Fix crashes caused by unaligned memory stores in stream_mem kernel

[fairydreaming]

This PR replaces movntdq instructions in stream_mem kernel with a combination of:

• unpcklpd instructions that unpack scalar double precision floating point values from FPR1 and FPR2 into FPR1 register and values from FPR3 and FPR4 into FPR3 register
• movntpd instructions that store values from FPR1 and FPR3 registers to memory

The original kernel caused segmentation faults due to unmet memory alignment requirement of movntdq (16 bytes) because the memory offsets in the kernel were increased by 8 bytes.

Also corrected the kernel description, INSTR_LOOP and UOPS values. Please check if the numbers are correct.

Fixes #649.

[TomTheBear]

Thx for the PR.

The INSTR_LOOP is correct. You have 16 instructions in the loop plus the loop increment, the compare and the jump instruction. For the UOPS, I'm counting only 22. Each of the instructions is at least one uop. The addsd instructions load from a memory location which adds a uop each. The loop increment is one uop but the compare&jump instructions are merged into one uop. The movnt instructions are one uop on all architectures I checked because we store from an xmm register (AMD Zen partly uses 2 uops if ymm or zmm register).

[fairydreaming]

The INSTR_LOOP is correct. You have 16 instructions in the loop plus the loop increment, the compare and the jump instruction. For the UOPS, I'm counting only 22. Each of the instructions is at least one uop. The addsd instructions load from a memory location which adds a uop each. The loop increment is one uop but the compare&jump instructions are merged into one uop. The movnt instructions are one uop on all architectures I checked because we store from an xmm register (AMD Zen partly uses 2 uops if ymm or zmm register).

OK, I corrected it to 22.