ac3enc: add SIMD-optimized shifting functions for use with the fixed-point AC3 encoder.

5 files changed, 120 insertions, 37 deletions

diff --git a/libavcodec/ac3dsp.c b/libavcodec/ac3dsp.c
index da3a123e9b..4d9db9be50 100644
--- a/libavcodec/ac3dsp.c
+++ b/libavcodec/ac3dsp.c
@@ -50,10 +50,47 @@ static int ac3_max_msb_abs_int16_c(const int16_t *src, int len)
     return v;
 }
 
+static void ac3_lshift_int16_c(int16_t *src, unsigned int len,
+                               unsigned int shift)
+{
+    uint32_t *src32 = (uint32_t *)src;
+    const uint32_t mask = ~(((1 << shift) - 1) << 16);
+    int i;
+    len >>= 1;
+    for (i = 0; i < len; i += 8) {
+        src32[i  ] = (src32[i  ] << shift) & mask;
+        src32[i+1] = (src32[i+1] << shift) & mask;
+        src32[i+2] = (src32[i+2] << shift) & mask;
+        src32[i+3] = (src32[i+3] << shift) & mask;
+        src32[i+4] = (src32[i+4] << shift) & mask;
+        src32[i+5] = (src32[i+5] << shift) & mask;
+        src32[i+6] = (src32[i+6] << shift) & mask;
+        src32[i+7] = (src32[i+7] << shift) & mask;
+    }
+}
+
+static void ac3_rshift_int32_c(int32_t *src, unsigned int len,
+                               unsigned int shift)
+{
+    do {
+        *src++ >>= shift;
+        *src++ >>= shift;
+        *src++ >>= shift;
+        *src++ >>= shift;
+        *src++ >>= shift;
+        *src++ >>= shift;
+        *src++ >>= shift;
+        *src++ >>= shift;
+        len -= 8;
+    } while (len > 0);
+}
+
 av_cold void ff_ac3dsp_init(AC3DSPContext *c)
 {
     c->ac3_exponent_min = ac3_exponent_min_c;
     c->ac3_max_msb_abs_int16 = ac3_max_msb_abs_int16_c;
+    c->ac3_lshift_int16 = ac3_lshift_int16_c;
+    c->ac3_rshift_int32 = ac3_rshift_int32_c;
 
     if (HAVE_MMX)
         ff_ac3dsp_init_x86(c);
diff --git a/libavcodec/ac3dsp.h b/libavcodec/ac3dsp.h
index a4f141f2f5..31a0af375d 100644
--- a/libavcodec/ac3dsp.h
+++ b/libavcodec/ac3dsp.h
@@ -46,6 +46,28 @@ typedef struct AC3DSPContext {
      * @return    a value with the same MSB as max(abs(src[]))
      */
     int (*ac3_max_msb_abs_int16)(const int16_t *src, int len);
+
+    /**
+     * Left-shift each value in an array of int16_t by a specified amount.
+     * @param src    input array
+     *               constraints: align 16
+     * @param len    number of values in the array
+     *               constraints: multiple of 32 greater than 0
+     * @param shift  left shift amount
+     *               constraints: range [0,15]
+     */
+    void (*ac3_lshift_int16)(int16_t *src, unsigned int len, unsigned int shift);
+
+    /**
+     * Right-shift each value in an array of int32_t by a specified amount.
+     * @param src    input array
+     *               constraints: align 16
+     * @param len    number of values in the array
+     *               constraints: multiple of 16 greater than 0
+     * @param shift  right shift amount
+     *               constraints: range [0,31]
+     */
+    void (*ac3_rshift_int32)(int32_t *src, unsigned int len, unsigned int shift);
 } AC3DSPContext;
 
 void ff_ac3dsp_init    (AC3DSPContext *c);
diff --git a/libavcodec/ac3enc_fixed.c b/libavcodec/ac3enc_fixed.c
index e750a39038..d1bb429ccc 100644
--- a/libavcodec/ac3enc_fixed.c
+++ b/libavcodec/ac3enc_fixed.c
@@ -278,40 +278,6 @@ static int log2_tab(AC3EncodeContext *s, int16_t *src, int len)
 
 
 /**
- * Left-shift each value in an array by a specified amount.
- * @param tab    input array
- * @param n      number of values in the array
- * @param lshift left shift amount
- */
-static void lshift_tab(int16_t *tab, int n, unsigned int lshift)
-{
-    int i;
-
-    if (lshift > 0) {
-        for (i = 0; i < n; i++)
-            tab[i] <<= lshift;
-    }
-}
-
-
-/**
- * Right-shift each value in an array of int32_t by a specified amount.
- * @param src    input array
- * @param len    number of values in the array
- * @param shift  right shift amount
- */
-static void ac3_rshift_int32_c(int32_t *src, unsigned int len, unsigned int shift)
-{
-    int i;
-
-    if (shift > 0) {
-        for (i = 0; i < len; i++)
-            src[i] >>= shift;
-    }
-}
-
-
-/**
  * Normalize the input samples to use the maximum available precision.
  * This assumes signed 16-bit input samples.
  *
@@ -320,7 +286,8 @@ static void ac3_rshift_int32_c(int32_t *src, unsigned int len, unsigned int shif
 static int normalize_samples(AC3EncodeContext *s)
 {
     int v = 14 - log2_tab(s, s->windowed_samples, AC3_WINDOW_SIZE);
-    lshift_tab(s->windowed_samples, AC3_WINDOW_SIZE, v);
+    if (v > 0)
+        s->ac3dsp.ac3_lshift_int16(s->windowed_samples, AC3_WINDOW_SIZE, v);
     /* +6 to right-shift from 31-bit to 25-bit */
     return v + 6;
 }
@@ -336,8 +303,8 @@ static void scale_coefficients(AC3EncodeContext *s)
     for (blk = 0; blk < AC3_MAX_BLOCKS; blk++) {
         AC3Block *block = &s->blocks[blk];
         for (ch = 0; ch < s->channels; ch++) {
-            ac3_rshift_int32_c(block->mdct_coef[ch], AC3_MAX_COEFS,
-                               block->coeff_shift[ch]);
+            s->ac3dsp.ac3_rshift_int32(block->mdct_coef[ch], AC3_MAX_COEFS,
+                                       block->coeff_shift[ch]);
         }
     }
 }
diff --git a/libavcodec/x86/ac3dsp.asm b/libavcodec/x86/ac3dsp.asm
index b1eeca9cd2..e281791b1e 100644
--- a/libavcodec/x86/ac3dsp.asm
+++ b/libavcodec/x86/ac3dsp.asm
@@ -133,3 +133,48 @@ INIT_XMM
 AC3_MAX_MSB_ABS_INT16 sse2, min_max
 %define ABS2 ABS2_SSSE3
 AC3_MAX_MSB_ABS_INT16 ssse3, or_abs
+
+;-----------------------------------------------------------------------------
+; macro used for ff_ac3_lshift_int16() and ff_ac3_rshift_int32()
+;-----------------------------------------------------------------------------
+
+%macro AC3_SHIFT 4 ; l/r, 16/32, shift instruction, instruction set
+cglobal ac3_%1shift_int%2_%4, 3,3,5, src, len, shift
+    movd      m0, shiftd
+.loop:
+    mova      m1, [srcq         ]
+    mova      m2, [srcq+mmsize  ]
+    mova      m3, [srcq+mmsize*2]
+    mova      m4, [srcq+mmsize*3]
+    %3        m1, m0
+    %3        m2, m0
+    %3        m3, m0
+    %3        m4, m0
+    mova  [srcq         ], m1
+    mova  [srcq+mmsize  ], m2
+    mova  [srcq+mmsize*2], m3
+    mova  [srcq+mmsize*3], m4
+    add     srcq, mmsize*4
+    sub     lend, mmsize*32/%2
+    ja .loop
+.end:
+    REP_RET
+%endmacro
+
+;-----------------------------------------------------------------------------
+; void ff_ac3_lshift_int16(int16_t *src, unsigned int len, unsigned int shift)
+;-----------------------------------------------------------------------------
+
+INIT_MMX
+AC3_SHIFT l, 16, psllw, mmx
+INIT_XMM
+AC3_SHIFT l, 16, psllw, sse2
+
+;-----------------------------------------------------------------------------
+; void ff_ac3_rshift_int32(int32_t *src, unsigned int len, unsigned int shift)
+;-----------------------------------------------------------------------------
+
+INIT_MMX
+AC3_SHIFT r, 32, psrad, mmx
+INIT_XMM
+AC3_SHIFT r, 32, psrad, sse2
diff --git a/libavcodec/x86/ac3dsp_mmx.c b/libavcodec/x86/ac3dsp_mmx.c
index d8af59ce8d..835b10696d 100644
--- a/libavcodec/x86/ac3dsp_mmx.c
+++ b/libavcodec/x86/ac3dsp_mmx.c
@@ -32,6 +32,12 @@ extern int ff_ac3_max_msb_abs_int16_mmxext(const int16_t *src, int len);
 extern int ff_ac3_max_msb_abs_int16_sse2  (const int16_t *src, int len);
 extern int ff_ac3_max_msb_abs_int16_ssse3 (const int16_t *src, int len);
 
+extern void ff_ac3_lshift_int16_mmx (int16_t *src, unsigned int len, unsigned int shift);
+extern void ff_ac3_lshift_int16_sse2(int16_t *src, unsigned int len, unsigned int shift);
+
+extern void ff_ac3_rshift_int32_mmx (int32_t *src, unsigned int len, unsigned int shift);
+extern void ff_ac3_rshift_int32_sse2(int32_t *src, unsigned int len, unsigned int shift);
+
 av_cold void ff_ac3dsp_init_x86(AC3DSPContext *c)
 {
     int mm_flags = av_get_cpu_flags();
@@ -40,6 +46,8 @@ av_cold void ff_ac3dsp_init_x86(AC3DSPContext *c)
     if (mm_flags & AV_CPU_FLAG_MMX) {
         c->ac3_exponent_min = ff_ac3_exponent_min_mmx;
         c->ac3_max_msb_abs_int16 = ff_ac3_max_msb_abs_int16_mmx;
+        c->ac3_lshift_int16 = ff_ac3_lshift_int16_mmx;
+        c->ac3_rshift_int32 = ff_ac3_rshift_int32_mmx;
     }
     if (mm_flags & AV_CPU_FLAG_MMX2 && HAVE_MMX2) {
         c->ac3_exponent_min = ff_ac3_exponent_min_mmxext;
@@ -48,6 +56,10 @@ av_cold void ff_ac3dsp_init_x86(AC3DSPContext *c)
     if (mm_flags & AV_CPU_FLAG_SSE2 && HAVE_SSE) {
         c->ac3_exponent_min = ff_ac3_exponent_min_sse2;
         c->ac3_max_msb_abs_int16 = ff_ac3_max_msb_abs_int16_sse2;
+        if (!(mm_flags & AV_CPU_FLAG_SSE2SLOW)) {
+            c->ac3_lshift_int16 = ff_ac3_lshift_int16_sse2;
+            c->ac3_rshift_int32 = ff_ac3_rshift_int32_sse2;
+        }
     }
     if (mm_flags & AV_CPU_FLAG_SSSE3 && HAVE_SSSE3) {
         c->ac3_max_msb_abs_int16 = ff_ac3_max_msb_abs_int16_ssse3;