or subtraction
1: divide

The binary loop is loading 3 bytes. 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	8
nb uops	8
loop length	32
used w registers	0
used x registers	4
used b registers	0
used h registers	0
used s registers	3
used d registers	0
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	1.00 cycles
front end	1.00 cycles

Back-end

	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11	P12	P13	P14	P15	P16
uops	1.00	1.00	0.50	0.50	0.50	0.50	0.50	0.50	1.00	0.75	0.75	0.50	0.83	0.50	0.67	0.00	0.00
cycles	1.00	1.00	0.50	0.50	0.50	0.50	0.50	0.50	1.00	0.75	0.75	0.50	0.83	0.50	0.67	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	0.87-1.13
Longest recurrence chain latency (RecMII)	1.00

Cycles summary

Front-end	1.00
Dispatch	1.00
DIV/SQRT	0.87-1.13
Data deps.	1.00
Overall L1	1.00-1.13

Vectorization ratios

INT

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

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	0%
other	NA (no other vectorizable/vectorized instructions)

INT+FP

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

Vector efficiency ratios

INT

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

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	25%
other	NA (no other vectorizable/vectorized instructions)

INT+FP

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

Cycles and memory resources usage

Assuming all data fit into the L1 cache, each iteration of the binary loop takes 1.13 cycles. At this rate:

2% of peak load performance is reached (2.66 out of 96.00 bytes loaded per cycle (GB/s @ 1GHz))
5% of peak store performance is reached (2.66 out of 48.00 bytes stored per cycle (GB/s @ 1GHz))

Front-end bottlenecks

Found no such bottlenecks.

ASM code

In the binary file, the address of the loop is: 411b20

Instruction	Nb FU	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11	P12	P13	P14	Latency	Recip. throughput	Vectorization
LDR S0, [X20, X25,LSL #2]	1	0	0	0	0	0	0	0	0	0	0	0	0	0.33	0.33	0.33	6	0.33	scal (25.0%)
FSUB S0, S0, S15	1	0	0	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	2	0.25	scal (25.0%)
BL 410140 <@plt_start@+0x120>	1	0.50	0.50	0.50	0.50	0	0	0	0	0	0	0	0	0	0	0	1	0.50	N/A
FDIV S0, S0, S14	1	0	0	0	0	0	0	0	0	0.50	0	0.50	0	0	0	0	7-10	0.87-1.13	scal (25.0%)
STR S0, [X21, X25,LSL #2]	1	0	0	0	0	0	0	0	0	0.50	0.50	0	0	0.50	0.50	0	2	0.50	scal (25.0%)
ADD X25, X25, #1	1	0	0	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	1	0.17	N/A
CMP X19, X25	1	0	0	0.25	0.25	0	0	0.25	0.25	0	0	0	0	0	0	0	1	0.33	N/A
B.NE 411b20 <_Z7softmaxPKfPfS1_i+0x140>	1	0.50	0.50	0	0	0	0	0	0	0	0	0	0	0	0	0	1	0.50	N/A

Function	softmax(float const, float, float*, int)
Source file and lines	attention_v2.cpp:55-56
Module	attention-gcc-native

The loop is defined in /home/eoseret/llm-attention/attention_v2.cpp:55-56.

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

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 1.13 to 1.00 cycles (1.13x speedup).

Workaround

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

Vectorization

Your loop is not vectorized. 4 data elements could be processed at once in vector registers. By vectorizing your loop, you can lower the cost of an iteration from 1.13 to 0.28 cycles (4.00x 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)

Expensive FP math instructions/calls

Detected performance impact from expensive FP math instructions/calls. By removing/reexpressing them, you can lower the cost of an iteration from 1.13 to 1.00 cycles (1.13x speedup).

CALL instructions

Detected function call instructions.

Details

unknown: 1 occurrences►
- /home/eoseret/llm-attention/attention_v2.cpp:56

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

The binary loop is composed of 2 FP arithmetical operations:

1: addition or subtraction
1: divide

The binary loop is loading 3 bytes. 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	8
nb uops	8
loop length	32
used w registers	0
used x registers	4
used b registers	0
used h registers	0
used s registers	3
used d registers	0
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	1.00 cycles
front end	1.00 cycles

Back-end

	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11	P12	P13	P14	P15	P16
uops	1.00	1.00	0.50	0.50	0.50	0.50	0.50	0.50	1.00	0.75	0.75	0.50	0.83	0.50	0.67	0.00	0.00
cycles	1.00	1.00	0.50	0.50	0.50	0.50	0.50	0.50	1.00	0.75	0.75	0.50	0.83	0.50	0.67	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	0.87-1.13
Longest recurrence chain latency (RecMII)	1.00

Cycles summary

Front-end	1.00
Dispatch	1.00
DIV/SQRT	0.87-1.13
Data deps.	1.00
Overall L1	1.00-1.13

Vectorization ratios

INT

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

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	0%
other	NA (no other vectorizable/vectorized instructions)

INT+FP

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

Vector efficiency ratios

INT

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

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	25%
other	NA (no other vectorizable/vectorized instructions)

INT+FP

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

Cycles and memory resources usage

Assuming all data fit into the L1 cache, each iteration of the binary loop takes 1.13 cycles. At this rate:

2% of peak load performance is reached (2.66 out of 96.00 bytes loaded per cycle (GB/s @ 1GHz))
5% of peak store performance is reached (2.66 out of 48.00 bytes stored per cycle (GB/s @ 1GHz))

Front-end bottlenecks

Found no such bottlenecks.

ASM code

In the binary file, the address of the loop is: 411b20

Instruction	Nb FU	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11	P12	P13	P14	Latency	Recip. throughput	Vectorization
LDR S0, [X20, X25,LSL #2]	1	0	0	0	0	0	0	0	0	0	0	0	0	0.33	0.33	0.33	6	0.33	scal (25.0%)
FSUB S0, S0, S15	1	0	0	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	2	0.25	scal (25.0%)
BL 410140 <@plt_start@+0x120>	1	0.50	0.50	0.50	0.50	0	0	0	0	0	0	0	0	0	0	0	1	0.50	N/A
FDIV S0, S0, S14	1	0	0	0	0	0	0	0	0	0.50	0	0.50	0	0	0	0	7-10	0.87-1.13	scal (25.0%)
STR S0, [X21, X25,LSL #2]	1	0	0	0	0	0	0	0	0	0.50	0.50	0	0	0.50	0.50	0	2	0.50	scal (25.0%)
ADD X25, X25, #1	1	0	0	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	1	0.17	N/A
CMP X19, X25	1	0	0	0.25	0.25	0	0	0.25	0.25	0	0	0	0	0	0	0	1	0.33	N/A
B.NE 411b20 <_Z7softmaxPKfPfS1_i+0x140>	1	0.50	0.50	0	0	0	0	0	0	0	0	0	0	0	0	0	1	0.50	N/A

Report Configuration

Code clean check

Workaround

Vectorization

Details

Workaround

Expensive FP math instructions/calls

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

Vectorization

Details

Workaround

Expensive FP math instructions/calls

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