DPC++ daily 2021-12-28

sycl-nightly/20211228

[ESIMD] Add infra to support half, bfloat, etc, support sycl::half. (…

DPC++ daily 2021-12-27

[SYCL][NFC] Optimize getKernelNamesUsingAssert (#5196)

Now it traverses reversed call graph by BFS algorithm from
__devicelib_assert_fail function up to SPIR kernels.

DPC++ daily 2021-12-25

sycl-nightly/20211225

[SYCL][CUDA][MATRIX] Remove `using namespace experimental` from heade…

DPC++ daily 2021-12-24

sycl-nightly/20211224

[BuildBot] Sync testlist.cfg with latest sycl-cts repo (#5201)

DPC++ daily 2021-12-23

sycl-nightly/20211223

[SPIR-V] Translate complex nested vector expressions instead of lower…

DPC++ daily 2021-12-22

sycl-nightly/20211222

[SYCL][XPTI] Report memory allocation info from SYCL runtime (#5172)

DPC++ daily 2021-12-21

[SYCL][NFC] Fix static code analysis concerns (#5189)

Found via a static-analysis tool:
Suspicious dereference of pointer in function call before NULL check

Inside checkAllowedSYCLInitializer() in SemaSYCL.cpp file:

    const Expr *Init = VD->getInit();
    bool ValueDependent = Init->isValueDependent(); --> 'Init' is dereferenced by being passed as argument 0 to function "isValueDependent
    bool isConstantInit =
       Init && !ValueDependent && Init->isConstantInitializer(Context, false);  --> 'Init' is checked for NULL here

This patch adds NULL value checking for 'Init' expression.

Signed-off-by: Soumi Manna <soumi.manna@intel.com>

DPC++ daily 2021-12-20

[CI] Add cache checkout script to docker containers (#5184)

Add a new cache checkout script to docker containers. General idea is to keep a local copy of repository and re-use existing objects to avoid long clone times.

Also add zstd to experiment with LLVM compression times and psutil, which is required to kill tests by timeout.

DPC++ daily 2021-12-19

sycl-nightly/20211219

[SYCL] Widen (u)int8/16 to (u)int32 and half to float in group_broadc…

DPC++ daily 2021-12-18

[LIBCLC] Fix NaN value for doubles (#5173)

A NaN is a floating point value with all exponent bits set to 1 and any
non-zero fraction, and the sign bit can be set or not.

For doubles the floating point is represented as one sign bit, eleven
exponent bits, and the fraction bits, so the value before this patch
breaks down as follows:

```
0x7ff0000000000000

0b0111111111110000000000000000000000000000000000000000000000000000

0b0 11111111111 0000000000000000000000000000000000000000000000000000
```

As you can see this value has all zeroes in the exponent, it is
therefore not a NaN.

Comparing to the value used for single precision, knowing that single
precision follows the same rule but has only 8 bits of exponent:

```
0x7fc00000

0b01111111110000000000000000000000

0b0 11111111 10000000000000000000000
```

As you can see the value used for single precision has all exponent bits
set to one and the most significant bit of the fraction set to one,
therefore it is indeed a NaN.

And so the correct value for the NaN constant for doubles is actually:

```
0b0 11111111111 1000000000000000000000000000000000000000000000000000

0b0111111111111000000000000000000000000000000000000000000000000000

0x7ff8000000000000
```

Which is what this patch is updating the constant to be.

The constant for half precision also correctly follows this pattern.

This fixes the `llvm-test-suite` `nan.cpp` test with the CUDA plugin.