Skip to content
Claudiu Zissulescu edited this page Nov 29, 2019 · 2 revisions

Impact of 64-bit integral operation on GCC toolchain

Intro

Some of the new 64 bit integral operations made available for ARCv2HS can be used to map the C-type long long. These are:

Operations Hardware option Possible compiler usage
LDD/STD LL64_OPTION Load/store 64 bit data type
Chained MPYM/MPYMU MPY_OPTION_{5,6} Implementation of 32x32->64 bit ops
MAC/MACU MPY_OPTION_7 Multiply and accumulate operations
MACD/MACDU MPY_OPTION_8+ Multiply and accumulate operations
MPYD/MPYDU MPY_OPTION_8+ Implementation of 32x32->64 bit ops.
VADD2 MPY_OPTION_9 Implement register to register move of a 64 bit data type

64-bit move operations

First step in efficiently supporting the long long data type is implementing an efficient way to move the 64 bit data type in and out register file as well as within register file. The LL64_OPTION provides us with the means for fast transfer of 64 bit data into a processor register pair. The LDD/STD can be used as well to implement a fast way to save/restore the registers in prologue/epilogue of a function.

The MPY_OPTION_9 also gives us means to move a register to another register or a 32-bit immediate into a 64 bit register. The 32-bit immediate is signed extended to match the 64 bit container. Hence, for a register to register move, we can use the following instruction:

VADD2	r0r1,r2r3,0

The above instruction takes 32 bits in the program memory as it uses the VADD2 A,B,u6 encoding. Although VADD2 supports predication, we cannot use it for register to register move due to ISA limitations (e.g., the source of the operands needs to be the input argument vadd2 .cc b,b,u6) If we want to move and sign extend a 32-bit immediate into a 64-bit register pair, we can use the following instruction:

VADD2	r0r1, 0xAFEF, 0

The above instruction takes 64 bits in the program memory as we use VADD2 A,limm,u6 encoding.

Multiplication Instructions

The implementation of multiplication instructions depends on the multiplier option used. A special care should be taken for chained operation when MPY_OPTION is either 5 or 6. In these configurations, the multiplier is blocking sequential, hence, the chained option improves the multiplication result. This, however, may be relevant for EM series as the HS will employ a fully pipelined multiplier.

In general, for 32x32bit -> 64 bit type of multiplier, we use the {mpy,mpym } instructions pair. However, when using MPY_OPTION larger than 7, we can make use of the MPYD/MPYDU instructions. These instructions are faster and are having a smaller impact on memory size than previous used solution. Please remark that the MPYD/MPYDU clobbers also the 64-bit accumulator register (ACCH,ACCL).

Multiply and Accumulate instructions

The ISAv2, provides a number of MAC operations. These are MAC/MACU for MPY_OPTION equals to 7, and additionally MACD/MACDU when using MPY_OPTION eight or more. The latter ones are interesting as they place the 64 bit result in a register pair. All the MAC operations are using the 64-bit accumulator register (ACCH,ACCL) to accumulate with, as well to place the result mac into.

Using a MAC operation needs to set up the accumulator register, as well as collecting the result from the accumulator and place it into a general purpose register. Hence,

Used instructions Single MAC (instructions) Multiple MACs, unroll case Throughput
MAC/MACU 4 (2 loads into ACCH,ACCL; 1 MAC; 1 move from ACCH to register) 4 + 1 for each unrolled MAC (2 to initialize ACCH,ACCL; 2 to move the accumulator) 3+ (output/anti-dependency on ACC), 1 (otherwise)
MACD/MACDU 3 (2 loads to ACCH,ACCL; 1 MAC) 2 + 1 for each unrolled MAC 3+ (output/anti-dependency on ACC), 1 (otherwise)
ADD/MPYD 3 ( 2 additions; 1 MPYD) 3 ops for each MAC 3
ADD/MPY 4 (2 additions; 2 multiplications) 4 ops for each MAC 4

Caveats

  • Having the implicit 64-bit accumulator as destination for MPYD/MPYDU operations complicate the generated code when we have an anti-dependency with a MAC operation on the accumulator register.
  • The accumulator register is used as input as well as output for the MAC operation, hence, using them in a pipelined fashion may be difficult (if, for example, between mac operations exist an output/anti-dependency). In this case, it is faster to use an implementation with ADD/MPYD operations.

Case study

Let us consider the following C-program:

long long foo (long long a, int b, int c)
{
  a += (long long )c * (long long )b;
  return a;
}
  • Using ADD/MPY
mpym     r5,r3,r2
mpy      r4,r3,r2
add.f    r0,r0,r4
adc      r1,r1,r5
  • Using ADD/MPYD
mpyd    r2,r3,r2
add.f   r0,r2,r0
adc     r1,r3,r1
  • Using MAC
mov       ACCL,r0
mov       ACCH,r1
mac       r0,r2,r3
mov       r1,ACCH
  • Using MACD (option 8)
mov       ACCL,r0
mov       ACCH,r1
macd      r0,r2,r3
  • Using MACD (oprion 9)
vadd2   ACC,r0,0
macd    r0,r2,r3

Implementation matrix used by GCC

Due to the accumulator caveats, I propose the following implementation matrix for MAC ops:

MPY_OPTION 2 3 4 5 6 7 8 9
ADD/MPY Y Y Y Y Y Y N N
ADD/MPYD N N N N N N Y N
MAC N N N N N N N N
MACD N N N N N N N Y