x86: Use *_FAST/*_SLOW CPU feature detection macros where appropriate
[ffmpeg.git] / libavcodec / x86 / ac3dsp_init.c
index d1c45dd..89044f4 100644 (file)
@@ -1,5 +1,5 @@
 /*
- * x86-optimized AC-3 DSP utils
+ * x86-optimized AC-3 DSP functions
  * Copyright (c) 2011 Justin Ruggles
  *
  * This file is part of Libav.
  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  */
 
+#include "libavutil/attributes.h"
+#include "libavutil/mem.h"
 #include "libavutil/x86/asm.h"
 #include "libavutil/x86/cpu.h"
-#include "dsputil_mmx.h"
+#include "libavcodec/ac3.h"
 #include "libavcodec/ac3dsp.h"
 
-extern void ff_ac3_exponent_min_mmx   (uint8_t *exp, int num_reuse_blocks, int nb_coefs);
-extern void ff_ac3_exponent_min_mmxext(uint8_t *exp, int num_reuse_blocks, int nb_coefs);
-extern void ff_ac3_exponent_min_sse2  (uint8_t *exp, int num_reuse_blocks, int nb_coefs);
+void ff_ac3_exponent_min_mmx   (uint8_t *exp, int num_reuse_blocks, int nb_coefs);
+void ff_ac3_exponent_min_mmxext(uint8_t *exp, int num_reuse_blocks, int nb_coefs);
+void ff_ac3_exponent_min_sse2  (uint8_t *exp, int num_reuse_blocks, int nb_coefs);
 
-extern int ff_ac3_max_msb_abs_int16_mmx  (const int16_t *src, int len);
-extern int ff_ac3_max_msb_abs_int16_mmx2 (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);
+int ff_ac3_max_msb_abs_int16_mmx  (const int16_t *src, int len);
+int ff_ac3_max_msb_abs_int16_mmxext(const int16_t *src, int len);
+int ff_ac3_max_msb_abs_int16_sse2 (const int16_t *src, int len);
+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);
+void ff_ac3_lshift_int16_mmx (int16_t *src, unsigned int len, unsigned int shift);
+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);
+void ff_ac3_rshift_int32_mmx (int32_t *src, unsigned int len, unsigned int shift);
+void ff_ac3_rshift_int32_sse2(int32_t *src, unsigned int len, unsigned int shift);
 
-extern void ff_float_to_fixed24_3dnow(int32_t *dst, const float *src, unsigned int len);
-extern void ff_float_to_fixed24_sse  (int32_t *dst, const float *src, unsigned int len);
-extern void ff_float_to_fixed24_sse2 (int32_t *dst, const float *src, unsigned int len);
+void ff_float_to_fixed24_3dnow(int32_t *dst, const float *src, unsigned int len);
+void ff_float_to_fixed24_sse  (int32_t *dst, const float *src, unsigned int len);
+void ff_float_to_fixed24_sse2 (int32_t *dst, const float *src, unsigned int len);
 
-extern int ff_ac3_compute_mantissa_size_sse2(uint16_t mant_cnt[6][16]);
+int ff_ac3_compute_mantissa_size_sse2(uint16_t mant_cnt[6][16]);
 
-extern void ff_ac3_extract_exponents_3dnow(uint8_t *exp, int32_t *coef, int nb_coefs);
-extern void ff_ac3_extract_exponents_sse2 (uint8_t *exp, int32_t *coef, int nb_coefs);
-extern void ff_ac3_extract_exponents_ssse3(uint8_t *exp, int32_t *coef, int nb_coefs);
+void ff_ac3_extract_exponents_sse2 (uint8_t *exp, int32_t *coef, int nb_coefs);
+void ff_ac3_extract_exponents_ssse3(uint8_t *exp, int32_t *coef, int nb_coefs);
+
+void ff_apply_window_int16_round_mmxext(int16_t *output, const int16_t *input,
+                                        const int16_t *window, unsigned int len);
+void ff_apply_window_int16_round_sse2(int16_t *output, const int16_t *input,
+                                      const int16_t *window, unsigned int len);
+void ff_apply_window_int16_mmxext(int16_t *output, const int16_t *input,
+                                  const int16_t *window, unsigned int len);
+void ff_apply_window_int16_sse2(int16_t *output, const int16_t *input,
+                                const int16_t *window, unsigned int len);
+void ff_apply_window_int16_ssse3(int16_t *output, const int16_t *input,
+                                 const int16_t *window, unsigned int len);
+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           8(%1), %%xmm6          \n"             \
+        "movss          24(%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),                                          \
+          "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)[2],
+                            int out_ch, int in_ch, int len)
+{
+    int (*matrix_cmp)[2] = (int(*)[2])matrix;
+    intptr_t i, j, k, m;
+
+    i = -len * sizeof(float);
+    if (in_ch == 5 && out_ch == 2 &&
+        !(matrix_cmp[0][1] | matrix_cmp[2][0]   |
+          matrix_cmp[3][1] | matrix_cmp[4][0]   |
+          (matrix_cmp[1][0] ^ matrix_cmp[1][1]) |
+          (matrix_cmp[0][0] ^ matrix_cmp[2][1]))) {
+        MIX5(IF0, IF1);
+    } else if (in_ch == 5 && out_ch == 1 &&
+               matrix_cmp[0][0] == matrix_cmp[2][0] &&
+               matrix_cmp[3][0] == matrix_cmp[4][0]) {
+        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);
+        __asm__ volatile (
+            "1:                                 \n"
+            "sub             $8, %0             \n"
+            "movss     (%2, %0), %%xmm4         \n"
+            "movss    4(%2, %0), %%xmm5         \n"
+            "shufps          $0, %%xmm4, %%xmm4 \n"
+            "shufps          $0, %%xmm5, %%xmm5 \n"
+            "movaps      %%xmm4,   (%1, %0, 4)  \n"
+            "movaps      %%xmm5, 16(%1, %0, 4)  \n"
+            "jg              1b                 \n"
+            : "+&r"(j)
+            : "r"(matrix_simd), "r"(matrix)
+            : "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 mm_flags = av_get_cpu_flags();
+    int cpu_flags = av_get_cpu_flags();
 
-    if (EXTERNAL_MMX(mm_flags)) {
+    if (EXTERNAL_MMX(cpu_flags)) {
         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 (EXTERNAL_AMD3DNOW(mm_flags)) {
-        c->extract_exponents = ff_ac3_extract_exponents_3dnow;
+    if (EXTERNAL_AMD3DNOW(cpu_flags)) {
         if (!bit_exact) {
             c->float_to_fixed24 = ff_float_to_fixed24_3dnow;
         }
     }
-    if (EXTERNAL_MMXEXT(mm_flags)) {
+    if (EXTERNAL_MMXEXT(cpu_flags)) {
         c->ac3_exponent_min = ff_ac3_exponent_min_mmxext;
-        c->ac3_max_msb_abs_int16 = ff_ac3_max_msb_abs_int16_mmx2;
+        c->ac3_max_msb_abs_int16 = ff_ac3_max_msb_abs_int16_mmxext;
+        if (bit_exact) {
+            c->apply_window_int16 = ff_apply_window_int16_mmxext;
+        } else {
+            c->apply_window_int16 = ff_apply_window_int16_round_mmxext;
+        }
     }
-    if (EXTERNAL_SSE(mm_flags)) {
+    if (EXTERNAL_SSE(cpu_flags)) {
         c->float_to_fixed24 = ff_float_to_fixed24_sse;
     }
-    if (EXTERNAL_SSE2(mm_flags)) {
+    if (EXTERNAL_SSE2(cpu_flags)) {
         c->ac3_exponent_min = ff_ac3_exponent_min_sse2;
         c->ac3_max_msb_abs_int16 = ff_ac3_max_msb_abs_int16_sse2;
         c->float_to_fixed24 = ff_float_to_fixed24_sse2;
         c->compute_mantissa_size = ff_ac3_compute_mantissa_size_sse2;
         c->extract_exponents = ff_ac3_extract_exponents_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 (bit_exact) {
+            c->apply_window_int16 = ff_apply_window_int16_sse2;
         }
     }
-    if (EXTERNAL_SSSE3(mm_flags)) {
+
+    if (EXTERNAL_SSE2_FAST(cpu_flags)) {
+        c->ac3_lshift_int16 = ff_ac3_lshift_int16_sse2;
+        c->ac3_rshift_int32 = ff_ac3_rshift_int32_sse2;
+        if (!bit_exact) {
+            c->apply_window_int16 = ff_apply_window_int16_round_sse2;
+        }
+    }
+
+    if (EXTERNAL_SSSE3(cpu_flags)) {
         c->ac3_max_msb_abs_int16 = ff_ac3_max_msb_abs_int16_ssse3;
-        if (!(mm_flags & AV_CPU_FLAG_ATOM)) {
+        if (cpu_flags & AV_CPU_FLAG_ATOM) {
+            c->apply_window_int16 = ff_apply_window_int16_ssse3_atom;
+        } else {
             c->extract_exponents = ff_ac3_extract_exponents_ssse3;
+            c->apply_window_int16 = ff_apply_window_int16_ssse3;
         }
     }
+
+#if HAVE_SSE_INLINE && HAVE_7REGS
+    if (INLINE_SSE(cpu_flags)) {
+        c->downmix = ac3_downmix_sse;
+    }
+#endif
 }