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 (by element, 4S)

Test 1: uops

Code:

  sqdmlal v0.4s, v1.4h, v2.h[1]
  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 v0.4s, v1.4h, v2.h[1]
  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
1020530066101071011000610210031300768905101002001000420030012110000100
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 simd uop inputs (81)? int output thing (e9)? simd retires (ee)? int retires (ef)
10024300331002121100002010000707689051002020100002030000111000010
10024300331002121100002010000707689051002020100002030000111000010
10024300331002121100002010000707689051002020100002030000111000010
10024300331002121100002010000707689051002020100002030000111000010
10024300331002121100002010000707689051002020100002030000111000010
10024300331002121100002010000707689051002020100002030000111000010
10024300331002121100002010000707689051002020100002030000111000010
10024300331002121100002010000707689051002020100002030000111000010
10024300331002121100002010000707689051002020100002030000111000010
10024300331002121100002010000707689051002020100002030000111000010

Test 3: Latency 1->2

Code:

  sqdmlal v0.4s, v0.4h, v1.h[1]
  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)
1020430033101011011000010010000300768905101002001000620230144210000100
1020430033101011011000010010000300768905101002001000420030018110000100
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 simd uop inputs (81)? int output thing (e9)? simd retires (ee)? int retires (ef)
10024300331002121100002010000707689051002020100062030000111000010
10024300331002121100002010000707689051002020100002030000111000010
10024300331002121100002010000707689051002020100002030000111000010
10024300331002121100002010000707689051002020100002030000111000010
10024300331002121100002010000707689051002020100002030000111000010
10024300331002121100002010000707689051002020100002030000111000010
10024300331002121100002010000707689051002020100002030000111000010
10024300331002121100002010000707689051002020100002030000111000010
10024300331002121100002010000707689051002020100002030000111000010
10024300331002121100002010000707689051002020100002030000111000010

Test 4: Latency 1->3

Code:

  sqdmlal v0.4s, v1.4h, v0.h[1]
  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
1020430033101011011000010010000300768905101002001000420030018110000100
1020430033101011011000010010000300768905101002001000420030135110000100
1020430033101011011000010010000300768905101002001000420030012110000100
1020430033101011011000010010000300768905101002001000420030012110000100
1020430033101011011000010010000300768905101002001000420030012110000100
1020430033101011011000010010000300768905101002001000420430489310000100
1020430241101391071003210610132307769997101682021008420030012110000100
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 simd uop inputs (81)? int output thing (e9)? simd retires (ee)? int retires (ef)
10024300331002121100002010000707689051002020100002030000111000010
10024300331002121100002010000707689051002020100002030000111000010
10024300331002121100002010000707689051002020100002030000111000010
10024300331002121100002010000707689051002020100002030000111000010
10024300331002121100002010000707689051002020100002030000111000010
10024300331002121100002010000707689051002020100002030000111000010
10024300331002121100002010000707689051002020100002030000111000010
10024300331002121100002010000707689051002020100002030000111000010
10024300331002121100002010000707689051002020100002030000111000010
10024300331002121100002010000707689051002020100002030000111000010

Test 5: throughput

Count: 8

Code:

  movi v0.16b, 0
  sqdmlal v0.4s, v8.4h, v9.h[1]
  movi v1.16b, 0
  sqdmlal v1.4s, v8.4h, v9.h[1]
  movi v2.16b, 0
  sqdmlal v2.4s, v8.4h, v9.h[1]
  movi v3.16b, 0
  sqdmlal v3.4s, v8.4h, v9.h[1]
  movi v4.16b, 0
  sqdmlal v4.4s, v8.4h, v9.h[1]
  movi v5.16b, 0
  sqdmlal v5.4s, v8.4h, v9.h[1]
  movi v6.16b, 0
  sqdmlal v6.4s, v8.4h, v9.h[1]
  movi v7.16b, 0
  sqdmlal v7.4s, v8.4h, v9.h[1]
  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)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)
1602044055380109101800081008001230032005680113200800132002400391160000100
1602044012180110101800091008001330032005280112200800122002400361160000100
1602044008680109101800081008001230032005280112200800122002400361160000100
1602044008680109101800081008001230032005280112200800122002400361160000100
1602044008680109101800081008001230032005280112200800122002400361160000100
1602044008680109101800081008001230032005280112200800122002400361160000100
1602044008680109101800081008001230032005280112200800122002400361160000100
1602044008680109101800081008001230032005280112200800122002401501160000100
1602044008680109101800081008001230032005280112200800122002400361160000100
1602044008680109101800081008001230032005280112200800122002400361160000100

1000 unrolls and 10 iterations

Result (median cycles for code divided by count): 0.5052

retire uop (01)cycle (02)schedule uop (52)schedule int uop (53)schedule simd uop (54)dispatch int uop (56)dispatch simd uop (57)dispatch ldst uop (58)int uops in schedulers (59)simd uops in schedulers (5a)ldst uops in schedulers (5b)dispatch uop (78)map int uop (7c)map ldst uop (7d)map simd uop (7e)map int uop inputs (7f)map simd uop inputs (81)? int output thing (e9)? simd retires (ee)? int retires (ef)
1600244384280019118000810800120300320000800102008000020240000116000010
1600244115080011118000010800000300320000800102008000020240000116000010
1600244042280011118000010800000300320000800102008000020240000116000010
1600244041880011118000010800000300320000800102008000020240000116000010
1600244041380011118000010800000300320000800102008000020240000116000010
1600244038880011118000010800000300320000800102008000020240000116000010
1600244040080011118000010800000300320000800102008000020240000116000010
1600244037080011118000010800000300320000800102008000020240000116000010
1600244041880011118000010800000300320000800102008000020240000116000010
1600254048580054118004310800470300320000800102008000020240000116000010

Test 6: throughput

Count: 12

Code:

  sqdmlal v0.4s, v12.4h, v13.h[1]
  sqdmlal v1.4s, v12.4h, v13.h[1]
  sqdmlal v2.4s, v12.4h, v13.h[1]
  sqdmlal v3.4s, v12.4h, v13.h[1]
  sqdmlal v4.4s, v12.4h, v13.h[1]
  sqdmlal v5.4s, v12.4h, v13.h[1]
  sqdmlal v6.4s, v12.4h, v13.h[1]
  sqdmlal v7.4s, v12.4h, v13.h[1]
  sqdmlal v8.4s, v12.4h, v13.h[1]
  sqdmlal v9.4s, v12.4h, v13.h[1]
  sqdmlal v10.4s, v12.4h, v13.h[1]
  sqdmlal v11.4s, v12.4h, v13.h[1]
  movi v12.16b, 13
  movi v13.16b, 14

(fused SUBS/B.cc loop)

100 unrolls and 100 iterations

Result (median cycles for code divided by count): 0.5003

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)
120204601121201051011200041001200083004800441201102001200142003600421120000100
120204600351201071011200061001200103004800361201082001200122003600361120000100
120204600351201051011200041001200083004800361201082001200122003600361120000100
120204600351201051011200041001200083004800361201082001200122003601981120000100
120204600451201071011200061001200103004800361201082001200122003600361120000100
120204600351201051011200041001200083004800361201082001200122003600361120000100
120204600351201051011200041001200083004800361201082001200122003600361120000100
120204600351201051011200041001200083004800361201082001200122003600361120000100
120204600351201051011200041001200083004800361201082001200122003600361120000100
120204600351201051011200041001200083004800361201082001200122003600361120000100

1000 unrolls and 10 iterations

Result (median cycles for code divided by count): 0.5003

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)
120024602301200151112000410120008304800001200102012000020360000112000010
120024600361200111112000010120000304802081200642012006020360000112000010
120024600351200111112000010120000304800001200102012000020360000112000010
120024600351200111112000010120000304800001200102012000020360000112000010
120024600351200111112000010120000304800001200102012000020360000112000010
120024600351200111112000010120000304800001200102012000020360000112000010
120024600351200111112000010120000304800001200102012000020360000112000010
120024600351200111112000010120000304800001200102012000020360000112000010
120024600351200111112000010120000304800001200102012000020360000112000010
120024600361200111112000010120000304800001200102012000020360000112000010