From 8967816df2b16192a435cbdfaa192138a5025810 Mon Sep 17 00:00:00 2001
From: Christophe Favergeon <Christophe.Favergeon@arm.com>
Date: Mon, 26 Apr 2021 15:05:54 +0200
Subject: [PATCH] CMSIS-DSP: Improved Helium implementation of fir q7.

---
 Source/FilteringFunctions/arm_fir_q7.c | 44 +++++++++++++++++++++++---
 1 file changed, 40 insertions(+), 4 deletions(-)
 mode change 100644 => 100755 Source/FilteringFunctions/arm_fir_q7.c

diff --git a/Source/FilteringFunctions/arm_fir_q7.c b/Source/FilteringFunctions/arm_fir_q7.c
old mode 100644
new mode 100755
index 3b5d8426..e2972cdb
--- a/Source/FilteringFunctions/arm_fir_q7.c
+++ b/Source/FilteringFunctions/arm_fir_q7.c
@@ -141,7 +141,27 @@
     while (blkCnt > 0);                                                                     \
 }
 
-static void arm_fir_q7_17_32_mve(const arm_fir_instance_q7 * S, 
+
+static void arm_fir_q7_49_64_mve(const arm_fir_instance_q7 * S,
+  const q7_t * __restrict pSrc,
+  q7_t * __restrict pDst, uint32_t blockSize)
+{
+    #define NBTAPS 64
+    FIR_Q7_MAIN_CORE();
+    #undef NBTAPS
+}
+
+
+void arm_fir_q7_33_48_mve(const arm_fir_instance_q7 * S,
+  const q7_t * __restrict pSrc,
+  q7_t * __restrict pDst, uint32_t blockSize)
+{
+    #define NBTAPS 48
+    FIR_Q7_MAIN_CORE();
+    #undef NBTAPS
+}
+
+static void arm_fir_q7_17_32_mve(const arm_fir_instance_q7 * S,
   const q7_t * __restrict pSrc,
   q7_t * __restrict pDst, uint32_t blockSize)
 {
@@ -151,8 +171,8 @@ static void arm_fir_q7_17_32_mve(const arm_fir_instance_q7 * S,
 }
 
 
-void arm_fir_q7_1_16_mve(const arm_fir_instance_q7 * S, 
-  const q7_t * __restrict pSrc, 
+void arm_fir_q7_1_16_mve(const arm_fir_instance_q7 * S,
+  const q7_t * __restrict pSrc,
   q7_t * __restrict pDst, uint32_t blockSize)
 {
     #define NBTAPS 16
@@ -196,6 +216,22 @@ void arm_fir_q7(
         arm_fir_q7_17_32_mve(S, pSrc, pDst, blockSize);
         return;
     }
+    else if (numTaps <= 48)
+    {
+        /*
+         * [33 to 48 taps] specialized routine
+         */
+        arm_fir_q7_33_48_mve(S, pSrc, pDst, blockSize);
+        return;
+    }
+    else if (numTaps <= 64)
+    {
+        /*
+         * [49 to 64 taps] specialized routine
+         */
+        arm_fir_q7_49_64_mve(S, pSrc, pDst, blockSize);
+        return;
+    }
 
     /*
      * pState points to state array which contains previous frame (numTaps - 1) samples
@@ -607,7 +643,7 @@ void arm_fir_q7(
     {
       acc0 += (q15_t) * (px++) * (*(pb++));
       i--;
-    } 
+    }
 
     /* The result is in 2.14 format. Convert to 1.7
        Then store the output in the destination buffer. */