OV - attention-aocc-znver5-512 - Loop 62 - attention-aocc-znver5-512

	Loop Id: 62	Module: attention-aocc-znver5-512	Source: attention_v2.cpp:26-33	Coverage: 0.30%

0x6070 MOV	0xa8(%RSP),%R9

0x6078 VCVTSD2SS	%XMM0,%XMM0,%XMM0

0x607c LEA	(%RSI,%RDI,1),%R8D

0x6080 INC	%RDI

0x6083 VMOVSS	%XMM0,(%R9,%R8,4)

0x6089 CMP	%R15,%RDI

0x608c JE	6020

0x608e VXORPS	%XMM0,%XMM0,%XMM0

0x6092 CMP	$0x4,%R15D

0x6096 JB	60ac

0x6098 MOV	%EDI,%R8D

0x609b ADD	0xb8(%RSP),%R8D

0x60a3 SETB	%R8B

0x60a7 OR	%DL,%R8B

0x60aa JE	60f0

0x60ac XOR	%R8D,%R8D

0x60af MOV	%R14,%R9

0x60b2 IMUL	%R8,%R9

0x60b6 ADD	%RDI,%R9

0x60b9 NOPL	(%RAX)

  (32) 0x60c0 LEA	(%RAX,%R8,1),%R10D

  (32) 0x60c4 INC	%R8

  (32) 0x60c7 VMOVSS	(%RBX,%R10,4),%XMM1

  (32) 0x60cd MOV	%R9D,%R10D

  (32) 0x60d0 VMOVSS	(%R12,%R10,4),%XMM2

  (32) 0x60d6 ADD	%R14,%R9

  (32) 0x60d9 VCVTSS2SD	%XMM1,%XMM1,%XMM1

  (32) 0x60dd VCVTSS2SD	%XMM2,%XMM2,%XMM2

  (32) 0x60e1 VFMADD231SD	%XMM2,%XMM1,%XMM0

  (32) 0x60e6 CMP	%R8,%R15

  (32) 0x60e9 JNE	60c0

0x60eb JMP	6070

0x60f0 CMP	$0x20,%R15D

0x60f4 JAE	60fe

0x60f6 XOR	%R9D,%R9D

0x60f9 JMP	61df

0x60fe MOV	0xa0(%RSP),%R10

0x6106 VXORPS	%XMM0,%XMM0,%XMM0

0x610a VXORPD	%XMM1,%XMM1,%XMM1

0x610e VXORPD	%XMM2,%XMM2,%XMM2

0x6112 VXORPD	%XMM3,%XMM3,%XMM3

0x6116 XOR	%R8D,%R8D

0x6119 NOPL	(%RAX)

  (63) 0x6120 LEA	(%RAX,%R8,1),%R9D

  (63) 0x6124 VCVTPS2PD	(%RBX,%R9,4),%ZMM4

  (63) 0x612b VCVTPS2PD	0x20(%RBX,%R9,4),%ZMM5

  (63) 0x6133 VCVTPS2PD	0x40(%RBX,%R9,4),%ZMM6

  (63) 0x613b VCVTPS2PD	0x60(%RBX,%R9,4),%ZMM7

  (63) 0x6143 LEA	(%RDI,%R8,1),%R9D

  (63) 0x6147 ADD	$0x20,%R8

  (63) 0x614b VCVTPS2PD	(%R12,%R9,4),%ZMM8

  (63) 0x6152 VCVTPS2PD	0x20(%R12,%R9,4),%ZMM9

  (63) 0x615a VFMADD231PD	%ZMM8,%ZMM4,%ZMM0

  (63) 0x6160 VCVTPS2PD	0x40(%R12,%R9,4),%ZMM8

  (63) 0x6168 VFMADD231PD	%ZMM9,%ZMM5,%ZMM1

  (63) 0x616e VCVTPS2PD	0x60(%R12,%R9,4),%ZMM5

  (63) 0x6176 VFMADD231PD	%ZMM8,%ZMM6,%ZMM2

  (63) 0x617c VFMADD231PD	%ZMM5,%ZMM7,%ZMM3

  (63) 0x6182 CMP	%R8,%R10

  (63) 0x6185 JNE	6120

0x6187 VADDPD	%ZMM0,%ZMM1,%ZMM0

0x618d VADDPD	%ZMM2,%ZMM3,%ZMM2

0x6193 MOV	0x220(%RSP),%R10

0x619b VADDPD	%ZMM0,%ZMM2,%ZMM0

0x61a1 VEXTRACTF64X4	$0x1,%ZMM0,%YMM1

0x61a8 VADDPD	%ZMM1,%ZMM0,%ZMM0

0x61ae VEXTRACTF128	$0x1,%YMM0,%XMM1

0x61b4 VADDPD	%XMM1,%XMM0,%XMM0

0x61b8 VSHUFPD	$0x1,%XMM0,%XMM0,%XMM1

0x61bd VADDSD	%XMM1,%XMM0,%XMM0

0x61c1 TEST	%R10,%R10

0x61c4 JE	6070

0x61ca MOV	0xa0(%RSP),%R8

0x61d2 MOV	%R8,%R9

0x61d5 CMP	$0x4,%R10D

0x61d9 JB	60af

0x61df MOV	0x228(%RSP),%R10

0x61e7 VMOVQ	%XMM0,%XMM0

0x61eb NOPL	(%RAX,%RAX,1)

  (64) 0x61f0 LEA	(%RAX,%R9,1),%R8D

  (64) 0x61f4 VCVTPS2PD	(%RBX,%R8,4),%YMM1

  (64) 0x61fa LEA	(%RDI,%R9,1),%R8D

  (64) 0x61fe ADD	$0x4,%R9

  (64) 0x6202 VCVTPS2PD	(%R12,%R8,4),%YMM2

  (64) 0x6208 VFMADD231PD	%YMM2,%YMM1,%YMM0

  (64) 0x620d CMP	%R9,%R10

  (64) 0x6210 JNE	61f0

0x6212 VEXTRACTF128	$0x1,%YMM0,%XMM1

0x6218 CMPQ	$0,0x230(%RSP)

0x6221 MOV	%R10,%R8

0x6224 VADDPD	%XMM1,%XMM0,%XMM0

0x6228 VSHUFPD	$0x1,%XMM0,%XMM0,%XMM1

0x622d VADDSD	%XMM1,%XMM0,%XMM0

0x6231 JNE	60af

0x6237 JMP	6070

/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

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	1.64
CQA speedup if FP arith vectorized	1.06
CQA speedup if fully vectorized	1.36
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.64
Bottlenecks	micro-operation queue,
Function	main
Source	attention_v2.cpp:26-27,attention_v2.cpp:30-30,attention_v2.cpp:33-33
Source loop unroll info	NA
Source loop unroll confidence level	NA
Unroll/vectorization loop type	NA
Unroll factor	NA
CQA cycles	7.38
CQA cycles if no scalar integer	4.50
CQA cycles if FP arith vectorized	6.97
CQA cycles if fully vectorized	5.41
Front-end cycles	7.38
P0 cycles	3.83
P1 cycles	3.83
P2 cycles	3.83
P3 cycles	3.83
P4 cycles	3.83
P5 cycles	3.83
P6 cycles	2.00
P7 cycles	2.00
P8 cycles	2.00
P9 cycles	2.00
P10 cycles	3.00
P11 cycles	3.00
P12 cycles	4.50
P13 cycles	4.50
P14 cycles	0.50
P15 cycles	0.50
DIV/SQRT cycles	0.00
Inter-iter dependencies cycles	NA
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	59.00
Nb uops	59.00
Nb loads	7.00
Nb stores	1.00
Nb stack references	6.00
FLOP/cycle	5.15
Nb FLOP add-sub	38.00
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	7.59
Bytes prefetched	0.00
Bytes loaded	52.00
Bytes stored	4.00
Stride 0	NA
Stride 1	NA
Stride n	NA
Stride unknown	NA
Stride indirect	NA
Vectorization ratio all	55.17
Vectorization ratio load	0.00
Vectorization ratio store	0.00
Vectorization ratio mul	NA
Vectorization ratio add_sub	75.00
Vectorization ratio fma	NA
Vectorization ratio div_sqrt	NA
Vectorization ratio other	58.82
Vector-efficiency ratio all	29.74
Vector-efficiency ratio load	12.50
Vector-efficiency ratio store	6.25
Vector-efficiency ratio mul	NA
Vector-efficiency ratio add_sub	59.38
Vector-efficiency ratio fma	NA
Vector-efficiency ratio div_sqrt	NA
Vector-efficiency ratio other	20.22

Metric	Value
CQA speedup if no scalar integer	1.64
CQA speedup if FP arith vectorized	1.06
CQA speedup if fully vectorized	1.36
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.64
Bottlenecks	micro-operation queue,
Function	main
Source	attention_v2.cpp:26-27,attention_v2.cpp:30-30,attention_v2.cpp:33-33
Source loop unroll info	NA
Source loop unroll confidence level	NA
Unroll/vectorization loop type	NA
Unroll factor	NA
CQA cycles	7.38
CQA cycles if no scalar integer	4.50
CQA cycles if FP arith vectorized	6.97
CQA cycles if fully vectorized	5.41
Front-end cycles	7.38
P0 cycles	3.83
P1 cycles	3.83
P2 cycles	3.83
P3 cycles	3.83
P4 cycles	3.83
P5 cycles	3.83
P6 cycles	2.00
P7 cycles	2.00
P8 cycles	2.00
P9 cycles	2.00
P10 cycles	3.00
P11 cycles	3.00
P12 cycles	4.50
P13 cycles	4.50
P14 cycles	0.50
P15 cycles	0.50
DIV/SQRT cycles	0.00
Inter-iter dependencies cycles	NA
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	59.00
Nb uops	59.00
Nb loads	7.00
Nb stores	1.00
Nb stack references	6.00
FLOP/cycle	5.15
Nb FLOP add-sub	38.00
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	7.59
Bytes prefetched	0.00
Bytes loaded	52.00
Bytes stored	4.00
Stride 0	NA
Stride 1	NA
Stride n	NA
Stride unknown	NA
Stride indirect	NA
Vectorization ratio all	55.17
Vectorization ratio load	0.00
Vectorization ratio store	0.00
Vectorization ratio mul	NA
Vectorization ratio add_sub	75.00
Vectorization ratio fma	NA
Vectorization ratio div_sqrt	NA
Vectorization ratio other	58.82
Vector-efficiency ratio all	29.74
Vector-efficiency ratio load	12.50
Vector-efficiency ratio store	6.25
Vector-efficiency ratio mul	NA
Vector-efficiency ratio add_sub	59.38
Vector-efficiency ratio fma	NA
Vector-efficiency ratio div_sqrt	NA
Vector-efficiency ratio other	20.22

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-aocc-znver5-512

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:26-27,30,33.

Warnings:

Non-innermost loop: analyzing only self part (ignoring child loops).
Ignoring paths for analysis
Too many paths. Rerun with max-paths=8
RecMII not computed since number of paths is unknown or > max_paths
Streams not analyzed since number of paths is unknown or > max_paths

Try to simplify control and/or increase the maximum number of paths per function/loop through the 'max-paths-nb' option.

This loop has 8 execution paths.

The presence of multiple execution paths is typically the main/first bottleneck.
Try to simplify control inside loop: ideally, try to remove all conditional expressions, for example by (if applicable):

hoisting them (moving them outside the loop)
turning them into conditional moves, MIN or MAX

Ex: if (x<0) x=0 => x = (x<0 ? 0 : x) (or MAX(0,x) after defining the corresponding macro)

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 7.38 to 4.50 cycles (1.64x 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 29% of vector register length is used (average across all SSE/AVX instructions). By fully vectorizing your loop, you can lower the cost of an iteration from 7.38 to 5.41 cycles (1.36x speedup).

Details

55% of SSE/AVX instructions are used in vector version (process two or more data elements in vector registers):

0% of SSE/AVX loads are used in vector version.
0% of SSE/AVX stores are used in vector version.
75% of SSE/AVX addition or subtraction instructions are used in vector version.
58% of SSE/AVX instructions that are not load, store, addition, subtraction nor multiply instructions are used in vector version.

Since your execution units are vector units, only a fully 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)

Execution units bottlenecks

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

No data for this section

Vector unaligned load/store instructions

Detected 3 suboptimal vector unaligned load/store instructions.

Details

VEXTRACTF128: 2 occurrences►
- /home/eoseret/llm-attention/attention_v2.cpp:30
- /home/eoseret/llm-attention/attention_v2.cpp:30
VEXTRACTF64X4: 1 occurrences►
- /home/eoseret/llm-attention/attention_v2.cpp:30

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.

Conversion instructions

Detected expensive conversion instructions.

Details

VCVTSD2SS (FP64 to FP32, scalar): 1 occurrences►
- /home/eoseret/llm-attention/attention_v2.cpp:33

Workaround

Avoid mixing data with different types. In particular, check if the type of constants is the same as array elements. Use double instead of single precision only when/where needed by numerical stability and avoid mixing precision. In C/C++, FP constants are double precision by default and must be suffixed by 'f' to make them single precision.

Type of elements and instruction set

2 SSE or AVX instructions are processing arithmetic or math operations on double precision FP elements in scalar mode (one at a time). 2 SSE or AVX instructions are processing arithmetic or math operations on double precision FP elements in vector mode (two at a time). 4 AVX-512 instructions are processing arithmetic or math operations on double 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 38 FP arithmetical operations:

38: addition or subtraction

The binary loop is loading 52 bytes (6 double precision FP elements). The binary loop is storing 4 bytes (0 double precision FP elements).

Arithmetic intensity

Arithmetic intensity is 0.68 FP operations per loaded or stored byte.

General properties

nb instructions	59
nb uops	59
loop length	277
used x86 registers	10
used mmx registers	0
used xmm registers	4
used ymm registers	2
used zmm registers	4
nb stack references	6

Front-end

MACRO FUSION NOT POSSIBLE FIT IN UOP CACHE

micro-operation queue	7.38 cycles
front end	7.38 cycles

Back-end

	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11	P12	P13	P14	P15
uops	3.83	3.83	3.83	3.83	3.83	3.83	2.00	2.00	2.00	2.00	3.00	3.00	4.50	4.50	0.50	0.50
cycles	3.83	3.83	3.83	3.83	3.83	3.83	2.00	2.00	2.00	2.00	3.00	3.00	4.50	4.50	0.50	0.50

Execution ports to units layout:

P0: ALU
P1: ALU
P2: ALU
P3: ALU, BRU
P4: ALU, BRU
P5: ALU, BRU
P6 (512 bits): store address, load
P7 (512 bits): store address, load
P8 (512 bits): store address, load
P9 (512 bits): store address, load
P10 (512 bits): VPU, DIV/SQRT
P11 (512 bits): VPU, DIV/SQRT
P12 (512 bits): VPU
P13 (512 bits): VPU
P14 (512 bits): FP store data
P15 (512 bits): FP store data

Cycles executing div or sqrt instructions

Cycles summary

Front-end	7.38
Dispatch	4.50
Overall L1	7.38

Vectorization ratios

INT

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)
other	0%

all	80%
load	NA (no load vectorizable/vectorized instructions)
store	0%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	75%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	90%

INT+FP

all	55%
load	0%
store	0%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	75%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	58%

Vector efficiency ratios

INT

all	10%
load	12%
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	9%

all	38%
load	NA (no load vectorizable/vectorized instructions)
store	6%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	59%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	26%

INT+FP

all	29%
load	12%
store	6%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	59%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	20%

Cycles and memory resources usage

Assuming all data fit into the L1 cache, each iteration of the binary loop takes 7.38 cycles. At this rate:

2% of peak load performance is reached (7.05 out of 256.00 bytes loaded per cycle (GB/s @ 1GHz))
0% of peak store performance is reached (0.54 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 7.38 to 4.50 cycles (1.64x speedup).

ASM code

In the binary file, the address of the loop is: 608e

Instruction	Nb FU	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11	P12	P13	P14	P15	Latency	Recip. throughput	Vectorization
MOV 0xa8(%RSP),%R9	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	3	0.25	N/A
VCVTSD2SS %XMM0,%XMM0,%XMM0	1	0	0	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	3	0.50	scal (12.5%)
LEA (%RSI,%RDI,1),%R8D	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	N/A
INC %RDI	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	N/A
VMOVSS %XMM0,(%R9,%R8,4)	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0.50	0.50	1	0.50	scal (6.3%)
CMP %R15,%RDI	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	scal (12.5%)
JE 6020 <main+0x16c0>	1	0	0	0	0.33	0.33	0.33	0	0	0	0	0	0	0	0	0	0	0	0.33-0.50	N/A
VXORPS %XMM0,%XMM0,%XMM0	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	vect (25.0%)
CMP $0x4,%R15D	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	scal (6.3%)
JB 60ac <main+0x174c>	1	0	0	0	0.33	0.33	0.33	0	0	0	0	0	0	0	0	0	0	0	0.33-0.50	N/A
MOV %EDI,%R8D	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.13	N/A
ADD 0xb8(%RSP),%R8D	1	0.17	0.17	0.17	0.17	0.17	0.17	0.25	0.25	0.25	0.25	0	0	0	0	0	0	1	0.25	N/A
SETB %R8B	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	N/A
OR %DL,%R8B	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	N/A
JE 60f0 <main+0x1790>	1	0	0	0	0.33	0.33	0.33	0	0	0	0	0	0	0	0	0	0	0	0.33-0.50	N/A
XOR %R8D,%R8D	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	N/A
MOV %R14,%R9	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.13	N/A
IMUL %R8,%R9	1	0.33	0.33	0.33	0	0	0	0	0	0	0	0	0	0	0	0	0	3	0.33	N/A
ADD %RDI,%R9	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	N/A
NOPL (%RAX)	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.13	N/A
JMP 6070 <main+0x1710>	1	0	0	0	0.33	0.33	0.33	0	0	0	0	0	0	0	0	0	0	0	0.50	N/A
CMP $0x20,%R15D	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	scal (6.3%)
JAE 60fe <main+0x179e>	1	0	0	0	0.33	0.33	0.33	0	0	0	0	0	0	0	0	0	0	0	0.33-0.50	N/A
XOR %R9D,%R9D	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	N/A
JMP 61df <main+0x187f>	1	0	0	0	0.33	0.33	0.33	0	0	0	0	0	0	0	0	0	0	0	0.50	N/A
MOV 0xa0(%RSP),%R10	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	3	0.25	scal (12.5%)
VXORPS %XMM0,%XMM0,%XMM0	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	vect (25.0%)
VXORPD %XMM1,%XMM1,%XMM1	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	vect (25.0%)
VXORPD %XMM2,%XMM2,%XMM2	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	vect (25.0%)
VXORPD %XMM3,%XMM3,%XMM3	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	vect (25.0%)
XOR %R8D,%R8D	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	N/A
NOPL (%RAX)	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.13	N/A
VADDPD %ZMM0,%ZMM1,%ZMM0	1	0	0	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	2	0.50	vect (100.0%)
VADDPD %ZMM2,%ZMM3,%ZMM2	1	0	0	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	2	0.50	vect (100.0%)
MOV 0x220(%RSP),%R10	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	3	0.25	scal (12.5%)
VADDPD %ZMM0,%ZMM2,%ZMM0	1	0	0	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	2	0.50	vect (100.0%)
VEXTRACTF64X4 $0x1,%ZMM0,%YMM1	1	0	0	0	0	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	4	0.25	vect (50.0%)
VADDPD %ZMM1,%ZMM0,%ZMM0	1	0	0	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	2	0.50	vect (100.0%)
VEXTRACTF128 $0x1,%YMM0,%XMM1	1	0	0	0	0	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	2	0.25	vect (25.0%)
VADDPD %XMM1,%XMM0,%XMM0	1	0	0	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	2	0.50	vect (25.0%)
VSHUFPD $0x1,%XMM0,%XMM0,%XMM1	1	0	0	0	0	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	1-2	0.25	vect (25.0%)
VADDSD %XMM1,%XMM0,%XMM0	1	0	0	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	2	0.50	scal (12.5%)
TEST %R10,%R10	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	scal (12.5%)
JE 6070 <main+0x1710>	1	0	0	0	0.33	0.33	0.33	0	0	0	0	0	0	0	0	0	0	0	0.33-0.50	N/A
MOV 0xa0(%RSP),%R8	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	3	0.25	N/A
MOV %R8,%R9	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.13	N/A
CMP $0x4,%R10D	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	scal (6.3%)
JB 60af <main+0x174f>	1	0	0	0	0.33	0.33	0.33	0	0	0	0	0	0	0	0	0	0	0	0.33-0.50	N/A
MOV 0x228(%RSP),%R10	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	3	0.25	scal (12.5%)
VMOVQ %XMM0,%XMM0	1	0	0	0	0	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	1	0.25	scal (12.5%)
NOPL (%RAX,%RAX,1)	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.13	N/A
VEXTRACTF128 $0x1,%YMM0,%XMM1	1	0	0	0	0	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	2	0.25	vect (25.0%)
CMPQ $0,0x230(%RSP)	1	0.17	0.17	0.17	0.17	0.17	0.17	0.25	0.25	0.25	0.25	0	0	0	0	0	0	1	0.25	N/A
MOV %R10,%R8	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.13	N/A
VADDPD %XMM1,%XMM0,%XMM0	1	0	0	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	2	0.50	vect (25.0%)
VSHUFPD $0x1,%XMM0,%XMM0,%XMM1	1	0	0	0	0	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	1-2	0.25	vect (25.0%)
VADDSD %XMM1,%XMM0,%XMM0	1	0	0	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	2	0.50	scal (12.5%)
JNE 60af <main+0x174f>	1	0	0	0	0.33	0.33	0.33	0	0	0	0	0	0	0	0	0	0	0	0.33-0.50	N/A
JMP 6070 <main+0x1710>	1	0	0	0	0.33	0.33	0.33	0	0	0	0	0	0	0	0	0	0	0	0.50	N/A

Function	main
Source file and lines	attention_v2.cpp:26-33
Module	attention-aocc-znver5-512

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:26-27,30,33.

Warnings:

Non-innermost loop: analyzing only self part (ignoring child loops).
Ignoring paths for analysis
Too many paths. Rerun with max-paths=8
RecMII not computed since number of paths is unknown or > max_paths
Streams not analyzed since number of paths is unknown or > max_paths

hoisting them (moving them outside the loop)
turning them into conditional moves, MIN or MAX

Ex: if (x<0) x=0 => x = (x<0 ? 0 : x) (or MAX(0,x) after defining the corresponding macro)

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 7.38 to 4.50 cycles (1.64x speedup).

Workaround

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

Vectorization

Details

55% of SSE/AVX instructions are used in vector version (process two or more data elements in vector registers):

0% of SSE/AVX loads are used in vector version.
0% of SSE/AVX stores are used in vector version.
75% of SSE/AVX addition or subtraction instructions are used in vector version.
58% of SSE/AVX instructions that are not load, store, addition, subtraction nor multiply instructions are used in vector version.

Since your execution units are vector units, only a fully 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)

Execution units bottlenecks

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

No data for this section

Vector unaligned load/store instructions

Detected 3 suboptimal vector unaligned load/store instructions.

Details

VEXTRACTF128: 2 occurrences►
- /home/eoseret/llm-attention/attention_v2.cpp:30
- /home/eoseret/llm-attention/attention_v2.cpp:30
VEXTRACTF64X4: 1 occurrences►
- /home/eoseret/llm-attention/attention_v2.cpp:30

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.

Conversion instructions

Detected expensive conversion instructions.

Details

VCVTSD2SS (FP64 to FP32, scalar): 1 occurrences►
- /home/eoseret/llm-attention/attention_v2.cpp:33

Workaround

Type of elements and instruction set

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

The binary loop is composed of 38 FP arithmetical operations:

38: addition or subtraction

The binary loop is loading 52 bytes (6 double precision FP elements). The binary loop is storing 4 bytes (0 double precision FP elements).

Arithmetic intensity

Arithmetic intensity is 0.68 FP operations per loaded or stored byte.

General properties

nb instructions	59
nb uops	59
loop length	277
used x86 registers	10
used mmx registers	0
used xmm registers	4
used ymm registers	2
used zmm registers	4
nb stack references	6

Front-end

MACRO FUSION NOT POSSIBLE FIT IN UOP CACHE

micro-operation queue	7.38 cycles
front end	7.38 cycles

Back-end

	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11	P12	P13	P14	P15
uops	3.83	3.83	3.83	3.83	3.83	3.83	2.00	2.00	2.00	2.00	3.00	3.00	4.50	4.50	0.50	0.50
cycles	3.83	3.83	3.83	3.83	3.83	3.83	2.00	2.00	2.00	2.00	3.00	3.00	4.50	4.50	0.50	0.50

Execution ports to units layout:

P0: ALU
P1: ALU
P2: ALU
P3: ALU, BRU
P4: ALU, BRU
P5: ALU, BRU
P6 (512 bits): store address, load
P7 (512 bits): store address, load
P8 (512 bits): store address, load
P9 (512 bits): store address, load
P10 (512 bits): VPU, DIV/SQRT
P11 (512 bits): VPU, DIV/SQRT
P12 (512 bits): VPU
P13 (512 bits): VPU
P14 (512 bits): FP store data
P15 (512 bits): FP store data

Cycles executing div or sqrt instructions

Cycles summary

Front-end	7.38
Dispatch	4.50
Overall L1	7.38

Vectorization ratios

INT

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)
other	0%

all	80%
load	NA (no load vectorizable/vectorized instructions)
store	0%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	75%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	90%

INT+FP

all	55%
load	0%
store	0%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	75%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	58%

Vector efficiency ratios

INT

all	10%
load	12%
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	9%

all	38%
load	NA (no load vectorizable/vectorized instructions)
store	6%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	59%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	26%

INT+FP

all	29%
load	12%
store	6%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	59%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	20%

Cycles and memory resources usage

Assuming all data fit into the L1 cache, each iteration of the binary loop takes 7.38 cycles. At this rate:

2% of peak load performance is reached (7.05 out of 256.00 bytes loaded per cycle (GB/s @ 1GHz))
0% of peak store performance is reached (0.54 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: 608e

Instruction	Nb FU	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11	P12	P13	P14	P15	Latency	Recip. throughput	Vectorization
MOV 0xa8(%RSP),%R9	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	3	0.25	N/A
VCVTSD2SS %XMM0,%XMM0,%XMM0	1	0	0	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	3	0.50	scal (12.5%)
LEA (%RSI,%RDI,1),%R8D	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	N/A
INC %RDI	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	N/A
VMOVSS %XMM0,(%R9,%R8,4)	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0.50	0.50	1	0.50	scal (6.3%)
CMP %R15,%RDI	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	scal (12.5%)
JE 6020 <main+0x16c0>	1	0	0	0	0.33	0.33	0.33	0	0	0	0	0	0	0	0	0	0	0	0.33-0.50	N/A
VXORPS %XMM0,%XMM0,%XMM0	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	vect (25.0%)
CMP $0x4,%R15D	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	scal (6.3%)
JB 60ac <main+0x174c>	1	0	0	0	0.33	0.33	0.33	0	0	0	0	0	0	0	0	0	0	0	0.33-0.50	N/A
MOV %EDI,%R8D	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.13	N/A
ADD 0xb8(%RSP),%R8D	1	0.17	0.17	0.17	0.17	0.17	0.17	0.25	0.25	0.25	0.25	0	0	0	0	0	0	1	0.25	N/A
SETB %R8B	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	N/A
OR %DL,%R8B	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	N/A
JE 60f0 <main+0x1790>	1	0	0	0	0.33	0.33	0.33	0	0	0	0	0	0	0	0	0	0	0	0.33-0.50	N/A
XOR %R8D,%R8D	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	N/A
MOV %R14,%R9	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.13	N/A
IMUL %R8,%R9	1	0.33	0.33	0.33	0	0	0	0	0	0	0	0	0	0	0	0	0	3	0.33	N/A
ADD %RDI,%R9	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	N/A
NOPL (%RAX)	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.13	N/A
JMP 6070 <main+0x1710>	1	0	0	0	0.33	0.33	0.33	0	0	0	0	0	0	0	0	0	0	0	0.50	N/A
CMP $0x20,%R15D	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	scal (6.3%)
JAE 60fe <main+0x179e>	1	0	0	0	0.33	0.33	0.33	0	0	0	0	0	0	0	0	0	0	0	0.33-0.50	N/A
XOR %R9D,%R9D	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	N/A
JMP 61df <main+0x187f>	1	0	0	0	0.33	0.33	0.33	0	0	0	0	0	0	0	0	0	0	0	0.50	N/A
MOV 0xa0(%RSP),%R10	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	3	0.25	scal (12.5%)
VXORPS %XMM0,%XMM0,%XMM0	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	vect (25.0%)
VXORPD %XMM1,%XMM1,%XMM1	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	vect (25.0%)
VXORPD %XMM2,%XMM2,%XMM2	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	vect (25.0%)
VXORPD %XMM3,%XMM3,%XMM3	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	vect (25.0%)
XOR %R8D,%R8D	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	N/A
NOPL (%RAX)	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.13	N/A
VADDPD %ZMM0,%ZMM1,%ZMM0	1	0	0	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	2	0.50	vect (100.0%)
VADDPD %ZMM2,%ZMM3,%ZMM2	1	0	0	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	2	0.50	vect (100.0%)
MOV 0x220(%RSP),%R10	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	3	0.25	scal (12.5%)
VADDPD %ZMM0,%ZMM2,%ZMM0	1	0	0	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	2	0.50	vect (100.0%)
VEXTRACTF64X4 $0x1,%ZMM0,%YMM1	1	0	0	0	0	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	4	0.25	vect (50.0%)
VADDPD %ZMM1,%ZMM0,%ZMM0	1	0	0	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	2	0.50	vect (100.0%)
VEXTRACTF128 $0x1,%YMM0,%XMM1	1	0	0	0	0	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	2	0.25	vect (25.0%)
VADDPD %XMM1,%XMM0,%XMM0	1	0	0	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	2	0.50	vect (25.0%)
VSHUFPD $0x1,%XMM0,%XMM0,%XMM1	1	0	0	0	0	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	1-2	0.25	vect (25.0%)
VADDSD %XMM1,%XMM0,%XMM0	1	0	0	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	2	0.50	scal (12.5%)
TEST %R10,%R10	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	scal (12.5%)
JE 6070 <main+0x1710>	1	0	0	0	0.33	0.33	0.33	0	0	0	0	0	0	0	0	0	0	0	0.33-0.50	N/A
MOV 0xa0(%RSP),%R8	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	3	0.25	N/A
MOV %R8,%R9	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.13	N/A
CMP $0x4,%R10D	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	scal (6.3%)
JB 60af <main+0x174f>	1	0	0	0	0.33	0.33	0.33	0	0	0	0	0	0	0	0	0	0	0	0.33-0.50	N/A
MOV 0x228(%RSP),%R10	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	3	0.25	scal (12.5%)
VMOVQ %XMM0,%XMM0	1	0	0	0	0	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	1	0.25	scal (12.5%)
NOPL (%RAX,%RAX,1)	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.13	N/A
VEXTRACTF128 $0x1,%YMM0,%XMM1	1	0	0	0	0	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	2	0.25	vect (25.0%)
CMPQ $0,0x230(%RSP)	1	0.17	0.17	0.17	0.17	0.17	0.17	0.25	0.25	0.25	0.25	0	0	0	0	0	0	1	0.25	N/A
MOV %R10,%R8	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.13	N/A
VADDPD %XMM1,%XMM0,%XMM0	1	0	0	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	2	0.50	vect (25.0%)
VSHUFPD $0x1,%XMM0,%XMM0,%XMM1	1	0	0	0	0	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	1-2	0.25	vect (25.0%)
VADDSD %XMM1,%XMM0,%XMM0	1	0	0	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	2	0.50	scal (12.5%)
JNE 60af <main+0x174f>	1	0	0	0	0.33	0.33	0.33	0	0	0	0	0	0	0	0	0	0	0	0.33-0.50	N/A
JMP 6070 <main+0x1710>	1	0	0	0	0.33	0.33	0.33	0	0	0	0	0	0	0	0	0	0	0	0.50	N/A

Report Configuration

Code clean check

Workaround

Vectorization

Details

Workaround

Execution units bottlenecks

Vector unaligned load/store instructions

Details

Workaround

Conversion 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

Vector unaligned load/store instructions

Details

Workaround

Conversion 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