From 0365b59188029dbc9070f16674c2dee13ccac4fc Mon Sep 17 00:00:00 2001
From: Christophe Favergeon <Christophe.Favergeon@arm.com>
Date: Mon, 15 Nov 2021 14:22:24 +0100
Subject: [PATCH] CMSIS-DSP: Python wrapper update

New function arm_mat_mult_opt_q31 added to the wrapper.
---
 PythonWrapper/cmsisdsp_pkg/src/cmsismodule.h | 39 ++++++++++++++++++++
 PythonWrapper/testdsp2.py                    | 17 +++++++--
 2 files changed, 52 insertions(+), 4 deletions(-)

diff --git a/PythonWrapper/cmsisdsp_pkg/src/cmsismodule.h b/PythonWrapper/cmsisdsp_pkg/src/cmsismodule.h
index b59a7253..cae19b20 100644
--- a/PythonWrapper/cmsisdsp_pkg/src/cmsismodule.h
+++ b/PythonWrapper/cmsisdsp_pkg/src/cmsismodule.h
@@ -7400,6 +7400,44 @@ cmsis_arm_mat_mult_q31(PyObject *obj, PyObject *args)
   return(NULL);
 }
 
+static PyObject *
+cmsis_arm_mat_mult_opt_q31(PyObject *obj, PyObject *args)
+{
+
+  PyObject *pSrcA=NULL; // input
+  arm_matrix_instance_q31 *pSrcA_converted=NULL; // input
+  PyObject *pSrcB=NULL; // input
+  arm_matrix_instance_q31 *pSrcB_converted=NULL; // input
+  PyObject *pState=NULL; // input
+  q31_t *pState_converted=NULL; // input
+
+  if (PyArg_ParseTuple(args,"OOO",&pSrcA,&pSrcB,&pState))
+  {
+
+    arm_matrix_instance_q31 *pSrcA_converted = q31MatrixFromNumpy(pSrcA);
+    arm_matrix_instance_q31 *pSrcB_converted = q31MatrixFromNumpy(pSrcB);
+    
+    GETARGUMENT(pState,NPY_INT32,int32_t,int32_t);
+    uint32_t row = pSrcA_converted->numRows ;
+    uint32_t column = pSrcB_converted->numCols ;
+    arm_matrix_instance_q31 *pDst_converted = createq31Matrix(row,column);
+
+    arm_status returnValue = arm_mat_mult_opt_q31(pSrcA_converted,pSrcB_converted,pDst_converted,pState_converted);
+    PyObject* theReturnOBJ=Py_BuildValue("i",returnValue);
+    PyObject* pDstOBJ=NumpyArrayFromq31Matrix(pDst_converted);
+
+    PyObject *pythonResult = Py_BuildValue("OO",theReturnOBJ,pDstOBJ);
+
+    Py_DECREF(theReturnOBJ);
+    FREEARGUMENT(pSrcA_converted);
+    FREEARGUMENT(pSrcB_converted);
+    Py_DECREF(pDstOBJ);
+    FREEARGUMENT(pState_converted);
+    return(pythonResult);
+
+  }
+  return(NULL);
+}
 
 static PyObject *
 cmsis_arm_mat_mult_fast_q31(PyObject *obj, PyObject *args)
@@ -17594,6 +17632,7 @@ static PyMethodDef CMSISMLMethods[] = {
 {"arm_mat_mult_q15",  cmsis_arm_mat_mult_q15, METH_VARARGS,""},
 {"arm_mat_mult_fast_q15",  cmsis_arm_mat_mult_fast_q15, METH_VARARGS,""},
 {"arm_mat_mult_q31",  cmsis_arm_mat_mult_q31, METH_VARARGS,""},
+{"arm_mat_mult_opt_q31",  cmsis_arm_mat_mult_opt_q31, METH_VARARGS,""},
 {"arm_mat_mult_fast_q31",  cmsis_arm_mat_mult_fast_q31, METH_VARARGS,""},
 {"arm_mat_sub_f32",  cmsis_arm_mat_sub_f32, METH_VARARGS,""},
 {"arm_mat_sub_q15",  cmsis_arm_mat_sub_q15, METH_VARARGS,""},
diff --git a/PythonWrapper/testdsp2.py b/PythonWrapper/testdsp2.py
index 99886f85..0a56bf32 100755
--- a/PythonWrapper/testdsp2.py
+++ b/PythonWrapper/testdsp2.py
@@ -162,6 +162,7 @@ print(c[1])
 
 printSubTitle("Fixed point")
 
+printSubTitle(" F32")
 normalizationFactor=2.0*np.sqrt(np.max(np.abs(c[1])))
 a = a / normalizationFactor
 b = b / normalizationFactor
@@ -171,9 +172,16 @@ print(c[1])
 print("")
 af = f.toQ31(a)
 bf = f.toQ31(b)
-c = dsp.arm_mat_mult_q31(af,bf)
-print(f.Q31toF32(c[1]))
+nbSamples = a.size
+tmp=np.zeros(nbSamples)
+c1 = dsp.arm_mat_mult_q31(af,bf)
+c2 = dsp.arm_mat_mult_opt_q31(af,bf,tmp)
+printSubTitle(" Q31")
+print(f.Q31toF32(c1[1]))
+print(f.Q31toF32(c2[1]))
 
+
+printSubTitle(" Q15")
 print("")
 af = f.toQ15(a)
 bf = f.toQ15(b)
@@ -183,6 +191,7 @@ tmp=np.zeros(nb)
 c = dsp.arm_mat_mult_q15(af,bf,tmp)
 print(f.Q15toF32(c[1]))
 
+printSubTitle(" Q7")
 print("")
 af = f.toQ7(a)
 bf = f.toQ7(b)
@@ -304,7 +313,7 @@ NBSAMPLES=10
 printSubTitle("u32")
 su32A=genBitvectors(NBSAMPLES,31)
 su32B=genBitvectors(NBSAMPLES,31)
-ffff = (np.ones(NBSAMPLES)*(-1)).astype(np.int)
+ffff = (np.ones(NBSAMPLES)*(-1)).astype(int)
 
 
 ref=np.bitwise_and(su32A, su32B)
@@ -331,7 +340,7 @@ printSubTitle("u16")
 su16A=genBitvectors(NBSAMPLES,15)
 su16B=genBitvectors(NBSAMPLES,15)
 
-ffff = (np.ones(NBSAMPLES)*(-1)).astype(np.int)
+ffff = (np.ones(NBSAMPLES)*(-1)).astype(int)
 
 
 ref=np.bitwise_and(su16A, su16B)