From cb15ffc3c756679092da19c96b9f9b4c0e2dc02c Mon Sep 17 00:00:00 2001
From: John Romein <romein@astron.nl>
Date: Fri, 11 Feb 2011 15:58:19 +0000
Subject: [PATCH] bug 225: Added little-endian int->float conversion routines,
necessary to bypass PPF bank.
---
.gitattributes | 1 +
RTCP/CNProc/src/FIR_Asm.S | 266 ++++++++++++++++++++++++++++++++
RTCP/CNProc/src/FIR_Asm.h | 3 +
RTCP/CNProc/src/PPF.h | 8 +-
RTCP/CNProc/test/CMakeLists.txt | 1 +
RTCP/CNProc/test/tFIR_Asm.cc | 83 ++++++++++
6 files changed, 358 insertions(+), 4 deletions(-)
create mode 100644 RTCP/CNProc/test/tFIR_Asm.cc
diff --git a/.gitattributes b/.gitattributes
index 66b43f80648..f871b594a5a 100644
--- a/.gitattributes
+++ b/.gitattributes
@@ -2079,6 +2079,7 @@ RTCP/CNProc/test/filterTestResult.ps -text
RTCP/CNProc/test/inversePPFTestOutput.dat -text
RTCP/CNProc/test/inversePPFTestResult.ps -text
RTCP/CNProc/test/tDedispersion.cc -text
+RTCP/CNProc/test/tFIR_Asm.cc -text
RTCP/CNProc/test/tInversePPF.cc -text
RTCP/CNProc/test/tPencilBeamFormer.cc -text
RTCP/CNProc/test/tPencilBeamFormer.sh -text
diff --git a/RTCP/CNProc/src/FIR_Asm.S b/RTCP/CNProc/src/FIR_Asm.S
index 320fc2df4db..a2ac5e52efc 100644
--- a/RTCP/CNProc/src/FIR_Asm.S
+++ b/RTCP/CNProc/src/FIR_Asm.S
@@ -2509,4 +2509,270 @@ _fast_memcpy:
lfpdux 14,1,8
addi 1,1,16
blr
+
+
+ # little endian i16complex -> float
+
+.global _ZN5LOFAR4RTCP8_convertISt7complexIsEEEvPS2_IfEPKT_j
+_ZN5LOFAR4RTCP8_convertISt7complexIsEEEvPS2_IfEPKT_j:
+
+ stwu 1,-32(1)
+
+ lis 8,sub_value@ha # load sub_values
+ li 9,sub_value@l
+ lfpdx 0,8,9
+
+ lis 12,0x0080
+ ori 12,12,0x0080
+
+ srwi 5,5,1
+ subi 5,5,2
+ mtctr 5
+
+ subi 3,3,8
+
+ li 8,8
+ li 9,4
+ li 10,14
+ li 11,6
+ addi 5,1,16
+
+ stfpdx 0,0,1 # initialize int->fp conversion area
+ stfpdx 0,0,5
+
+ lwz 6,0(4)
+
+ lwzux 7,4,9
+ xor 6,6,12
+ sthbrx 6,10,1
+ srwi 6,6,16
+ sthbrx 6,11,1
+
+ lwzux 6,4,9
+ xor 7,7,12
+ sthbrx 7,10,5
+ srwi 7,7,16
+ sthbrx 7,11,5
+ lfpdx 1,0,1
+
+ lwzux 7,4,9
+ xor 6,6,12
+ sthbrx 6,10,1
+ srwi 6,6,16
+ sthbrx 6,11,1
+ fpsub 1,1,0
+ lfpdx 2,0,5
+
+0:
+ lwzux 6,4,9
+ xor 7,7,12
+ sthbrx 7,10,5
+ srwi 7,7,16
+ sthbrx 7,11,5
+ fpsub 2,2,0
+ stfpsux 1,3,8
+ lfpdx 1,0,1
+
+ lwzux 7,4,9
+ xor 6,6,12
+ sthbrx 6,10,1
+ srwi 6,6,16
+ sthbrx 6,11,1
+ fpsub 1,1,0
+ stfpsux 2,3,8
+ lfpdx 2,0,5
+
+ bdnz 0b
+
+ xor 7,7,12
+ sthbrx 7,10,5
+ srwi 7,7,16
+ sthbrx 7,11,5
+ fpsub 2,2,0
+ stfpsux 1,3,8
+ lfpdx 1,0,1
+
+ fpsub 1,1,0
+ stfpsux 2,3,8
+ lfpdx 2,0,5
+
+ fpsub 2,2,0
+ stfpsux 1,3,8
+
+ stfpsux 2,3,8
+
+ addi 1,1,32
+ blr
+
+
+ # i8complex -> float
+
+.global _ZN5LOFAR4RTCP8_convertISt7complexIaEEEvPS2_IfEPKT_j
+_ZN5LOFAR4RTCP8_convertISt7complexIaEEEvPS2_IfEPKT_j:
+
+ lis 11,_ZN5LOFAR4RTCP13_FIR_fp_tableE@ha
+ la 11,_ZN5LOFAR4RTCP13_FIR_fp_tableE@l(11)
+
+ srwi 5,5,2
+ subi 5,5,2
+ mtctr 5
+
+ subi 3,3,8
+
+ li 10,8
+ li 9,2
+
+ lhz 5,0(4)
+ lhzux 6,4,9
+ lhzux 7,4,9
+ lhzux 8,4,9
+
+ slwi 5,5,3
+ lfpsx 0,11,5
+ lhzux 5,4,9
+
+ slwi 6,6,3
+ lfpsx 1,11,6
+ lhzux 6,4,9
+
+ slwi 7,7,3
+ lfpsx 2,11,7
+ lhzux 7,4,9
+
+ slwi 8,8,3
+ lfpsx 3,11,8
+ lhzux 8,4,9
+
+0:
+ slwi 5,5,3
+ stfpsux 0,3,10
+ lfpsx 0,11,5
+ lhzux 5,4,9
+
+ slwi 6,6,3
+ stfpsux 1,3,10
+ lfpsx 1,11,6
+ lhzux 6,4,9
+
+ slwi 7,7,3
+ stfpsux 2,3,10
+ lfpsx 2,11,7
+ lhzux 7,4,9
+
+ slwi 8,8,3
+ stfpsux 3,3,10
+ lfpsx 3,11,8
+ lhzux 8,4,9
+
+ bdnz 0b
+
+ slwi 5,5,3
+ stfpsux 0,3,10
+ lfpsx 0,11,5
+
+ slwi 6,6,3
+ stfpsux 1,3,10
+ lfpsx 1,11,6
+
+ slwi 7,7,3
+ stfpsux 2,3,10
+ lfpsx 2,11,7
+
+ slwi 8,8,3
+ stfpsux 3,3,10
+ lfpsx 3,11,8
+
+ stfpsux 0,3,10
+ stfpsux 1,3,10
+ stfpsux 2,3,10
+ stfpsux 3,3,10
+
+ blr
+
+
+ # i4complex -> float
+
+.global _ZN5LOFAR4RTCP8_convertINS_5TYPES9i4complexEEEvPSt7complexIfEPKT_j
+_ZN5LOFAR4RTCP8_convertINS_5TYPES9i4complexEEEvPSt7complexIfEPKT_j:
+
+ lis 11,_ZN5LOFAR4RTCP13_FIR_fp_tableE@ha
+ la 11,_ZN5LOFAR4RTCP13_FIR_fp_tableE@l(11)
+
+ srwi 5,5,2
+ subi 5,5,2
+ mtctr 5
+
+ subi 3,3,8
+
+ li 10,8
+ li 9,1
+
+ lbz 5,0(4)
+ lbzux 6,4,9
+ lbzux 7,4,9
+ lbzux 8,4,9
+
+ slwi 5,5,3
+ lfpsx 0,11,5
+ lbzux 5,4,9
+
+ slwi 6,6,3
+ lfpsx 1,11,6
+ lbzux 6,4,9
+
+ slwi 7,7,3
+ lfpsx 2,11,7
+ lbzux 7,4,9
+
+ slwi 8,8,3
+ lfpsx 3,11,8
+ lbzux 8,4,9
+
+0:
+ slwi 5,5,3
+ stfpsux 0,3,10
+ lfpsx 0,11,5
+ lbzux 5,4,9
+
+ slwi 6,6,3
+ stfpsux 1,3,10
+ lfpsx 1,11,6
+ lbzux 6,4,9
+
+ slwi 7,7,3
+ stfpsux 2,3,10
+ lfpsx 2,11,7
+ lbzux 7,4,9
+
+ slwi 8,8,3
+ stfpsux 3,3,10
+ lfpsx 3,11,8
+ lbzux 8,4,9
+
+ bdnz 0b
+
+ slwi 5,5,3
+ stfpsux 0,3,10
+ lfpsx 0,11,5
+
+ slwi 6,6,3
+ stfpsux 1,3,10
+ lfpsx 1,11,6
+
+ slwi 7,7,3
+ stfpsux 2,3,10
+ lfpsx 2,11,7
+
+ slwi 8,8,3
+ stfpsux 3,3,10
+ lfpsx 3,11,8
+
+ stfpsux 0,3,10
+ stfpsux 1,3,10
+ stfpsux 2,3,10
+ stfpsux 3,3,10
+
+ blr
+
+
#endif
diff --git a/RTCP/CNProc/src/FIR_Asm.h b/RTCP/CNProc/src/FIR_Asm.h
index 947bfd57973..1d85d51da62 100644
--- a/RTCP/CNProc/src/FIR_Asm.h
+++ b/RTCP/CNProc/src/FIR_Asm.h
@@ -22,6 +22,7 @@
#define LOFAR_CNPROC_FIR_ASM_H
#if defined HAVE_BGP
+#include <Common/lofar_complex.h>
#include <Interface/Config.h>
namespace LOFAR {
@@ -38,6 +39,8 @@ template <typename SAMPLE_TYPE> extern void _filter(unsigned nrChannels,
fcomplex out[],
int nr_samples_div_16);
+template <typename SAMPLE_TYPE> extern void _convert(fcomplex out[], const SAMPLE_TYPE samples[], unsigned count);
+
extern "C" {
void _transpose_4x8(fcomplex *out,
const fcomplex *in,
diff --git a/RTCP/CNProc/src/PPF.h b/RTCP/CNProc/src/PPF.h
index d4f8769e95b..7db2e8436f8 100644
--- a/RTCP/CNProc/src/PPF.h
+++ b/RTCP/CNProc/src/PPF.h
@@ -38,13 +38,13 @@ template <typename SAMPLE_TYPE> class PPF: boost::noncopyable
void computeFlags(unsigned stat, const SubbandMetaData *metaData, FilteredData *);
void filter(unsigned stat, double centerFrequency, const SubbandMetaData *metaData, const TransposedData<SAMPLE_TYPE> *, FilteredData *);
- private:
- void init_fft(), destroy_fft();
-
#if !defined PPF_C_IMPLEMENTATION
- void initConstantTable();
+ static void initConstantTable();
#endif
+ private:
+ void init_fft(), destroy_fft();
+
#if defined PPF_C_IMPLEMENTATION
fcomplex phaseShift(unsigned time, unsigned chan, double baseFrequency, double delayAtBegin, double delayAfterEnd) const;
#else
diff --git a/RTCP/CNProc/test/CMakeLists.txt b/RTCP/CNProc/test/CMakeLists.txt
index 5e0f7013410..c0dcf82b448 100644
--- a/RTCP/CNProc/test/CMakeLists.txt
+++ b/RTCP/CNProc/test/CMakeLists.txt
@@ -7,6 +7,7 @@ include_directories(${PACKAGE_SOURCE_DIR}/src)
lofar_add_test(tCN_Processing tCN_Processing.cc)
lofar_add_test(tBeamForming tBeamForming.cc)
lofar_add_test(tDedispersion tDedispersion.cc)
+lofar_add_test(tFIR_Asm tFIR_Asm.cc)
lofar_add_test(tPencilBeamFormer tPencilBeamFormer.cc)
lofar_add_test(tStokes tStokes.cc)
lofar_add_test(tInversePPF tInversePPF.cc)
diff --git a/RTCP/CNProc/test/tFIR_Asm.cc b/RTCP/CNProc/test/tFIR_Asm.cc
new file mode 100644
index 00000000000..a9fbc2eda8e
--- /dev/null
+++ b/RTCP/CNProc/test/tFIR_Asm.cc
@@ -0,0 +1,83 @@
+#include <lofar_config.h>
+
+#include <FIR_Asm.h>
+#include <PPF.h>
+#include <Common/Timer.h>
+
+#include <iostream>
+
+
+#if defined HAVE_BGP
+using namespace LOFAR;
+using namespace LOFAR::RTCP;
+using namespace LOFAR::TYPES;
+#endif
+
+#define SIZE 131072
+
+int main()
+{
+#if defined HAVE_BGP
+ {
+ i16complex in[SIZE] = {
+ makei16complex(0x0100, 0x0200),
+ makei16complex(0x0300, 0x0400),
+ makei16complex(0x0500, 0x0600),
+ };
+
+ in[SIZE - 1] = makei16complex(0x0700, 0x0801);
+
+ fcomplex out[SIZE];
+
+ NSTimer timer("little endian i16complex -> float", true);
+ timer.start();
+ _convert(out, in, SIZE);
+ timer.stop();
+
+ std::cout << out[0] << ' ' << out[1] << ' ' << out[2] << ' ' << out[SIZE - 1] << std::endl;
+ }
+
+ {
+ PPF<i8complex>::initConstantTable();
+
+ i8complex in[SIZE] = {
+ makei8complex(1, 2),
+ makei8complex(3, 4),
+ makei8complex(5, 6),
+ };
+
+ in[SIZE - 1] = makei8complex(7, 8);
+
+ fcomplex out[SIZE];
+
+ NSTimer timer("little endian i8complex -> float", true);
+ timer.start();
+ _convert(out, in, SIZE);
+ timer.stop();
+
+ std::cout << out[0] << ' ' << out[1] << ' ' << out[2] << ' ' << out[SIZE - 1] << std::endl;
+ }
+
+ {
+ PPF<i4complex>::initConstantTable();
+
+ i4complex in[SIZE] = {
+ makei4complex(0.5, 1.5),
+ makei4complex(2.5, 3.5),
+ makei4complex(4.5, 5.5),
+ };
+
+ in[SIZE - 1] = makei4complex(-1.5, -0.5);
+
+ fcomplex out[SIZE];
+
+ NSTimer timer("little endian i4complex -> float", true);
+ timer.start();
+ _convert(out, in, SIZE);
+ timer.stop();
+
+ std::cout << out[0] << ' ' << out[1] << ' ' << out[2] << ' ' << out[SIZE - 1] << std::endl;
+ }
+#endif
+ return 0;
+}
--
GitLab