Skip to content

Commit

Permalink
Add casts to ARM NEON code
Browse files Browse the repository at this point in the history 
(cherry picked from commit b61b50e)
  • Loading branch information
@madebr
madebr committed Oct 15, 2024
1 parent 62e9c82 commit a3d0895
Showing 1 changed file with 52 additions and 51 deletions.
103 changes: 52 additions & 51 deletions SDL_ttf.c
View file
Original file line number Diff line number Diff line change
Expand Up @@ -775,33 +775,33 @@ static SDL_INLINE void BG_Blended_Opaque_NEON(const TTF_Image *image, Uint32 *de

uint32x4_t s, d0, d1, d2, d3, r0, r1, r2, r3;
uint8x16x2_t sx, sx01, sx23;
uint32x4_t zero = vmovq_n_u32(0);
const uint8x16_t zero = vdupq_n_u8(0);

while (height--) {
/* *INDENT-OFF* */
DUFFS_LOOP4(
/* Read 4 Uint32 and put 16 Uint8 into uint32x4x2_t (uint8x16x2_t)
* takes advantage of vzipq_u8 which produces two lanes */

s = vld1q_u32(src); // load
d0 = vld1q_u32(dst); // load
d1 = vld1q_u32(dst + 4); // load
d2 = vld1q_u32(dst + 8); // load
d3 = vld1q_u32(dst + 12); // load

sx = vzipq_u8(zero, s); // interleave
sx01 = vzipq_u8(zero, sx.val[0]); // interleave
sx23 = vzipq_u8(zero, sx.val[1]); // interleave
// already shifted by 24
r0 = vorrq_u32(d0, sx01.val[0]); // or
r1 = vorrq_u32(d1, sx01.val[1]); // or
r2 = vorrq_u32(d2, sx23.val[0]); // or
r3 = vorrq_u32(d3, sx23.val[1]); // or

vst1q_u32(dst, r0); // store
vst1q_u32(dst + 4, r1); // store
vst1q_u32(dst + 8, r2); // store
vst1q_u32(dst + 12, r3); // store
s = vld1q_u32(src); // load
d0 = vld1q_u32(dst); // load
d1 = vld1q_u32(dst + 4); // load
d2 = vld1q_u32(dst + 8); // load
d3 = vld1q_u32(dst + 12); // load

sx = vzipq_u8(zero, (uint8x16_t)s); // interleave
sx01 = vzipq_u8(zero, sx.val[0]); // interleave
sx23 = vzipq_u8(zero, sx.val[1]); // interleave
// already shifted by 24
r0 = vorrq_u32(d0, (uint32x4_t)sx01.val[0]); // or
r1 = vorrq_u32(d1, (uint32x4_t)sx01.val[1]); // or
r2 = vorrq_u32(d2, (uint32x4_t)sx23.val[0]); // or
r3 = vorrq_u32(d3, (uint32x4_t)sx23.val[1]); // or

vst1q_u32(dst, r0); // store
vst1q_u32(dst + 4, r1); // store
vst1q_u32(dst + 8, r2); // store
vst1q_u32(dst + 12, r3); // store

dst += 16;
src += 4;
Expand All @@ -823,60 +823,61 @@ static SDL_INLINE void BG_Blended_NEON(const TTF_Image *image, Uint32 *destinati
uint32x4_t s, d0, d1, d2, d3, r0, r1, r2, r3;
uint16x8_t Ls8, Hs8;
uint8x16x2_t sx, sx01, sx23;
uint16x8x2_t sxm;

const uint16x8_t alpha = vmovq_n_u16(fg_alpha);
const uint16x8_t one = vmovq_n_u16(1);
const uint32x4_t zero = vmovq_n_u32(0);
const uint8x16_t zero = vdupq_n_u8(0);

while (height--) {
/* *INDENT-OFF* */
DUFFS_LOOP4(
/* Read 4 Uint32 and put 16 Uint8 into uint32x4x2_t (uint8x16x2_t)
* takes advantage of vzipq_u8 which produces two lanes */

s = vld1q_u32(src); // load
d0 = vld1q_u32(dst); // load
d1 = vld1q_u32(dst + 4); // load
d2 = vld1q_u32(dst + 8); // load
d3 = vld1q_u32(dst + 12); // load
s = vld1q_u32(src); // load
d0 = vld1q_u32(dst); // load
d1 = vld1q_u32(dst + 4); // load
d2 = vld1q_u32(dst + 8); // load
d3 = vld1q_u32(dst + 12); // load

sx = vzipq_u8(s, zero); // interleave, no shifting
// enough room to multiply
sx = vzipq_u8((uint8x16_t)s, zero); // interleave, no shifting
// enough room to multiply

/* Apply: alpha_table[i] = ((i * fg.a / 255) << 24; */
/* Divide by 255 is done as: (x + 1 + (x >> 8)) >> 8 */

sx.val[0] = vmulq_u16(sx.val[0], alpha); // x := i * fg.a
sx.val[1] = vmulq_u16(sx.val[1], alpha);
sxm.val[0] = vmulq_u16((uint16x8_t)sx.val[0], alpha); // x := i * fg.a
sxm.val[1] = vmulq_u16((uint16x8_t)sx.val[1], alpha);

Ls8 = vshrq_n_u16(sx.val[0], 8); // x >> 8
Hs8 = vshrq_n_u16(sx.val[1], 8);
Ls8 = vshrq_n_u16(sxm.val[0], 8); // x >> 8
Hs8 = vshrq_n_u16(sxm.val[1], 8);

sx.val[0] = vaddq_u16(sx.val[0], one); // x + 1
sx.val[1] = vaddq_u16(sx.val[1], one);
sxm.val[0] = vaddq_u16(sxm.val[0], one); // x + 1
sxm.val[1] = vaddq_u16(sxm.val[1], one);

sx.val[0] = vaddq_u16(sx.val[0], Ls8); // x + 1 + (x >> 8)
sx.val[1] = vaddq_u16(sx.val[1], Hs8);
sxm.val[0] = vaddq_u16(sxm.val[0], Ls8); // x + 1 + (x >> 8)
sxm.val[1] = vaddq_u16(sxm.val[1], Hs8);

sx.val[0] = vshrq_n_u16(sx.val[0], 8); // ((x + 1 + (x >> 8)) >> 8
sx.val[1] = vshrq_n_u16(sx.val[1], 8);
sxm.val[0] = vshrq_n_u16(sxm.val[0], 8); // ((x + 1 + (x >> 8)) >> 8
sxm.val[1] = vshrq_n_u16(sxm.val[1], 8);

sx.val[0] = vshlq_n_u16(sx.val[0], 8); // shift << 8, so we're prepared
sx.val[1] = vshlq_n_u16(sx.val[1], 8); // to have final format << 24
sxm.val[0] = vshlq_n_u16(sxm.val[0], 8); // shift << 8, so we're prepared
sxm.val[1] = vshlq_n_u16(sxm.val[1], 8); // to have final format << 24

sx01 = vzipq_u8(zero, sx.val[0]); // interleave
sx23 = vzipq_u8(zero, sx.val[1]); // interleave
// already shifted by 24
sx01 = vzipq_u8(zero, (uint8x16_t)sxm.val[0]); // interleave
sx23 = vzipq_u8(zero, (uint8x16_t)sxm.val[1]); // interleave
// already shifted by 24

r0 = vorrq_u32(d0, sx01.val[0]); // or
r1 = vorrq_u32(d1, sx01.val[1]); // or
r2 = vorrq_u32(d2, sx23.val[0]); // or
r3 = vorrq_u32(d3, sx23.val[1]); // or
r0 = vorrq_u32(d0, (uint32x4_t)sx01.val[0]); // or
r1 = vorrq_u32(d1, (uint32x4_t)sx01.val[1]); // or
r2 = vorrq_u32(d2, (uint32x4_t)sx23.val[0]); // or
r3 = vorrq_u32(d3, (uint32x4_t)sx23.val[1]); // or

vst1q_u32(dst, r0); // store
vst1q_u32(dst + 4, r1); // store
vst1q_u32(dst + 8, r2); // store
vst1q_u32(dst + 12, r3); // store
vst1q_u32(dst, r0); // store
vst1q_u32(dst + 4, r1); // store
vst1q_u32(dst + 8, r2); // store
vst1q_u32(dst + 12, r3); // store

dst += 16;
src += 4;
Expand Down

0 comments on commit a3d0895

Please sign in to comment.