OV - attention-gcc-native - Loop 19

0x410fe0 MOVZ	W13, #0

  (18) 0x410fe4 MOVI	D19, #0

  (18) 0x410fe8 ORR	W10, WZR, W13

  (18) 0x410fec MOVZ	W0, #0

    (17) 0x410ff0 ADD	W12, W0, W3

    (17) 0x410ff4 LDR	S17, [X21, W10,UXTW #2]

    (17) 0x410ff8 ADD	W0, W0, #1

    (17) 0x410ffc ADD	W10, W10, W4

    (17) 0x411000 LDR	S18, [X26, X12,LSL #2]

    (17) 0x411004 FCVT	D17, S17

    (17) 0x411008 FCVT	D18, S18

    (17) 0x41100c FMADD	D19, D18, D17, D19

    (17) 0x411010 CMP	W5, W0

    (17) 0x411014 B.HI	410ff0

  (18) 0x411018 FCVT	S19, D19

  (18) 0x41101c ADD	W0, W13, W11

  (18) 0x411020 ADD	W13, W13, #1

  (18) 0x411024 STR	S19, [X14, X0,LSL #2]

  (18) 0x411028 CMP	W13, W4

  (18) 0x41102c B.CC	410fe4

0x411030 ADD	W1, W1, #1

0x411034 ADD	W11, W11, W4

0x411038 ADD	W3, W3, W5

0x41103c CMP	W5, W1

0x411040 B.HI	410fe0

/home/eoseret/llm-attention/attention_v2.cpp: 26 - 33

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

26:    for (unsigned int i = 0; i < M; ++i) {

27:       for (unsigned int j = 0; j < N; ++j) {  //vectorized

28:          double sum = 0.0;

29: #pragma clang loop vectorize(enable)

30:          for (unsigned int k = 0; k < K; ++k) { //vectorized

31:             sum += (double)A[i * K + k] * (double)B[k * N + j];

32:          }

33:          C[i * N + j] = alpha * static_cast<float>(sum);

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	1.00
CQA speedup if FP arith vectorized	1.00
CQA speedup if fully vectorized	8.00
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.67
Bottlenecks	P2, P3, P4, P5,
Function	main
Source	attention_v2.cpp:26-27
Source loop unroll info	NA
Source loop unroll confidence level	NA
Unroll/vectorization loop type	NA
Unroll factor	NA
CQA cycles	1.25
CQA cycles if no scalar integer	1.25
CQA cycles if FP arith vectorized	1.25
CQA cycles if fully vectorized	0.16
Front-end cycles	0.75
P0 cycles	0.50
P1 cycles	0.50
P2 cycles	1.25
P3 cycles	1.25
P4 cycles	1.25
P5 cycles	1.25
P6 cycles	0.00
P7 cycles	0.00
P8 cycles	0.00
P9 cycles	0.00
P10 cycles	0.00
P11 cycles	0.00
P12 cycles	0.00
P13 cycles	0.00
P14 cycles	0.00
DIV/SQRT cycles	0.00
Inter-iter dependencies cycles	1
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	6.00
Nb uops	6.00
Nb loads	NA
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	0.00
Bytes prefetched	0.00
Bytes loaded	0.00
Bytes stored	0.00
Stride 0	0.00
Stride 1	0.00
Stride n	0.00
Stride unknown	2.00
Stride indirect	0.00
Vectorization ratio all	0.00
Vectorization ratio load	NA
Vectorization ratio store	NA
Vectorization ratio mul	NA
Vectorization ratio add_sub	0.00
Vectorization ratio fma	NA
Vectorization ratio div_sqrt	NA
Vectorization ratio other	0.00
Vector-efficiency ratio all	12.50
Vector-efficiency ratio load	NA
Vector-efficiency ratio store	NA
Vector-efficiency ratio mul	NA
Vector-efficiency ratio add_sub	12.50
Vector-efficiency ratio fma	NA
Vector-efficiency ratio div_sqrt	NA
Vector-efficiency ratio other	12.50

Metric	Value
CQA speedup if no scalar integer	1.00
CQA speedup if FP arith vectorized	1.00
CQA speedup if fully vectorized	8.00
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.67
Bottlenecks	P2, P3, P4, P5,
Function	main
Source	attention_v2.cpp:26-27
Source loop unroll info	NA
Source loop unroll confidence level	NA
Unroll/vectorization loop type	NA
Unroll factor	NA
CQA cycles	1.25
CQA cycles if no scalar integer	1.25
CQA cycles if FP arith vectorized	1.25
CQA cycles if fully vectorized	0.16
Front-end cycles	0.75
P0 cycles	0.50
P1 cycles	0.50
P2 cycles	1.25
P3 cycles	1.25
P4 cycles	1.25
P5 cycles	1.25
P6 cycles	0.00
P7 cycles	0.00
P8 cycles	0.00
P9 cycles	0.00
P10 cycles	0.00
P11 cycles	0.00
P12 cycles	0.00
P13 cycles	0.00
P14 cycles	0.00
DIV/SQRT cycles	0.00
Inter-iter dependencies cycles	1
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	6.00
Nb uops	6.00
Nb loads	NA
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	0.00
Bytes prefetched	0.00
Bytes loaded	0.00
Bytes stored	0.00
Stride 0	0.00
Stride 1	0.00
Stride n	0.00
Stride unknown	2.00
Stride indirect	0.00
Vectorization ratio all	0.00
Vectorization ratio load	NA
Vectorization ratio store	NA
Vectorization ratio mul	NA
Vectorization ratio add_sub	0.00
Vectorization ratio fma	NA
Vectorization ratio div_sqrt	NA
Vectorization ratio other	0.00
Vector-efficiency ratio all	12.50
Vector-efficiency ratio load	NA
Vector-efficiency ratio store	NA
Vector-efficiency ratio mul	NA
Vector-efficiency ratio add_sub	12.50
Vector-efficiency ratio fma	NA
Vector-efficiency ratio div_sqrt	NA
Vector-efficiency ratio other	12.50

Path /

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

Function	main
Source file and lines	attention_v2.cpp:26-33
Module	attention-gcc-native

The loop is defined in /home/eoseret/llm-attention/attention_v2.cpp:26-33.

Analyzed code is defined in /home/eoseret/llm-attention/attention_v2.cpp:26-27.

Warnings:
Non-innermost loop: analyzing only self part (ignoring child loops).

gain
potential
hint
expert

Vectorization

Your loop is not vectorized. 8 data elements could be processed at once in vector registers. By vectorizing your loop, you can lower the cost of an iteration from 1.25 to 0.16 cycles (8.00x 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)

No data for this section

Slow data structures access

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

Details

Constant unknown stride: 2 occurrence(s)

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

Workaround

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

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 does not load or store any data.

General properties

nb instructions	6
nb uops	6
loop length	24
used w registers	6
used x registers	0
used b registers	0
used h registers	0
used s registers	0
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	0.75 cycles
front end	0.75 cycles

Back-end

	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11	P12	P13	P14
uops	0.50	0.50	1.25	1.25	1.25	1.25	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00
cycles	0.50	0.50	1.25	1.25	1.25	1.25	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00

Execution ports to units layout:

P0: BRU
P1: BRU
P2: ALU
P3: ALU
P4: ALU
P5: ALU
P6 (128 bits): VPU, FP store data, ALU, DIV/SQRT
P7 (128 bits): VPU, ALU, FP store data
P8 (128 bits): VPU, ALU, DIV/SQRT
P9 (128 bits): ALU, VPU
P10 (256 bits): store address, load
P11 (256 bits): store address, load
P12 (256 bits): load
P13 (64 bits): store data
P14 (64 bits): store data

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

Cycles summary

Front-end	0.75
Dispatch	1.25
Data deps.	1.00
Overall L1	1.25

Vectorization ratios

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

Vector efficiency ratios

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

Cycles and memory resources usage

Assuming all data fit into the L1 cache, each iteration of the binary loop takes 1.25 cycles.

Front-end bottlenecks

Found no such bottlenecks.

ASM code

In the binary file, the address of the loop is: 410fe0

Instruction	Nb FU	P0	P1	P2	P3	P4	P5	Latency	Recip. throughput	Vectorization
MOVZ W13, #0	1	0	0	0.25	0.25	0.25	0.25	1	0.25	scal (12.5%)
ADD W1, W1, #1	1	0	0	0.25	0.25	0.25	0.25	1	0.25	N/A
ADD W11, W11, W4	1	0	0	0.25	0.25	0.25	0.25	1	0.25	scal (12.5%)
ADD W3, W3, W5	1	0	0	0.25	0.25	0.25	0.25	1	0.25	scal (12.5%)
CMP W5, W1	1	0	0	0.25	0.25	0.25	0.25	1	0.33	N/A
B.HI 410fe0 <main+0xe20>	1	0.50	0.50	0	0	0	0	1	0.50	N/A

Function	main
Source file and lines	attention_v2.cpp:26-33
Module	attention-gcc-native

The loop is defined in /home/eoseret/llm-attention/attention_v2.cpp:26-33.

Analyzed code is defined in /home/eoseret/llm-attention/attention_v2.cpp:26-27.

Warnings:
Non-innermost loop: analyzing only self part (ignoring child loops).

gain
potential
hint
expert

Vectorization

Your loop is not vectorized. 8 data elements could be processed at once in vector registers. By vectorizing your loop, you can lower the cost of an iteration from 1.25 to 0.16 cycles (8.00x 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)

No data for this section

Slow data structures access

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

Details

Constant unknown stride: 2 occurrence(s)

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

Workaround

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

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 does not load or store any data.

General properties

nb instructions	6
nb uops	6
loop length	24
used w registers	6
used x registers	0
used b registers	0
used h registers	0
used s registers	0
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	0.75 cycles
front end	0.75 cycles

Back-end

	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11	P12	P13	P14
uops	0.50	0.50	1.25	1.25	1.25	1.25	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00
cycles	0.50	0.50	1.25	1.25	1.25	1.25	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00

Execution ports to units layout:

P0: BRU
P1: BRU
P2: ALU
P3: ALU
P4: ALU
P5: ALU
P6 (128 bits): VPU, FP store data, ALU, DIV/SQRT
P7 (128 bits): VPU, ALU, FP store data
P8 (128 bits): VPU, ALU, DIV/SQRT
P9 (128 bits): ALU, VPU
P10 (256 bits): store address, load
P11 (256 bits): store address, load
P12 (256 bits): load
P13 (64 bits): store data
P14 (64 bits): store data

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

Cycles summary

Front-end	0.75
Dispatch	1.25
Data deps.	1.00
Overall L1	1.25

Vectorization ratios

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

Vector efficiency ratios

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

Cycles and memory resources usage

Assuming all data fit into the L1 cache, each iteration of the binary loop takes 1.25 cycles.

Front-end bottlenecks

Found no such bottlenecks.

ASM code

In the binary file, the address of the loop is: 410fe0

Instruction	Nb FU	P0	P1	P2	P3	P4	P5	Latency	Recip. throughput	Vectorization
MOVZ W13, #0	1	0	0	0.25	0.25	0.25	0.25	1	0.25	scal (12.5%)
ADD W1, W1, #1	1	0	0	0.25	0.25	0.25	0.25	1	0.25	N/A
ADD W11, W11, W4	1	0	0	0.25	0.25	0.25	0.25	1	0.25	scal (12.5%)
ADD W3, W3, W5	1	0	0	0.25	0.25	0.25	0.25	1	0.25	scal (12.5%)
CMP W5, W1	1	0	0	0.25	0.25	0.25	0.25	1	0.33	N/A
B.HI 410fe0 <main+0xe20>	1	0.50	0.50	0	0	0	0	1	0.50	N/A

Report Configuration

Vectorization

Details

Workaround

Slow data structures access

Details

Workaround

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

Vectorization

Details

Workaround

Slow data structures access

Details

Workaround

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