OV - attention-clang-gnr512 - Loop 98 - attention-clang-gnr512

	Loop Id: 98	Module: attention-clang-gnr512	Source: random.tcc:401-3367 [...]	Coverage: 0.34%

0x4470 MOV	%RBX,%RDX

0x4473 INC	%RBX

0x4476 MOV	%RBX,0x1938(%RSP)

0x447e MOV	0x5b8(%RSP,%RDX,8),%RDX

0x4486 MOV	%RDX,%RSI

0x4489 SHR	$0xb,%RSI

0x448d MOV	%ESI,%ESI

0x448f XOR	%RDX,%RSI

0x4492 MOV	%ESI,%EDX

0x4494 SAL	$0x7,%EDX

0x4497 AND	$-0x62d3a980,%EDX

0x449d XOR	%RSI,%RDX

0x44a0 MOV	%EDX,%ESI

0x44a2 SAL	$0xf,%ESI

0x44a5 AND	$-0x103a0000,%ESI

0x44ab XOR	%RDX,%RSI

0x44ae MOV	%RSI,%RDX

0x44b1 SHR	$0x12,%RDX

0x44b5 XOR	%RSI,%RDX

0x44b8 VCVTUSI2SS	%RDX,%XMM15,%XMM2

0x44be VFMADD231SS	%XMM2,%XMM1,%XMM0

0x44c3 VMULSS	0x4b3d(%RIP),%XMM1,%XMM1

0x44cb DEC	%RCX

0x44ce JE	4840

0x44d4 CMP	$0x270,%RBX

0x44db JB	4470

0x44dd VPBROADCASTQ	%R13,%ZMM2

0x44e3 XOR	%EDX,%EDX

0x44e5 VPBROADCASTQ	0x4b51(%RIP),%ZMM14

0x44ef VPBROADCASTQ	0x4b4f(%RIP),%ZMM15

0x44f9 VPBROADCASTQ	0x4b4d(%RIP),%ZMM16

0x4503 VPBROADCASTQ	0x4b4b(%RIP),%ZMM17

0x450d NOPL	(%RAX)

  (99) 0x4510 VMOVDQU64	0x5c0(%RSP,%RDX,8),%ZMM3

  (99) 0x4518 VMOVDQU64	0x600(%RSP,%RDX,8),%ZMM4

  (99) 0x4520 VMOVDQU64	0x640(%RSP,%RDX,8),%ZMM5

  (99) 0x4528 VALIGNQ	$0x7,%ZMM2,%ZMM3,%ZMM6

  (99) 0x452f VMOVDQU64	0x680(%RSP,%RDX,8),%ZMM2

  (99) 0x4537 VALIGNQ	$0x7,%ZMM3,%ZMM4,%ZMM7

  (99) 0x453e VALIGNQ	$0x7,%ZMM4,%ZMM5,%ZMM8

  (99) 0x4545 VALIGNQ	$0x7,%ZMM5,%ZMM2,%ZMM9

  (99) 0x454c VPANDQ	%ZMM15,%ZMM3,%ZMM10

  (99) 0x4552 VPANDQ	%ZMM15,%ZMM4,%ZMM11

  (99) 0x4558 VPANDQ	%ZMM15,%ZMM5,%ZMM12

  (99) 0x455e VPANDQ	%ZMM15,%ZMM2,%ZMM13

  (99) 0x4564 VPTERNLOGQ	$-0x8,%ZMM14,%ZMM6,%ZMM10

  (99) 0x456b VPTERNLOGQ	$-0x8,%ZMM14,%ZMM7,%ZMM11

  (99) 0x4572 VPTERNLOGQ	$-0x8,%ZMM14,%ZMM8,%ZMM12

  (99) 0x4579 VPTERNLOGQ	$-0x8,%ZMM14,%ZMM9,%ZMM13

  (99) 0x4580 VPSRLQ	$0x1,%ZMM10,%ZMM6

  (99) 0x4587 VPSRLQ	$0x1,%ZMM11,%ZMM7

  (99) 0x458e VPSRLQ	$0x1,%ZMM12,%ZMM8

  (99) 0x4595 VPSRLQ	$0x1,%ZMM13,%ZMM9

  (99) 0x459c VPXORQ	0x1220(%RSP,%RDX,8),%ZMM6,%ZMM6

  (99) 0x45a7 VPXORQ	0x1260(%RSP,%RDX,8),%ZMM7,%ZMM7

  (99) 0x45b2 VPXORQ	0x12a0(%RSP,%RDX,8),%ZMM8,%ZMM8

  (99) 0x45bd VPXORQ	0x12e0(%RSP,%RDX,8),%ZMM9,%ZMM9

  (99) 0x45c8 VPTESTMQ	%ZMM16,%ZMM3,%K1

  (99) 0x45ce VPTESTMQ	%ZMM16,%ZMM4,%K2

  (99) 0x45d4 VPTESTMQ	%ZMM16,%ZMM5,%K3

  (99) 0x45da VPTESTMQ	%ZMM16,%ZMM2,%K4

  (99) 0x45e0 VPXORQ	%ZMM17,%ZMM6,%ZMM6{%K1}

  (99) 0x45e6 VPXORQ	%ZMM17,%ZMM7,%ZMM7{%K2}

  (99) 0x45ec VPXORQ	%ZMM17,%ZMM8,%ZMM8{%K3}

  (99) 0x45f2 VPXORQ	%ZMM17,%ZMM9,%ZMM9{%K4}

  (99) 0x45f8 VMOVDQU64	%ZMM6,0x5b8(%RSP,%RDX,8)

  (99) 0x4603 VMOVDQU64	%ZMM7,0x5f8(%RSP,%RDX,8)

  (99) 0x460e VMOVDQU64	%ZMM8,0x638(%RSP,%RDX,8)

  (99) 0x4619 VMOVDQU64	%ZMM9,0x678(%RSP,%RDX,8)

  (99) 0x4624 ADD	$0x20,%RDX

  (99) 0x4628 CMP	$0xe0,%RDX

  (99) 0x462f JNE	4510

0x4635 VEXTRACTI32X4	$0x3,%ZMM2,%XMM2

0x463c VPEXTRQ	$0x1,%XMM2,%RDI

0x4642 AND	$-0x80000000,%RDI

0x4649 MOV	0xcc0(%RSP),%RSI

0x4651 MOV	0xcc8(%RSP),%RDX

0x4659 MOV	%ESI,%R8D

0x465c AND	$0x7ffffffe,%R8D

0x4663 OR	%RDI,%R8

0x4666 SHR	$0x1,%R8

0x4669 XOR	0x1920(%RSP),%R8

0x4671 MOV	%ESI,%EDI

0x4673 AND	$0x1,%EDI

0x4676 NEG	%EDI

0x4678 AND	%R12D,%EDI

0x467b XOR	%R8,%RDI

0x467e MOV	%RDI,0xcb8(%RSP)

0x4686 AND	$-0x80000000,%RSI

0x468d MOV	%EDX,%EDI

0x468f AND	$0x7ffffffe,%EDI

0x4695 OR	%RSI,%RDI

0x4698 SHR	$0x1,%RDI

0x469b XOR	0x1928(%RSP),%RDI

0x46a3 MOV	%EDX,%ESI

0x46a5 AND	$0x1,%ESI

0x46a8 NEG	%ESI

0x46aa AND	%R12D,%ESI

0x46ad XOR	%RDI,%RSI

0x46b0 MOV	%RSI,0xcc0(%RSP)

0x46b8 AND	$-0x80000000,%RDX

0x46bf MOV	0xcd0(%RSP),%RSI

0x46c7 MOV	%ESI,%EDI

0x46c9 VPBROADCASTQ	%RSI,%XMM2

0x46cf AND	$0x7ffffffe,%ESI

0x46d5 OR	%RDX,%RSI

0x46d8 SHR	$0x1,%RSI

0x46db XOR	0x1930(%RSP),%RSI

0x46e3 AND	$0x1,%EDI

0x46e6 NEG	%EDI

0x46e8 AND	%R12D,%EDI

0x46eb XOR	%RSI,%RDI

0x46ee MOV	%RDI,0xcc8(%RSP)

0x46f6 MOV	$0xe8,%EDX

0x46fb VPBROADCASTQ	0x493c(%RIP),%XMM7

0x4704 VPBROADCASTQ	0x493b(%RIP),%XMM8

0x470d VPBROADCASTQ	0x493a(%RIP),%XMM9

0x4716 VPBROADCASTQ	0x4939(%RIP),%XMM10

0x471f NOP

  (100) 0x4720 VMOVDQU	0x598(%RSP,%RDX,8),%XMM3

  (100) 0x4729 VMOVDQU	0x5a8(%RSP,%RDX,8),%XMM4

  (100) 0x4732 VPALIGNR	$0x8,%XMM2,%XMM3,%XMM2

  (100) 0x4738 VMOVDQU	0x5b8(%RSP,%RDX,8),%XMM5

  (100) 0x4741 VPAND	%XMM3,%XMM8,%XMM6

  (100) 0x4745 VPTERNLOGQ	$-0x8,%XMM7,%XMM2,%XMM6

  (100) 0x474c VPSRLQ	$0x1,%XMM6,%XMM2

  (100) 0x4751 VPXOR	-0x188(%RSP,%RDX,8),%XMM2,%XMM2

  (100) 0x475a VPTESTMQ	%XMM9,%XMM3,%K1

  (100) 0x4760 VPXORQ	%XMM10,%XMM2,%XMM2{%K1}

  (100) 0x4766 VMOVDQU	%XMM2,0x590(%RSP,%RDX,8)

  (100) 0x476f VPALIGNR	$0x8,%XMM3,%XMM4,%XMM2

  (100) 0x4775 VPAND	%XMM4,%XMM8,%XMM3

  (100) 0x4779 VPTERNLOGQ	$-0x8,%XMM7,%XMM2,%XMM3

  (100) 0x4780 VPSRLQ	$0x1,%XMM3,%XMM2

  (100) 0x4785 VPXOR	-0x178(%RSP,%RDX,8),%XMM2,%XMM2

  (100) 0x478e VPTESTMQ	%XMM9,%XMM4,%K1

  (100) 0x4794 VPXORQ	%XMM10,%XMM2,%XMM2{%K1}

  (100) 0x479a VMOVDQU	%XMM2,0x5a0(%RSP,%RDX,8)

  (100) 0x47a3 VPALIGNR	$0x8,%XMM4,%XMM5,%XMM2

  (100) 0x47a9 VPAND	%XMM5,%XMM8,%XMM3

  (100) 0x47ad VPTERNLOGQ	$-0x8,%XMM7,%XMM2,%XMM3

  (100) 0x47b4 VPSRLQ	$0x1,%XMM3,%XMM2

  (100) 0x47b9 VPXOR	-0x168(%RSP,%RDX,8),%XMM2,%XMM2

  (100) 0x47c2 VPTESTMQ	%XMM9,%XMM5,%K1

  (100) 0x47c8 VPXORQ	%XMM10,%XMM2,%XMM2{%K1}

  (100) 0x47ce VMOVDQU	%XMM2,0x5b0(%RSP,%RDX,8)

  (100) 0x47d7 ADD	$0x6,%RDX

  (100) 0x47db VMOVDQA	%XMM5,%XMM2

  (100) 0x47df CMP	$0x274,%RDX

  (100) 0x47e6 JNE	4720

0x47ec MOV	0x1930(%RSP),%RDX

0x47f4 MOV	$-0x80000000,%RSI

0x47fb AND	%RSI,%RDX

0x47fe MOV	0x5b8(%RSP),%R13

0x4806 MOV	%R13D,%ESI

0x4809 AND	$0x7ffffffe,%ESI

0x480f OR	%RDX,%RSI

0x4812 SHR	$0x1,%RSI

0x4815 XOR	0x1218(%RSP),%RSI

0x481d MOV	%R13D,%EDX

0x4820 AND	$0x1,%EDX

0x4823 NEG	%EDX

0x4825 AND	%R12D,%EDX

0x4828 XOR	%RSI,%RDX

0x482b MOV	%RDX,0x1930(%RSP)

0x4833 XOR	%EBX,%EBX

0x4835 JMP	4470

/usr/lib/gcc/x86_64-redhat-linux/11/../../../../include/c++/11/bits/random.tcc: 401 - 3367

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

401:       for (size_t __k = 0; __k < (__n - __m); ++__k)

402:         {

403: 	  _UIntType __y = ((_M_x[__k] & __upper_mask)

404: 			   | (_M_x[__k + 1] & __lower_mask));

405: 	  _M_x[__k] = (_M_x[__k + __m] ^ (__y >> 1)

406: 		       ^ ((__y & 0x01) ? __a : 0));

407:         }

408:

409:       for (size_t __k = (__n - __m); __k < (__n - 1); ++__k)

410: 	{

411: 	  _UIntType __y = ((_M_x[__k] & __upper_mask)

412: 			   | (_M_x[__k + 1] & __lower_mask));

413: 	  _M_x[__k] = (_M_x[__k + (__m - __n)] ^ (__y >> 1)

414: 		       ^ ((__y & 0x01) ? __a : 0));

415: 	}

416:

417:       _UIntType __y = ((_M_x[__n - 1] & __upper_mask)

418: 		       | (_M_x[0] & __lower_mask));

419:       _M_x[__n - 1] = (_M_x[__m - 1] ^ (__y >> 1)

420: 		       ^ ((__y & 0x01) ? __a : 0));

[...]

455:       if (_M_p >= state_size)

456: 	_M_gen_rand();

457:

458:       // Calculate o(x(i)).

459:       result_type __z = _M_x[_M_p++];

460:       __z ^= (__z >> __u) & __d;

461:       __z ^= (__z << __s) & __b;

462:       __z ^= (__z << __t) & __c;

463:       __z ^= (__z >> __l);

[...]

3364:       for (size_t __k = __m; __k != 0; --__k)

3365: 	{

3366: 	  __sum += _RealType(__urng() - __urng.min()) * __tmp;

3367: 	  __tmp *= __r;

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	2.35
CQA speedup if FP arith vectorized	1.84
CQA speedup if fully vectorized	11.86
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.34
Bottlenecks
Function	main
Source	random.tcc:401-406,random.tcc:409-409,random.tcc:413-413,random.tcc:417-420,random.tcc:455-455,random.tcc:459-463,random.tcc:3364-3367
Source loop unroll info	NA
Source loop unroll confidence level	NA
Unroll/vectorization loop type	NA
Unroll factor	NA
CQA cycles	10.58
CQA cycles if no scalar integer	4.50
CQA cycles if FP arith vectorized	5.75
CQA cycles if fully vectorized	0.89
Front-end cycles	10.58
P0 cycles	7.85
P1 cycles	7.70
P2 cycles	3.50
P3 cycles	3.50
P4 cycles	1.50
P5 cycles	7.80
P6 cycles	7.85
P7 cycles	1.50
P8 cycles	1.50
P9 cycles	1.50
P10 cycles	7.80
P11 cycles	3.50
DIV/SQRT cycles	0.00
Inter-iter dependencies cycles	4
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	61.50
Nb uops	63.50
Nb loads	10.50
Nb stores	3.00
Nb stack references	5.50
FLOP/cycle	0.28
Nb FLOP add-sub	0.00
Nb FLOP mul	1.00
Nb FLOP fma	1.00
Nb FLOP div	0.00
Nb FLOP rcp	0.00
Nb FLOP sqrt	0.00
Nb FLOP rsqrt	0.00
Bytes/cycle	7.86
Bytes prefetched	0.00
Bytes loaded	80.00
Bytes stored	24.00
Stride 0	0.00
Stride 1	0.00
Stride n	0.00
Stride unknown	2.00
Stride indirect	0.00
Vectorization ratio all	0.72
Vectorization ratio load	0.00
Vectorization ratio store	0.00
Vectorization ratio mul	0.00
Vectorization ratio add_sub	0.00
Vectorization ratio fma	0.00
Vectorization ratio div_sqrt	NA
Vectorization ratio other	0.86
Vector-efficiency ratio all	9.97
Vector-efficiency ratio load	9.18
Vector-efficiency ratio store	12.50
Vector-efficiency ratio mul	6.25
Vector-efficiency ratio add_sub	12.50
Vector-efficiency ratio fma	6.25
Vector-efficiency ratio div_sqrt	NA
Vector-efficiency ratio other	9.93

Metric	Value
CQA speedup if no scalar integer	3.33
CQA speedup if FP arith vectorized	1.75
CQA speedup if fully vectorized	11.41
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.36
Bottlenecks	micro-operation queue,
Function	main
Source	random.tcc:401-406,random.tcc:409-409,random.tcc:413-413,random.tcc:417-420,random.tcc:455-455,random.tcc:459-463,random.tcc:3364-3367
Source loop unroll info	NA
Source loop unroll confidence level	NA
Unroll/vectorization loop type	NA
Unroll factor	NA
CQA cycles	16.67
CQA cycles if no scalar integer	5.00
CQA cycles if FP arith vectorized	9.50
CQA cycles if fully vectorized	1.46
Front-end cycles	16.67
P0 cycles	12.20
P1 cycles	12.10
P2 cycles	6.33
P3 cycles	6.33
P4 cycles	2.50
P5 cycles	12.20
P6 cycles	12.30
P7 cycles	2.50
P8 cycles	2.50
P9 cycles	2.50
P10 cycles	12.20
P11 cycles	6.33
DIV/SQRT cycles	0.00
Inter-iter dependencies cycles	4
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	97.00
Nb uops	100.00
Nb loads	19.00
Nb stores	5.00
Nb stack references	10.00
FLOP/cycle	0.18
Nb FLOP add-sub	0.00
Nb FLOP mul	1.00
Nb FLOP fma	1.00
Nb FLOP div	0.00
Nb FLOP rcp	0.00
Nb FLOP sqrt	0.00
Nb FLOP rsqrt	0.00
Bytes/cycle	11.28
Bytes prefetched	0.00
Bytes loaded	148.00
Bytes stored	40.00
Stride 0	0.00
Stride 1	0.00
Stride n	0.00
Stride unknown	2.00
Stride indirect	0.00
Vectorization ratio all	1.45
Vectorization ratio load	0.00
Vectorization ratio store	0.00
Vectorization ratio mul	0.00
Vectorization ratio add_sub	0.00
Vectorization ratio fma	0.00
Vectorization ratio div_sqrt	NA
Vectorization ratio other	1.72
Vector-efficiency ratio all	10.33
Vector-efficiency ratio load	12.11
Vector-efficiency ratio store	12.50
Vector-efficiency ratio mul	6.25
Vector-efficiency ratio add_sub	12.50
Vector-efficiency ratio fma	6.25
Vector-efficiency ratio div_sqrt	NA
Vector-efficiency ratio other	10.13

Metric	Value
CQA speedup if no scalar integer	1.13
CQA speedup if FP arith vectorized	2.25
CQA speedup if fully vectorized	13.87
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.29
Bottlenecks	micro-operation queue,
Function	main
Source	random.tcc:401-406,random.tcc:409-409,random.tcc:413-413,random.tcc:417-420,random.tcc:455-455,random.tcc:459-463,random.tcc:3364-3367
Source loop unroll info	NA
Source loop unroll confidence level	NA
Unroll/vectorization loop type	NA
Unroll factor	NA
CQA cycles	4.50
CQA cycles if no scalar integer	4.00
CQA cycles if FP arith vectorized	2.00
CQA cycles if fully vectorized	0.32
Front-end cycles	4.50
P0 cycles	3.50
P1 cycles	3.30
P2 cycles	0.67
P3 cycles	0.67
P4 cycles	0.50
P5 cycles	3.40
P6 cycles	3.40
P7 cycles	0.50
P8 cycles	0.50
P9 cycles	0.50
P10 cycles	3.40
P11 cycles	0.67
DIV/SQRT cycles	0.00
Inter-iter dependencies cycles	4
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	26.00
Nb uops	27.00
Nb loads	2.00
Nb stores	1.00
Nb stack references	1.00
FLOP/cycle	0.67
Nb FLOP add-sub	0.00
Nb FLOP mul	1.00
Nb FLOP fma	1.00
Nb FLOP div	0.00
Nb FLOP rcp	0.00
Nb FLOP sqrt	0.00
Nb FLOP rsqrt	0.00
Bytes/cycle	4.44
Bytes prefetched	0.00
Bytes loaded	12.00
Bytes stored	8.00
Stride 0	NA
Stride 1	NA
Stride n	NA
Stride unknown	NA
Stride indirect	NA
Vectorization ratio all	0.00
Vectorization ratio load	0.00
Vectorization ratio store	0.00
Vectorization ratio mul	0.00
Vectorization ratio add_sub	0.00
Vectorization ratio fma	0.00
Vectorization ratio div_sqrt	NA
Vectorization ratio other	0.00
Vector-efficiency ratio all	9.62
Vector-efficiency ratio load	6.25
Vector-efficiency ratio store	12.50
Vector-efficiency ratio mul	6.25
Vector-efficiency ratio add_sub	12.50
Vector-efficiency ratio fma	6.25
Vector-efficiency ratio div_sqrt	NA
Vector-efficiency ratio other	9.72

Path /

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

Function	main
Source file and lines	random.tcc:401-3367
Module	attention-clang-gnr512

The loop is defined in /usr/lib/gcc/x86_64-redhat-linux/11/../../../../include/c++/11/bits/random.tcc:401-420,455-463,3364-3367.

Analyzed code is defined in /usr/lib/gcc/x86_64-redhat-linux/11/../../../../include/c++/11/bits/random.tcc:401-406,409,413,417-420,455,459-463,3364-3367.

Warnings:
Non-innermost loop: analyzing only self part (ignoring child loops).
The structure of this loop is probably <if then [else] end>.

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 probably not vectorized. Only 9% 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 10.58 to 0.89 cycles (11.86x 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)

FMA

Detected 1 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.

VCVTUSI2SS: 0 occurrences
VPEXTRQ: 0 occurrences

Slow data structures access

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

Details

Constant unknown stride: 2 occurrence(s)

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

Workaround

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

Conversion instructions

Detected expensive conversion instructions.

Details

VCVTUSI2SS (INT32/64 to FP32, scalar): 0 occurrences

Workaround

Avoid mixing data with different types. In particular, check if the type of constants is the same as array elements.

Type of elements and instruction set

2 SSE or AVX instructions are processing arithmetic or math operations on single precision FP elements in scalar mode (one at a time).

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

The binary loop is composed of 3 FP arithmetical operations:

1: addition or subtraction (all inside FMA instructions)
2: multiply (1 inside FMA instructions)

The binary loop is loading 80 bytes (20 single precision FP elements). The binary loop is storing 24 bytes (6 single precision FP elements).

Arithmetic intensity

Arithmetic intensity is 0.03 FP operations per loaded or stored byte.

General properties

nb instructions	61.50
nb uops	63.50
loop length	291
used x86 registers	7.50
used mmx registers	0
used xmm registers	6
used ymm registers	0
used zmm registers	2.50
nb stack references	5.50

Front-end

MACRO FUSION NOT POSSIBLE

micro-operation queue	10.58 cycles
front end	10.58 cycles

Back-end

	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11
uops	7.85	7.70	3.50	3.50	1.50	7.80	7.85	1.50	1.50	1.50	7.80	3.50
cycles	7.85	7.70	3.50	3.50	1.50	7.80	7.85	1.50	1.50	1.50	7.80	3.50

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)	4.00

Cycles summary

Front-end	10.58
Dispatch	7.90
Data deps.	4.00
Overall L1	10.58

Vectorization ratios

INT

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

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

INT+FP

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

Vector efficiency ratios

INT

all	10%
load	12%
store	12%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	12%
fma	NA (no fma vectorizable/vectorized instructions)
other	9%

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

INT+FP

all	9%
load	9%
store	12%
mul	6%
add-sub	12%
fma	6%
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	9%

Cycles and memory resources usage

Assuming all data fit into the L1 cache, each iteration of the binary loop takes 10.58 cycles. At this rate:

3% of peak load performance is reached (5.77 out of 192.00 bytes loaded per cycle (GB/s @ 1GHz))
1% of peak store performance is reached (2.09 out of 128.00 bytes stored per cycle (GB/s @ 1GHz))

Function	main
Source file and lines	random.tcc:401-3367
Module	attention-clang-gnr512

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 16.67 to 5.00 cycles (3.33x speedup).

Workaround

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

Vectorization

Your loop is probably not vectorized. Only 10% 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 16.67 to 1.46 cycles (11.41x speedup).

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.

FMA

Detected 1 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.

VCVTUSI2SS: 1 occurrences►
- /usr/lib/gcc/x86_64-redhat-linux/11/../../../../include/c++/11/bits/random.tcc:3366
VPEXTRQ: 1 occurrences►
- /usr/lib/gcc/x86_64-redhat-linux/11/../../../../include/c++/11/bits/random.tcc:404

Slow data structures access

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

Details

Constant unknown stride: 2 occurrence(s)

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

Workaround

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

Conversion instructions

Detected expensive conversion instructions.

Details

VCVTUSI2SS (INT32/64 to FP32, scalar): 1 occurrences►
- /usr/lib/gcc/x86_64-redhat-linux/11/../../../../include/c++/11/bits/random.tcc:3366

Workaround

Avoid mixing data with different types. In particular, check if the type of constants is the same as array elements.

Type of elements and instruction set

2 SSE or AVX instructions are processing arithmetic or math operations on single precision FP elements in scalar mode (one at a time).

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

The binary loop is composed of 3 FP arithmetical operations:

1: addition or subtraction (all inside FMA instructions)
2: multiply (1 inside FMA instructions)

The binary loop is loading 148 bytes (37 single precision FP elements). The binary loop is storing 40 bytes (10 single precision FP elements).

Arithmetic intensity

Arithmetic intensity is 0.02 FP operations per loaded or stored byte.

General properties

nb instructions	97
nb uops	100
loop length	473
used x86 registers	10
used mmx registers	0
used xmm registers	8
used ymm registers	0
used zmm registers	5
nb stack references	10

Front-end

MACRO FUSION NOT POSSIBLE FIT IN UOP CACHE

micro-operation queue	16.67 cycles
front end	16.67 cycles

Back-end

	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11
uops	12.20	12.10	6.33	6.33	2.50	12.20	12.30	2.50	2.50	2.50	12.20	6.33
cycles	12.20	12.10	6.33	6.33	2.50	12.20	12.30	2.50	2.50	2.50	12.20	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)	4.00

Cycles summary

Front-end	16.67
Dispatch	12.30
Data deps.	4.00
Overall L1	16.67

Vectorization ratios

INT

all	1%
load	0%
store	0%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	0%
fma	NA (no fma vectorizable/vectorized instructions)
other	1%

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

INT+FP

all	1%
load	0%
store	0%
mul	0%
add-sub	0%
fma	0%
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	1%

Vector efficiency ratios

INT

all	10%
load	12%
store	12%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	12%
fma	NA (no fma vectorizable/vectorized instructions)
other	10%

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

INT+FP

all	10%
load	12%
store	12%
mul	6%
add-sub	12%
fma	6%
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	10%

Cycles and memory resources usage

Assuming all data fit into the L1 cache, each iteration of the binary loop takes 16.67 cycles. At this rate:

4% of peak load performance is reached (8.88 out of 192.00 bytes loaded per cycle (GB/s @ 1GHz))
1% of peak store performance is reached (2.40 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 16.67 to 12.30 cycles (1.36x speedup).

ASM code

In the binary file, the address of the loop is: 44d4

Instruction	Nb FU	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11	Latency	Recip. throughput	Vectorization
MOV %RBX,%RDX	1	0	0	0	0	0	0	0	0	0	0	0	0	1	0.17	N/A
INC %RBX	1	0	0	0	0	0	0	0	0	0	0	0	0	1	0.17	N/A
MOV %RBX,0x1938(%RSP)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	1	0.50	scal (12.5%)
MOV 0x5b8(%RSP,%RDX,8),%RDX	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	1	0.33	N/A
MOV %RDX,%RSI	1	0	0	0	0	0	0	0	0	0	0	0	0	1	0.17	scal (12.5%)
SHR $0xb,%RSI	1	0.50	0	0	0	0	0	0.50	0	0	0	0	0	0-2	0.50	scal (12.5%)
MOV %ESI,%ESI	1	0	0	0	0	0	0	0	0	0	0	0	0	1	0.17	scal (6.3%)
XOR %RDX,%RSI	1	0.20	0.20	0	0	0	0.20	0.20	0	0	0	0.20	0	1-2	0.20	scal (12.5%)
MOV %ESI,%EDX	1	0	0	0	0	0	0	0	0	0	0	0	0	1	0.17	N/A
SAL $0x7,%EDX	1	0.50	0	0	0	0	0	0.50	0	0	0	0	0	0-2	0.50	N/A
AND $-0x62d3a980,%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
XOR %RSI,%RDX	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 %EDX,%ESI	1	0	0	0	0	0	0	0	0	0	0	0	0	1	0.17	scal (6.3%)
SAL $0xf,%ESI	1	0.50	0	0	0	0	0	0.50	0	0	0	0	0	0-2	0.50	scal (6.3%)
AND $-0x103a0000,%ESI	1	0.20	0.20	0	0	0	0.20	0.20	0	0	0	0.20	0	1-2	0.20	scal (6.3%)
XOR %RDX,%RSI	1	0.20	0.20	0	0	0	0.20	0.20	0	0	0	0.20	0	1-2	0.20	scal (12.5%)
MOV %RSI,%RDX	1	0	0	0	0	0	0	0	0	0	0	0	0	1	0.17	N/A
SHR $0x12,%RDX	1	0.50	0	0	0	0	0	0.50	0	0	0	0	0	0-2	0.50	N/A
XOR %RSI,%RDX	1	0.20	0.20	0	0	0	0.20	0.20	0	0	0	0.20	0	1-2	0.20	N/A
VCVTUSI2SS %RDX,%XMM15,%XMM2	3	0.50	0.50	0	0	0	2	0	0	0	0	0	0	8	2	scal (12.5%)
VFMADD231SS %XMM2,%XMM1,%XMM0	1	0.50	0.50	0	0	0	0	0	0	0	0	0	0	4	0.50	scal (6.3%)
VMULSS 0x4b3d(%RIP),%XMM1,%XMM1	1	0.50	0.50	0.33	0.33	0	0	0	0	0	0	0	0.33	4	0.50	scal (6.3%)
DEC %RCX	1	0	0	0	0	0	0	0	0	0	0	0	0	1	0.17	scal (12.5%)
JE 4840 <main+0xfc0>	1	0.50	0	0	0	0	0	0.50	0	0	0	0	0	0	0.50	N/A
CMP $0x270,%RBX	1	0.20	0.20	0	0	0	0.20	0.20	0	0	0	0.20	0	1	0.20	N/A
JB 4470 <main+0xbf0>	1	0.50	0	0	0	0	0	0.50	0	0	0	0	0	0	0.50	N/A
VPBROADCASTQ %R13,%ZMM2	1	0	0	0	0	0	1	0	0	0	0	0	0	3	1	scal (12.5%)
XOR %EDX,%EDX	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	N/A
VPBROADCASTQ 0x4b51(%RIP),%ZMM14	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	3	0.33	scal (12.5%)
VPBROADCASTQ 0x4b4f(%RIP),%ZMM15	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	3	0.33	scal (12.5%)
VPBROADCASTQ 0x4b4d(%RIP),%ZMM16	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	3	0.33	scal (12.5%)
VPBROADCASTQ 0x4b4b(%RIP),%ZMM17	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	3	0.33	scal (12.5%)
NOPL (%RAX)	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	N/A
VEXTRACTI32X4 $0x3,%ZMM2,%XMM2	1	0	0	0	0	0	1	0	0	0	0	0	0	3	1	vect (25.0%)
VPEXTRQ $0x1,%XMM2,%RDI	2	1	0.50	0	0	0	0.50	0	0	0	0	0	0	4	1	scal (12.5%)
AND $-0x80000000,%RDI	1	0.20	0.20	0	0	0	0.20	0.20	0	0	0	0.20	0	1-2	0.20	scal (12.5%)
MOV 0xcc0(%RSP),%RSI	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	1	0.33	scal (12.5%)
MOV 0xcc8(%RSP),%RDX	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	1	0.33	N/A
MOV %ESI,%R8D	1	0	0	0	0	0	0	0	0	0	0	0	0	1	0.17	scal (6.3%)
AND $0x7ffffffe,%R8D	1	0.20	0.20	0	0	0	0.20	0.20	0	0	0	0.20	0	1-2	0.20	scal (6.3%)
OR %RDI,%R8	1	0.20	0.20	0	0	0	0.20	0.20	0	0	0	0.20	0	1-2	0.20	scal (12.5%)
SHR $0x1,%R8	1	0.50	0	0	0	0	0	0.50	0	0	0	0	0	0-2	0.50	scal (12.5%)
XOR 0x1920(%RSP),%R8	1	0.20	0.20	0.33	0.33	0	0.20	0.20	0	0	0	0.20	0.33	1-2	0.33	scal (12.5%)
MOV %ESI,%EDI	1	0	0	0	0	0	0	0	0	0	0	0	0	1	0.17	scal (6.3%)
AND $0x1,%EDI	1	0.20	0.20	0	0	0	0.20	0.20	0	0	0	0.20	0	1-2	0.20	scal (6.3%)
NEG %EDI	1	0.20	0.20	0	0	0	0.20	0.20	0	0	0	0.20	0	1	0.20	scal (6.3%)
AND %R12D,%EDI	1	0.20	0.20	0	0	0	0.20	0.20	0	0	0	0.20	0	1-2	0.20	scal (6.3%)
XOR %R8,%RDI	1	0.20	0.20	0	0	0	0.20	0.20	0	0	0	0.20	0	1-2	0.20	scal (12.5%)
MOV %RDI,0xcb8(%RSP)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	1	0.50	scal (12.5%)
AND $-0x80000000,%RSI	1	0.20	0.20	0	0	0	0.20	0.20	0	0	0	0.20	0	1-2	0.20	scal (12.5%)
MOV %EDX,%EDI	1	0	0	0	0	0	0	0	0	0	0	0	0	1	0.17	scal (6.3%)
AND $0x7ffffffe,%EDI	1	0.20	0.20	0	0	0	0.20	0.20	0	0	0	0.20	0	1-2	0.20	scal (6.3%)
OR %RSI,%RDI	1	0.20	0.20	0	0	0	0.20	0.20	0	0	0	0.20	0	1-2	0.20	scal (12.5%)
SHR $0x1,%RDI	1	0.50	0	0	0	0	0	0.50	0	0	0	0	0	0-2	0.50	scal (12.5%)
XOR 0x1928(%RSP),%RDI	1	0.20	0.20	0.33	0.33	0	0.20	0.20	0	0	0	0.20	0.33	1-2	0.33	scal (12.5%)
MOV %EDX,%ESI	1	0	0	0	0	0	0	0	0	0	0	0	0	1	0.17	scal (6.3%)
AND $0x1,%ESI	1	0.20	0.20	0	0	0	0.20	0.20	0	0	0	0.20	0	1-2	0.20	scal (6.3%)
NEG %ESI	1	0.20	0.20	0	0	0	0.20	0.20	0	0	0	0.20	0	1	0.20	scal (6.3%)
AND %R12D,%ESI	1	0.20	0.20	0	0	0	0.20	0.20	0	0	0	0.20	0	1-2	0.20	scal (6.3%)
XOR %RDI,%RSI	1	0.20	0.20	0	0	0	0.20	0.20	0	0	0	0.20	0	1-2	0.20	scal (12.5%)
MOV %RSI,0xcc0(%RSP)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	1	0.50	scal (12.5%)
AND $-0x80000000,%RDX	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 0xcd0(%RSP),%RSI	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	1	0.33	scal (12.5%)
MOV %ESI,%EDI	1	0	0	0	0	0	0	0	0	0	0	0	0	1	0.17	scal (6.3%)
VPBROADCASTQ %RSI,%XMM2	1	0	0	0	0	0	1	0	0	0	0	0	0	3	1	scal (12.5%)
AND $0x7ffffffe,%ESI	1	0.20	0.20	0	0	0	0.20	0.20	0	0	0	0.20	0	1-2	0.20	scal (6.3%)
OR %RDX,%RSI	1	0.20	0.20	0	0	0	0.20	0.20	0	0	0	0.20	0	1-2	0.20	scal (12.5%)
SHR $0x1,%RSI	1	0.50	0	0	0	0	0	0.50	0	0	0	0	0	0-2	0.50	scal (12.5%)
XOR 0x1930(%RSP),%RSI	1	0.20	0.20	0.33	0.33	0	0.20	0.20	0	0	0	0.20	0.33	1-2	0.33	scal (12.5%)
AND $0x1,%EDI	1	0.20	0.20	0	0	0	0.20	0.20	0	0	0	0.20	0	1-2	0.20	scal (6.3%)
NEG %EDI	1	0.20	0.20	0	0	0	0.20	0.20	0	0	0	0.20	0	1	0.20	scal (6.3%)
AND %R12D,%EDI	1	0.20	0.20	0	0	0	0.20	0.20	0	0	0	0.20	0	1-2	0.20	scal (6.3%)
XOR %RSI,%RDI	1	0.20	0.20	0	0	0	0.20	0.20	0	0	0	0.20	0	1-2	0.20	scal (12.5%)
MOV %RDI,0xcc8(%RSP)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	1	0.50	scal (12.5%)
MOV $0xe8,%EDX	1	0.20	0.20	0	0	0	0.20	0.20	0	0	0	0.20	0	1	0.20	N/A
VPBROADCASTQ 0x493c(%RIP),%XMM7	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	3	0.33	scal (12.5%)
VPBROADCASTQ 0x493b(%RIP),%XMM8	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	3	0.33	scal (12.5%)
VPBROADCASTQ 0x493a(%RIP),%XMM9	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	3	0.33	scal (12.5%)
VPBROADCASTQ 0x4939(%RIP),%XMM10	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	3	0.33	scal (12.5%)
NOP	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	N/A
MOV 0x1930(%RSP),%RDX	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	1	0.33	N/A
MOV $-0x80000000,%RSI	1	0.20	0.20	0	0	0	0.20	0.20	0	0	0	0.20	0	1	0.20	scal (6.3%)
AND %RSI,%RDX	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 0x5b8(%RSP),%R13	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	1	0.33	scal (12.5%)
MOV %R13D,%ESI	1	0	0	0	0	0	0	0	0	0	0	0	0	1	0.17	scal (6.3%)
AND $0x7ffffffe,%ESI	1	0.20	0.20	0	0	0	0.20	0.20	0	0	0	0.20	0	1-2	0.20	scal (6.3%)
OR %RDX,%RSI	1	0.20	0.20	0	0	0	0.20	0.20	0	0	0	0.20	0	1-2	0.20	scal (12.5%)
SHR $0x1,%RSI	1	0.50	0	0	0	0	0	0.50	0	0	0	0	0	0-2	0.50	scal (12.5%)
XOR 0x1218(%RSP),%RSI	1	0.20	0.20	0.33	0.33	0	0.20	0.20	0	0	0	0.20	0.33	1-2	0.33	scal (12.5%)
MOV %R13D,%EDX	1	0	0	0	0	0	0	0	0	0	0	0	0	1	0.17	N/A
AND $0x1,%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
NEG %EDX	1	0.20	0.20	0	0	0	0.20	0.20	0	0	0	0.20	0	1	0.20	N/A
AND %R12D,%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
XOR %RSI,%RDX	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 %RDX,0x1930(%RSP)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	1	0.50	scal (12.5%)
XOR %EBX,%EBX	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	N/A
JMP 4470 <main+0xbf0>	1	0	0	0	0	0	0	0	0	0	0	0	0	0	2.08	N/A

Function	main
Source file and lines	random.tcc:401-3367
Module	attention-clang-gnr512

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 4.50 to 4.00 cycles (1.12x speedup).

Workaround

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

Vectorization

Your loop is not vectorized. Only 9% 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 4.50 to 0.32 cycles (13.87x speedup).

Details

All 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.

FMA

Detected 1 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.

VCVTUSI2SS: 1 occurrences►
- /usr/lib/gcc/x86_64-redhat-linux/11/../../../../include/c++/11/bits/random.tcc:3366

Conversion instructions

Detected expensive conversion instructions.

Details

VCVTUSI2SS (INT32/64 to FP32, scalar): 1 occurrences►
- /usr/lib/gcc/x86_64-redhat-linux/11/../../../../include/c++/11/bits/random.tcc:3366

Workaround

Avoid mixing data with different types. In particular, check if the type of constants is the same as array elements.

Type of elements and instruction set

2 SSE or AVX instructions are processing arithmetic or math operations on single precision FP elements in scalar mode (one at a time).

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

The binary loop is composed of 3 FP arithmetical operations:

1: addition or subtraction (all inside FMA instructions)
2: multiply (1 inside FMA instructions)

The binary loop is loading 12 bytes (3 single precision FP elements). The binary loop is storing 8 bytes (2 single precision FP elements).

Arithmetic intensity

Arithmetic intensity is 0.15 FP operations per loaded or stored byte.

General properties

nb instructions	26
nb uops	27
loop length	109
used x86 registers	5
used mmx registers	0
used xmm registers	4
used ymm registers	0
used zmm registers	0
nb stack references	1

Front-end

ASSUMED MACRO FUSION FIT IN UOP CACHE

micro-operation queue	4.50 cycles
front end	4.50 cycles

Back-end

	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11
uops	3.50	3.30	0.67	0.67	0.50	3.40	3.40	0.50	0.50	0.50	3.40	0.67
cycles	3.50	3.30	0.67	0.67	0.50	3.40	3.40	0.50	0.50	0.50	3.40	0.67

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)	4.00

Cycles summary

Front-end	4.50
Dispatch	3.50
Data deps.	4.00
Overall L1	4.50

Vectorization ratios

INT

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

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

INT+FP

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

Vector efficiency ratios

INT

all	10%
load	NA (no load vectorizable/vectorized instructions)
store	12%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	12%
fma	NA (no fma vectorizable/vectorized instructions)
other	9%

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

INT+FP

all	9%
load	6%
store	12%
mul	6%
add-sub	12%
fma	6%
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	9%

Cycles and memory resources usage

Assuming all data fit into the L1 cache, each iteration of the binary loop takes 4.50 cycles. At this rate:

1% of peak load performance is reached (2.67 out of 192.00 bytes loaded per cycle (GB/s @ 1GHz))
1% of peak store performance is reached (1.78 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: 44d4

Instruction	Nb FU	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11	Latency	Recip. throughput	Vectorization
MOV %RBX,%RDX	1	0	0	0	0	0	0	0	0	0	0	0	0	1	0.17	N/A
INC %RBX	1	0	0	0	0	0	0	0	0	0	0	0	0	1	0.17	N/A
MOV %RBX,0x1938(%RSP)	1	0	0	0	0	0.50	0	0	0.50	0.50	0.50	0	0	1	0.50	scal (12.5%)
MOV 0x5b8(%RSP,%RDX,8),%RDX	1	0	0	0.33	0.33	0	0	0	0	0	0	0	0.33	1	0.33	N/A
MOV %RDX,%RSI	1	0	0	0	0	0	0	0	0	0	0	0	0	1	0.17	scal (12.5%)
SHR $0xb,%RSI	1	0.50	0	0	0	0	0	0.50	0	0	0	0	0	0-2	0.50	scal (12.5%)
MOV %ESI,%ESI	1	0	0	0	0	0	0	0	0	0	0	0	0	1	0.17	scal (6.3%)
XOR %RDX,%RSI	1	0.20	0.20	0	0	0	0.20	0.20	0	0	0	0.20	0	1-2	0.20	scal (12.5%)
MOV %ESI,%EDX	1	0	0	0	0	0	0	0	0	0	0	0	0	1	0.17	N/A
SAL $0x7,%EDX	1	0.50	0	0	0	0	0	0.50	0	0	0	0	0	0-2	0.50	N/A
AND $-0x62d3a980,%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
XOR %RSI,%RDX	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 %EDX,%ESI	1	0	0	0	0	0	0	0	0	0	0	0	0	1	0.17	scal (6.3%)
SAL $0xf,%ESI	1	0.50	0	0	0	0	0	0.50	0	0	0	0	0	0-2	0.50	scal (6.3%)
AND $-0x103a0000,%ESI	1	0.20	0.20	0	0	0	0.20	0.20	0	0	0	0.20	0	1-2	0.20	scal (6.3%)
XOR %RDX,%RSI	1	0.20	0.20	0	0	0	0.20	0.20	0	0	0	0.20	0	1-2	0.20	scal (12.5%)
MOV %RSI,%RDX	1	0	0	0	0	0	0	0	0	0	0	0	0	1	0.17	N/A
SHR $0x12,%RDX	1	0.50	0	0	0	0	0	0.50	0	0	0	0	0	0-2	0.50	N/A
XOR %RSI,%RDX	1	0.20	0.20	0	0	0	0.20	0.20	0	0	0	0.20	0	1-2	0.20	N/A
VCVTUSI2SS %RDX,%XMM15,%XMM2	3	0.50	0.50	0	0	0	2	0	0	0	0	0	0	8	2	scal (12.5%)
VFMADD231SS %XMM2,%XMM1,%XMM0	1	0.50	0.50	0	0	0	0	0	0	0	0	0	0	4	0.50	scal (6.3%)
VMULSS 0x4b3d(%RIP),%XMM1,%XMM1	1	0.50	0.50	0.33	0.33	0	0	0	0	0	0	0	0.33	4	0.50	scal (6.3%)
DEC %RCX	1	0	0	0	0	0	0	0	0	0	0	0	0	1	0.17	scal (12.5%)
JE 4840 <main+0xfc0>	1	0.50	0	0	0	0	0	0.50	0	0	0	0	0	0	0.50	N/A
CMP $0x270,%RBX	1	0.20	0.20	0	0	0	0.20	0.20	0	0	0	0.20	0	1	0.20	N/A
JB 4470 <main+0xbf0>	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

FMA

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

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

FMA

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

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

FMA

Complex instructions

Details

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