DSP: Update DSP support source code

Signed-off-by: Huaqi Fang <578567190@qq.com>
Nuclei-Software · Jul 24, 2020 · 9ba880a · 9ba880a
1 parent 72f443d
commit 9ba880a
Show file tree

Hide file tree

Showing 13 changed files with 57 additions and 28 deletions.
diff --git a/NMSIS/DSP/Source/FilteringFunctions/riscv_conv_fast_opt_q15.c b/NMSIS/DSP/Source/FilteringFunctions/riscv_conv_fast_opt_q15.c
@@ -116,7 +116,7 @@ void riscv_conv_fast_opt_q15(
   }
 #if __RISCV_XLEN == 64
   /* Pointer to take end of scratch2 buffer */
-  pScr2 = pScratch2 + srcBLen - 4;
+  pScr2 = pScratch2 + srcBLen - 3;
 #else
   /* Pointer to take end of scratch2 buffer */
   pScr2 = pScratch2 + srcBLen - 1;

diff --git a/NMSIS/DSP/Source/FilteringFunctions/riscv_conv_fast_q15.c b/NMSIS/DSP/Source/FilteringFunctions/riscv_conv_fast_q15.c
@@ -215,7 +215,9 @@ void riscv_conv_fast_q15(
       /* Decrement loop counter */
       k--;
     }
-
+#if defined RISCV_MATH_DSP && __RISCV_XLEN == 64
+    py += 2;
+#endif /* defined RISCV_MATH_DSP && __RISCV_XLEN == 64 */
     /* For the next MAC operations, the pointer py is used without SIMD
      * So, py is incremented by 1 */
     py = py + 1U;

diff --git a/NMSIS/DSP/Source/FilteringFunctions/riscv_conv_opt_q15.c b/NMSIS/DSP/Source/FilteringFunctions/riscv_conv_opt_q15.c
@@ -122,7 +122,7 @@ void riscv_conv_opt_q15(
   /* Loop unrolling: Compute 4 outputs at a time */
   k = srcBLen >> 2U;
 #if __RISCV_XLEN == 64
-  pScr2 -= 4;
+  pScr2 -= 3;
 #endif /* __RISCV_XLEN == 64 */
   /* Copy smaller length input sequence in reverse order into second scratch buffer */
   while (k > 0U)
@@ -140,7 +140,9 @@ void riscv_conv_opt_q15(
     /* Decrement loop counter */
     k--;
   }
-
+#if __RISCV_XLEN == 64
+  pScr2 += 3;
+#endif /* __RISCV_XLEN == 64 */
   /* Loop unrolling: Compute remaining outputs */
   k = srcBLen % 0x4U;
 

diff --git a/NMSIS/DSP/Source/FilteringFunctions/riscv_conv_partial_fast_opt_q15.c b/NMSIS/DSP/Source/FilteringFunctions/riscv_conv_partial_fast_opt_q15.c
@@ -126,7 +126,7 @@ riscv_status riscv_conv_partial_fast_opt_q15(
 
 #if __RISCV_XLEN == 64
   /* Pointer to take end of scratch2 buffer */
-  pScr2 = pScratch2 + srcBLen - 5;
+  pScr2 = pScratch2 + srcBLen - 4;
 #else
    /* Pointer to take end of scratch2 buffer */
     pScr2 = pScratch2 + srcBLen - 1;

diff --git a/NMSIS/DSP/Source/FilteringFunctions/riscv_conv_partial_fast_q15.c b/NMSIS/DSP/Source/FilteringFunctions/riscv_conv_partial_fast_q15.c
@@ -227,7 +227,9 @@ riscv_status riscv_conv_partial_fast_q15(
         /* Decrement loop counter */
         k--;
       }
-
+#if defined RISCV_MATH_DSP && __RISCV_XLEN == 64
+       py += 2;
+#endif /* RISCV_MATH_DSP && __RISCV_XLEN == 64 */
       /* For the next MAC operations, the pointer py is used without SIMD
          So, py is incremented by 1 */
       py = py + 1U;

diff --git a/NMSIS/DSP/Source/FilteringFunctions/riscv_conv_partial_opt_q15.c b/NMSIS/DSP/Source/FilteringFunctions/riscv_conv_partial_opt_q15.c
@@ -133,7 +133,7 @@ riscv_status riscv_conv_partial_opt_q15(
 
 #if defined (RISCV_MATH_LOOPUNROLL)
 #if __RISCV_XLEN == 64
-  pScr2 -= 4;
+  pScr2 -= 3;
 #endif /* __RISCV_XLEN == 64 */
     /* Loop unrolling: Compute 4 outputs at a time */
     k = srcBLen >> 2U;
@@ -153,7 +153,9 @@ riscv_status riscv_conv_partial_opt_q15(
       /* Decrement loop counter */
       k--;
     }
-
+#if __RISCV_XLEN == 64
+  pScr2 += 3;
+#endif /* __RISCV_XLEN == 64 */
     /* Loop unrolling: Compute remaining outputs */
     k = srcBLen % 0x4U;
 

diff --git a/NMSIS/DSP/Source/FilteringFunctions/riscv_conv_partial_opt_q7.c b/NMSIS/DSP/Source/FilteringFunctions/riscv_conv_partial_opt_q7.c
@@ -168,9 +168,6 @@ riscv_status riscv_conv_partial_opt_q7(
      ** a second loop below copies for the remaining 1 to 3 samples. */
     while (k > 0U)
     {
-#if __RISCV_XLEN == 64
-    write_q15x4_ia(&pScr2,read_q15x4_da((q15_t **)&px));
-#else
       /* copy second buffer in reversal manner */
       x4 = (q15_t) *pIn1++;
       *pScr1++ = x4;
@@ -180,7 +177,6 @@ riscv_status riscv_conv_partial_opt_q7(
       *pScr1++ = x4;
       x4 = (q15_t) *pIn1++;
       *pScr1++ = x4;
-#endif /* __RISCV_XLEN == 64 */
       /* Decrement loop counter */
       k--;
     }

diff --git a/NMSIS/DSP/Source/FilteringFunctions/riscv_conv_partial_q15.c b/NMSIS/DSP/Source/FilteringFunctions/riscv_conv_partial_q15.c
@@ -260,9 +260,9 @@ riscv_status riscv_conv_partial_q15(
       /* Update the inputA and inputB pointers for next MAC calculation */
       py = ++pSrc2 - 1U;
       px = pIn1;
-#if __RISCV_XLEN == 64
-      py = py - 2;
-#endif /* __RISCV_XLEN == 64 */
+// #if __RISCV_XLEN == 64
+//       py = py - 2;
+// #endif /* __RISCV_XLEN == 64 */
       /* Increment MAC count */
       count++;
 

diff --git a/NMSIS/DSP/Source/FilteringFunctions/riscv_conv_partial_q31.c b/NMSIS/DSP/Source/FilteringFunctions/riscv_conv_partial_q31.c
@@ -209,7 +209,9 @@ riscv_status riscv_conv_partial_q31(
         /* Decrement loop counter */
         k--;
       }
-
+#if __RISCV_XLEN == 64
+    py++;
+#endif /* __RISCV_XLEN == 64 */
       /* Loop unrolling: Compute remaining outputs */
       k = count % 0x4U;
 
@@ -414,7 +416,9 @@ riscv_status riscv_conv_partial_q31(
 
         /* Loop unrolling: Compute 4 outputs at a time */
         k = srcBLen >> 2U;
-
+#if __RISCV_XLEN == 64
+    py--;
+#endif /* __RISCV_XLEN == 64 */
         while (k > 0U)
         {
 #if __RISCV_XLEN == 64
@@ -435,7 +439,9 @@ riscv_status riscv_conv_partial_q31(
           /* Decrement loop counter */
           k--;
         }
-
+#if __RISCV_XLEN == 64
+    py++;
+#endif /* __RISCV_XLEN == 64 */
         /* Loop unrolling: Compute remaining outputs */
         k = srcBLen % 0x4U;
 
@@ -572,7 +578,9 @@ riscv_status riscv_conv_partial_q31(
         /* Decrement loop counter */
         k--;
       }
-
+#if __RISCV_XLEN == 64
+    py++;
+#endif /* __RISCV_XLEN == 64 */
       /* Loop unrolling: Compute remaining outputs */
       k = count % 0x4U;
 

diff --git a/NMSIS/DSP/Source/FilteringFunctions/riscv_conv_partial_q7.c b/NMSIS/DSP/Source/FilteringFunctions/riscv_conv_partial_q7.c
@@ -224,7 +224,9 @@ riscv_status riscv_conv_partial_q7(
         /* Decrement loop counter */
         k--;
       }
-
+#if __RISCV_XLEN == 64
+    py += 7;
+#endif /* __RISCV_XLEN == 64 */
       /* Loop unrolling: Compute remaining outputs */
       k = count % 0x4U;
 

diff --git a/NMSIS/DSP/Source/FilteringFunctions/riscv_conv_q31.c b/NMSIS/DSP/Source/FilteringFunctions/riscv_conv_q31.c
@@ -194,7 +194,9 @@ void riscv_conv_q31(
       /* Decrement loop counter */
       k--;
     }
-
+#if __RISCV_XLEN == 64
+    py++;
+#endif /* __RISCV_XLEN == 64 */
     /* Loop unrolling: Compute remaining outputs */
     k = count % 0x4U;
 
@@ -410,7 +412,9 @@ void riscv_conv_q31(
         /* Decrement loop counter */
         k--;
       }
-
+#if __RISCV_XLEN == 64
+    py++;
+#endif /* __RISCV_XLEN == 64 */
       /* Loop unrolling: Compute remaining outputs */
       k = srcBLen % 0x4U;
 
@@ -548,7 +552,9 @@ void riscv_conv_q31(
       /* Decrement loop counter */
       k--;
     }
-
+#if __RISCV_XLEN == 64
+    py++;
+#endif /* __RISCV_XLEN == 64 */
     /* Loop unrolling: Compute remaining outputs */
     k = blockSize3 % 0x4U;
 

diff --git a/NMSIS/DSP/Source/FilteringFunctions/riscv_fir_decimate_fast_q15.c b/NMSIS/DSP/Source/FilteringFunctions/riscv_fir_decimate_fast_q15.c
@@ -82,7 +82,7 @@ void riscv_fir_decimate_fast_q15(
 #if defined (RISCV_MATH_LOOPUNROLL)
         q31_t c1;                                      /* Temporary variables to hold state and coefficient values */
 #if __RISCV_XLEN == 64
-        q63_t x064, x164, c064;
+        q63_t x064, x164, c064, sum064 = 0;
 #endif /* __RISCV_XLEN == 64 */
 #endif
 
@@ -228,6 +228,15 @@ void riscv_fir_decimate_fast_q15(
 
     while (tapCnt > 0U)
     {
+#if __RISCV_XLEN == 64
+      /* Read the b[numTaps-1] and b[numTaps-2] coefficients */
+      c064 = read_q15x4_ia ((q15_t **) &pb);
+
+      /* Read x[n-numTaps-1] and x[n-numTaps-2] sample */
+      x064 = read_q15x4_ia ((q15_t **) &px);
+
+      sum064 = __SMLAD(x064, c064, sum064);
+#else
       /* Read the b[numTaps-1] and b[numTaps-2] coefficients */
       c0 = read_q15x2_ia ((q15_t **) &pb);
 
@@ -245,11 +254,13 @@ void riscv_fir_decimate_fast_q15(
 
       /* Perform the multiply-accumulate */
       sum0 = __SMLAD(x0, c1, sum0);
-
+#endif
       /* Decrement loop counter */
       tapCnt--;
     }
-
+#if __RISCV_XLEN == 64
+    sum0 =(q31_t)((sum064 + (sum064<<32u))>>32u);
+#endif
     /* Loop unrolling: Compute remaining taps */
     tapCnt = numTaps % 0x4U;
 

diff --git a/NMSIS/DSP/Test/ControllerFunctions/sincospart/riscv_controller.c b/NMSIS/DSP/Test/ControllerFunctions/sincospart/riscv_controller.c
@@ -29,7 +29,6 @@ static int DSP_SIN_COS_F32(void)
     riscv_sin_cos_f32(0, &pSinVal, &pCosVal);
     BENCH_END(riscv_sin_cos_f32);
     ref_sin_cos_f32(0, &pSinVal_ref, &pCosVal_ref);
-    printf("%-20s : cycle: %d\n", "riscv_sin_cos_f32", cycle);
     if ((fabs(pSinVal - pSinVal_ref) > DELTAF32) ||
         (fabs(pCosVal - pCosVal_ref) > DELTAF32)) {
         BENCH_ERROR(riscv_sin_cos_f32);
@@ -50,7 +49,6 @@ static int DSP_SIN_COS_Q31(void)
     riscv_sin_cos_q31(0, &pSinVal, &pCosVal);
     BENCH_END(riscv_sin_cos_q31);
     ref_sin_cos_q31(0, &pSinVal_ref, &pCosVal_ref);
-    printf("%-20s : cycle: %d\n", "riscv_sin_cos_q31", cycle);
     if ((labs(pSinVal - pSinVal_ref) > DELTAQ31) ||
         (labs(pCosVal - pCosVal_ref) > DELTAQ31)) {
         BENCH_ERROR(riscv_sin_cos_q31);