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

SQRDMLAH (by element, H)

Test 1: uops

Code:

  sqrdmlah h0, h1, 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)
1005306610071100610317590510001000300011000
1004303310011100010007590510001000300011000
1004303310011100010007590510001000300011000
1004303310011100010007590510001000300011000
1004303310011100010007590510001000300011000
1004303310011100010007590510001000300011000
1004303310011100010007590510001000300011000
1004303310011100010007590510001000300011000
1004303310011100010007590510001000300011000
1004303310011100010007590510001000300011000

Test 2: Latency 1->1

Code:

  sqrdmlah h0, h1, 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
1020430033101011011000010010000300768905101002001000420030012110000100
1020430033101011011000010010000300768905101002001000420030012110000100
1020430033101011011000010010000300768905101002001000420030012110000100
1020430033101011011000010010000300768905101002001000420030012110000100
1020430033101011011000010010000300769247101312001004420030012110000100

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)
10024300331002121100002010000707689051002020100042030012111000010
10024300331002121100002010000707689051002020100002030144111000010
10024300331002121100002010000707689051002020100002030000111000010
10024300331002121100002010000707689051002020100002030000111000010
10025300661002721100062010031707689051002020100042030018111000010
10024300331002121100002010000707689051002020100002030000111000010
10024300331002121100002010000707689051002020100002030000111000010
10024300331002121100002010000707689051002020100002030000111000010
10024300331002121100002010000707689051002020100002030000111000010
10024300331002121100002010000707689051002020100002030000111000010

Test 3: Latency 1->2

Code:

  sqrdmlah h0, h0, 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)
1020430033101011011000010010000300768905101002001000620030018110000100
1020430033101011011000010010000300768905101002001000420030012110000100
1020430033101011011000010010000300768905101002001000420030012110000100
1020430033101011011000010010000300768905101002001000420030012110000100
1020430033101011011000010010000300768905101002001000420030012110000100
1020430033101011011000010010000300768905101002001000420030012110000100
1020430033101011011000010010000300768905101002001000420030012110000100
1020430033101011011000010010000300768905101002001000420030012110000100
1020430033101011011000010010000300768905101002001000420030018110000100
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
10024300331002121100002010000697689051002020100042030018111000010

Test 4: Latency 1->3

Code:

  sqrdmlah h0, h1, 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)
1020430033101011011000010010000300768905101002001000620030135110000100
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 simd uop inputs (81)? int output thing (e9)? simd retires (ee)? int retires (ef)
10024300331002121100002010000707689051002020100002030000111000010
10024300331002121100002010000707689051002020100002030000111000010
10024302351005321100322010132707689051002020100002030000111000010
10024300331002121100002010000707689051002020100002030000111000010
10024300331002121100002010000707689051002020100002030000111000010
10024300331002121100002010000707689051002020100002030000111000010
10024300331002121100002010000707689051002020100002030000111000010
10024300331002121100002010000707689051002020100002030000111000010
10024302351005321100322010132707689051002020100002030000111000010
10024300331002121100002010000707689051002020100002030000111000010

Test 5: throughput

Count: 8

Code:

  movi v0.16b, 0
  sqrdmlah h0, h8, v9.h[1]
  movi v1.16b, 0
  sqrdmlah h1, h8, v9.h[1]
  movi v2.16b, 0
  sqrdmlah h2, h8, v9.h[1]
  movi v3.16b, 0
  sqrdmlah h3, h8, v9.h[1]
  movi v4.16b, 0
  sqrdmlah h4, h8, v9.h[1]
  movi v5.16b, 0
  sqrdmlah h5, h8, v9.h[1]
  movi v6.16b, 0
  sqrdmlah h6, h8, v9.h[1]
  movi v7.16b, 0
  sqrdmlah h7, h8, 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)
1602044052080109101800081008001230032005680113200800132002400391160000100
1602044011980110101800091008001330032005280112200800122002400361160000100
1602044008680109101800081008001230032005280112200800122002400361160000100
1602044008680109101800081008001230032005280112200800122002400361160000100
1602044008680109101800081008001230032005280112200800122002400361160000100
1602054013180144101800431008004730032005280112200800122002400361160000100
1602044008680109101800081008001230032005280112200800122002400361160000100
1602044008680109101800081008001230032005280112200800122002400361160000100
1602044008680109101800081008001230032005280112200800122002400361160000100
1602044008680109101800081008001230032005280112200800122002400361160000100

1000 unrolls and 10 iterations

Result (median cycles for code divided by count): 0.5054

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)
1600244383480019118000810800123032005680023208001320240000116000010
1600244123080011118000010800003032000080010208000020240000116000010
1600244039180011118000010800003032000080010208000020240147116000010
1600244050280011118000010800003032000080010208000020240000116000010
1600244036480011118000010800003032000080010208000020240000116000010
1600244042480011118000010800003032000080010208000020240000116000010
1600244042580011118000010800003032000080010208000020240036116000010
1600244113780011118000010800003032000080010208000020240000116000010
1600244041080011118000010800003032000080010208000020240000116000010
1600244042780011118000010800003032000080010208000020240000116000010

Test 6: throughput

Count: 12

Code:

  sqrdmlah h0, h12, v13.h[1]
  sqrdmlah h1, h12, v13.h[1]
  sqrdmlah h2, h12, v13.h[1]
  sqrdmlah h3, h12, v13.h[1]
  sqrdmlah h4, h12, v13.h[1]
  sqrdmlah h5, h12, v13.h[1]
  sqrdmlah h6, h12, v13.h[1]
  sqrdmlah h7, h12, v13.h[1]
  sqrdmlah h8, h12, v13.h[1]
  sqrdmlah h9, h12, v13.h[1]
  sqrdmlah h10, h12, v13.h[1]
  sqrdmlah h11, h12, 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)
120204600561201071011200061001200103004800441201102001200142003600421120000100
120204600451201071011200061001200103004800361201082001200122003600361120000100
120204600351201051011200041001200083004800361201082001200122003600361120000100
120204600351201051011200041001200083004800361201082001200122003601921120000100
120204600441201051011200041001200083004800361201082001200122003600361120000100
120204600351201051011200041001200083004800361201082001200122003600361120000100
120204600351201051011200041001200083004802161201562001200662003600361120000100
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)
120024601601200161112000510120009304800441200202012001420360000112000010
120024600411200111112000010120000304800001200102012000020360000112000010
120024600351200111112000010120000304800001200102012000020360000112000010
120024600351200111112000010120000304800001200102012000020360000112000010
120024600351200111112000010120000304800001200102012000020360000112000010
120024600351200111112000010120000304800001200102012000020360000112000010
120024600351200111112000010120000304800001200102012000020360000112000010
120024600351200111112000010120000304800001200102012000020360000112000010
120024600351200111112000010120000304800001200102012000020360186112000010
120024623061200161112000510120009304800521200222012001520360042112000010