OV - attention-clang-skl256 - Loop 47 - attention-clang-skl256

0x4e30 VMOVSS	(%RDI,%RBX,4),%XMM0    [1]

0x4e35 VSUBSS	%XMM2,%XMM0,%XMM0

0x4e39 MOV	%R12,%R13

0x4e3c MOV	%RCX,%R12

0x4e3f VZEROUPPER

0x4e42 CALL	1130 <expf@plt>

0x4e47 VMOVAPS	0xe0(%RSP),%XMM2    [2]

0x4e50 MOV	%R12,%RCX

0x4e53 MOV	%R13,%R12

0x4e56 VMOVAPS	0xc0(%RSP),%XMM1    [2]

0x4e5f MOV	(%RSP),%RDI    [2]

0x4e63 VDIVSS	%XMM1,%XMM0,%XMM0

0x4e67 VMOVSS	%XMM0,(%R13,%RBX,4)    [3]

0x4e6e INC	%RBX

0x4e71 CMP	%RBX,%RCX

0x4e74 JNE	4e30

/home/eoseret/Applications/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	Module
►100.00+	__libc_init_first	libc.so.6
○	__libc_start_main	libc.so.6
○	_start	attention-clang-skl256

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	1.46
CQA speedup if FP arith vectorized	2.81
CQA speedup if fully vectorized	6.33
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	NA
Bottlenecks	micro-operation queue,
Function	main
Source	attention_v2.cpp:55-56
Source loop unroll info	unrolled by 8
Source loop unroll confidence level	max
Unroll/vectorization loop type	peel/tail
Unroll factor	1
CQA cycles	4.75
CQA cycles if no scalar integer	3.25
CQA cycles if FP arith vectorized	1.69
CQA cycles if fully vectorized	0.75
Front-end cycles	4.75
P0 cycles	1.50
P1 cycles	1.00
P2 cycles	2.00
P3 cycles	2.00
P4 cycles	2.00
P5 cycles	1.00
P6 cycles	1.50
P7 cycles	2.00
DIV/SQRT cycles	3.00
Inter-iter dependencies cycles	1
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	16.00
Nb uops	19.00
Nb loads	4.00
Nb stores	1.00
Nb stack references	3.00
FLOP/cycle	0.42
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	10.11
Bytes prefetched	0.00
Bytes loaded	44.00
Bytes stored	4.00
Stride 0	1.00
Stride 1	2.00
Stride n	0.00
Stride unknown	0.00
Stride indirect	0.00
Vectorization ratio all	33.33
Vectorization ratio load	66.67
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	33.33
Vector-efficiency ratio all	13.89
Vector-efficiency ratio load	18.75
Vector-efficiency ratio store	6.25
Vector-efficiency ratio mul	NA
Vector-efficiency ratio add_sub	6.25
Vector-efficiency ratio fma	NA
Vector-efficiency ratio div_sqrt	6.25
Vector-efficiency ratio other	16.67

Metric	Value
CQA speedup if no scalar integer	1.46
CQA speedup if FP arith vectorized	2.81
CQA speedup if fully vectorized	6.33
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	NA
Bottlenecks	micro-operation queue,
Function	main
Source	attention_v2.cpp:55-56
Source loop unroll info	unrolled by 8
Source loop unroll confidence level	max
Unroll/vectorization loop type	peel/tail
Unroll factor	1
CQA cycles	4.75
CQA cycles if no scalar integer	3.25
CQA cycles if FP arith vectorized	1.69
CQA cycles if fully vectorized	0.75
Front-end cycles	4.75
P0 cycles	1.50
P1 cycles	1.00
P2 cycles	2.00
P3 cycles	2.00
P4 cycles	2.00
P5 cycles	1.00
P6 cycles	1.50
P7 cycles	2.00
DIV/SQRT cycles	3.00
Inter-iter dependencies cycles	1
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	16.00
Nb uops	19.00
Nb loads	4.00
Nb stores	1.00
Nb stack references	3.00
FLOP/cycle	0.42
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	10.11
Bytes prefetched	0.00
Bytes loaded	44.00
Bytes stored	4.00
Stride 0	1.00
Stride 1	2.00
Stride n	0.00
Stride unknown	0.00
Stride indirect	0.00
Vectorization ratio all	33.33
Vectorization ratio load	66.67
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	33.33
Vector-efficiency ratio all	13.89
Vector-efficiency ratio load	18.75
Vector-efficiency ratio store	6.25
Vector-efficiency ratio mul	NA
Vector-efficiency ratio add_sub	6.25
Vector-efficiency ratio fma	NA
Vector-efficiency ratio div_sqrt	6.25
Vector-efficiency ratio other	16.67

Path /

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

Function	main
Source file and lines	attention_v2.cpp:55-56
Module	attention-clang-skl256

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

It is peel/tail loop of related source loop which is unrolled by 8 (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 4.75 to 3.25 cycles (1.46x 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 probably not vectorized. Only 13% of vector register length is used (average across all SSE/AVX instructions).

Details

Store and arithmetical SSE/AVX instructions are used in scalar version (process only one data element in vector registers).

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

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.

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

Type of elements and instruction set

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

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 44 bytes (11 single precision FP elements). The binary loop is storing 4 bytes (1 single precision FP elements).

Arithmetic intensity

Arithmetic intensity is 0.04 FP operations per loaded or stored byte.

General properties

nb instructions	16
nb uops	19
loop length	70
used x86 registers	6
used mmx registers	0
used xmm registers	3
used ymm registers	0
used zmm registers	0
nb stack references	3

Front-end

ASSUMED MACRO FUSION FIT IN UOP CACHE

micro-operation queue	4.75 cycles
front end	4.75 cycles

Back-end

	P0	P1	P2	P3	P4	P5	P6	P7
uops	1.50	1.00	2.00	2.00	2.00	1.00	1.50	2.00
cycles	1.50	1.00	2.00	2.00	2.00	1.00	1.50	2.00

Execution ports to units layout:

P0 (256 bits): VPU, BRU, ALU, DIV/SQRT
P1 (256 bits): ALU, VPU
P2 (512 bits): store address, load
P3 (512 bits): store address, load
P4 (512 bits): store data
P5 (512 bits): ALU, VPU
P6: ALU, BRU
P7: store address

Cycles executing div or sqrt instructions	3.00
Longest recurrence chain latency (RecMII)	1.00

Cycles summary

Front-end	4.75
Dispatch	2.00
DIV/SQRT	3.00
Data deps.	1.00
Overall L1	4.75

Vectorization ratios

INT

all	33%
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	NA (no fma vectorizable/vectorized instructions)
other	33%

FP

all	33%
load	66%
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)

INT+FP

all	33%
load	66%
store	0%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	0%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	0%
other	33%

Vector efficiency ratios

INT

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	NA (no fma vectorizable/vectorized instructions)
other	16%

FP

all	12%
load	18%
store	6%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	6%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	6%
other	NA (no other vectorizable/vectorized instructions)

INT+FP

all	13%
load	18%
store	6%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	6%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	6%
other	16%

Cycles and memory resources usage

Assuming all data fit into the L1 cache, each iteration of the binary loop takes 4.75 cycles. At this rate:

7% of peak load performance is reached (9.26 out of 128.00 bytes loaded per cycle (GB/s @ 1GHz))
1% of peak store performance is reached (0.84 out of 64.00 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: 4e30

Instruction	Nb FU	P0	P1	P2	P3	P4	P5	P6	P7	Latency	Recip. throughput	Vectorization
VMOVSS (%RDI,%RBX,4),%XMM0	1	0	0	0.50	0.50	0	0	0	0	4-5	0.50	scal (6.3%)
VSUBSS %XMM2,%XMM0,%XMM0	1	0.50	0.50	0	0	0	0	0	0	4	0.50	scal (6.3%)
MOV %R12,%R13	1	0	0	0	0	0	0	0	0	0	0.25	N/A
MOV %RCX,%R12	1	0	0	0	0	0	0	0	0	0	0.25	scal (12.5%)
VZEROUPPER	4	0	0	0	0	0	0	0	0	0	1	vect (25.0%)
CALL 1130 <expf@plt>	2	0	0	0.33	0.33	1	0	1	0.33	0	1	N/A
VMOVAPS 0xe0(%RSP),%XMM2	1	0	0	0.50	0.50	0	0	0	0	4-5	0.50	vect (25.0%)
MOV %R12,%RCX	1	0	0	0	0	0	0	0	0	0	0.25	N/A
MOV %R13,%R12	1	0	0	0	0	0	0	0	0	0	0.25	scal (12.5%)
VMOVAPS 0xc0(%RSP),%XMM1	1	0	0	0.50	0.50	0	0	0	0	4-5	0.50	vect (25.0%)
MOV (%RSP),%RDI	1	0	0	0.50	0.50	0	0	0	0	4-5	0.50	N/A
VDIVSS %XMM1,%XMM0,%XMM0	1	1	0	0	0	0	0	0	0	11	3	scal (6.3%)
VMOVSS %XMM0,(%R13,%RBX,4)	1	0	0	0.33	0.33	1	0	0	0.33	3	1	scal (6.3%)
INC %RBX	1	0.25	0.25	0	0	0	0.25	0.25	0	1	0.25	N/A
CMP %RBX,%RCX	1	0.25	0.25	0	0	0	0.25	0.25	0	1	0.25	N/A
JNE 4e30 <main+0x2950>	1	0.50	0	0	0	0	0	0.50	0	0	0.50-1	N/A

Function	main
Source file and lines	attention_v2.cpp:55-56
Module	attention-clang-skl256

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

It is peel/tail loop of related source loop which is unrolled by 8 (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 4.75 to 3.25 cycles (1.46x 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 probably not vectorized. Only 13% of vector register length is used (average across all SSE/AVX instructions).

Details

Store and arithmetical SSE/AVX instructions are used in scalar version (process only one data element in vector registers).

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

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.

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

Type of elements and instruction set

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

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 44 bytes (11 single precision FP elements). The binary loop is storing 4 bytes (1 single precision FP elements).

Arithmetic intensity

Arithmetic intensity is 0.04 FP operations per loaded or stored byte.

General properties

nb instructions	16
nb uops	19
loop length	70
used x86 registers	6
used mmx registers	0
used xmm registers	3
used ymm registers	0
used zmm registers	0
nb stack references	3

Front-end

ASSUMED MACRO FUSION FIT IN UOP CACHE

micro-operation queue	4.75 cycles
front end	4.75 cycles

Back-end

	P0	P1	P2	P3	P4	P5	P6	P7
uops	1.50	1.00	2.00	2.00	2.00	1.00	1.50	2.00
cycles	1.50	1.00	2.00	2.00	2.00	1.00	1.50	2.00

Execution ports to units layout:

P0 (256 bits): VPU, BRU, ALU, DIV/SQRT
P1 (256 bits): ALU, VPU
P2 (512 bits): store address, load
P3 (512 bits): store address, load
P4 (512 bits): store data
P5 (512 bits): ALU, VPU
P6: ALU, BRU
P7: store address

Cycles executing div or sqrt instructions	3.00
Longest recurrence chain latency (RecMII)	1.00

Cycles summary

Front-end	4.75
Dispatch	2.00
DIV/SQRT	3.00
Data deps.	1.00
Overall L1	4.75

Vectorization ratios

INT

all	33%
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	NA (no fma vectorizable/vectorized instructions)
other	33%

FP

all	33%
load	66%
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)

INT+FP

all	33%
load	66%
store	0%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	0%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	0%
other	33%

Vector efficiency ratios

INT

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	NA (no fma vectorizable/vectorized instructions)
other	16%

FP

all	12%
load	18%
store	6%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	6%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	6%
other	NA (no other vectorizable/vectorized instructions)

INT+FP

all	13%
load	18%
store	6%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	6%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	6%
other	16%

Cycles and memory resources usage

Assuming all data fit into the L1 cache, each iteration of the binary loop takes 4.75 cycles. At this rate:

7% of peak load performance is reached (9.26 out of 128.00 bytes loaded per cycle (GB/s @ 1GHz))
1% of peak store performance is reached (0.84 out of 64.00 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: 4e30

Instruction	Nb FU	P0	P1	P2	P3	P4	P5	P6	P7	Latency	Recip. throughput	Vectorization
VMOVSS (%RDI,%RBX,4),%XMM0	1	0	0	0.50	0.50	0	0	0	0	4-5	0.50	scal (6.3%)
VSUBSS %XMM2,%XMM0,%XMM0	1	0.50	0.50	0	0	0	0	0	0	4	0.50	scal (6.3%)
MOV %R12,%R13	1	0	0	0	0	0	0	0	0	0	0.25	N/A
MOV %RCX,%R12	1	0	0	0	0	0	0	0	0	0	0.25	scal (12.5%)
VZEROUPPER	4	0	0	0	0	0	0	0	0	0	1	vect (25.0%)
CALL 1130 <expf@plt>	2	0	0	0.33	0.33	1	0	1	0.33	0	1	N/A
VMOVAPS 0xe0(%RSP),%XMM2	1	0	0	0.50	0.50	0	0	0	0	4-5	0.50	vect (25.0%)
MOV %R12,%RCX	1	0	0	0	0	0	0	0	0	0	0.25	N/A
MOV %R13,%R12	1	0	0	0	0	0	0	0	0	0	0.25	scal (12.5%)
VMOVAPS 0xc0(%RSP),%XMM1	1	0	0	0.50	0.50	0	0	0	0	4-5	0.50	vect (25.0%)
MOV (%RSP),%RDI	1	0	0	0.50	0.50	0	0	0	0	4-5	0.50	N/A
VDIVSS %XMM1,%XMM0,%XMM0	1	1	0	0	0	0	0	0	0	11	3	scal (6.3%)
VMOVSS %XMM0,(%R13,%RBX,4)	1	0	0	0.33	0.33	1	0	0	0.33	3	1	scal (6.3%)
INC %RBX	1	0.25	0.25	0	0	0	0.25	0.25	0	1	0.25	N/A
CMP %RBX,%RCX	1	0.25	0.25	0	0	0	0.25	0.25	0	1	0.25	N/A
JNE 4e30 <main+0x2950>	1	0.50	0	0	0	0	0	0.50	0	0	0.50-1	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