OV - attention-avx512 - Loop 70

0x3d50 MOV	0x40(%RSP),%EAX

0x3d54 INC	%EAX

0x3d56 ADD	%R12,%R10

0x3d59 CMP	0x18(%RSP),%EAX

0x3d5d JE	3d20

0x3d5f MOV	%EAX,0x40(%RSP)

0x3d63 IMUL	%R12D,%EAX

0x3d67 MOV	%RAX,0x30(%RSP)

0x3d6c XOR	%EAX,%EAX

0x3d6e JMP	3d98

  (71) 0x3d70 VCVTSD2SS	%XMM0,%XMM0,%XMM0

  (71) 0x3d74 MOV	0x30(%RSP),%RCX

  (71) 0x3d79 ADD	%EAX,%ECX

  (71) 0x3d7b MOV	0xc8(%RSP),%R9

  (71) 0x3d83 VMOVSS	%XMM0,(%R9,%RCX,4)

  (71) 0x3d89 INC	%RAX

  (71) 0x3d8c CMP	0x28(%RSP),%RAX

  (71) 0x3d91 MOV	0x50(%RSP),%R15

  (71) 0x3d96 JE	3d50

  (71) 0x3d98 CMPL	$0x3,0x80(%RSP)

  (71) 0x3da0 JAE	3dc0

  (71) 0x3da2 VXORPS	%XMM0,%XMM0,%XMM0

  (71) 0x3da6 XOR	%R13D,%R13D

  (71) 0x3da9 MOV	0x78(%RSP),%R9

  (71) 0x3dae JMP	3e72

  (71) 0x3dc0 VXORPS	%XMM0,%XMM0,%XMM0

  (71) 0x3dc4 MOV	%RAX,%R11

  (71) 0x3dc7 XOR	%R13D,%R13D

  (71) 0x3dca MOV	0x78(%RSP),%R9

  (71) 0x3dcf NOP

    (68) 0x3dd0 LEA	(%R10,%R13,1),%R15D

    (68) 0x3dd4 VMOVSS	(%R14,%R15,4),%XMM1

    (68) 0x3dda VCVTSS2SD	%XMM1,%XMM1,%XMM1

    (68) 0x3dde MOV	%R11D,%R15D

    (68) 0x3de1 VMOVSS	(%R9,%R15,4),%XMM2

    (68) 0x3de7 VCVTSS2SD	%XMM2,%XMM2,%XMM2

    (68) 0x3deb LEA	0x1(%R10,%R13,1),%R15D

    (68) 0x3df0 VMOVSS	(%R14,%R15,4),%XMM3

    (68) 0x3df6 VCVTSS2SD	%XMM3,%XMM3,%XMM3

    (68) 0x3dfa LEA	(%R12,%R11,1),%R15D

    (68) 0x3dfe VMOVSS	(%R9,%R15,4),%XMM4

    (68) 0x3e04 VCVTSS2SD	%XMM4,%XMM4,%XMM4

    (68) 0x3e08 LEA	0x2(%R10,%R13,1),%R15D

    (68) 0x3e0d VMOVSS	(%R14,%R15,4),%XMM5

    (68) 0x3e13 VFMADD213SD	%XMM0,%XMM1,%XMM2

    (68) 0x3e18 VCVTSS2SD	%XMM5,%XMM5,%XMM0

    (68) 0x3e1c LEA	(%R8,%R11,1),%R15D

    (68) 0x3e20 VMOVSS	(%R9,%R15,4),%XMM1

    (68) 0x3e26 VFMADD213SD	%XMM2,%XMM3,%XMM4

    (68) 0x3e2b VCVTSS2SD	%XMM1,%XMM1,%XMM1

    (68) 0x3e2f LEA	0x3(%R10,%R13,1),%R15D

    (68) 0x3e34 VMOVSS	(%R14,%R15,4),%XMM2

    (68) 0x3e3a VFMADD213SD	%XMM4,%XMM0,%XMM1

    (68) 0x3e3f VCVTSS2SD	%XMM2,%XMM2,%XMM2

    (68) 0x3e43 LEA	(%RSI,%R11,1),%R15D

    (68) 0x3e47 VMOVSS	(%R9,%R15,4),%XMM0

    (68) 0x3e4d VCVTSS2SD	%XMM0,%XMM0,%XMM0

    (68) 0x3e51 VFMADD213SD	%XMM1,%XMM2,%XMM0

    (68) 0x3e56 ADD	$0x4,%R13

    (68) 0x3e5a ADD	%RDI,%R11

    (68) 0x3e5d CMP	%R13,%RDX

    (68) 0x3e60 JNE	3dd0

  (71) 0x3e66 CMPQ	$0,0x10(%RSP)

  (71) 0x3e6c JE	3d70

  (71) 0x3e72 MOV	%R12,%R11

  (71) 0x3e75 IMUL	%R13,%R11

  (71) 0x3e79 ADD	%RAX,%R11

  (71) 0x3e7c ADD	%R10D,%R13D

  (71) 0x3e7f MOV	0x10(%RSP),%R15

  (71) 0x3e84 NOPW	%CS:(%RAX,%RAX,1)

    (72) 0x3e90 MOV	%R13D,%ECX

    (72) 0x3e93 VMOVSS	(%R14,%RCX,4),%XMM1

    (72) 0x3e99 VCVTSS2SD	%XMM1,%XMM1,%XMM1

    (72) 0x3e9d MOV	%R11D,%ECX

    (72) 0x3ea0 VMOVSS	(%R9,%RCX,4),%XMM2

    (72) 0x3ea6 VCVTSS2SD	%XMM2,%XMM2,%XMM2

    (72) 0x3eaa VFMADD231SD	%XMM2,%XMM1,%XMM0

    (72) 0x3eaf ADD	%R12,%R11

    (72) 0x3eb2 INC	%R13D

    (72) 0x3eb5 DEC	%R15

    (72) 0x3eb8 JNE	3e90

  (71) 0x3eba JMP	3d70

/home/eoseret/Applications/llm-attention/attention.cpp: 26 - 193

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

26:    for (unsigned int i = 0; i < M; ++i) {

27:       for (unsigned int j = 0; j < N; ++j) {  //vectorized

28:          double sum = 0.0;

29: #pragma clang loop vectorize(enable)

30:          for (unsigned int k = 0; k < K; ++k) { //vectorized

31:             sum += (double)A[i * K + k] * (double)B[k * N + j];

32:          }

33:          C[i * N + j] = alpha * static_cast<float>(sum);

[...]

193:    for (size_t r = 0; r < rept; r++)

Coverage (%)	Name	Module
►100.00+	__libc_init_first	libc.so.6
○	__libc_start_main	libc.so.6
○	_start	attention-avx512

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	1.00
CQA speedup if FP arith vectorized	1.00
CQA speedup if fully vectorized	13.33
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.25
Bottlenecks	micro-operation queue,
Function	main
Source	attention.cpp:26-26,attention.cpp:193-193
Source loop unroll info	NA
Source loop unroll confidence level	NA
Unroll/vectorization loop type	NA
Unroll factor	NA
CQA cycles	2.50
CQA cycles if no scalar integer	2.50
CQA cycles if FP arith vectorized	2.50
CQA cycles if fully vectorized	0.19
Front-end cycles	2.50
P0 cycles	1.50
P1 cycles	1.75
P2 cycles	1.33
P3 cycles	1.33
P4 cycles	2.00
P5 cycles	1.50
P6 cycles	1.50
P7 cycles	1.33
DIV/SQRT cycles	0.00
Inter-iter dependencies cycles	0
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	10.00
Nb uops	10.00
Nb loads	2.00
Nb stores	2.00
Nb stack references	3.00
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	8.00
Bytes prefetched	0.00
Bytes loaded	8.00
Bytes stored	12.00
Stride 0	0.00
Stride 1	0.00
Stride n	0.00
Stride unknown	2.00
Stride indirect	0.00
Vectorization ratio all	0.00
Vectorization ratio load	NA
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	NA
Vector-efficiency ratio all	10.42
Vector-efficiency ratio load	NA
Vector-efficiency ratio store	9.38
Vector-efficiency ratio mul	NA
Vector-efficiency ratio add_sub	12.50
Vector-efficiency ratio fma	NA
Vector-efficiency ratio div_sqrt	NA
Vector-efficiency ratio other	NA

Metric	Value
CQA speedup if no scalar integer	1.00
CQA speedup if FP arith vectorized	1.00
CQA speedup if fully vectorized	13.33
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.25
Bottlenecks	micro-operation queue,
Function	main
Source	attention.cpp:26-26,attention.cpp:193-193
Source loop unroll info	NA
Source loop unroll confidence level	NA
Unroll/vectorization loop type	NA
Unroll factor	NA
CQA cycles	2.50
CQA cycles if no scalar integer	2.50
CQA cycles if FP arith vectorized	2.50
CQA cycles if fully vectorized	0.19
Front-end cycles	2.50
P0 cycles	1.50
P1 cycles	1.75
P2 cycles	1.33
P3 cycles	1.33
P4 cycles	2.00
P5 cycles	1.50
P6 cycles	1.50
P7 cycles	1.33
DIV/SQRT cycles	0.00
Inter-iter dependencies cycles	0
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	10.00
Nb uops	10.00
Nb loads	2.00
Nb stores	2.00
Nb stack references	3.00
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	8.00
Bytes prefetched	0.00
Bytes loaded	8.00
Bytes stored	12.00
Stride 0	0.00
Stride 1	0.00
Stride n	0.00
Stride unknown	2.00
Stride indirect	0.00
Vectorization ratio all	0.00
Vectorization ratio load	NA
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	NA
Vector-efficiency ratio all	10.42
Vector-efficiency ratio load	NA
Vector-efficiency ratio store	9.38
Vector-efficiency ratio mul	NA
Vector-efficiency ratio add_sub	12.50
Vector-efficiency ratio fma	NA
Vector-efficiency ratio div_sqrt	NA
Vector-efficiency ratio other	NA

Path /

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

Function	main
Source file and lines	attention.cpp:26-193
Module	attention-avx512

The loop is defined in /home/eoseret/Applications/llm-attention/attention.cpp:26-33,193.

Analyzed code is defined in /home/eoseret/Applications/llm-attention/attention.cpp:26,193.

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

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

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

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

gain
potential
hint
expert

Vectorization

Your loop is 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 2.50 to 0.19 cycles (13.33x 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
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

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)

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 8 bytes. The binary loop is storing 12 bytes.

General properties

nb instructions	10
nb uops	10
loop length	32
used x86 registers	4
used mmx registers	0
used xmm registers	0
used ymm registers	0
used zmm registers	0
nb stack references	3

Front-end

MACRO FUSION NOT POSSIBLE FIT IN UOP CACHE

micro-operation queue	2.50 cycles
front end	2.50 cycles

Back-end

	P0	P1	P2	P3	P4	P5	P6	P7
uops	1.50	1.50	1.33	1.33	2.00	1.50	1.50	1.33
cycles	1.50	1.75	1.33	1.33	2.00	1.50	1.50	1.33

Execution ports to units layout:

P0 (256 bits): VPU, BRU, ALU, DIV/SQRT
P1 (256 bits): ALU, VPU
P2 (512 bits): store address, load
P3 (512 bits): store address, load
P4 (512 bits): store data
P5 (512 bits): ALU, VPU
P6: ALU, BRU
P7: store address

Cycles executing div or sqrt instructions	NA
Longest recurrence chain latency (RecMII)	0.00

Cycles summary

Front-end	2.50
Dispatch	2.00
Data deps.	0.00
Overall L1	2.50

Vectorization ratios

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)
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	NA (no other vectorizable/vectorized instructions)

Vector efficiency ratios

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

Cycles and memory resources usage

Assuming all data fit into the L1 cache, each iteration of the binary loop takes 2.50 cycles. At this rate:

2% of peak load performance is reached (3.20 out of 128.00 bytes loaded per cycle (GB/s @ 1GHz))
7% of peak store performance is reached (4.80 out of 64.00 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 2.50 to 2.00 cycles (1.25x speedup).

ASM code

In the binary file, the address of the loop is: 3d5f

Instruction	Nb FU	P0	P1	P2	P3	P4	P5	P6	P7	Latency	Recip. throughput	Vectorization
MOV 0x40(%RSP),%EAX	1	0	0	0.50	0.50	0	0	0	0	4-5	0.50	N/A
INC %EAX	1	0.25	0.25	0	0	0	0.25	0.25	0	1	0.25	N/A
ADD %R12,%R10	1	0.25	0.25	0	0	0	0.25	0.25	0	1	0.25	scal (12.5%)
CMP 0x18(%RSP),%EAX	1	0.25	0.25	0.50	0.50	0	0.25	0.25	0	1	0.50	N/A
JE 3d20 <main+0x1730>	1	0.50	0	0	0	0	0	0.50	0	0	0.50-1	N/A
MOV %EAX,0x40(%RSP)	1	0	0	0.33	0.33	1	0	0	0.33	3	1	scal (6.3%)
IMUL %R12D,%EAX	1	0	1	0	0	0	0	0	0	3	1	N/A
MOV %RAX,0x30(%RSP)	1	0	0	0.33	0.33	1	0	0	0.33	3	1	scal (12.5%)
XOR %EAX,%EAX	1	0	0	0	0	0	0	0	0	0	0.25	N/A
JMP 3d98 <main+0x17a8>	1	0	0	0	0	0	0	1	0	0	1-2	N/A

Function	main
Source file and lines	attention.cpp:26-193
Module	attention-avx512

The loop is defined in /home/eoseret/Applications/llm-attention/attention.cpp:26-33,193.

Analyzed code is defined in /home/eoseret/Applications/llm-attention/attention.cpp:26,193.

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

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

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

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

gain
potential
hint
expert

Vectorization

Your loop is 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 2.50 to 0.19 cycles (13.33x 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
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

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)

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 8 bytes. The binary loop is storing 12 bytes.

General properties

nb instructions	10
nb uops	10
loop length	32
used x86 registers	4
used mmx registers	0
used xmm registers	0
used ymm registers	0
used zmm registers	0
nb stack references	3

Front-end

MACRO FUSION NOT POSSIBLE FIT IN UOP CACHE

micro-operation queue	2.50 cycles
front end	2.50 cycles

Back-end

	P0	P1	P2	P3	P4	P5	P6	P7
uops	1.50	1.50	1.33	1.33	2.00	1.50	1.50	1.33
cycles	1.50	1.75	1.33	1.33	2.00	1.50	1.50	1.33

Execution ports to units layout:

P0 (256 bits): VPU, BRU, ALU, DIV/SQRT
P1 (256 bits): ALU, VPU
P2 (512 bits): store address, load
P3 (512 bits): store address, load
P4 (512 bits): store data
P5 (512 bits): ALU, VPU
P6: ALU, BRU
P7: store address

Cycles executing div or sqrt instructions	NA
Longest recurrence chain latency (RecMII)	0.00

Cycles summary

Front-end	2.50
Dispatch	2.00
Data deps.	0.00
Overall L1	2.50

Vectorization ratios

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)
div/sqrt	NA (no div/sqrt vectorizable/vectorized instructions)
other	NA (no other vectorizable/vectorized instructions)

Vector efficiency ratios

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

Cycles and memory resources usage

Assuming all data fit into the L1 cache, each iteration of the binary loop takes 2.50 cycles. At this rate:

2% of peak load performance is reached (3.20 out of 128.00 bytes loaded per cycle (GB/s @ 1GHz))
7% of peak store performance is reached (4.80 out of 64.00 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 2.50 to 2.00 cycles (1.25x speedup).

ASM code

In the binary file, the address of the loop is: 3d5f

Instruction	Nb FU	P0	P1	P2	P3	P4	P5	P6	P7	Latency	Recip. throughput	Vectorization
MOV 0x40(%RSP),%EAX	1	0	0	0.50	0.50	0	0	0	0	4-5	0.50	N/A
INC %EAX	1	0.25	0.25	0	0	0	0.25	0.25	0	1	0.25	N/A
ADD %R12,%R10	1	0.25	0.25	0	0	0	0.25	0.25	0	1	0.25	scal (12.5%)
CMP 0x18(%RSP),%EAX	1	0.25	0.25	0.50	0.50	0	0.25	0.25	0	1	0.50	N/A
JE 3d20 <main+0x1730>	1	0.50	0	0	0	0	0	0.50	0	0	0.50-1	N/A
MOV %EAX,0x40(%RSP)	1	0	0	0.33	0.33	1	0	0	0.33	3	1	scal (6.3%)
IMUL %R12D,%EAX	1	0	1	0	0	0	0	0	0	3	1	N/A
MOV %RAX,0x30(%RSP)	1	0	0	0.33	0.33	1	0	0	0.33	3	1	scal (12.5%)
XOR %EAX,%EAX	1	0	0	0	0	0	0	0	0	0	0.25	N/A
JMP 3d98 <main+0x17a8>	1	0	0	0	0	0	0	1	0	0	1-2	N/A

Report Configuration

Vectorization

Details

Workaround

Execution units bottlenecks

Slow data structures access

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

Slow data structures access

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