Adding dotprod instruction to ARM rtcd

Used for DNN matrix multiplies

4 files changed, 57 insertions, 13 deletions

diff --git a/celt/arm/arm_celt_map.c b/celt/arm/arm_celt_map.c
index ca988b66..cbaea495 100644
--- a/celt/arm/arm_celt_map.c
+++ b/celt/arm/arm_celt_map.c
@@ -40,7 +40,8 @@ opus_val32 (*const CELT_INNER_PROD_IMPL[OPUS_ARCHMASK+1])(const opus_val16 *x, c
   celt_inner_prod_c,   /* ARMv4 */
   celt_inner_prod_c,   /* EDSP */
   celt_inner_prod_c,   /* Media */
-  celt_inner_prod_neon /* NEON */
+  celt_inner_prod_neon,/* NEON */
+  celt_inner_prod_neon /* DOTPROD */
 };
 
 void (*const DUAL_INNER_PROD_IMPL[OPUS_ARCHMASK+1])(const opus_val16 *x, const opus_val16 *y01, const opus_val16 *y02,
@@ -48,7 +49,8 @@ void (*const DUAL_INNER_PROD_IMPL[OPUS_ARCHMASK+1])(const opus_val16 *x, const o
   dual_inner_prod_c,   /* ARMv4 */
   dual_inner_prod_c,   /* EDSP */
   dual_inner_prod_c,   /* Media */
-  dual_inner_prod_neon /* NEON */
+  dual_inner_prod_neon,/* NEON */
+  dual_inner_prod_neon /* DOTPROD */
 };
 # endif
 
@@ -61,7 +63,8 @@ opus_val32 (*const CELT_PITCH_XCORR_IMPL[OPUS_ARCHMASK+1])(const opus_val16 *,
   celt_pitch_xcorr_c,               /* ARMv4 */
   MAY_HAVE_EDSP(celt_pitch_xcorr),  /* EDSP */
   MAY_HAVE_MEDIA(celt_pitch_xcorr), /* Media */
-  MAY_HAVE_NEON(celt_pitch_xcorr)   /* NEON */
+  MAY_HAVE_NEON(celt_pitch_xcorr),  /* NEON */
+  MAY_HAVE_NEON(celt_pitch_xcorr)   /* DOTPROD */
 };
 
 #  endif
@@ -72,7 +75,8 @@ void (*const CELT_PITCH_XCORR_IMPL[OPUS_ARCHMASK+1])(const opus_val16 *,
   celt_pitch_xcorr_c,              /* ARMv4 */
   celt_pitch_xcorr_c,              /* EDSP */
   celt_pitch_xcorr_c,              /* Media */
-  celt_pitch_xcorr_float_neon      /* Neon */
+  celt_pitch_xcorr_float_neon,     /* Neon */
+  celt_pitch_xcorr_float_neon      /* DOTPROD */
 };
 #  endif
 # endif /* FIXED_POINT */
@@ -90,6 +94,7 @@ void (*const XCORR_KERNEL_IMPL[OPUS_ARCHMASK + 1])(
   xcorr_kernel_c,                /* EDSP */
   xcorr_kernel_c,                /* Media */
   xcorr_kernel_neon_fixed,       /* Neon */
+  xcorr_kernel_neon_fixed        /* DOTPROD */
 };
 
 #endif
@@ -101,14 +106,16 @@ int (*const OPUS_FFT_ALLOC_ARCH_IMPL[OPUS_ARCHMASK+1])(kiss_fft_state *st) = {
    opus_fft_alloc_arch_c,        /* ARMv4 */
    opus_fft_alloc_arch_c,        /* EDSP */
    opus_fft_alloc_arch_c,        /* Media */
-   opus_fft_alloc_arm_neon       /* Neon with NE10 library support */
+   opus_fft_alloc_arm_neon,      /* Neon with NE10 library support */
+   opus_fft_alloc_arm_neon       /* DOTPROD with NE10 library support */
 };
 
 void (*const OPUS_FFT_FREE_ARCH_IMPL[OPUS_ARCHMASK+1])(kiss_fft_state *st) = {
    opus_fft_free_arch_c,         /* ARMv4 */
    opus_fft_free_arch_c,         /* EDSP */
    opus_fft_free_arch_c,         /* Media */
-   opus_fft_free_arm_neon        /* Neon with NE10 */
+   opus_fft_free_arm_neon,       /* Neon with NE10 */
+   opus_fft_free_arm_neon        /* DOTPROD with NE10 */
 };
 #   endif /* CUSTOM_MODES */
 
@@ -118,7 +125,8 @@ void (*const OPUS_FFT[OPUS_ARCHMASK+1])(const kiss_fft_state *cfg,
    opus_fft_c,                   /* ARMv4 */
    opus_fft_c,                   /* EDSP */
    opus_fft_c,                   /* Media */
-   opus_fft_neon                 /* Neon with NE10 */
+   opus_fft_neon,                /* Neon with NE10 */
+   opus_fft_neon                 /* DOTPROD with NE10 */
 };
 
 void (*const OPUS_IFFT[OPUS_ARCHMASK+1])(const kiss_fft_state *cfg,
@@ -127,7 +135,8 @@ void (*const OPUS_IFFT[OPUS_ARCHMASK+1])(const kiss_fft_state *cfg,
    opus_ifft_c,                   /* ARMv4 */
    opus_ifft_c,                   /* EDSP */
    opus_ifft_c,                   /* Media */
-   opus_ifft_neon                 /* Neon with NE10 */
+   opus_ifft_neon,                /* Neon with NE10 */
+   opus_ifft_neon                 /* DOTPROD with NE10 */
 };
 
 void (*const CLT_MDCT_FORWARD_IMPL[OPUS_ARCHMASK+1])(const mdct_lookup *l,
@@ -139,7 +148,8 @@ void (*const CLT_MDCT_FORWARD_IMPL[OPUS_ARCHMASK+1])(const mdct_lookup *l,
    clt_mdct_forward_c,           /* ARMv4 */
    clt_mdct_forward_c,           /* EDSP */
    clt_mdct_forward_c,           /* Media */
-   clt_mdct_forward_neon         /* Neon with NE10 */
+   clt_mdct_forward_neon,        /* Neon with NE10 */
+   clt_mdct_forward_neon         /* DOTPROD with NE10 */
 };
 
 void (*const CLT_MDCT_BACKWARD_IMPL[OPUS_ARCHMASK+1])(const mdct_lookup *l,
@@ -151,7 +161,8 @@ void (*const CLT_MDCT_BACKWARD_IMPL[OPUS_ARCHMASK+1])(const mdct_lookup *l,
    clt_mdct_backward_c,           /* ARMv4 */
    clt_mdct_backward_c,           /* EDSP */
    clt_mdct_backward_c,           /* Media */
-   clt_mdct_backward_neon         /* Neon with NE10 */
+   clt_mdct_backward_neon,        /* Neon with NE10 */
+   clt_mdct_backward_neon         /* DOTPROD with NE10 */
 };
 
 #  endif /* HAVE_ARM_NE10 */
diff --git a/celt/arm/armcpu.c b/celt/arm/armcpu.c
index c7d16e6d..6fbfb904 100644
--- a/celt/arm/armcpu.c
+++ b/celt/arm/armcpu.c
@@ -43,6 +43,7 @@
 #define OPUS_CPU_ARM_EDSP_FLAG  (1<<OPUS_ARCH_ARM_EDSP)
 #define OPUS_CPU_ARM_MEDIA_FLAG (1<<OPUS_ARCH_ARM_MEDIA)
 #define OPUS_CPU_ARM_NEON_FLAG  (1<<OPUS_ARCH_ARM_NEON)
+#define OPUS_CPU_ARM_DOTPROD_FLAG  (1<<OPUS_ARCH_ARM_DOTPROD)
 
 #if defined(_MSC_VER)
 /*For GetExceptionCode() and EXCEPTION_ILLEGAL_INSTRUCTION.*/
@@ -127,6 +128,11 @@ opus_uint32 opus_cpu_capabilities(void)
         if(p != NULL && (p[5] == ' ' || p[5] == '\n'))
           flags |= OPUS_CPU_ARM_NEON_FLAG;
 #  endif
+#  if defined(OPUS_ARM_MAY_HAVE_DOTPROD)
+        p = strstr(buf, " asimddp");
+        if(p != NULL && (p[8] == ' ' || p[8] == '\n'))
+          flags |= OPUS_CPU_ARM_DOTPROD_FLAG;
+#  endif
       }
 # endif
 
@@ -144,6 +150,13 @@ opus_uint32 opus_cpu_capabilities(void)
 # endif
     }
 
+#if defined(OPUS_ARM_PRESUME_AARCH64_NEON_INTR)
+    flags |= OPUS_CPU_ARM_EDSP_FLAG | OPUS_CPU_ARM_MEDIA_FLAG | OPUS_CPU_ARM_NEON_FLAG;
+# if defined(OPUS_ARM_PRESUME_DOTPROD)
+    flags |= OPUS_CPU_ARM_DOTPROD_FLAG;
+# endif
+#endif
+
     fclose(cpuinfo);
   }
   return flags;
@@ -180,7 +193,13 @@ static int opus_select_arch_impl(void)
   }
   arch++;
 
-  celt_assert(arch == OPUS_ARCH_ARM_NEON);
+  if(!(flags & OPUS_CPU_ARM_DOTPROD_FLAG)) {
+    celt_assert(arch == OPUS_ARCH_ARM_NEON);
+    return arch;
+  }
+  arch++;
+
+  celt_assert(arch == OPUS_ARCH_ARM_DOTPROD);
   return arch;
 }
 
diff --git a/celt/arm/armcpu.h b/celt/arm/armcpu.h
index 820262ff..6d5803d8 100644
--- a/celt/arm/armcpu.h
+++ b/celt/arm/armcpu.h
@@ -46,6 +46,12 @@
 #  define MAY_HAVE_NEON(name) MAY_HAVE_MEDIA(name)
 # endif
 
+# if defined(OPUS_ARM_MAY_HAVE_DOTPROD)
+#  define MAY_HAVE_DOTPROD(name) name ## _dotprod
+# else
+#  define MAY_HAVE_DOTPROD(name) MAY_HAVE_NEON(name)
+# endif
+
 # if defined(OPUS_ARM_PRESUME_EDSP)
 #  define PRESUME_EDSP(name) name ## _edsp
 # else
@@ -64,6 +70,12 @@
 #  define PRESUME_NEON(name) PRESUME_MEDIA(name)
 # endif
 
+# if defined(OPUS_ARM_PRESUME_DOTPROD)
+#  define PRESUME_DOTPROD(name) name ## _dotprod
+# else
+#  define PRESUME_DOTPROD(name) PRESUME_NEON(name)
+# endif
+
 # if defined(OPUS_HAVE_RTCD)
 int opus_select_arch(void);
 
@@ -71,6 +83,7 @@ int opus_select_arch(void);
 #define OPUS_ARCH_ARM_EDSP  (1)
 #define OPUS_ARCH_ARM_MEDIA (2)
 #define OPUS_ARCH_ARM_NEON  (3)
+#define OPUS_ARCH_ARM_DOTPROD  (4)
 
 # endif
 
diff --git a/celt/cpu_support.h b/celt/cpu_support.h
index fdd9fb64..9f13d8ae 100644
--- a/celt/cpu_support.h
+++ b/celt/cpu_support.h
@@ -35,13 +35,14 @@
   (defined(OPUS_ARM_ASM) || defined(OPUS_ARM_MAY_HAVE_NEON_INTR))
 #include "arm/armcpu.h"
 
-/* We currently support 4 ARM variants:
+/* We currently support 5 ARM variants:
  * arch[0] -> ARMv4
  * arch[1] -> ARMv5E
  * arch[2] -> ARMv6
  * arch[3] -> NEON
+ * arch[4] -> NEON+DOTPROD
  */
-#define OPUS_ARCHMASK 3
+#define OPUS_ARCHMASK 7
 
 #elif defined(OPUS_HAVE_RTCD) && \
   ((defined(OPUS_X86_MAY_HAVE_SSE) && !defined(OPUS_X86_PRESUME_SSE)) || \