SSE: Vector Floats for x86

CS 441 Lecture, Dr. Lawlor

We covered an old, simple, but powerful idea today-- "SIMD", which stands for Single Instruction Multiple Data:
You can do lots interesting SIMD work without using any special instructions--plain old C will do, if you treat an "int" as 32 completely independent bits, because any normal "int" instructions will operate on all the bits at once.  This use of bitwise operations is often called "SIMD within a register (SWAR)" or "word-SIMD"; see Sean Anderson's "Bit Twiddling Hacks" for a variety of amazing examples.

Back in the 1980's, "vector" machines were quite popular in supercomputing centers.  For example, the 1988 Cray Y-MP was a typical vector machine.  When I was an undergraduate, ARSC still had a Y-MP vector machine.  The Y-MP had eight "vector" registers, each of which held 64 doubles (that's a total of 4KB of registers!).  A single Y-MP machine language instruction could add all 64 corresponding numbers in two vector registers, which enabled the Y-MP to achieve the (then) mind-blowing speed of *millions* floating-point operations per second.  Vector machines have now almost completely died out; the NEC SV-1 and the Japanese "Earth Simulator" are the last of this breed.  Vector machines are classic SIMD, because one instruction can modify 64 doubles.

But the most common form of SIMD today are the "multimedia" instruction set extensions in normal CPUs.  The usual arrangment for multimedia instructions is for single instructions to operate on four 32-bit floats.  These four-float instructions exist almost everywhere nowdays:
We'll look at the x86 version below.

SSE Assembly

SSE instructions were first introduced with the Intel Pentium II, but they're now found on all modern x86 processors, including the 64-bit versions.  SSE introduces 8 new registers, called xmm0 through xmm7, that each contain four 32-bit single-precision floats.  New instructions that operate on these registers have the suffix "ps", for "Packed Single-precision".  See the x86 reference manual for a complete list of SSE instructions.

For example, "add" adds two integer registers, like eax and ebx.  "addps" adds two SSE registers, like xmm3 and xmm6.  There are  SSE versions of most other arithmetic operations: subps, mulps, divps, etc.

There are actually two flavors of the SSE move instruction: "movups" moves a value between *unaligned* addresses (not a multiple of 16); "movaps" moves a value between *aligned* addresses.  The aligned move is substantially faster, but it will segfault if the address you give isn't a multiple of 16!   Luckily, most "malloc" implementations return you 16-byte aligned data, and the stack is aligned to 16 bytes with most compilers.  But for maximum portability, for the aligned move to work you sometimes have to manually add padding (using pointer arithmetic), which is really painful.

Here's a simple SSE assembly example where we load up a constant, add it to itself with SSE, and then write it out:
	movups xmm1,[thing1]; <- copy the four floats into xmm1
addps xmm1,xmm1; <- add floats to themselves
movups [retval],xmm1; <- move that constant into the global "retval"

; Print out retval
extern farray_print
push 4 ;<- number of floats to print
push retval ;<- points to array of floats
call farray_print
add esp,8 ; <- pop off arguments
ret

section .data
thing1: dd 10.2, 100.2, 1000.2, 10000.2;<- source constant
retval: dd 0.0, 0.0, 0.0, 0.0 ;<- our return value

(Try this in NetRun now!)

Here's a version that loads up two separate float constants, and adds them:

	movups xmm1,[thing1]; <- copy the four floats into xmm1
movups xmm6,[thing2]; <- copy the four floats into xmm1
addps xmm1,xmm6; <- add floats
movups [retval],xmm1; <- move that constant into the global "retval"

; Print out retval
extern farray_print
push 4 ;<- number of floats to print
push retval ;<- points to array of floats
call farray_print
add esp,8 ; <- pop off arguments
ret

section .data
thing1: dd 10.2, 100.2, 1000.2, 10000.2;<- source constant
thing2: dd 1.2, 2.2, 3.2, 4.2;<- source constant
retval: dd 0.0, 0.0, 0.0, 0.0 ;<- our return value

(Try this in NetRun now!)

The "dd" lines declare a 32-bit constant.  NASM is smart enough to automatically use float format if you type a constant with a decimal point!

There's one curious operation called "shufps", which does a  permutation of the incoming floats.  The permutation is controlled by 4 2-bit "select" fields, one for each destination float, that determines which incoming float goes to that slot.  For example, a swizzle can copy the first float into all four destination slots, with a select field of 0x00.  To copy the last float into all four destination slots, the select field is 0xff.  A select field of 00010111, or 0x17, would set the first float from the last (float 3), copy float 1 into the middle two slots, and fill the last slot with the first float (float 0).  This is the *only* operation which rearranges floats between different slots in the registers.  (In graphics language, this operation is called a "swizzle".)   For example, if xmm1 contains the four floats "(0.0, 1.1, 2.2, 3.3)", then:
      shufps xmm1,xmm1, 0xff
will set all 4 of xmm1's floats to the last value (index 11).

Scalar
Single-precision
(float)
Scalar
Double-precision
(double)
Packed
Single-precision
(4 floats)
Packed
Double-precision
(2 doubles)
Comments
add
addss
addsd
addps
addpd
sub, mul, div all work the same way
min
minss
minsd
minps
minpd
max works the same way
sqrt
sqrtss
sqrtsd
sqrtps
sqrtpd
Square root (sqrt), reciprocal (rcp), and reciprocal-square-root (rsqrt) all work the same way
mov
movss
movsd
movaps (aligned)
movups (unaligned)
movapd (aligned)
movupd (unaligned)
Aligned loads are up to 4x faster, but will crash if given an unaligned address!  Stack is always 16-byte aligned specifically for this instruction. Use "align 16" directive for static data.
cvt cvtss2sd
cvtss2si
cvttss2si

cvtsd2ss
cvtsd2si
cvttsd2si
cvtps2pd
cvtps2dq
cvttps2dq
cvtpd2ps
cvtpd2dq
cvttpd2dq
Convert to ("2", get it?) Single Integer (si, stored in register like eax) or four DWORDs (dq, stored in xmm register).  "cvtt" versions do truncation (round down); "cvt" versions round to nearest.
com
ucomiss
ucomisd
n/a
n/a
Sets CPU flags like normal x86 "cmp" instruction, from SSE registers.
cmp
cmpeqss
cmpeqsd
cmpeqps
cmpeqpd
Compare for equality ("lt", "le", "neq", "nlt", "nle" versions work the same way).  Sets all bits of float to zero if false (0.0), or all bits to ones if true (a NaN).  Result is used as a bitmask for the bitwise AND and OR operations.
and
n/a
n/a
andps
andnps
andpd
andnpd
Bitwise AND operation.  "andn" versions are bitwise AND-NOT operations (A=(~A) & B).  "or" version works the same way.

The easy way to get SSE output is to just convert to integer, like this:
movss xmm3,[pi]; load up constant
addss xmm3,xmm3 ; add pi to itself
cvtss2si eax,xmm3 ; round to integer
ret
section .data
pi: dd 3.14159265358979 ; constant

(Try this in NetRun now!)

It's annoyingly tricky to display full floating-point values.  The trouble here is that our function "foo" returns an int to main, so we have to call a function to print floating-point values.  Also, with SSE floating-point, on a 64-bit machine you're supposed to keep the stack aligned to a 16-byte boundary (the SSE "movaps" instruction crashes if it's not given a 16-byte aligned value).  Sadly, the "call" instruction messes up your stack's alignment by pushing an 8-byte return address, so we've got to use up another 8 bytes of stack space purely for stack alignment, like this.
movss xmm3,[pi]; load up constant
addss xmm3,xmm3 ; add pi to itself
movss [output],xmm3; write register out to memory

; Print floating-point output
mov rdi,output ; first parameter: pointer to floats
mov rsi,1 ; second parameter: number of floats
sub rsp,8 ; keep stack 16-byte aligned (else get crash!)
extern farray_print
call farray_print
add rsp,8

ret

section .data
pi: dd 3.14159265358979 ; constant
output: dd 0.0 ; overwritten at runtime

(Try this in NetRun now!)


SSE in C/C++

The x86 SSE instructions can be accessed from C/C++ via the header <xmmintrin.h>.  (Corresonding Apple headers exist for PowerPC AltiVec; the AltiVec instructions have different names but are almost identical.)   The xmmintrin header exists and works out-of-the-box with most modern compilers:
Documentation can be found by just reading the xmmintrin header.  The underlying instructions have different names, which are listed in the x86 reference manual and summarized below and at this graphical and helpful SSE site.  Keep in mind that the names of these things are absurdly ugly; I take this as a hint that you should wrap them in a nicer interface, such as a "vec4" class of your own making.

The C version of an SSE register is the type "__m128".  gcc provides factory overloads for arithmetic on __m128's.  All the instructions start with "_mm_" (i.e., MultiMedia).  The suffix indicates the data type; in these examples, we'll just be talking about 4 floats, which use the suffix "_ps" (Packed Single-precision floats).  SSE supports other data types in the same 128 bits, but 4 floats seems to be the sweet spot.  So, for example, "_mm_load_ps" loads up 4 floats into a __m128, "_mm_add_ps" adds 4 corresponding floats together, etc.  Major useful operations are:
__m128 _mm_load_ps(float *src)
Load 4 floats from a 16-byte aligned address.  WARNING: Segfaults if the address isn't a multiple of 16!
__m128 _mm_loadu_ps(float *src) Load 4 floats from an unaligned address (4x slower!)
__m128 _mm_load1_ps(float *src) Load 1 individual float into all 4 fields of an __m128
__m128 _mm_setr_ps(float a,float b,float c,float d)
Load 4 separate floats from parameters into an __m128
void _mm_store_ps(float *dest,__m128 src)
Store 4 floats to an aligned address.
void _mm_storeu_ps(float *dest,__m128 src) Store 4 floats to unaligned address
__m128 _mm_add_ps(__m128 a,__m128 b)
Add corresponding floats (also "sub")
__m128 _mm_mul_ps(__m128 a,__m128 b) Multiply corresponding floats (also "div", but it's slow)
__m128 _mm_min_ps(__m128 a,__m128 b) Take corresponding minimum (also "max")
__m128 _mm_sqrt_ps(__m128 a) Take square roots of 4 floats (12ns, slow like divide)
__m128 _mm_rcp_ps(__m128 a) Compute rough (12-bit accuracy) reciprocal of all 4 floats (as fast as an add!)
__m128 _mm_rsqrt_ps(__m128 a) Rough (12-bit) reciprocal-square-root of all 4 floats (fast)
__m128 _mm_shuffle_ps(__m128 lo,__m128 hi,
       _MM_SHUFFLE(hi3,hi2,lo1,lo0))
Interleave inputs into low 2 floats and high 2 floats of output. Basically
   out[0]=lo[lo0];
   out[1]=lo[lo1];
   out[2]=hi[hi2];
   out[3]=hi[hi3];
For example, _mm_shuffle_ps(a,a,_MM_SHUFFLE(i,i,i,i)) copies the float a[i] into all 4 output floats.
There are also instructions for integer conversion, comparsions, various bitwise operations, and cache-friendly prefetches and streaming store operations.

So take your classic "loop over floats":
	for (int i=0;i<n_vals;i++) { 
        vals[i]=vals[i]*a+b;
}
(executable NetRun link)

This takes about 4.5 ns per float.

Step 1 is to unroll this loop 4 times, since SSE works on blocks of 4 floats at a time:
	for (int i=0;i<n_vals;i+=4) { 
        vals[i+0]=vals[i+0]*a+b;
        vals[i+1]=vals[i+1]*a+b;
        vals[i+2]=vals[i+2]*a+b;
        vals[i+3]=vals[i+3]*a+b;
}
(Try this in NetRun now!)

This alone speeds the code up by about 2x, because we don't have to check the loop counter as often.

We can then replace the guts of the loop with SSE instructions:
	__m128 SSEa=_mm_load1_ps(&a);
__m128 SSEb=_mm_load1_ps(&b);
__m128 v=_mm_load_ps(&vec[i]);
v=_mm_add_ps(_mm_mul_ps(v,SSEa),SSEb);
_mm_store_ps(&vec[i],v);

(Try this in NetRun now!)

This gives us about a 4x speedup over the original, and still a 2x speedup over the unrolled version!
(Note: I'd previously called "_mm_load_ps1" in the above example, which works, but is a bit slower.)

Note that if we use the unaligned loads and stores (loadu and storeu), we lose almost all the performance benefit from using SSE in the first place!

Apple points out that even when you're doing arithmetic on actual 3D vectors (for example), you might *not* want to store one vector per SSE register.  The reason is that not every vector operation maps nicely onto SSE, for example the final summation in a dot product takes two calls to the ever-so-funky haddps instruction.  It's sometimes better to just unroll each of your vector loops by four, collect up four adjacent vector X, Y, or Z coordinates, and then do your usual computation on these blocks of adjacent vectors (all X's get added to all Y's, for example).   This question is usually phrased as "Array of Structures, or Structure of Arrays?"  There's a good description of the tradeoffs on page 5&6 of this PDF.

AoS: Array of structures: four 3-vectors takes four registers, one per vector.  Not clear what to put in the 4th float of each vector (maybe w?), so may need padding.
a.x
a.y
a.z
??
b.x
b.y
b.z
??
c.x
c.y
c.z
??
d.x
d.y
d.z
??

SoA: Structure of arrays: four 3-vectors takes just three registers, one per component.
a.x
b.x
c.x
d.x
a.y
b.y
c.y
d.y
a.z
b.z
c.z
d.z

AoS vs SoA: SSE Matrix-Vector Multiply

It's informative to look at the performance of matrix-vector multiply.  I'll pick a 4x4 matrix, just to match SSE data sizes.  To start with, the naive float version takes 45ns on a Pentium 4, and quite nearly the same speed on a newer Q6600 (serial performance of newer processors is pretty much identical).
enum {n=4};
float mat[n][n];
float vec[n];
float outvector[n];

int foo(void) {
for (int row=0;row<4;row++) {
float sum=0.0;
for (int col=0;col<n;col++) {
float m=mat[row][col];
float v=vec[col];
sum+=m*v;
}
outvector[row]=sum;
}
return 0;
}

(Try this in NetRun now!)

Unrolling the inner loop, as ugly as it is, speeds things up substantially, to 26ns:
enum {n=4};
float mat[n][n];
float vec[n];
float outvector[n];

int foo(void) {
for (int row=0;row<4;row++) {
float sum=0.0, m,v;
m=mat[row][0];
v=vec[0];
sum+=m*v;
m=mat[row][1];
v=vec[1];
sum+=m*v;
m=mat[row][2];
v=vec[2];
sum+=m*v;
m=mat[row][3];
v=vec[3];
sum+=m*v;
outvector[row]=sum;
}
return 0;
}

(Try this in NetRun now!)

Making a line-by-line transformation to SSE doesn't really buy any performance, at 25ns:
#include <pmmintrin.h>

enum {n=4};
__m128 mat[n]; /* rows */
__m128 vec;
float outvector[n];

int foo(void) {
for (int row=0;row<n;row++) {
__m128 mrow=mat[row];
__m128 v=vec;
__m128 sum=mrow*v;
sum=_mm_hadd_ps(sum,sum); /* adds adjacent-two floats */
_mm_store_ss(&outvector[row],_mm_hadd_ps(sum,sum)); /* adds those floats */
}
return 0;
}

(Try this in NetRun now!)

The trouble here is that we can cheaply operate on 4-vectors, but summing up the elements of those 4-vectors (with the hadd instruction) is expensive.  We can eliminate that horizontal summation by operating on columns, although now we need a new matrix layout.  This is down to 19ns on a Pentium 4, and just 12ns on the Q6600!
#include <xmmintrin.h>

enum {n=4};
__m128 mat[n]; /* by column */
float vec[n];
__m128 outvector;

int foo(void) {
float z=0.0;
__m128 sum=_mm_load1_ps(&z);
for (int col=0;col<n;col++) {
__m128 mcol=mat[col];
float v=vec[col];
sum+=mcol*_mm_load1_ps(&v);
}
outvector=sum;
return 0;
}

(Try this in NetRun now!)

The bottom line is that, as is *typical*, just making the inner loop parallel requires extensive changes to the data layout and processing throughout the program.

Per-Float Branching in SSE

There are a really curious set of instructions in SSE to support per-float branching:
Compare-and-AND is actually useful to simulate branches.  The situation where these are useful is when you're trying to convert a loop like this to SSE:
	for (int i=0;i<n;i++) { 
        if (vec[i]<7)
vec[i]=vec[i]*a+b;
else
vec[i]=c;
}
(Try this in NetRun now!)

You can implement this branch by setting a mask indicating where vals[i]<7, and then using the mask to pick the correct side of the branch to squash:
	for (int i=0;i<n;i++) { 
        unsigned int mask=(vec[i]<7)?0xffFFffFF:0;
vec[i]=((vec[i]*a+b)&mask) | (c&~mask);
}
Written in ordinary sequential code, this is actually a slowdown, not a speedup!  But in SSE this branch-to-logical transformation means you can keep barreling along in parallel, without having to switch to sequential floating point to do the branches:
	__m128 A=_mm_load1_ps(&a), B=_mm_load1_ps(&b), C=_mm_load1_ps(&c);
__m128 Thresh=_mm_load1_ps(&thresh);
for (int i=0;i<n;i+=4) {
__m128 V=_mm_load_ps(&vec[i]);
__m128 mask=_mm_cmplt_ps(V,Thresh); // Do all four comparisons
__m128 V_then=_mm_add_ps(_mm_mul_ps(V,A),B); // "then" half of "if"
__m128 V_else=C; // "else" half of "if"
V=_mm_or_ps( _mm_and_ps(mask,V_then), _mm_andnot_ps(mask,V_else) );
_mm_store_ps(&vec[i],V);
}

(Try this in NetRun now!)

This gives about a 3.8x speedup over the original loop on my machine!

Intel hinted in their Larrabee paper that NVIDIA is actually doing this exact float-to-SSE branch transformation in CUDA, NVIDIA's very high-performance language for running sequential-looking code in parallel on the graphics card.

Hiding SSE Nastiness In a Wrapper Class

SSE is just ugly; comparisons doubly so.  You can hide the ugliness inside a "wrapper class":
#include <xmmintrin.h>

class fourfloats; /* forward declaration */

/* Wrapper around four bitmasks: 0 if false, all-ones (NAN) if true.
This class is used to implement comparisons on SSE registers.
*/
class fourmasks {
__m128 mask;
public:
fourmasks(__m128 m) {mask=m;}
__m128 if_then_else(fourfloats dthen,fourfloats delse);
};

/* Nice wrapper around __m128:
it represents four floating point values. */
class fourfloats {
__m128 v;
public:
fourfloats(float onevalue) { v=_mm_load1_ps(&onevalue); }

fourfloats(__m128 ssevalue) {v=ssevalue;} // put in an SSE value
operator __m128 () const {return v;} // take out an SSE value

fourfloats(const float *fourvalues) { v=_mm_load_ps(fourvalues);}
void store(float *fourvalues) {_mm_store_ps(fourvalues,v);}

/* arithmetic operations return blocks of floats */
fourfloats operator+(const fourfloats &right) {
return _mm_add_ps(v,right.v);
}

/* comparison operations return blocks of masks (bools) */
fourmasks operator<(const fourfloats &right) {
return _mm_cmplt_ps(v,right.v);
}
};

inline __m128 fourmasks::if_then_else(fourfloats dthen,fourfloats delse) {
return _mm_and_ps(mask,dthen)+
_mm_andnot_ps(mask,delse);
}

float src[4]={1.0,5.0,3.0,4.0};
float dest[4];
int foo(void) {
/*
// Serial code
for (int i=0;i<4;i++) {
if (src[i]<4.0) dest[i]=src[i]*2.0; else dest[i]=17.0;
}
*/
// Parallel code
fourfloats s(src);
fourfloats d=(s<4.0).if_then_else(s+s,17.0);
d.store(dest);

//farray_print(dest,4); // <- for debugging
return 0;
}

(Try this in NetRun now!)

Compilers are now good enough that there is zero speed penalty due to the nice "fourfloats" class: the nice syntax comes for free!