Compiler/PROCESSOR-SDK-DRA8X-TDA4X: DSP vector

Guoxun

Intellectual 670 points

Part Number: PROCESSOR-SDK-DRA8X-TDA4X

Tool/software: TI C/C++ Compiler

About C66x dsp in TDA4x,

there is datatype named float16(__float16 is equivalent to a vector of 16 float) in head file C6x_vec.h,

but function units support up to 128-bit source operands.

How is float16 handled?

Are there some examples about float16?

Is dataType like float16 supported on TDA4?

over 4 years ago

0 George Mock over 4 years ago

TI__Guru**** 249095 points

Guoxun said:
How is float16 handled?

Like a built-in type that contains 16 float values, each of which is 32-bits wide.

Guoxun said:
Are there some examples about float16?

Here is some C code that uses float16 ...

void vector_multiply(float const *x, float const *y, float *z, int len )
{
   int i;
   float16 * restrict v1;
   float16 * restrict v2;
   float16 * restrict v3;

   v1 = (float16*)x;
   v2 = (float16*)y;
   v3 = (float16*)z;

   #pragma MUST_ITERATE(1,,)
   for (i = 0; i < len / 16; i++)
      *v3++ = *v1++ * *v2++;
}

Guoxun said:
Is dataType like float16 supported on TDA4?

Yes.

Thanks and regards,

-George

0 Guoxun over 4 years ago in reply to George Mock

Intellectual 670 points

Hi,

For C66x DSP:

The .M units now support up to 128-bit source operands.

The .L/.S units now support up to 64-bit operands.

"Like a built-in type that contains 16 float values, each of which is 32-bits wide."

so float16 is 32 * 16 = 512 bit, can C66x DSP process 16 floats at a time?

Or, float16 is processed by C71x DSP in TDA4x?

Thanks,

Guoxun

0 George Mock over 4 years ago in reply to Guoxun

TI__Guru**** 249095 points

Please follow this very similar thread. -George

Processors

Processors forum

Compiler/PROCESSOR-SDK-DRA8X-TDA4X: DSP vector