dcadec: simplify decoding of VQ high frequencies

The vector dequantization has a test in a loop preventing effective SIMD
implementation. By moving it out of the loop, this loop can be DSPized.

Therefore, modify the current DSP implementation. In particular, the
DSP implementation no longer has to handle null loop sizes.

The decode_hf implementations have following timings:

For x86 Arrandale:
        C  SSE SSE2 SSE4
win32: 260 162  119  104
win64: 242 N/A   89   72

The arm NEON optimizations follow in a later patch as external asm. The
now unused check for the y modifier in arm inline asm is removed from
configure.

6 files changed, 80 insertions, 74 deletions

diff --git a/libavcodec/arm/dca.h b/libavcodec/arm/dca.h
index 580bd75ffe..4aed57603e 100644
--- a/libavcodec/arm/dca.h
+++ b/libavcodec/arm/dca.h
@@ -81,27 +81,4 @@ static inline int decode_blockcodes(int code1, int code2, int levels,
 
 #endif
 
-#if HAVE_NEON_INLINE && HAVE_ASM_MOD_Y
-
-#define int8x8_fmul_int32 int8x8_fmul_int32
-static inline void int8x8_fmul_int32(av_unused DCADSPContext *dsp,
-                                     float *dst, const int8_t *src, int scale)
-{
-    __asm__ ("vcvt.f32.s32 %2,  %2,  #4         \n"
-             "vld1.8       {d0},     [%1,:64]   \n"
-             "vmovl.s8     q0,  d0              \n"
-             "vmovl.s16    q1,  d1              \n"
-             "vmovl.s16    q0,  d0              \n"
-             "vcvt.f32.s32 q0,  q0              \n"
-             "vcvt.f32.s32 q1,  q1              \n"
-             "vmul.f32     q0,  q0,  %y2        \n"
-             "vmul.f32     q1,  q1,  %y2        \n"
-             "vst1.32      {q0-q1},  [%m0,:128] \n"
-             : "=Um"(*(float (*)[8])dst)
-             : "r"(src), "x"(scale)
-             : "d0", "d1", "d2", "d3");
-}
-
-#endif
-
 #endif /* AVCODEC_ARM_DCA_H */
diff --git a/libavcodec/dcadec.c b/libavcodec/dcadec.c
index 5ede61db33..92edce7e42 100644
--- a/libavcodec/dcadec.c
+++ b/libavcodec/dcadec.c
@@ -50,14 +50,10 @@
 #if ARCH_ARM
 #   include "arm/dca.h"
 #endif
-#if ARCH_X86
-#   include "x86/dca.h"
-#endif
 
 //#define TRACE
 
 #define DCA_PRIM_CHANNELS_MAX  (7)
-#define DCA_SUBBANDS          (32)
 #define DCA_ABITS_MAX         (32)      /* Should be 28 */
 #define DCA_SUBSUBFRAMES_MAX   (4)
 #define DCA_SUBFRAMES_MAX     (16)
@@ -340,7 +336,7 @@ typedef struct {
     int prediction_vq[DCA_PRIM_CHANNELS_MAX][DCA_SUBBANDS];      ///< prediction VQ coefs
     int bitalloc[DCA_PRIM_CHANNELS_MAX][DCA_SUBBANDS];           ///< bit allocation index
     int transition_mode[DCA_PRIM_CHANNELS_MAX][DCA_SUBBANDS];    ///< transition mode (transients)
-    int scale_factor[DCA_PRIM_CHANNELS_MAX][DCA_SUBBANDS][2];    ///< scale factors (2 if transient)
+    int32_t scale_factor[DCA_PRIM_CHANNELS_MAX][DCA_SUBBANDS][2];///< scale factors (2 if transient)
     int joint_huff[DCA_PRIM_CHANNELS_MAX];                       ///< joint subband scale factors codebook
     int joint_scale_factor[DCA_PRIM_CHANNELS_MAX][DCA_SUBBANDS]; ///< joint subband scale factors
     float downmix_coef[DCA_PRIM_CHANNELS_MAX + 1][2];            ///< stereo downmix coefficients
@@ -353,7 +349,7 @@ typedef struct {
     uint8_t  core_downmix_amode;                                 ///< audio channel arrangement of embedded downmix
     uint16_t core_downmix_codes[DCA_PRIM_CHANNELS_MAX + 1][4];   ///< embedded downmix coefficients (9-bit codes)
 
-    int high_freq_vq[DCA_PRIM_CHANNELS_MAX][DCA_SUBBANDS];       ///< VQ encoded high frequency subbands
+    int32_t  high_freq_vq[DCA_PRIM_CHANNELS_MAX][DCA_SUBBANDS];  ///< VQ encoded high frequency subbands
 
     float lfe_data[2 * DCA_LFE_MAX * (DCA_BLOCKS_MAX + 4)];      ///< Low frequency effect data
     int lfe_scale_factor;
@@ -1088,14 +1084,6 @@ static int decode_blockcodes(int code1, int code2, int levels, int32_t *values)
 static const uint8_t abits_sizes[7]  = { 7, 10, 12, 13, 15, 17, 19 };
 static const uint8_t abits_levels[7] = { 3,  5,  7,  9, 13, 17, 25 };
 
-#ifndef int8x8_fmul_int32
-static inline void int8x8_fmul_int32(DCADSPContext *dsp, float *dst,
-                                     const int8_t *src, int scale)
-{
-    dsp->int8x8_fmul_int32(dst, src, scale);
-}
-#endif
-
 static int dca_subsubframe(DCAContext *s, int base_channel, int block_index)
 {
     int k, l;
@@ -1220,20 +1208,16 @@ static int dca_subsubframe(DCAContext *s, int base_channel, int block_index)
         /*
          * Decode VQ encoded high frequencies
          */
-        for (l = s->vq_start_subband[k]; l < s->subband_activity[k]; l++) {
-            /* 1 vector -> 32 samples but we only need the 8 samples
-             * for this subsubframe. */
-            int hfvq = s->high_freq_vq[k][l];
-
+        if (s->subband_activity[k] > s->vq_start_subband[k]) {
             if (!s->debug_flag & 0x01) {
                 av_log(s->avctx, AV_LOG_DEBUG,
                        "Stream with high frequencies VQ coding\n");
                 s->debug_flag |= 0x01;
             }
-
-            int8x8_fmul_int32(&s->dcadsp, subband_samples[k][l],
-                              &high_freq_vq[hfvq][subsubframe * 8],
-                              s->scale_factor[k][l][0]);
+            s->dcadsp.decode_hf(subband_samples[k], s->high_freq_vq[k],
+                                high_freq_vq, subsubframe * 8,
+                                s->scale_factor[k], s->vq_start_subband[k],
+                                s->subband_activity[k]);
         }
     }
 
diff --git a/libavcodec/dcadsp.c b/libavcodec/dcadsp.c
index 294a1b2736..18732dc80a 100644
--- a/libavcodec/dcadsp.c
+++ b/libavcodec/dcadsp.c
@@ -24,12 +24,22 @@
 #include "libavutil/intreadwrite.h"
 #include "dcadsp.h"
 
-static void int8x8_fmul_int32_c(float *dst, const int8_t *src, int scale)
+static void decode_hf_c(float dst[DCA_SUBBANDS][8],
+                        const int32_t vq_num[DCA_SUBBANDS],
+                        const int8_t hf_vq[1024][32], intptr_t vq_offset,
+                        int32_t scale[DCA_SUBBANDS][2],
+                        intptr_t start, intptr_t end)
 {
-    float fscale = scale / 16.0;
-    int i;
-    for (i = 0; i < 8; i++)
-        dst[i] = src[i] * fscale;
+    int i, l;
+
+    for (l = start; l < end; l++) {
+        /* 1 vector -> 32 samples but we only need the 8 samples
+         * for this subsubframe. */
+        const int8_t *ptr = &hf_vq[vq_num[l]][vq_offset];
+        float fscale = scale[l][0] * (1 / 16.0);
+        for (i = 0; i < 8; i++)
+            dst[l][i] = ptr[i] * fscale;
+    }
 }
 
 static inline void
@@ -96,7 +106,7 @@ av_cold void ff_dcadsp_init(DCADSPContext *s)
     s->lfe_fir[0] = dca_lfe_fir0_c;
     s->lfe_fir[1] = dca_lfe_fir1_c;
     s->qmf_32_subbands = dca_qmf_32_subbands;
-    s->int8x8_fmul_int32 = int8x8_fmul_int32_c;
+    s->decode_hf = decode_hf_c;
     if (ARCH_ARM) ff_dcadsp_init_arm(s);
     if (ARCH_X86) ff_dcadsp_init_x86(s);
 }
diff --git a/libavcodec/dcadsp.h b/libavcodec/dcadsp.h
index e4c1bc7f05..0fa75a5ed6 100644
--- a/libavcodec/dcadsp.h
+++ b/libavcodec/dcadsp.h
@@ -22,6 +22,8 @@
 #include "avfft.h"
 #include "synth_filter.h"
 
+#define DCA_SUBBANDS 32
+
 typedef struct DCADSPContext {
     void (*lfe_fir[2])(float *out, const float *in, const float *coefs);
     void (*qmf_32_subbands)(float samples_in[32][8], int sb_act,
@@ -30,7 +32,11 @@ typedef struct DCADSPContext {
                             int *synth_buf_offset, float synth_buf2[32],
                             const float window[512], float *samples_out,
                             float raXin[32], float scale);
-    void (*int8x8_fmul_int32)(float *dst, const int8_t *src, int scale);
+    void (*decode_hf)(float dst[DCA_SUBBANDS][8],
+                      const int32_t vq_num[DCA_SUBBANDS],
+                      const int8_t hf_vq[1024][32], intptr_t vq_offset,
+                      int32_t scale[DCA_SUBBANDS][2],
+                      intptr_t start, intptr_t end);
 } DCADSPContext;
 
 void ff_dcadsp_init(DCADSPContext *s);
diff --git a/libavcodec/x86/dcadsp.asm b/libavcodec/x86/dcadsp.asm
index c1db118676..56039baa6f 100644
--- a/libavcodec/x86/dcadsp.asm
+++ b/libavcodec/x86/dcadsp.asm
@@ -26,18 +26,35 @@ pf_inv16:  times 4 dd 0x3D800000 ; 1/16
 
 SECTION_TEXT
 
-; void int8x8_fmul_int32_sse2(float *dst, const int8_t *src, int scale)
-%macro INT8X8_FMUL_INT32 0
-cglobal int8x8_fmul_int32, 3,3,5, dst, src, scale
-    cvtsi2ss    m0, scalem
+; void decode_hf(float dst[DCA_SUBBANDS][8], const int32_t vq_num[DCA_SUBBANDS],
+;                const int8_t hf_vq[1024][32], intptr_t vq_offset,
+;                int32_t scale[DCA_SUBBANDS][2], intptr_t start, intptr_t end)
+
+%macro DECODE_HF 0
+cglobal decode_hf, 6,6,5, dst, num, src, offset, scale, start, end
+    lea       srcq, [srcq + offsetq]
+    shl     startq, 2
+    mov    offsetd, endm
+%define DICT offsetq
+    shl    offsetq, 2
+    mov       endm, offsetq
+.loop:
+%if ARCH_X86_64
+    mov    offsetd, [scaleq + 2 * startq]
+    cvtsi2ss    m0, offsetd
+%else
+    cvtsi2ss    m0, [scaleq + 2 * startq]
+%endif
+    mov    offsetd, [numq + startq]
     mulss       m0, [pf_inv16]
+    shl       DICT, 5
     shufps      m0, m0, 0
 %if cpuflag(sse2)
 %if cpuflag(sse4)
-    pmovsxbd    m1, [srcq+0]
-    pmovsxbd    m2, [srcq+4]
+    pmovsxbd    m1, [srcq + DICT + 0]
+    pmovsxbd    m2, [srcq + DICT + 4]
 %else
-    movq        m1, [srcq]
+    movq        m1, [srcq + DICT]
     punpcklbw   m1, m1
     mova        m2, m1
     punpcklwd   m1, m1
@@ -48,8 +65,8 @@ cglobal int8x8_fmul_int32, 3,3,5, dst, src, scale
     cvtdq2ps    m1, m1
     cvtdq2ps    m2, m2
 %else
-    movd       mm0, [srcq+0]
-    movd       mm1, [srcq+4]
+    movd       mm0, [srcq + DICT + 0]
+    movd       mm1, [srcq + DICT + 4]
     punpcklbw  mm0, mm0
     punpcklbw  mm1, mm1
     movq       mm2, mm0
@@ -67,27 +84,33 @@ cglobal int8x8_fmul_int32, 3,3,5, dst, src, scale
     cvtpi2ps    m3, mm2
     cvtpi2ps    m4, mm3
     shufps      m0, m0, 0
-    emms
     shufps      m1, m3, q1010
     shufps      m2, m4, q1010
 %endif
     mulps       m1, m0
     mulps       m2, m0
-    mova [dstq+ 0], m1
-    mova [dstq+16], m2
+    mova [dstq + 8 * startq +  0], m1
+    mova [dstq + 8 * startq + 16], m2
+    add     startq, 4
+    cmp     startq, endm
+    jl       .loop
+.end:
+%if notcpuflag(sse2)
+    emms
+%endif
     REP_RET
 %endmacro
 
 %if ARCH_X86_32
 INIT_XMM sse
-INT8X8_FMUL_INT32
+DECODE_HF
 %endif
 
 INIT_XMM sse2
-INT8X8_FMUL_INT32
+DECODE_HF
 
 INIT_XMM sse4
-INT8X8_FMUL_INT32
+DECODE_HF
 
 ; %1=v0/v1  %2=in1  %3=in2
 %macro FIR_LOOP 2-3
diff --git a/libavcodec/x86/dcadsp_init.c b/libavcodec/x86/dcadsp_init.c
index 5f6e8c5a19..3821892166 100644
--- a/libavcodec/x86/dcadsp_init.c
+++ b/libavcodec/x86/dcadsp_init.c
@@ -23,9 +23,15 @@
 #include "libavutil/x86/cpu.h"
 #include "libavcodec/dcadsp.h"
 
-void ff_int8x8_fmul_int32_sse(float *dst, const int8_t *src, int scale);
-void ff_int8x8_fmul_int32_sse2(float *dst, const int8_t *src, int scale);
-void ff_int8x8_fmul_int32_sse4(float *dst, const int8_t *src, int scale);
+void ff_decode_hf_sse(float dst[DCA_SUBBANDS][8], const int vq_num[DCA_SUBBANDS],
+                      const int8_t hf_vq[1024][32], intptr_t vq_offset,
+                      int scale[DCA_SUBBANDS][2], intptr_t start, intptr_t end);
+void ff_decode_hf_sse2(float dst[DCA_SUBBANDS][8], const int vq_num[DCA_SUBBANDS],
+                       const int8_t hf_vq[1024][32], intptr_t vq_offset,
+                       int scale[DCA_SUBBANDS][2], intptr_t start, intptr_t end);
+void ff_decode_hf_sse4(float dst[DCA_SUBBANDS][8], const int vq_num[DCA_SUBBANDS],
+                       const int8_t hf_vq[1024][32], intptr_t vq_offset,
+                       int scale[DCA_SUBBANDS][2], intptr_t start, intptr_t end);
 void ff_dca_lfe_fir0_sse(float *out, const float *in, const float *coefs);
 void ff_dca_lfe_fir1_sse(float *out, const float *in, const float *coefs);
 
@@ -35,18 +41,18 @@ av_cold void ff_dcadsp_init_x86(DCADSPContext *s)
 
     if (EXTERNAL_SSE(cpu_flags)) {
 #if ARCH_X86_32
-        s->int8x8_fmul_int32 = ff_int8x8_fmul_int32_sse;
+        s->decode_hf = ff_decode_hf_sse;
 #endif
         s->lfe_fir[0]        = ff_dca_lfe_fir0_sse;
         s->lfe_fir[1]        = ff_dca_lfe_fir1_sse;
     }
 
     if (EXTERNAL_SSE2(cpu_flags)) {
-        s->int8x8_fmul_int32 = ff_int8x8_fmul_int32_sse2;
+        s->decode_hf = ff_decode_hf_sse2;
     }
 
     if (EXTERNAL_SSE4(cpu_flags)) {
-        s->int8x8_fmul_int32 = ff_int8x8_fmul_int32_sse4;
+        s->decode_hf = ff_decode_hf_sse4;
     }
 }