Dynamic scheduling

• Tomasulo's approach :
• A technique to allow execution to proceed in the presence of hazards .
• This was first introduced in the IBM 360/91.

 

• Applied only to floating-point operations (including FP loads & stores).

 

• We have already seen that the compiler can rename registers (statically) to avoid WAW and WAR hazards.

 

• Tomasulo's scheme performs this function dynamically.
• It buffers operands of instructions waiting to issue, fetching them as soon as they are available, avoiding the register file.

 

• The register specifiers of instructions are renamed to reservation station numbers as they are issued, eliminating WAW and WAR hazards.

Tomasulo's approach

• Differences between scoreboarding and Tomasulo's approach:
• Register renaming
• Register renaming is used to eliminate WAR and WAW hazards.
• In contrast to scoreboarding which must wait for WAR and WAW hazards to clear.

 

• Distributed control
• Hazard detection and execution control are distributed to each functional unit.
• In contrast to scoreboarding, in which it is centralized.

 

• Common Data Bus
• Is used to forward results directly to the functional units without going through the register file.

Tomasulo's approach

• Reservation stations:
• The reservation stations are the heart of Tomasulo's approach.
• They are located at each functional unit and determine when an instruction can begin execution.

Tomasulo's approach

• Operation steps:
• Issue
• Take an instruction from the FP operation's queue.

 

• If there's a station available for it, send the instruction to the station.
• Otherwise, stall for a structural hazard.

 

• Also, this step checks to see if the source operands will be produced by a current instruction.
• If so, renaming is performed by checking to see if the desired register is being written by an instruction already at a reservation station.

 

• If the value is not being generated by a functional unit, it is fetched from the register file.

 

• If the operation is a load or a store, it can issue if there is an available load or store buffer.

Tomasulo's approach

• Operation steps:
• Execute
• If at least one operand is missing, monitor the CDB until it is generated.

 

• When a needed operand is put out onto the CDB, it is placed into the appropriate reservation station.

 

• When both operands are ready, the operation is executed.

 

• RAW hazards are handled here.

 

• Write result
• When the result is ready, write it on the CDB and into the register file and any waiting reservation station.

 

• Obviously, only one value can be written on the CDB in any single cycle.

 

• Also, indicate that the reservation station is no longer busy.

Tomasulo's approach

• The reservation station fields:
• Operation (Op)
• The operation to be performed.

 

• Operand sources (Q _j , Q _k )
• The reservation stations that will produce the values for the two operands.
• A 0 in either slot means the source operand is already in V _j or V _k , or that the slot is not needed.

 

• Operand values (V _j , V _k )
• The values for the two operands.
• They are valid iff the corresponding Q is 0.

 

• Busy
• Indicates the reservation station and the accompanying functional unit are busy.

Tomasulo's approach

• Other components:
• The register file and store buffer have a field for all registers: Q _i .

 

• This is the number of the reservation station that contains the operation that will eventually write the register/buffer.

 

• If no operation is pending, this value is 0 (blank).

 

• Consider the same code sequence:

Tomasulo's approach

• An example

• The information in the instruction status table is actually distributed in the hardware.

Tomasulo's approach

• Two major advantages:
• The distribution of the hazard detection logic.
• If multiple instructions are waiting on the second of two operands, the instructions can be released simultaneously by a broadcast on the CDB.

 

• The elimination of WAW and WAR hazards is possible because:
• Register renaming is performed using the reservation stations.
• Operands are stored into the reservation tables as soon as they are available.

 

• In the example, the WAR hazard was eliminated because the reservation station held the value of F6 for the DIVD instruction.

 

• If the `LD F6, 34(R2)' had not completed before the DIVD had issued, the WAR hazard (and the possible WAW hazard) is still eliminated since Q _k would point to the Load1 reservation table for the value of F6.

Tomasulo's approach

• Loop unrolling is performed dynamically !

 

• With only 4 FP registers, WAW and WAR hazards would severly limit loop unrolling, even by the compiler.

 

• The virtual registers provided by the reservation stations make it possible to execute multiple iterations of some loops simultaneously (see text for an example).

 

• Memory disambiguation:
• Since the store functional unit keeps a memory address as well as a value, it's possible to do disambiguation.
• When a memory operation is issued, check to see if that location is already involved in an operation.

 

• Therefore, LDs and SDs from different iterations of the loop can be executed non-sequentially .
•