Core DSP operation

In many DSP algorithms, the Sum of Product or Multiply-Accumulate (MAC) operations are very common. A DSP CPU is designed to handle the math-intensive calculations necessary for DSP algorithms. For efficient implementation of the MAC operations, the C6211 CPU has two multipliers and each of them can perform a 16-bit multiplication in each clock cycle. For example, if we want to compute the dot product of two length-40 vectors a n a n and x n x n , we need to compute ∑n=140 a n x n n 1 40 a n x n . (For example, the FIR filtering algorithm is exactly same as this dot product operation.) When a n a n and x n x n are stored in memory, starting from n=1 n 1 , we need to compute a n x n a n x n and add it to yy (yy is initially 00) and repeat this up to n=40 n 40 . In the C62x assembly, this MAC operation can be written as

	    
	    MPY .M a,x,prod
	    ADD .L y,prod,y
	    
	  

Register Files

Where are the numbers stored in the CPU? In C62x, the numbers used in operations are stored in the registers. Because the registers are directly accessible through the data bus of the CPU, accessing the registers are much faster than accessing data in the external memory.

The C62x CPU has two register files consisting of sixteen 32-bit registers each. There are two separate register files (A and B). Each of these files contains sixteen 32-bit registers (A0-A15 for file A and B0-B15 for file B). The general-purpose registers can be used for data, data address pointers, or condition registers.

The general-purpose register files support data ranging in size from 16-bit data through 40-bit fixed-point. Values larger than 32 bits, such as 40-bit long quantities, are stored in register pairs. In a register pair, the 32 LSBs of data are placed in an even-numbered register and the remaining 8 MSBs in the next upper register (which is always an odd-numbered register). In assembly language syntax, a colon between two register names denotes the register pairs, and the odd-numbered register is specified first. For example, A1:A0 represents the register pair consisting of A0 and A1. But you don't need to be concerned with the 40-bit numbers too much. Throughout this course, you will be mostly handling either 16 or 32-bit values stored in a single register. Let's for now focus on file A only. The registers in the register file A are named A0 to A15. Each register can store a 32-bit binary number. The numbers such as a,x,prod,y above are stored in these registers. For example, register A0 stores a. For now, let's assume we interpret all 32-bit numbers stored in registers as unsigned integer. Therefore, the range of values we can represent is 0 to 232-1 2 32 1 . (For representation of real numbers using binary bits, we will learn about the Q format numbers for fixed-point representation of real numbers.) Let's assume the numbers a,x,prod,y are in the registers A0,A1,A3,A4, respectively. Then, the above assembly instructions can be written specifically

	    
	    MPY .M1 A0,A1,A3
	    ADD .L1 A4,A3,A4
	    
	  

The TI C62x CPU has a load/store architecture. This means that all the numbers must be stored in the registers for being used as operands for the operations for instructions such as MPY and ADD. The numbers can be read from a memory location to a register (using, for example, LDW, LDB instructions) or a register can be loaded with a constant value. The content of a register can be stored to a memory location (using, for example, STW, STB instructions).

In addition to the general-purpose register files, the CPU has a separate register file for the control registers. The control registers are used to control various CPU functions such as addressing mode, interrupts, etc. You will learn more about some of the control registers when we learn each individual topic.

Functional units

Then, where do the actual operations such as multiplication and addition take place? The C62x CPU has several functional units that perform the actual operations. Each register file has 4 functional units named .M, .L, .S, and .D. (See Figure 1-1). The 4 functional units connected to the register file A are named .L1, .S1, .D1, and .M1. Those connected to the register file B are named .L2, .S2, .D2, and .M2. See Figure 1-1. For example, the functional unit .M1 performs multiplication on the operands that are in register file A. When the CPU executes the MPY .M1 A0,A1,A3 above, the functional unit .M1 takes the values stored in A0 and A1, multiply them together and stores the result to A3. The .M1 in MPY .M1 A0,A1,A3 indicates that this operation is performed in the .M1 unit. The .M1 unit has a 16 bit multiplier and all the multiplications are performed by the .M1 unit.

Similarly, the ADD operation can be executed by the .L1 unit. The .L1 can perform all the logical operations such as bitwise AND operation (AND instruction) as well as basic addition (ADD instruction) and subtraction (SUB instruction).

For complete list of instructions executed by each function unit, see Table 3-2 in the handout TMS320C62x/C64x/C67x Fixed-Point Instruction Set. We will later learn more about assigning the functional units for assembly instructions.

A0*(A1+A2)+A3