OV - attention-avx512 - Loop 42

	Loop Id: 42	Module: attention-avx512	Source: attention.cpp:27-98 [...]	Coverage: 6.30%

0x403df0 VPBROADCASTQ	%R12,%ZMM7

0x403df6 VPSUBQ	%ZMM7,%ZMM11,%ZMM8

0x403dfc VPCMPNLEUQ	%ZMM0,%ZMM8,%K1

0x403e03 VPMOVQD	%ZMM7,%YMM7

0x403e09 VPADDD	%YMM1,%YMM7,%YMM7

0x403e0d VPADDD	%YMM7,%YMM4,%YMM8

0x403e11 VPSUBD	%YMM2,%YMM8,%YMM8

0x403e15 KMOVQ	%K1,%K2

0x403e1a VXORPS	%XMM9,%XMM9,%XMM9

0x403e1f VGATHERDPS	(%R8,%YMM8,4),%YMM9{%K2}

0x403e26 VCVTPS2PD	%YMM9,%ZMM8

0x403e2c VPMULLD	%YMM7,%YMM10,%YMM7

0x403e31 VPADDD	%YMM7,%YMM6,%YMM6

0x403e35 VPSUBD	%YMM2,%YMM6,%YMM6

0x403e39 KMOVQ	%K1,%K2

0x403e3e VPXOR	%XMM7,%XMM7,%XMM7

0x403e42 VGATHERDPS	(%R9,%YMM6,4),%YMM7{%K2}

0x403e49 VCVTPS2PD	%YMM7,%ZMM6

0x403e4f VFMADD231PD	%ZMM8,%ZMM6,%ZMM5{%K1}

0x403e55 VEXTRACTF64X4	$0x1,%ZMM5,%YMM6

0x403e5c VADDPD	%ZMM6,%ZMM5,%ZMM5

0x403e62 VEXTRACTF128	$0x1,%YMM5,%XMM6

0x403e68 VADDPD	%XMM6,%XMM5,%XMM5

0x403e6c VSHUFPD	$0x1,%XMM5,%XMM5,%XMM6

0x403e71 VADDSD	%XMM6,%XMM5,%XMM5

0x403e75 VCVTSD2SS	%XMM5,%XMM5,%XMM5

0x403e79 LEA	(%R14,%R15,1),%R12D

0x403e7d MOV	0xc8(%RSP),%RAX

0x403e85 VMOVSS	%XMM5,(%RAX,%R12,4)

0x403e8b LEA	0x1(%R15),%R12

0x403e8f CMP	%RSI,%R15

0x403e92 MOV	%R12,%R15

0x403e95 JE	403dc0

0x403e9b VPBROADCASTD	%R15D,%YMM6

0x403ea1 VXORPD	%XMM5,%XMM5,%XMM5

0x403ea5 XOR	%R12D,%R12D

0x403ea8 TEST	%RDI,%RDI

0x403eab JE	403df0

0x403eb1 NOPW	%CS:(%RAX,%RAX,1)

  (43) 0x403ec0 LEA	(%R10,%R12,1),%R13D

  (43) 0x403ec4 VPBROADCASTD	%R13D,%YMM7

  (43) 0x403eca VPADDD	%YMM3,%YMM7,%YMM7

  (43) 0x403ece VPCMPEQB	%XMM0,%XMM0,%K1

  (43) 0x403ed4 VXORPD	%XMM8,%XMM8,%XMM8

  (43) 0x403ed9 VGATHERDPS	(%R8,%YMM7,4),%YMM8{%K1}

  (43) 0x403ee0 VCVTPS2PD	%YMM8,%ZMM7

  (43) 0x403ee6 VPBROADCASTD	%R12D,%YMM8

  (43) 0x403eec VPADDD	%YMM1,%YMM8,%YMM8

  (43) 0x403ef0 VPMULLD	%YMM8,%YMM10,%YMM8

  (43) 0x403ef5 VPADDD	%YMM6,%YMM8,%YMM8

  (43) 0x403ef9 VPSUBD	%YMM2,%YMM8,%YMM8

  (43) 0x403efd VPCMPEQB	%XMM0,%XMM0,%K1

  (43) 0x403f03 VXORPS	%XMM9,%XMM9,%XMM9

  (43) 0x403f08 VGATHERDPS	(%R9,%YMM8,4),%YMM9{%K1}

  (43) 0x403f0f VCVTPS2PD	%YMM9,%ZMM8

  (43) 0x403f15 VFMADD231PD	%ZMM8,%ZMM7,%ZMM5

  (43) 0x403f1b ADD	$0x8,%R12

  (43) 0x403f1f CMP	%RDI,%R12

  (43) 0x403f22 JB	403ec0

0x403f24 MOV	%RDI,%R12

0x403f27 CMP	%EDI,0x38(%RSP)

0x403f2b MOV	0x88(%RSP),%R13

0x403f33 JNE	403df0

0x403f39 JMP	403e55

/home/eoseret/llm-attention/attention.cpp: 27 - 98

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

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);

[...]

98:    if (argc < 4) {

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	1.07
CQA speedup if FP arith vectorized	1.07
CQA speedup if fully vectorized	1.84
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.03
Bottlenecks
Function	main
Source	attention.cpp:27-27,attention.cpp:30-33,attention.cpp:98-98
Source loop unroll info	NA
Source loop unroll confidence level	NA
Unroll/vectorization loop type	NA
Unroll factor	NA
CQA cycles	8.67
CQA cycles if no scalar integer	8.11
CQA cycles if FP arith vectorized	8.13
CQA cycles if fully vectorized	4.72
Front-end cycles	7.11
P0 cycles	8.26
P1 cycles	8.11
P2 cycles	4.33
P3 cycles	4.33
P4 cycles	0.50
P5 cycles	8.33
P6 cycles	2.83
P7 cycles	0.50
P8 cycles	0.50
P9 cycles	0.50
P10 cycles	2.80
P11 cycles	4.33
DIV/SQRT cycles	0.00
Inter-iter dependencies cycles	0
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	35.33
Nb uops	42.67
Nb loads	3.67
Nb stores	1.00
Nb stack references	2.33
FLOP/cycle	2.50
Nb FLOP add-sub	11.00
Nb FLOP mul	0.00
Nb FLOP fma	5.33
Nb FLOP div	0.00
Nb FLOP rcp	0.00
Nb FLOP sqrt	0.00
Nb FLOP rsqrt	0.00
Bytes/cycle	6.89
Bytes prefetched	0.00
Bytes loaded	58.67
Bytes stored	4.00
Stride 0	NA
Stride 1	NA
Stride n	NA
Stride unknown	NA
Stride indirect	NA
Vectorization ratio all	66.02
Vectorization ratio load	55.56
Vectorization ratio store	0.00
Vectorization ratio mul	100.00
Vectorization ratio add_sub	81.48
Vectorization ratio fma	100.00
Vectorization ratio div_sqrt	NA
Vectorization ratio other	65.20
Vector-efficiency ratio all	37.67
Vector-efficiency ratio load	33.33
Vector-efficiency ratio store	6.25
Vector-efficiency ratio mul	50.00
Vector-efficiency ratio add_sub	51.39
Vector-efficiency ratio fma	100.00
Vector-efficiency ratio div_sqrt	NA
Vector-efficiency ratio other	32.30

Metric	Value
CQA speedup if no scalar integer	1.07
CQA speedup if FP arith vectorized	1.07
CQA speedup if fully vectorized	1.97
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.02
Bottlenecks	P0,
Function	main
Source	attention.cpp:27-27,attention.cpp:30-33,attention.cpp:98-98
Source loop unroll info	NA
Source loop unroll confidence level	NA
Unroll/vectorization loop type	NA
Unroll factor	NA
CQA cycles	10.83
CQA cycles if no scalar integer	10.17
CQA cycles if FP arith vectorized	10.17
CQA cycles if fully vectorized	5.50
Front-end cycles	8.00
P0 cycles	10.83
P1 cycles	10.67
P2 cycles	5.67
P3 cycles	5.67
P4 cycles	0.50
P5 cycles	10.50
P6 cycles	2.00
P7 cycles	0.50
P8 cycles	0.50
P9 cycles	0.50
P10 cycles	2.00
P11 cycles	5.67
DIV/SQRT cycles	0.00
Inter-iter dependencies cycles	0
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	38.00
Nb uops	48.00
Nb loads	3.00
Nb stores	1.00
Nb stack references	1.00
FLOP/cycle	2.49
Nb FLOP add-sub	11.00
Nb FLOP mul	0.00
Nb FLOP fma	8.00
Nb FLOP div	0.00
Nb FLOP rcp	0.00
Nb FLOP sqrt	0.00
Nb FLOP rsqrt	0.00
Bytes/cycle	7.02
Bytes prefetched	0.00
Bytes loaded	72.00
Bytes stored	4.00
Stride 0	NA
Stride 1	NA
Stride n	NA
Stride unknown	NA
Stride indirect	NA
Vectorization ratio all	78.57
Vectorization ratio load	100.00
Vectorization ratio store	0.00
Vectorization ratio mul	100.00
Vectorization ratio add_sub	88.89
Vectorization ratio fma	100.00
Vectorization ratio div_sqrt	NA
Vectorization ratio other	75.00
Vector-efficiency ratio all	45.09
Vector-efficiency ratio load	50.00
Vector-efficiency ratio store	6.25
Vector-efficiency ratio mul	50.00
Vector-efficiency ratio add_sub	54.17
Vector-efficiency ratio fma	100.00
Vector-efficiency ratio div_sqrt	NA
Vector-efficiency ratio other	38.67

Metric	Value
CQA speedup if no scalar integer	1.08
CQA speedup if FP arith vectorized	1.07
CQA speedup if fully vectorized	1.56
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.08
Bottlenecks	micro-operation queue,
Function	main
Source	attention.cpp:27-27,attention.cpp:30-33,attention.cpp:98-98
Source loop unroll info	NA
Source loop unroll confidence level	NA
Unroll/vectorization loop type	NA
Unroll factor	NA
CQA cycles	4.33
CQA cycles if no scalar integer	4.00
CQA cycles if FP arith vectorized	4.05
CQA cycles if fully vectorized	2.77
Front-end cycles	4.33
P0 cycles	3.10
P1 cycles	3.00
P2 cycles	1.00
P3 cycles	1.00
P4 cycles	0.50
P5 cycles	4.00
P6 cycles	2.90
P7 cycles	0.50
P8 cycles	0.50
P9 cycles	0.50
P10 cycles	3.00
P11 cycles	1.00
DIV/SQRT cycles	0.00
Inter-iter dependencies cycles	0
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	25.00
Nb uops	26.00
Nb loads	3.00
Nb stores	1.00
Nb stack references	3.00
FLOP/cycle	2.54
Nb FLOP add-sub	11.00
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	5.54
Bytes prefetched	0.00
Bytes loaded	20.00
Bytes stored	4.00
Stride 0	0.00
Stride 1	0.00
Stride n	0.00
Stride unknown	0.00
Stride indirect	2.00
Vectorization ratio all	46.15
Vectorization ratio load	0.00
Vectorization ratio store	0.00
Vectorization ratio mul	NA
Vectorization ratio add_sub	66.67
Vectorization ratio fma	NA
Vectorization ratio div_sqrt	NA
Vectorization ratio other	50.00
Vector-efficiency ratio all	25.00
Vector-efficiency ratio load	12.50
Vector-efficiency ratio store	6.25
Vector-efficiency ratio mul	NA
Vector-efficiency ratio add_sub	45.83
Vector-efficiency ratio fma	NA
Vector-efficiency ratio div_sqrt	NA
Vector-efficiency ratio other	21.09

Metric	Value
CQA speedup if no scalar integer	1.07
CQA speedup if FP arith vectorized	1.07
CQA speedup if fully vectorized	1.84
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.02
Bottlenecks	P0,
Function	main
Source	attention.cpp:27-27,attention.cpp:30-33,attention.cpp:98-98
Source loop unroll info	NA
Source loop unroll confidence level	NA
Unroll/vectorization loop type	NA
Unroll factor	NA
CQA cycles	10.83
CQA cycles if no scalar integer	10.17
CQA cycles if FP arith vectorized	10.17
CQA cycles if fully vectorized	5.89
Front-end cycles	9.00
P0 cycles	10.83
P1 cycles	10.67
P2 cycles	6.33
P3 cycles	6.33
P4 cycles	0.50
P5 cycles	10.50
P6 cycles	3.60
P7 cycles	0.50
P8 cycles	0.50
P9 cycles	0.50
P10 cycles	3.40
P11 cycles	6.33
DIV/SQRT cycles	0.00
Inter-iter dependencies cycles	0
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	43.00
Nb uops	54.00
Nb loads	5.00
Nb stores	1.00
Nb stack references	3.00
FLOP/cycle	2.49
Nb FLOP add-sub	11.00
Nb FLOP mul	0.00
Nb FLOP fma	8.00
Nb FLOP div	0.00
Nb FLOP rcp	0.00
Nb FLOP sqrt	0.00
Nb FLOP rsqrt	0.00
Bytes/cycle	8.12
Bytes prefetched	0.00
Bytes loaded	84.00
Bytes stored	4.00
Stride 0	0.00
Stride 1	0.00
Stride n	0.00
Stride unknown	0.00
Stride indirect	2.00
Vectorization ratio all	73.33
Vectorization ratio load	66.67
Vectorization ratio store	0.00
Vectorization ratio mul	100.00
Vectorization ratio add_sub	88.89
Vectorization ratio fma	100.00
Vectorization ratio div_sqrt	NA
Vectorization ratio other	70.59
Vector-efficiency ratio all	42.92
Vector-efficiency ratio load	37.50
Vector-efficiency ratio store	6.25
Vector-efficiency ratio mul	50.00
Vector-efficiency ratio add_sub	54.17
Vector-efficiency ratio fma	100.00
Vector-efficiency ratio div_sqrt	NA
Vector-efficiency ratio other	37.13

Path /

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

Function	main
Source file and lines	attention.cpp:27-98
Module	attention-avx512

The loop is defined in /home/eoseret/llm-attention/attention.cpp:27-33,98.

Analyzed code is defined in /home/eoseret/llm-attention/attention.cpp:27,30-33,98.

Warnings:
Non-innermost loop: analyzing only self part (ignoring child loops).
This loop has 3 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

Vectorization

Your loop is partially vectorized. Only 37% 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 8.67 to 4.72 cycles (1.84x speedup).

Details

66% of SSE/AVX instructions are used in vector version (process two or more data elements in vector registers):

55% of SSE/AVX loads are used in vector version.
0% of SSE/AVX stores are used in vector version.
81% of SSE/AVX addition or subtraction instructions are used in vector version.
65% 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)

Masked instructions

Detected masked instructions.

Details

Vector registers are partially exploited, which is expected if your loop is irregular or mixes elements of different sizes.

Workaround

If your loop is irregular, try to remove or hoist conditional structures out of your loop. If it mixes elements of different sizes, try to uniformize them.

FMA

Detected 5.33 FMA (fused multiply-add) operations.

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.

VCVTPS2PD: 0 occurrences
VCVTSD2SS: 0 occurrences
VGATHERDPS: 0 occurrences
VPMULLD: 0 occurrences

Gather/scatter instructions

Detected gather/scatter instructions (typically caused by indirect accesses).

Details

VGATHERDPS: 0 occurrences

Workaround

Try to simplify your code and/or replace indirect accesses with unit-stride ones.

Conversion instructions

Detected expensive conversion instructions.

Details

VCVTPS2PD (FP32 to FP64, SIMD): 0 occurrences
VCVTSD2SS (FP64 to FP32, scalar): 0 occurrences
VPMOVQD: 0 occurrences

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

1 SSE or AVX instructions are processing arithmetic or math operations on double precision FP elements in scalar mode (one at a time). 1 SSE or AVX instructions are processing arithmetic or math operations on double precision FP elements in vector mode (two at a time). 1 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 21.666666666667 FP arithmetical operations:

16.33: addition or subtraction (5 inside FMA instructions)
5.33: multiply (all inside FMA instructions)

The binary loop is loading 58 bytes (7 double precision FP elements). The binary loop is storing 4 bytes (0 double precision FP elements).

Arithmetic intensity

Arithmetic intensity is 0.35 FP operations per loaded or stored byte.

General properties

nb instructions	35.33
nb uops	42.67
loop length	185
used x86 registers	9
used mmx registers	0
used xmm registers	3.33
used ymm registers	6.67
used zmm registers	4.67
nb stack references	2.33

Front-end

MACRO FUSION NOT POSSIBLE

micro-operation queue	7.11 cycles
front end	7.11 cycles

Back-end

	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11
uops	8.26	8.11	4.33	4.33	0.50	8.33	2.83	0.50	0.50	0.50	2.80	4.33
cycles	8.26	8.11	4.33	4.33	0.50	8.33	2.83	0.50	0.50	0.50	2.80	4.33

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	NA
Longest recurrence chain latency (RecMII)	0.00

Cycles summary

Front-end	7.11
Dispatch	8.56
Data deps.	0.00
Overall L1	8.67

Vectorization ratios

INT

all	47%
load	0%
store	NA (no store vectorizable/vectorized instructions)
mul	100%
add-sub	100%
fma	NA (no fma vectorizable/vectorized instructions)
other	30%

all	75%
load	100%
store	0%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	66%
fma	100%
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	86%

INT+FP

all	66%
load	55%
store	0%
mul	100%
add-sub	81%
fma	100%
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	65%

Vector efficiency ratios

INT

all	35%
load	12%
store	NA (no store vectorizable/vectorized instructions)
mul	50%
add-sub	58%
fma	NA (no fma vectorizable/vectorized instructions)
other	30%

all	37%
load	50%
store	6%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	45%
fma	100%
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	33%

INT+FP

all	37%
load	33%
store	6%
mul	50%
add-sub	51%
fma	100%
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	32%

Cycles and memory resources usage

Detected masked instructions: assuming all mask elements are active. Assuming all data fit into the L1 cache, each iteration of the binary loop takes 8.67 cycles. At this rate:

3% of peak load performance is reached (6.34 out of 192.00 bytes loaded per cycle (GB/s @ 1GHz))
0% of peak store performance is reached (0.55 out of 128.00 bytes stored per cycle (GB/s @ 1GHz))

Function	main
Source file and lines	attention.cpp:27-98
Module	attention-avx512

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 10.83 to 10.17 cycles (1.07x 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 45% 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 10.83 to 5.50 cycles (1.97x speedup).

Details

78% of SSE/AVX instructions are used in vector version (process two or more data elements in vector registers):

0% of SSE/AVX stores are used in vector version.
88% of SSE/AVX addition or subtraction instructions are used in vector version.
75% 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

Performance is limited by execution of FP multiply or FMA (fused multiply-add) operations (the FP multiply/FMA unit is a bottleneck).

Workaround

Reduce the number of FP multiply/FMA instructions

Masked instructions

Detected masked instructions.

Details

Vector registers are partially exploited, which is expected if your loop is irregular or mixes elements of different sizes.

Workaround

If your loop is irregular, try to remove or hoist conditional structures out of your loop. If it mixes elements of different sizes, try to uniformize them.

FMA

Detected 8 FMA (fused multiply-add) operations.

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.

VCVTPS2PD: 2 occurrences►
- /home/eoseret/llm-attention/attention.cpp:31
- /home/eoseret/llm-attention/attention.cpp:31
VCVTSD2SS: 1 occurrences►
- /home/eoseret/llm-attention/attention.cpp:33
VGATHERDPS: 2 occurrences►
- /home/eoseret/llm-attention/attention.cpp:31
- /home/eoseret/llm-attention/attention.cpp:31
VPMULLD: 1 occurrences►
- /home/eoseret/llm-attention/attention.cpp:31

Vector unaligned load/store instructions

Detected 1 suboptimal vector unaligned load/store instructions.

Details

VEXTRACTF128: 1 occurrences►
- /home/eoseret/llm-attention/attention.cpp:31

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.

Gather/scatter instructions

Detected gather/scatter instructions (typically caused by indirect accesses).

Details

VGATHERDPS: 2 occurrences►
- /home/eoseret/llm-attention/attention.cpp:31
- /home/eoseret/llm-attention/attention.cpp:31

Workaround

Try to simplify your code and/or replace indirect accesses with unit-stride ones.

Conversion instructions

Detected expensive conversion instructions.

Details

VCVTPS2PD (FP32 to FP64, SIMD): 2 occurrences►
- /home/eoseret/llm-attention/attention.cpp:31
- /home/eoseret/llm-attention/attention.cpp:31
VCVTSD2SS (FP64 to FP32, scalar): 1 occurrences►
- /home/eoseret/llm-attention/attention.cpp:33
VPMOVQD: 1 occurrences►
- /home/eoseret/llm-attention/attention.cpp:31

Workaround

Type of elements and instruction set

1 SSE or AVX instructions are processing arithmetic or math operations on double precision FP elements in scalar mode (one at a time). 1 SSE or AVX instructions are processing arithmetic or math operations on double precision FP elements in vector mode (two at a time). 2 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 27 FP arithmetical operations:

19: addition or subtraction (8 inside FMA instructions)
8: multiply (all inside FMA instructions)

The binary loop is loading 72 bytes (9 double precision FP elements). The binary loop is storing 4 bytes (0 double precision FP elements).

Arithmetic intensity

Arithmetic intensity is 0.36 FP operations per loaded or stored byte.

General properties

nb instructions	38
nb uops	48
loop length	193
used x86 registers	9
used mmx registers	0
used xmm registers	4
used ymm registers	9
used zmm registers	6
nb stack references	1

Front-end

ASSUMED MACRO FUSION FIT IN UOP CACHE

micro-operation queue	8.00 cycles
front end	8.00 cycles

Back-end

	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11
uops	10.83	10.67	5.67	5.67	0.50	10.50	2.00	0.50	0.50	0.50	2.00	5.67
cycles	10.83	10.67	5.67	5.67	0.50	10.50	2.00	0.50	0.50	0.50	2.00	5.67

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	NA
Longest recurrence chain latency (RecMII)	0.00

Cycles summary

Front-end	8.00
Dispatch	10.83
Data deps.	0.00
Overall L1	10.83

Vectorization ratios

INT

all	76%
load	NA (no load vectorizable/vectorized instructions)
store	NA (no store vectorizable/vectorized instructions)
mul	100%
add-sub	100%
fma	NA (no fma vectorizable/vectorized instructions)
other	50%

all	80%
load	100%
store	0%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	66%
fma	100%
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	90%

INT+FP

all	78%
load	100%
store	0%
mul	100%
add-sub	88%
fma	100%
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	75%

Vector efficiency ratios

INT

all	50%
load	NA (no load vectorizable/vectorized instructions)
store	NA (no store vectorizable/vectorized instructions)
mul	50%
add-sub	58%
fma	NA (no fma vectorizable/vectorized instructions)
other	42%

all	40%
load	50%
store	6%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	45%
fma	100%
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	36%

INT+FP

all	45%
load	50%
store	6%
mul	50%
add-sub	54%
fma	100%
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	38%

Cycles and memory resources usage

Detected masked instructions: assuming all mask elements are active. Assuming all data fit into the L1 cache, each iteration of the binary loop takes 10.83 cycles. At this rate:

3% of peak load performance is reached (6.65 out of 192.00 bytes loaded per cycle (GB/s @ 1GHz))
0% of peak store performance is reached (0.37 out of 128.00 bytes stored per cycle (GB/s @ 1GHz))

Front-end bottlenecks

Found no such bottlenecks.

ASM code

In the binary file, the address of the loop is: 403e9b

Instruction	Nb FU	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11	Latency	Recip. throughput	Vectorization
VPBROADCASTQ %R12,%ZMM7	1	0	0	0	0	0	1	0	0	0	0	0	0	3	1	scal (12.5%)
VPSUBQ %ZMM7,%ZMM11,%ZMM8	1	0.50	0	0	0	0	0.50	0	0	0	0	0	0	0-1	0.50	vect (100.0%)
VPCMPNLEUQ %ZMM0,%ZMM8,%K1																vect (100.0%)
VPMOVQD %ZMM7,%YMM7	1	0	0	0	0	0	1	0	0	0	0	0	0	3	1	vect (100.0%)
VPADDD %YMM1,%YMM7,%YMM7	1	0.33	0.33	0	0	0	0.33	0	0	0	0	0	0	1	0.33	vect (50.0%)
VPADDD %YMM7,%YMM4,%YMM8	1	0.33	0.33	0	0	0	0.33	0	0	0	0	0	0	1	0.33	vect (50.0%)
VPSUBD %YMM2,%YMM8,%YMM8	1	0.33	0.33	0	0	0	0.33	0	0	0	0	0	0	0-1	0.33	vect (50.0%)
KMOVQ %K1,%K2	1	1	0	0	0	0	0	0	0	0	0	0	0	0	1	N/A
VXORPS %XMM9,%XMM9,%XMM9	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	vect (25.0%)
VGATHERDPS (%R8,%YMM8,4),%YMM9{%K2}	5	1	1	2.67	2.67	0	1	0	0	0	0	0	2.67	29	3	vect (50.0%)
VCVTPS2PD %YMM9,%ZMM8	2	1	0	0	0	0	1	0	0	0	0	0	0	7	1	vect (50.0%)
VPMULLD %YMM7,%YMM10,%YMM7	2	1	1	0	0	0	0	0	0	0	0	0	0	10	1	vect (50.0%)
VPADDD %YMM7,%YMM6,%YMM6	1	0.33	0.33	0	0	0	0.33	0	0	0	0	0	0	1	0.33	vect (50.0%)
VPSUBD %YMM2,%YMM6,%YMM6	1	0.33	0.33	0	0	0	0.33	0	0	0	0	0	0	0-1	0.33	vect (50.0%)
KMOVQ %K1,%K2	1	1	0	0	0	0	0	0	0	0	0	0	0	0	1	N/A
VPXOR %XMM7,%XMM7,%XMM7	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	vect (25.0%)
VGATHERDPS (%R9,%YMM6,4),%YMM7{%K2}	5	1	1	2.67	2.67	0	1	0	0	0	0	0	2.67	29	3	vect (50.0%)
VCVTPS2PD %YMM7,%ZMM6	2	1	0	0	0	0	1	0	0	0	0	0	0	7	1	vect (50.0%)
VFMADD231PD %ZMM8,%ZMM6,%ZMM5{%K1}	1	0.50	0	0	0	0	0.50	0	0	0	0	0	0	4	0.50	vect (100.0%)
VEXTRACTF64X4 $0x1,%ZMM5,%YMM6	1	0	0	0	0	0	1	0	0	0	0	0	0	3	1	vect (50.0%)
VADDPD %ZMM6,%ZMM5,%ZMM5	1	0.50	0	0	0	0	0.50	0	0	0	0	0	0	3	0.50	vect (100.0%)
VEXTRACTF128 $0x1,%YMM5,%XMM6	1	0	0	0	0	0	1	0	0	0	0	0	0	3	1	vect (25.0%)
VADDPD %XMM6,%XMM5,%XMM5	1	0	0.50	0	0	0	0.50	0	0	0	0	0	0	3	0.50	vect (25.0%)
VSHUFPD $0x1,%XMM5,%XMM5,%XMM6	1	0	0.50	0	0	0	0.50	0	0	0	0	0	0	1	0.50	vect (25.0%)
VADDSD %XMM6,%XMM5,%XMM5	1	0	0.50	0	0	0	0.50	0	0	0	0	0	0	3	0.50	scal (12.5%)
VCVTSD2SS %XMM5,%XMM5,%XMM5	2	0.50	0.50	0	0	0	1	0	0	0	0	0	0	5	1	scal (12.5%)
LEA (%R14,%R15,1),%R12D	1	0.20	0.20	0	0	0	0.20	0.20	0	0	0	0.20	0	1-2	0.20	N/A
MOV 0xc8(%RSP),%RAX	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	1	0.33	N/A
VMOVSS %XMM5,(%RAX,%R12,4)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	1	0.50	scal (6.3%)
LEA 0x1(%R15),%R12	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	N/A
CMP %RSI,%R15	1	0.20	0.20	0	0	0	0.20	0.20	0	0	0	0.20	0	1	0.20	scal (12.5%)
MOV %R12,%R15	1	0	0	0	0	0	0	0	0	0	0	0	0	1	0.17	N/A
JE 403dc0 <main+0xb00>	1	0.50	0	0	0	0	0	0.50	0	0	0	0	0	0	0.50	N/A
VPBROADCASTD %R15D,%YMM6	1	0	0	0	0	0	1	0	0	0	0	0	0	3	1	scal (6.3%)
VXORPD %XMM5,%XMM5,%XMM5	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	vect (25.0%)
XOR %R12D,%R12D	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	N/A
TEST %RDI,%RDI	1	0.20	0.20	0	0	0	0.20	0.20	0	0	0	0.20	0	2	0.20	N/A
JE 403df0 <main+0xb30>	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.cpp:27-98
Module	attention-avx512

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 4.33 to 4.00 cycles (1.08x speedup).

Workaround

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

Vectorization

Your loop is poorly vectorized. Only 25% 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 4.33 to 2.77 cycles (1.56x speedup).

Details

46% 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.
66% of SSE/AVX addition or subtraction instructions are used in vector version.
50% 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

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.

VCVTSD2SS: 1 occurrences►
- /home/eoseret/llm-attention/attention.cpp:33

Slow data structures access

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

Details

Irregular (variable stride) or indirect: 2 occurrence(s)

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

Workaround

Try to remove indirect accesses. If applicable, precompute elements out of the innermost loop.

Vector unaligned load/store instructions

Detected 1 suboptimal vector unaligned load/store instructions.

Details

VEXTRACTF128: 1 occurrences►
- /home/eoseret/llm-attention/attention.cpp:31

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.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 11 FP arithmetical operations:

11: addition or subtraction

The binary loop is loading 20 bytes (2 double precision FP elements). The binary loop is storing 4 bytes (0 double precision FP elements).

Arithmetic intensity

Arithmetic intensity is 0.46 FP operations per loaded or stored byte.

General properties

nb instructions	25
nb uops	26
loop length	133
used x86 registers	8
used mmx registers	0
used xmm registers	2
used ymm registers	2
used zmm registers	2
nb stack references	3

Front-end

MACRO FUSION NOT POSSIBLE FIT IN UOP CACHE

micro-operation queue	4.33 cycles
front end	4.33 cycles

Back-end

	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11
uops	3.10	3.00	1.00	1.00	0.50	4.00	2.90	0.50	0.50	0.50	3.00	1.00
cycles	3.10	3.00	1.00	1.00	0.50	4.00	2.90	0.50	0.50	0.50	3.00	1.00

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	NA
Longest recurrence chain latency (RecMII)	0.00

Cycles summary

Front-end	4.33
Dispatch	4.00
Data deps.	0.00
Overall L1	4.33

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	66%
load	NA (no load vectorizable/vectorized instructions)
store	0%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	66%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	80%

INT+FP

all	46%
load	0%
store	0%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	66%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	50%

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	10%

all	31%
load	NA (no load vectorizable/vectorized instructions)
store	6%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	45%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	27%

INT+FP

all	25%
load	12%
store	6%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	45%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	21%

Cycles and memory resources usage

Assuming all data fit into the L1 cache, each iteration of the binary loop takes 4.33 cycles. At this rate:

2% of peak load performance is reached (4.62 out of 192.00 bytes loaded per cycle (GB/s @ 1GHz))
0% of peak store performance is reached (0.92 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 4.33 to 4.00 cycles (1.08x speedup).

ASM code

In the binary file, the address of the loop is: 403e9b

Instruction	Nb FU	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11	Latency	Recip. throughput	Vectorization
VEXTRACTF64X4 $0x1,%ZMM5,%YMM6	1	0	0	0	0	0	1	0	0	0	0	0	0	3	1	vect (50.0%)
VADDPD %ZMM6,%ZMM5,%ZMM5	1	0.50	0	0	0	0	0.50	0	0	0	0	0	0	3	0.50	vect (100.0%)
VEXTRACTF128 $0x1,%YMM5,%XMM6	1	0	0	0	0	0	1	0	0	0	0	0	0	3	1	vect (25.0%)
VADDPD %XMM6,%XMM5,%XMM5	1	0	0.50	0	0	0	0.50	0	0	0	0	0	0	3	0.50	vect (25.0%)
VSHUFPD $0x1,%XMM5,%XMM5,%XMM6	1	0	0.50	0	0	0	0.50	0	0	0	0	0	0	1	0.50	vect (25.0%)
VADDSD %XMM6,%XMM5,%XMM5	1	0	0.50	0	0	0	0.50	0	0	0	0	0	0	3	0.50	scal (12.5%)
VCVTSD2SS %XMM5,%XMM5,%XMM5	2	0.50	0.50	0	0	0	1	0	0	0	0	0	0	5	1	scal (12.5%)
LEA (%R14,%R15,1),%R12D	1	0.20	0.20	0	0	0	0.20	0.20	0	0	0	0.20	0	1-2	0.20	N/A
MOV 0xc8(%RSP),%RAX	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	1	0.33	N/A
VMOVSS %XMM5,(%RAX,%R12,4)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	1	0.50	scal (6.3%)
LEA 0x1(%R15),%R12	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	N/A
CMP %RSI,%R15	1	0.20	0.20	0	0	0	0.20	0.20	0	0	0	0.20	0	1	0.20	scal (12.5%)
MOV %R12,%R15	1	0	0	0	0	0	0	0	0	0	0	0	0	1	0.17	N/A
JE 403dc0 <main+0xb00>	1	0.50	0	0	0	0	0	0.50	0	0	0	0	0	0	0.50	N/A
VPBROADCASTD %R15D,%YMM6	1	0	0	0	0	0	1	0	0	0	0	0	0	3	1	scal (6.3%)
VXORPD %XMM5,%XMM5,%XMM5	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	vect (25.0%)
XOR %R12D,%R12D	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	N/A
TEST %RDI,%RDI	1	0.20	0.20	0	0	0	0.20	0.20	0	0	0	0.20	0	2	0.20	scal (12.5%)
JE 403df0 <main+0xb30>	1	0.50	0	0	0	0	0	0.50	0	0	0	0	0	0	0.50	N/A
NOPW %CS:(%RAX,%RAX,1)	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	N/A
MOV %RDI,%R12	1	0	0	0	0	0	0	0	0	0	0	0	0	1	0.17	N/A
CMP %EDI,0x38(%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
MOV 0x88(%RSP),%R13	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	1	0.33	scal (12.5%)
JNE 403df0 <main+0xb30>	1	0.50	0	0	0	0	0	0.50	0	0	0	0	0	0	0.50	N/A
JMP 403e55 <main+0xb95>	1	0	0	0	0	0	0	0	0	0	0	0	0	0	2.08	N/A

Function	main
Source file and lines	attention.cpp:27-98
Module	attention-avx512

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

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 42% 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 10.83 to 5.89 cycles (1.84x speedup).

Details

73% of SSE/AVX instructions are used in vector version (process two or more data elements in vector registers):

66% of SSE/AVX loads are used in vector version.
0% of SSE/AVX stores are used in vector version.
88% of SSE/AVX addition or subtraction instructions are used in vector version.
70% 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

Performance is limited by execution of FP multiply or FMA (fused multiply-add) operations (the FP multiply/FMA unit is a bottleneck).

Workaround

Reduce the number of FP multiply/FMA instructions

Masked instructions

Detected masked instructions.

Details

Vector registers are partially exploited, which is expected if your loop is irregular or mixes elements of different sizes.

Workaround

If your loop is irregular, try to remove or hoist conditional structures out of your loop. If it mixes elements of different sizes, try to uniformize them.

FMA

Detected 8 FMA (fused multiply-add) operations.

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.

VCVTPS2PD: 2 occurrences►
- /home/eoseret/llm-attention/attention.cpp:31
- /home/eoseret/llm-attention/attention.cpp:31
VCVTSD2SS: 1 occurrences►
- /home/eoseret/llm-attention/attention.cpp:33
VGATHERDPS: 2 occurrences►
- /home/eoseret/llm-attention/attention.cpp:31
- /home/eoseret/llm-attention/attention.cpp:31
VPMULLD: 1 occurrences►
- /home/eoseret/llm-attention/attention.cpp:31

Slow data structures access

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

Details

Irregular (variable stride) or indirect: 2 occurrence(s)

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

Workaround

Try to remove indirect accesses. If applicable, precompute elements out of the innermost loop.

Vector unaligned load/store instructions

Detected 1 suboptimal vector unaligned load/store instructions.

Details

VEXTRACTF128: 1 occurrences►
- /home/eoseret/llm-attention/attention.cpp:31

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.

Gather/scatter instructions

Detected gather/scatter instructions (typically caused by indirect accesses).

Details

VGATHERDPS: 2 occurrences►
- /home/eoseret/llm-attention/attention.cpp:31
- /home/eoseret/llm-attention/attention.cpp:31

Workaround

Try to simplify your code and/or replace indirect accesses with unit-stride ones.

Conversion instructions

Detected expensive conversion instructions.

Details

VCVTPS2PD (FP32 to FP64, SIMD): 2 occurrences►
- /home/eoseret/llm-attention/attention.cpp:31
- /home/eoseret/llm-attention/attention.cpp:31
VCVTSD2SS (FP64 to FP32, scalar): 1 occurrences►
- /home/eoseret/llm-attention/attention.cpp:33
VPMOVQD: 1 occurrences►
- /home/eoseret/llm-attention/attention.cpp:31

Workaround

Type of elements and instruction set

1 SSE or AVX instructions are processing arithmetic or math operations on double precision FP elements in scalar mode (one at a time). 1 SSE or AVX instructions are processing arithmetic or math operations on double precision FP elements in vector mode (two at a time). 2 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 27 FP arithmetical operations:

19: addition or subtraction (8 inside FMA instructions)
8: multiply (all inside FMA instructions)

The binary loop is loading 84 bytes (10 double precision FP elements). The binary loop is storing 4 bytes (0 double precision FP elements).

Arithmetic intensity

Arithmetic intensity is 0.31 FP operations per loaded or stored byte.

General properties

nb instructions	43
nb uops	54
loop length	229
used x86 registers	10
used mmx registers	0
used xmm registers	4
used ymm registers	9
used zmm registers	6
nb stack references	3

Front-end

MACRO FUSION NOT POSSIBLE FIT IN UOP CACHE

micro-operation queue	9.00 cycles
front end	9.00 cycles

Back-end

	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11
uops	10.83	10.67	6.33	6.33	0.50	10.50	3.60	0.50	0.50	0.50	3.40	6.33
cycles	10.83	10.67	6.33	6.33	0.50	10.50	3.60	0.50	0.50	0.50	3.40	6.33

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	NA
Longest recurrence chain latency (RecMII)	0.00

Cycles summary

Front-end	9.00
Dispatch	10.83
Data deps.	0.00
Overall L1	10.83

Vectorization ratios

INT

all	66%
load	0%
store	NA (no store vectorizable/vectorized instructions)
mul	100%
add-sub	100%
fma	NA (no fma vectorizable/vectorized instructions)
other	42%

all	80%
load	100%
store	0%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	66%
fma	100%
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	90%

INT+FP

all	73%
load	66%
store	0%
mul	100%
add-sub	88%
fma	100%
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	70%

Vector efficiency ratios

INT

all	45%
load	12%
store	NA (no store vectorizable/vectorized instructions)
mul	50%
add-sub	58%
fma	NA (no fma vectorizable/vectorized instructions)
other	38%

all	40%
load	50%
store	6%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	45%
fma	100%
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	36%

INT+FP

all	42%
load	37%
store	6%
mul	50%
add-sub	54%
fma	100%
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	37%

Cycles and memory resources usage

Detected masked instructions: assuming all mask elements are active. Assuming all data fit into the L1 cache, each iteration of the binary loop takes 10.83 cycles. At this rate:

4% of peak load performance is reached (7.75 out of 192.00 bytes loaded per cycle (GB/s @ 1GHz))
0% of peak store performance is reached (0.37 out of 128.00 bytes stored per cycle (GB/s @ 1GHz))

Front-end bottlenecks

Found no such bottlenecks.

ASM code

In the binary file, the address of the loop is: 403e9b

Instruction	Nb FU	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11	Latency	Recip. throughput	Vectorization
VPBROADCASTQ %R12,%ZMM7	1	0	0	0	0	0	1	0	0	0	0	0	0	3	1	scal (12.5%)
VPSUBQ %ZMM7,%ZMM11,%ZMM8	1	0.50	0	0	0	0	0.50	0	0	0	0	0	0	0-1	0.50	vect (100.0%)
VPCMPNLEUQ %ZMM0,%ZMM8,%K1																vect (100.0%)
VPMOVQD %ZMM7,%YMM7	1	0	0	0	0	0	1	0	0	0	0	0	0	3	1	vect (100.0%)
VPADDD %YMM1,%YMM7,%YMM7	1	0.33	0.33	0	0	0	0.33	0	0	0	0	0	0	1	0.33	vect (50.0%)
VPADDD %YMM7,%YMM4,%YMM8	1	0.33	0.33	0	0	0	0.33	0	0	0	0	0	0	1	0.33	vect (50.0%)
VPSUBD %YMM2,%YMM8,%YMM8	1	0.33	0.33	0	0	0	0.33	0	0	0	0	0	0	0-1	0.33	vect (50.0%)
KMOVQ %K1,%K2	1	1	0	0	0	0	0	0	0	0	0	0	0	0	1	N/A
VXORPS %XMM9,%XMM9,%XMM9	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	vect (25.0%)
VGATHERDPS (%R8,%YMM8,4),%YMM9{%K2}	5	1	1	2.67	2.67	0	1	0	0	0	0	0	2.67	29	3	vect (50.0%)
VCVTPS2PD %YMM9,%ZMM8	2	1	0	0	0	0	1	0	0	0	0	0	0	7	1	vect (50.0%)
VPMULLD %YMM7,%YMM10,%YMM7	2	1	1	0	0	0	0	0	0	0	0	0	0	10	1	vect (50.0%)
VPADDD %YMM7,%YMM6,%YMM6	1	0.33	0.33	0	0	0	0.33	0	0	0	0	0	0	1	0.33	vect (50.0%)
VPSUBD %YMM2,%YMM6,%YMM6	1	0.33	0.33	0	0	0	0.33	0	0	0	0	0	0	0-1	0.33	vect (50.0%)
KMOVQ %K1,%K2	1	1	0	0	0	0	0	0	0	0	0	0	0	0	1	N/A
VPXOR %XMM7,%XMM7,%XMM7	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	vect (25.0%)
VGATHERDPS (%R9,%YMM6,4),%YMM7{%K2}	5	1	1	2.67	2.67	0	1	0	0	0	0	0	2.67	29	3	vect (50.0%)
VCVTPS2PD %YMM7,%ZMM6	2	1	0	0	0	0	1	0	0	0	0	0	0	7	1	vect (50.0%)
VFMADD231PD %ZMM8,%ZMM6,%ZMM5{%K1}	1	0.50	0	0	0	0	0.50	0	0	0	0	0	0	4	0.50	vect (100.0%)
VEXTRACTF64X4 $0x1,%ZMM5,%YMM6	1	0	0	0	0	0	1	0	0	0	0	0	0	3	1	vect (50.0%)
VADDPD %ZMM6,%ZMM5,%ZMM5	1	0.50	0	0	0	0	0.50	0	0	0	0	0	0	3	0.50	vect (100.0%)
VEXTRACTF128 $0x1,%YMM5,%XMM6	1	0	0	0	0	0	1	0	0	0	0	0	0	3	1	vect (25.0%)
VADDPD %XMM6,%XMM5,%XMM5	1	0	0.50	0	0	0	0.50	0	0	0	0	0	0	3	0.50	vect (25.0%)
VSHUFPD $0x1,%XMM5,%XMM5,%XMM6	1	0	0.50	0	0	0	0.50	0	0	0	0	0	0	1	0.50	vect (25.0%)
VADDSD %XMM6,%XMM5,%XMM5	1	0	0.50	0	0	0	0.50	0	0	0	0	0	0	3	0.50	scal (12.5%)
VCVTSD2SS %XMM5,%XMM5,%XMM5	2	0.50	0.50	0	0	0	1	0	0	0	0	0	0	5	1	scal (12.5%)
LEA (%R14,%R15,1),%R12D	1	0.20	0.20	0	0	0	0.20	0.20	0	0	0	0.20	0	1-2	0.20	N/A
MOV 0xc8(%RSP),%RAX	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	1	0.33	N/A
VMOVSS %XMM5,(%RAX,%R12,4)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	1	0.50	scal (6.3%)
LEA 0x1(%R15),%R12	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	N/A
CMP %RSI,%R15	1	0.20	0.20	0	0	0	0.20	0.20	0	0	0	0.20	0	1	0.20	scal (12.5%)
MOV %R12,%R15	1	0	0	0	0	0	0	0	0	0	0	0	0	1	0.17	N/A
JE 403dc0 <main+0xb00>	1	0.50	0	0	0	0	0	0.50	0	0	0	0	0	0	0.50	N/A
VPBROADCASTD %R15D,%YMM6	1	0	0	0	0	0	1	0	0	0	0	0	0	3	1	scal (6.3%)
VXORPD %XMM5,%XMM5,%XMM5	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	vect (25.0%)
XOR %R12D,%R12D	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	N/A
TEST %RDI,%RDI	1	0.20	0.20	0	0	0	0.20	0.20	0	0	0	0.20	0	2	0.20	scal (12.5%)
JE 403df0 <main+0xb30>	1	0.50	0	0	0	0	0	0.50	0	0	0	0	0	0	0.50	N/A
NOPW %CS:(%RAX,%RAX,1)	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	N/A
MOV %RDI,%R12	1	0	0	0	0	0	0	0	0	0	0	0	0	1	0.17	N/A
CMP %EDI,0x38(%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
MOV 0x88(%RSP),%R13	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	1	0.33	scal (12.5%)
JNE 403df0 <main+0xb30>	1	0.50	0	0	0	0	0	0.50	0	0	0	0	0	0	0.50	N/A

Report Configuration

Vectorization

Details

Workaround

Masked instructions

Details

Workaround

FMA

Complex instructions

Details

Gather/scatter 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

Code clean check

Workaround

Vectorization

Details

Workaround

Execution units bottlenecks

Workaround

Masked instructions

Details

Workaround

FMA

Complex instructions

Details

Vector unaligned load/store instructions

Details

Workaround

Gather/scatter 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

Complex instructions

Details

Slow data structures access

Details

Workaround

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