OV - attention-clang-znver5-512 - Loop 94 - attention-clang-znver5-512

	Loop Id: 94	Module: attention-clang-znver5-512	Source: random.tcc:401-3367 [...]	Coverage: 0.17%

0x54a0 MOV	%R13,%RCX

0x54a3 INC	%R13

0x54a6 MOV	$0x200b,%EDX

0x54ab MOV	%R13,0x19b8(%RSP)

0x54b3 MOV	0x638(%RSP,%RCX,8),%RCX

0x54bb BEXTR	%RDX,%RCX,%RDX

0x54c0 XOR	%RCX,%RDX

0x54c3 MOV	%EDX,%ECX

0x54c5 SAL	$0x7,%ECX

0x54c8 AND	$-0x62d3a980,%ECX

0x54ce XOR	%RDX,%RCX

0x54d1 MOV	%ECX,%EDX

0x54d3 SAL	$0xf,%EDX

0x54d6 AND	$-0x103a0000,%EDX

0x54dc XOR	%RCX,%RDX

0x54df MOV	%RDX,%RCX

0x54e2 SHR	$0x12,%RCX

0x54e6 XOR	%RDX,%RCX

0x54e9 DEC	%RAX

0x54ec VCVTUSI2SS	%RCX,%XMM15,%XMM2

0x54f2 VFMADD231SS	%XMM2,%XMM1,%XMM0

0x54f7 VMULSS	0x3b09(%RIP),%XMM1,%XMM1

0x54ff JE	5920

0x5505 CMP	$0x270,%R13

0x550c JB	54a0

0x550e VPBROADCASTQ	0x3b28(%RIP),%ZMM14

0x5518 VPBROADCASTQ	0x3b26(%RIP),%ZMM15

0x5522 VPBROADCASTQ	0x3b24(%RIP),%ZMM16

0x552c VPBROADCASTQ	0x3b22(%RIP),%ZMM17

0x5536 VPBROADCASTQ	%RBX,%ZMM2

0x553c XOR	%ECX,%ECX

0x553e XCHG	%AX,%AX

  (95) 0x5540 VMOVDQU64	0x640(%RSP,%RCX,8),%ZMM3

  (95) 0x5548 VMOVDQU64	0x680(%RSP,%RCX,8),%ZMM4

  (95) 0x5550 VMOVDQU64	0x6c0(%RSP,%RCX,8),%ZMM5

  (95) 0x5558 VALIGNQ	$0x7,%ZMM2,%ZMM3,%ZMM6

  (95) 0x555f VMOVDQU64	0x700(%RSP,%RCX,8),%ZMM2

  (95) 0x5567 VALIGNQ	$0x7,%ZMM3,%ZMM4,%ZMM7

  (95) 0x556e VALIGNQ	$0x7,%ZMM4,%ZMM5,%ZMM8

  (95) 0x5575 VPANDQ	%ZMM15,%ZMM3,%ZMM10

  (95) 0x557b VPANDQ	%ZMM15,%ZMM4,%ZMM11

  (95) 0x5581 VPANDQ	%ZMM15,%ZMM5,%ZMM12

  (95) 0x5587 VPTESTMQ	%ZMM16,%ZMM3,%K1

  (95) 0x558d VPTESTMQ	%ZMM16,%ZMM4,%K2

  (95) 0x5593 VPTESTMQ	%ZMM16,%ZMM5,%K3

  (95) 0x5599 VPTERNLOGQ	$-0x8,%ZMM14,%ZMM6,%ZMM10

  (95) 0x55a0 VPTERNLOGQ	$-0x8,%ZMM14,%ZMM7,%ZMM11

  (95) 0x55a7 VPTERNLOGQ	$-0x8,%ZMM14,%ZMM8,%ZMM12

  (95) 0x55ae VPSRLQ	$0x1,%ZMM10,%ZMM6

  (95) 0x55b5 VPSRLQ	$0x1,%ZMM11,%ZMM7

  (95) 0x55bc VPSRLQ	$0x1,%ZMM12,%ZMM8

  (95) 0x55c3 VPXORQ	0x12a0(%RSP,%RCX,8),%ZMM6,%ZMM6

  (95) 0x55ce VPXORQ	0x12e0(%RSP,%RCX,8),%ZMM7,%ZMM7

  (95) 0x55d9 VPXORQ	0x1320(%RSP,%RCX,8),%ZMM8,%ZMM8

  (95) 0x55e4 VALIGNQ	$0x7,%ZMM5,%ZMM2,%ZMM9

  (95) 0x55eb VPANDQ	%ZMM15,%ZMM2,%ZMM13

  (95) 0x55f1 VPTESTMQ	%ZMM16,%ZMM2,%K4

  (95) 0x55f7 VPTERNLOGQ	$-0x8,%ZMM14,%ZMM9,%ZMM13

  (95) 0x55fe VPSRLQ	$0x1,%ZMM13,%ZMM9

  (95) 0x5605 VPXORQ	0x1360(%RSP,%RCX,8),%ZMM9,%ZMM9

  (95) 0x5610 VPXORQ	%ZMM17,%ZMM6,%ZMM6{%K1}

  (95) 0x5616 VPXORQ	%ZMM17,%ZMM7,%ZMM7{%K2}

  (95) 0x561c VPXORQ	%ZMM17,%ZMM8,%ZMM8{%K3}

  (95) 0x5622 VMOVDQU64	%ZMM6,0x638(%RSP,%RCX,8)

  (95) 0x562d VMOVDQU64	%ZMM7,0x678(%RSP,%RCX,8)

  (95) 0x5638 VMOVDQU64	%ZMM8,0x6b8(%RSP,%RCX,8)

  (95) 0x5643 VPXORQ	%ZMM17,%ZMM9,%ZMM9{%K4}

  (95) 0x5649 VMOVDQU64	%ZMM9,0x6f8(%RSP,%RCX,8)

  (95) 0x5654 ADD	$0x20,%RCX

  (95) 0x5658 CMP	$0xe0,%RCX

  (95) 0x565f JNE	5540

0x5665 MOV	0xd40(%RSP),%RDX

0x566d VEXTRACTI32X4	$0x3,%ZMM2,%XMM2

0x5674 MOV	0xd48(%RSP),%RCX

0x567c VPBROADCASTQ	0x39bb(%RIP),%XMM8

0x5685 VPBROADCASTQ	0x39ba(%RIP),%XMM9

0x568e VPBROADCASTQ	0x39b9(%RIP),%XMM10

0x5697 VPBROADCASTQ	0x39b8(%RIP),%XMM11

0x56a0 VPEXTRQ	$0x1,%XMM2,%RSI

0x56a6 AND	$-0x80000000,%RSI

0x56ad MOV	%EDX,%EDI

0x56af AND	$0x7ffffffe,%EDI

0x56b5 OR	%RSI,%RDI

0x56b8 MOV	%EDX,%ESI

0x56ba AND	$0x1,%ESI

0x56bd AND	$-0x80000000,%RDX

0x56c4 SHR	$0x1,%RDI

0x56c7 XOR	0x19a0(%RSP),%RDI

0x56cf NEG	%ESI

0x56d1 AND	%R12D,%ESI

0x56d4 XOR	%RDI,%RSI

0x56d7 MOV	%RSI,0xd38(%RSP)

0x56df MOV	%ECX,%ESI

0x56e1 AND	$0x7ffffffe,%ESI

0x56e7 OR	%RDX,%RSI

0x56ea MOV	%ECX,%EDX

0x56ec AND	$0x1,%EDX

0x56ef AND	$-0x80000000,%RCX

0x56f6 SHR	$0x1,%RSI

0x56f9 XOR	0x19a8(%RSP),%RSI

0x5701 NEG	%EDX

0x5703 AND	%R12D,%EDX

0x5706 XOR	%RSI,%RDX

0x5709 MOV	%RDX,0xd40(%RSP)

0x5711 MOV	0xd50(%RSP),%RDX

0x5719 MOV	%EDX,%ESI

0x571b VPBROADCASTQ	%RDX,%XMM2

0x5721 AND	$0x7ffffffe,%EDX

0x5727 AND	$0x1,%ESI

0x572a OR	%RCX,%RDX

0x572d NEG	%ESI

0x572f MOV	$0xee,%ECX

0x5734 SHR	$0x1,%RDX

0x5737 XOR	0x19b0(%RSP),%RDX

0x573f AND	%R12D,%ESI

0x5742 XOR	%RDX,%RSI

0x5745 MOV	%RSI,0xd48(%RSP)

0x574d NOPL	(%RAX)

  (96) 0x5750 VMOVDQU	0x5e8(%RSP,%RCX,8),%XMM3

  (96) 0x5759 VMOVDQU	0x5f8(%RSP,%RCX,8),%XMM4

  (96) 0x5762 VMOVDQU	0x608(%RSP,%RCX,8),%XMM5

  (96) 0x576b VMOVDQU	0x618(%RSP,%RCX,8),%XMM6

  (96) 0x5774 VPALIGNR	$0x8,%XMM2,%XMM3,%XMM2

  (96) 0x577a VPAND	%XMM3,%XMM9,%XMM7

  (96) 0x577e VPTESTMQ	%XMM10,%XMM3,%K1

  (96) 0x5784 VPTERNLOGQ	$-0x8,%XMM8,%XMM2,%XMM7

  (96) 0x578b VPSRLQ	$0x1,%XMM7,%XMM2

  (96) 0x5790 VPXOR	-0x138(%RSP,%RCX,8),%XMM2,%XMM2

  (96) 0x5799 VPXORQ	%XMM11,%XMM2,%XMM2{%K1}

  (96) 0x579f VPTESTMQ	%XMM10,%XMM4,%K1

  (96) 0x57a5 VMOVDQU	%XMM2,0x5e0(%RSP,%RCX,8)

  (96) 0x57ae VPALIGNR	$0x8,%XMM3,%XMM4,%XMM2

  (96) 0x57b4 VPAND	%XMM4,%XMM9,%XMM3

  (96) 0x57b8 VPTERNLOGQ	$-0x8,%XMM8,%XMM2,%XMM3

  (96) 0x57bf VPSRLQ	$0x1,%XMM3,%XMM2

  (96) 0x57c4 VPXOR	-0x128(%RSP,%RCX,8),%XMM2,%XMM2

  (96) 0x57cd VPAND	%XMM5,%XMM9,%XMM3

  (96) 0x57d1 VPXORQ	%XMM11,%XMM2,%XMM2{%K1}

  (96) 0x57d7 VPTESTMQ	%XMM10,%XMM5,%K1

  (96) 0x57dd VMOVDQU	%XMM2,0x5f0(%RSP,%RCX,8)

  (96) 0x57e6 VPALIGNR	$0x8,%XMM4,%XMM5,%XMM2

  (96) 0x57ec VPTERNLOGQ	$-0x8,%XMM8,%XMM2,%XMM3

  (96) 0x57f3 VPSRLQ	$0x1,%XMM3,%XMM2

  (96) 0x57f8 VPXOR	-0x118(%RSP,%RCX,8),%XMM2,%XMM2

  (96) 0x5801 VPAND	%XMM6,%XMM9,%XMM3

  (96) 0x5805 VPXORQ	%XMM11,%XMM2,%XMM2{%K1}

  (96) 0x580b VPTESTMQ	%XMM10,%XMM6,%K1

  (96) 0x5811 VMOVDQU	%XMM2,0x600(%RSP,%RCX,8)

  (96) 0x581a VPALIGNR	$0x8,%XMM5,%XMM6,%XMM2

  (96) 0x5820 VPTERNLOGQ	$-0x8,%XMM8,%XMM2,%XMM3

  (96) 0x5827 VPSRLQ	$0x1,%XMM3,%XMM2

  (96) 0x582c VPXOR	-0x108(%RSP,%RCX,8),%XMM2,%XMM2

  (96) 0x5835 VPXORQ	%XMM11,%XMM2,%XMM2{%K1}

  (96) 0x583b VMOVDQU	%XMM2,0x610(%RSP,%RCX,8)

  (96) 0x5844 VMOVDQU	0x628(%RSP,%RCX,8),%XMM3

  (96) 0x584d VPALIGNR	$0x8,%XMM6,%XMM3,%XMM2

  (96) 0x5853 VPAND	%XMM3,%XMM9,%XMM4

  (96) 0x5857 VPTESTMQ	%XMM10,%XMM3,%K1

  (96) 0x585d VPTERNLOGQ	$-0x8,%XMM8,%XMM2,%XMM4

  (96) 0x5864 VPSRLQ	$0x1,%XMM4,%XMM2

  (96) 0x5869 VPXOR	-0xf8(%RSP,%RCX,8),%XMM2,%XMM2

  (96) 0x5872 VPXORQ	%XMM11,%XMM2,%XMM2{%K1}

  (96) 0x5878 VMOVDQU	%XMM2,0x620(%RSP,%RCX,8)

  (96) 0x5881 VMOVDQU	0x638(%RSP,%RCX,8),%XMM2

  (96) 0x588a VPALIGNR	$0x8,%XMM3,%XMM2,%XMM3

  (96) 0x5890 VPAND	%XMM2,%XMM9,%XMM4

  (96) 0x5894 VPTESTMQ	%XMM10,%XMM2,%K1

  (96) 0x589a VPTERNLOGQ	$-0x8,%XMM8,%XMM3,%XMM4

  (96) 0x58a1 VPSRLQ	$0x1,%XMM4,%XMM3

  (96) 0x58a6 VPXOR	-0xe8(%RSP,%RCX,8),%XMM3,%XMM3

  (96) 0x58af VPXORQ	%XMM11,%XMM3,%XMM3{%K1}

  (96) 0x58b5 VMOVDQU	%XMM3,0x630(%RSP,%RCX,8)

  (96) 0x58be ADD	$0xc,%RCX

  (96) 0x58c2 CMP	$0x27a,%RCX

  (96) 0x58c9 JNE	5750

0x58cf MOV	0x19b0(%RSP),%RCX

0x58d7 MOV	0x638(%RSP),%RBX

0x58df XOR	%R13D,%R13D

0x58e2 MOV	%EBX,%EDX

0x58e4 AND	%R14,%RCX

0x58e7 AND	$0x7ffffffe,%EDX

0x58ed OR	%RCX,%RDX

0x58f0 MOV	%EBX,%ECX

0x58f2 AND	$0x1,%ECX

0x58f5 SHR	$0x1,%RDX

0x58f8 XOR	0x1298(%RSP),%RDX

0x5900 NEG	%ECX

0x5902 AND	%R12D,%ECX

0x5905 XOR	%RDX,%RCX

0x5908 MOV	%RCX,0x19b0(%RSP)

0x5910 JMP	54a0

/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	1.93
CQA speedup if FP arith vectorized	2.04
CQA speedup if fully vectorized	12.09
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.35
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	8.19
CQA cycles if no scalar integer	4.25
CQA cycles if FP arith vectorized	4.01
CQA cycles if fully vectorized	0.68
Front-end cycles	7.75
P0 cycles	5.75
P1 cycles	5.75
P2 cycles	5.75
P3 cycles	5.75
P4 cycles	5.75
P5 cycles	5.75
P6 cycles	3.38
P7 cycles	3.38
P8 cycles	3.38
P9 cycles	3.38
P10 cycles	2.13
P11 cycles	2.75
P12 cycles	2.25
P13 cycles	1.88
P14 cycles	0.25
P15 cycles	0.25
DIV/SQRT cycles	0.00
Inter-iter dependencies cycles	4
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	60.00
Nb uops	62.00
Nb loads	10.50
Nb stores	3.00
Nb stack references	5.50
FLOP/cycle	0.37
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	10.10
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	0.77
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	0.91
Vector-efficiency ratio all	9.69
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	9.52

Metric	Value
CQA speedup if no scalar integer	2.75
CQA speedup if FP arith vectorized	1.78
CQA speedup if fully vectorized	11.52
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.38
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	12.38
CQA cycles if no scalar integer	4.50
CQA cycles if FP arith vectorized	6.94
CQA cycles if fully vectorized	1.07
Front-end cycles	12.38
P0 cycles	9.00
P1 cycles	9.00
P2 cycles	9.00
P3 cycles	9.00
P4 cycles	9.00
P5 cycles	9.00
P6 cycles	6.00
P7 cycles	6.00
P8 cycles	6.00
P9 cycles	6.00
P10 cycles	3.25
P11 cycles	4.50
P12 cycles	4.00
P13 cycles	3.25
P14 cycles	0.50
P15 cycles	0.50
DIV/SQRT cycles	0.00
Inter-iter dependencies cycles	4
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	95.00
Nb uops	99.00
Nb loads	19.00
Nb stores	5.00
Nb stack references	10.00
FLOP/cycle	0.24
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	15.19
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	1.54
Vectorization ratio load	0.00
Vectorization ratio store	0.00
Vectorization ratio mul	0.00
Vectorization ratio add_sub	NA
Vectorization ratio fma	0.00
Vectorization ratio div_sqrt	NA
Vectorization ratio other	1.82
Vector-efficiency ratio all	10.29
Vector-efficiency ratio load	12.11
Vector-efficiency ratio store	12.50
Vector-efficiency ratio mul	6.25
Vector-efficiency ratio add_sub	NA
Vector-efficiency ratio fma	6.25
Vector-efficiency ratio div_sqrt	NA
Vector-efficiency ratio other	10.11

Metric	Value
CQA speedup if no scalar integer	1.00
CQA speedup if FP arith vectorized	3.72
CQA speedup if fully vectorized	14.29
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.28
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	4.00
CQA cycles if no scalar integer	4.00
CQA cycles if FP arith vectorized	1.07
CQA cycles if fully vectorized	0.28
Front-end cycles	3.13
P0 cycles	2.50
P1 cycles	2.50
P2 cycles	2.50
P3 cycles	2.50
P4 cycles	2.50
P5 cycles	2.50
P6 cycles	0.75
P7 cycles	0.75
P8 cycles	0.75
P9 cycles	0.75
P10 cycles	1.00
P11 cycles	1.00
P12 cycles	0.50
P13 cycles	0.50
P14 cycles	0.00
P15 cycles	0.00
DIV/SQRT cycles	0.00
Inter-iter dependencies cycles	4
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	25.00
Nb uops	25.00
Nb loads	2.00
Nb stores	1.00
Nb stack references	1.00
FLOP/cycle	0.75
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	5.00
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	9.09
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	8.93

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-znver5-512

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 8.19 to 0.68 cycles (12.09x 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
VPBROADCASTQ: 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	60
nb uops	62
loop length	287.50
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	7.75 cycles
front end	7.75 cycles

Back-end

	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11	P12	P13	P14	P15
uops	5.75	5.75	5.75	5.75	5.75	5.75	3.38	3.38	3.38	3.38	2.13	2.75	2.25	1.88	0.25	0.25
cycles	5.75	5.75	5.75	5.75	5.75	5.75	3.38	3.38	3.38	3.38	2.13	2.75	2.25	1.88	0.25	0.25

Execution ports to units layout:

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

Cycles executing div or sqrt instructions	NA
Longest recurrence chain latency (RecMII)	4.00

Cycles summary

Front-end	7.75
Dispatch	5.75
Data deps.	4.00
Overall L1	8.19

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 8.19 cycles. At this rate:

2% of peak load performance is reached (7.48 out of 256.00 bytes loaded per cycle (GB/s @ 1GHz))
2% of peak store performance is reached (2.62 out of 128.00 bytes stored per cycle (GB/s @ 1GHz))

Function	main
Source file and lines	random.tcc:401-3367
Module	attention-clang-znver5-512

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 12.38 to 4.50 cycles (2.75x 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 12.38 to 1.07 cycles (11.52x 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
VPBROADCASTQ: 2 occurrences►
- /usr/lib/gcc/x86_64-redhat-linux/11/../../../../include/c++/11/bits/random.tcc:401
- /usr/lib/gcc/x86_64-redhat-linux/11/../../../../include/c++/11/bits/random.tcc:409
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	95
nb uops	99
loop length	465
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	12.38 cycles
front end	12.38 cycles

Back-end

	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11	P12	P13	P14	P15
uops	9.00	9.00	9.00	9.00	9.00	9.00	6.00	6.00	6.00	6.00	3.25	4.50	4.00	3.25	0.50	0.50
cycles	9.00	9.00	9.00	9.00	9.00	9.00	6.00	6.00	6.00	6.00	3.25	4.50	4.00	3.25	0.50	0.50

Execution ports to units layout:

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

Cycles executing div or sqrt instructions	NA
Longest recurrence chain latency (RecMII)	4.00

Cycles summary

Front-end	12.38
Dispatch	9.00
Data deps.	4.00
Overall L1	12.38

Vectorization ratios

INT

all	1%
load	0%
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)
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	NA (no add-sub vectorizable/vectorized instructions)
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	NA (no add-sub vectorizable/vectorized instructions)
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	NA (no add-sub vectorizable/vectorized instructions)
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 12.38 cycles. At this rate:

4% of peak load performance is reached (11.96 out of 256.00 bytes loaded per cycle (GB/s @ 1GHz))
2% of peak store performance is reached (3.23 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 12.38 to 9.00 cycles (1.37x speedup).

ASM code

In the binary file, the address of the loop is: 5505

Instruction	Nb FU	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11	P12	P13	P14	P15	Latency	Recip. throughput	Vectorization
MOV %R13,%RCX	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.13	N/A
INC %R13	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	N/A
MOV $0x200b,%EDX	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	scal (6.3%)
MOV %R13,0x19b8(%RSP)	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	4	0.50	scal (12.5%)
MOV 0x638(%RSP,%RCX,8),%RCX	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	3	0.25	N/A
BEXTR %RDX,%RCX,%RDX	1	0	0	0	0.33	0.33	0.33	0	0	0	0	0	0	0	0	0	0	1	0.33	N/A
XOR %RCX,%RDX	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	scal (12.5%)
MOV %EDX,%ECX	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.13	N/A
SAL $0x7,%ECX	1	0	0	0	0.33	0.33	0.33	0	0	0	0	0	0	0	0	0	0	1	0.33	N/A
AND $-0x62d3a980,%ECX	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	N/A
XOR %RDX,%RCX	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	N/A
MOV %ECX,%EDX	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.13	scal (6.3%)
SAL $0xf,%EDX	1	0	0	0	0.33	0.33	0.33	0	0	0	0	0	0	0	0	0	0	1	0.33	scal (6.3%)
AND $-0x103a0000,%EDX	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	scal (6.3%)
XOR %RCX,%RDX	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	scal (12.5%)
MOV %RDX,%RCX	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.13	N/A
SHR $0x12,%RCX	1	0	0	0	0.33	0.33	0.33	0	0	0	0	0	0	0	0	0	0	1	0.33	N/A
XOR %RDX,%RCX	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	N/A
DEC %RAX	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	N/A
VCVTUSI2SS %RCX,%XMM15,%XMM2	2	0	0	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	3-9	1	scal (12.5%)
VFMADD231SS %XMM2,%XMM1,%XMM0	1	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	0	0	4	0.50	scal (6.3%)
VMULSS 0x3b09(%RIP),%XMM1,%XMM1	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0.50	0.50	0	0	0	0	3	0.50	scal (6.3%)
JE 5920 <main+0x1f70>	1	0	0	0	0.33	0.33	0.33	0	0	0	0	0	0	0	0	0	0	0	0.33-0.50	N/A
CMP $0x270,%R13	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	N/A
JB 54a0 <main+0x1af0>	1	0	0	0	0.33	0.33	0.33	0	0	0	0	0	0	0	0	0	0	0	0.33-0.50	N/A
VPBROADCASTQ 0x3b28(%RIP),%ZMM14	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0.50	0.50	0	0	0	2	0.50	scal (12.5%)
VPBROADCASTQ 0x3b26(%RIP),%ZMM15	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0.50	0.50	0	0	0	2	0.50	scal (12.5%)
VPBROADCASTQ 0x3b24(%RIP),%ZMM16	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0.50	0.50	0	0	0	2	0.50	scal (12.5%)
VPBROADCASTQ 0x3b22(%RIP),%ZMM17	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0.50	0.50	0	0	0	2	0.50	scal (12.5%)
VPBROADCASTQ %RBX,%ZMM2	2	0	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	0	2	1	scal (12.5%)
XOR %ECX,%ECX	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	N/A
XCHG %AX,%AX	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.13	N/A
MOV 0xd40(%RSP),%RDX	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	3	0.25	scal (12.5%)
VEXTRACTI32X4 $0x3,%ZMM2,%XMM2	1	0	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	0	5	0.50	vect (25.0%)
MOV 0xd48(%RSP),%RCX	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	3	0.25	N/A
VPBROADCASTQ 0x39bb(%RIP),%XMM8	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0.25	0.25	0.25	0.25	0	0	2	0.50	scal (12.5%)
VPBROADCASTQ 0x39ba(%RIP),%XMM9	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0.25	0.25	0.25	0.25	0	0	2	0.50	scal (12.5%)
VPBROADCASTQ 0x39b9(%RIP),%XMM10	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0.25	0.25	0.25	0.25	0	0	2	0.50	scal (12.5%)
VPBROADCASTQ 0x39b8(%RIP),%XMM11	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0.25	0.25	0.25	0.25	0	0	2	0.50	scal (12.5%)
VPEXTRQ $0x1,%XMM2,%RSI	2	0	0	0	0	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0.50	0.50	7	0.50	scal (12.5%)
AND $-0x80000000,%RSI	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	scal (12.5%)
MOV %EDX,%EDI	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.13	scal (6.3%)
AND $0x7ffffffe,%EDI	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	scal (6.3%)
OR %RSI,%RDI	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	scal (12.5%)
MOV %EDX,%ESI	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.13	scal (6.3%)
AND $0x1,%ESI	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	scal (6.3%)
AND $-0x80000000,%RDX	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	scal (12.5%)
SHR $0x1,%RDI	1	0	0	0	0.33	0.33	0.33	0	0	0	0	0	0	0	0	0	0	1	0.33	scal (12.5%)
XOR 0x19a0(%RSP),%RDI	1	0.17	0.17	0.17	0.17	0.17	0.17	0.25	0.25	0.25	0.25	0	0	0	0	0	0	1	0.25	scal (12.5%)
NEG %ESI	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	scal (6.3%)
AND %R12D,%ESI	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	scal (6.3%)
XOR %RDI,%RSI	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	scal (12.5%)
MOV %RSI,0xd38(%RSP)	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	4	0.50	scal (12.5%)
MOV %ECX,%ESI	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.13	scal (6.3%)
AND $0x7ffffffe,%ESI	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	scal (6.3%)
OR %RDX,%RSI	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	scal (12.5%)
MOV %ECX,%EDX	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.13	scal (6.3%)
AND $0x1,%EDX	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	scal (6.3%)
AND $-0x80000000,%RCX	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	N/A
SHR $0x1,%RSI	1	0	0	0	0.33	0.33	0.33	0	0	0	0	0	0	0	0	0	0	1	0.33	scal (12.5%)
XOR 0x19a8(%RSP),%RSI	1	0.17	0.17	0.17	0.17	0.17	0.17	0.25	0.25	0.25	0.25	0	0	0	0	0	0	1	0.25	scal (12.5%)
NEG %EDX	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	scal (6.3%)
AND %R12D,%EDX	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	scal (6.3%)
XOR %RSI,%RDX	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	scal (12.5%)
MOV %RDX,0xd40(%RSP)	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	4	0.50	scal (12.5%)
MOV 0xd50(%RSP),%RDX	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	3	0.25	scal (12.5%)
MOV %EDX,%ESI	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.13	scal (6.3%)
VPBROADCASTQ %RDX,%XMM2	2	0	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	0	2	1	scal (12.5%)
AND $0x7ffffffe,%EDX	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	scal (6.3%)
AND $0x1,%ESI	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	scal (6.3%)
OR %RCX,%RDX	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	scal (12.5%)
NEG %ESI	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	scal (6.3%)
MOV $0xee,%ECX	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	N/A
SHR $0x1,%RDX	1	0	0	0	0.33	0.33	0.33	0	0	0	0	0	0	0	0	0	0	1	0.33	scal (12.5%)
XOR 0x19b0(%RSP),%RDX	1	0.17	0.17	0.17	0.17	0.17	0.17	0.25	0.25	0.25	0.25	0	0	0	0	0	0	1	0.25	scal (12.5%)
AND %R12D,%ESI	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	scal (6.3%)
XOR %RDX,%RSI	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	scal (12.5%)
MOV %RSI,0xd48(%RSP)	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	4	0.50	scal (12.5%)
NOPL (%RAX)	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.13	N/A
MOV 0x19b0(%RSP),%RCX	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	3	0.25	N/A
MOV 0x638(%RSP),%RBX	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	3	0.25	scal (12.5%)
XOR %R13D,%R13D	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.17	N/A
MOV %EBX,%EDX	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.13	scal (6.3%)
AND %R14,%RCX	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	N/A
AND $0x7ffffffe,%EDX	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	scal (6.3%)
OR %RCX,%RDX	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	scal (12.5%)
MOV %EBX,%ECX	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.13	N/A
AND $0x1,%ECX	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	N/A
SHR $0x1,%RDX	1	0	0	0	0.33	0.33	0.33	0	0	0	0	0	0	0	0	0	0	1	0.33	scal (12.5%)
XOR 0x1298(%RSP),%RDX	1	0.17	0.17	0.17	0.17	0.17	0.17	0.25	0.25	0.25	0.25	0	0	0	0	0	0	1	0.25	scal (12.5%)
NEG %ECX	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	N/A
AND %R12D,%ECX	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	N/A
XOR %RDX,%RCX	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	N/A
MOV %RCX,0x19b0(%RSP)	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	4	0.50	scal (12.5%)
JMP 54a0 <main+0x1af0>	1	0	0	0	0.33	0.33	0.33	0	0	0	0	0	0	0	0	0	0	0	0.50	N/A

Function	main
Source file and lines	random.tcc:401-3367
Module	attention-clang-znver5-512

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 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 4.00 to 0.28 cycles (14.29x 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	25
nb uops	25
loop length	110
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	3.13 cycles
front end	3.13 cycles

Back-end

	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11	P12	P13	P14	P15
uops	2.50	2.50	2.50	2.50	2.50	2.50	0.75	0.75	0.75	0.75	1.00	1.00	0.50	0.50	0.00	0.00
cycles	2.50	2.50	2.50	2.50	2.50	2.50	0.75	0.75	0.75	0.75	1.00	1.00	0.50	0.50	0.00	0.00

Execution ports to units layout:

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

Cycles executing div or sqrt instructions	NA
Longest recurrence chain latency (RecMII)	4.00

Cycles summary

Front-end	3.13
Dispatch	2.50
Data deps.	4.00
Overall L1	4.00

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	9%
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	8%

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	8%

Cycles and memory resources usage

Assuming all data fit into the L1 cache, each iteration of the binary loop takes 4.00 cycles. At this rate:

1% of peak load performance is reached (3.00 out of 256.00 bytes loaded per cycle (GB/s @ 1GHz))
1% of peak store performance is reached (2.00 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: 5505

Instruction	Nb FU	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11	P12	P13	Latency	Recip. throughput	Vectorization
MOV %R13,%RCX	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.13	N/A
INC %R13	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	1	0.17	N/A
MOV $0x200b,%EDX	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	1	0.17	scal (6.3%)
MOV %R13,0x19b8(%RSP)	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	4	0.50	scal (12.5%)
MOV 0x638(%RSP,%RCX,8),%RCX	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	3	0.25	N/A
BEXTR %RDX,%RCX,%RDX	1	0	0	0	0.33	0.33	0.33	0	0	0	0	0	0	0	0	1	0.33	N/A
XOR %RCX,%RDX	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	1	0.17	scal (12.5%)
MOV %EDX,%ECX	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.13	N/A
SAL $0x7,%ECX	1	0	0	0	0.33	0.33	0.33	0	0	0	0	0	0	0	0	1	0.33	N/A
AND $-0x62d3a980,%ECX	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	1	0.17	N/A
XOR %RDX,%RCX	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	1	0.17	N/A
MOV %ECX,%EDX	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.13	scal (6.3%)
SAL $0xf,%EDX	1	0	0	0	0.33	0.33	0.33	0	0	0	0	0	0	0	0	1	0.33	scal (6.3%)
AND $-0x103a0000,%EDX	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	1	0.17	scal (6.3%)
XOR %RCX,%RDX	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	1	0.17	scal (12.5%)
MOV %RDX,%RCX	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.13	N/A
SHR $0x12,%RCX	1	0	0	0	0.33	0.33	0.33	0	0	0	0	0	0	0	0	1	0.33	N/A
XOR %RDX,%RCX	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	1	0.17	N/A
DEC %RAX	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	1	0.17	scal (12.5%)
VCVTUSI2SS %RCX,%XMM15,%XMM2	2	0	0	0	0	0	0	0	0	0	0	0	0	0.50	0.50	3-9	1	scal (12.5%)
VFMADD231SS %XMM2,%XMM1,%XMM0	1	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	4	0.50	scal (6.3%)
VMULSS 0x3b09(%RIP),%XMM1,%XMM1	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0.50	0.50	0	0	3	0.50	scal (6.3%)
JE 5920 <main+0x1f70>	1	0	0	0	0.33	0.33	0.33	0	0	0	0	0	0	0	0	0	0.33-0.50	N/A
CMP $0x270,%R13	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	1	0.17	N/A
JB 54a0 <main+0x1af0>	1	0	0	0	0.33	0.33	0.33	0	0	0	0	0	0	0	0	0	0.33-0.50	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

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