DSP execution time
- Thread starter ob1
- Start date
antime
Extra Hard Mid Boss
The SOA#8 technical note claims that a multiplication is performed in one cycle, so I would assume that all other instructions do as well. The patent on the DMA mechanism states that if the DSP program tries to access the same memory while it's being written to by DMA, the program will be paused. It's not immediately clear at least to me if this covers the whole memory or just a single memory location.
I haven't opened Saturn Orbit yet (maybe my boss won't like it). Nevertheless, the SCU is clear enough. You write your assembly program let's say, on paper, then, with the doc, you convert it yourself in machine language. You get some 32-bits words. This is the program you're willing to send to the DSP Program Data Port.
Yesterday, I started implementing IEE754 single precision on it. I already can extract sign, exponent and mantissa from my numbers. IEE754 won't be very useful on the Saturn, but it's a pretty good exercise.
Yesterday, I started implementing IEE754 single precision on it. I already can extract sign, exponent and mantissa from my numbers. IEE754 won't be very useful on the Saturn, but it's a pretty good exercise.
According to the SCU manual, it runs at half the SH2 speed. It also states that each step takes about 70 ns. According to Charles MacDonald's Saturn hardware notes, the SCU is affected by the selectable clock speed that's used for the SH2's and VDP's. So basically the SCU runs at either 13Mhz or 14Mhz depending on the setting.
OH !!! Got it !!!
I won't forget you. As soon as I find a little bit of time, I'll restart working on my assembler. It will be in Java if you don't mind. Well, maybe in C. Don't know really yet.
You guy make me want to dev more !
Even if, OMG, it seems to me that the Saturn is gonna be a hell to code, compared to the Genny !!!
CU
Olivier
I won't forget you. As soon as I find a little bit of time, I'll restart working on my assembler. It will be in Java if you don't mind. Well, maybe in C. Don't know really yet.
You guy make me want to dev more !
Even if, OMG, it seems to me that the Saturn is gonna be a hell to code, compared to the Genny !!!
CU
Olivier
From the SOA#8 (sattech.pdf, SEGA Saturn Technical Bulletins & all),
the MUL takes 1 cycle to complete, but the ALU computation is done in the same cycle.
The DSP can handle multiple operations simultaneously. They just have to be Operation commands and concern ALU or any different bus.
I can have an ALU computation, while a D1 bus access, while a X access, while a Y access ! 4 operations meanwhile, talk about parallelism ! All I have to do is to ensure that they are operation commands, that they do not concern the same bus and to write these operations on the same line :
In the same cycle, M0 is loaded into X register, and M1 is loaded into Y register. Meanwhile, the MUL is computing X * Y.
On the next step, MUL has finished and you can do
P = X * Y, then you can use P.
With, the ALU, you can do even faster :
In just a single cycle, you have done M2[CT2++] = M0 + M1
the MUL takes 1 cycle to complete, but the ALU computation is done in the same cycle.
The DSP can handle multiple operations simultaneously. They just have to be Operation commands and concern ALU or any different bus.
I can have an ALU computation, while a D1 bus access, while a X access, while a Y access ! 4 operations meanwhile, talk about parallelism ! All I have to do is to ensure that they are operation commands, that they do not concern the same bus and to write these operations on the same line :
Code:
MOV M0,X MOV M1,YIn the same cycle, M0 is loaded into X register, and M1 is loaded into Y register. Meanwhile, the MUL is computing X * Y.
On the next step, MUL has finished and you can do
Code:
MOV MUL,PP = X * Y, then you can use P.
With, the ALU, you can do even faster :
Code:
MOV M0,P MOV M1,A
ADD MOV ALU,A MOV ALL,MC2In just a single cycle, you have done M2[CT2++] = M0 + M1
(like every Sega doc translated from japanese to english Similar threads
