void ar_stride3_vminall_u8_neon(const uint8_t* a,
uint32_t n,
uint8_t* line_results)
{
uint8x16x3_t a_loaded;
uint8x16_t line0_min = vdupq_n_u8(255);
uint8x16_t line1_min = vdupq_n_u8(255);
uint8x16_t line2_min = vdupq_n_u8(255);
line_results[0] = 255;
line_results[1] = 255;
line_results[2] = 255;
uint8_t line0_array[16];
uint8_t line1_array[16];
uint8_t line2_array[16];
for (uint32_t i = 0; i < n; i += 48) {
a_loaded = vld3q_u8(&(a[i]));
line0_min = vminq_u8(a_loaded.val[0], line0_min);
line1_min = vminq_u8(a_loaded.val[1], line1_min);
line2_min = vminq_u8(a_loaded.val[2], line2_min);
}
vst1q_u8(line0_array, line0_min);
vst1q_u8(line1_array, line1_min);
vst1q_u8(line2_array, line2_min);
for (uint32_t i = 0; i < 16; i++) {
line_results[0] = ar_min_u8(line_results[0], line0_array[i]);
line_results[1] = ar_min_u8(line_results[1], line1_array[i]);
line_results[2] = ar_min_u8(line_results[2], line2_array[i]);
}
}
/* Build an RGBA palette from the RGB and separate alpha palettes. */
void
png_riffle_palette_rgba(png_structrp png_ptr, png_row_infop row_info)
{
png_const_colorp palette = png_ptr->palette;
png_bytep riffled_palette = png_ptr->riffled_palette;
png_const_bytep trans_alpha = png_ptr->trans_alpha;
int num_trans = png_ptr->num_trans;
int i;
/* Initially black, opaque. */
uint8x16x4_t w = {{
vdupq_n_u8(0x00),
vdupq_n_u8(0x00),
vdupq_n_u8(0x00),
vdupq_n_u8(0xff),
}};
if (row_info->bit_depth != 8)
{
png_error(png_ptr, "bit_depth must be 8 for png_riffle_palette_rgba");
return;
}
/* First, riffle the RGB colours into a RGBA palette, the A value is
* set to opaque for now.
*/
for (i = 0; i < (1 << row_info->bit_depth); i += 16)
{
uint8x16x3_t v = vld3q_u8((png_const_bytep)(palette + i));
w.val[0] = v.val[0];
w.val[1] = v.val[1];
w.val[2] = v.val[2];
vst4q_u8(riffled_palette + (i << 2), w);
}
/* Fix up the missing transparency values. */
for (i = 0; i < num_trans; i++)
riffled_palette[(i << 2) + 3] = trans_alpha[i];
}
void test_vld3Qu8 (void)
{
uint8x16x3_t out_uint8x16x3_t;
out_uint8x16x3_t = vld3q_u8 (0);
}
// If we have ARM NEON support, pick off 48 bytes at a time:
while (srclen >= 48)
{
uint8x16x3_t str;
uint8x16x4_t res;
// Load 48 bytes and deinterleave:
str = vld3q_u8((uint8_t *)c);
// Reshuffle:
res = enc_reshuffle(str);
// Translate reshuffled bytes to the Base64 alphabet:
res = enc_translate(res);
// Interleave and store result:
vst4q_u8((uint8_t *)o, res);
c += 48; // 3 * 16 bytes of input
o += 64; // 4 * 16 bytes of output
outl += 64;
srclen -= 48;
}
inline uint8x16x3_t vld3q(const u8 * ptr) { return vld3q_u8(ptr); }
