summaryrefslogtreecommitdiff
path: root/libavcodec/x86/dsputil_rnd_template.c
diff options
context:
space:
mode:
authorRonald S. Bultje <rsbultje@gmail.com>2013-03-10 15:37:59 -0700
committerMartin Storsjö <martin@martin.st>2013-04-19 23:18:53 +0300
commit8db00081a37d5b7e23918ee500bb16bc59b57197 (patch)
tree1f9fba986571a05fd4e75666867a2d0dffb4cca6 /libavcodec/x86/dsputil_rnd_template.c
parent28bc406c84b04a5f1458b90ff52ddbec73e46202 (diff)
x86: hpeldsp: Move half-pel assembly from dsputil to hpeldsp
Signed-off-by: Martin Storsjö <martin@martin.st>
Diffstat (limited to 'libavcodec/x86/dsputil_rnd_template.c')
-rw-r--r--libavcodec/x86/dsputil_rnd_template.c373
1 files changed, 0 insertions, 373 deletions
diff --git a/libavcodec/x86/dsputil_rnd_template.c b/libavcodec/x86/dsputil_rnd_template.c
index 6ce926c6d1..d441950fc1 100644
--- a/libavcodec/x86/dsputil_rnd_template.c
+++ b/libavcodec/x86/dsputil_rnd_template.c
@@ -25,212 +25,6 @@
*/
// put_pixels
-static void DEF(put, pixels8_x2)(uint8_t *block, const uint8_t *pixels, ptrdiff_t line_size, int h)
-{
- MOVQ_BFE(mm6);
- __asm__ volatile(
- "lea (%3, %3), %%"REG_a" \n\t"
- ".p2align 3 \n\t"
- "1: \n\t"
- "movq (%1), %%mm0 \n\t"
- "movq 1(%1), %%mm1 \n\t"
- "movq (%1, %3), %%mm2 \n\t"
- "movq 1(%1, %3), %%mm3 \n\t"
- PAVGBP(%%mm0, %%mm1, %%mm4, %%mm2, %%mm3, %%mm5)
- "movq %%mm4, (%2) \n\t"
- "movq %%mm5, (%2, %3) \n\t"
- "add %%"REG_a", %1 \n\t"
- "add %%"REG_a", %2 \n\t"
- "movq (%1), %%mm0 \n\t"
- "movq 1(%1), %%mm1 \n\t"
- "movq (%1, %3), %%mm2 \n\t"
- "movq 1(%1, %3), %%mm3 \n\t"
- PAVGBP(%%mm0, %%mm1, %%mm4, %%mm2, %%mm3, %%mm5)
- "movq %%mm4, (%2) \n\t"
- "movq %%mm5, (%2, %3) \n\t"
- "add %%"REG_a", %1 \n\t"
- "add %%"REG_a", %2 \n\t"
- "subl $4, %0 \n\t"
- "jnz 1b \n\t"
- :"+g"(h), "+S"(pixels), "+D"(block)
- :"r"((x86_reg)line_size)
- :REG_a, "memory");
-}
-
-static void av_unused DEF(put, pixels8_l2)(uint8_t *dst, uint8_t *src1, uint8_t *src2, int dstStride, int src1Stride, int h)
-{
- MOVQ_BFE(mm6);
- __asm__ volatile(
- "testl $1, %0 \n\t"
- " jz 1f \n\t"
- "movq (%1), %%mm0 \n\t"
- "movq (%2), %%mm1 \n\t"
- "add %4, %1 \n\t"
- "add $8, %2 \n\t"
- PAVGB(%%mm0, %%mm1, %%mm4, %%mm6)
- "movq %%mm4, (%3) \n\t"
- "add %5, %3 \n\t"
- "decl %0 \n\t"
- ".p2align 3 \n\t"
- "1: \n\t"
- "movq (%1), %%mm0 \n\t"
- "movq (%2), %%mm1 \n\t"
- "add %4, %1 \n\t"
- "movq (%1), %%mm2 \n\t"
- "movq 8(%2), %%mm3 \n\t"
- "add %4, %1 \n\t"
- PAVGBP(%%mm0, %%mm1, %%mm4, %%mm2, %%mm3, %%mm5)
- "movq %%mm4, (%3) \n\t"
- "add %5, %3 \n\t"
- "movq %%mm5, (%3) \n\t"
- "add %5, %3 \n\t"
- "movq (%1), %%mm0 \n\t"
- "movq 16(%2), %%mm1 \n\t"
- "add %4, %1 \n\t"
- "movq (%1), %%mm2 \n\t"
- "movq 24(%2), %%mm3 \n\t"
- "add %4, %1 \n\t"
- "add $32, %2 \n\t"
- PAVGBP(%%mm0, %%mm1, %%mm4, %%mm2, %%mm3, %%mm5)
- "movq %%mm4, (%3) \n\t"
- "add %5, %3 \n\t"
- "movq %%mm5, (%3) \n\t"
- "add %5, %3 \n\t"
- "subl $4, %0 \n\t"
- "jnz 1b \n\t"
-#if !HAVE_EBX_AVAILABLE //Note "+bm" and "+mb" are buggy too (with gcc 3.2.2 at least) and cannot be used
- :"+m"(h), "+a"(src1), "+c"(src2), "+d"(dst)
-#else
- :"+b"(h), "+a"(src1), "+c"(src2), "+d"(dst)
-#endif
- :"S"((x86_reg)src1Stride), "D"((x86_reg)dstStride)
- :"memory");
-}
-
-static void DEF(put, pixels16_x2)(uint8_t *block, const uint8_t *pixels, ptrdiff_t line_size, int h)
-{
- MOVQ_BFE(mm6);
- __asm__ volatile(
- "lea (%3, %3), %%"REG_a" \n\t"
- ".p2align 3 \n\t"
- "1: \n\t"
- "movq (%1), %%mm0 \n\t"
- "movq 1(%1), %%mm1 \n\t"
- "movq (%1, %3), %%mm2 \n\t"
- "movq 1(%1, %3), %%mm3 \n\t"
- PAVGBP(%%mm0, %%mm1, %%mm4, %%mm2, %%mm3, %%mm5)
- "movq %%mm4, (%2) \n\t"
- "movq %%mm5, (%2, %3) \n\t"
- "movq 8(%1), %%mm0 \n\t"
- "movq 9(%1), %%mm1 \n\t"
- "movq 8(%1, %3), %%mm2 \n\t"
- "movq 9(%1, %3), %%mm3 \n\t"
- PAVGBP(%%mm0, %%mm1, %%mm4, %%mm2, %%mm3, %%mm5)
- "movq %%mm4, 8(%2) \n\t"
- "movq %%mm5, 8(%2, %3) \n\t"
- "add %%"REG_a", %1 \n\t"
- "add %%"REG_a", %2 \n\t"
- "movq (%1), %%mm0 \n\t"
- "movq 1(%1), %%mm1 \n\t"
- "movq (%1, %3), %%mm2 \n\t"
- "movq 1(%1, %3), %%mm3 \n\t"
- PAVGBP(%%mm0, %%mm1, %%mm4, %%mm2, %%mm3, %%mm5)
- "movq %%mm4, (%2) \n\t"
- "movq %%mm5, (%2, %3) \n\t"
- "movq 8(%1), %%mm0 \n\t"
- "movq 9(%1), %%mm1 \n\t"
- "movq 8(%1, %3), %%mm2 \n\t"
- "movq 9(%1, %3), %%mm3 \n\t"
- PAVGBP(%%mm0, %%mm1, %%mm4, %%mm2, %%mm3, %%mm5)
- "movq %%mm4, 8(%2) \n\t"
- "movq %%mm5, 8(%2, %3) \n\t"
- "add %%"REG_a", %1 \n\t"
- "add %%"REG_a", %2 \n\t"
- "subl $4, %0 \n\t"
- "jnz 1b \n\t"
- :"+g"(h), "+S"(pixels), "+D"(block)
- :"r"((x86_reg)line_size)
- :REG_a, "memory");
-}
-
-static void av_unused DEF(put, pixels16_l2)(uint8_t *dst, uint8_t *src1, uint8_t *src2, int dstStride, int src1Stride, int h)
-{
- MOVQ_BFE(mm6);
- __asm__ volatile(
- "testl $1, %0 \n\t"
- " jz 1f \n\t"
- "movq (%1), %%mm0 \n\t"
- "movq (%2), %%mm1 \n\t"
- "movq 8(%1), %%mm2 \n\t"
- "movq 8(%2), %%mm3 \n\t"
- "add %4, %1 \n\t"
- "add $16, %2 \n\t"
- PAVGBP(%%mm0, %%mm1, %%mm4, %%mm2, %%mm3, %%mm5)
- "movq %%mm4, (%3) \n\t"
- "movq %%mm5, 8(%3) \n\t"
- "add %5, %3 \n\t"
- "decl %0 \n\t"
- ".p2align 3 \n\t"
- "1: \n\t"
- "movq (%1), %%mm0 \n\t"
- "movq (%2), %%mm1 \n\t"
- "movq 8(%1), %%mm2 \n\t"
- "movq 8(%2), %%mm3 \n\t"
- "add %4, %1 \n\t"
- PAVGBP(%%mm0, %%mm1, %%mm4, %%mm2, %%mm3, %%mm5)
- "movq %%mm4, (%3) \n\t"
- "movq %%mm5, 8(%3) \n\t"
- "add %5, %3 \n\t"
- "movq (%1), %%mm0 \n\t"
- "movq 16(%2), %%mm1 \n\t"
- "movq 8(%1), %%mm2 \n\t"
- "movq 24(%2), %%mm3 \n\t"
- "add %4, %1 \n\t"
- PAVGBP(%%mm0, %%mm1, %%mm4, %%mm2, %%mm3, %%mm5)
- "movq %%mm4, (%3) \n\t"
- "movq %%mm5, 8(%3) \n\t"
- "add %5, %3 \n\t"
- "add $32, %2 \n\t"
- "subl $2, %0 \n\t"
- "jnz 1b \n\t"
-#if !HAVE_EBX_AVAILABLE //Note "+bm" and "+mb" are buggy too (with gcc 3.2.2 at least) and cannot be used
- :"+m"(h), "+a"(src1), "+c"(src2), "+d"(dst)
-#else
- :"+b"(h), "+a"(src1), "+c"(src2), "+d"(dst)
-#endif
- :"S"((x86_reg)src1Stride), "D"((x86_reg)dstStride)
- :"memory");
-}
-
-static void DEF(put, pixels8_y2)(uint8_t *block, const uint8_t *pixels, ptrdiff_t line_size, int h)
-{
- MOVQ_BFE(mm6);
- __asm__ volatile(
- "lea (%3, %3), %%"REG_a" \n\t"
- "movq (%1), %%mm0 \n\t"
- ".p2align 3 \n\t"
- "1: \n\t"
- "movq (%1, %3), %%mm1 \n\t"
- "movq (%1, %%"REG_a"),%%mm2 \n\t"
- PAVGBP(%%mm1, %%mm0, %%mm4, %%mm2, %%mm1, %%mm5)
- "movq %%mm4, (%2) \n\t"
- "movq %%mm5, (%2, %3) \n\t"
- "add %%"REG_a", %1 \n\t"
- "add %%"REG_a", %2 \n\t"
- "movq (%1, %3), %%mm1 \n\t"
- "movq (%1, %%"REG_a"),%%mm0 \n\t"
- PAVGBP(%%mm1, %%mm2, %%mm4, %%mm0, %%mm1, %%mm5)
- "movq %%mm4, (%2) \n\t"
- "movq %%mm5, (%2, %3) \n\t"
- "add %%"REG_a", %1 \n\t"
- "add %%"REG_a", %2 \n\t"
- "subl $4, %0 \n\t"
- "jnz 1b \n\t"
- :"+g"(h), "+S"(pixels), "+D"(block)
- :"r"((x86_reg)line_size)
- :REG_a, "memory");
-}
-
static void DEF(put, pixels8_xy2)(uint8_t *block, const uint8_t *pixels, ptrdiff_t line_size, int h)
{
MOVQ_ZERO(mm7);
@@ -297,27 +91,6 @@ static void DEF(put, pixels8_xy2)(uint8_t *block, const uint8_t *pixels, ptrdiff
:REG_a, "memory");
}
-// avg_pixels
-static void av_unused DEF(avg, pixels4)(uint8_t *block, const uint8_t *pixels, ptrdiff_t line_size, int h)
-{
- MOVQ_BFE(mm6);
- JUMPALIGN();
- do {
- __asm__ volatile(
- "movd %0, %%mm0 \n\t"
- "movd %1, %%mm1 \n\t"
- OP_AVG(%%mm0, %%mm1, %%mm2, %%mm6)
- "movd %%mm2, %0 \n\t"
- :"+m"(*block)
- :"m"(*pixels)
- :"memory");
- pixels += line_size;
- block += line_size;
- }
- while (--h);
-}
-
-#ifndef NO_RND
// in case more speed is needed - unroling would certainly help
static void DEF(avg, pixels8)(uint8_t *block, const uint8_t *pixels, ptrdiff_t line_size, int h)
{
@@ -337,7 +110,6 @@ static void DEF(avg, pixels8)(uint8_t *block, const uint8_t *pixels, ptrdiff_t l
}
while (--h);
}
-#endif // NO_RND
static void DEF(avg, pixels16)(uint8_t *block, const uint8_t *pixels, ptrdiff_t line_size, int h)
{
@@ -362,141 +134,6 @@ static void DEF(avg, pixels16)(uint8_t *block, const uint8_t *pixels, ptrdiff_t
while (--h);
}
-#ifndef NO_RND
-static void DEF(avg, pixels8_x2)(uint8_t *block, const uint8_t *pixels, ptrdiff_t line_size, int h)
-{
- MOVQ_BFE(mm6);
- JUMPALIGN();
- do {
- __asm__ volatile(
- "movq %1, %%mm0 \n\t"
- "movq 1%1, %%mm1 \n\t"
- "movq %0, %%mm3 \n\t"
- PAVGB(%%mm0, %%mm1, %%mm2, %%mm6)
- OP_AVG(%%mm3, %%mm2, %%mm0, %%mm6)
- "movq %%mm0, %0 \n\t"
- :"+m"(*block)
- :"m"(*pixels)
- :"memory");
- pixels += line_size;
- block += line_size;
- } while (--h);
-}
-#endif // NO_RND
-
-static av_unused void DEF(avg, pixels8_l2)(uint8_t *dst, uint8_t *src1, uint8_t *src2, int dstStride, int src1Stride, int h)
-{
- MOVQ_BFE(mm6);
- JUMPALIGN();
- do {
- __asm__ volatile(
- "movq %1, %%mm0 \n\t"
- "movq %2, %%mm1 \n\t"
- "movq %0, %%mm3 \n\t"
- PAVGB(%%mm0, %%mm1, %%mm2, %%mm6)
- OP_AVG(%%mm3, %%mm2, %%mm0, %%mm6)
- "movq %%mm0, %0 \n\t"
- :"+m"(*dst)
- :"m"(*src1), "m"(*src2)
- :"memory");
- dst += dstStride;
- src1 += src1Stride;
- src2 += 8;
- } while (--h);
-}
-
-static void DEF(avg, pixels16_x2)(uint8_t *block, const uint8_t *pixels, ptrdiff_t line_size, int h)
-{
- MOVQ_BFE(mm6);
- JUMPALIGN();
- do {
- __asm__ volatile(
- "movq %1, %%mm0 \n\t"
- "movq 1%1, %%mm1 \n\t"
- "movq %0, %%mm3 \n\t"
- PAVGB(%%mm0, %%mm1, %%mm2, %%mm6)
- OP_AVG(%%mm3, %%mm2, %%mm0, %%mm6)
- "movq %%mm0, %0 \n\t"
- "movq 8%1, %%mm0 \n\t"
- "movq 9%1, %%mm1 \n\t"
- "movq 8%0, %%mm3 \n\t"
- PAVGB(%%mm0, %%mm1, %%mm2, %%mm6)
- OP_AVG(%%mm3, %%mm2, %%mm0, %%mm6)
- "movq %%mm0, 8%0 \n\t"
- :"+m"(*block)
- :"m"(*pixels)
- :"memory");
- pixels += line_size;
- block += line_size;
- } while (--h);
-}
-
-static av_unused void DEF(avg, pixels16_l2)(uint8_t *dst, uint8_t *src1, uint8_t *src2, int dstStride, int src1Stride, int h)
-{
- MOVQ_BFE(mm6);
- JUMPALIGN();
- do {
- __asm__ volatile(
- "movq %1, %%mm0 \n\t"
- "movq %2, %%mm1 \n\t"
- "movq %0, %%mm3 \n\t"
- PAVGB(%%mm0, %%mm1, %%mm2, %%mm6)
- OP_AVG(%%mm3, %%mm2, %%mm0, %%mm6)
- "movq %%mm0, %0 \n\t"
- "movq 8%1, %%mm0 \n\t"
- "movq 8%2, %%mm1 \n\t"
- "movq 8%0, %%mm3 \n\t"
- PAVGB(%%mm0, %%mm1, %%mm2, %%mm6)
- OP_AVG(%%mm3, %%mm2, %%mm0, %%mm6)
- "movq %%mm0, 8%0 \n\t"
- :"+m"(*dst)
- :"m"(*src1), "m"(*src2)
- :"memory");
- dst += dstStride;
- src1 += src1Stride;
- src2 += 16;
- } while (--h);
-}
-
-static void DEF(avg, pixels8_y2)(uint8_t *block, const uint8_t *pixels, ptrdiff_t line_size, int h)
-{
- MOVQ_BFE(mm6);
- __asm__ volatile(
- "lea (%3, %3), %%"REG_a" \n\t"
- "movq (%1), %%mm0 \n\t"
- ".p2align 3 \n\t"
- "1: \n\t"
- "movq (%1, %3), %%mm1 \n\t"
- "movq (%1, %%"REG_a"), %%mm2 \n\t"
- PAVGBP(%%mm1, %%mm0, %%mm4, %%mm2, %%mm1, %%mm5)
- "movq (%2), %%mm3 \n\t"
- OP_AVG(%%mm3, %%mm4, %%mm0, %%mm6)
- "movq (%2, %3), %%mm3 \n\t"
- OP_AVG(%%mm3, %%mm5, %%mm1, %%mm6)
- "movq %%mm0, (%2) \n\t"
- "movq %%mm1, (%2, %3) \n\t"
- "add %%"REG_a", %1 \n\t"
- "add %%"REG_a", %2 \n\t"
-
- "movq (%1, %3), %%mm1 \n\t"
- "movq (%1, %%"REG_a"), %%mm0 \n\t"
- PAVGBP(%%mm1, %%mm2, %%mm4, %%mm0, %%mm1, %%mm5)
- "movq (%2), %%mm3 \n\t"
- OP_AVG(%%mm3, %%mm4, %%mm2, %%mm6)
- "movq (%2, %3), %%mm3 \n\t"
- OP_AVG(%%mm3, %%mm5, %%mm1, %%mm6)
- "movq %%mm2, (%2) \n\t"
- "movq %%mm1, (%2, %3) \n\t"
- "add %%"REG_a", %1 \n\t"
- "add %%"REG_a", %2 \n\t"
-
- "subl $4, %0 \n\t"
- "jnz 1b \n\t"
- :"+g"(h), "+S"(pixels), "+D"(block)
- :"r"((x86_reg)line_size)
- :REG_a, "memory");
-}
-
// this routine is 'slightly' suboptimal but mostly unused
static void DEF(avg, pixels8_xy2)(uint8_t *block, const uint8_t *pixels, ptrdiff_t line_size, int h)
{
@@ -573,21 +210,11 @@ static void DEF(avg, pixels8_xy2)(uint8_t *block, const uint8_t *pixels, ptrdiff
}
//FIXME optimize
-static void DEF(put, pixels16_y2)(uint8_t *block, const uint8_t *pixels, ptrdiff_t line_size, int h){
- DEF(put, pixels8_y2)(block , pixels , line_size, h);
- DEF(put, pixels8_y2)(block+8, pixels+8, line_size, h);
-}
-
static void DEF(put, pixels16_xy2)(uint8_t *block, const uint8_t *pixels, ptrdiff_t line_size, int h){
DEF(put, pixels8_xy2)(block , pixels , line_size, h);
DEF(put, pixels8_xy2)(block+8, pixels+8, line_size, h);
}
-static void DEF(avg, pixels16_y2)(uint8_t *block, const uint8_t *pixels, ptrdiff_t line_size, int h){
- DEF(avg, pixels8_y2)(block , pixels , line_size, h);
- DEF(avg, pixels8_y2)(block+8, pixels+8, line_size, h);
-}
-
static void DEF(avg, pixels16_xy2)(uint8_t *block, const uint8_t *pixels, ptrdiff_t line_size, int h){
DEF(avg, pixels8_xy2)(block , pixels , line_size, h);
DEF(avg, pixels8_xy2)(block+8, pixels+8, line_size, h);