Skip to content

Commit

Permalink
smolscale-avx2: Pass YMM return values by reference
Browse files Browse the repository at this point in the history 
Workaround for GCC bug #54412:
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=54412

GCC does not enforce 32-byte alignment for return values placed on the
stack in 64-bit MS Windows builds, yet it emits vmovdqa instructions
that require alignment. This would cause crashes when resampling big
images.

Fixes #210 (GitHub).
  • Loading branch information
@hpjansson
hpjansson committed Sep 8, 2024
1 parent 9533e05 commit 7a0ea95
Showing 1 changed file with 30 additions and 26 deletions.
56 changes: 30 additions & 26 deletions chafa/internal/smolscale/smolscale-avx2.c
View file
Original file line number Diff line number Diff line change
Expand Up @@ -1553,8 +1553,8 @@ hadd_pixels_16x_to_8x_64bpp (__m256i i0, __m256i i1, __m256i i2, __m256i i3,
*o1 = t1;
}

static SMOL_INLINE __m256i
hadd_pixels_8x_to_4x_64bpp (__m256i i0, __m256i i1)
static SMOL_INLINE void
hadd_pixels_8x_to_4x_64bpp (__m256i i0, __m256i i1, __m256i * SMOL_RESTRICT o0)
{
__m256i t0, t1;

Expand All @@ -1568,7 +1568,7 @@ hadd_pixels_8x_to_4x_64bpp (__m256i i0, __m256i i1)

t0 = _mm256_permute4x64_epi64 (t0, CONTROL_4X2BIT_3_1_2_0);

return t0;
*o0 = t0;
}

static SMOL_INLINE void
Expand Down Expand Up @@ -1691,30 +1691,32 @@ interp_horizontal_bilinear_batch_64bpp (const uint64_t * SMOL_RESTRICT row_parts
*o3 = _mm256_permute4x64_epi64 (m3, CONTROL_4X2BIT_3_1_2_0);
}

static __m256i
static void
interp_horizontal_bilinear_batch_to_4x_64bpp (const uint64_t * SMOL_RESTRICT row_parts_in,
const uint16_t * SMOL_RESTRICT precalc_x)
const uint16_t * SMOL_RESTRICT precalc_x,
__m256i * SMOL_RESTRICT o0)
{
__m256i m0, m1, m2, m3, s0, s1;

interp_horizontal_bilinear_batch_64bpp (row_parts_in, precalc_x, &m0, &m1, &m2, &m3);
hadd_pixels_16x_to_8x_64bpp (m0, m1, m2, m3, &s0, &s1);
return hadd_pixels_8x_to_4x_64bpp (s0, s1);
hadd_pixels_8x_to_4x_64bpp (s0, s1, o0);
}

static __m256i
static void
interp_horizontal_bilinear_4x_batch_to_4x_64bpp (const uint64_t * SMOL_RESTRICT row_parts_in,
const uint16_t * SMOL_RESTRICT precalc_x)
const uint16_t * SMOL_RESTRICT precalc_x,
__m256i * SMOL_RESTRICT o0)
{
__m256i t0, t1, t2, t3;

t0 = interp_horizontal_bilinear_batch_to_4x_64bpp (row_parts_in, precalc_x);
t1 = interp_horizontal_bilinear_batch_to_4x_64bpp (row_parts_in, precalc_x + 32);
t2 = interp_horizontal_bilinear_batch_to_4x_64bpp (row_parts_in, precalc_x + 64);
t3 = interp_horizontal_bilinear_batch_to_4x_64bpp (row_parts_in, precalc_x + 96);
interp_horizontal_bilinear_batch_to_4x_64bpp (row_parts_in, precalc_x, &t0);
interp_horizontal_bilinear_batch_to_4x_64bpp (row_parts_in, precalc_x + 32, &t1);
interp_horizontal_bilinear_batch_to_4x_64bpp (row_parts_in, precalc_x + 64, &t2);
interp_horizontal_bilinear_batch_to_4x_64bpp (row_parts_in, precalc_x + 96, &t3);

hadd_pixels_16x_to_8x_64bpp (t0, t1, t2, t3, &t0, &t1);
return hadd_pixels_8x_to_4x_64bpp (t0, t1);
hadd_pixels_8x_to_4x_64bpp (t0, t1, o0);
}

/* Note that precalc_x must point to offsets and factors interleaved one by one, i.e.
Expand Down Expand Up @@ -1824,7 +1826,9 @@ interp_horizontal_bilinear_2h_64bpp (const SmolScaleCtx *scale_ctx,

while (row_parts_out + 4 <= row_parts_out_max)
{
__m256i t = interp_horizontal_bilinear_batch_to_4x_64bpp (row_parts_in, precalc_x);
__m256i t;

interp_horizontal_bilinear_batch_to_4x_64bpp (row_parts_in, precalc_x, &t);
t = _mm256_srli_epi16 (t, 2);
_mm256_store_si256 ((__m256i *) row_parts_out, t);

Expand All @@ -1851,10 +1855,10 @@ interp_horizontal_bilinear_3h_64bpp (const SmolScaleCtx *scale_ctx,
{
__m256i s0, s1;

s0 = interp_horizontal_bilinear_batch_to_4x_64bpp (row_parts_in, precalc_x);
s1 = interp_horizontal_bilinear_batch_to_4x_64bpp (row_parts_in, precalc_x + 32);
interp_horizontal_bilinear_batch_to_4x_64bpp (row_parts_in, precalc_x, &s0);
interp_horizontal_bilinear_batch_to_4x_64bpp (row_parts_in, precalc_x + 32, &s1);

s0 = hadd_pixels_8x_to_4x_64bpp (s0, s1);
hadd_pixels_8x_to_4x_64bpp (s0, s1, &s0);
s0 = _mm256_srli_epi16 (s0, 3);
_mm256_store_si256 ((__m256i *) row_parts_out, s0);

Expand All @@ -1881,7 +1885,7 @@ interp_horizontal_bilinear_4h_64bpp (const SmolScaleCtx *scale_ctx,
{
__m256i t0;

t0 = interp_horizontal_bilinear_4x_batch_to_4x_64bpp (row_parts_in, precalc_x);
interp_horizontal_bilinear_4x_batch_to_4x_64bpp (row_parts_in, precalc_x, &t0);
t0 = _mm256_srli_epi16 (t0, 4);
_mm256_store_si256 ((__m256i *) row_parts_out, t0);

Expand All @@ -1908,10 +1912,10 @@ interp_horizontal_bilinear_5h_64bpp (const SmolScaleCtx *scale_ctx,
{
__m256i t0, t1;

t0 = interp_horizontal_bilinear_4x_batch_to_4x_64bpp (row_parts_in, precalc_x);
t1 = interp_horizontal_bilinear_4x_batch_to_4x_64bpp (row_parts_in, precalc_x + 128);
interp_horizontal_bilinear_4x_batch_to_4x_64bpp (row_parts_in, precalc_x, &t0);
interp_horizontal_bilinear_4x_batch_to_4x_64bpp (row_parts_in, precalc_x + 128, &t1);

t0 = hadd_pixels_8x_to_4x_64bpp (t0, t1);
hadd_pixels_8x_to_4x_64bpp (t0, t1, &t0);
t0 = _mm256_srli_epi16 (t0, 5);
_mm256_store_si256 ((__m256i *) row_parts_out, t0);

Expand All @@ -1938,13 +1942,13 @@ interp_horizontal_bilinear_6h_64bpp (const SmolScaleCtx *scale_ctx,
{
__m256i t0, t1, t2, t3;

t0 = interp_horizontal_bilinear_4x_batch_to_4x_64bpp (row_parts_in, precalc_x);
t1 = interp_horizontal_bilinear_4x_batch_to_4x_64bpp (row_parts_in, precalc_x + 128);
t2 = interp_horizontal_bilinear_4x_batch_to_4x_64bpp (row_parts_in, precalc_x + 256);
t3 = interp_horizontal_bilinear_4x_batch_to_4x_64bpp (row_parts_in, precalc_x + 384);
interp_horizontal_bilinear_4x_batch_to_4x_64bpp (row_parts_in, precalc_x, &t0);
interp_horizontal_bilinear_4x_batch_to_4x_64bpp (row_parts_in, precalc_x + 128, &t1);
interp_horizontal_bilinear_4x_batch_to_4x_64bpp (row_parts_in, precalc_x + 256, &t2);
interp_horizontal_bilinear_4x_batch_to_4x_64bpp (row_parts_in, precalc_x + 384, &t3);

hadd_pixels_16x_to_8x_64bpp (t0, t1, t2, t3, &t0, &t1);
t0 = hadd_pixels_8x_to_4x_64bpp (t0, t1);
hadd_pixels_8x_to_4x_64bpp (t0, t1, &t0);

t0 = _mm256_srli_epi16 (t0, 6);
_mm256_store_si256 ((__m256i *) row_parts_out, t0);
Expand Down

0 comments on commit 7a0ea95

Please sign in to comment.