Real Arithmetic
Computer Organization and Assembly Languages Yung-Yu Chuang
2006/12/11
Announcement
• Homework #4 is extended for two days
• Homework #5 will be announced today, due two weeks later.
• Midterm re-grading by this Thursday.
IA-32 floating point architecture
• Original 8086 only has integers. It is possible to simulate real arithmetic using software, but it is slow.
• 8087 floating-point processor (and 80287, 80387) was sold separately at early time.
• Since 80486, FPU (floating-point unit) was
integrated into CPU.
FPU data types
• Three floating-point types
FPU data types
• Four integer types
FPU registers
• Data register
• Control register
• Status register
• Tag register
Data registers
• Load: push, TOP--
• Store: pop, TOP++
• Instructions access the stack using ST(i)
relative to TOP
• If TOP=0 and push, TOP wraps to R7
• If TOP=7 and pop, TOP wraps to R0
• When overwriting occurs, generate an exception
• Real values are transferred to and from memory and stored in 10-byte temporary format. When storing, convert back to integer, long, real, long real.
79 0
R0 R1 R2 R3 R4 R5 R6 R7
ST(0) ST(1) ST(2)
010 TOP
Postfix expression
• (5*6)-4 → 5 6 * 4 -
5 5
5 6
6
30
*
30 4
4
26 -
Special-purpose registers
Special-purpose registers
• Last data pointer stores the memory address of the operand for the last non-control instruction.
Last instruction pointer stored the address of the last non-control instruction. Both are 48 bits, 32 for offset, 16 for segment selector.
1 1 0 1 1
Control register
Initial 037Fh
The instruction FINIT will initialize it to 037Fh.
Rounding
• FPU attempts to round an infinitely accurate result from a floating-point calculation
– Round to nearest even: round toward to the closest one; if both are equally close, round to the even one – Round down: round toward to -∞
– Round up: round toward to +∞
– Truncate: round toward to zero
• Example
– suppose 3 fractional bits can be stored, and a calculated value equals +1.0111.
– rounding up by adding .0001 produces 1.100
– rounding down by subtracting .0001 produces 1.011
Rounding
1.011 1.0111
Truncate
1.100 1.0111
Round up
1.011 1.0111
Round down
1.100 1.0111
Round to nearest even
rounded value original value
method
-1.011 -1.0111
Truncate
-1.011 -1.0111
Round up
-1.100 -1.0111
Round down
-1.100 -1.0111
Round to nearest even
rounded value original value
method
• Six types of exception conditions
– #I: Invalid operation – #Z: Divide by zero
– #D: Denormalized operand – #O: Numeric overflow
– #U: Numeric underflow – #P: Inexact precision
• Each has a corresponding mask bit
– if set when an exception occurs, the exception is handled automatically by FPU
– if clear when an exception occurs, a software exception handler is invoked
Floating-Point Exceptions
detect before execution
detect after execution
Status register
C3-C0: condition bits after comparisons
.data
bigVal REAL10 1.212342342234234243E+864 .code
Fld bigVal
FPU data types
FPU instruction set
• Instruction mnemonics begin with letter F
• Second letter identifies data type of memory operand
– B = bcd – I = integer
– no letter: floating point
• Examples
– FLBD load binary coded decimal – FISTP store integer and pop stack
– FMUL multiply floating-point operands
FPU instruction set
• Operands
– zero, one, or two
– no immediate operands
– no general-purpose registers (EAX, EBX, ...) (FSTSW is the only exception which stores FPU status word to AX)
– if an instruction has two operands, one must be a FPU register
– integers must be loaded from memory onto the stack and converted to floating-point before being used in calculations
Instruction format
{…}: implied operands
Classic stack
• ST(0) as source, ST(1) as destination. Result is
stored at ST(1) and ST(0) is popped, leaving the
result on the top.
Real memory and integer memory
• ST(0) as the implied destination. The second
operand is from memory.
Register and register pop
• Register: operands are FP data registers, one must be ST.
• Register pop: the same as register with a ST
pop afterwards.
Example: evaluating an expression
Load
FLDPI stores π
FLDL2T stores log2(10) FLDL2E stores log2(e) FLDLG2 stores log10(2) FLDLN2 stores ln(2)
load
.data
array REAL8 10 DUP(?) .code
fld array ; direct
fld [array+16] ; direct-offset
fld REAL8 PTR[esi] ; indirect
fld array[esi] ; indexed
fld array[esi*8] ; indexed, scaled
fld REAL8 PTR[ebx+esi]; base-index
fld array[ebx+esi] ; base-index-displacement
Store
Store
fst dblOne
fst dblTwo
fstp dblThree
fstp dblFour
Register
Arithmetic instructions
FCHS ; change sign of ST FABS ; ST=|ST|
Floating-Point add
• FADD
– adds source to destination
– No-operand version pops the FPU stack after addition
• Examples:
Floating-Point subtract
• FSUB
– subtracts source from destination.
– No-operand version pops the FPU stack after subtracting
• Example:
fsub mySingle ; ST -= mySingle
fsub array[edi*8] ; ST -= array[edi*8]
Floating-point multiply/divide
• FMUL
– Multiplies source by destination, stores product in destination
• FDIV
– Divides destination by source, then pops the stack
Example: compute distance
; compute D=sqrt(x^2+y^2)
fld x ; load x
fld st(0) ; duplicate x
fmul ; x*x
fld y ; load y
fld st(0) ; duplicate y
fmul ; y*y
fadd ; x*x+y*y
fsqrt fst D
Example: expression
; expression:valD = –valA + (valB * valC).
.data
valA REAL8 1.5 valB REAL8 2.5 valC REAL8 3.0
valD REAL8 ? ; will be +6.0 .code
fld valA ; ST(0) = valA
fchs ; change sign of ST(0) fld valB ; load valB into ST(0) fmul valC ; ST(0) *= valC
fadd ; ST(0) += ST(1)
fstp valD ; store ST(0) to valD
Example: array sum
.data N = 20
array REAL8 N DUP(1.0) sum REAL8 0.0
.code
mov ecx, N
mov esi, OFFSET array
fldz ; ST0 = 0
lp: fadd REAL8 PTR [esi]; ST0 += *(esi)
add esi, 8 ; move to next double loop lp
fstp sum ; store result
Comparisons
Comparisons
• The above instructions change FPU’s status register of FPU and the following instructions are used to transfer them to CPU.
• SAHF copies C
0into carry, C
2into parity and C
3to zero. Since the sign and overflow flags are
not set, use conditional jumps for unsigned
integers (ja, jae, jb, jbe, je, jz).
Comparisons
Branching after FCOM
• Required steps:
1. Use the FSTSW instruction to move the FPU status word into AX.
2. Use the SAHF instruction to copy AH into the EFLAGS register.
3. Use JA, JB, etc to do the branching.
• Pentium Pro supports two new comparison
instructions that directly modify CPU’s FLAGS.
FCOMI ST(0), src ; src=STn FCOMIP ST(0), src
Example
fcomi ST(0), ST(1) jnb Label1
Example: comparison
.data
x REAL8 1.0 y REAL8 2.0 .code
; if (x>y) return 1 else return 0
fld x ; ST0 = x
fcomp y ; compare ST0 and y
fstsw ax ; move C bits into FLAGS sahf
jna else_part ; if x not above y, ...
then_part:
mov eax, 1 jmp end_if else_part:
mov eax, 0 end_if:
Example: comparison
.data
x REAL8 1.0 y REAL8 2.0 .code
; if (x>y) return 1 else return 0
fld y ; ST0 = y
fld x ; ST0 = x ST1 = y fcomi ST(0), ST(1)
jna else_part ; if x not above y, ...
then_part:
mov eax, 1 jmp end_if else_part:
mov eax, 0 end_if:
Comparing for Equality
• Not to compare floating-point values directly because of precision limit. For example,
sqrt(2.0)*sqrt(2.0) != 2.0
ST(0): +4.4408921E-016 fsub two
ST(0): +2.0000000E+000 fmul ST(0), ST(1)
ST(0): +1.4142135+000 fsqrt
ST(0): +2.0000000E+000 fld two
FPU stack instruction
Comparing for Equality
• Calculate the absolute value of the difference between two floating-point values
.data
epsilon REAL8 1.0E-12 ; difference value val2 REAL8 0.0 ; value to compare
val3 REAL8 1.001E-13 ; considered equal to val2 .code
; if( val2 == val3 ), display "Values are equal".
fld epsilon fld val2
fsub val3 fabs
fcomi ST(0),ST(1) ja skip
mWrite <"Values are equal",0dh,0ah>
skip:
Miscellaneous instructions
.data
x REAL4 2.75
five REAL4 5.2 .code
fld five ; ST0=5.2
fld x ; ST0=2.75, ST1=5.2
fscale ; ST0=2.75*32=88
; ST1=5.2
Example: quadratic formula
Example: quadratic formula
Example: quadratic formula
Other instructions
• F2XM1 ; ST=2
ST(0)-1; ST in [-1,1]
• FYL2X ; ST=ST(1)*log
2(ST(0))
• FYL2XP1 ; ST=ST(1)*log
2(ST(0)+1)
• FPTAN ; ST(0)=1;ST(1)=tan(ST)
• FPATAN ; ST=arctan(ST(1)/ST(0))
• FSIN ; ST=sin(ST) in radius
• FCOS ; ST=sin(ST) in radius
• FSINCOS ; ST(0)=cos(ST);ST(1)=sin(ST)
Exception synchronization
• Main CPU and FPU can execute instructions concurrently
– if an unmasked exception occurs, the current FPU instruction is interrupted and the FPU signals an exception
– But the main CPU does not check for pending FPU exceptions. It might use a memory value that the interrupted FPU instruction was supposed to set.
– Example:
.data
intVal DWORD 25 .code
fild intVal ; load integer into ST(0) inc intVal ; increment the integer
Exception synchronization
• Exception is issued and pended; the next floating-point instruction checks exceptions before it executes.
• For safety, insert a fwait instruction, which tells the CPU to wait for the FPU's exception handler to finish:
.data
intVal DWORD 25 .code
fild intVal ; load integer into ST(0)
fwait ; wait for pending exceptions inc intVal ; increment the integer
Mixed-mode arithmetic
• Combining integers and reals.
– Integer arithmetic instructions such as ADD and MUL cannot handle reals
– FPU has instructions that promote integers to reals and load the values onto the floating point stack.
• Example: Z = N + X
.data
N SDWORD 20 X REAL8 3.5 Z REAL8 ? .code
fild N ; load integer into ST(0) fwait ; wait for exceptions
fadd X ; add mem to ST(0) fstp Z ; store ST(0) to mem
Mixed-mode arithmetic
int N=20;
double X=3.5;
int Z=(int)(N+X);
fild N fadd X
fist Z ; Z=24
fstcw ctrlWord
or ctrlWord, 110000000000b ; round mode=truncate fldcw ctrlWord
fild N fadd X
fist Z ; Z=23
Masking and unmasking exceptions
• Exceptions are masked by default
– Divide by zero just generates infinity, without halting the program
• If you unmask an exception
– processor executes an appropriate exception handler – Unmask the divide by zero exception by clearing bit 2
.data
ctrlWord WORD ? .code
fstcw ctrlWord ; get the control word and ctrlWord,1111111111111011b ; unmask #Z
fldcw ctrlWord ; load it back into FPU
Homework #5
for (i=0; i<m; i++)
for (j=0; j<p; j++) { C[I,j]=0;
for (r=0; r<n; r++)
C[i,j]+=A[i,r]*B[r,j];
}
• Strassen’s algorithm?
• Coppersmith & Winograd O(n2.376)
• Memory coherence
m
n
n
p
Homework #5
A C
k i
B
k j
i
j
/* ijk */
for (i=0; i<n; i++) { for (j=0; j<n; j++) {
sum = 0.0;
for (k=0; k<n; k++)
sum += a[i][k] * b[k][j];
c[i][j] = sum;
} }
/* ijk */
for (i=0; i<n; i++) { for (j=0; j<n; j++) {
sum = 0.0;
for (k=0; k<n; k++)
sum += a[i][k] * b[k][j];
c[i][j] = sum;
} }
/* jik */
for (j=0; j<n; j++) { for (i=0; i<n; i++) {
sum = 0.0;
for (k=0; k<n; k++)
sum += a[i][k] * b[k][j];
c[i][j] = sum }
}
/* jik */
for (j=0; j<n; j++) { for (i=0; i<n; i++) {
sum = 0.0;
for (k=0; k<n; k++)
sum += a[i][k] * b[k][j];
c[i][j] = sum }
}
Homework #5
/* kij */
for (k=0; k<n; k++) { for (i=0; i<n; i++) {
r = a[i][k];
for (j=0; j<n; j++)
c[i][j] += r * b[k][j];
} }
/* kij */
for (k=0; k<n; k++) { for (i=0; i<n; i++) {
r = a[i][k];
for (j=0; j<n; j++)
c[i][j] += r * b[k][j];
} }
/* ikj */
for (i=0; i<n; i++) { for (k=0; k<n; k++) {
r = a[i][k];
for (j=0; j<n; j++)
c[i][j] += r * b[k][j];
} }
/* ikj */
for (i=0; i<n; i++) { for (k=0; k<n; k++) {
r = a[i][k];
for (j=0; j<n; j++)
c[i][j] += r * b[k][j];
} }
/* jki */
for (j=0; j<n; j++) { for (k=0; k<n; k++) {
r = b[k][j];
for (i=0; i<n; i++)
c[i][j] += a[i][k] * r;
} }
/* jki */
for (j=0; j<n; j++) { for (k=0; k<n; k++) {
r = b[k][j];
for (i=0; i<n; i++)
c[i][j] += a[i][k] * r;
} }
/* kji */
for (k=0; k<n; k++) { for (j=0; j<n; j++) {
r = b[k][j];
for (i=0; i<n; i++)
c[i][j] += a[i][k] * r;
} }
/* kji */
for (k=0; k<n; k++) { for (j=0; j<n; j++) {
r = b[k][j];
for (i=0; i<n; i++)
c[i][j] += a[i][k] * r;
} }
Homework #5
0 10 20 30 40 50 60
25 50 75 100 125 150 175 200 225 250 275 300 325 350 375 400 Array size (n)
Cycles/iteration
kji jki kij ikj jik ijk
kji & jki
kij & ikj
jik & ijk
Blocked array
0 10 20 30 40 50 60
25 50 75 100 125 150 175
200 225 250 275
300 325 350 375 400 Array size (n)
Cycles/iteration
kji jki kij ikj jik ijk
bijk (bsize = 25) bikj (bsize = 25)