OV - attention-armclang-native - Loop 63 - attention-armclang-native

0x12600 LDR	X10, [SP, #584]

0x12604 ADD	W8, W8, #1

0x12608 ADD	W9, W9, W24

0x1260c CMP	W8, W10

0x12610 B.EQ	126cc

0x12614 MUL	W11, W8, W24

0x12618 ORR	X10, XZR, XZR

0x1261c B	12660

  (64) 0x12620 MOVI	D0, #0

  (64) 0x12624 ADD	W13, W11, W12

  (64) 0x12628 MADD	W12, W12, W24, W10

  (64) 0x1262c LDR	S1, [X19, W13,UXTW #2]

  (64) 0x12630 LDR	S2, [X25, W12,UXTW #2]

  (64) 0x12634 FCVT	D1, S1

  (64) 0x12638 FCVT	D2, S2

  (64) 0x1263c FMADD	D0, D1, D2, D0

  (64) 0x12640 LDR	X13, [SP, #568]

  (64) 0x12644 FCVT	S0, D0

  (64) 0x12648 ADD	W12, W11, W10

  (64) 0x1264c ADD	X10, X10, #1

  (64) 0x12650 STR	S0, [X13, W12,UXTW #2]

  (64) 0x12654 AND	X12, X24, #0x7fffffff

  (64) 0x12658 CMP	X10, X12

  (64) 0x1265c B.EQ	12600

  (64) 0x12660 LDUR	X12, [X29, #472]

  (64) 0x12664 CBZ	X12, 12620

  (64) 0x12668 MOVI	D0, #0

  (64) 0x1266c AND	X16, X24, #0x0

  (64) 0x12670 ORR	X12, XZR, XZR

  (64) 0x12674 ORR	X13, XZR, X10

  (64) 0x12678 ORR	W14, WZR, W9

  (64) 0x1267c HINT	#0

    (65) 0x12680 SUB	W15, W14, #1

    (65) 0x12684 LDR	S2, [X25, W13,UXTW #2]

    (65) 0x12688 ADD	X12, X12, #2

    (65) 0x1268c LDR	S1, [X19, W15,UXTW #2]

    (65) 0x12690 ADD	W15, W24, W13

    (65) 0x12694 ADD	X13, X13, X20

    (65) 0x12698 FCVT	D2, S2

    (65) 0x1269c FCVT	D1, S1

    (65) 0x126a0 FMADD	D0, D1, D2, D0

    (65) 0x126a4 LDR	S1, [X19, W14,UXTW #2]

    (65) 0x126a8 LDR	S2, [X25, W15,UXTW #2]

    (65) 0x126ac ADD	W14, W14, #2

    (65) 0x126b0 FCVT	D1, S1

    (65) 0x126b4 FCVT	D2, S2

    (65) 0x126b8 FMADD	D0, D1, D2, D0

    (65) 0x126bc CMP	X16, X12

    (65) 0x126c0 B.NE	12680

  (64) 0x126c4 TBNZ	W24, #0, 12624

  (64) 0x126c8 B	12640

/home/eoseret/llm-attention/attention_v2.cpp: 26 - 194

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

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);

[...]

194:       start = std::chrono::steady_clock::now();

Coverage (%)	Name	Module
►100.00+	__libc_start_call_main	libc.so.6
○	__libc_start_main	libc.so.6
○	_start	attention-armclang-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	4.00
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.20
Bottlenecks	micro-operation queue, P0, P1,
Function	main
Source	attention_v2.cpp:26-26,attention_v2.cpp:194-194
Source loop unroll info	NA
Source loop unroll confidence level	NA
Unroll/vectorization loop type	NA
Unroll factor	NA
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	0.25
Front-end cycles	1.00
P0 cycles	1.00
P1 cycles	1.00
P2 cycles	0.83
P3 cycles	0.83
P4 cycles	0.83
P5 cycles	0.83
P6 cycles	0.83
P7 cycles	0.83
P8 cycles	0.00
P9 cycles	0.00
P10 cycles	0.00
P11 cycles	0.00
P12 cycles	0.33
P13 cycles	0.33
P14 cycles	0.33
P15 cycles	0.00
P16 cycles	0.00
DIV/SQRT cycles	0.00
Inter-iter dependencies cycles	NA
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	8.00
Nb uops	8.00
Nb loads	NA
Nb stores	0.00
Nb stack references	1.00
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	NA
Stride 1	NA
Stride n	NA
Stride unknown	NA
Stride indirect	NA
Vectorization ratio all	0.00
Vectorization ratio load	NA
Vectorization ratio store	NA
Vectorization ratio mul	0.00
Vectorization ratio add_sub	0.00
Vectorization ratio fma	NA
Vectorization ratio div_sqrt	NA
Vectorization ratio other	NA
Vector-efficiency ratio all	25.00
Vector-efficiency ratio load	NA
Vector-efficiency ratio store	NA
Vector-efficiency ratio mul	25.00
Vector-efficiency ratio add_sub	25.00
Vector-efficiency ratio fma	NA
Vector-efficiency ratio div_sqrt	NA
Vector-efficiency ratio other	NA

Metric	Value
CQA speedup if no scalar integer	1.00
CQA speedup if FP arith vectorized	1.00
CQA speedup if fully vectorized	4.00
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.20
Bottlenecks	micro-operation queue, P0, P1,
Function	main
Source	attention_v2.cpp:26-26,attention_v2.cpp:194-194
Source loop unroll info	NA
Source loop unroll confidence level	NA
Unroll/vectorization loop type	NA
Unroll factor	NA
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	0.25
Front-end cycles	1.00
P0 cycles	1.00
P1 cycles	1.00
P2 cycles	0.83
P3 cycles	0.83
P4 cycles	0.83
P5 cycles	0.83
P6 cycles	0.83
P7 cycles	0.83
P8 cycles	0.00
P9 cycles	0.00
P10 cycles	0.00
P11 cycles	0.00
P12 cycles	0.33
P13 cycles	0.33
P14 cycles	0.33
P15 cycles	0.00
P16 cycles	0.00
DIV/SQRT cycles	0.00
Inter-iter dependencies cycles	NA
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	8.00
Nb uops	8.00
Nb loads	NA
Nb stores	0.00
Nb stack references	1.00
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	NA
Stride 1	NA
Stride n	NA
Stride unknown	NA
Stride indirect	NA
Vectorization ratio all	0.00
Vectorization ratio load	NA
Vectorization ratio store	NA
Vectorization ratio mul	0.00
Vectorization ratio add_sub	0.00
Vectorization ratio fma	NA
Vectorization ratio div_sqrt	NA
Vectorization ratio other	NA
Vector-efficiency ratio all	25.00
Vector-efficiency ratio load	NA
Vector-efficiency ratio store	NA
Vector-efficiency ratio mul	25.00
Vector-efficiency ratio add_sub	25.00
Vector-efficiency ratio fma	NA
Vector-efficiency ratio div_sqrt	NA
Vector-efficiency ratio other	NA

Path /

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

Function	main
Source file and lines	attention_v2.cpp:26-194
Module	attention-armclang-native

The loop is defined in /home/eoseret/llm-attention/attention_v2.cpp:26-33,194.

Analyzed code is defined in /home/eoseret/llm-attention/attention_v2.cpp:26,194.

Warnings:

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

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 1.00 to 0.25 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:
- recompile with ffast-math (included in Ofast) to extend loop vectorization to FP reductions.
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

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	8
nb uops	8
loop length	32
used w registers	5
used x registers	2
used b registers	0
used h registers	0
used s registers	1
used d registers	0
used q registers	0
used v registers	0
used z registers	0
nb stack references	1

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	0.83	0.83	0.83	0.83	0.83	0.83	0.00	0.00	0.00	0.00	0.33	0.33	0.33	0.00	0.00
cycles	1.00	1.00	0.83	0.83	0.83	0.83	0.83	0.83	0.00	0.00	0.00	0.00	0.33	0.33	0.33	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

NA

Cycles summary

Front-end	1.00
Dispatch	1.00
Overall L1	1.00

Vectorization ratios

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

Vector efficiency ratios

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

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 0.83 cycles (1.20x speedup).

ASM code

In the binary file, the address of the loop is: 12614

Instruction	Nb FU	P0	P1	P2	P3	P4	P5	P6	P7	P12	P13	P14	Latency	Recip. throughput	Vectorization
LDR X10, [SP, #584]	1	0	0	0	0	0	0	0	0	0.33	0.33	0.33	4	0.33	N/A
ADD W8, W8, #1	1	0	0	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	1	0.17	N/A
ADD W9, W9, W24	1	0	0	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	1	0.17	scal (25.0%)
CMP W8, W10	1	0	0	0.25	0.25	0	0	0.25	0.25	0	0	0	1	0.33	N/A
B.EQ 126cc <main+0x14ac>	1	0.50	0.50	0	0	0	0	0	0	0	0	0	1	0.50	N/A
MUL W11, W8, W24	1	0	0	0	0	0	0	0.50	0.50	0	0	0	2	0.50	scal (25.0%)
ORR X10, XZR, XZR	1	0	0	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	1	0.17	N/A
B 12660 <main+0x1440>	1	0.50	0.50	0	0	0	0	0	0	0	0	0	1	0.50	N/A

Function	main
Source file and lines	attention_v2.cpp:26-194
Module	attention-armclang-native

The loop is defined in /home/eoseret/llm-attention/attention_v2.cpp:26-33,194.

Analyzed code is defined in /home/eoseret/llm-attention/attention_v2.cpp:26,194.

Warnings:

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

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 1.00 to 0.25 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:
- recompile with ffast-math (included in Ofast) to extend loop vectorization to FP reductions.
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

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	8
nb uops	8
loop length	32
used w registers	5
used x registers	2
used b registers	0
used h registers	0
used s registers	1
used d registers	0
used q registers	0
used v registers	0
used z registers	0
nb stack references	1

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	0.83	0.83	0.83	0.83	0.83	0.83	0.00	0.00	0.00	0.00	0.33	0.33	0.33	0.00	0.00
cycles	1.00	1.00	0.83	0.83	0.83	0.83	0.83	0.83	0.00	0.00	0.00	0.00	0.33	0.33	0.33	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

NA

Cycles summary

Front-end	1.00
Dispatch	1.00
Overall L1	1.00

Vectorization ratios

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

Vector efficiency ratios

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

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 0.83 cycles (1.20x speedup).

ASM code

In the binary file, the address of the loop is: 12614

Instruction	Nb FU	P0	P1	P2	P3	P4	P5	P6	P7	P12	P13	P14	Latency	Recip. throughput	Vectorization
LDR X10, [SP, #584]	1	0	0	0	0	0	0	0	0	0.33	0.33	0.33	4	0.33	N/A
ADD W8, W8, #1	1	0	0	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	1	0.17	N/A
ADD W9, W9, W24	1	0	0	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	1	0.17	scal (25.0%)
CMP W8, W10	1	0	0	0.25	0.25	0	0	0.25	0.25	0	0	0	1	0.33	N/A
B.EQ 126cc <main+0x14ac>	1	0.50	0.50	0	0	0	0	0	0	0	0	0	1	0.50	N/A
MUL W11, W8, W24	1	0	0	0	0	0	0	0.50	0.50	0	0	0	2	0.50	scal (25.0%)
ORR X10, XZR, XZR	1	0	0	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	1	0.17	N/A
B 12660 <main+0x1440>	1	0.50	0.50	0	0	0	0	0	0	0	0	0	1	0.50	N/A

Report Configuration

Vectorization

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

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