NMSIS/DSP: Fix riscv_scale_q15 functional error with rv64p optimization

If the scale factor is negtive, the hight 16 bits of q31_t are all "1",
pack two 16 bits value by "OR" will lead to missing the high part value.

Signed-off-by: Jiandong Qiu <qiujiandong@nucleisys.com>

Author: qiujiandong

NMSIS/DSP/Source/BasicMathFunctions/riscv_scale_q15.c

@@ -81,7 +81,7 @@ RISCV_DSP_ATTRIBUTE void riscv_scale_q15(
 #if __RISCV_XLEN == 64
   q31_t out32, tmpScale32;
   q63_t out64;
-  tmpScale32 = (q31_t)scaleFract | ((q31_t)scaleFract << 16);
+  tmpScale32 = __RV_PKBB16(scaleFract, scaleFract);
 #endif /* __RISCV_XLEN == 64 */
   q31_t inA1, inA2;
   q31_t out1, out2, out3, out4;                  /* Temporary output variables */

NMSIS/doc/source/changelog.rst

@@ -50,6 +50,7 @@ This is the version of ``V1.5.0`` release.
   - Fix ``riscv_cfft_radix4_q31.c`` producing incorrect results
   - Fix ``riscv_power_q31`` output scaling, which should be in Q2.48 format
   - Fix ``riscv_cfft_q15`` using the wrong variable
+  - Fix ``riscv_scale_q15`` functional error when optimized with rv64 P extension
   - Add ``riscv_vec_math.h`` and ``riscv_vec_math_f16.h`` providing f16/f32 RVV-accelerated math functions (exp, log, sin, cos, pow, tanh, etc.)
   - Add multiple library build configurations for Nuclei n100 and n200 CPU
   - Synchronize with the CMSIS-DSP v1.17.0 release, incorporating the following updates: