x86/flacdsp: add SSE2 and AVX decorrelate functions

Two to four times faster depending on instruction set, block size and channel count.

7 files changed, 328 insertions, 12 deletions

diff --git a/libavcodec/arm/flacdsp_init_arm.c b/libavcodec/arm/flacdsp_init_arm.c
index 9ddb26829d..df1b19c1e8 100644
--- a/libavcodec/arm/flacdsp_init_arm.c
+++ b/libavcodec/arm/flacdsp_init_arm.c
@@ -24,7 +24,7 @@
 void ff_flac_lpc_16_arm(int32_t *samples, const int coeffs[32], int order,
                         int qlevel, int len);
 
-av_cold void ff_flacdsp_init_arm(FLACDSPContext *c, enum AVSampleFormat fmt,
+av_cold void ff_flacdsp_init_arm(FLACDSPContext *c, enum AVSampleFormat fmt, int channels,
                                  int bps)
 {
     if (bps <= 16 && CONFIG_FLAC_DECODER)
diff --git a/libavcodec/flacdec.c b/libavcodec/flacdec.c
index f1316658e7..9b1e4280b6 100644
--- a/libavcodec/flacdec.c
+++ b/libavcodec/flacdec.c
@@ -111,7 +111,7 @@ static av_cold int flac_decode_init(AVCodecContext *avctx)
     if (ret < 0)
         return ret;
     flac_set_bps(s);
-    ff_flacdsp_init(&s->dsp, avctx->sample_fmt, s->bps);
+    ff_flacdsp_init(&s->dsp, avctx->sample_fmt, s->channels, s->bps);
     s->got_streaminfo = 1;
 
     return 0;
@@ -173,7 +173,7 @@ static int parse_streaminfo(FLACContext *s, const uint8_t *buf, int buf_size)
     if (ret < 0)
         return ret;
     flac_set_bps(s);
-    ff_flacdsp_init(&s->dsp, s->avctx->sample_fmt, s->bps);
+    ff_flacdsp_init(&s->dsp, s->avctx->sample_fmt, s->channels, s->bps);
     s->got_streaminfo = 1;
 
     return 0;
@@ -472,7 +472,7 @@ static int decode_frame(FLACContext *s)
         ret = allocate_buffers(s);
         if (ret < 0)
             return ret;
-        ff_flacdsp_init(&s->dsp, s->avctx->sample_fmt, s->bps);
+        ff_flacdsp_init(&s->dsp, s->avctx->sample_fmt, s->channels, s->bps);
         s->got_streaminfo = 1;
         dump_headers(s->avctx, (FLACStreaminfo *)s);
     }
diff --git a/libavcodec/flacdsp.c b/libavcodec/flacdsp.c
index b15bc7476b..a83eb833b5 100644
--- a/libavcodec/flacdsp.c
+++ b/libavcodec/flacdsp.c
@@ -85,7 +85,7 @@ static void flac_lpc_32_c(int32_t *decoded, const int coeffs[32],
 
 }
 
-av_cold void ff_flacdsp_init(FLACDSPContext *c, enum AVSampleFormat fmt,
+av_cold void ff_flacdsp_init(FLACDSPContext *c, enum AVSampleFormat fmt, int channels,
                              int bps)
 {
     if (bps > 16) {
@@ -127,7 +127,7 @@ av_cold void ff_flacdsp_init(FLACDSPContext *c, enum AVSampleFormat fmt,
     }
 
     if (ARCH_ARM)
-        ff_flacdsp_init_arm(c, fmt, bps);
+        ff_flacdsp_init_arm(c, fmt, channels, bps);
     if (ARCH_X86)
-        ff_flacdsp_init_x86(c, fmt, bps);
+        ff_flacdsp_init_x86(c, fmt, channels, bps);
 }
diff --git a/libavcodec/flacdsp.h b/libavcodec/flacdsp.h
index 14f3466b0b..417381cc70 100644
--- a/libavcodec/flacdsp.h
+++ b/libavcodec/flacdsp.h
@@ -31,8 +31,8 @@ typedef struct FLACDSPContext {
                        const int32_t coefs[32], int shift);
 } FLACDSPContext;
 
-void ff_flacdsp_init(FLACDSPContext *c, enum AVSampleFormat fmt, int bps);
-void ff_flacdsp_init_arm(FLACDSPContext *c, enum AVSampleFormat fmt, int bps);
-void ff_flacdsp_init_x86(FLACDSPContext *c, enum AVSampleFormat fmt, int bps);
+void ff_flacdsp_init(FLACDSPContext *c, enum AVSampleFormat fmt, int channels, int bps);
+void ff_flacdsp_init_arm(FLACDSPContext *c, enum AVSampleFormat fmt, int channels, int bps);
+void ff_flacdsp_init_x86(FLACDSPContext *c, enum AVSampleFormat fmt, int channels, int bps);
 
 #endif /* AVCODEC_FLACDSP_H */
diff --git a/libavcodec/flacenc.c b/libavcodec/flacenc.c
index 3b72888966..e66ef3db5e 100644
--- a/libavcodec/flacenc.c
+++ b/libavcodec/flacenc.c
@@ -428,7 +428,7 @@ static av_cold int flac_encode_init(AVCodecContext *avctx)
                       s->options.max_prediction_order, FF_LPC_TYPE_LEVINSON);
 
     ff_bswapdsp_init(&s->bdsp);
-    ff_flacdsp_init(&s->flac_dsp, avctx->sample_fmt,
+    ff_flacdsp_init(&s->flac_dsp, avctx->sample_fmt, channels,
                     avctx->bits_per_raw_sample);
 
     dprint_compression_options(s);
diff --git a/libavcodec/x86/flacdsp.asm b/libavcodec/x86/flacdsp.asm
index 37ee87b163..d60a6e4a06 100644
--- a/libavcodec/x86/flacdsp.asm
+++ b/libavcodec/x86/flacdsp.asm
@@ -2,6 +2,7 @@
 ;* FLAC DSP SIMD optimizations
 ;*
 ;* Copyright (C) 2014 Loren Merritt
+;* Copyright (C) 2014 James Almer
 ;*
 ;* This file is part of FFmpeg.
 ;*
@@ -72,3 +73,257 @@ ALIGN 16
 LPC_32 xop
 %endif
 LPC_32 sse4
+
+;----------------------------------------------------------------------------------
+;void ff_flac_decorrelate_[lrm]s_16_sse2(uint8_t **out, int32_t **in, int channels,
+;                                                   int len, int shift);
+;----------------------------------------------------------------------------------
+%macro FLAC_DECORRELATE_16 3-4
+cglobal flac_decorrelate_%1_16, 2, 4, 4, out, in0, in1, len
+%if ARCH_X86_32 || WIN64
+    movd       m3, r4m
+%if ARCH_X86_32
+    mov      lend, lenm
+%endif
+%else ; UNIX64
+    movd       m3, r4d
+%endif
+    shl      lend, 2
+    mov      in1q, [in0q + gprsize]
+    mov      in0q, [in0q]
+    mov      outq, [outq]
+    add      in1q, lenq
+    add      in0q, lenq
+    add      outq, lenq
+    neg      lenq
+
+align 16
+.loop:
+    mova       m0, [in0q + lenq]
+    mova       m1, [in1q + lenq]
+%ifidn %1, ms
+    psrad      m2, m1, 1
+    psubd      m0, m2
+%endif
+%ifnidn %1, indep2
+    p%4d       m2, m0, m1
+%endif
+    packssdw  m%2, m%2
+    packssdw  m%3, m%3
+    punpcklwd m%2, m%3
+    psllw     m%2, m3
+    mova [outq + lenq], m%2
+    add      lenq, 16
+    jl .loop
+    REP_RET
+%endmacro
+
+INIT_XMM sse2
+FLAC_DECORRELATE_16 ls, 0, 2, sub
+FLAC_DECORRELATE_16 rs, 2, 1, add
+FLAC_DECORRELATE_16 ms, 2, 0, add
+
+;----------------------------------------------------------------------------------
+;void ff_flac_decorrelate_[lrm]s_32_sse2(uint8_t **out, int32_t **in, int channels,
+;                                        int len, int shift);
+;----------------------------------------------------------------------------------
+%macro FLAC_DECORRELATE_32 5
+cglobal flac_decorrelate_%1_32, 2, 4, 4, out, in0, in1, len
+%if ARCH_X86_32 || WIN64
+    movd       m3, r4m
+%if ARCH_X86_32
+    mov      lend, lenm
+%endif
+%else ; UNIX64
+    movd       m3, r4d
+%endif
+    mov      in1q, [in0q + gprsize]
+    mov      in0q, [in0q]
+    mov      outq, [outq]
+    sub      in1q, in0q
+
+align 16
+.loop:
+    mova       m0, [in0q]
+    mova       m1, [in0q + in1q]
+%ifidn %1, ms
+    psrad      m2, m1, 1
+    psubd      m0, m2
+%endif
+    p%5d       m2, m0, m1
+    pslld     m%2, m3
+    pslld     m%3, m3
+
+    SBUTTERFLY dq, %2, %3, %4
+
+    mova  [outq         ], m%2
+    mova  [outq + mmsize], m%3
+
+    add      in0q, mmsize
+    add      outq, mmsize*2
+    sub      lend, mmsize/4
+    jg .loop
+    REP_RET
+%endmacro
+
+INIT_XMM sse2
+FLAC_DECORRELATE_32 ls, 0, 2, 1, sub
+FLAC_DECORRELATE_32 rs, 2, 1, 0, add
+FLAC_DECORRELATE_32 ms, 2, 0, 1, add
+
+;-----------------------------------------------------------------------------------------
+;void ff_flac_decorrelate_indep<ch>_<bps>_<opt>(uint8_t **out, int32_t **in, int channels,
+;                                            int len, int shift);
+;-----------------------------------------------------------------------------------------
+%macro TRANSPOSE8x4D 9
+    SBUTTERFLY dq,  %1, %2, %9
+    SBUTTERFLY dq,  %3, %4, %9
+    SBUTTERFLY dq,  %5, %6, %9
+    SBUTTERFLY dq,  %7, %8, %9
+    SBUTTERFLY qdq, %1, %3, %9
+    SBUTTERFLY qdq, %2, %4, %9
+    SBUTTERFLY qdq, %5, %7, %9
+    SBUTTERFLY qdq, %6, %8, %9
+    SWAP %2, %5
+    SWAP %4, %7
+%endmacro
+
+;%1 = bps
+;%2 = channels
+;%3 = last xmm reg used
+;%4 = word/dword (shift instruction)
+%macro FLAC_DECORRELATE_INDEP 4
+%define REPCOUNT %2/(32/%1) ; 16bits = channels / 2; 32bits = channels
+cglobal flac_decorrelate_indep%2_%1, 2, %2+2, %3+1, out, in0, in1, len, in2, in3, in4, in5, in6, in7
+%if ARCH_X86_32
+    movd      m%3, r4m
+%if %2 == 6
+    DEFINE_ARGS out, in0, in1, in2, in3, in4, in5
+    %define  lend  dword r3m
+%else
+    mov      lend, lenm
+%endif
+%elif WIN64
+    movd      m%3, r4m
+%else ; UNIX64
+    movd      m%3, r4d
+%endif
+
+%assign %%i 1
+%rep %2-1
+    mov      in %+ %%i %+ q, [in0q+%%i*gprsize]
+%assign %%i %%i+1
+%endrep
+
+    mov      in0q, [in0q]
+    mov      outq, [outq]
+
+%assign %%i 1
+%rep %2-1
+    sub      in %+ %%i %+ q, in0q
+%assign %%i %%i+1
+%endrep
+
+align 16
+.loop:
+    mova       m0, [in0q]
+
+%assign %%i 1
+%rep REPCOUNT-1
+    mova     m %+ %%i, [in0q + in %+ %%i %+ q]
+%assign %%i %%i+1
+%endrep
+
+%if %1 == 32
+
+%if %2 == 8
+    TRANSPOSE8x4D 0, 1, 2, 3, 4, 5, 6, 7, 8
+%elif %2 == 6
+    SBUTTERFLY dq, 0, 1, 6
+    SBUTTERFLY dq, 2, 3, 6
+    SBUTTERFLY dq, 4, 5, 6
+
+    punpcklqdq m6, m0, m2
+    punpckhqdq m2, m4
+    shufps     m4, m0, 0xe4
+    punpcklqdq m0, m1, m3
+    punpckhqdq m3, m5
+    shufps     m5, m1, 0xe4
+    SWAP 0,6,1,4,5,3
+%elif %2 == 4
+    TRANSPOSE4x4D 0, 1, 2, 3, 4
+%else ; %2 == 2
+    SBUTTERFLY dq, 0, 1, 2
+%endif
+
+%else ; %1 == 16
+
+%if %2 == 8
+    packssdw   m0, [in0q + in4q]
+    packssdw   m1, [in0q + in5q]
+    packssdw   m2, [in0q + in6q]
+    packssdw   m3, [in0q + in7q]
+    TRANSPOSE2x4x4W 0, 1, 2, 3, 4
+%elif %2 == 6
+    packssdw   m0, [in0q + in3q]
+    packssdw   m1, [in0q + in4q]
+    packssdw   m2, [in0q + in5q]
+    pshufd     m3, m0,     q1032
+    punpcklwd  m0, m1
+    punpckhwd  m1, m2
+    punpcklwd  m2, m3
+
+    shufps     m3, m0, m2, q2020
+    shufps     m0, m1,     q2031
+    shufps     m2, m1,     q3131
+    shufps     m1, m2, m3, q3120
+    shufps     m3, m0,     q0220
+    shufps     m0, m2,     q3113
+    SWAP 2, 0, 3
+%else ; %2 == 4
+    packssdw   m0, [in0q + in2q]
+    packssdw   m1, [in0q + in3q]
+    SBUTTERFLY wd, 0, 1, 2
+    SBUTTERFLY dq, 0, 1, 2
+%endif
+
+%endif
+
+%assign %%i 0
+%rep REPCOUNT
+    psll%4   m %+ %%i, m%3
+%assign %%i %%i+1
+%endrep
+
+%assign %%i 0
+%rep REPCOUNT
+    mova [outq + %%i*mmsize], m %+ %%i
+%assign %%i %%i+1
+%endrep
+
+    add      in0q, mmsize
+    add      outq, mmsize*REPCOUNT
+    sub      lend, mmsize/4
+    jg .loop
+    REP_RET
+%endmacro
+
+INIT_XMM sse2
+FLAC_DECORRELATE_16 indep2, 0, 1 ; Reuse stereo 16bits macro
+FLAC_DECORRELATE_INDEP 32, 2, 3, d
+FLAC_DECORRELATE_INDEP 16, 4, 3, w
+FLAC_DECORRELATE_INDEP 32, 4, 5, d
+FLAC_DECORRELATE_INDEP 16, 6, 4, w
+FLAC_DECORRELATE_INDEP 32, 6, 7, d
+%if ARCH_X86_64
+FLAC_DECORRELATE_INDEP 16, 8, 5, w
+FLAC_DECORRELATE_INDEP 32, 8, 9, d
+%endif
+
+INIT_XMM avx
+FLAC_DECORRELATE_INDEP 32, 4, 5, d
+FLAC_DECORRELATE_INDEP 32, 6, 7, d
+%if ARCH_X86_64
+FLAC_DECORRELATE_INDEP 16, 8, 5, w
+FLAC_DECORRELATE_INDEP 32, 8, 9, d
+%endif
diff --git a/libavcodec/x86/flacdsp_init.c b/libavcodec/x86/flacdsp_init.c
index 33b31f12e4..d04af4511d 100644
--- a/libavcodec/x86/flacdsp_init.c
+++ b/libavcodec/x86/flacdsp_init.c
@@ -29,17 +29,78 @@ void ff_flac_lpc_32_xop(int32_t *samples, const int coeffs[32], int order,
 
 void ff_flac_enc_lpc_16_sse4(int32_t *, const int32_t *, int, int, const int32_t *,int);
 
-av_cold void ff_flacdsp_init_x86(FLACDSPContext *c, enum AVSampleFormat fmt,
+#define DECORRELATE_FUNCS(fmt, opt)                                                      \
+void ff_flac_decorrelate_ls_##fmt##_##opt(uint8_t **out, int32_t **in, int channels,     \
+                                          int len, int shift);                           \
+void ff_flac_decorrelate_rs_##fmt##_##opt(uint8_t **out, int32_t **in, int channels,     \
+                                          int len, int shift);                           \
+void ff_flac_decorrelate_ms_##fmt##_##opt(uint8_t **out, int32_t **in, int channels,     \
+                                          int len, int shift);                           \
+void ff_flac_decorrelate_indep2_##fmt##_##opt(uint8_t **out, int32_t **in, int channels, \
+                                             int len, int shift);                        \
+void ff_flac_decorrelate_indep4_##fmt##_##opt(uint8_t **out, int32_t **in, int channels, \
+                                              int len, int shift);                       \
+void ff_flac_decorrelate_indep6_##fmt##_##opt(uint8_t **out, int32_t **in, int channels, \
+                                              int len, int shift);                       \
+void ff_flac_decorrelate_indep8_##fmt##_##opt(uint8_t **out, int32_t **in, int channels, \
+                                              int len, int shift)
+
+DECORRELATE_FUNCS(16, sse2);
+DECORRELATE_FUNCS(16,  avx);
+DECORRELATE_FUNCS(32, sse2);
+DECORRELATE_FUNCS(32,  avx);
+
+av_cold void ff_flacdsp_init_x86(FLACDSPContext *c, enum AVSampleFormat fmt, int channels,
                                  int bps)
 {
 #if HAVE_YASM
     int cpu_flags = av_get_cpu_flags();
 
 #if CONFIG_FLAC_DECODER
+    if (EXTERNAL_SSE2(cpu_flags)) {
+        if (fmt == AV_SAMPLE_FMT_S16) {
+            if (channels == 2)
+                c->decorrelate[0] = ff_flac_decorrelate_indep2_16_sse2;
+            else if (channels == 4)
+                c->decorrelate[0] = ff_flac_decorrelate_indep4_16_sse2;
+            else if (channels == 6)
+                c->decorrelate[0] = ff_flac_decorrelate_indep6_16_sse2;
+            else if (ARCH_X86_64 && channels == 8)
+                c->decorrelate[0] = ff_flac_decorrelate_indep8_16_sse2;
+            c->decorrelate[1] = ff_flac_decorrelate_ls_16_sse2;
+            c->decorrelate[2] = ff_flac_decorrelate_rs_16_sse2;
+            c->decorrelate[3] = ff_flac_decorrelate_ms_16_sse2;
+        } else if (fmt == AV_SAMPLE_FMT_S32) {
+            if (channels == 2)
+                c->decorrelate[0] = ff_flac_decorrelate_indep2_32_sse2;
+            else if (channels == 4)
+                c->decorrelate[0] = ff_flac_decorrelate_indep4_32_sse2;
+            else if (channels == 6)
+                c->decorrelate[0] = ff_flac_decorrelate_indep6_32_sse2;
+            else if (ARCH_X86_64 && channels == 8)
+                c->decorrelate[0] = ff_flac_decorrelate_indep8_32_sse2;
+            c->decorrelate[1] = ff_flac_decorrelate_ls_32_sse2;
+            c->decorrelate[2] = ff_flac_decorrelate_rs_32_sse2;
+            c->decorrelate[3] = ff_flac_decorrelate_ms_32_sse2;
+        }
+    }
     if (EXTERNAL_SSE4(cpu_flags)) {
         if (bps > 16)
             c->lpc = ff_flac_lpc_32_sse4;
     }
+    if (EXTERNAL_AVX(cpu_flags)) {
+        if (fmt == AV_SAMPLE_FMT_S16) {
+            if (ARCH_X86_64 && channels == 8)
+                c->decorrelate[0] = ff_flac_decorrelate_indep8_16_avx;
+        } else if (fmt == AV_SAMPLE_FMT_S32) {
+            if (channels == 4)
+                c->decorrelate[0] = ff_flac_decorrelate_indep4_32_avx;
+            else if (channels == 6)
+                c->decorrelate[0] = ff_flac_decorrelate_indep6_32_avx;
+            else if (ARCH_X86_64 && channels == 8)
+                c->decorrelate[0] = ff_flac_decorrelate_indep8_32_avx;
+        }
+    }
     if (EXTERNAL_XOP(cpu_flags)) {
         if (bps > 16)
             c->lpc = ff_flac_lpc_32_xop;