OV - attention-avx512 - Loop 57

	Loop Id: 57	Module: attention-avx512	Source: attention.cpp:27-33	Coverage: 0.42%

0x404750 LEA	(,%R11,8),%EDX

0x404758 VXORPD	%XMM7,%XMM7,%XMM7

0x40475c MOV	-0x40(%RBP),%R10D

0x404760 MOV	%R9D,%ESI

0x404763 MOV	-0x70(%RBP),%RDI

0x404767 VXORPD	%XMM8,%XMM8,%XMM8

0x40476c VXORPD	%XMM9,%XMM9,%XMM9

0x404771 VXORPD	%XMM10,%XMM10,%XMM10

0x404776 VXORPD	%XMM11,%XMM11,%XMM11

0x40477b VXORPD	%XMM12,%XMM12,%XMM12

0x404780 VXORPD	%XMM13,%XMM13,%XMM13

0x404785 VXORPD	%XMM14,%XMM14,%XMM14

0x40478a NOPW	(%RAX,%RAX,1)

  (56) 0x404790 MOV	%R10D,%R14D

  (56) 0x404793 VMOVSS	(%R15,%R14,4),%XMM15

  (56) 0x404799 VCVTSS2SD	%XMM15,%XMM15,%XMM15

  (56) 0x40479e LEA	-0x7(%RSI),%R14D

  (56) 0x4047a2 VMOVSS	(%R13,%R14,4),%XMM16

  (56) 0x4047aa VCVTSS2SD	%XMM16,%XMM16,%XMM16

  (56) 0x4047b0 VFMADD231SD	%XMM16,%XMM15,%XMM14

  (56) 0x4047b6 LEA	-0x6(%RSI),%R14D

  (56) 0x4047ba VMOVSS	(%R13,%R14,4),%XMM16

  (56) 0x4047c2 VCVTSS2SD	%XMM16,%XMM16,%XMM16

  (56) 0x4047c8 VFMADD231SD	%XMM16,%XMM15,%XMM13

  (56) 0x4047ce LEA	-0x5(%RSI),%R14D

  (56) 0x4047d2 VMOVSS	(%R13,%R14,4),%XMM16

  (56) 0x4047da VCVTSS2SD	%XMM16,%XMM16,%XMM16

  (56) 0x4047e0 LEA	-0x4(%RSI),%R14D

  (56) 0x4047e4 VMOVSS	(%R13,%R14,4),%XMM17

  (56) 0x4047ec VCVTSS2SD	%XMM17,%XMM17,%XMM17

  (56) 0x4047f2 VFMADD231SD	%XMM16,%XMM15,%XMM12

  (56) 0x4047f8 VFMADD231SD	%XMM17,%XMM15,%XMM11

  (56) 0x4047fe LEA	-0x3(%RSI),%R14D

  (56) 0x404802 VMOVSS	(%R13,%R14,4),%XMM16

  (56) 0x40480a VCVTSS2SD	%XMM16,%XMM16,%XMM16

  (56) 0x404810 VFMADD231SD	%XMM16,%XMM15,%XMM10

  (56) 0x404816 LEA	-0x2(%RSI),%R14D

  (56) 0x40481a VMOVSS	(%R13,%R14,4),%XMM16

  (56) 0x404822 VCVTSS2SD	%XMM16,%XMM16,%XMM16

  (56) 0x404828 VFMADD231SD	%XMM16,%XMM15,%XMM9

  (56) 0x40482e LEA	-0x1(%RSI),%R14D

  (56) 0x404832 VMOVSS	(%R13,%R14,4),%XMM16

  (56) 0x40483a VCVTSS2SD	%XMM16,%XMM16,%XMM16

  (56) 0x404840 VFMADD231SD	%XMM16,%XMM15,%XMM8

  (56) 0x404846 MOV	%ESI,%R14D

  (56) 0x404849 VMOVSS	(%R13,%R14,4),%XMM16

  (56) 0x404851 VCVTSS2SD	%XMM16,%XMM16,%XMM16

  (56) 0x404857 VFMADD231SD	%XMM16,%XMM15,%XMM7

  (56) 0x40485d ADD	%R8D,%ESI

  (56) 0x404860 INC	%R10D

  (56) 0x404863 DEC	%RDI

  (56) 0x404866 JNE	404790

0x40486c VCVTSD2SS	%XMM14,%XMM14,%XMM14

0x404871 LEA	(%RAX,%RDX,1),%ESI

0x404874 MOV	-0xc0(%RBP),%RDI

0x40487b VMOVSS	%XMM14,(%RDI,%RSI,4)

0x404880 VCVTSD2SS	%XMM13,%XMM13,%XMM13

0x404885 LEA	0x1(%RAX,%RDX,1),%ESI

0x404889 VMOVSS	%XMM13,(%RDI,%RSI,4)

0x40488e VCVTSD2SS	%XMM12,%XMM12,%XMM12

0x404893 LEA	0x2(%RAX,%RDX,1),%ESI

0x404897 VMOVSS	%XMM12,(%RDI,%RSI,4)

0x40489c VCVTSD2SS	%XMM11,%XMM11,%XMM11

0x4048a1 LEA	0x3(%RAX,%RDX,1),%ESI

0x4048a5 VMOVSS	%XMM11,(%RDI,%RSI,4)

0x4048aa VCVTSD2SS	%XMM10,%XMM10,%XMM10

0x4048af LEA	0x4(%RAX,%RDX,1),%ESI

0x4048b3 VMOVSS	%XMM10,(%RDI,%RSI,4)

0x4048b8 VCVTSD2SS	%XMM9,%XMM9,%XMM9

0x4048bd LEA	0x5(%RAX,%RDX,1),%ESI

0x4048c1 VMOVSS	%XMM9,(%RDI,%RSI,4)

0x4048c6 VCVTSD2SS	%XMM8,%XMM8,%XMM8

0x4048cb LEA	0x6(%RAX,%RDX,1),%ESI

0x4048cf VMOVSS	%XMM8,(%RDI,%RSI,4)

0x4048d4 VCVTSD2SS	%XMM7,%XMM7,%XMM7

0x4048d8 LEA	0x7(%RAX,%RDX,1),%EDX

0x4048dc VMOVSS	%XMM7,(%RDI,%RDX,4)

0x4048e1 ADD	$0x8,%R9D

0x4048e5 CMP	-0x60(%RBP),%R11

0x4048e9 LEA	0x1(%R11),%R11

0x4048ed JNE	404750

/home/eoseret/llm-attention/attention.cpp: 27 - 33

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

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	1.04
CQA speedup if FP arith vectorized	1.00
CQA speedup if fully vectorized	8.33
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.04
Bottlenecks	micro-operation queue,
Function	main
Source	attention.cpp:27-27,attention.cpp:31-33
Source loop unroll info	NA
Source loop unroll confidence level	NA
Unroll/vectorization loop type	NA
Unroll factor	NA
CQA cycles	8.33
CQA cycles if no scalar integer	8.00
CQA cycles if FP arith vectorized	8.33
CQA cycles if fully vectorized	1.00
Front-end cycles	8.33
P0 cycles	7.50
P1 cycles	7.80
P2 cycles	1.33
P3 cycles	1.33
P4 cycles	4.00
P5 cycles	8.00
P6 cycles	2.60
P7 cycles	4.00
P8 cycles	4.00
P9 cycles	4.00
P10 cycles	2.40
P11 cycles	1.33
DIV/SQRT cycles	0.00
Inter-iter dependencies cycles	1
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	42.00
Nb uops	50.00
Nb loads	4.00
Nb stores	8.00
Nb stack references	4.00
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	7.20
Bytes prefetched	0.00
Bytes loaded	28.00
Bytes stored	32.00
Stride 0	0.00
Stride 1	0.00
Stride n	1.00
Stride unknown	1.00
Stride indirect	0.00
Vectorization ratio all	30.77
Vectorization ratio load	0.00
Vectorization ratio store	0.00
Vectorization ratio mul	NA
Vectorization ratio add_sub	0.00
Vectorization ratio fma	NA
Vectorization ratio div_sqrt	NA
Vectorization ratio other	50.00
Vector-efficiency ratio all	13.94
Vector-efficiency ratio load	6.25
Vector-efficiency ratio store	6.25
Vector-efficiency ratio mul	NA
Vector-efficiency ratio add_sub	6.25
Vector-efficiency ratio fma	NA
Vector-efficiency ratio div_sqrt	NA
Vector-efficiency ratio other	18.75

Metric	Value
CQA speedup if no scalar integer	1.04
CQA speedup if FP arith vectorized	1.00
CQA speedup if fully vectorized	8.33
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.04
Bottlenecks	micro-operation queue,
Function	main
Source	attention.cpp:27-27,attention.cpp:31-33
Source loop unroll info	NA
Source loop unroll confidence level	NA
Unroll/vectorization loop type	NA
Unroll factor	NA
CQA cycles	8.33
CQA cycles if no scalar integer	8.00
CQA cycles if FP arith vectorized	8.33
CQA cycles if fully vectorized	1.00
Front-end cycles	8.33
P0 cycles	7.50
P1 cycles	7.80
P2 cycles	1.33
P3 cycles	1.33
P4 cycles	4.00
P5 cycles	8.00
P6 cycles	2.60
P7 cycles	4.00
P8 cycles	4.00
P9 cycles	4.00
P10 cycles	2.40
P11 cycles	1.33
DIV/SQRT cycles	0.00
Inter-iter dependencies cycles	1
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	42.00
Nb uops	50.00
Nb loads	4.00
Nb stores	8.00
Nb stack references	4.00
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	7.20
Bytes prefetched	0.00
Bytes loaded	28.00
Bytes stored	32.00
Stride 0	0.00
Stride 1	0.00
Stride n	1.00
Stride unknown	1.00
Stride indirect	0.00
Vectorization ratio all	30.77
Vectorization ratio load	0.00
Vectorization ratio store	0.00
Vectorization ratio mul	NA
Vectorization ratio add_sub	0.00
Vectorization ratio fma	NA
Vectorization ratio div_sqrt	NA
Vectorization ratio other	50.00
Vector-efficiency ratio all	13.94
Vector-efficiency ratio load	6.25
Vector-efficiency ratio store	6.25
Vector-efficiency ratio mul	NA
Vector-efficiency ratio add_sub	6.25
Vector-efficiency ratio fma	NA
Vector-efficiency ratio div_sqrt	NA
Vector-efficiency ratio other	18.75

Path /

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

Function	main
Source file and lines	attention.cpp:27-33
Module	attention-avx512

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

Analyzed code is defined in /home/eoseret/llm-attention/attention.cpp:27,31-33.

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

gain
potential
hint
expert

Vectorization

Your loop is probably not vectorized. Only 13% of vector register length is used (average across all SSE/AVX instructions). By vectorizing your loop, you can lower the cost of an iteration from 8.33 to 1.00 cycles (8.33x speedup).

Details

Store and arithmetical SSE/AVX 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)

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

Slow data structures access

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

Details

Constant unknown stride: 1 occurrence(s)
Constant non-unit stride: 1 occurrence(s)

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

Workaround

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

Conversion instructions

Detected expensive conversion instructions.

Details

VCVTSD2SS (FP64 to FP32, scalar): 8 occurrences►
- /home/eoseret/llm-attention/attention.cpp:33
- /home/eoseret/llm-attention/attention.cpp:33
- /home/eoseret/llm-attention/attention.cpp:33
- /home/eoseret/llm-attention/attention.cpp:33
- /home/eoseret/llm-attention/attention.cpp:33
- /home/eoseret/llm-attention/attention.cpp:33
- /home/eoseret/llm-attention/attention.cpp:33
- /home/eoseret/llm-attention/attention.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

No instructions are processing arithmetic or math operations on FP elements. This loop is probably writing/copying data or processing integer elements.

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 is loading 28 bytes. The binary loop is storing 32 bytes.

General properties

nb instructions	42
nb uops	50
loop length	199
used x86 registers	8
used mmx registers	0
used xmm registers	8
used ymm registers	0
used zmm registers	0
nb stack references	4

Front-end

MACRO FUSION NOT POSSIBLE FIT IN UOP CACHE

micro-operation queue	8.33 cycles
front end	8.33 cycles

Back-end

	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11
uops	7.50	7.50	1.33	1.33	4.00	8.00	2.60	4.00	4.00	4.00	2.40	1.33
cycles	7.50	7.80	1.33	1.33	4.00	8.00	2.60	4.00	4.00	4.00	2.40	1.33

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)	1.00

Cycles summary

Front-end	8.33
Dispatch	8.00
Data deps.	1.00
Overall L1	8.33

Vectorization ratios

INT

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

all	33%
load	NA (no load vectorizable/vectorized instructions)
store	0%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	NA (no add-sub vectorizable/vectorized instructions)
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	50%

INT+FP

all	30%
load	0%
store	0%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	0%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	50%

Vector efficiency ratios

INT

all	6%
load	6%
store	NA (no store vectorizable/vectorized instructions)
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	6%
fma	NA (no fma vectorizable/vectorized instructions)
other	NA (no other vectorizable/vectorized instructions)

all	14%
load	NA (no load vectorizable/vectorized instructions)
store	6%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	NA (no add-sub vectorizable/vectorized instructions)
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	18%

INT+FP

all	13%
load	6%
store	6%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	6%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	18%

Cycles and memory resources usage

Assuming all data fit into the L1 cache, each iteration of the binary loop takes 8.33 cycles. At this rate:

1% of peak load performance is reached (3.36 out of 192.00 bytes loaded per cycle (GB/s @ 1GHz))
3% of peak store performance is reached (3.84 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).

ASM code

In the binary file, the address of the loop is: 404750

Instruction	Nb FU	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11	Latency	Recip. throughput	Vectorization
LEA (,%R11,8),%EDX	1	0.20	0.20	0	0	0	0.20	0.20	0	0	0	0.20	0	1-2	0.20	N/A
VXORPD %XMM7,%XMM7,%XMM7	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	vect (25.0%)
MOV -0x40(%RBP),%R10D	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	1	0.33	scal (6.3%)
MOV %R9D,%ESI	1	0	0	0	0	0	0	0	0	0	0	0	0	1	0.17	N/A
MOV -0x70(%RBP),%RDI	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	1	0.33	N/A
VXORPD %XMM8,%XMM8,%XMM8	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	vect (25.0%)
VXORPD %XMM9,%XMM9,%XMM9	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	vect (25.0%)
VXORPD %XMM10,%XMM10,%XMM10	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	vect (25.0%)
VXORPD %XMM11,%XMM11,%XMM11	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	vect (25.0%)
VXORPD %XMM12,%XMM12,%XMM12	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	vect (25.0%)
VXORPD %XMM13,%XMM13,%XMM13	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	vect (25.0%)
VXORPD %XMM14,%XMM14,%XMM14	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	vect (25.0%)
NOPW (%RAX,%RAX,1)	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	N/A
VCVTSD2SS %XMM14,%XMM14,%XMM14	2	0.50	0.50	0	0	0	1	0	0	0	0	0	0	5	1	scal (12.5%)
LEA (%RAX,%RDX,1),%ESI	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 -0xc0(%RBP),%RDI	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	1	0.33	N/A
VMOVSS %XMM14,(%RDI,%RSI,4)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	1	0.50	scal (6.3%)
VCVTSD2SS %XMM13,%XMM13,%XMM13	2	0.50	0.50	0	0	0	1	0	0	0	0	0	0	5	1	scal (12.5%)
LEA 0x1(%RAX,%RDX,1),%ESI	1	0	1	0	0	0	0	0	0	0	0	0	0	3	1	N/A
VMOVSS %XMM13,(%RDI,%RSI,4)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	1	0.50	scal (6.3%)
VCVTSD2SS %XMM12,%XMM12,%XMM12	2	0.50	0.50	0	0	0	1	0	0	0	0	0	0	5	1	scal (12.5%)
LEA 0x2(%RAX,%RDX,1),%ESI	1	0	1	0	0	0	0	0	0	0	0	0	0	3	1	N/A
VMOVSS %XMM12,(%RDI,%RSI,4)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	1	0.50	scal (6.3%)
VCVTSD2SS %XMM11,%XMM11,%XMM11	2	0.50	0.50	0	0	0	1	0	0	0	0	0	0	5	1	scal (12.5%)
LEA 0x3(%RAX,%RDX,1),%ESI	1	0	1	0	0	0	0	0	0	0	0	0	0	3	1	N/A
VMOVSS %XMM11,(%RDI,%RSI,4)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	1	0.50	scal (6.3%)
VCVTSD2SS %XMM10,%XMM10,%XMM10	2	0.50	0.50	0	0	0	1	0	0	0	0	0	0	5	1	scal (12.5%)
LEA 0x4(%RAX,%RDX,1),%ESI	1	0	1	0	0	0	0	0	0	0	0	0	0	3	1	N/A
VMOVSS %XMM10,(%RDI,%RSI,4)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	1	0.50	scal (6.3%)
VCVTSD2SS %XMM9,%XMM9,%XMM9	2	0.50	0.50	0	0	0	1	0	0	0	0	0	0	5	1	scal (12.5%)
LEA 0x5(%RAX,%RDX,1),%ESI	1	0	1	0	0	0	0	0	0	0	0	0	0	3	1	N/A
VMOVSS %XMM9,(%RDI,%RSI,4)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	1	0.50	scal (6.3%)
VCVTSD2SS %XMM8,%XMM8,%XMM8	2	0.50	0.50	0	0	0	1	0	0	0	0	0	0	5	1	scal (12.5%)
LEA 0x6(%RAX,%RDX,1),%ESI	1	0	1	0	0	0	0	0	0	0	0	0	0	3	1	N/A
VMOVSS %XMM8,(%RDI,%RSI,4)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	1	0.50	scal (6.3%)
VCVTSD2SS %XMM7,%XMM7,%XMM7	2	0.50	0.50	0	0	0	1	0	0	0	0	0	0	5	1	scal (12.5%)
LEA 0x7(%RAX,%RDX,1),%EDX	1	0	1	0	0	0	0	0	0	0	0	0	0	3	1	N/A
VMOVSS %XMM7,(%RDI,%RDX,4)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	1	0.50	scal (6.3%)
ADD $0x8,%R9D	1	0.20	0.20	0	0	0	0.20	0.20	0	0	0	0.20	0	1	0.20	scal (6.3%)
CMP -0x60(%RBP),%R11	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
LEA 0x1(%R11),%R11	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	N/A
JNE 404750 <main+0x950>	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:27-33
Module	attention-avx512

gain
potential
hint
expert

Vectorization

Details

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

Slow data structures access

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

Details

Constant unknown stride: 1 occurrence(s)
Constant non-unit stride: 1 occurrence(s)

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

Workaround

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

Conversion instructions

Detected expensive conversion instructions.

Details

VCVTSD2SS (FP64 to FP32, scalar): 8 occurrences►
- /home/eoseret/llm-attention/attention.cpp:33
- /home/eoseret/llm-attention/attention.cpp:33
- /home/eoseret/llm-attention/attention.cpp:33
- /home/eoseret/llm-attention/attention.cpp:33
- /home/eoseret/llm-attention/attention.cpp:33
- /home/eoseret/llm-attention/attention.cpp:33
- /home/eoseret/llm-attention/attention.cpp:33
- /home/eoseret/llm-attention/attention.cpp:33

Workaround

Type of elements and instruction set

No instructions are processing arithmetic or math operations on FP elements. This loop is probably writing/copying data or processing integer elements.

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 is loading 28 bytes. The binary loop is storing 32 bytes.

General properties

nb instructions	42
nb uops	50
loop length	199
used x86 registers	8
used mmx registers	0
used xmm registers	8
used ymm registers	0
used zmm registers	0
nb stack references	4

Front-end

MACRO FUSION NOT POSSIBLE FIT IN UOP CACHE

micro-operation queue	8.33 cycles
front end	8.33 cycles

Back-end

	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11
uops	7.50	7.50	1.33	1.33	4.00	8.00	2.60	4.00	4.00	4.00	2.40	1.33
cycles	7.50	7.80	1.33	1.33	4.00	8.00	2.60	4.00	4.00	4.00	2.40	1.33

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)	1.00

Cycles summary

Front-end	8.33
Dispatch	8.00
Data deps.	1.00
Overall L1	8.33

Vectorization ratios

INT

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

all	33%
load	NA (no load vectorizable/vectorized instructions)
store	0%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	NA (no add-sub vectorizable/vectorized instructions)
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	50%

INT+FP

all	30%
load	0%
store	0%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	0%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	50%

Vector efficiency ratios

INT

all	6%
load	6%
store	NA (no store vectorizable/vectorized instructions)
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	6%
fma	NA (no fma vectorizable/vectorized instructions)
other	NA (no other vectorizable/vectorized instructions)

all	14%
load	NA (no load vectorizable/vectorized instructions)
store	6%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	NA (no add-sub vectorizable/vectorized instructions)
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	18%

INT+FP

all	13%
load	6%
store	6%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	6%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	18%

Cycles and memory resources usage

Assuming all data fit into the L1 cache, each iteration of the binary loop takes 8.33 cycles. At this rate:

1% of peak load performance is reached (3.36 out of 192.00 bytes loaded per cycle (GB/s @ 1GHz))
3% of peak store performance is reached (3.84 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).

ASM code

In the binary file, the address of the loop is: 404750

Instruction	Nb FU	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11	Latency	Recip. throughput	Vectorization
LEA (,%R11,8),%EDX	1	0.20	0.20	0	0	0	0.20	0.20	0	0	0	0.20	0	1-2	0.20	N/A
VXORPD %XMM7,%XMM7,%XMM7	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	vect (25.0%)
MOV -0x40(%RBP),%R10D	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	1	0.33	scal (6.3%)
MOV %R9D,%ESI	1	0	0	0	0	0	0	0	0	0	0	0	0	1	0.17	N/A
MOV -0x70(%RBP),%RDI	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	1	0.33	N/A
VXORPD %XMM8,%XMM8,%XMM8	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	vect (25.0%)
VXORPD %XMM9,%XMM9,%XMM9	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	vect (25.0%)
VXORPD %XMM10,%XMM10,%XMM10	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	vect (25.0%)
VXORPD %XMM11,%XMM11,%XMM11	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	vect (25.0%)
VXORPD %XMM12,%XMM12,%XMM12	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	vect (25.0%)
VXORPD %XMM13,%XMM13,%XMM13	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	vect (25.0%)
VXORPD %XMM14,%XMM14,%XMM14	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	vect (25.0%)
NOPW (%RAX,%RAX,1)	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	N/A
VCVTSD2SS %XMM14,%XMM14,%XMM14	2	0.50	0.50	0	0	0	1	0	0	0	0	0	0	5	1	scal (12.5%)
LEA (%RAX,%RDX,1),%ESI	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 -0xc0(%RBP),%RDI	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	1	0.33	N/A
VMOVSS %XMM14,(%RDI,%RSI,4)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	1	0.50	scal (6.3%)
VCVTSD2SS %XMM13,%XMM13,%XMM13	2	0.50	0.50	0	0	0	1	0	0	0	0	0	0	5	1	scal (12.5%)
LEA 0x1(%RAX,%RDX,1),%ESI	1	0	1	0	0	0	0	0	0	0	0	0	0	3	1	N/A
VMOVSS %XMM13,(%RDI,%RSI,4)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	1	0.50	scal (6.3%)
VCVTSD2SS %XMM12,%XMM12,%XMM12	2	0.50	0.50	0	0	0	1	0	0	0	0	0	0	5	1	scal (12.5%)
LEA 0x2(%RAX,%RDX,1),%ESI	1	0	1	0	0	0	0	0	0	0	0	0	0	3	1	N/A
VMOVSS %XMM12,(%RDI,%RSI,4)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	1	0.50	scal (6.3%)
VCVTSD2SS %XMM11,%XMM11,%XMM11	2	0.50	0.50	0	0	0	1	0	0	0	0	0	0	5	1	scal (12.5%)
LEA 0x3(%RAX,%RDX,1),%ESI	1	0	1	0	0	0	0	0	0	0	0	0	0	3	1	N/A
VMOVSS %XMM11,(%RDI,%RSI,4)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	1	0.50	scal (6.3%)
VCVTSD2SS %XMM10,%XMM10,%XMM10	2	0.50	0.50	0	0	0	1	0	0	0	0	0	0	5	1	scal (12.5%)
LEA 0x4(%RAX,%RDX,1),%ESI	1	0	1	0	0	0	0	0	0	0	0	0	0	3	1	N/A
VMOVSS %XMM10,(%RDI,%RSI,4)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	1	0.50	scal (6.3%)
VCVTSD2SS %XMM9,%XMM9,%XMM9	2	0.50	0.50	0	0	0	1	0	0	0	0	0	0	5	1	scal (12.5%)
LEA 0x5(%RAX,%RDX,1),%ESI	1	0	1	0	0	0	0	0	0	0	0	0	0	3	1	N/A
VMOVSS %XMM9,(%RDI,%RSI,4)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	1	0.50	scal (6.3%)
VCVTSD2SS %XMM8,%XMM8,%XMM8	2	0.50	0.50	0	0	0	1	0	0	0	0	0	0	5	1	scal (12.5%)
LEA 0x6(%RAX,%RDX,1),%ESI	1	0	1	0	0	0	0	0	0	0	0	0	0	3	1	N/A
VMOVSS %XMM8,(%RDI,%RSI,4)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	1	0.50	scal (6.3%)
VCVTSD2SS %XMM7,%XMM7,%XMM7	2	0.50	0.50	0	0	0	1	0	0	0	0	0	0	5	1	scal (12.5%)
LEA 0x7(%RAX,%RDX,1),%EDX	1	0	1	0	0	0	0	0	0	0	0	0	0	3	1	N/A
VMOVSS %XMM7,(%RDI,%RDX,4)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	1	0.50	scal (6.3%)
ADD $0x8,%R9D	1	0.20	0.20	0	0	0	0.20	0.20	0	0	0	0.20	0	1	0.20	scal (6.3%)
CMP -0x60(%RBP),%R11	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
LEA 0x1(%R11),%R11	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	N/A
JNE 404750 <main+0x950>	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

Execution units bottlenecks

Complex instructions

Details

Slow data structures access

Details

Workaround

Conversion instructions

Details

Workaround

Type of elements and instruction set

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

Execution units bottlenecks

Complex instructions

Details

Slow data structures access

Details

Workaround

Conversion instructions

Details

Workaround

Type of elements and instruction set

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