-void x265_cvt32to16_shr_sse2(int16_t * dst, int *src, intptr_t, int, int);
-void x265_cvt32to16_shl_4_sse2(int16_t * dst, int *src, intptr_t, int);
-void x265_cvt32to16_shl_8_sse2(int16_t * dst, int *src, intptr_t, int);
-void x265_cvt32to16_shl_16_sse2(int16_t * dst, int *src, intptr_t, int);
-void x265_cvt32to16_shl_32_sse2(int16_t * dst, int *src, intptr_t, int);
-void x265_cvt32to16_shl_4_avx2(int16_t * dst, int *src, intptr_t, int);
-void x265_cvt32to16_shl_8_avx2(int16_t * dst, int *src, intptr_t, int);
-void x265_cvt32to16_shl_16_avx2(int16_t * dst, int *src, intptr_t, int);
-void x265_cvt32to16_shl_32_avx2(int16_t * dst, int *src, intptr_t, int);
-void x265_cvt16to32_shl_sse4(int32_t * dst, int16_t * src, intptr_t, int32_t, int32_t);
-void x265_cvt16to32_shr_4_sse4(int32_t * dst, int16_t * src, intptr_t, int32_t, int32_t);
-void x265_cvt16to32_shr_8_sse4(int32_t * dst, int16_t * src, intptr_t, int32_t, int32_t);
-void x265_cvt16to32_shr_16_sse4(int32_t * dst, int16_t * src, intptr_t, int32_t, int32_t);
-void x265_cvt16to32_shr_32_sse4(int32_t * dst, int16_t * src, intptr_t, int32_t, int32_t);
-void x265_copy_shr_sse4(int16_t * dst, int16_t *src, intptr_t, int, int);
-void x265_copy_shl_4_sse2(int16_t * dst, int16_t *src, intptr_t, int);
-void x265_copy_shl_8_sse2(int16_t * dst, int16_t *src, intptr_t, int);
-void x265_copy_shl_16_sse2(int16_t * dst, int16_t *src, intptr_t, int);
-void x265_copy_shl_32_sse2(int16_t * dst, int16_t *src, intptr_t, int);
-uint32_t x265_copy_cnt_4_sse4(int16_t * dst, int16_t * src, intptr_t);
-uint32_t x265_copy_cnt_8_sse4(int16_t * dst, int16_t * src, intptr_t);
-uint32_t x265_copy_cnt_16_sse4(int16_t * dst, int16_t * src, intptr_t);
-uint32_t x265_copy_cnt_32_sse4(int16_t * dst, int16_t * src, intptr_t);
-uint32_t x265_copy_cnt_4_avx2(int16_t * dst, int16_t * src, intptr_t);
-uint32_t x265_copy_cnt_8_avx2(int16_t * dst, int16_t * src, intptr_t);
-uint32_t x265_copy_cnt_16_avx2(int16_t * dst, int16_t * src, intptr_t);
-uint32_t x265_copy_cnt_32_avx2(int16_t * dst, int16_t * src, intptr_t);
+void x265_cpy2Dto1D_shl_4_sse2(int16_t* dst, const int16_t* src, intptr_t srcStride, int shift);
+void x265_cpy2Dto1D_shl_8_sse2(int16_t* dst, const int16_t* src, intptr_t srcStride, int shift);
+void x265_cpy2Dto1D_shl_16_sse2(int16_t* dst, const int16_t* src, intptr_t srcStride, int shift);
+void x265_cpy2Dto1D_shl_32_sse2(int16_t* dst, const int16_t* src, intptr_t srcStride, int shift);
+void x265_cpy2Dto1D_shr_4_sse2(int16_t* dst, const int16_t* src, intptr_t srcStride, int shift);
+void x265_cpy2Dto1D_shr_8_sse2(int16_t* dst, const int16_t* src, intptr_t srcStride, int shift);
+void x265_cpy2Dto1D_shr_16_sse2(int16_t* dst, const int16_t* src, intptr_t srcStride, int shift);
+void x265_cpy2Dto1D_shr_32_sse2(int16_t* dst, const int16_t* src, intptr_t srcStride, int shift);
+void x265_cpy1Dto2D_shl_4_avx2(int16_t* dst, const int16_t* src, intptr_t srcStride, int shift);
+void x265_cpy1Dto2D_shl_8_avx2(int16_t* dst, const int16_t* src, intptr_t srcStride, int shift);
+void x265_cpy1Dto2D_shl_16_avx2(int16_t* dst, const int16_t* src, intptr_t srcStride, int shift);
+void x265_cpy1Dto2D_shl_32_avx2(int16_t* dst, const int16_t* src, intptr_t srcStride, int shift);
+void x265_cpy1Dto2D_shl_4_sse2(int16_t* dst, const int16_t* src, intptr_t dstStride, int shift);
+void x265_cpy1Dto2D_shl_8_sse2(int16_t* dst, const int16_t* src, intptr_t dstStride, int shift);
+void x265_cpy1Dto2D_shl_16_sse2(int16_t* dst, const int16_t* src, intptr_t dstStride, int shift);
+void x265_cpy1Dto2D_shl_32_sse2(int16_t* dst, const int16_t* src, intptr_t dstStride, int shift);
+void x265_cpy1Dto2D_shr_4_avx2(int16_t* dst, const int16_t* src, intptr_t srcStride, int shift);
+void x265_cpy1Dto2D_shr_8_avx2(int16_t* dst, const int16_t* src, intptr_t dstStride, int shift);
+void x265_cpy1Dto2D_shr_16_avx2(int16_t* dst, const int16_t* src, intptr_t dstStride, int shift);
+void x265_cpy1Dto2D_shr_32_avx2(int16_t* dst, const int16_t* src, intptr_t dstStride, int shift);
+void x265_cpy1Dto2D_shr_4_sse2(int16_t* dst, const int16_t* src, intptr_t dstStride, int shift);
+void x265_cpy1Dto2D_shr_8_sse2(int16_t* dst, const int16_t* src, intptr_t dstStride, int shift);
+void x265_cpy1Dto2D_shr_16_sse2(int16_t* dst, const int16_t* src, intptr_t dstStride, int shift);
+void x265_cpy1Dto2D_shr_32_sse2(int16_t* dst, const int16_t* src, intptr_t dstStride, int shift);
+uint32_t x265_copy_cnt_4_sse4(int16_t* dst, const int16_t* src, intptr_t srcStride);
+uint32_t x265_copy_cnt_8_sse4(int16_t* dst, const int16_t* src, intptr_t srcStride);
+uint32_t x265_copy_cnt_16_sse4(int16_t* dst, const int16_t* src, intptr_t srcStride);
+uint32_t x265_copy_cnt_32_sse4(int16_t* dst, const int16_t* src, intptr_t srcStride);
+uint32_t x265_copy_cnt_4_avx2(int16_t* dst, const int16_t* src, intptr_t srcStride);
+uint32_t x265_copy_cnt_8_avx2(int16_t* dst, const int16_t* src, intptr_t srcStride);
+uint32_t x265_copy_cnt_16_avx2(int16_t* dst, const int16_t* src, intptr_t srcStride);
+uint32_t x265_copy_cnt_32_avx2(int16_t* dst, const int16_t* src, intptr_t srcStride);