ESE 345 Computer Architecture Project

 

Pipelined SIMD multimedia unit design with the VHDL/Verilog hardware description language

 

 

1        Introduction

Purpose: To learn a use of VHDL/Verilog hardware description language and modern CAD tools for the structural and behavioral design of a four-stage pipelined multimedia unit with a reduced set of multimedia instructions similar to those in the Sony Cell SPU and Intel SSE architectures.

 

CAD Tools: Mentor Graphics Modelsim at the Undergraduate CAD Lab (room 281 Light Eng. Bldg.) or any other VHDL/Verilog simulator (e.g. Aldec, Vivado).

 

It is a one/two-students project.

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

2       Procedure

1.      It is suggested to read Chapter 3.6-3.8 on subword parallelism to understand the concept of multimedia processing introduced as the MMX architecture for Intel processors in the 1990s.

2.      Refresh your knowledge of VHDL/Verilog in the HDL design of digital circuits by reading Chapter 4.14 (Verilog) and this VHDL tutorial.

3.   Part 1. Develop and submit behavioral HDL code and its verification results for all multimedia ALU operations at the 3^rd stage. (No knowledge of pipelining & forwarding is expected/used at that step.

4.    Develop the HDL model of the four-stage multimedia unit and its modules. As an example, look how the Verilog code is used to describe the operation of the 5-stage MIPS pipeline.

5.      Verify individual modules of your design with their testbenches before instantiating them in higher order modules. Verify the final model with a testbench module and generate file Results showing the status of each stage of the unit during execution.

 

3       Requirements

The complete 4-stage pipelined design is to be developed in a structural/RTL manner with several modules operating simultaneously. Each module represents a pipelined stage with its interstage register. The major units inside those stages modules are described below.

1.     Multimedia ALU

Takes up to three inputs from the Register File, and calculates the result based on the current instruction to be performed.

The ALU must be implemented as behavioral model in VHDL or continuous assignment (dataflow models in Verilog).

2.     Register File

The register file has 32 128-bit registers. On any cycle, there can be 3 reads and 1 write. When executing instructions, each cycle two/three 128-bit register values are read, and one 128-bit result can be written if a write signal is valid. This register write signal must be explicitly declared so it can be checked during simulation and demonstration of your design. The register module must be implemented as a behavioral model in VHDL (dataflow/RTL model in Verilog).

3.     Instruction Buffer

The instruction buffer can store 64 25-bit instructions. The contents of the buffer should be loaded by the testbench instructions from a test file at the start of simulation. On each cycle, the instruction specified by the Program Counter (PC) is fetched, and the value of PC is incremented by 1.

The Instruction Buffer module must be implemented as a behavioral model in VHDL (dataflow/RTL model in Verilog).

4.     Forwarding Unit

Every instruction must use the most recent value of a register, even if this value has not yet been written to the Register File. Be mindful of the ordering of instructions; the most recent value should be used, in the event of two consecutive writes to a register, followed by a read from that same register. Your processor should never stall in the event of hazards.

Take extra care of which instructions require forwarding, and which ones do not. Namely, NOP and the instructions with Immediate fields do not contain one/two register sources. Only valid data and source/destination registers should be considered for forwarding.

5.     Four-Stage Pipelined Multimedia Unit

Clock edge-sensitive pipeline registers separate the IF, ID, EXE, and WB stages. Data should be written to the Register File after the WB Stage.

All instructions (including li) take four cycles to complete. This pipeline must be implemented as a structural model with modules for each corresponding pipeline stages and their interstage registers. Four instructions can be at different stages of the pipeline at every cycle.

6.      Testbench This module loads the instruction buffer using data loaded from a file, begins simulation, and upon completion, compares the contents of the register file to a file containing the expected results. This expected results file does not need to be auto-generated. Instead, this can be manually entered when designing a test program.

This must be implemented as a behavioral model.

 

7.      Assembler This is a separate program written in any language your team prefers (i.e. Java, C++, Python). Its purpose is to convert an assembly file to the binary format for the Instruction Buffer. This assembler does not need to be robust, and can assume very specific syntax rules that you as a team decide.

8.      Results File This file must show the status of the pipeline for each cycle during program execution. It should include the opcodes, input operand, and results of the execution of instructions, as well as all relevant control signals and forwarding information. This should be carried out by your testbench.

 

 

4       Instruction Formats and Opcode Description

4.1       Load Immediate

 

 

 

li: Load a 16-bit Immediate value from the [20:5] instruction field into the 16-bit field specified by the Load Index field [23:21] of the 128-bit register rd. Other fields of register rd are not changed. Note that a LI instruction first reads register rd and then (after inserting an immediate value into one of its fields) writes it back to register rd, i.e., register rd is both a source and destination register of the LI instruction!

 

4.2      Multiply-Add and Multiply-Subtract R4-Instruction Format

 

 

 

Signed operations are performed with saturated rounding that takes the result, and sets a floor and ceiling corresponding to the max range for that data size. This means that instead of over/underflow wrapping, the max/min values are used.

 

Size (Num Bits)	Min	Max
Long (64)	-263	+263 − 1
Int (32)	-231	+231 − 1

 

The tables below show the description for each operation:

LI/SA/HL [22:20]	Description of Instruction Code
000	Signed Integer Multiply-Add Low with Saturation: Multiply low 16-bit-fields of each 32-bit field of registers rs3 and rs2, then add 32-bit products to 32-bit fields of register rs1, and save result in register rd
001	Signed Integer Multiply-Add High with Saturation: Multiply high 16-bit-fields of each 32-bit field of registers rs3 and rs2, then add 32-bit products to 32-bit fields of register rs1, and save result in register rd
010	Signed Integer Multiply-Subtract Low with Saturation: Multiply low 16-bit-fields of each 32-bit field of registers rs3 and rs2, then subtract 32-bit products from 32-bit fields of register rs1, and save result in register rd
011	Signed Integer Multiply-Subtract High with Saturation: Multiply high 16-bit- fields of each 32-bit field of registers rs3 and rs2, then subtract 32-bit products from 32-bit fields of register rs1, and save result in register rd
100	Signed Long Integer Multiply-Add Low with Saturation: Multiply low 32-bit- fields of each 64-bit field of registers rs3 and rs2, then add 64-bit products to 64-bit fields of register rs1, and save result in register rd
101	Signed Long Integer Multiply-Add High with Saturation: Multiply high 32-bit- fields of each 64-bit field of registers rs3 and rs2, then add 64-bit products to 64-bit fields of register rs1, and save result in register rd
110	Signed Long Integer Multiply-Subtract Low with Saturation: Multiply low 32- bit-fields of each 64-bit field of registers rs3 and rs2, then subtract 64-bit products from 64-bit fields of register rs1, and save result in register rd
111	Signed Long Integer Multiply-Subtract High with Saturation: Multiply high 32- bit-fields of each 64-bit field of registers rs3 and rs2, then subtract 64-bit products from 64-bit fields of register rs1, and save result in register rd

 

4.3      R3-Instruction Format

 

 

 

 

In the table below, 16-bit signed integer add (AHS), and subtract (SFHS) operations are performed with saturation to signed halfword rounding that takes a 16-bit signed integer X, and converts it to -32768 (the most negative 16-bit signed value) if it is less than -32768, to +32767 (the highest positive 16-bit signed value) if it is greater than 32767, and leaves it unchanged otherwise.

 

Opcode [22:15]	Description of Instruction Opcode
xxxx0000	NOP: no operation. Make sure that a NOP instruction does not write anything to the register file!
xxxx0001	ABSDB: absolute difference of bytes: the contents of each of the sixteen byte slots in register rs2 is subtracted from the contents of the corresponding byte slots in register rs1. The absolute value of each of the results is placed into the corresponding byte slot in register rd. (Comments: 16 separate 8-bit values in each 128-bit register)
xxxx0010	AU: add word unsigned: packed 32-bit unsigned addition of the contents of registers rs1 and rs2 (Comments: 4 separate 32-bit values in each 128-bit register)
xxxx0011	CNT1W: count 1s in words           :: count 1s in each packed 32-bit word of the contents of register rs1. The results are placed into corresponding slots in register rd .  (Comments: 4 separate 32-bit values in each 128-bit register)
xxxx0100	AHS: add halfword saturated : packed 16-bit halfword signed addition with saturation of the contents of registers rs1 and rs2 . (Comments: 8 separate 16-bit values in each 128-bit register)
xxxx0101	AND: bitwise logical and of the contents of registers rs1 and rs2
xxxx0110	BCW: broadcast word : broadcast the rightmost 32-bit word of register rs1 to each of the four 32-bit words of register rd
xxxx0111	MAXWS: max signed word: for each of the four 32-bit word slots, place the maximum signed value between rs1 and rs2 in register rd. (Comments: 4 separate 32-bit values in each128-bit register)
xxx01000	MINWS: min signed word: for each of the four 32-bit word slots, place the minimum  signed value between rs1 and rs2 in register rd . (Comments: 4 separate 32-bit values in each 128-bit register)
xxxx1001	MLHU: multiply low unsigned: the 16 rightmost bits of each of the four 32-bit slots in register rs1 are multiplied by the 16 rightmost bits of the corresponding 32-bit slots in register rs2, treating both operands as unsigned. The four 32-bit products are placed into the corresponding slots of register rd . (Comments: 4 separate 32-bit values in each 128-bit register)
xxxx1010	MLHCU: multiply low by constant unsigned: the 16 rightmost bits of each of the four 32-bit slots in register rs1 are multiplied by a 5-bit value in the rs2 field of the instruction, treating both operands as unsigned. The four 32-bit products are placed into the corresponding slots of register rd . (Comments: 4 separate 32-bit values in each 128-bit register)
xxxx1011	OR: bitwise logical or of the contents of registers rs1 and rs2
xxxx1100	CNTLZ: count leading zeroes in words: for each of the four 32-bit word slots in register rs1, count the number of zero bits to the left of the first “1”. If the word slot in register rs1 is zero, the result is 32. The foure results are placed into the corresponding 32-bit word slots in register rd. (Comments: 4 separate 32-bit values in each 128-bit register)
xxxx1101	ROTW: rotate bits in word : the contents of each 32-bit field in register rs1 are rotated to the right according to the value of the 5 least significant bits of the corresponding 32-bit field in register rs2. The results are placed in register rd. Bits rotated out of the right end of each word are rotated in on the left end of the same 32-bit word field. (Comments: 4 separate 32-bit word values in each 128-bit register)
xxxx1110	SFWU: subtract from word unsigned: packed 32-bit word unsigned subtract of the contentsof rs1 from rs2 (rd = rs2 - rs1). (Comments: 4 separate 32-bit values in each 128-bit register)
xxxx1111	SFHS: subtract from halfword saturated: packed 16-bit halfword signed subtraction with saturation of the contents of rs1 from rs2 (rd = rs2 - rs1). (Comments: 8 separate 16-bit values in each 128-bit register)

 

5        Expected Results