This was related to the source allocator being passed into the
constructor potentially having a different type than allocator being
constructed.
We simply need to provide a constructor to handle this case.
This resolves issues related to the allocator causing debug builds on
MSVC to fail.
Implement VOTE using Nvidia's intrinsics. Documentation about these can
be found here
https://developer.nvidia.com/reading-between-threads-shader-intrinsics
Instead of using portable ARB instructions I opted to use Nvidia
intrinsics because these are the closest we have to how Tegra X1
hardware renders.
To stub VOTE on non-Nvidia drivers (including nouveau) this commit
simulates a GPU with a warp size of one, returning what is meaningful
for the instruction being emulated:
* anyThreadNV(value) -> value
* allThreadsNV(value) -> value
* allThreadsEqualNV(value) -> true
ballotARB, also known as "uint64_t(activeThreadsNV())", emits
VOTE.ANY Rd, PT, PT;
on nouveau's compiler. This doesn't match exactly to Nvidia's code
VOTE.ALL Rd, PT, PT;
Which is emulated with activeThreadsNV() by this commit. In theory this
shouldn't really matter since .ANY, .ALL and .EQ affect the predicates
(set to PT on those cases) and not the registers.
This commit ensures that all backing memory allocated for the Guest CPU
is aligned to 256 bytes. This due to how gpu memory works and the heavy
constraints it has in the alignment of physical memory.
Instead of storing all block width, height and depths in their shifted
form:
block_width = 1U << block_shift;
Store them like they are provided by the emulated hardware (their
block_shift form). This way we can avoid doing the costly
Common::AlignUp operation to align texture sizes and drop CPU integer
divisions with bitwise logic (defined in Common::AlignBits).
Avoids potentially performing multiple reallocations (depending on the
size of the input data) by reserving the necessary amount of memory
ahead of time.
This is trivially doable, so there's no harm in it.
These can be generified together by using a concept type to designate
them. This also has the benefit of not making copies of potentially very
large arrays.
Allows for things such as:
auto rect = Common::Rectangle{0, 0, 0, 0};
as opposed to being required to explicitly write out the underlying
type, such as:
auto rect = Common::Rectangle<int>{0, 0, 0, 0};
The only requirement for the deduction is that all constructor arguments
be the same type.
nullptr was being returned in the error case, which, at a glance may
seem perfectly OK... until you realize that std::string has the
invariant that it may not be constructed from a null pointer. This
means that if this error case was ever hit, then the application would
most likely crash from a thrown exception in std::string's constructor.
Instead, we can change the function to return an optional value,
indicating if a failure occurred.
Makes the parameter ordering consistent, and also makes the filename
parameter a std::string. A std::string would be constructed anyways with
the previous code, as IOFile's only constructor with a filepath is one
taking a std::string.
We can also make WriteStringToFile's string parameter utilize a
std::string_view for the string, making use of our previous changes to
IOFile.
We don't need to force the usage of a std::string here, and can instead
use a std::string_view, which allows writing out other forms of strings
(e.g. C-style strings) without any unnecessary heap allocations.
Uses arithmetic that can be identified more trivially by compilers for
optimizations. e.g. Rather than shifting the halves of the value and
then swapping and combining them, we can swap them in place.
e.g. for the original swap32 code on x86-64, clang 8.0 would generate:
mov ecx, edi
rol cx, 8
shl ecx, 16
shr edi, 16
rol di, 8
movzx eax, di
or eax, ecx
ret
while GCC 8.3 would generate the ideal:
mov eax, edi
bswap eax
ret
now both generate the same optimal output.
MSVC used to generate the following with the old code:
mov eax, ecx
rol cx, 8
shr eax, 16
rol ax, 8
movzx ecx, cx
movzx eax, ax
shl ecx, 16
or eax, ecx
ret 0
Now MSVC also generates a similar, but equally optimal result as clang/GCC:
bswap ecx
mov eax, ecx
ret 0
====
In the swap64 case, for the original code, clang 8.0 would generate:
mov eax, edi
bswap eax
shl rax, 32
shr rdi, 32
bswap edi
or rax, rdi
ret
(almost there, but still missing the mark)
while, again, GCC 8.3 would generate the more ideal:
mov rax, rdi
bswap rax
ret
now clang also generates the optimal sequence for this fallback as well.
This is a case where MSVC unfortunately falls short, despite the new
code, this one still generates a doozy of an output.
mov r8, rcx
mov r9, rcx
mov rax, 71776119061217280
mov rdx, r8
and r9, rax
and edx, 65280
mov rax, rcx
shr rax, 16
or r9, rax
mov rax, rcx
shr r9, 16
mov rcx, 280375465082880
and rax, rcx
mov rcx, 1095216660480
or r9, rax
mov rax, r8
and rax, rcx
shr r9, 16
or r9, rax
mov rcx, r8
mov rax, r8
shr r9, 8
shl rax, 16
and ecx, 16711680
or rdx, rax
mov eax, -16777216
and rax, r8
shl rdx, 16
or rdx, rcx
shl rdx, 16
or rax, rdx
shl rax, 8
or rax, r9
ret 0
which is pretty unfortunate.
Allows the compiler to inform when the result of a swap function is
being ignored (which is 100% a bug in all usage scenarios). We also mark
them noexcept to allow other functions using them to be able to be
marked as noexcept and play nicely with things that potentially inspect
"nothrowability".
Including every OS' own built-in byte swapping functions is kind of
undesirable, since it adds yet another build path to ensure compilation
succeeds on.
Given we only support clang, GCC, and MSVC for the time being, we can
utilize their built-in functions directly instead of going through the
OS's API functions.
This shrinks the overall code down to just
if (msvc)
use msvc's functions
else if (clang or gcc)
use clang/gcc's builtins
else
use the slow path
The template type here is actually a forwarding reference, not an rvalue
reference in this case, so it's more appropriate to use std::forward to
preserve the value category of the type being moved.
Makes the return type consistently uniform (like the intrinsics we're
wrapping). This also conveniently silences a truncation warning within
the kernel multi_level_queue.
