Skip to content
Snippets Groups Projects
Commit e0cde58d authored by John Romein's avatar John Romein
Browse files

Updated for e4m3/e5m2 support.

parent 2da65b31
No related branches found
No related tags found
1 merge request!15Fp8
Pipeline #114275 failed
......@@ -32,36 +32,49 @@ how to use the CUDA driver API (wrappers); and
OpenCL program. `test/CorrelatorTest/CorrelatorTest.cc` is a much more versatile,
robust (and complex) example than `test/SimpleExample/SimpleExample.cu`.
Input and output data types are defined as follows:
The TCC accepts the following input data types:
- half precision floating point (a.k.a. fp16), starting from Volta (sm\_70)
- e4m3 and e5m2 (a.k.a. fp8), starting from Hopper (sm\_90)
- 8-bit integers (i8), starting from the Jetson Xavier (sm\_72)
- 4-bit integers (i4), only natively supported on Ampere and Ada
```
#if NR_BITS == 4
#if INPUT_FORMAT == FORMAT_I4
#define NR_TIMES_PER_BLOCK 32
typedef complex_int4_t Sample;
typedef std::complex<int32_t> Visibility;
#elif NR_BITS == 8
#elif INPUT_FORMAT == FORMAT_I8
#define NR_TIMES_PER_BLOCK 16
typedef std::complex<int8_t> Sample;
typedef std::complex<int32_t> Visibility;
#elif NR_BITS == 16
#elif INPUT_FORMAT == FORMAT_E4M3
#define NR_TIMES_PER_BLOCK 16
typedef std::complex<__nv_fp8_e4m3> Sample;
typedef std::complex<float> Visibility;
#elif INPUT_FORMAT == FORMAT_E5M2
#define NR_TIMES_PER_BLOCK 16
typedef std::complex<__nv_fp8_e5m2> Sample;
typedef std::complex<float> Visibility;
#elif INPUT_FORMAT == FORMAT_FP16
#define NR_TIMES_PER_BLOCK 8
typedef std::complex<__half> Sample;
typedef std::complex<float> Visibility;
#endif
#define NR_TIMES_PER_BLOCK (128 / NR_BITS)
typedef Sample Samples[NR_CHANNELS][NR_SAMPLES_PER_CHANNEL / NR_TIMES_PER_BLOCK][NR_RECEIVERS][NR_POLARIZATIONS][NR_TIMES_PER_BLOCK];
typedef Visibility Visibilities[NR_CHANNELS][NR_BASELINES][NR_POLARIZATIONS][NR_POLARIZATIONS];
```
Note that in 4-bit and 8-bit mode, the input samples may not contain -8 or -128
Note that with FORMAT\_I4 and FORMAT\_I8, the input samples may not contain -8 or -128
respectively, as these values cannot be conjugated properly.
The input data type (`Samples`) is a weird format, but this seemed to be the only
format that yields good performance (tensor cores are very unforgiving).
Limitations:
- `NR_POLARIZATIONS` must be 2
- `NR_BITS` must be 4, 8, or 16
- the amount of samples over which is integrated) must be a multiple of 128 / `NR_BITS`
(i.e., 32, 16, or 8 for 4-bit, 8-bit, or 16-bit input, respectively).
- the amount of samples over which is integrated must be a multiple of
NR\_TIMES\_PER\_BLOCK.
## Building, testing, and installation
Clone the repository:
......
0% Loading or .
You are about to add 0 people to the discussion. Proceed with caution.
Please register or to comment