ac3dsp: x86: Replace inline asm for in-decoder downmixing with standalone asm

Adds a wrapper function for downmixing which detects channel count changes
and updates the selected downmix function accordingly.

Simplification and porting to current x86inc infrastructure by Diego Biurrun.

Signed-off-by: Diego Biurrun <diego@biurrun.de>

1 files changed, 41 insertions, 133 deletions

diff --git a/libavcodec/x86/ac3dsp_init.c b/libavcodec/x86/ac3dsp_init.c
index 9036389296..6d049b37cf 100644
--- a/libavcodec/x86/ac3dsp_init.c
+++ b/libavcodec/x86/ac3dsp_init.c
@@ -63,135 +63,6 @@ void ff_apply_window_int16_ssse3(int16_t *output, const int16_t *input,
 void ff_apply_window_int16_ssse3_atom(int16_t *output, const int16_t *input,
                                       const int16_t *window, unsigned int len);
 
-#if HAVE_SSE_INLINE && HAVE_7REGS
-
-#define IF1(x) x
-#define IF0(x)
-
-#define MIX5(mono, stereo)                                      \
-    __asm__ volatile (                                          \
-        "movss           0(%1), %%xmm5          \n"             \
-        "movss           4(%1), %%xmm6          \n"             \
-        "movss          12(%1), %%xmm7          \n"             \
-        "shufps     $0, %%xmm5, %%xmm5          \n"             \
-        "shufps     $0, %%xmm6, %%xmm6          \n"             \
-        "shufps     $0, %%xmm7, %%xmm7          \n"             \
-        "1:                                     \n"             \
-        "movaps       (%0, %2), %%xmm0          \n"             \
-        "movaps       (%0, %3), %%xmm1          \n"             \
-        "movaps       (%0, %4), %%xmm2          \n"             \
-        "movaps       (%0, %5), %%xmm3          \n"             \
-        "movaps       (%0, %6), %%xmm4          \n"             \
-        "mulps          %%xmm5, %%xmm0          \n"             \
-        "mulps          %%xmm6, %%xmm1          \n"             \
-        "mulps          %%xmm5, %%xmm2          \n"             \
-        "mulps          %%xmm7, %%xmm3          \n"             \
-        "mulps          %%xmm7, %%xmm4          \n"             \
- stereo("addps          %%xmm1, %%xmm0          \n")            \
-        "addps          %%xmm1, %%xmm2          \n"             \
-        "addps          %%xmm3, %%xmm0          \n"             \
-        "addps          %%xmm4, %%xmm2          \n"             \
-   mono("addps          %%xmm2, %%xmm0          \n")            \
-        "movaps         %%xmm0, (%0, %2)        \n"             \
- stereo("movaps         %%xmm2, (%0, %3)        \n")            \
-        "add               $16, %0              \n"             \
-        "jl                 1b                  \n"             \
-        : "+&r"(i)                                              \
-        : "r"(matrix[0]),                                          \
-          "r"(samples[0] + len),                                \
-          "r"(samples[1] + len),                                \
-          "r"(samples[2] + len),                                \
-          "r"(samples[3] + len),                                \
-          "r"(samples[4] + len)                                 \
-        : XMM_CLOBBERS("%xmm0", "%xmm1", "%xmm2", "%xmm3",      \
-                      "%xmm4", "%xmm5", "%xmm6", "%xmm7",)      \
-         "memory"                                               \
-    );
-
-#define MIX_MISC(stereo)                                        \
-    __asm__ volatile (                                          \
-        "mov              %5, %2            \n"                 \
-        "1:                                 \n"                 \
-        "mov -%c7(%6, %2, %c8), %3          \n"                 \
-        "movaps     (%3, %0), %%xmm0        \n"                 \
- stereo("movaps       %%xmm0, %%xmm1        \n")                \
-        "mulps        %%xmm4, %%xmm0        \n"                 \
- stereo("mulps        %%xmm5, %%xmm1        \n")                \
-        "2:                                 \n"                 \
-        "mov   (%6, %2, %c8), %1            \n"                 \
-        "movaps     (%1, %0), %%xmm2        \n"                 \
- stereo("movaps       %%xmm2, %%xmm3        \n")                \
-        "mulps   (%4, %2, 8), %%xmm2        \n"                 \
- stereo("mulps 16(%4, %2, 8), %%xmm3        \n")                \
-        "addps        %%xmm2, %%xmm0        \n"                 \
- stereo("addps        %%xmm3, %%xmm1        \n")                \
-        "add              $4, %2            \n"                 \
-        "jl               2b                \n"                 \
-        "mov              %5, %2            \n"                 \
- stereo("mov   (%6, %2, %c8), %1            \n")                \
-        "movaps       %%xmm0, (%3, %0)      \n"                 \
- stereo("movaps       %%xmm1, (%1, %0)      \n")                \
-        "add             $16, %0            \n"                 \
-        "jl               1b                \n"                 \
-        : "+&r"(i), "=&r"(j), "=&r"(k), "=&r"(m)                \
-        : "r"(matrix_simd + in_ch),                             \
-          "g"((intptr_t) - 4 * (in_ch - 1)),                    \
-          "r"(samp + in_ch),                                    \
-          "i"(sizeof(float *)), "i"(sizeof(float *)/4)          \
-        : "memory"                                              \
-    );
-
-static void ac3_downmix_sse(float **samples, float **matrix,
-                            int out_ch, int in_ch, int len)
-{
-    int **matrix_cmp = (int **)matrix;
-    intptr_t i, j, k, m;
-
-    i = -len * sizeof(float);
-    if (in_ch == 5 && out_ch == 2 &&
-        !(matrix_cmp[1][0] | matrix_cmp[0][2]   |
-          matrix_cmp[1][3] | matrix_cmp[0][4]   |
-          (matrix_cmp[0][1] ^ matrix_cmp[1][1]) |
-          (matrix_cmp[0][0] ^ matrix_cmp[1][2]))) {
-        MIX5(IF0, IF1);
-    } else if (in_ch == 5 && out_ch == 1 &&
-               matrix_cmp[0][0] == matrix_cmp[0][2] &&
-               matrix_cmp[0][3] == matrix_cmp[0][4]) {
-        MIX5(IF1, IF0);
-    } else {
-        DECLARE_ALIGNED(16, float, matrix_simd)[AC3_MAX_CHANNELS][2][4];
-        float *samp[AC3_MAX_CHANNELS];
-
-        for (j = 0; j < in_ch; j++)
-            samp[j] = samples[j] + len;
-
-        j = 2 * in_ch * sizeof(float);
-        k =     in_ch * sizeof(float);
-        __asm__ volatile (
-            "1:                                 \n"
-            "sub             $4, %1             \n"
-            "sub             $8, %0             \n"
-            "movss     (%3, %1), %%xmm4         \n"
-            "movss     (%4, %1), %%xmm5         \n"
-            "shufps          $0, %%xmm4, %%xmm4 \n"
-            "shufps          $0, %%xmm5, %%xmm5 \n"
-            "movaps      %%xmm4,   (%2, %0, 4)  \n"
-            "movaps      %%xmm5, 16(%2, %0, 4)  \n"
-            "jg              1b                 \n"
-            : "+&r"(j), "+&r"(k)
-            : "r"(matrix_simd), "r"(matrix[0]), "r"(matrix[1])
-            : "memory"
-        );
-        if (out_ch == 2) {
-            MIX_MISC(IF1);
-        } else {
-            MIX_MISC(IF0);
-        }
-    }
-}
-
-#endif /* HAVE_SSE_INLINE && HAVE_7REGS */
-
 av_cold void ff_ac3dsp_init_x86(AC3DSPContext *c, int bit_exact)
 {
     int cpu_flags = av_get_cpu_flags();
@@ -247,10 +118,47 @@ av_cold void ff_ac3dsp_init_x86(AC3DSPContext *c, int bit_exact)
             c->apply_window_int16 = ff_apply_window_int16_ssse3;
         }
     }
+}
+
+#define DOWNMIX_FUNC_OPT(ch, opt)                                       \
+void ff_ac3_downmix_ ## ch ## _to_1_ ## opt(float **samples,            \
+                                            float **matrix, int len);   \
+void ff_ac3_downmix_ ## ch ## _to_2_ ## opt(float **samples,            \
+                                            float **matrix, int len);
+
+#define DOWNMIX_FUNCS(opt)   \
+    DOWNMIX_FUNC_OPT(3, opt) \
+    DOWNMIX_FUNC_OPT(4, opt) \
+    DOWNMIX_FUNC_OPT(5, opt) \
+    DOWNMIX_FUNC_OPT(6, opt)
+
+DOWNMIX_FUNCS(sse)
+DOWNMIX_FUNCS(avx)
+DOWNMIX_FUNCS(fma3)
+
+void ff_ac3dsp_set_downmix_x86(AC3DSPContext *c)
+{
+    int cpu_flags = av_get_cpu_flags();
+
+#define SET_DOWNMIX(ch, suf, SUF)                                       \
+    if (ch == c->in_channels) {                                         \
+        if (EXTERNAL_ ## SUF (cpu_flags)) {                             \
+            if (c->out_channels == 1)                                   \
+                c->downmix = ff_ac3_downmix_ ## ch ## _to_1_ ## suf;    \
+            else                                                        \
+                c->downmix = ff_ac3_downmix_ ## ch ## _to_2_ ## suf;    \
+        }                                                               \
+    }
+
+#define SET_DOWNMIX_ALL(suf, SUF)                   \
+    SET_DOWNMIX(3, suf, SUF)                        \
+    SET_DOWNMIX(4, suf, SUF)                        \
+    SET_DOWNMIX(5, suf, SUF)                        \
+    SET_DOWNMIX(6, suf, SUF)
 
-#if HAVE_SSE_INLINE && HAVE_7REGS
-    if (INLINE_SSE(cpu_flags)) {
-        c->downmix = ac3_downmix_sse;
+    SET_DOWNMIX_ALL(sse,  SSE)
+    if (!(cpu_flags & AV_CPU_FLAG_AVXSLOW)) {
+        SET_DOWNMIX_ALL(avx,  AVX)
+        SET_DOWNMIX_ALL(fma3, FMA3)
     }
-#endif
 }