OV - attention-clang-skl256 - Loop 49 - attention-clang-skl256

0x4a40 VMAXSS	(%RDI,%RAX,4),%XMM1,%XMM1    [1]

0x4a45 INC	%RAX

0x4a48 CMP	%RAX,%RBX

0x4a4b JNE	4a40

/home/eoseret/Applications/llm-attention/attention_v2.cpp: 47 - 48

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

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

48:          if (S_row[idx] > max_val) max_val = S_row[idx];

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	1.00
CQA speedup if FP arith vectorized	1.00
CQA speedup if fully vectorized	16.00
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.33
Bottlenecks	micro-operation queue, P0,
Function	main
Source	attention_v2.cpp:47-48
Source loop unroll info	unrolled by 32
Source loop unroll confidence level	max
Unroll/vectorization loop type	peel/tail
Unroll factor	1
CQA cycles	1.00
CQA cycles if no scalar integer	1.00
CQA cycles if FP arith vectorized	1.00
CQA cycles if fully vectorized	0.06
Front-end cycles	1.00
P0 cycles	1.00
P1 cycles	0.75
P2 cycles	0.50
P3 cycles	0.50
P4 cycles	0.00
P5 cycles	0.75
P6 cycles	0.50
P7 cycles	0.00
DIV/SQRT cycles	0.00
Inter-iter dependencies cycles	NA
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	4.00
Nb uops	3.00
Nb loads	1.00
Nb stores	0.00
Nb stack references	0.00
FLOP/cycle	0.00
Nb FLOP add-sub	0.00
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	4.00
Bytes prefetched	0.00
Bytes loaded	4.00
Bytes stored	0.00
Stride 0	NA
Stride 1	NA
Stride n	NA
Stride unknown	NA
Stride indirect	NA
Vectorization ratio all	0.00
Vectorization ratio load	0.00
Vectorization ratio store	NA
Vectorization ratio mul	NA
Vectorization ratio add_sub	NA
Vectorization ratio fma	NA
Vectorization ratio div_sqrt	NA
Vectorization ratio other	0.00
Vector-efficiency ratio all	6.25
Vector-efficiency ratio load	6.25
Vector-efficiency ratio store	NA
Vector-efficiency ratio mul	NA
Vector-efficiency ratio add_sub	NA
Vector-efficiency ratio fma	NA
Vector-efficiency ratio div_sqrt	NA
Vector-efficiency ratio other	6.25

Metric	Value
CQA speedup if no scalar integer	1.00
CQA speedup if FP arith vectorized	1.00
CQA speedup if fully vectorized	16.00
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.33
Bottlenecks	micro-operation queue, P0,
Function	main
Source	attention_v2.cpp:47-48
Source loop unroll info	unrolled by 32
Source loop unroll confidence level	max
Unroll/vectorization loop type	peel/tail
Unroll factor	1
CQA cycles	1.00
CQA cycles if no scalar integer	1.00
CQA cycles if FP arith vectorized	1.00
CQA cycles if fully vectorized	0.06
Front-end cycles	1.00
P0 cycles	1.00
P1 cycles	0.75
P2 cycles	0.50
P3 cycles	0.50
P4 cycles	0.00
P5 cycles	0.75
P6 cycles	0.50
P7 cycles	0.00
DIV/SQRT cycles	0.00
Inter-iter dependencies cycles	NA
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	4.00
Nb uops	3.00
Nb loads	1.00
Nb stores	0.00
Nb stack references	0.00
FLOP/cycle	0.00
Nb FLOP add-sub	0.00
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	4.00
Bytes prefetched	0.00
Bytes loaded	4.00
Bytes stored	0.00
Stride 0	NA
Stride 1	NA
Stride n	NA
Stride unknown	NA
Stride indirect	NA
Vectorization ratio all	0.00
Vectorization ratio load	0.00
Vectorization ratio store	NA
Vectorization ratio mul	NA
Vectorization ratio add_sub	NA
Vectorization ratio fma	NA
Vectorization ratio div_sqrt	NA
Vectorization ratio other	0.00
Vector-efficiency ratio all	6.25
Vector-efficiency ratio load	6.25
Vector-efficiency ratio store	NA
Vector-efficiency ratio mul	NA
Vector-efficiency ratio add_sub	NA
Vector-efficiency ratio fma	NA
Vector-efficiency ratio div_sqrt	NA
Vector-efficiency ratio other	6.25

Path /

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

Function	main
Source file and lines	attention_v2.cpp:47-48
Module	attention-clang-skl256

The loop is defined in /home/eoseret/Applications/llm-attention/attention_v2.cpp:47-48.

It is peel/tail loop of related source loop which is unrolled by 32 (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

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. 16 data elements could be processed at once in vector registers.

Details

All 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

Type of elements and instruction set

1 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 does not contain any FP arithmetical operations. The binary loop is loading 4 bytes (1 single precision FP elements).

General properties

nb instructions	4
nb uops	3
loop length	13
used x86 registers	3
used mmx registers	0
used xmm registers	1
used ymm registers	0
used zmm registers	0
nb stack references	0

Front-end

ASSUMED MACRO FUSION 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
uops	1.00	0.75	0.50	0.50	0.00	0.75	0.50	0.00
cycles	1.00	0.75	0.50	0.50	0.00	0.75	0.50	0.00

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

NA

Cycles summary

Front-end	1.00
Dispatch	1.00
Overall L1	1.00

Vectorization ratios

all	0%
load	0%
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)
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	0%

Vector efficiency ratios

all	6%
load	6%
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)
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	6%

Cycles and memory resources usage

Assuming all data fit into the L1 cache, each iteration of the binary loop takes 1.00 cycles. At this rate:

3% of peak load performance is reached (4.00 out of 128.00 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 1.00 to 0.75 cycles (1.33x speedup).

ASM code

In the binary file, the address of the loop is: 4a40

Instruction	Nb FU	P0	P1	P2	P3	P5	P6	Latency	Recip. throughput	Vectorization
VMAXSS (%RDI,%RAX,4),%XMM1,%XMM1	1	0.50	0.50	0.50	0.50	0	0	4	0.50	scal (6.3%)
INC %RAX	1	0.25	0.25	0	0	0.25	0.25	1	0.25	N/A
CMP %RAX,%RBX	1	0.25	0.25	0	0	0.25	0.25	1	0.25	N/A
JNE 4a40 <main+0x2560>	1	0.50	0	0	0	0	0.50	0	0.50-1	N/A

Function	main
Source file and lines	attention_v2.cpp:47-48
Module	attention-clang-skl256

The loop is defined in /home/eoseret/Applications/llm-attention/attention_v2.cpp:47-48.

It is peel/tail loop of related source loop which is unrolled by 32 (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

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. 16 data elements could be processed at once in vector registers.

Details

All 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

Type of elements and instruction set

1 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 does not contain any FP arithmetical operations. The binary loop is loading 4 bytes (1 single precision FP elements).

General properties

nb instructions	4
nb uops	3
loop length	13
used x86 registers	3
used mmx registers	0
used xmm registers	1
used ymm registers	0
used zmm registers	0
nb stack references	0

Front-end

ASSUMED MACRO FUSION 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
uops	1.00	0.75	0.50	0.50	0.00	0.75	0.50	0.00
cycles	1.00	0.75	0.50	0.50	0.00	0.75	0.50	0.00

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

NA

Cycles summary

Front-end	1.00
Dispatch	1.00
Overall L1	1.00

Vectorization ratios

all	0%
load	0%
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)
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	0%

Vector efficiency ratios

all	6%
load	6%
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)
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	6%

Cycles and memory resources usage

Assuming all data fit into the L1 cache, each iteration of the binary loop takes 1.00 cycles. At this rate:

3% of peak load performance is reached (4.00 out of 128.00 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 1.00 to 0.75 cycles (1.33x speedup).

ASM code

In the binary file, the address of the loop is: 4a40

Instruction	Nb FU	P0	P1	P2	P3	P5	P6	Latency	Recip. throughput	Vectorization
VMAXSS (%RDI,%RAX,4),%XMM1,%XMM1	1	0.50	0.50	0.50	0.50	0	0	4	0.50	scal (6.3%)
INC %RAX	1	0.25	0.25	0	0	0.25	0.25	1	0.25	N/A
CMP %RAX,%RBX	1	0.25	0.25	0	0	0.25	0.25	1	0.25	N/A
JNE 4a40 <main+0x2560>	1	0.50	0	0	0	0	0.50	0	0.50-1	N/A

Report Configuration

Unrolling/vectorization cost

Details

Workaround

Vectorization

Details

Execution units bottlenecks

Type of elements and instruction set

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

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

Execution units bottlenecks

Type of elements and instruction set

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

General properties

Front-end

Back-end

Cycles summary

Vectorization ratios

Vector efficiency ratios

Cycles and memory resources usage

Front-end bottlenecks

ASM code