MLP DSP functions x86-optimized.
authorRamiro Polla <ramiro.polla@gmail.com>
Sat, 23 May 2009 00:23:30 +0000 (00:23 +0000)
committerRamiro Polla <ramiro.polla@gmail.com>
Sat, 23 May 2009 00:23:30 +0000 (00:23 +0000)
12.59% overall speedup in x86_32
 9.98% overall speedup in x86_64
compared to gcc 4.3.3

Originally committed as revision 18903 to svn://svn.ffmpeg.org/ffmpeg/trunk

libavcodec/Makefile
libavcodec/mlpdsp.c
libavcodec/x86/mlpdsp.c [new file with mode: 0644]

index 551576a..18ca9ea 100644 (file)
@@ -445,8 +445,10 @@ MMX-OBJS-$(CONFIG_CAVS_DECODER)        += x86/cavsdsp_mmx.o
 MMX-OBJS-$(CONFIG_ENCODERS)            += x86/dsputilenc_mmx.o
 MMX-OBJS-$(CONFIG_FLAC_ENCODER)        += x86/flacdsp_mmx.o
 MMX-OBJS-$(CONFIG_GPL)                 += x86/idct_mmx.o
+MMX-OBJS-$(CONFIG_MLP_DECODER)         += x86/mlpdsp.o
 MMX-OBJS-$(CONFIG_SNOW_DECODER)        += x86/snowdsp_mmx.o
 MMX-OBJS-$(CONFIG_THEORA_DECODER)      += x86/vp3dsp_mmx.o x86/vp3dsp_sse2.o
+MMX-OBJS-$(CONFIG_TRUEHD_DECODER)      += x86/mlpdsp.o
 MMX-OBJS-$(CONFIG_VC1_DECODER)         += x86/vc1dsp_mmx.o
 MMX-OBJS-$(CONFIG_VP3_DECODER)         += x86/vp3dsp_mmx.o x86/vp3dsp_sse2.o
 MMX-OBJS-$(CONFIG_VP5_DECODER)         += x86/vp3dsp_mmx.o x86/vp3dsp_sse2.o
index 6519b16..79059d9 100644 (file)
@@ -55,7 +55,11 @@ static void ff_mlp_filter_channel(int32_t *state, const int32_t *coeff,
     }
 }
 
+void ff_mlp_init_x86(DSPContext* c, AVCodecContext *avctx);
+
 void ff_mlp_init(DSPContext* c, AVCodecContext *avctx)
 {
     c->mlp_filter_channel = ff_mlp_filter_channel;
+    if (ARCH_X86)
+        ff_mlp_init_x86(c, avctx);
 }
diff --git a/libavcodec/x86/mlpdsp.c b/libavcodec/x86/mlpdsp.c
new file mode 100644 (file)
index 0000000..7577b09
--- /dev/null
@@ -0,0 +1,190 @@
+/*
+ * MLP DSP functions x86-optimized
+ * Copyright (c) 2009 Ramiro Polla
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#include "libavutil/x86_cpu.h"
+#include "libavcodec/dsputil.h"
+#include "libavcodec/mlp.h"
+
+#if HAVE_7REGS && HAVE_TEN_OPERANDS
+
+extern void ff_mlp_firorder_8;
+extern void ff_mlp_firorder_7;
+extern void ff_mlp_firorder_6;
+extern void ff_mlp_firorder_5;
+extern void ff_mlp_firorder_4;
+extern void ff_mlp_firorder_3;
+extern void ff_mlp_firorder_2;
+extern void ff_mlp_firorder_1;
+extern void ff_mlp_firorder_0;
+
+extern void ff_mlp_iirorder_4;
+extern void ff_mlp_iirorder_3;
+extern void ff_mlp_iirorder_2;
+extern void ff_mlp_iirorder_1;
+extern void ff_mlp_iirorder_0;
+
+static const void *firtable[9] = { &ff_mlp_firorder_0, &ff_mlp_firorder_1,
+                                   &ff_mlp_firorder_2, &ff_mlp_firorder_3,
+                                   &ff_mlp_firorder_4, &ff_mlp_firorder_5,
+                                   &ff_mlp_firorder_6, &ff_mlp_firorder_7,
+                                   &ff_mlp_firorder_8 };
+static const void *iirtable[5] = { &ff_mlp_iirorder_0, &ff_mlp_iirorder_1,
+                                   &ff_mlp_iirorder_2, &ff_mlp_iirorder_3,
+                                   &ff_mlp_iirorder_4 };
+
+#if ARCH_X86_64
+
+#define MLPMUL(label, offset, offs, offc)   \
+    MANGLE(label)":                   \n\t" \
+    "movslq "offset"+"offs"(%0), %%rax\n\t" \
+    "movslq "offset"+"offc"(%1), %%rdx\n\t" \
+    "imul                 %%rdx, %%rax\n\t" \
+    "add                  %%rax, %%rsi\n\t"
+
+#define FIRMULREG(label, offset, firc)\
+    MANGLE(label)":             \n\t" \
+    "movslq "#offset"(%0), %%rax\n\t" \
+    "imul        %"#firc", %%rax\n\t" \
+    "add            %%rax, %%rsi\n\t"
+
+#define CLEAR_ACCUM                   \
+    "xor            %%rsi, %%rsi\n\t"
+
+#define SHIFT_ACCUM                   \
+    "shr     %%cl,         %%rsi\n\t"
+
+#define ACCUM    "%%rdx"
+#define RESULT   "%%rsi"
+#define RESULT32 "%%esi"
+
+#define READVAL "r"
+#define RDWRVAL "+r"
+#define COUNTER "c"
+#define ECXUSED
+
+#else /* if ARCH_X86_32 */
+
+#define MLPMUL(label, offset, offs, offc)  \
+    MANGLE(label)":                  \n\t" \
+    "mov   "offset"+"offs"(%0), %%eax\n\t" \
+    "imull "offset"+"offc"(%1)       \n\t" \
+    "add                %%eax , %%esi\n\t" \
+    "adc                %%edx , %%ecx\n\t"
+
+#define FIRMULREG(label, offset, firc)  \
+    MLPMUL(label, #offset, "0", "0")
+
+#define CLEAR_ACCUM                  \
+    "xor           %%esi, %%esi\n\t" \
+    "xor           %%ecx, %%ecx\n\t"
+
+#define SHIFT_ACCUM                  \
+    "mov           %%ecx, %%edx\n\t" \
+    "mov           %%esi, %%eax\n\t" \
+    "movzbl        %7   , %%ecx\n\t" \
+    "shrd    %%cl, %%edx, %%eax\n\t" \
+
+#define ACCUM    "%%edx"
+#define RESULT   "%%eax"
+#define RESULT32 "%%eax"
+
+#define READVAL "m"
+#define RDWRVAL "+m"
+#define COUNTER "m"
+#define ECXUSED , "ecx"
+
+#endif /* !ARCH_X86_64 */
+
+#define BINC  AV_STRINGIFY(4* MAX_CHANNELS)
+#define IOFFS AV_STRINGIFY(4*(MAX_FIR_ORDER + MAX_BLOCKSIZE))
+#define IOFFC AV_STRINGIFY(4* MAX_FIR_ORDER)
+
+#define FIRMUL(label, offset) MLPMUL(label, #offset,   "0",   "0")
+#define IIRMUL(label, offset) MLPMUL(label, #offset, IOFFS, IOFFC)
+
+static void mlp_filter_channel_x86(int32_t *state, const int32_t *coeff,
+                                   int firorder, int iirorder,
+                                   unsigned int filter_shift, int32_t mask,
+                                   int blocksize, int32_t *sample_buffer)
+{
+    const void *firjump = firtable[firorder];
+    const void *iirjump = iirtable[iirorder];
+
+    blocksize = -blocksize;
+
+    __asm__ volatile(
+        "1:                           \n\t"
+        CLEAR_ACCUM
+        "jmp  *%5                     \n\t"
+        FIRMUL   (ff_mlp_firorder_8, 0x1c   )
+        FIRMUL   (ff_mlp_firorder_7, 0x18   )
+        FIRMUL   (ff_mlp_firorder_6, 0x14   )
+        FIRMUL   (ff_mlp_firorder_5, 0x10   )
+        FIRMUL   (ff_mlp_firorder_4, 0x0c   )
+        FIRMULREG(ff_mlp_firorder_3, 0x08,10)
+        FIRMULREG(ff_mlp_firorder_2, 0x04, 9)
+        FIRMULREG(ff_mlp_firorder_1, 0x00, 8)
+        MANGLE   (ff_mlp_firorder_0)":\n\t"
+        "jmp  *%6                     \n\t"
+        IIRMUL   (ff_mlp_iirorder_4, 0x0c   )
+        IIRMUL   (ff_mlp_iirorder_3, 0x08   )
+        IIRMUL   (ff_mlp_iirorder_2, 0x04   )
+        IIRMUL   (ff_mlp_iirorder_1, 0x00   )
+        MANGLE   (ff_mlp_iirorder_0)":\n\t"
+        SHIFT_ACCUM
+        "mov  "RESULT"  ,"ACCUM"      \n\t"
+        "add  (%2)      ,"RESULT"     \n\t"
+        "and   %4       ,"RESULT"     \n\t"
+        "sub   $4       ,  %0         \n\t"
+        "mov  "RESULT32", (%0)        \n\t"
+        "mov  "RESULT32", (%2)        \n\t"
+        "add $"BINC"    ,  %2         \n\t"
+        "sub  "ACCUM"   ,"RESULT"     \n\t"
+        "mov  "RESULT32","IOFFS"(%0)  \n\t"
+        "incl              %3         \n\t"
+        "js 1b                        \n\t"
+        : /* 0*/"+r"(state),
+          /* 1*/"+r"(coeff),
+          /* 2*/"+r"(sample_buffer),
+          /* 3*/RDWRVAL(blocksize)
+        :
+          /* 4*/READVAL((x86_reg)mask),
+          /* 5*/READVAL(firjump),
+          /* 6*/READVAL(iirjump),
+          /* 7*/COUNTER(filter_shift)
+#if ARCH_X86_64
+        , /* 8*/"r"((int64_t)coeff[0])
+        , /* 9*/"r"((int64_t)coeff[1])
+        , /*10*/"r"((int64_t)coeff[2])
+#endif /* ARCH_X86_64 */
+        : REG_a, REG_d, REG_S
+          ECXUSED
+    );
+}
+
+#endif /* HAVE_7REGS && HAVE_TEN_OPERANDS */
+
+void ff_mlp_init_x86(DSPContext* c, AVCodecContext *avctx)
+{
+#if HAVE_7REGS && HAVE_TEN_OPERANDS
+    c->mlp_filter_channel = mlp_filter_channel_x86;
+#endif
+}