New APM32F103CB

[blue-man]

for the FPU, my guess is one would need to just ask for it (the documentation) from the manufacturer expressing interest.

Referring to my question

i am asking me what would have happened, if you would have explained this dependencies to the manufacturer?

it seems to be my part to do this.
But first i will only ask for further documentation for the FPU.

I will post here if i get an answer ...

[blue-man]

Up to now there is no response from Apexmic.

But in a german forum the basic usage of the FPU have been decoded with disassembling of the library sc_math.lib from the SDK

Code: Select all

Ok, 32-Bit-Float,
5 32-Bit-Register:
0: Opcode
1: Status,  Bit0: Busy
2: Result
3: Op1
4: Op2

Opcodes:
  1: asin
  3: acos
  5: atan
  7: atan2
  9: sin
 11: cos
 13: tan
...
289: fadd
293: fsub
295: fmul
 45: fdiv
...

and the example

Code: Select all

00000000 <sc_math_sin>:
   0:  4601        mov  r1, r0
   2:  2009        movs  r0, #9
   4:  4afe        ldr  r2, [pc, #1016]  ; (400 <__wrap___ltsf2+0x2a>)
   6:  6010        str  r0, [r2, #0]
   8:  4610        mov  r0, r2
   a:  60c1        str  r1, [r0, #12]
   c:  bf00        nop
   e:  48fc        ldr  r0, [pc, #1008]  ; (400 <__wrap___ltsf2+0x2a>)
  10:  6840        ldr  r0, [r0, #4]
  12:  f000 0001   and.w  r0, r0, #1
  16:  2800        cmp  r0, #0
  18:  d0f9        beq.n  e <sc_math_sin+0xe>
  1a:  48f9        ldr  r0, [pc, #996]  ; (400 <__wrap___ltsf2+0x2a>)
  1c:  6880        ldr  r0, [r0, #8]
  1e:  4770        bx  lr

at:

400:  40024000   .word  0x40024000

is the base address of the FPU hardware.

[Pito]

Here is the (Google) translation for page 14:

3.10. Floating Point Unit (FPU)
The product has a built-in independent FPU floating-point arithmetic processing unit, it supports the IEEE754 standard single-precision floating-point arithmetic.

As it is memory-mapped on the AHB bus (0x40024000-0x400243FF), it is probably not the standard ARM FPv4-SP core-integrated that can be found in the M4 or M7 cores.

As a noob I added a Cordic unit to my 32bit MicroBlaze in about 1 hour, therefore it cannot be a big problem for the guys who mastered the APM32F103 to add an FPU unit, where you write 1-2 32bit operands somewhere, set a bit somewhere and in 1-14clocks you read a result from somewhere.

[Squonk42]

Good!

So, just a simple single-precision floating point ALU with Cordic exposed in peripheral address space.

No SIMD vector or MAC operations and no saturation, so almost useless for all DSP usage.

Better than nothing

[blue-man]

It's just another MCU that is STM32 compatible, with this bad documented FPU and QSPI goodie.

[Pito]

In 1.0.2 version of their datasheet they list APM32E103x as the chips with an FPU, in 1.0.5 they list APM32F103x as the chips with an FPU.

The above "sc_math_sin asm" does not show the FPU usage at all, you have to dig deeper into asm to find out the actual sine() calculation. The FPU has got most probably only mul/div/add/sub in its most simple version, the full FPU in 4xx and above has got abs() and sqrt(), plus some fused DSP operations, plus conversions from int/fp and vice versa.
Even an single precision FPU with mul/div/add/sub would help. I would expect 5-7x speed up in an average.

[blue-man]

We have ordered some of them and want to find it out ...

[Pito]

You need libraries for their FPU usage, otherwise you would be lost..

PS: a search through the SDK shows FPU, with one global FPU interrupt, FPU clock enable flag and FPU_DIV flag.

[blue-man]

There is a sc_math.lib in the SDK:

Code: Select all

extern float sc_math_sin(float x);
extern float sc_math_cos(float x);
extern float sc_math_tan(float x);
extern float sc_math_asin(float x);
extern float sc_math_acos(float x);
extern float sc_math_atan(float x);
extern float sc_math_atan2(float y, float x);

extern float sc_math_invsqrt(float x);
extern float sc_math_mac(float x, float y, float z);
extern float sc_math_sum_N(float* x, unsigned char n);
extern float sc_math_sub_N(float* x, unsigned char n);
extern float sc_math_prdct(float* x, unsigned char n);
extern float sc_math_sumsq(float* x, unsigned char n);

[Pito]

That is good.
From my search above the FPU could work with 4 primitives add/sub/mult/div, while add/sub/mult is 1 cycle and div is more cycles (like 14 in F4xx).
The interrupt fires when div finished, imho.