shithub: opus

Info  •  Files  •  Log  •  Branches

small fix for supporting neon for dnn in Windwos ARM64

Signed-off-by: Jean-Marc Valin <jmvalin@jmvalin.ca>

--- a/cmake/OpusConfig.cmake

+++ b/cmake/OpusConfig.cmake

@@ -56,7 +56,7 @@

   else()

     set(OPUS_CPU_X86 1)

   endif()

-elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "(arm|aarch64)")

+elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "(arm|aarch64|ARM)")

   set(OPUS_CPU_ARM 1)

 endif()

--- a/cmake/OpusFunctions.cmake

+++ b/cmake/OpusFunctions.cmake

@@ -134,7 +134,7 @@

 endfunction()

 function(opus_detect_neon COMPILER_SUPPORT_NEON)

-  if(CMAKE_SYSTEM_PROCESSOR MATCHES "(arm|aarch64)")

+  if(CMAKE_SYSTEM_PROCESSOR MATCHES "(arm|aarch64|ARM)")

     message(STATUS "Check NEON support by compiler")

     check_include_file(arm_neon.h HAVE_ARM_NEON_H)

     if(HAVE_ARM_NEON_H)

--- a/dnn/vec_neon.h

+++ b/dnn/vec_neon.h

@@ -402,12 +402,12 @@

       for (;j<cols-4;j+=8)

       {

          int8x16_t vw0, vw1, vw2, vw3, vx0, vx1;

-         vx0 = (int8x16_t)vld1q_dup_s32((int*)(void*)&x[j]);

+         vx0 = vreinterpretq_s8_s32(vld1q_dup_s32((int*)(void*)&x[j]));

          vw0 = vld1q_s8(w);

          vw1 = vld1q_s8(&w[16]);

          acc0 = vdotprod(acc0, vw0, vx0);

          acc1 = vdotprod(acc1, vw1, vx0);

-         vx1 = (int8x16_t)vld1q_dup_s32((int*)(void*)&x[j+4]);

+         vx1 = vreinterpretq_s8_s32(vld1q_dup_s32((int*)(void*)&x[j+4]));

          vw2 = vld1q_s8(&w[32]);

          vw3 = vld1q_s8(&w[48]);

          acc2 = vdotprod(acc2, vw2, vx1);

@@ -419,7 +419,7 @@

       for (;j<cols;j+=4)

       {

          int8x16_t vw0, vw1, vx;

-         vx = (int8x16_t)vld1q_dup_s32((int*)(void*)&x[j]);

+         vx = vreinterpretq_s8_s32(vld1q_dup_s32((int*)(void*)&x[j]));

          vw0 = vld1q_s8(w);

          vw1 = vld1q_s8(&w[16]);

          acc0 = vdotprod(acc0, vw0, vx);

@@ -457,7 +457,7 @@

          int pos;

          pos = (*idx++);

          int8x16_t vw0, vw1, vx;

-         vx = (int8x16_t)vld1q_dup_s32((int*)(void*)&x[pos]);

+         vx = vreinterpretq_s8_s32(vld1q_dup_s32((int*)(void*)&x[pos]));

          vw0 = vld1q_s8(w);

          vw1 = vld1q_s8(&w[16]);

          acc0 = vdotprod(acc0, vw0, vx);

-- 

⑨