SIMD

A 4-way SIMD unit

First introduced on the Illiac IV, and not to be confused with vector processing.

Compilers and Assemblers

• GCC can issue SIMD four different ways:
  • Target-specific code generation will sometimes use scalar SIMD
  • Target-specific builtins, using vector extensions
  • Inline assembly, using syntax of the GNU Assembler and a baroque register specification convention
  • Auto-vectorization from gcc's autovect branch attempts to autovectorize some loops

Data Types (taken from SSE specs)

These do not necessarily map to the C data types of the same name, for any given compiler!

• half precision: 16-bit IEEE 754 floating-point (bias-15) (IEEE 754 2008 binary16)
• single: 32-bit IEEE 754 floating-point (bias-127) (IEEE 754 2008 binary32)
• double: 64-bit IEEE 754 floating-point (bias-1023) (IEEE 754 2008 binary64)
• long double: 80-bit "double extended" IEEE 754-1985 floating-point (bias-16383)
  • not an actual SIMD type, but an artifact of x87
• word: 32-bit two's complement integer
• doubleword, dword: 64-bit two's complement integer

x86 YMM

The AVX (Advanced Vector eXtensions) were introduced on Intel's Sandy Bridge (2010) and AMD's Bulldozer (2011). They operate on the 256-bit YMM registers (YMM0..YMM15), which are aliased by the XMM registers. They are encoded using the VEX scheme. Support for AVX can be determined via:

• Determine that CPUID.1:ECX.OSXSAVE[bit 27] is set (XGETBV is enabled for application use)
• Issue XGETBV and verify that XFEATURE_ENABLED_MASK[2:1] = 0x3 (XMM state and YMM state are enabled by OS)
• Determine that CPUID.1:ECX.AVX[bit 28] is set (AVX instructions supported)

x86 XMM

The Streaming SIMD Extensions operate on the 128-bit XMM registers (XMM0..XMM7 in 32-bit mode, XMM0..XMM15 in 64-bit mode). In its original incarnation on the PIII, execution units (but not registers) were shared with the x87 floating-point architecture. The execution units were separated in the NetBurst microarchitecture. In the Core microarchitecture, the execution engine has been widened for greater SSE throughput.

SSE5 (AMD)

• Unimplemented extensions competing with SSE4, encoded using a method incompatible with VEX
• Withdrawn, converted into VEX-compatible encodings, and split into:
  • FMA4: Fused floating-point multiply-add (compare Intel's FMA)
  • XOP: Fused integer multiply-add, byte permutations, shifts, rotates, integer vector horizontal operations (compare Intel's SSE4)
  • CVT16: Half-precision conversion

SSE4 (Intel)

SSE4.1

DPPD instruction dataflow

• Introduced on Penryn
• dpps -- dot product of two vectors having four single components each
• dppd -- dot product of two vectors having two double components each
• insertps

SSE4.2

Intel Nehalem microarchitecture

• Introduced on Nehalem
• crc32
• pcmpestri
• pcmpestrm
• pcmpistri
• pcmpistrm
• pcmpgtq
• popcnt

SSE4a (AMD)

• lzcnt
• popcnt
• extrq
• insertq
• movntsd
• movntss

SSE3 (PNI)

• Originally known as Prescott New Instructions, and introduced on P4-Prescott
• movddup -- move a double from a 8-byte-aligned memory location or lower half of XMM register to upper half, then duplicate upper half to lower half
• MWAIT/MONITOR

SSSE3 (TNI/MNI)

• Introduced with the Core microarchitecture. Sometimes referred to as Tejas New Instructions or Merom New Instructions
• pmaddwd -- multiply packed words, then horizontally sum pairs, accumulating into doublewords

SSE2

• Introduced with the P4.
• Extends the MMX instructions to XMM registers.
• movapd -- move two packed doubles from a 16-byte-aligned memory location to XMM registers, or vice versa, or between two XMM registers.
  • movupd -- movapd safe for unaligned memory references, with far inferior performance.
• mulpd -- multiply two packed doubles. the multiplier is a 16-byte-aligned memory location or XMM register. the target XMM register serves as the multiplicand.
• addpd -- add two packed doubles. the addend is a 16-byte-aligned memory location or XMM register. the target XMM register serves as the augend.

SSE (KNI/ISSE)

• Introduced with the PIII. Sometimes referred to as Katmai New Instructions, and branded for some time as the Internet Streaming SIMD Extensions
• movaps -- move four packed singles from a 16-byte-aligned memory location to XMM registers, or vice versa, or between two XMM registers.
  • movups -- movaps safe for unaligned memory references, with far inferior performance.
• mulps -- multiply four packed singles. the multiplier is a 16-byte-aligned memory location or XMM register. the target XMM register serves as the multiplicand.
• addps -- add four packed singles. the addend is a 16-byte-aligned memory location or XMM register. the target XMM register serves as the augend.

Future Directions

• The FMA instruction set extension to x86 should hit around 2011, providing floating-point fused multiply-add
  • AMD appears to call this FMA4, part of what was SSE5

Other Architectures

• PowerPC implements AltiVec
• SPARC implements VIS, the Visual Instruction Set
• PA-RISC implements MAX, the Multimedia Acceleration eXtensions
• ARM implements NEON
• Alpha implemented MVI, the Motion Video Instructions
• SWAR: SIMD Within a Register (bit-parallel methods)

See Also

• "Why no FMA in AVX in Sandy Bridge?", Intel Developers Forum
• SSE5 guide at AMD
• SSE4 reference at Intel
• 2007-04-19 post to http://virtualdub.org, "SSE4 finally adds dot products"
• AMD64 Architecture Programmer’s Manual Volume 6: 128-Bit and 256-Bit XOP, FMA4 and CVT16 Instructions
• Agner Fog's x86 Instruction Tables
• General architecture page