On Mon, Apr 14, 2014 at 03:13:22PM +1000, Matt Oliver wrote:

This is a patch to enable intel compiler on Windows to compile the libmpcodec inline asm functions. It is based of similar work previously done in FFmpeg (and in fact was initially posted there) and extends some previous patches that were applied to libmpcodecs a few months ago. This is required to compile ffmpeg under icl with asm and was previously posted (probably incorrectly) on the ffmpeg mailing list but im now posting it here in the correct location.

Matt

vf_fspp.c | 18 ++++++++++++------ 1 file changed, 12 insertions(+), 6 deletions(-) b54f42435c23df435d2243e8e11f2be539ee014c 4-4-Fix-libmpcodecs-inline-asm-on-ICL-part-2.patch From 0a314fb3faa063a008e654862375e2ef8a674658 Mon Sep 17 00:00:00 2001 From: Matt Oliver <protogonoi@gmail.com> Date: Sun, 9 Feb 2014 22:25:15 +1100 Subject: [PATCH] Fix libmpcodecs inline asm on ICL.

--- libavfilter/libmpcodecs/vf_fspp.c | 18 ++++++++++++------ 1 file changed, 12 insertions(+), 6 deletions(-)

diff --git a/libavfilter/libmpcodecs/vf_fspp.c b/libavfilter/libmpcodecs/vf_fspp.c index dadfae7..65af77d 100644 --- a/libavfilter/libmpcodecs/vf_fspp.c +++ b/libavfilter/libmpcodecs/vf_fspp.c @@ -1598,6 +1598,10 @@ static void column_fidct_mmx(int16_t* thr_adr, int16_t *data, int16_t *output,

: "+S"(data), "+D"(output), "+c"(cnt), "=o"(temps) : "d"(thr_adr) + NAMED_CONSTRAINTS_ADD(ff_MM_FIX_0_707106781,MM_2,MM_FIX_1_414213562_A,MM_FIX_1_414213562,MM_FIX_0_382683433, + ff_MM_FIX_0_541196100,MM_FIX_1_306562965,MM_FIX_0_847759065) + NAMED_CONSTRAINTS_ADD(MM_FIX_0_566454497,MM_FIX_0_198912367,MM_FIX_2_613125930,MM_FIX_1_847759065, + MM_FIX_1_082392200,ff_MM_FIX_0_541196100,MM_FIX_1_306562965) : "%"REG_a ); } @@ -1867,6 +1871,8 @@ static void row_idct_mmx (int16_t* workspace,

: "+S"(workspace), "+D"(output_adr), "+c"(cnt), "=o"(temps) : "a"(output_stride*sizeof(short)) + NAMED_CONSTRAINTS_ADD(MM_FIX_1_414213562_A,MM_FIX_2_613125930,MM_FIX_1_847759065,MM_FIX_1_082392200, + MM_FIX_1_414213562,MM_DESCALE_RND) : "%"REG_d ); } @@ -1974,10 +1980,10 @@ static void row_fdct_mmx(int16_t *data, const uint8_t *pixels, int line_size, "movd (%%"REG_S",%%"REG_a",2), %%mm3 \n\t" //5 "paddw %%mm4, %%mm1 \n\t"

- "movq %%mm5, 0*8+%3 \n\t" //t7 + "movq %%mm5, %3 \n\t" //t7 "punpcklbw %%mm7, %%mm3 \n\t"

- "movq %%mm6, 1*8+%3 \n\t" //t6 + "movq %%mm6, %4 \n\t" //t6 "movq %%mm2, %%mm4 \n\t"

"movd (%%"REG_S"), %%mm5 \n\t" //3 @@ -2023,7 +2029,7 @@ static void row_fdct_mmx(int16_t *data, const uint8_t *pixels, int line_size, "psubw %%mm1, %%mm5 \n\t" //d1 "movq %%mm0, %%mm6 \n\t"

- "movq 1*8+%3, %%mm1 \n\t" + "movq %4, %%mm1 \n\t" "punpcklwd %%mm5, %%mm0 \n\t"

"punpckhwd %%mm5, %%mm6 \n\t" @@ -2047,7 +2053,7 @@ static void row_fdct_mmx(int16_t *data, const uint8_t *pixels, int line_size, "movq %%mm7, "DCTSIZE_S"*3*2(%%"REG_D") \n\t" "psllw $2, %%mm3 \n\t" //t10

- "movq 0*8+%3, %%mm2 \n\t" + "movq %3, %%mm2 \n\t" "psllw $2, %%mm4 \n\t" //t11

"pmulhw "MANGLE(ff_MM_FIX_0_707106781)", %%mm4 \n\t" //z3

@@ -2110,8 +2116,8 @@ static void row_fdct_mmx(int16_t *data, const uint8_t *pixels, int line_size, "dec %%"REG_c" \n\t" "jnz 6b \n\t"

- : "+S"(pixels), "+D"(data), "+c"(cnt), "=o"(temps) - : "a"(line_size) + : "+S"(pixels), "+D"(data), "+c"(cnt), "=o"(temps), "=o"(temps[1]) + : NAMED_CONSTRAINTS(ff_MM_FIX_0_707106781,ff_MM_FIX_0_541196100,MM_FIX_0_382683433,MM_FIX_1_306562965) : "%"REG_d);

line_size is lost here, this doesnt work make sure you test the code ... [...] -- Michael GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB Asymptotically faster algorithms should always be preferred if you have asymptotical amounts of data