Apple Microarchitecture Research by Dougall Johnson

M1/A14 P-core (Firestorm): Overview | Base Instructions | SIMD and FP Instructions
M1/A14 E-core (Icestorm):  Overview | Base Instructions | SIMD and FP Instructions

SQDMLAL (scalar, S)

Test 1: uops

Code:

  sqdmlal s0, h1, h2
  movi v0.16b, 1
  movi v1.16b, 2
  movi v2.16b, 3

(no loop instructions)

1000 unrolls and 1 iteration

Retires: 1.000

Issues: 1.000

Integer unit issues: 0.001

Load/store unit issues: 0.000

SIMD/FP unit issues: 1.000

retire uop (01)cycle (02)schedule uop (52)schedule int uop (53)schedule simd uop (54)dispatch simd uop (57)ldst uops in schedulers (5b)dispatch uop (78)map simd uop (7e)map simd uop inputs (81)? int output thing (e9)? simd retires (ee)
1004303310011100010007590510001000300011000
1004303310011100010007590510001000300011000
1004303310011100010007590510001000300011000
1004303310011100010007590510001000300011000
1004303310011100010007590510001000300011000
1004303310011100010007590510001000300011000
1004303310011100010007590510001000300011000
1004303310011100010007590510001000300011000
1004303310011100010007590510001000300011000
1004303310011100010007590510001000300011000

Test 2: Latency 1->1

Code:

  sqdmlal s0, h1, h2
  movi v0.16b, 1
  movi v1.16b, 2
  movi v2.16b, 3

(fused SUBS/B.cc loop)

100 unrolls and 100 iterations

Result (median cycles for code): 3.0033

retire uop (01)cycle (02)schedule uop (52)schedule int uop (53)schedule simd uop (54)dispatch int uop (56)dispatch simd uop (57)int uops in schedulers (59)ldst uops in schedulers (5b)dispatch uop (78)map int uop (7c)map simd uop (7e)map int uop inputs (7f)map simd uop inputs (81)? int output thing (e9)? simd retires (ee)? int retires (ef)
1020430033101011011000010010000300768905101002001000420030012110000100
1020430033101011011000010010000300768905101002001000420030012110000100
1020430033101011011000010010000300768905101002001000420030012110000100
1020430033101011011000010010000300768905101002001000420030012110000100
1020430033101011011000010010000300768905101002001000420030012110000100
1020430033101011011000010010000300768905101002001000420030012110000100
1020430033101011011000010010000300768905101002001000420030012110000100
1020430033101011011000010010000300768905101002001000420030012110000100
1020430033101011011000010010000300768905101002001000420030012110000100
1020530066101091031000610210031300768905101002001000420030012110000100

1000 unrolls and 10 iterations

Result (median cycles for code): 3.0033

retire uop (01)cycle (02)schedule uop (52)schedule int uop (53)schedule simd uop (54)dispatch int uop (56)dispatch simd uop (57)int uops in schedulers (59)ldst uops in schedulers (5b)dispatch uop (78)map int uop (7c)map simd uop (7e)map int uop inputs (7f)map simd uop inputs (81)? int output thing (e9)? simd retires (ee)? int retires (ef)
10024300331002121100002010000707689051002020100042030000111000010
10024300331002121100002010000707689051002020100002030000111000010
10024300331002121100002010000707689051002020100002030000111000010
10024300331002121100002010000707689051002020100002030000111000010
10024300331002121100002010000707689051002020100002030000111000010
10024300331002121100002010000707689051002020100002030000111000010
10024300331002121100002010000707689051002020100002030000111000010
10024300331002121100002010000707689051002020100002030000111000010
10024300331002121100002010000707689051002020100002030000111000010
10024300331002121100002010000707689051002020100002030000111000010

Test 3: Latency 1->2

Code:

  sqdmlal s0, h0, h1
  movi v0.16b, 1
  movi v1.16b, 2

(fused SUBS/B.cc loop)

100 unrolls and 100 iterations

Result (median cycles for code): 3.0033

retire uop (01)cycle (02)schedule uop (52)schedule int uop (53)schedule simd uop (54)schedule ldst uop (55)dispatch int uop (56)dispatch simd uop (57)int uops in schedulers (59)ldst uops in schedulers (5b)dispatch uop (78)map int uop (7c)map simd uop (7e)map int uop inputs (7f)map simd uop inputs (81)? int output thing (e9)? simd retires (ee)? int retires (ef)
10204300331010110110000010010000300768905101002001000620030012110000100
10204300331010110110000010010000300769247101312001004420030012110000100
10204300331010110110000010010000300768905101002001000420030012110000100
10204300331010110110000010010000300768905101002001000420030012110000100
10204300331010110110000010010000300768905101002001000420030012110000100
10204300331010110110000010010000300768905101002001000420030012110000100
10204300331010110110000010010000300768905101002001000420030012110000100
10204300331010110110000010010000300768905101002001000420030012110000100
10204300331010110110000010010000300768905101002001000420030012110000100
10204300331010110110000010010000300768905101002001000420030012110000100

1000 unrolls and 10 iterations

Result (median cycles for code): 3.0033

retire uop (01)cycle (02)schedule uop (52)schedule int uop (53)schedule simd uop (54)dispatch int uop (56)dispatch simd uop (57)int uops in schedulers (59)ldst uops in schedulers (5b)dispatch uop (78)map int uop (7c)map simd uop (7e)map int uop inputs (7f)map simd uop inputs (81)? int output thing (e9)? simd retires (ee)? int retires (ef)
10024300331002121100002010000707689051002020100062030000111000010
10024300331002121100002010000707689051002020100002030000111000010
10024300331002121100002010000707689051002020100002030000111000010
10024300331002121100002010000707689051002020100002030000111000010
10024300331002121100002010000707689051002020100002030000111000010
10024300331002121100002010000707689051002020100002030000111000010
10024300331002121100002010000707689051002020100002030000111000010
10025300661002721100062010031707689051002020100002030000111000010
10024300331002121100002010000707689051002020100002030000111000010
10024300331002121100002010000707689051002020100002030000111000010

Test 4: Latency 1->3

Code:

  sqdmlal s0, h1, h0
  movi v0.16b, 1
  movi v1.16b, 2

(fused SUBS/B.cc loop)

100 unrolls and 100 iterations

Result (median cycles for code): 3.0033

retire uop (01)cycle (02)schedule uop (52)schedule int uop (53)schedule simd uop (54)dispatch int uop (56)dispatch simd uop (57)int uops in schedulers (59)ldst uops in schedulers (5b)dispatch uop (78)map int uop (7c)map simd uop (7e)map int uop inputs (7f)map simd uop inputs (81)? int output thing (e9)? simd retires (ee)? int retires (ef)
1020430033101011011000010010000300768905101002001000620030012110000100
1020430033101011011000010010000300768905101002001000420030012110000100
1020430033101011011000010010000300768905101002001000420030012110000100
1020430033101011011000010010000300768905101002001000420030012110000100
1020430033101011011000010010000300768905101002001000420030012110000100
1020430033101011011000010010000300768905101002001000420030012110000100
1020430033101011011000010010000300768905101002001000420030012110000100
1020430033101011011000010010000300768905101002001000420030012110000100
1020430033101011011000010010000300768905101002001000420030012110000100
1020430033101011011000010010000300768905101002001000420030012110000100

1000 unrolls and 10 iterations

Result (median cycles for code): 3.0033

retire uop (01)cycle (02)schedule uop (52)schedule int uop (53)schedule simd uop (54)dispatch int uop (56)dispatch simd uop (57)int uops in schedulers (59)ldst uops in schedulers (5b)dispatch uop (78)map int uop (7c)map simd uop (7e)map int uop inputs (7f)map ldst uop inputs (80)map simd uop inputs (81)? int output thing (e9)? ldst retires (ed)? simd retires (ee)? int retires (ef)
1002430033100212110000201000070768905100202010006200300121101000010
1002430033100212110000201000070768905100202010000200300001101000010
100243003310021211000020100007076890510020201000044642824185062168114360922293
1002430033100212110000201000070768905100202010000200300001101000010
1002430033100212110000201000068769247100512010045200300001101000010
1002430033100212110000201000069768905100202010004200300001101000010
1002430033100212110000201000070768905100202010000200300001101000010
1002430033100212110000201000070768905100202010000200300001101000010
1002430033100212110000201000070768905100202010000200300001101000010
1002430033100212110000201000070768905100202010000200300001101000010

Test 5: throughput

Count: 8

Code:

  movi v0.16b, 0
  sqdmlal s0, h8, h9
  movi v1.16b, 0
  sqdmlal s1, h8, h9
  movi v2.16b, 0
  sqdmlal s2, h8, h9
  movi v3.16b, 0
  sqdmlal s3, h8, h9
  movi v4.16b, 0
  sqdmlal s4, h8, h9
  movi v5.16b, 0
  sqdmlal s5, h8, h9
  movi v6.16b, 0
  sqdmlal s6, h8, h9
  movi v7.16b, 0
  sqdmlal s7, h8, h9
  movi v8.16b, 9
  movi v9.16b, 10

(fused SUBS/B.cc loop)

100 unrolls and 100 iterations

Result (median cycles for code divided by count): 0.5011

retire uop (01)cycle (02)schedule uop (52)schedule int uop (53)schedule simd uop (54)schedule ldst uop (55)dispatch int uop (56)dispatch simd uop (57)int uops in schedulers (59)ldst uops in schedulers (5b)dispatch uop (78)map int uop (7c)map simd uop (7e)map int uop inputs (7f)map simd uop inputs (81)? int output thing (e9)? simd retires (ee)? int retires (ef)
16020440493801091018000801008001230032005680113200800132002400391160000100
16020440121801101018000901008001330032005280112200800122002400361160000100
16020440086801091018000801008001230032005280112200800122002400361160000100
16020440086801091018000801008001230032005280112200800122002400361160000100
16020440086801091018000801008001230032005280112200800122002400361160000100
16020440086801091018000801008001230032005280112200800122002400361160000100
16020440086801091018000801008001230032005280112200800122002400361160000100
16020440086801091018000801008001230032005280112200800122002400361160000100
16020440086801091018000801008001230032005280112200800122002400361160000100
16020440086801091018000801008001230032005280112200800122002400361160000100

1000 unrolls and 10 iterations

Result (median cycles for code divided by count): 0.5053

retire uop (01)cycle (02)schedule uop (52)schedule int uop (53)schedule simd uop (54)dispatch int uop (56)dispatch simd uop (57)int uops in schedulers (59)ldst uops in schedulers (5b)dispatch uop (78)map int uop (7c)map simd uop (7e)map int uop inputs (7f)map simd uop inputs (81)? int output thing (e9)? simd retires (ee)? int retires (ef)
1600244390180020118000910800133032005680023208001320240000116000010
1600244136880011118000010800003032000080010208000020240000116000010
1600244042680011118000010800003032000080010208000020240000116000010
1600244041880011118000010800003032000080010208000020240000116000010
1600244042880011118000010800003032000080010208000020240000116000010
1600244036180011118000010800003032000080010208000020240000116000010
1600244042880011118000010800003032000080010208000020240000116000010
1600244038980011118000010800003032000080010208000020240000116000010
1600244043680011118000010800003032000080010208000020240000116000010
1600244042880011118000010800003032000080010208000020240000116000010

Test 6: throughput

Count: 16

Code:

  sqdmlal s0, h16, h17
  sqdmlal s1, h16, h17
  sqdmlal s2, h16, h17
  sqdmlal s3, h16, h17
  sqdmlal s4, h16, h17
  sqdmlal s5, h16, h17
  sqdmlal s6, h16, h17
  sqdmlal s7, h16, h17
  sqdmlal s8, h16, h17
  sqdmlal s9, h16, h17
  sqdmlal s10, h16, h17
  sqdmlal s11, h16, h17
  sqdmlal s12, h16, h17
  sqdmlal s13, h16, h17
  sqdmlal s14, h16, h17
  sqdmlal s15, h16, h17
  movi v16.16b, 17
  movi v17.16b, 18

(fused SUBS/B.cc loop)

100 unrolls and 100 iterations

Result (median cycles for code divided by count): 0.5002

retire uop (01)cycle (02)schedule uop (52)schedule int uop (53)schedule simd uop (54)schedule ldst uop (55)dispatch int uop (56)dispatch simd uop (57)int uops in schedulers (59)ldst uops in schedulers (5b)dispatch uop (78)map int uop (7c)map simd uop (7e)map int uop inputs (7f)map ldst uop inputs (80)map simd uop inputs (81)? int output thing (e9)? ldst retires (ed)? simd retires (ee)? int retires (ef)
160204801011601051011600040100160008300640044160110200160014200048003910160000100
160204800351601071011600060100160010300640044160110200160014200048003610160000100
160204800351601071011600060100160010300640036160108200160012200048003910160000100
160204800351601071011600060100160010300640036160108200160012200048062710160000100
160204800351601051011600040100160008300640240160159200160063200048019210160000100
160204801111601541011600530100160057300640036160108200160012200048003610160000100
160204800351601051011600040100160008300640036160108200160012200048003610160000100
160204800351601051011600040100160008300640036160108200160012200048003610160000100
160204801931601561011600550100160059300640844160310200160214200048078310160000100
160204809561605541031604510102160455300640036160108200160012200048003610160000100

1000 unrolls and 10 iterations

Result (median cycles for code divided by count): 0.5002

retire uop (01)cycle (02)schedule uop (52)schedule int uop (53)schedule simd uop (54)dispatch int uop (56)dispatch simd uop (57)int uops in schedulers (59)ldst uops in schedulers (5b)dispatch uop (78)map int uop (7c)map simd uop (7e)map int uop inputs (7f)map simd uop inputs (81)? int output thing (e9)? simd retires (ee)? int retires (ef)
160024801571600171116000610160010306400441600202016001420480042116000010
160024800361600111116000010160000306400001600102016000020480147116000010
160024801101600601116004910160049306400001600102016000020480000116000010
160024801281600111116000010160000306400521600222016001520480000116000010
160024805321602581116024710160247306409921602582016024820480000116000010
160024800371600111116000010160000306401961600592016004920480000116000010
160024800351600111116000010160000306400001600102016000020480000116000010
160024800351600111116000010160000306400001600102016000020480000116000010
160024800351600111116000010160000306402081600642016006220480000116000010
160024800351600111116000010160000306400001600102016000020480000116000010