Add code to check that a function doesn't accidentally overwrite
anything in the area located just above the current stack frame.

This allows selecting at runtime if placebo should use OpenGL
or Vulkan for rendering.

To un-clutter the main dav1dplay.c, move the fifo to its own
file and header.

If the maximum number of arguments (currently 15) is changed into
an even number, and a function actually takes the full number of
arguments, we would have the situation where the checked spot on
the stack is at the same place as we store an inverted copy of it.

We already allocate enough space for two values though (for stack
alignment purposes, 16 bytes on arm64 and 8 bytes on arm32) so by
storing the reference in the upper half of this, the lower half of
it works as canary and isn't overwritten.

Henrik Gramner authored 5 years ago and Henrik Gramner committed 5 years ago

Use 'unsigned' instead of 'unsigned int' for consistency.
Add 'const' to a few variables.
Make proper use of C99 features.

Henrik Gramner authored 5 years ago and Henrik Gramner committed 5 years ago

Also skip the AVX warmup.

Matthieu Bouron authored 5 years ago and Martin Storsjö committed 5 years ago

If functions return a float value, this value is stored in this
register.

We should just use a normal bl here, and the linker will add the 'x'
bit if necessary.

This fixes calling the checkasm_fail_func on windows, where the
code is built in thumb mode (and the linker doesn't clear the 'x'
bit in the blx instruction).

Matthias Dressel authored 5 years ago and Jean-Baptiste Kempf committed 5 years ago

* The build from 'build-debian' is reused. 'logging' is not disabled
  since that would trigger an almost full rebuild.
* All ASM tests are merged into one job which is expected to
  seldomly fail, therefore ease of debugging is traded in for
  efficiency.

When benchmarking, the functions are called with a fixed width
of 64x32 or 32x16, while the test itself is run with a random size
in the range up to 128x32.

In 16 bpc mode, the source pixels must be within the valid range,
because they otherwise cause accesses out of bounds in the scaling
array.

Henrik Gramner authored 5 years ago and Jean-Baptiste Kempf committed 5 years ago

Also avoid integer overflows by using 64-bit intermediate precision.

Henrik Gramner authored 5 years ago and Henrik Gramner committed 5 years ago

Allows for macro-op fusion.

Henrik Gramner authored 5 years ago and Henrik Gramner committed 5 years ago

Eliminates the x86-64 check from the meson configuration file to be
consistent with how other x86-64-exclusive code is handled.

Henrik Gramner authored 5 years ago and Henrik Gramner committed 5 years ago

Allows for constant propagation and tail call elimination in the
msac initialization, which is performed in each tile.

Henrik Gramner authored 5 years ago and Henrik Gramner committed 5 years ago

Utilize the unsigned representation of a signed integer to skip
the refill code if the count was already negative to begin with,
which saves a few clock cycles at the end of each tile.

Add an element size specifier to the existing individual transform
functions for 8 bpc, naming them e.g. inv_dct_8h_x8_neon, to clarify
that they operate on input vectors of 8h, and make the symbols
public, to let the 10 bpc case call them from a different object file.
The same convention is used in the new itx16.S, like inv_dct_4s_x8_neon.

Make the existing itx.S compiled regardless of whether 8 bpc support
is enabled. For builds with 8 bpc support disabled, this does include
the unused frontend functions though, but this is hopefully tolerable
to avoid having to split the file into a sharable file for transforms
and a separate one for frontends.

This only implements the 10 bpc case, as that case can use transforms
operating on 16 bit coefficients in the second pass.

Relative speedup vs C for a few functions:

                                     Cortex A53    A72    A73
inv_txfm_add_4x4_dct_dct_0_10bpc_neon:     4.14   4.06   4.49
inv_txfm_add_4x4_dct_dct_1_10bpc_neon:     6.51   6.49   6.42
inv_txfm_add_8x8_dct_dct_0_10bpc_neon:     5.02   4.63   6.23
inv_txfm_add_8x8_dct_dct_1_10bpc_neon:     8.54   7.13  11.96
inv_txfm_add_16x16_dct_dct_0_10bpc_neon:   5.52   6.60   8.03
inv_txfm_add_16x16_dct_dct_1_10bpc_neon:  11.27   9.62  12.22
inv_txfm_add_16x16_dct_dct_2_10bpc_neon:   9.60   6.97   8.59
inv_txfm_add_32x32_dct_dct_0_10bpc_neon:   2.60   3.48   3.19
inv_txfm_add_32x32_dct_dct_1_10bpc_neon:  14.65  12.64  16.86
inv_txfm_add_32x32_dct_dct_2_10bpc_neon:  11.57   8.80  12.68
inv_txfm_add_32x32_dct_dct_3_10bpc_neon:   8.79   8.00   9.21
inv_txfm_add_32x32_dct_dct_4_10bpc_neon:   7.58   6.21   7.80
inv_txfm_add_64x64_dct_dct_0_10bpc_neon:   2.41   2.85   2.75
inv_txfm_add_64x64_dct_dct_1_10bpc_neon:  12.91  10.27  12.24
inv_txfm_add_64x64_dct_dct_2_10bpc_neon:  10.96   7.97  10.31
inv_txfm_add_64x64_dct_dct_3_10bpc_neon:   8.95   7.42   9.55
inv_txfm_add_64x64_dct_dct_4_10bpc_neon:   7.97   6.12   7.82

This matches what is done in C by -fvisibility=hidden.

This avoids issues with relocations against other symbols exported
from another assembly file.

Before this, we never did the early exit from the first pass.

Before:                               Cortex A53      A72      A73
inv_txfm_add_64x16_dct_dct_1_8bpc_neon:   7275.7   5198.3   5250.9
inv_txfm_add_64x16_dct_dct_2_8bpc_neon:   7276.1   5197.0   5251.3
inv_txfm_add_64x16_dct_dct_3_8bpc_neon:   7275.8   5196.2   5254.5
inv_txfm_add_64x16_dct_dct_4_8bpc_neon:   7273.6   5198.8   5254.2
After:
inv_txfm_add_64x16_dct_dct_1_8bpc_neon:   5187.8   3763.8   3735.0
inv_txfm_add_64x16_dct_dct_2_8bpc_neon:   7280.6   5185.6   5256.3
inv_txfm_add_64x16_dct_dct_3_8bpc_neon:   7270.7   5179.8   5250.3
inv_txfm_add_64x16_dct_dct_4_8bpc_neon:   7271.7   5212.4   5256.4

The other related variants didn't have this bug and properly exited
early when possible.

Unify some loads and stores, avoiding some extra pointer moving.

This gives a couple cycles speedup.

This isn't used for a sqrdmulh in its current form here.

The one left in idct_coeffs[1] isn't used within the idct itself,
but inv_txfm_horz_scale_dct_32x8 relies on it being left there for
use with sqrdmulh scaling later.

These cases were removed from x86 to save space and simplify the code
in e0b88bd2, as those cases
were essentially unused in real world bitstreams.