summaryrefslogtreecommitdiff
path: root/src/libffmpeg/libavcodec/alpha/dsputil_alpha.c
diff options
context:
space:
mode:
Diffstat (limited to 'src/libffmpeg/libavcodec/alpha/dsputil_alpha.c')
-rw-r--r--src/libffmpeg/libavcodec/alpha/dsputil_alpha.c246
1 files changed, 143 insertions, 103 deletions
diff --git a/src/libffmpeg/libavcodec/alpha/dsputil_alpha.c b/src/libffmpeg/libavcodec/alpha/dsputil_alpha.c
index 5e1aa2093..9a3fb1eac 100644
--- a/src/libffmpeg/libavcodec/alpha/dsputil_alpha.c
+++ b/src/libffmpeg/libavcodec/alpha/dsputil_alpha.c
@@ -22,6 +22,8 @@
void simple_idct_axp(DCTELEM *block);
+void put_pixels_axp_asm(uint8_t *block, const uint8_t *pixels,
+ int line_size, int h);
void put_pixels_clamped_mvi_asm(const DCTELEM *block, uint8_t *pixels,
int line_size);
void add_pixels_clamped_mvi_asm(const DCTELEM *block, uint8_t *pixels,
@@ -103,145 +105,183 @@ void add_pixels_clamped_mvi(const DCTELEM *block, uint8_t *pixels,
}
#endif
-/* Average 8 unsigned bytes in parallel: (b1 + b2) >> 1
- Since the immediate result could be greater than 255, we do the
- shift first. The result is too low by one if the bytes were both
- odd, so we need to add (l1 & l2) & BYTE_VEC(0x01). */
-static inline UINT64 avg2_no_rnd(UINT64 l1, UINT64 l2)
-{
- UINT64 correction = (l1 & l2) & BYTE_VEC(0x01);
- l1 = (l1 & ~BYTE_VEC(0x01)) >> 1;
- l2 = (l2 & ~BYTE_VEC(0x01)) >> 1;
- return l1 + l2 + correction;
+static void clear_blocks_axp(DCTELEM *blocks) {
+ uint64_t *p = (uint64_t *) blocks;
+ int n = sizeof(DCTELEM) * 6 * 64;
+
+ do {
+ p[0] = 0;
+ p[1] = 0;
+ p[2] = 0;
+ p[3] = 0;
+ p[4] = 0;
+ p[5] = 0;
+ p[6] = 0;
+ p[7] = 0;
+ p += 8;
+ n -= 8 * 8;
+ } while (n);
}
-/* Average 8 bytes with rounding: (b1 + b2 + 1) >> 1
- The '1' only has an effect when one byte is even and the other odd,
- i. e. we also need to add (l1 ^ l2) & BYTE_VEC(0x01).
- Incidentally, that is equivalent to (l1 | l2) & BYTE_VEC(0x01). */
-static inline UINT64 avg2(UINT64 l1, UINT64 l2)
+static inline uint64_t avg2_no_rnd(uint64_t a, uint64_t b)
{
- UINT64 correction = (l1 | l2) & BYTE_VEC(0x01);
- l1 = (l1 & ~BYTE_VEC(0x01)) >> 1;
- l2 = (l2 & ~BYTE_VEC(0x01)) >> 1;
- return l1 + l2 + correction;
+ return (a & b) + (((a ^ b) & BYTE_VEC(0xfe)) >> 1);
}
-static inline UINT64 avg4(UINT64 l1, UINT64 l2, UINT64 l3, UINT64 l4)
+static inline uint64_t avg2(uint64_t a, uint64_t b)
{
- UINT64 r1 = ((l1 & ~BYTE_VEC(0x03)) >> 2)
- + ((l2 & ~BYTE_VEC(0x03)) >> 2)
- + ((l3 & ~BYTE_VEC(0x03)) >> 2)
- + ((l4 & ~BYTE_VEC(0x03)) >> 2);
- UINT64 r2 = (( (l1 & BYTE_VEC(0x03))
- + (l2 & BYTE_VEC(0x03))
- + (l3 & BYTE_VEC(0x03))
- + (l4 & BYTE_VEC(0x03))
- + BYTE_VEC(0x02)) >> 2) & BYTE_VEC(0x03);
- return r1 + r2;
+ return (a | b) - (((a ^ b) & BYTE_VEC(0xfe)) >> 1);
}
-static inline UINT64 avg4_no_rnd(UINT64 l1, UINT64 l2, UINT64 l3, UINT64 l4)
+#if 0
+/* The XY2 routines basically utilize this scheme, but reuse parts in
+ each iteration. */
+static inline uint64_t avg4(uint64_t l1, uint64_t l2, uint64_t l3, uint64_t l4)
{
- UINT64 r1 = ((l1 & ~BYTE_VEC(0x03)) >> 2)
- + ((l2 & ~BYTE_VEC(0x03)) >> 2)
- + ((l3 & ~BYTE_VEC(0x03)) >> 2)
- + ((l4 & ~BYTE_VEC(0x03)) >> 2);
- UINT64 r2 = (( (l1 & BYTE_VEC(0x03))
- + (l2 & BYTE_VEC(0x03))
- + (l3 & BYTE_VEC(0x03))
- + (l4 & BYTE_VEC(0x03))
- + BYTE_VEC(0x01)) >> 2) & BYTE_VEC(0x03);
+ uint64_t r1 = ((l1 & ~BYTE_VEC(0x03)) >> 2)
+ + ((l2 & ~BYTE_VEC(0x03)) >> 2)
+ + ((l3 & ~BYTE_VEC(0x03)) >> 2)
+ + ((l4 & ~BYTE_VEC(0x03)) >> 2);
+ uint64_t r2 = (( (l1 & BYTE_VEC(0x03))
+ + (l2 & BYTE_VEC(0x03))
+ + (l3 & BYTE_VEC(0x03))
+ + (l4 & BYTE_VEC(0x03))
+ + BYTE_VEC(0x02)) >> 2) & BYTE_VEC(0x03);
return r1 + r2;
}
+#endif
-#define PIXOPNAME(suffix) put ## suffix
-#define BTYPE UINT8
-#define AVG2 avg2
-#define AVG4 avg4
-#define STORE(l, b) stq(l, b)
-#include "pixops.h"
-#undef PIXOPNAME
-#undef BTYPE
-#undef AVG2
-#undef AVG4
-#undef STORE
+#define OP(LOAD, STORE) \
+ do { \
+ STORE(LOAD(pixels), block); \
+ pixels += line_size; \
+ block += line_size; \
+ } while (--h)
-#define PIXOPNAME(suffix) put_no_rnd ## suffix
-#define BTYPE UINT8
-#define AVG2 avg2_no_rnd
-#define AVG4 avg4_no_rnd
-#define STORE(l, b) stq(l, b)
-#include "pixops.h"
-#undef PIXOPNAME
-#undef BTYPE
-#undef AVG2
-#undef AVG4
-#undef STORE
+#define OP_X2(LOAD, STORE) \
+ do { \
+ uint64_t pix1, pix2; \
+ \
+ pix1 = LOAD(pixels); \
+ pix2 = pix1 >> 8 | ((uint64_t) pixels[8] << 56); \
+ STORE(AVG2(pix1, pix2), block); \
+ pixels += line_size; \
+ block += line_size; \
+ } while (--h)
-/* The following functions are untested. */
-#if 0
+#define OP_Y2(LOAD, STORE) \
+ do { \
+ uint64_t pix = LOAD(pixels); \
+ do { \
+ uint64_t next_pix; \
+ \
+ pixels += line_size; \
+ next_pix = LOAD(pixels); \
+ STORE(AVG2(pix, next_pix), block); \
+ block += line_size; \
+ pix = next_pix; \
+ } while (--h); \
+ } while (0)
+
+#define OP_XY2(LOAD, STORE) \
+ do { \
+ uint64_t pix1 = LOAD(pixels); \
+ uint64_t pix2 = pix1 >> 8 | ((uint64_t) pixels[8] << 56); \
+ uint64_t pix_l = (pix1 & BYTE_VEC(0x03)) \
+ + (pix2 & BYTE_VEC(0x03)); \
+ uint64_t pix_h = ((pix1 & ~BYTE_VEC(0x03)) >> 2) \
+ + ((pix2 & ~BYTE_VEC(0x03)) >> 2); \
+ \
+ do { \
+ uint64_t npix1, npix2; \
+ uint64_t npix_l, npix_h; \
+ uint64_t avg; \
+ \
+ pixels += line_size; \
+ npix1 = LOAD(pixels); \
+ npix2 = npix1 >> 8 | ((uint64_t) pixels[8] << 56); \
+ npix_l = (npix1 & BYTE_VEC(0x03)) \
+ + (npix2 & BYTE_VEC(0x03)); \
+ npix_h = ((npix1 & ~BYTE_VEC(0x03)) >> 2) \
+ + ((npix2 & ~BYTE_VEC(0x03)) >> 2); \
+ avg = (((pix_l + npix_l + AVG4_ROUNDER) >> 2) & BYTE_VEC(0x03)) \
+ + pix_h + npix_h; \
+ STORE(avg, block); \
+ \
+ block += line_size; \
+ pix_l = npix_l; \
+ pix_h = npix_h; \
+ } while (--h); \
+ } while (0)
+
+#define MAKE_OP(OPNAME, SUFF, OPKIND, STORE) \
+static void OPNAME ## _pixels ## SUFF ## _axp \
+ (uint8_t *restrict block, const uint8_t *restrict pixels, \
+ int line_size, int h) \
+{ \
+ if ((size_t) pixels & 0x7) { \
+ OPKIND(uldq, STORE); \
+ } else { \
+ OPKIND(ldq, STORE); \
+ } \
+}
-#define PIXOPNAME(suffix) avg ## suffix
-#define BTYPE UINT8
+#define PIXOP(OPNAME, STORE) \
+ MAKE_OP(OPNAME, , OP, STORE) \
+ MAKE_OP(OPNAME, _x2, OP_X2, STORE) \
+ MAKE_OP(OPNAME, _y2, OP_Y2, STORE) \
+ MAKE_OP(OPNAME, _xy2, OP_XY2, STORE)
+
+/* Rounding primitives. */
#define AVG2 avg2
#define AVG4 avg4
+#define AVG4_ROUNDER BYTE_VEC(0x02)
+#define STORE(l, b) stq(l, b)
+PIXOP(put, STORE);
+
+#undef STORE
#define STORE(l, b) stq(AVG2(l, ldq(b)), b);
-#include "pixops.h"
-#undef PIXOPNAME
-#undef BTYPE
+PIXOP(avg, STORE);
+
+/* Not rounding primitives. */
#undef AVG2
#undef AVG4
+#undef AVG4_ROUNDER
#undef STORE
-
-#define PIXOPNAME(suffix) avg_no_rnd ## suffix
-#define BTYPE UINT8
#define AVG2 avg2_no_rnd
#define AVG4 avg4_no_rnd
-#define STORE(l, b) stq(AVG2(l, ldq(b)), b);
-#include "pixops.h"
-#undef PIXOPNAME
-#undef BTYPE
-#undef AVG2
-#undef AVG4
-#undef STORE
+#define AVG4_ROUNDER BYTE_VEC(0x01)
+#define STORE(l, b) stq(l, b)
+PIXOP(put_no_rnd, STORE);
-#define PIXOPNAME(suffix) sub ## suffix
-#define BTYPE DCTELEM
-#define AVG2 avg2
-#define AVG4 avg4
-#define STORE(l, block) do { \
- UINT64 xxx = l; \
- (block)[0] -= (xxx >> 0) & 0xff; \
- (block)[1] -= (xxx >> 8) & 0xff; \
- (block)[2] -= (xxx >> 16) & 0xff; \
- (block)[3] -= (xxx >> 24) & 0xff; \
- (block)[4] -= (xxx >> 32) & 0xff; \
- (block)[5] -= (xxx >> 40) & 0xff; \
- (block)[6] -= (xxx >> 48) & 0xff; \
- (block)[7] -= (xxx >> 56) & 0xff; \
-} while (0)
-#include "pixops.h"
-#undef PIXOPNAME
-#undef BTYPE
-#undef AVG2
-#undef AVG4
#undef STORE
-
-#endif
+#define STORE(l, b) stq(AVG2(l, ldq(b)), b);
+PIXOP(avg_no_rnd, STORE);
void dsputil_init_alpha(void)
{
- put_pixels_tab[0] = put_pixels_axp;
+ put_pixels_tab[0] = put_pixels_axp_asm;
put_pixels_tab[1] = put_pixels_x2_axp;
put_pixels_tab[2] = put_pixels_y2_axp;
put_pixels_tab[3] = put_pixels_xy2_axp;
- put_no_rnd_pixels_tab[0] = put_pixels_axp;
+ put_no_rnd_pixels_tab[0] = put_pixels_axp_asm;
put_no_rnd_pixels_tab[1] = put_no_rnd_pixels_x2_axp;
put_no_rnd_pixels_tab[2] = put_no_rnd_pixels_y2_axp;
put_no_rnd_pixels_tab[3] = put_no_rnd_pixels_xy2_axp;
+ avg_pixels_tab[0] = avg_pixels_axp;
+ avg_pixels_tab[1] = avg_pixels_x2_axp;
+ avg_pixels_tab[2] = avg_pixels_y2_axp;
+ avg_pixels_tab[3] = avg_pixels_xy2_axp;
+
+ avg_no_rnd_pixels_tab[0] = avg_no_rnd_pixels_axp;
+ avg_no_rnd_pixels_tab[1] = avg_no_rnd_pixels_x2_axp;
+ avg_no_rnd_pixels_tab[2] = avg_no_rnd_pixels_y2_axp;
+ avg_no_rnd_pixels_tab[3] = avg_no_rnd_pixels_xy2_axp;
+
+ clear_blocks = clear_blocks_axp;
+
/* amask clears all bits that correspond to present features. */
if (amask(AMASK_MVI) == 0) {
put_pixels_clamped = put_pixels_clamped_mvi_asm;
generated by cgit v1.2.3 (git 2.25.1) at 2025-05-05 03:50:03 +0000