Chapter14 - THE ARITHMETIC COPROCESSOR, MMX, AND SIMD TECHNOLOGIES

Chapter 14: The Arithmetic Coprocessor, MMX, and
SIMID Technologies
Introduction
• The Intel family of arithmetic coprocessors
includes the 8087, 80287, 80387SX, 80387DX,
and 80487SX for use
with the 80486SX microprocessor.
• This chapter provides detail on the entire
family of arithmetic coprocessors.
• Many programs now require or at least
benefit from a coprocessor.
The Intel Microprocessors: 8086/8088, 80186/80188, 80286, 80386, 80486 Pentium,

Pentium Pro Processor, Pentium II, Pentium, 4, and Core2 with 64-bit Extensions Copyright ©2009 by Pearson Education, Inc.
Architecture, Programming, and Interfacing, Eighth Edition Upper Saddle River, New Jersey 07458 • All rights reserved.
Barry B. Brey
Chapter Objectives
Upon completion of this chapter, you will be able to:
• Convert between decimal data and signed

integer, BCD, and floating-point data for use
by the arithmetic coprocessor, MMX, and
SIMD technologies.
• Explain the operation of the 80X87 arithmetic
coprocessor and the MMX and SIMD units.

Barry B. Brey
Chapter Objectives (cont.)
Upon completion of this chapter, you will be able to:
• Explain the operation and addressing modes

of each arithmetic coprocessor, MMX, and
SSE instruction.
• Develop programs that solve complex
arithmetic problems using the arithmetic
coprocessor, MMX, and SIMD instructions.

Barry B. Brey
14–1 DATA FORMATS FOR THE
ARITHMETIC COPROCESSOR
• This section presents types of data used with
all arithmetic coprocessor family members.
• These data types include signed integer,
BCD, and floating-point.
• Each has a specific use in a system, and
many systems require all three data types.

Barry B. Brey
Signed Integers
• When used with the coprocessor, signed
integers are 16- (word), 32- (doubleword
integer), or 64-bits (quadword integer) wide.
• Positive numbers are stored in true form
with a leftmost sign-bit of 0.
• Negative numbers are stored in two’s
complement form with a leftmost sign-bit of 1.

Barry B. Brey
• Word integers range in value from –32,768
to +32,767.
– the doubleword integer range is ±2  109
– the quadword integer range is ±9  1018
• Data are stored in memory using assembler
directives described in earlier chapters.
– DW directive defines words
– DD defines doubleword integers
– DQ defines quadword integers
• Figure 14–1 shows these three forms of
signed integer data.
Barry B. Brey
Figure 14–1 Integer formats for the 8087 family of arithmetic coprocessors: (a) word,
(b) short, and (c) long.

Barry B. Brey
Binary-Coded Decimal (BCD)
• Each number is stored as an 18-digit packed
integer in nine bytes of memory as two digits
per byte.
– the tenth byte contains only a sign-bit for
the 18-digit signed BCD number
– BCD requires 80 bits of memory
• Fig 14–2 shows the format of the BCD
number used with the arithmetic coprocessor.

Barry B. Brey
Figure 14–2 BCD data format for the 80X87 family of arithmetic coprocessors.
– note both positive and negative numbers

are stored in true form and never in ten’s
complement form
– the DT directive stores BCD data in the memory
as illustrated in Example 14–2
– this form is rarely used because it is unique to
the Intel coprocessor

Barry B. Brey
Floating-Point Numbers
• Often called real numbers because they hold
signed integers, fractions, & mixed numbers.
• Floating-point numbers are written in scientific
binary notation and have three parts:
– a sign-bit, a biased exponent, and a significand
• Intel coprocessors support three types of
floating-point numbers:
– single (32 bits), double (64 bits) and
temporary (80 bits)
• See Fig 14–3 for the three forms.
Barry B. Brey
Figure 14–3 Floating-point (real) format for the 80X87 family of arithmetic
coprocessors. (a) Short (single-precision) with a bias of 7FH, (b) long (double-
precision) with a bias of 3FFH, and (c) temporary (extended-precision) with a bias of
3FFFH.

Barry B. Brey
Converting to Floating-Point Form.
• Converting decimal to floating-point is
accomplished through the following steps:
– 1. Convert the decimal number to binary
– 2. Normalize the binary number
– 3. Calculate the biased exponent
– 4. Store the number in floating-point format

Barry B. Brey
• Some special rules apply to a few numbers.
• The number 0 is stored as all zeros except
for the sign-bit.
– which can be logic 1 to represent a negative zero
• Plus and minus infinity is stored as logic 1s
in the exponent with a significand of all zeros
and the sign-bit that represents plus or minus.
• A NAN (not-a-number) is an invalid floating-
point result with all ones in the exponent with
a significand that is not all zeros.

Barry B. Brey
Converting from Floating-Point
Form
• Conversion to a decimal from floating-point:
– 1. Separate the sign-bit, biased exponent, and
significand
– 2. Convert the biased exponent into a true
exponent by subtracting the bias
– 3. Write as a normalized binary number
– 4. Convert it to a denormalized binary number
– 5. Convert the denormalized binary to decimal

Barry B. Brey
Storing Floating-Point Data in
Memory
• Some examples of floating-point data storage
are shown in Example 14–5.
• The Microsoft macro assembler contains an
error that does not allow a plus sign to be
used with positive floating-point numbers.
– a +92.45 must be defined as 92.45 for the
assembler to function correctly
• Corrected in version 6.11 of MASM.
Barry B. Brey
14–2 80X87 ARCHITECTURE
• 80X87 is an external integrated circuit
designed to operate concurrently with the
microprocessor.
– 80486DX–Core2 have internal versions of 80387
• The coprocessor is a special-purpose
microprocessor designed to efficiently execute
arithmetic and transcendental operations
• The processor executes all normal instructions
and 80X87 executes coprocessor instructions.
– 80X87 executes 68 different instructions
Barry B. Brey
• The microprocessor intercepts and executes
the normal instruction set.
• The coprocessor intercepts and executes
only the coprocessor instructions.
• Coprocessor instructions are actually escape
(ESC) instructions.
– used by the microprocessor to generate a
memory address for the coprocessor so the
coprocessor can execute a coprocessor
instruction

Barry B. Brey
Internal Structure of the 80X87
• 80X87 is divided into two major sections:
– the control unit and numeric execution unit
• Both devices monitor the instruction stream.
– if the instruction is an ESCape (coprocessor)
instruction, the coprocessor executes it
– if not, the microprocessor executes it
• The control unit interfaces the coprocessor
to the microprocessor-system data bus.
• The numeric execution unit (NEU) is
executes all coprocessor instructions.
Barry B. Brey
• The NEU has an eight-register stack that
holds operands for arithmetic instructions
and the results of arithmetic instructions.
– instructions address data in specific stack data
registers or use a push-and-pop mechanism to
store and retrieve data on the top of the stack
• Other registers in the NEU are status, control,
tag, and exception pointers.
• See Figure 14–4.

Barry B. Brey
Figure 14–4 The internal structure of the 80X87 arithmetic coprocessor.

Barry B. Brey
• FSTSW AX is the only coprocessor instruction
available that allows direct communication to
the microprocessor through the AX register.
– 8087 does not contain FSTSW AX, but newer
coprocessors do
• The stack within the coprocessor contains
eight registers that are each 80 bits wide.
– these always contain an 80-bit extended-
precision floating-point number
• The only time data appear as any other form
is when they reside in the memory system.
Barry B. Brey
Status Register
• The status register reflects the overall
operation of the coprocessor.
– accessed by executing the instruction (FSTSW)
• FSTSW AX copies the status register directly
to the microprocessor’s AX register.
– once stored, bit positions of the status register
can be examined by normal software

Barry B. Brey
Figure 14–5 The 80X87 arithmetic coprocessor status register.
– coprocessor/microprocessor communications are

carried out through I/O ports 00FAH–00FFH on
80187/80287, & ports 800000FAH–800000FFH
80386-Pentium 4
– never use these I/O ports for interfacing I/O
devices to the microprocessor
Barry B. Brey
Coprocessor Status Bits
B
• The busy bit indicates the coprocessor is
busy executing a task.
• Busy is tested by examining the status
register or by using the FWAIT instruction.
– newer coprocessors automatically synchronize
with the microprocessor
– busy need not be tested before performing
additional coprocessor tasks

Barry B. Brey
C0–C3
• The condition code bits indicate conditions
about the coprocessor.
• These bits have different meanings for
different instructions
– the top of the stack is denoted as ST

Barry B. Brey
TOP
• The top-of-stack (ST) bit indicates the
current register addressed as the top-of-the-
stack (ST).
– This is normally register ST(0).

Barry B. Brey
ES
• The error summary bit is set if any
unmasked error bit is set.
– PE, UE, OE, ZE, DE, or IE
• In the 8087 coprocessor, the error summary
also caused a coprocessor interrupt.
– since 80187, the coprocessor interrupt has
been absent from the family

Barry B. Brey
PE
• The precision error indicates that the result
or operands exceed the selected precision.
UE
• An underflow error indicates a nonzero
result too small to represent with the current
precision selected by the control word.

Barry B. Brey
OE
• An overflow error indicates a result too large
to be represented.
– if this error is masked, the coprocessor
generates infinity for an overflow error
ZE
• A zero error indicates the divisor was zero
while the dividend is a noninfinity or nonzero
number.
Barry B. Brey
DE
• A denormalized error indicates at least
one of the operands is denormalized.
IE
• An invalid error indicates a stack overflow or
underflow, indeterminate form (0 ÷ 0, +∞, –∞,
etc.), or use of a NAN as an operand.
– indicates errors such as produced by taking the
square root of a negative number, etc.
Barry B. Brey
Control Register
• The control register selects the precision,
rounding control, and infinity control.
– also masks and unmasks exception bits
corresponding to the rightmost 6 bits of
the status register
• FLDCW is used to load a value into the
control register.

Barry B. Brey
Control Register Bits
IC
• Infinity control selects either affine or
projective infinity.
– affine allows positive and negative infinity
– projective assumes infinity is unsigned
RC
• Rounding control determines the type of
rounding, as defined in Figure 14–6.
Barry B. Brey
Control Register Bits
PC
• The precision control sets the precision
of the result, as defined in Figure 14–6.
Exception Masks
• Determine whether the error indicated by the
exception affects the error bit in the status
register.
– if logic 1 is placed in an exception control bits,
the corresponding status register bit is masked
Barry B. Brey
Figure 14–6 The 80X87 arithmetic coprocessor control register.

Barry B. Brey
Tag Register
• The tag register indicates the contents of
each location in the coprocessor stack.
• The tag indicates whether a register is valid;
zero; invalid or infinity; or empty.
• Figure 14–7 illustrates the tag register and
the status indicated by each tag.

Barry B. Brey
Figure 14–7 The 80X87 arithmetic coprocessor tag register.
– a program views the tag register by storing the

coprocessor environment using FSTENV,
FSAVE, or FRSTOR
– each instruction stores the tag register along
with other coprocessor data

Barry B. Brey
14–3 INSTRUCTION SET
• When a coprocessor instruction references
memory, the processor generates the memory
address for the instruction.
• The coprocessor uses the data bus for data
transfers during coprocessor instructions.
– the processor uses it during normal instructions
• When assembler encounters a coprocessor
mnemonic opcode, it converts it into a
machine language ESC instruction.
– ESC represents an opcode to the coprocessor
Barry B. Brey
Data Transfer Instructions
• There are three basic data transfers:
– floating-point
– signed integer
– BCD
• The only time that data ever appear in the
signed integer or BCD form is in the memory.
• Inside the coprocessor, data are always stored
as an 80-bit extended-precision floating-point
number.

Barry B. Brey
Floating-Point Data Transfers
• There are four traditional floating-point
data transfer instructions in the coprocessor
instruction set:
– FLD (load real)
– FST (store real)
– FSTP (store real and pop)
– FXCH (exchange)

Barry B. Brey
• FLD loads floating-point memory data to the
top of the internal stack.
– referred to as ST (stack top)
• FST stores a copy of the top of the stack to
the memory location or coprocessor register
indicated by the operand.
• FSTP stores a copy of the top of the stack to
memory or any coprocessor register, then
pops the data from the top of the stack.
– one might think of FST as a copy instruction and
FSTP as a removal instruction
Barry B. Brey
• FXCH exchanges the register indicated by
the operand with the top of the stack.
– for example, FXCH ST(2) exchanges the top
of the stack with register 2
• A new instruction in Pentium Pro-Core2
uses opcode FCMOV with a floating-point
condition.
– called a conditional floating-point move
instruction

Barry B. Brey
Integer Data Transfer Instructions
• The coprocessor supports three integer data
transfer instructions:
– FILD (load integer)
– FIST (store integer)
– FISTP (store integer and pop)
• These function as do FLD, FST, and FSTP,
except data transferred are integer data.
• The coprocessor automatically converts the
internal extended temporary-precision floating-
point data to integer data.
Barry B. Brey
BCD Data Transfer Instructions
• Two instructions load or store BCD signed-
integer data.
• FBLD loads the top of the stack with BCD
memory data.
• FBSTP stores the top of the stack and
does a pop.

Barry B. Brey
The FCMOV Instruction
• Pentium Pro–Pentium 4 contain a new
instruction called FCMOV
– which also contains a condition
• If the condition is true, FCMOV copies the
source to the destination.
– conditions tested by FCMOV and the opcodes
used with FCMOV appear in Table 14–4
• The conditions check for either an ordered or
unordered condition.

Barry B. Brey
Arithmetic Instructions
• Arithmetic instructions for the coprocessor
include addition, subtraction, multiplication,
division, and calculating square roots.
– arithmetic-related instructions are scaling,
rounding, absolute value, and changing sign
• All arithmetic operations are floating-point,
except some cases in which memory data
are referenced as an operand.

Barry B. Brey
• Classic stack addressing operand data uses
the top of the stack as the source operand
and the next to the top of the stack as the
destination operand.
– afterward, a pop removes the source datum
– only the result in the destination register
remains at the top of the stack
• The top of the stack is always used as the
destination for the memory-addressing mode
because the coprocessor is a stack-oriented
machine.
Barry B. Brey
Arithmetic Operations
• The letter P in an opcode specifies a register
pop after the operation.
– FADDP compared to FADD
• The letter R in an opcode (subtraction and
division only) indicates reverse mode.
– subtracts the top of the stack from memory
and stores the result in the top of the stack
• Letter I as a second letter in an opcode
indicates the memory operand is an integer.
– also applies to FADD, FSUB, FMUL, and FDIV
Barry B. Brey
Arithmetic-Related Operations
FSQRT
• Finds the square root of the top of the
stack and leaves the resultant square
root at the top of the stack.
– an invalid error occurs for the square
root of a negative number
• The IE bit of the status register should be
tested when an invalid result can occur.
– test IE by loading the status register to AX with
FSTSW AX, followed by TEST AX,1 to test the bit
Barry B. Brey
FSCALE
• Adds the contents of ST(1) (interpreted as
an integer) to the exponent at the top of the
stack.
• FSCALE multiplies or divides rapidly by
powers of two.
– value in ST(1) must be between 2–15 and 2+15

Barry B. Brey
FPREM/FPREM1
• Performs modulo division of ST by ST(1).
• The resultant remainder is found in the top
of the stack and has the same sign as the
original dividend.
– a modulo division results in a remainder
without a quotient
• FPREM is supported for 8086 & 80287
– FPREM1 should be used in newer coprocessors

Barry B. Brey
FXTRACT
• Decomposes the number at the top of the
stack into separate parts representing value
of the unbiased exponent and significand.
• The extracted significand is found at the top
of the stack and the unbiased exponent at
ST(1).
– often used to convert a floating-point to a form
that can be printed as a mixed number

Barry B. Brey
FRNDINT
• Rounds the top of the stack to an integer.
FABS
• Changes the sign of the top of the stack to
positive.
FCHS
• Changes the sign from positive to negative
or negative to positive.
Barry B. Brey
Comparison Instructions
• Examine data at the top of the stack in
relation to another element and return
the result of the comparison in the status
register condition code bits C3–C0.
FCOM
• Compares the floating-point data at the top
of the stack with an operand, which may be
any register or any memory operand.
– if not coded with the instruction, the next stack
element ST(1) is compared with the stack top ST
Barry B. Brey
FCOMP/FCOMPP
• Both perform as FCOM, but also pop one
or two data from the stack.
FICOM/FICOMP
• The top of the stack is compared with the
integer stored at a memory operand.
– FICOMP also pops the top of the stack

Barry B. Brey
FTST
• Tests the contents of the top of the stack
against a zero.
• Result of the comparison is coded in the
status register condition code bits,
– as illustrated in Table 14–2
• Refer to Table 14–3 for a way of using
SAHF and the conditional jump instruction
with FTST.

Barry B. Brey
FXAM
• Examines the stack top and modifies the
condition code bits to indicate whether the
contents are positive, negative, normalized,
and so on.
• Refer to the status register in Table 14–2.

Barry B. Brey
FCOMI/FUCOMI (Pentium Pro-P4)
• Compares the same manner as FCOM with
one additional feature:
– it moves the floating-point flags into the flag
register, as do FNSTSW AX and SAHF
• Intel combined FCOM, FNSTSW AX, & SAHF
instructions to form FCOMI.
– also available is unordered compare FUCOMI
• All available with a pop by appending the
opcode with a P.
Barry B. Brey
Transcendental Operations
FPTAN
• Finds the partial tangent of Y/X = tan θ.
• The value of θ is at the top of the stack.
– must be between 0 & n/4 radians for 8087/80287
– less than 263 for the 80387 & above
• Result a ratio found as ST = X and ST(1) = Y.
– if outside allowable range, an invalid error occurs
indicated by the status register IE bit
• ST(7) must be empty for this instruction to
function properly.
Barry B. Brey
FPATAN
• Finds the partial arctangent as
θ = ARCTAN X/Y.
– the value of X is at the top of the stack
– Y is at ST(1)
• The values of X and Y must be 0 ≤ Y < X <∞.
• The instruction pops the stack and leaves θ
in radians at the top of the stack.

Barry B. Brey
F2XM1
• Finds the function 2X – 1.
– value of X is taken from the top of the stack and
the result is returned to the top of the stack
• To obtain 2X add one to the result at the top
of the stack.
– value of X must be in the range of –1 and +1
• Constants log2 10 and log2 ε are built in as
standard values for the coprocessor.
Barry B. Brey
FSIN/FCOS
• Finds sine or cosine of the argument located
in ST expressed in radians (360°=2π
radians), with the result found in ST.
– values of ST must be less than 263
FSINCOS
• Finds sine and cosine of ST, expressed in
radians, and leaves results as ST = sine and
ST(1) = cosine.
Barry B. Brey
FYL2X
• Finds Y log2 X.
– value X is taken from the stack top
– Y is taken from ST(1)
• Result is found at the stack top after a pop.
– value of X must range between 0 and ∞
– value of Y must be between –∞ and +∞
• A logarithm with any positive base (b) is found
by equation LOGb X = (LOG2 b)–1  LOG2 X.
Barry B. Brey
FYL2P1
• Finds Finds Y log2 (X + 1).
• The value of X is taken from the stack top
and Y is taken from ST(1).
• The result is found at the top of the stack
after a pop.

Barry B. Brey
Constant Operations
• The coprocessor instruction set includes
opcodes that return constants to the top
of the stack.
– FLDZ +0.0
– FLD1 +1.0
– FLDPI 3.14159
– FLDL2T log210
– FLDL2E log2 ε
– FLDLG2 log10 2
– FLDLN2 logε 2
Barry B. Brey
Coprocessor Control Instructions
• The coprocessor has control instructions for
initialization, exception handling, and task
switching.
• All control instructions have two forms.

Barry B. Brey
FINIT/FNINIT
• Performs a reset (initialize) operation on the
arithmetic coprocessor.
• The coprocessor operates with a closure of
projective (unsigned infinity), rounds to the
nearest or even, and uses extended-precision
when reset or initialized.
– also sets register 0 as the top of the stack

Barry B. Brey
FSETPM
• Changes the coprocessor to the protected-
addressing mode.
– used when the microprocessor is protected mode
• Protected mode can only be exited by a
hardware reset.
– or in 80386-Pentium 4, with a change to the
control register

Barry B. Brey
FLDCW
• Loads the control register with the
word addressed by the operand.
FSTCW
• Stores the control register into the
word-sized memory operand.

Barry B. Brey
FSTSW AX
• Copies the contents of the control register
to the AX register.
– not available to 8087
FCLEX
• Clears the error flags in the status register
and also the busy flag.

Barry B. Brey
FSAVE
• Writes the entire state of the machine to
memory.
– Fig 14–8 shows memory layout for this instruction
FRSTOR
• Restores the state of the machine from
memory. This instruction is used to restore
the information saved by FSAVE.
Barry B. Brey
Figure 14–8 Memory format when the 80X87 registers are saved with the FSAVE
instruction.

Barry B. Brey
FSTENV
• Stores the environment of the coprocessor.
– as shown in Figure 14–9
FLDENV
• Reloads the environment saved by FSTENV.
FINCSP
• Increments the stack pointer.
Barry B. Brey
Figure 14–9 Memory format for the FSTENV instruction: (a) real mode and (b)
protected mode.

Barry B. Brey
FDECSP
• Decrements the stack pointer.
FFREE
• Frees a register by changing the destination
register’s tag to empty.
– does not affect the contents of the register

Barry B. Brey
FNOP
• Floating-point coprocessor NOP.
FWAIT
• Causes the microprocessor to wait for the
coprocessor to finish an operation.
– should be used before the microprocessor
accesses memory data affected by the
coprocessor
Barry B. Brey
Coprocessor Instructions
• Newer coprocessors contain the same basic
instructions provided by earlier versions, with
a few additional instructions
• 80387-Core2 contain additional instructions:
– FCOS (cosine)
– FPREM1 (partial remainder)
– FSIN (sine)
– FSINCOS (sine and cosine)
– FUCOM/FUCOMP/FUCOMPP (unordered
compare)
Barry B. Brey
14–4 Programming with the
Arithmetic Coprocessor
• This section provides programming
examples for the arithmetic coprocessor.
• Each example is chosen to illustrate a
programming technique for the coprocessor.

Barry B. Brey
Calculating the Area of a Circle
• This first example illustrates a simple method
of addressing the coprocessor stack.
• The equation for calculating the area of a
circle is A = πR2.
• A program that performs this calculation is
listed in Example 14–8.
• The program takes test data from array
RAD that contains five sample radii.

Barry B. Brey
Figure 14–10 Operation of the coprocessor stack for one iteration of the loop in
Example 14–8.

Barry B. Brey
Finding the Resonant Frequency
• An equation commonly used in electronics is
the formula for determining the resonant
frequency of an LC circuit.
• This example uses L1 for the inductance L, C1
for the capacitor C, and RES for the resultant
resonant frequency.
• The equation solved by the program illustrated
in Example 14–9 is:

Barry B. Brey
Finding the Roots Using the
Quadratic Equation
• This example illustrates how to find the roots
of a polynomial expression (ax2 + bx + c = 0)
by using the quadratic equation:
• Example 14–10 illustrates a program that finds

the roots (R1 and R2) for the quadratic
equation.
– constants are stored in locations A1, B1, and C1
Barry B. Brey
Using a Memory Array to Store
Results
• Example 14–11 shows a program that
calculates 100 values of inductive reactance.
• Equation for inductive reactance is XL = 2pFL.
– frequency range is 10 Hz to 1000 Hz for F
– inductance of 4 mH
• FSTP XL[ECX*4 + 4] stores the reactance for
each frequency
• FCOMP clears the stack just before the RET
Barry B. Brey
Converting a Single-Precision
Floating-Point Number to a String
• This example takes floating-point contents of
a 32-bit single-precision floating-point number
and stores it as an ASCII character string.
– the procedure converts the floating-point number
as a mixed number with an integer part and a
fractional part, separated by a decimal point
• Example 14–12 lists a procedure to convert
the contents of memory location NUMB to a
string stored in the STR array.
Barry B. Brey
• A limit is placed on the mixed number size.
– integer portion is a 32-bit binary number (±2 G)
– fraction is a 24-bit binary number (1/16M)
– the procedure will not function properly for
larger or smaller numbers
• The procedure first tests the sign of the
number and stores a minus sign for a
negative number.
– after storing a minus sign the number is
made positive by the FABS instruction

Barry B. Brey
• Next, it is divided into integer and fractional
parts and stored at WHOLE and FRACT.
• The FRNDINT instruction is used to round
(using the chop mode) the top of the stack
to form the whole number part of NUMB.
– the whole number part is subtracted from the
original number to generate the fractional part

Barry B. Brey
14–5 INTRO TO MMX TECHNOLOGY
• MMX (multimedia extensions) technology
adds 57 new instructions to the instruction
set of the Pentium–P4 microprocessors.
– also new general-purpose instructions
• The new instructions are designed for
applications such as motion video, combined
graphics with video, image processing, audio
synthesis, speech synthesis and compression,
telephony, video conferencing, 2D graphics,
and 3D graphics.
Barry B. Brey
Data Types
• MMX architecture new packed data types:
– eight packed, consecutive 8-bit bytes
– four packed, consecutive 16-bit words
– two packed, consecutive 32-bit doublewords
• Bytes in this multibyte form have consecutive
memory addresses and use the little endian
form, as with other Intel data.
• See Figure 14–11 for the format for these
new data types.

Barry B. Brey
Figure 14–11 The structure of data stored in the MMX registers.

Barry B. Brey
• MMX technology registers have the same
format as a 64-bit quantity in memory and
have two data access modes:
– 64-bit access mode, for 64-bit memory and
registers transfers for most instructions
– 32-bit access mode for 32-bit memory and
register transfers for most instructions
• 32-bit transfers occur between processor
registers.
• 64-bit transfers occur between floating-point
coprocessor registers.
Barry B. Brey
• Fig 14–12 shows the internal register set of
MMX technology extensions and how it uses
the floating-point coprocessor register set.
– called aliasing because the floating-point
registers are shared as the MMX registers
• MMX registers (MM0–MM7) are the same as
the floating-point registers.
– the MMX register set is 64 bits wide and
uses the rightmost 64 bits of the floating-
point register set

Barry B. Brey
Figure 14–12 The structure of the MMX register set. Note that MM0 and FF0 through
MM7 and FP7 interchange with each other.

Barry B. Brey
Instruction Set
• The instruction for MMX technology includes
arithmetic, comparison, conversion, logical,
shift, and data transfer instructions.
• The instruction types are similar to the
microprocessor’s instruction set.
– MMX instructions use data types shown in
Fig 14–11 instead of the normal data types
used with the microprocessor

Barry B. Brey
Arithmetic Instructions
• Addition, subtraction, multiplication, a special
multiplication with an addition, and others.
• PADD and PSUB instructions add or subtract
packed signed or unsigned packed bytes,
packed words, or packed doubleword data.
– the add instructions are appended by B, W, or D to
select size, as in PADDB for a byte, PADDW
for a word, and PADDD for a doubleword
• The same is true for the PSUB instruction.

Barry B. Brey
• PMULHW & PMULLW perform multiplication
on four pairs of 16-bit operands, producing
32-bit results.
– PMULHW multiplies the high-order 16 bits
– PMULLW multiplies the low-order 16 bits
• PMADDWD multiplies and adds.
– after multiplying, the four 32-bit results are
added to produce two 32-bit doubleword results
• The MMX instructions use operands just as
the integer or floating-point instructions do.

Barry B. Brey
• There are two comparison instructions:
– PCMPEQ (equal) & PCMPGT (greater than)
• There are three versions of each instruction:
– PCMPEQUB (compares bytes)
– PCMPEQUW (compares words)
– PCMPEQUD (compares doublewords)
• These instructions do not change processor
flag bits
– the result is all ones for a true condition
– all zeros for a false condition.
Barry B. Brey
Conversion Instructions
• Two basic conversion instructions:
• PACK, available as PACKSS (signed
saturation) and PACKUS (unsigned saturation).
• PUNPCK, available as PUNPCKH (unpack
high data) and PUNPCKL (unpack low data)
• Appended with B, W, or D for byte, word, and
doubleword pack and unpack.
– but must be used in combinations WB (word to
byte) or DW (doubleword to word)

Barry B. Brey
Logic Instructions
• PAND (AND)
• PANDN (NAND)
• POR (OR)
• PXOR (Exclusive-OR).
• These do not have size extensions & perform
bit-wise operations on all 64 bits of the data.

Barry B. Brey
Shift Instruction
• Logical shifts and arithmetic shift right
instructions.
• Logic shifts are PSLL (left) and PSRL (right).
– variations are word (W), doubleword (D), and
quadword (Q)
• The PSRA (arithmetic right shift) instruction
functions in the same manner as the logical
shifts.
– except the sign-bit is preserved

Barry B. Brey
Data Transfer Instructions
• MOVED and MOVEQ.
– allow transfers between registers and
between a register and memory
• MOVED transfers 32 bits of data between an
integer register or memory location and an
MMX register.
• MOVEQ copies all 64 bits of an MMX register
between memory or another MMX register.
– MOVEQ MM2, MM3 transfers all 64 bits of MM3
into MM2
Barry B. Brey
EMMS Instruction
• EMMS (empty MMX-state) sets (11) all the
tags in the floating-point unit, so the floating-
point registers are listed as empty.
• Must be executed before the return instruction
at the end of any MMX procedure.
– or a subsequent floating-point operation will
cause a floating-point interrupt error, crashing
Windows or any other application
• All other MMX instructions clear the tags,
indicating all floating-point registers are in use.
Barry B. Brey
Instruction Listing
• Table 14–10 lists all MMX instructions with
the machine code so they can be used with
the assembler.
• MASM does not support these instructions
unless you have the latest version (6.15).
• Any MMX instruction can be used inside
Visual C++ using the inline assembler.

Barry B. Brey
Programming Example
• Example 14–13 uses the MMAX instructions
to perform a task that takes eight times longer
using normal microprocessor instruction.
– an array of 1000 bytes of data (BLOCKA) is
added to a second array of 1000 bytes (BLOCKB)
– result is stored in a third array called BLOCKC
• Example 14–13(a) uses traditional assembly
language to perform the addition.
• Example 14–13(b) shows the same process
using MMX instructions.
Barry B. Brey
14–6 INTRO TO SSE TECHNOLOGY
• The latest type added to the P4 instruction set
of is SIMD (single-instruction multiple data).
• MMX instructions function with integers
– SIMD instructions function with floating-point
numbers as well as integers
• First appeared in Pentium III as SSE
(streaming SIMD extensions) instructions.
– later SSE 2 instructions were added to P4
– SSE 3 extensions are found in the 90
nanometer P4 and Core2
Barry B. Brey
• MMX instructions shared registers with the
arithmetic coprocessor.
– SSE instructions use a register array called
XMM registers (XMM0–XMM7)
• To accommodate the 128-bit-wide data size, a
new keyword, OWORD is added
– octalword designates a 128-bit variable, as
in OWORD PTR for the SSE instruction set
• A double quadword is also used at times to
specify a 128-bit number.
• See Fig 14–13.
Barry B. Brey
Figure 14–13 The XMM registers used by the SSE instructions.

Barry B. Brey
• Figure 14–14 illustrates the multiple data
types that can appear in any XMM register
for various SSE instructions.
– an XMM register can hold four single-precision
floating-point numbers or two double-precision
floating-point numbers
– XMM registers can also hold sixteen 8-bit,
eight 16-bit, four 32-bit , or two 64-bit integers
• This is a twofold increase in the capacity of
and execution speeds of integers compared
to the integers contained in MMX registers.

Barry B. Brey
Figure 14–14 Data formats for the SSE 2 and SSE 3 instructions.

Barry B. Brey
Floating-Point Data
• Floating point data are operated upon as
packed or scalar, and single- or double-
precision.
– packed operation is performed on all sections
at a time
– scalar only on the rightmost section of the
register contents
• Figure 14–15 shows both the packed and
scalar operations on SSE data in XMM
registers.
Barry B. Brey
Figure 14–15 Packed (a) and scalar (b) operations for single-precision floating-point
numbers.

Barry B. Brey
The Instruction Set
• SSE instructions have a few new types added.
– the floating-point unit does not have a reciprocal
instruction, used to solve complex equations
• The reciprocal instruction (1/n) now appears
in the SSE extensions as the RCP instruction.
– written as RCPPS, RCPSS, RCPPD & RCPSD
• The reciprocal of a square root instruction,
called RSQRT is written as:
– RSQRTPS, RSQRTSS, RSQRTPD, & RSQRTSD

Barry B. Brey
The Control/Status Register
• Accessed in the SSE unit as MXCSR, shown
in Fig 14–16.
• This register sets precision and rounding
modes for the coprocessor.
– as does the control register for the arithmetic
coprocessor
– provides information about SSE operation
• Loaded using LDMXCSR and FXRSTOR.
• Stored using STMXCSR and FXSAVE.

Barry B. Brey
Figure 14–16 The MXCSR (control/status) register of the SSE unit.

Barry B. Brey
Programming Examples
• Suppose capacitive reactance is needed for
a circuit that contains a 1.0 µF capacitor at
frequencies from 100 Hz to 10,000 Hz in 100
Hz steps.
• The equation used to calculate capacitive
reactance is:

Barry B. Brey
• Example 14–15 illustrates a procedure that
generates the 100 outcomes for this equation
using the SSE unit and single-precision
floating-point data.
– 14–15(a) uses the SSE unit to perform four
calculations per iteration
– 14–15(b) uses the floating-point coprocessor
to calculate XC one at a time
– 14–15(c) is an example in C++

Barry B. Brey
• Example 14–16 uses integer operation to
add BlockA to BlockB and store the sum in
BlockC.
– each block contains 4000 eight-bit numbers
• Both generate 4000 sums
– the second example using SSE unit does it by
looping 250 times, while first example takes 4000
• SSE uses the same commands as MMX
– except the registers are different
– MMX uses 64-bit-wide registers
– SSE 128-bit-wide XMM registers.
Barry B. Brey
Optimization
• The compiler in Visual C++ does have
optimization for the SSE unit, but does not
optimize the examples presented here.
• It will attempt to optimize a single equation in
a statement if the SSE unit can be utilized for
the equation.
• Until a compiler and extensions are developed,
programs that require high speeds will require
hand-coded assembly language for
optimization.
Barry B. Brey
SUMMARY
• The arithmetic coprocessor functions in
parallel with the microprocessor.
• This means that the microprocessor and
coprocessor can execute their respective
instructions simultaneously.
• The data types manipulated by the
coprocessor include signed integer,
floating-point, and binary-coded decimal
(BCD).

Barry B. Brey
SUMMARY (cont.)
• Three forms of integers are used with the
coprocessor: word (16 bits), short (32 bits),
and long (64 bits).
• Each integer contains a signed number in
true magnitude for positive numbers and
two's complement form for negative
numbers.

Barry B. Brey
SUMMARY (cont.)
• A BCD number is stored as an 18-digit
number in 10 bytes of memory.
• The most significant byte contains the sign-
bit, and the remaining nine bytes contain an
18-digit packed BCD number.
• The coprocessor supports three types of
floating-point numbers: single-precision (32
bits), double-precision (64 bits), and
temporary extended-precision (80 bits).

Barry B. Brey
SUMMARY (cont.)
• A floating-point number has three parts: the
sign, biased exponent, and significant.
• In the coprocessor, the exponent is biased
with a constant and the integer bit of the
normalized number is not stored in the
significant, except in the temporary
extended-precision form.

Barry B. Brey
SUMMARY (cont.)
• Decimal numbers are converted to floating-
point numbers by (a) converting the number
to binary, (b) normalizing the binary number,
(c) adding the bias to the exponent, and (d)
storing the number in floating-point form.
• Floating-point numbers are converted to
decimal by (a) subtracting the bias from the
exponent, (b) un-normalizing the number,
and (c) converting it to decimal.

Barry B. Brey
SUMMARY (cont.)
• The 80287 uses I/O space for the execution
of some of its instructions. This space is
invisible to the program and is used
internally by the 80286/80287 sys-tem.
• These 16-bit I/O addresses (00F8H-00FFH)
must not be used for I/O data transfers in a
system that contains an 80287.
• The 80387, 80486/7, and Pentium through
Core2 use I/O addresses 800000F8H-
800000FFH.
Barry B. Brey
SUMMARY (cont.)
• The coprocessor contains a status register
that indicates busy, the conditions that
follow a compare or test, the location of the
top of the stack, and the state of the error
bits.
• The FSTSW AX instruction, followed by
SAHF, is often used with conditional jump
instructions to test for some coprocessor
conditions.

Barry B. Brey
SUMMARY (cont.)
• The control register of the coprocessor
contains control bits that select infinity,
rounding, precision, and error masks.
• The following directives are often used with
the coprocessor for storing data: DW
(defineword), DD (define doubleword), DQ
(define quadword), and DT (de-fine 10
bytes).

Barry B. Brey
SUMMARY (cont.)
• The coprocessor uses a stack to transfer
data between itself and the memory.
• Generally, data are loaded to the top of the
stack or removed from the top of the stack
for storage.
• All internal coprocessor data are always in
the 80-bit extended-precision form.
• The only time data are in any other form is
when stored or loaded from the memory.

Barry B. Brey
SUMMARY (cont.)
• The coprocessor addressing modes include
the classic stack mode, register, register
with apop, and memory.
• Stack addressing is implied. The data at ST
become the source, at ST(1) the destination,
and the result is found in ST after a pop.
• The coprocessor's arithmetic operations
include addition, subtraction, multiplication,
division, and square root calculation.

Barry B. Brey
SUMMARY (cont.)
• There are transcendental functions in the
coprocessor's instruction set.
• These functions find the partial tangent or
arctangent, 2X - 1, Y log2 X, and Y log2 (X
+ 1). 80387, 80486/7, and Pentium-Core2
also include sine and cosine functions.
• Constants are stored inside the
coprocessor that provide +0.0, +1.0, π,
log2 10, log2 ε, log2 2, and logε 2.
Barry B. Brey
SUMMARY (cont.)
• The 80387 functions with the 80386
microprocessor and the 80487SX functions
with the 80486SX microprocessor, but the
80486DX and Pentium-Core2 contain their
own internal arithmetic coprocessor.
• The instructions performed by the earlier
versions are available on these
coprocessors.
• 80387, 80486/7, and Pentium-Core2 also
can find the sine and cosine.
Barry B. Brey
SUMMARY (cont.)
• The Pentium Pro through Core2 contain
two new floating-point instructions: FCMOV
and FCOMI.
• The FCMOV instruction is a conditional
move and the FCOMI performs the same
task as FCOM, but it also places the
floating-point flags into the system flag
register.

Barry B. Brey
SUMMARY (cont.)
• The MMX extension uses the arithmetic
coprocessor registers for MM0-MM7.
• Therefore, it is important that coprocessor
software and MMX software do not try to
use them at the same time.
• The instructions for the MMX extensions
perform arithmetic and logic operations on
bytes (eight at a time), words (four at a
time), doublewords (two at a time), and
quadwords.
Barry B. Brey
SUMMARY (cont.)
• The operations performed are addition,
subtraction, multiplication, division, AND,
OR, Exclusive-OR, and NAND.
• Both the MMX unit and the SSE unit
employ SIMD techniques to perform
parallel operations on multiple data
with a single instruction.

Barry B. Brey
SUMMARY (cont.)
• The SSE unit performs operations on
integers and floating-point numbers.
• The registers in the SSE unit are 128 bits in
width and can hold (SSE 2 or newer) 16
bytes at a time or four single-precision
floating-point numbers.
• The SSE unit contains registers XMM0-
XMM7.

Barry B. Brey
SUMMARY
• New applications written for the Pentium 4
should contain SSE instructions in place of
MMX instructions.
• The OWORD pointer has been added to
address 128-bit-wide numbers, which are
referred to as octal words or double
quadwords.

Barry B. Brey

Chapter14 - THE ARITHMETIC COPROCESSOR, MMX, AND SIMD TECHNOLOGIES

Uploaded by

Copyright:

Available Formats

Chapter14 - THE ARITHMETIC COPROCESSOR, MMX, AND SIMD TECHNOLOGIES

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Chapter14 - THE ARITHMETIC COPROCESSOR, MMX, AND SIMD TECHNOLOGIES

Uploaded by

Copyright:

Available Formats

Chapter 14: The Arithmetic Coprocessor, MMX, and

The Intel Microprocessors: 8086/8088, 80186/80188, 80286, 80386, 80486 Pentium,

• Convert between decimal data and signed

The Intel Microprocessors: 8086/8088, 80186/80188, 80286, 80386, 80486 Pentium,

• Explain the operation and addressing modes

The Intel Microprocessors: 8086/8088, 80186/80188, 80286, 80386, 80486 Pentium,

The Intel Microprocessors: 8086/8088, 80186/80188, 80286, 80386, 80486 Pentium,

The Intel Microprocessors: 8086/8088, 80186/80188, 80286, 80386, 80486 Pentium,

The Intel Microprocessors: 8086/8088, 80186/80188, 80286, 80386, 80486 Pentium,

The Intel Microprocessors: 8086/8088, 80186/80188, 80286, 80386, 80486 Pentium,

– note both positive and negative numbers

The Intel Microprocessors: 8086/8088, 80186/80188, 80286, 80386, 80486 Pentium,

The Intel Microprocessors: 8086/8088, 80186/80188, 80286, 80386, 80486 Pentium,

The Intel Microprocessors: 8086/8088, 80186/80188, 80286, 80386, 80486 Pentium,

The Intel Microprocessors: 8086/8088, 80186/80188, 80286, 80386, 80486 Pentium,

The Intel Microprocessors: 8086/8088, 80186/80188, 80286, 80386, 80486 Pentium,

The Intel Microprocessors: 8086/8088, 80186/80188, 80286, 80386, 80486 Pentium,

The Intel Microprocessors: 8086/8088, 80186/80188, 80286, 80386, 80486 Pentium,

The Intel Microprocessors: 8086/8088, 80186/80188, 80286, 80386, 80486 Pentium,

The Intel Microprocessors: 8086/8088, 80186/80188, 80286, 80386, 80486 Pentium,

– coprocessor/microprocessor communications are

The Intel Microprocessors: 8086/8088, 80186/80188, 80286, 80386, 80486 Pentium,

The Intel Microprocessors: 8086/8088, 80186/80188, 80286, 80386, 80486 Pentium,

The Intel Microprocessors: 8086/8088, 80186/80188, 80286, 80386, 80486 Pentium,

The Intel Microprocessors: 8086/8088, 80186/80188, 80286, 80386, 80486 Pentium,

The Intel Microprocessors: 8086/8088, 80186/80188, 80286, 80386, 80486 Pentium,

The Intel Microprocessors: 8086/8088, 80186/80188, 80286, 80386, 80486 Pentium,

The Intel Microprocessors: 8086/8088, 80186/80188, 80286, 80386, 80486 Pentium,

The Intel Microprocessors: 8086/8088, 80186/80188, 80286, 80386, 80486 Pentium,

– a program views the tag register by storing the

The Intel Microprocessors: 8086/8088, 80186/80188, 80286, 80386, 80486 Pentium,

The Intel Microprocessors: 8086/8088, 80186/80188, 80286, 80386, 80486 Pentium,

The Intel Microprocessors: 8086/8088, 80186/80188, 80286, 80386, 80486 Pentium,

The Intel Microprocessors: 8086/8088, 80186/80188, 80286, 80386, 80486 Pentium,

The Intel Microprocessors: 8086/8088, 80186/80188, 80286, 80386, 80486 Pentium,

The Intel Microprocessors: 8086/8088, 80186/80188, 80286, 80386, 80486 Pentium,

The Intel Microprocessors: 8086/8088, 80186/80188, 80286, 80386, 80486 Pentium,

The Intel Microprocessors: 8086/8088, 80186/80188, 80286, 80386, 80486 Pentium,

The Intel Microprocessors: 8086/8088, 80186/80188, 80286, 80386, 80486 Pentium,

The Intel Microprocessors: 8086/8088, 80186/80188, 80286, 80386, 80486 Pentium,

The Intel Microprocessors: 8086/8088, 80186/80188, 80286, 80386, 80486 Pentium,

The Intel Microprocessors: 8086/8088, 80186/80188, 80286, 80386, 80486 Pentium,

The Intel Microprocessors: 8086/8088, 80186/80188, 80286, 80386, 80486 Pentium,

The Intel Microprocessors: 8086/8088, 80186/80188, 80286, 80386, 80486 Pentium,

The Intel Microprocessors: 8086/8088, 80186/80188, 80286, 80386, 80486 Pentium,

The Intel Microprocessors: 8086/8088, 80186/80188, 80286, 80386, 80486 Pentium,

The Intel Microprocessors: 8086/8088, 80186/80188, 80286, 80386, 80486 Pentium,

The Intel Microprocessors: 8086/8088, 80186/80188, 80286, 80386, 80486 Pentium,

The Intel Microprocessors: 8086/8088, 80186/80188, 80286, 80386, 80486 Pentium,

The Intel Microprocessors: 8086/8088, 80186/80188, 80286, 80386, 80486 Pentium,

The Intel Microprocessors: 8086/8088, 80186/80188, 80286, 80386, 80486 Pentium,

The Intel Microprocessors: 8086/8088, 80186/80188, 80286, 80386, 80486 Pentium,

The Intel Microprocessors: 8086/8088, 80186/80188, 80286, 80386, 80486 Pentium,

The Intel Microprocessors: 8086/8088, 80186/80188, 80286, 80386, 80486 Pentium,

The Intel Microprocessors: 8086/8088, 80186/80188, 80286, 80386, 80486 Pentium,

The Intel Microprocessors: 8086/8088, 80186/80188, 80286, 80386, 80486 Pentium,

The Intel Microprocessors: 8086/8088, 80186/80188, 80286, 80386, 80486 Pentium,

• Example 14–10 illustrates a program that finds

The Intel Microprocessors: 8086/8088, 80186/80188, 80286, 80386, 80486 Pentium,