Datapoint 2200 IA-32. main memory. components. implemented by Intel in the Nicholas FitzRoy-Dale

Transcription

1 Datapoint 2200 IA-32 Nicholas FitzRoy-Dale At the forefront of the computer revolution - Intel Difficult to explain and impossible to love - Hennessy and Patterson! Released 1970! 2K shift register main memory! CPU: ~100 TTL components! Instruction set implemented by Intel in the 8008 History History DP2200, 8008 Precursors to x86 Pentium II SIMD extensions. 8088, 8086, 81016, Before IA-32: "x86". Rather dull. Pentium III SIMD extensions # IA bit addressing. Paging. Pentium 4 Netburst microarchitecture. SIMD extensions #3 and #4, virtualisation, 64-bit support i486 On-chip L1. Pipelining. Pentium M Pentium M microarchitecture. Pentium Superscalar! Branch prediction. SMP. Core family SIMD extensions try #5. Pentium Pro P6 microarchitecture. Dynamic execution. Speculation.

2 Legacy Microarchitectures! CISC ISA! Lack of registers! Four modes of operation! Segmentation! New one every two years! Focus on P6! Superscalar! Complex addressing modes P6 microarchitecture External Bus L2 Cache P6 microarchitecture Bus Interface Unit Instruction Fetch Unit Memory Reorder Buffer Data Cache Unit Memory Interface Unit! As RISC as possible! Register renaming Instruction Decoder Branch Target Buffer Microinstruction Sequencer Register Alias Table Reservation Station Address Generation Unit Integer Execution Unit Floating Point Execution Unit Reorder Buffer and Retirement Register File! Superscalar! Out-of-order execution! Speculation Image from

3 P6: As RISC as possible Instruction decode L1 I-cache! Pentium 4: Trace cache! Pentium M, Core: µ-op fusion Complex instruction decoder Simple instruction decoder 0 Simple instruction decoder 1 µ-op sequencer! ARM: Single instruction per clock! IA-64: Two identical decoders (currently) µ-op queue P6 family instruction decoder P6: Register renaming Avoiding stalls! Performed by ROB! 40 physical registers in RAT! ARM: Lots of GPRs! IA-64: Lots of GPRs and register rotation! Out-of-order execution! Branch prediction! Speculation

4 Out-of-order execution P6: Superscalar µ-op queue 6 µ-ops Reservation station Reservation station Reorder buffer 20 entries 40 entries Store data Store address Load address FPU CIU SIU P6 family OOO architecture Memory units Integer units Superscalar: Netburst They can fricassee or roast a duck in half the time! - Intel! ARM: Superscalar! Varies per-core: Port 0 Port 1 Port 2 Port 3! None ALU 0 Double speed ADD/SUB Logic Store Data Branches FP Move FP Move FP Store Data FXCH ALU 1 Double speed ADD/SUB Integer Operation Normal speed Shift/Rotate FP Execute FP_ADD FP_MUL FP_DIV FP_MISC MMX_SHFT MMX_ALU MMX_MISC Memory Load All Loads LEA Prefetch Memory Store Store Address! Separate ALU, MAC and LS (ARM11)! IA-64:! 2 I-units, 2 F-units, 3 B-units! But very few templates avoid split issue Source:

5 P6: Branch prediction P6: Speculation! Branch target buffer: 512 entries! Branch history and predicted address! Mispredicts: cycles! Static prediction! ~90% hit rate! Pentium M: Loop detection Reservation station ROB Execution units ROB Write P6 speculative execution Registers Floating-point EAX ECX EDX EBX EBP ESP ESI EDI EIP EFLAGS AH CH DH BH BP SP SI DI IP FLAGS CS SS DS ES FS GS AL CL DL BL Accumulator Count Data Base of data Base of stack Stack pointer String source idx String dest idx Instruction pointer CPU flags Code segment Stack segment Data segment Extra data segment Extra data segment 2 Extra data segment 3 79 FP stack FP status Source: Hennessy and Patterson Source: Hennessy and Patterson

6 MMX SSE, SSE2, SSE3, SSE4 79 FPR 63 0 MM0 MMX GP register 0 FPR MM1 MMX GP register 1 FPR MM2 MMX GP register 2 FPR MM3 MMX GP register 3 FPR MM4 MMX GP register 4 FPR MM5 MMX GP register 5 FPR MM6 MMX GP register 6 FPR MM7 MMX GP register XMM0 SSE GP register 0 XMM1 SSE GP register 1 XMM2 SSE GP register 2 XMM3 SSE GP register 3 XMM4 SSE GP register 4 XMM5 SSE GP register 5 XMM6 SSE GP register 6 XMM7 SSE GP register 7 IA-32 Instructions Instruction counts 550 Instructions (x86) 516 #! $ ()%$ )*+ #! # (),' What everyone (usually) targets #!! #! &# #! &$ #! &% #! &' i486 PentiumPPro P. MMX P2 P3 P4 #! &"! "! #!! #"! $!! $"! %!! %"! $ python2.4 instrcount.py /bin/* `which gcc` `which vim`

7 Instruction counts IA-32 vs ARM Percentage IA-32 gcc4.1 IA-32 gcc3.3 P4 IA-32 gcc4.1 P4 IA-32 icc9.0 P4 ARM mov call jmp je test lea push nop cmp pop 0 health power perimeter em3d em3d core compute_nodes() The code ARM IA-32 other_value = localnode->from_values[i]; cur_value -= coeff * value; ldr r4, [r0, #12] ldr r2, [r4, lr, lsl #2] mufd f0, f2, f1 sufd f3, f3, f0 mov ecx,dword PTR [eax+12] mov eax,dword PTR [ecx+ebx*4] mulsd xmm1,xmm2 subsd xmm0,xmm1 <function prologue> stmdb sp!, {r4, r5, lr} push ebp mov ebp, esp and esp, 0xfffffff8 sub esp, 0x10 mov DWORD PTR [esp], esi mov DWORD PTR [esp+4], edi mov DWORD PTR [esp+8], ebx <function epilogue> ldmia sp!, {r4, r5, pc} mov esi, DWORD PTR [esp] mov edi, DWORD PTR [esp+4] mov ebx, DWORD PTR [esp+8] mov esp, ebp pop ebp ret SIMD These instructions tend to be solutions, not primitives... - H&P! MMX: Integer only. Renames the FP registers! SSE: 8 new registers, each with 4 ints or 4 single-precision floats. Cache control.! SSE2: Lots more data types! SSE3: Horizontal operations! SSE4: Unknown

8 Floating point Thread-level parallelism! SMP! 8087: Stack architecture! P III, P4, Core: SSE, SSE2, SSE3, SSE4! Pentium Extreme Edition! Core Duo! Hyperthreading SMP Hyperthreading Instruction stream Execution units! P III, 4, Core: MESI (Athlon: MOESI)! Bus arbitration via dedicated lines! Interrupts: Whichever gets them! Replicate register renaming, architectural registers! Partition re-order buffer! Share caches and execution units! Relatively large performance gain Execution units Instruction streams Image: arstechnica.com Image: arstechnica.com

9 Performance The future of IA-32! Speed demon P4 was disappointing! Focus on ILP / TLP 120,! Dual-core designs! Hyperthreading a, 100-! EMT64! Virtualisable! Multicore I a7 io i Relative area (percentage) Source: IEEE Micro seuqnoit!!?s