OV - attention-clang-gnr256 - Loop 49 - attention-clang-gnr256

	Loop Id: 49	Module: attention-clang-gnr256	Source: attention_v2.cpp:55-56	Coverage: 2.66%

0x6045 VMOVUPS	(%R8,%R13,4),%YMM0    [2]

0x604b VSUBPS	0x3a0(%RSP),%YMM0,%YMM0    [1]

0x6054 VMOVAPS	%YMM0,0x60(%RSP)    [1]

0x605a VEXTRACTF128	$0x1,%YMM0,%XMM0

0x6060 VMOVAPS	%XMM0,0xa0(%RSP)    [1]

0x6069 VZEROUPPER

0x606c CALL	1160 <expf@plt>

0x6071 VMOVAPS	%XMM0,0x100(%RSP)    [1]

0x607a VMOVSHDUP	0xa0(%RSP),%XMM0    [1]

0x6083 CALL	1160 <expf@plt>

0x6088 VMOVAPS	0x100(%RSP),%XMM1    [1]

0x6091 VINSERTPS	$0x10,%XMM0,%XMM1,%XMM0

0x6097 VMOVAPS	%XMM0,0x100(%RSP)    [1]

0x60a0 VPERMILPD	$0x1,0xa0(%RSP),%XMM0    [1]

0x60ab CALL	1160 <expf@plt>

0x60b0 VMOVAPS	0x100(%RSP),%XMM1    [1]

0x60b9 VINSERTPS	$0x20,%XMM0,%XMM1,%XMM0

0x60bf VMOVAPS	%XMM0,0x100(%RSP)    [1]

0x60c8 VPSHUFD	$-0x1,0xa0(%RSP),%XMM0    [1]

0x60d2 CALL	1160 <expf@plt>

0x60d7 VMOVAPS	0x100(%RSP),%XMM1    [1]

0x60e0 VINSERTPS	$0x30,%XMM0,%XMM1,%XMM0

0x60e6 VMOVAPS	%XMM0,0xa0(%RSP)    [1]

0x60ef VMOVAPS	0x60(%RSP),%YMM0    [1]

0x60f5 VZEROUPPER

0x60f8 CALL	1160 <expf@plt>

0x60fd VMOVAPS	%XMM0,0x100(%RSP)    [1]

0x6106 VMOVSHDUP	0x60(%RSP),%XMM0    [1]

0x610c CALL	1160 <expf@plt>

0x6111 VMOVAPS	0x100(%RSP),%XMM1    [1]

0x611a VINSERTPS	$0x10,%XMM0,%XMM1,%XMM0

0x6120 VMOVAPS	%XMM0,0x100(%RSP)    [1]

0x6129 VPERMILPD	$0x1,0x60(%RSP),%XMM0    [1]

0x6131 CALL	1160 <expf@plt>

0x6136 VMOVAPS	0x100(%RSP),%XMM1    [1]

0x613f VINSERTPS	$0x20,%XMM0,%XMM1,%XMM0

0x6145 VMOVAPS	%XMM0,0x100(%RSP)    [1]

0x614e VPSHUFD	$-0x1,0x60(%RSP),%XMM0    [1]

0x6155 CALL	1160 <expf@plt>

0x615a MOV	0x230(%RSP),%RAX    [1]

0x6162 MOV	0x20(%RSP),%R8    [1]

0x6167 VMOVAPS	0x100(%RSP),%XMM1    [1]

0x6170 VINSERTPS	$0x30,%XMM0,%XMM1,%XMM0

0x6176 VINSERTF128	$0x1,0xa0(%RSP),%YMM0,%YMM0    [1]

0x6181 VDIVPS	0x3c0(%RSP),%YMM0,%YMM0    [1]

0x618a VMOVUPS	%YMM0,(%RAX,%R13,4)    [3]

0x6190 ADD	$0x8,%R13

0x6194 CMP	%R13,0xe0(%RSP)    [1]

0x619c JNE	6045

/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

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	1.17
CQA speedup if FP arith vectorized	1.17
CQA speedup if fully vectorized	2.10
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.17
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	main
Unroll factor	8
CQA cycles	10.50
CQA cycles if no scalar integer	9.00
CQA cycles if FP arith vectorized	9.00
CQA cycles if fully vectorized	5.00
Front-end cycles	10.50
P0 cycles	2.00
P1 cycles	3.00
P2 cycles	6.67
P3 cycles	6.67
P4 cycles	9.00
P5 cycles	9.00
P6 cycles	1.00
P7 cycles	9.00
P8 cycles	9.00
P9 cycles	9.00
P10 cycles	0.00
P11 cycles	6.67
DIV/SQRT cycles	5.00
Inter-iter dependencies cycles	1
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	49.00
Nb uops	63.00
Nb loads	20.00
Nb stores	10.00
Nb stack references	8.00
FLOP/cycle	1.52
Nb FLOP add-sub	8.00
Nb FLOP mul	0.00
Nb FLOP fma	0.00
Nb FLOP div	8.00
Nb FLOP rcp	0.00
Nb FLOP sqrt	0.00
Nb FLOP rsqrt	0.00
Bytes/cycle	52.57
Bytes prefetched	0.00
Bytes loaded	360.00
Bytes stored	192.00
Stride 0	1.00
Stride 1	2.00
Stride n	0.00
Stride unknown	0.00
Stride indirect	0.00
Vectorization ratio all	83.33
Vectorization ratio load	100.00
Vectorization ratio store	100.00
Vectorization ratio mul	NA
Vectorization ratio add_sub	100.00
Vectorization ratio fma	NA
Vectorization ratio div_sqrt	100.00
Vectorization ratio other	57.14
Vector-efficiency ratio all	25.35
Vector-efficiency ratio load	29.41
Vector-efficiency ratio store	30.00
Vector-efficiency ratio mul	NA
Vector-efficiency ratio add_sub	50.00
Vector-efficiency ratio fma	NA
Vector-efficiency ratio div_sqrt	50.00
Vector-efficiency ratio other	16.96

Metric	Value
CQA speedup if no scalar integer	1.17
CQA speedup if FP arith vectorized	1.17
CQA speedup if fully vectorized	2.10
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.17
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	main
Unroll factor	8
CQA cycles	10.50
CQA cycles if no scalar integer	9.00
CQA cycles if FP arith vectorized	9.00
CQA cycles if fully vectorized	5.00
Front-end cycles	10.50
P0 cycles	2.00
P1 cycles	3.00
P2 cycles	6.67
P3 cycles	6.67
P4 cycles	9.00
P5 cycles	9.00
P6 cycles	1.00
P7 cycles	9.00
P8 cycles	9.00
P9 cycles	9.00
P10 cycles	0.00
P11 cycles	6.67
DIV/SQRT cycles	5.00
Inter-iter dependencies cycles	1
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	49.00
Nb uops	63.00
Nb loads	20.00
Nb stores	10.00
Nb stack references	8.00
FLOP/cycle	1.52
Nb FLOP add-sub	8.00
Nb FLOP mul	0.00
Nb FLOP fma	0.00
Nb FLOP div	8.00
Nb FLOP rcp	0.00
Nb FLOP sqrt	0.00
Nb FLOP rsqrt	0.00
Bytes/cycle	52.57
Bytes prefetched	0.00
Bytes loaded	360.00
Bytes stored	192.00
Stride 0	1.00
Stride 1	2.00
Stride n	0.00
Stride unknown	0.00
Stride indirect	0.00
Vectorization ratio all	83.33
Vectorization ratio load	100.00
Vectorization ratio store	100.00
Vectorization ratio mul	NA
Vectorization ratio add_sub	100.00
Vectorization ratio fma	NA
Vectorization ratio div_sqrt	100.00
Vectorization ratio other	57.14
Vector-efficiency ratio all	25.35
Vector-efficiency ratio load	29.41
Vector-efficiency ratio store	30.00
Vector-efficiency ratio mul	NA
Vector-efficiency ratio add_sub	50.00
Vector-efficiency ratio fma	NA
Vector-efficiency ratio div_sqrt	50.00
Vector-efficiency ratio other	16.96

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-gnr256

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

It is main 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 10.50 to 9.00 cycles (1.17x speedup).

Workaround

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

Vectorization

Your loop is partially vectorized. Only 25% of vector register length is used (average across all SSE/AVX instructions).

Details

83% of SSE/AVX instructions are used in vector version (process two or more data elements in vector registers):

57% of SSE/AVX instructions that are not load, store, addition, subtraction nor multiply instructions are used in vector version.

Workaround

Read the "512-bits vectorization" report at "Potential" confidence level.

Execution units bottlenecks

Found no such bottlenecks but see expert reports for more complex bottlenecks.

512-bits vectorization

On some x86 processors supporting 512-bits vectorization, compilers are often too conservative and limit vectorization to 256 bits. Performance can then be improved by enforcing 512-bits vectorization, especially with many vectorized and high trip count loops. 512-bits vectorization performance overhead (compared to 256-bits) is generally lower on newer processors.

Workaround

Recompile with -mprefer-vector-width=512

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

VINSERTF128: 1 occurrences►
- /home/eoseret/llm-attention/attention_v2.cpp:56
VPERMILPD: 2 occurrences►
- /home/eoseret/llm-attention/attention_v2.cpp:56
- /home/eoseret/llm-attention/attention_v2.cpp:56
VPSHUFD: 2 occurrences►
- /home/eoseret/llm-attention/attention_v2.cpp:56
- /home/eoseret/llm-attention/attention_v2.cpp:56
VZEROUPPER: 2 occurrences►
- /home/eoseret/llm-attention/attention_v2.cpp:56
- /home/eoseret/llm-attention/attention_v2.cpp:56

Vector unaligned load/store instructions

Detected 2 suboptimal vector unaligned load/store instructions.

Details

VEXTRACTF128: 1 occurrences►
- /home/eoseret/llm-attention/attention_v2.cpp:56
VINSERTF128: 1 occurrences►
- /home/eoseret/llm-attention/attention_v2.cpp:56

Workaround

Pass to your compiler a micro-architecture specialization option:
- use march=native
Use vector aligned instructions:
1. align your arrays on 64 bytes boundaries: replace { void *p = malloc (size); } with { void *p; posix_memalign (&p, 64, size); }.
2. inform your compiler that your arrays are vector aligned: if array 'foo' is 64 bytes-aligned, define a pointer 'p_foo' as __builtin_assume_aligned (foo, 64) and use it instead of 'foo' in the loop.

Type of elements and instruction set

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

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

The binary loop is composed of 16 FP arithmetical operations:

8: addition or subtraction
8: divide

The binary loop is loading 360 bytes (90 single precision FP elements). The binary loop is storing 192 bytes (48 single precision FP elements).

Arithmetic intensity

Arithmetic intensity is 0.03 FP operations per loaded or stored byte.

General properties

nb instructions	49
nb uops	63
loop length	349
used x86 registers	4
used mmx registers	0
used xmm registers	2
used ymm registers	1
used zmm registers	0
nb stack references	8

Front-end

ASSUMED MACRO FUSION FIT IN UOP CACHE

micro-operation queue	10.50 cycles
front end	10.50 cycles

Back-end

	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11
uops	2.00	3.00	6.67	6.67	9.00	9.00	1.00	9.00	9.00	9.00	0.00	6.67
cycles	2.00	3.00	6.67	6.67	9.00	9.00	1.00	9.00	9.00	9.00	0.00	6.67

Execution ports to units layout:

P0 (256 bits): VPU, BRU, ALU, DIV/SQRT
P1 (256 bits): ALU, VPU
P2 (512 bits): load
P3 (512 bits): load
P4 (512 bits): store data
P5 (512 bits): ALU, VPU
P6: ALU, BRU
P7: store address
P8: store address
P9 (512 bits): store data
P10: ALU
P11 (512 bits): load

Cycles executing div or sqrt instructions	5.00
Longest recurrence chain latency (RecMII)	1.00

Cycles summary

Front-end	10.50
Dispatch	9.00
DIV/SQRT	5.00
Data deps.	1.00
Overall L1	10.50

Vectorization ratios

INT

all	100%
load	100%
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	100%

all	81%
load	100%
store	100%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	100%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	100%
other	40%

INT+FP

all	83%
load	100%
store	100%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	100%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	100%
other	57%

Vector efficiency ratios

INT

all	25%
load	25%
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	25%

all	25%
load	30%
store	30%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	50%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	50%
other	13%

INT+FP

all	25%
load	29%
store	30%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	50%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	50%
other	16%

Cycles and memory resources usage

Assuming all data fit into the L1 cache, each iteration of the binary loop takes 10.50 cycles. At this rate:

17% of peak load performance is reached (34.29 out of 192.00 bytes loaded per cycle (GB/s @ 1GHz))
14% of peak store performance is reached (18.29 out of 128.00 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 10.50 to 9.00 cycles (1.17x speedup).

ASM code

In the binary file, the address of the loop is: 6045

Instruction	Nb FU	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11	Latency	Recip. throughput	Vectorization
VMOVUPS (%R8,%R13,4),%YMM0	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	0-1	0.33	vect (50.0%)
VSUBPS 0x3a0(%RSP),%YMM0,%YMM0	1	0	0.50	0.33	0.33	0	0.50	0	0	0	0	0	0.33	3	0.50	vect (50.0%)
VMOVAPS %YMM0,0x60(%RSP)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	0-1	0.50	vect (50.0%)
VEXTRACTF128 $0x1,%YMM0,%XMM0	1	0	0	0	0	0	1	0	0	0	0	0	0	3	1	vect (25.0%)
VMOVAPS %XMM0,0xa0(%RSP)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	0-1	0.50	vect (25.0%)
VZEROUPPER	2	0	0	0	0	0	0	0	0	0	0	0	0	0	1	vect (25.0%)
CALL 1160 <expf@plt>	2	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	0	1	N/A
VMOVAPS %XMM0,0x100(%RSP)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	0-1	0.50	vect (25.0%)
VMOVSHDUP 0xa0(%RSP),%XMM0	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	1	0.33	vect (12.5%)
CALL 1160 <expf@plt>	2	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	0	1	N/A
VMOVAPS 0x100(%RSP),%XMM1	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	0-1	0.33	vect (25.0%)
VINSERTPS $0x10,%XMM0,%XMM1,%XMM0	1	0	0	0	0	0	1	0	0	0	0	0	0	1	1	scal (6.3%)
VMOVAPS %XMM0,0x100(%RSP)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	0-1	0.50	vect (25.0%)
VPERMILPD $0x1,0xa0(%RSP),%XMM0	2	0	0	0.33	0.33	0	1	0	0	0	0	0	0.33	1	1	vect (25.0%)
CALL 1160 <expf@plt>	2	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	0	1	N/A
VMOVAPS 0x100(%RSP),%XMM1	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	0-1	0.33	vect (25.0%)
VINSERTPS $0x20,%XMM0,%XMM1,%XMM0	1	0	0	0	0	0	1	0	0	0	0	0	0	1	1	scal (6.3%)
VMOVAPS %XMM0,0x100(%RSP)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	0-1	0.50	vect (25.0%)
VPSHUFD $-0x1,0xa0(%RSP),%XMM0	2	0	0.50	0.33	0.33	0	0.50	0	0	0	0	0	0.33	1	0.50	vect (25.0%)
CALL 1160 <expf@plt>	2	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	0	1	N/A
VMOVAPS 0x100(%RSP),%XMM1	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	0-1	0.33	vect (25.0%)
VINSERTPS $0x30,%XMM0,%XMM1,%XMM0	1	0	0	0	0	0	1	0	0	0	0	0	0	1	1	scal (6.3%)
VMOVAPS %XMM0,0xa0(%RSP)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	0-1	0.50	vect (25.0%)
VMOVAPS 0x60(%RSP),%YMM0	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	0-1	0.33	vect (50.0%)
VZEROUPPER	2	0	0	0	0	0	0	0	0	0	0	0	0	0	1	vect (25.0%)
CALL 1160 <expf@plt>	2	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	0	1	N/A
VMOVAPS %XMM0,0x100(%RSP)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	0-1	0.50	vect (25.0%)
VMOVSHDUP 0x60(%RSP),%XMM0	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	1	0.33	vect (12.5%)
CALL 1160 <expf@plt>	2	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	0	1	N/A
VMOVAPS 0x100(%RSP),%XMM1	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	0-1	0.33	vect (25.0%)
VINSERTPS $0x10,%XMM0,%XMM1,%XMM0	1	0	0	0	0	0	1	0	0	0	0	0	0	1	1	scal (6.3%)
VMOVAPS %XMM0,0x100(%RSP)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	0-1	0.50	vect (25.0%)
VPERMILPD $0x1,0x60(%RSP),%XMM0	2	0	0	0.33	0.33	0	1	0	0	0	0	0	0.33	1	1	vect (25.0%)
CALL 1160 <expf@plt>	2	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	0	1	N/A
VMOVAPS 0x100(%RSP),%XMM1	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	0-1	0.33	vect (25.0%)
VINSERTPS $0x20,%XMM0,%XMM1,%XMM0	1	0	0	0	0	0	1	0	0	0	0	0	0	1	1	scal (6.3%)
VMOVAPS %XMM0,0x100(%RSP)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	0-1	0.50	vect (25.0%)
VPSHUFD $-0x1,0x60(%RSP),%XMM0	2	0	0.50	0.33	0.33	0	0.50	0	0	0	0	0	0.33	1	0.50	vect (25.0%)
CALL 1160 <expf@plt>	2	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	0	1	N/A
MOV 0x230(%RSP),%RAX	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	1	0.33	N/A
MOV 0x20(%RSP),%R8	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	1	0.33	N/A
VMOVAPS 0x100(%RSP),%XMM1	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	0-1	0.33	vect (25.0%)
VINSERTPS $0x30,%XMM0,%XMM1,%XMM0	1	0	0	0	0	0	1	0	0	0	0	0	0	1	1	scal (6.3%)
VINSERTF128 $0x1,0xa0(%RSP),%YMM0,%YMM0	2	0.33	0.33	0.33	0.33	0	0.33	0	0	0	0	0	0.33	3	0.33	vect (25.0%)
VDIVPS 0x3c0(%RSP),%YMM0,%YMM0	1	1	0	0.33	0.33	0	0	0	0	0	0	0	0.33	11-12	5	vect (50.0%)
VMOVUPS %YMM0,(%RAX,%R13,4)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	0-1	0.50	vect (50.0%)
ADD $0x8,%R13	1	0	0	0	0	0	0	0	0	0	0	0	0	1	0.17	N/A
CMP %R13,0xe0(%RSP)	1	0.20	0.20	0.33	0.33	0	0.20	0.20	0	0	0	0.20	0.33	1	0.33	N/A
JNE 6045 <main+0x3335>	1	0.50	0	0	0	0	0	0.50	0	0	0	0	0	0	0.50	N/A

Function	main
Source file and lines	attention_v2.cpp:55-56
Module	attention-clang-gnr256

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

It is main 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 10.50 to 9.00 cycles (1.17x speedup).

Workaround

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

Vectorization

Your loop is partially vectorized. Only 25% of vector register length is used (average across all SSE/AVX instructions).

Details

83% of SSE/AVX instructions are used in vector version (process two or more data elements in vector registers):

57% of SSE/AVX instructions that are not load, store, addition, subtraction nor multiply instructions are used in vector version.

Workaround

Read the "512-bits vectorization" report at "Potential" confidence level.

Execution units bottlenecks

Found no such bottlenecks but see expert reports for more complex bottlenecks.

512-bits vectorization

Workaround

Recompile with -mprefer-vector-width=512

CALL instructions

Detected function call instructions.

Details

expf@plt: 8 occurrences►
- /home/eoseret/llm-attention/attention_v2.cpp:56
- /home/eoseret/llm-attention/attention_v2.cpp:56
- /home/eoseret/llm-attention/attention_v2.cpp:56
- /home/eoseret/llm-attention/attention_v2.cpp:56
- /home/eoseret/llm-attention/attention_v2.cpp:56
- /home/eoseret/llm-attention/attention_v2.cpp:56
- /home/eoseret/llm-attention/attention_v2.cpp:56
- /home/eoseret/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.

VINSERTF128: 1 occurrences►
- /home/eoseret/llm-attention/attention_v2.cpp:56
VPERMILPD: 2 occurrences►
- /home/eoseret/llm-attention/attention_v2.cpp:56
- /home/eoseret/llm-attention/attention_v2.cpp:56
VPSHUFD: 2 occurrences►
- /home/eoseret/llm-attention/attention_v2.cpp:56
- /home/eoseret/llm-attention/attention_v2.cpp:56
VZEROUPPER: 2 occurrences►
- /home/eoseret/llm-attention/attention_v2.cpp:56
- /home/eoseret/llm-attention/attention_v2.cpp:56

Vector unaligned load/store instructions

Detected 2 suboptimal vector unaligned load/store instructions.

Details

VEXTRACTF128: 1 occurrences►
- /home/eoseret/llm-attention/attention_v2.cpp:56
VINSERTF128: 1 occurrences►
- /home/eoseret/llm-attention/attention_v2.cpp:56

Workaround

Pass to your compiler a micro-architecture specialization option:
- use march=native
Use vector aligned instructions:
1. align your arrays on 64 bytes boundaries: replace { void *p = malloc (size); } with { void *p; posix_memalign (&p, 64, size); }.
2. inform your compiler that your arrays are vector aligned: if array 'foo' is 64 bytes-aligned, define a pointer 'p_foo' as __builtin_assume_aligned (foo, 64) and use it instead of 'foo' in the loop.

Type of elements and instruction set

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

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

The binary loop is composed of 16 FP arithmetical operations:

8: addition or subtraction
8: divide

The binary loop is loading 360 bytes (90 single precision FP elements). The binary loop is storing 192 bytes (48 single precision FP elements).

Arithmetic intensity

Arithmetic intensity is 0.03 FP operations per loaded or stored byte.

General properties

nb instructions	49
nb uops	63
loop length	349
used x86 registers	4
used mmx registers	0
used xmm registers	2
used ymm registers	1
used zmm registers	0
nb stack references	8

Front-end

ASSUMED MACRO FUSION FIT IN UOP CACHE

micro-operation queue	10.50 cycles
front end	10.50 cycles

Back-end

	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11
uops	2.00	3.00	6.67	6.67	9.00	9.00	1.00	9.00	9.00	9.00	0.00	6.67
cycles	2.00	3.00	6.67	6.67	9.00	9.00	1.00	9.00	9.00	9.00	0.00	6.67

Execution ports to units layout:

P0 (256 bits): VPU, BRU, ALU, DIV/SQRT
P1 (256 bits): ALU, VPU
P2 (512 bits): load
P3 (512 bits): load
P4 (512 bits): store data
P5 (512 bits): ALU, VPU
P6: ALU, BRU
P7: store address
P8: store address
P9 (512 bits): store data
P10: ALU
P11 (512 bits): load

Cycles executing div or sqrt instructions	5.00
Longest recurrence chain latency (RecMII)	1.00

Cycles summary

Front-end	10.50
Dispatch	9.00
DIV/SQRT	5.00
Data deps.	1.00
Overall L1	10.50

Vectorization ratios

INT

all	100%
load	100%
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	100%

all	81%
load	100%
store	100%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	100%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	100%
other	40%

INT+FP

all	83%
load	100%
store	100%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	100%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	100%
other	57%

Vector efficiency ratios

INT

all	25%
load	25%
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	25%

all	25%
load	30%
store	30%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	50%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	50%
other	13%

INT+FP

all	25%
load	29%
store	30%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	50%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	50%
other	16%

Cycles and memory resources usage

Assuming all data fit into the L1 cache, each iteration of the binary loop takes 10.50 cycles. At this rate:

17% of peak load performance is reached (34.29 out of 192.00 bytes loaded per cycle (GB/s @ 1GHz))
14% of peak store performance is reached (18.29 out of 128.00 bytes stored per cycle (GB/s @ 1GHz))

Front-end bottlenecks

ASM code

In the binary file, the address of the loop is: 6045

Instruction	Nb FU	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11	Latency	Recip. throughput	Vectorization
VMOVUPS (%R8,%R13,4),%YMM0	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	0-1	0.33	vect (50.0%)
VSUBPS 0x3a0(%RSP),%YMM0,%YMM0	1	0	0.50	0.33	0.33	0	0.50	0	0	0	0	0	0.33	3	0.50	vect (50.0%)
VMOVAPS %YMM0,0x60(%RSP)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	0-1	0.50	vect (50.0%)
VEXTRACTF128 $0x1,%YMM0,%XMM0	1	0	0	0	0	0	1	0	0	0	0	0	0	3	1	vect (25.0%)
VMOVAPS %XMM0,0xa0(%RSP)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	0-1	0.50	vect (25.0%)
VZEROUPPER	2	0	0	0	0	0	0	0	0	0	0	0	0	0	1	vect (25.0%)
CALL 1160 <expf@plt>	2	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	0	1	N/A
VMOVAPS %XMM0,0x100(%RSP)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	0-1	0.50	vect (25.0%)
VMOVSHDUP 0xa0(%RSP),%XMM0	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	1	0.33	vect (12.5%)
CALL 1160 <expf@plt>	2	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	0	1	N/A
VMOVAPS 0x100(%RSP),%XMM1	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	0-1	0.33	vect (25.0%)
VINSERTPS $0x10,%XMM0,%XMM1,%XMM0	1	0	0	0	0	0	1	0	0	0	0	0	0	1	1	scal (6.3%)
VMOVAPS %XMM0,0x100(%RSP)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	0-1	0.50	vect (25.0%)
VPERMILPD $0x1,0xa0(%RSP),%XMM0	2	0	0	0.33	0.33	0	1	0	0	0	0	0	0.33	1	1	vect (25.0%)
CALL 1160 <expf@plt>	2	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	0	1	N/A
VMOVAPS 0x100(%RSP),%XMM1	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	0-1	0.33	vect (25.0%)
VINSERTPS $0x20,%XMM0,%XMM1,%XMM0	1	0	0	0	0	0	1	0	0	0	0	0	0	1	1	scal (6.3%)
VMOVAPS %XMM0,0x100(%RSP)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	0-1	0.50	vect (25.0%)
VPSHUFD $-0x1,0xa0(%RSP),%XMM0	2	0	0.50	0.33	0.33	0	0.50	0	0	0	0	0	0.33	1	0.50	vect (25.0%)
CALL 1160 <expf@plt>	2	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	0	1	N/A
VMOVAPS 0x100(%RSP),%XMM1	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	0-1	0.33	vect (25.0%)
VINSERTPS $0x30,%XMM0,%XMM1,%XMM0	1	0	0	0	0	0	1	0	0	0	0	0	0	1	1	scal (6.3%)
VMOVAPS %XMM0,0xa0(%RSP)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	0-1	0.50	vect (25.0%)
VMOVAPS 0x60(%RSP),%YMM0	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	0-1	0.33	vect (50.0%)
VZEROUPPER	2	0	0	0	0	0	0	0	0	0	0	0	0	0	1	vect (25.0%)
CALL 1160 <expf@plt>	2	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	0	1	N/A
VMOVAPS %XMM0,0x100(%RSP)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	0-1	0.50	vect (25.0%)
VMOVSHDUP 0x60(%RSP),%XMM0	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	1	0.33	vect (12.5%)
CALL 1160 <expf@plt>	2	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	0	1	N/A
VMOVAPS 0x100(%RSP),%XMM1	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	0-1	0.33	vect (25.0%)
VINSERTPS $0x10,%XMM0,%XMM1,%XMM0	1	0	0	0	0	0	1	0	0	0	0	0	0	1	1	scal (6.3%)
VMOVAPS %XMM0,0x100(%RSP)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	0-1	0.50	vect (25.0%)
VPERMILPD $0x1,0x60(%RSP),%XMM0	2	0	0	0.33	0.33	0	1	0	0	0	0	0	0.33	1	1	vect (25.0%)
CALL 1160 <expf@plt>	2	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	0	1	N/A
VMOVAPS 0x100(%RSP),%XMM1	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	0-1	0.33	vect (25.0%)
VINSERTPS $0x20,%XMM0,%XMM1,%XMM0	1	0	0	0	0	0	1	0	0	0	0	0	0	1	1	scal (6.3%)
VMOVAPS %XMM0,0x100(%RSP)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	0-1	0.50	vect (25.0%)
VPSHUFD $-0x1,0x60(%RSP),%XMM0	2	0	0.50	0.33	0.33	0	0.50	0	0	0	0	0	0.33	1	0.50	vect (25.0%)
CALL 1160 <expf@plt>	2	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	0	1	N/A
MOV 0x230(%RSP),%RAX	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	1	0.33	N/A
MOV 0x20(%RSP),%R8	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	1	0.33	N/A
VMOVAPS 0x100(%RSP),%XMM1	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	0-1	0.33	vect (25.0%)
VINSERTPS $0x30,%XMM0,%XMM1,%XMM0	1	0	0	0	0	0	1	0	0	0	0	0	0	1	1	scal (6.3%)
VINSERTF128 $0x1,0xa0(%RSP),%YMM0,%YMM0	2	0.33	0.33	0.33	0.33	0	0.33	0	0	0	0	0	0.33	3	0.33	vect (25.0%)
VDIVPS 0x3c0(%RSP),%YMM0,%YMM0	1	1	0	0.33	0.33	0	0	0	0	0	0	0	0.33	11-12	5	vect (50.0%)
VMOVUPS %YMM0,(%RAX,%R13,4)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	0-1	0.50	vect (50.0%)
ADD $0x8,%R13	1	0	0	0	0	0	0	0	0	0	0	0	0	1	0.17	N/A
CMP %R13,0xe0(%RSP)	1	0.20	0.20	0.33	0.33	0	0.20	0.20	0	0	0	0.20	0.33	1	0.33	N/A
JNE 6045 <main+0x3335>	1	0.50	0	0	0	0	0	0.50	0	0	0	0	0	0	0.50	N/A

Report Configuration

Code clean check

Workaround

Vectorization

Details

Workaround

Execution units bottlenecks

512-bits vectorization

Workaround

CALL instructions

Details

Complex instructions

Details

Vector unaligned load/store instructions

Details

Workaround

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

Vectorization

Details

Workaround

Execution units bottlenecks

512-bits vectorization

Workaround

CALL instructions

Details

Complex instructions

Details

Vector unaligned load/store instructions

Details

Workaround

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