Apple M1 Microarchitecture Research by Dougall Johnson

Firestorm: Overview | Base Instructions | SIMD and FP Instructions
Icestorm:  Overview | Base Instructions | SIMD and FP Instructions

SQDMULL (by element, 4S)

Test 1: uops

Code:

  sqdmull v0.4s, v0.4h, v1.h[1]
  movi v0.16b, 1
  movi v1.16b, 2

(no loop instructions)

1000 unrolls and 1 iteration

Retires: 1.000

Issues: 1.000

Integer unit issues: 0.000

Load/store unit issues: 0.000

SIMD/FP unit issues: 1.000

retire (01)cycle (02)03181e3f4e51inst issue (52)~issue fp/simd (54)~dispatch fp/simd (57)huge thing fp/simd (5b)60696d6edispatch stall (70)72scheduler rewind (75)scheduler stall (76)~dispatch op (78)~map op fp/simd (7e)~map lookup fp/simd (81)8283pipeline redirect (84)85inst all (8c)inst neon or fp (9a)a8accfd5d6ddinst fetch restart (de)e0? fp/simd (ee)f5f6f7f8fd
10043037230061254825100010001000398313130183037303724150328951000100020003037303711100110000073316332630100030383038303830383038
10043037230061254825100010001000398313130183037303724150328951000100020003037303711100110002073316332630100030383038303830383038
10043037230061254825100010001000398313130183037303724150328951000100020003037303711100110000073316332630100030383038303830383038
10043037220061254825100010001000398313130183037303724150328951000100020003037303711100110000073316332630100030383038303830383086
10043037220061254825100010001000398313130183037303724150328951000100020003037303711100110000073316332630100030383038303830383038
10043037220061254825100010001000398313130183037303724150328951000100020003037303711100110000073316332630100030383038303830383038
100430372300200254825100010001000398313130183037303724150328951000100020003037303711100110000073316332630100030383038303830383038
100430372300307254825100010001000398313130183037303724150328951000100020003037303711100110000073316332630100030383038303830383038
10043037230061254825100010001000398313130183037303724150328951000100020003037303711100110000073316332630100030383038303830383038
100430372309061254825100010001000398313130183037303724150328951000100020003037303711100110001073316332630100030383038303830383038

Test 2: Latency 1->2

Code:

  sqdmull 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): 3.0037

retire (01)cycle (02)0318191e3f4e51inst issue (52)~issue int (53)~issue fp/simd (54)~dispatch int (56)~dispatch fp/simd (57)huge thing int (59)huge thing fp/simd (5b)60696d6edispatch stall (70)scheduler rewind (75)scheduler stall (76)~dispatch op (78)~map op int (7c)~map op fp/simd (7e)~map lookup int (7f)~map lookup fp/simd (81)8283pipeline redirect (84)85inst all (8c)inst branch (8d)inst branch taken (90)inst b.cc (94)inst integer (97)inst neon or fp (9a)9fa6a8a9acc2cfd5d6ddinst fetch restart (de)e0? int output thing (e9)? fp/simd (ee)gpr retires (ef)f5f6f7f8fd
1020430037225000929295482510100100100001001000050042773130300183003730037282653287451010020010000200200003003730037111020110099100100100001000500071011611296340100001003003830038300383003830038
1020430228225000166295482510159107100161001000050042773131300183003730037282653287451010020010000200200003003730037111020110099100100100001000000071011611296340100001003003830038300383003830038
1020430037225000277295482510100100100001001000050042773131300183003730037282653287451010020010000200200003003730037111020110099100100100001000000271011611296340100001003003830038300383003830038
1020430037225000126295482510100100100001001000050042773131300183003730037282653287451010020010000200200003003730037111020110099100100100001000000071011611296340100001003003830038300383003830038
1020430037225000191295482510100100100001001000050042773131300183003730037282653287451010020010000200200003003730037111020110099100100100001000000071011611296340100001003003830038301343008630038
1020430037225000146295482510100100100001001000050042813490300183003730037282653287451010020010000200200003003730037111020110099100100100001000000071031611296340100001003003830038300383003830038
1020430037225000187295482510100100100001001000050042773130300183003730037282653287451010020010000200200003003730037111020110099100100100001000003071011611296340100001003003830038300383003830038
10204300372253101452954825101001001000010010000500427731313001830037300372826532874510100200100002002000030037300375110201100991001001000010001039071011611296340100001003003830038300383003830038
1020430037225000147295482510100100100001001000050042773131300183003730037282653287451010020010000206200003003730037111020110099100100100001002022780471011611296670100001003003830038300383008630038
1020430037226000147295484510100100100081001014950042786701300183003730037282653287451010020010000200200003003730037111020110099100100100001000003071011611296340100001003003830038300383003830038

1000 unrolls and 10 iterations

Result (median cycles for code): 3.0037

retire (01)cycle (02)03080b18191e1f3f4e51inst issue (52)~issue int (53)~issue fp/simd (54)~dispatch int (56)~dispatch fp/simd (57)huge thing int (59)huge thing fp/simd (5b)696b6d6edispatch stall (70)72scheduler rewind (75)scheduler stall (76)~dispatch op (78)~map op int (7c)~map op fp/simd (7e)~map lookup int (7f)~map lookup fp/simd (81)8283pipeline redirect (84)85inst all (8c)inst branch (8d)inst branch taken (90)inst b.cc (94)inst integer (97)inst neon or fp (9a)9fa0a1a6a7a8a9acc2cfd5d6ddinst fetch restart (de)e0? int output thing (e9)? fp/simd (ee)gpr retires (ef)f5f6f7f8fd
10024300372250000006129548251001010100001010000504277313300180300373003728287032876710010201000020200003003730037111002110910101000010000000006403163329630010000103003830038300383003830038
10024300372240000006729548251001010100001010000504277313300180300373003728287032876710010201000020200003003730037111002110910101000010000000006403163329630010000103003830038300383003830038
100243003722500000040629548251001010100001010000504277313300180300373003728287032876710010201000020200003003730037111002110910101000010000000006403163329630010000103003830038300383003830038
10024300372250000008229548251001010100001010000504277313300180300373003728287032876710010201000020200003003730037111002110910101000010000000006403163329630010000103003830038300383003830038
10024300372250000008229548251001010100001010000504277313300183300373003728287032876710010201000020200003003730037111002110910101000010000000006403163329630410000103007930038300383003830038
10024300372250000008229548251001010100001010000504277313300180300373003728287032876710010201000020200003003730037111002110910101000010000000006403163329630010000103003830038300383003830038
100243003722500000010329548251001010100001010000504277313300180300373003728287032876710010201000020200003003730037111002110910101000010000000006403163329630010000103003830038300383003830038
10024300372250000008429548251001010100001010000504277313300180300373003728287732876710010201000020200003003730037111002110910101000010000000006403163329630010000103003830038300383003830038
100243003722400000017029548251001010100001010000504277313300180300373003728287032876710010201000020200003003730037111002110910101000010000000006403163329630010000103003830038300383003830038
10024300372250000008229548251001010100001010000504277313300180300373003728287032876710010201000020200003003730037111002110910101000010000000006403163329630010000103003830038300383003830038

Test 3: Latency 1->3

Code:

  sqdmull 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): 3.0037

retire (01)cycle (02)031e3f4e51inst issue (52)~issue int (53)~issue fp/simd (54)~dispatch int (56)~dispatch fp/simd (57)huge thing int (59)huge thing fp/simd (5b)60696d6edispatch stall (70)scheduler rewind (75)scheduler stall (76)~dispatch op (78)~map op int (7c)~map op fp/simd (7e)~map lookup int (7f)~map lookup fp/simd (81)8283pipeline redirect (84)85inst all (8c)inst branch (8d)inst branch taken (90)inst b.cc (94)inst integer (97)inst neon or fp (9a)9fa8accdcfd5d6ddinst fetch restart (de)e0? fp/simd (ee)gpr retires (ef)f5f6f7f8fd
1020430037225061295482510100100100001001000050042773130300183003730037282653287451010020010000200200003003730037111020110099100100100001000007102161129634100001003003830038300383003830038
10204300372250124295482510100100100001001000050042773130300183003730037282653287451010020010000200200003003730037111020110099100100100001000007101161129634100001003003830038300383003830038
1020430037225061295482510100100100001001000050042773130300183003730037282653287451010020010000200200003003730037111020110099100100100001000007101161129634100001003003830038300383003830038
1020430037224061295482510100100100001001000050042773130300183003730037282653287451010020010000200200003003730037111020110099100100100001000007101161129634100001003003830038300383003830038
1020430037225089295482510100100100001001000050042773130300183003730037282653287451010020010000200200003003730037111020110099100100100001000007101161129634100001003003830038300383003830038
1020430037225061295482510100100100001001000050042773130300183003730037282653287451010020010000200200003003730037111020110099100100100001000007101161129634100001003003830038300383003830038
1020530037225061295482510100100100001001000050042773130300183003730037282653287451010020010000200200003003730037111020110099100100100001000007101161129634100001003003830038300383003830038
1020430037225061295482510100100100001001000050042773130300183003730037282653287451010020010000200200003003730037111020110099100100100001000007101161129634100001003003830038300383003830038
1020430037225061295482510100100100001001000050042773130300183003730037282653287451010020010000200200003003730037111020110099100100100001000007101161129634100001003003830038300383003830038
1020430037224061295482510100100100001021000050042773131300183003730037282653287451010020010000200200003003730037111020110099100100100001000007101161129634100001003003830038300383003830038

1000 unrolls and 10 iterations

Result (median cycles for code): 3.0037

retire (01)cycle (02)03080b18191e1f3f4e51inst issue (52)~issue int (53)~issue fp/simd (54)~dispatch int (56)~dispatch fp/simd (57)huge thing int (59)huge thing fp/simd (5b)60696d6edispatch stall (70)72scheduler rewind (75)scheduler stall (76)~dispatch op (78)~map op int (7c)~map op fp/simd (7e)~map lookup int (7f)~map lookup fp/simd (81)8283pipeline redirect (84)85inst all (8c)inst branch (8d)inst branch taken (90)inst b.cc (94)inst integer (97)inst neon or fp (9a)9fa1a6a7a8a9acc2cfd5d6ddinst fetch restart (de)e0? int output thing (e9)? fp/simd (ee)gpr retires (ef)f5f6f7f8fd
1002430037225000000842954825100101010000101000050427731313001830037300372828703287671001020100002020000300373003711100211091010100001000000006403163329630010000103003830038300383003830038
10024300372250000001832954825100101010000101000050427731313001830037300372828703287671001020100002020000300373003721100211091010100001000000006403163329630010000103003830038300383003830038
10024300372250000007262954825100101010000101000050427731313001830037300372828703287671001020100002020000300373003711100211091010100001000000006403163329630010000103003830038300383003830038
10024300372250000001562954825100101010000101000050427731313001830037300372828703287671001020100002020344300373003711100211091010100001000000006403163329630010000103003830038300383003830038
1002430037225000000612954825100101010000101000050427731313001830037300372828703287671001020100002020000300373003711100211091010100001000010306403163329630010000103003830038300383003830038
1002430037225000000612954825100101010000101000050427731303001830037300372828703287671001020100002020000300373003711100211091010100001000000006403163329630010000103003830038300383003830038
1002430037225000000612954825100101010000101000050427731303001830037300372828703287671001020100002020000300373003711100211091010100001000000006403163329630010000103003830038300383003830038
100243003722500000108612954825100101010000101000050427731303001830037300372828703287671001020100002020000300373003711100211091010100001000000006403163329700010000103003830038300383003830038
1002430037225000000612954825100101010000101000050427731313001830037300372828703287671001020100002020000300373003711100211091010100001000000006403163329630010000103003830038300383003830083
10024300372250000001472954825100101010000101000050427731303001830037300372828703287671001020100002020000300373003711100211091010100001000000006403163329630010000103003830038300383003830038

Test 4: throughput

Count: 8

Code:

  sqdmull v0.4s, v8.4h, v9.h[1]
  sqdmull v1.4s, v8.4h, v9.h[1]
  sqdmull v2.4s, v8.4h, v9.h[1]
  sqdmull v3.4s, v8.4h, v9.h[1]
  sqdmull v4.4s, v8.4h, v9.h[1]
  sqdmull v5.4s, v8.4h, v9.h[1]
  sqdmull v6.4s, v8.4h, v9.h[1]
  sqdmull 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): 0.2505

retire (01)cycle (02)03081e3f51inst issue (52)~issue int (53)~issue fp/simd (54)~dispatch int (56)~dispatch fp/simd (57)huge thing int (59)huge thing fp/simd (5b)60696d6edispatch stall (70)scheduler rewind (75)scheduler stall (76)~dispatch op (78)~map op int (7c)~map op fp/simd (7e)~map lookup int (7f)~map lookup fp/simd (81)8283pipeline redirect (84)85inst all (8c)inst branch (8d)inst branch taken (90)inst b.cc (94)inst integer (97)inst neon or fp (9a)9fa7a8acc5cfd5d6ddinst fetch restart (de)e0? int output thing (e9)? fp/simd (ee)gpr retires (ef)f5f6f7f8fd
8020420058150004125801001008000010080000500640000020020200392003999733999780100200800002001600002003920039118020110099100100800001000000511021611200360800001002004020040200402004020040
8020420039150006225801001008000010080000500640000020020200392003999733999780100200800002001600002003920039118020110099100100800001000000511031611200360800001002004020040200402004020040
8020420039150004125801001008000010080000500640000020020200392003999733999780100200800002001600002003920039118020110099100100800001000090511011611200360800001002004020040200402004020040
8020420039150004125801001008000010080000500640000020020200392003999733999780100200800002001600002003920039118020110099100100800001000000511011611200360800001002004020040200402004020040
8020420039150004125801001008000010080000500640000020020200392003999733999780100200800002001600002003920039118020110099100100800001000000511021611200360800001002009620040200402004020040
8020420039150004125801001008000010080000500640000020020200392003999733999780100200800002001600002003920039118020110099100100800001000000511021611200360800001002004020040200402004020040
8020420039150004125801001008000010080000500640000020020200392003999733999780100200800002001600002003920039118020110099100100800001000000511021611200360800001002004020040200402004020040
8020420039150004125801001008000010080000500640000020020200392003999733999780100200800002001600002003920039118020110099100100800001000000511011611200360800001002004020040200402004020040
8020420039150004125801001008000010080000500640000020020200392003999733999780100200800002001600002003920039118020110099100100800001000000511011611200360800001002004020040200402004020040
8020420039150004125801001008000010080000500640000020020200392003999733999780100200800002001600002003920039118020110099100100800001000000511021611200360800001002004020040200402004020040

1000 unrolls and 10 iterations

Result (median cycles for code divided by count): 0.2505

retire (01)cycle (02)03080b18191e1f3f51inst issue (52)~issue int (53)~issue fp/simd (54)~dispatch int (56)~dispatch fp/simd (57)huge thing int (59)huge thing fp/simd (5b)60696d6edispatch stall (70)scheduler rewind (75)scheduler stall (76)~dispatch op (78)~map op int (7c)~map op fp/simd (7e)~map lookup int (7f)~map lookup fp/simd (81)8283pipeline redirect (84)85inst all (8c)inst branch (8d)inst branch taken (90)inst b.cc (94)inst integer (97)inst neon or fp (9a)9fa8accfd2d5d6d9daddinst fetch restart (de)e0? int output thing (e9)? fp/simd (ee)gpr retires (ef)f5f6f7f8fd
800242003915000000010525800101080000108000050640000020020200392003999963100198001020800002016000020039200391180021109101080000100050200416004620036080000102004020040200402004020040
800242003915000000040258001010800001080000506400000200202003920039999631001980010208000020160000200392003911800211091010800001000502002160051820036080000102004020040200402004020040
800242003915000000040258001010800001080000506400001200202003920039999631001980010208000020160000200392003911800211091010800001000502004160041520036080000102004020040200402004020040
800242003915000000040258001010800001080000506400001200202003920039999631001980010208000020160000200392003911800211091010800001000502004160041920036080000102004020040200402004020040
800242003915000000040258001010800001080000506400000200202003920039999631004680010208000020160000200392003911800211091010800001000502004160041320036080000102004020040200402004020040
8002420039150000000103258001010800001080000506400000200202003920039999631001980010208000020160000200392003911800211091010800001000502004160021120036080000102004020040200402004020040
8002420039160000000221258001010800001080000506400000200202003920039999631001980010208000020160000200392003911800211091010800001000502004160041720036080000102004020040200402004020040
8002420039150000015063258001010800001080000506400000200202003920039999631001980010208000020160000200392003911800211091010800001000502004160041720036080000102004020040200402004020040
800242003915000000040258001010800001080000506400000200202003920039999631001980010208000020160000200392003911800211091010800001000502004160041420036080000102004020040200402004020040
800242003915000000040258001010800001080000506400001200202003920039999631001980010208000020160000200392003911800211091010800001000502002160041820036080000102004020040200402004020040