merge from gr36_ni branch

25 files changed, 230 insertions, 0 deletions

diff --git a/volk/orc/volk_16i_s32f_deinterleave_32f_x2_a_orc_impl.orc b/volk/orc/volk_16i_s32f_deinterleave_32f_x2_a_orc_impl.orc
new file mode 100644
index 000000000..fd8915da0
--- /dev/null
+++ b/volk/orc/volk_16i_s32f_deinterleave_32f_x2_a_orc_impl.orc
@@ -0,0 +1,12 @@
+.function volk_16ic_s32f_deinterleave_32f_x2_a_orc_impl
+.dest 4 idst
+.dest 4 qdst
+.source 4 src
+.floatparam 4 scalar
+.temp 8 iql
+.temp 8 iqf
+
+x2 convswl iql, src
+x2 convlf iqf, iql
+x2 divf iqf, iqf, scalar
+splitql qdst, idst, iqf
diff --git a/volk/orc/volk_16ic_deinterleave_16i_x2_a_orc_impl.orc b/volk/orc/volk_16ic_deinterleave_16i_x2_a_orc_impl.orc
new file mode 100644
index 000000000..76faa936a
--- /dev/null
+++ b/volk/orc/volk_16ic_deinterleave_16i_x2_a_orc_impl.orc
@@ -0,0 +1,5 @@
+.function volk_16ic_deinterleave_16i_x2_a_orc_impl
+.dest 2 idst
+.dest 2 qdst
+.source 4 src
+splitlw qdst, idst, src
diff --git a/volk/orc/volk_16ic_deinterleave_real_8i_a_orc_impl.orc b/volk/orc/volk_16ic_deinterleave_real_8i_a_orc_impl.orc
new file mode 100644
index 000000000..8db49fd7c
--- /dev/null
+++ b/volk/orc/volk_16ic_deinterleave_real_8i_a_orc_impl.orc
@@ -0,0 +1,6 @@
+.function volk_16ic_deinterleave_real_8i_a_orc_impl
+.dest 1 dst
+.source 4 src
+.temp 2 iw
+select0lw iw, src
+convhwb dst, iw
diff --git a/volk/orc/volk_16ic_magnitude_16i_a_orc_impl.orc b/volk/orc/volk_16ic_magnitude_16i_a_orc_impl.orc
new file mode 100644
index 000000000..fbaebc46d
--- /dev/null
+++ b/volk/orc/volk_16ic_magnitude_16i_a_orc_impl.orc
@@ -0,0 +1,23 @@
+.function volk_16ic_magnitude_16i_a_orc_impl
+.source 4 src
+.dest 2 dst
+.floatparam 4 scalar
+.temp 8 iql
+.temp 8 iqf
+.temp 8 prodiqf
+.temp 4 qf
+.temp 4 if
+.temp 4 sumf
+.temp 4 rootf
+.temp 4 rootl
+
+x2 convswl iql, src
+x2 convlf iqf, iql
+x2 divf iqf, iqf, scalar
+x2 mulf prodiqf, iqf, iqf
+splitql qf, if, prodiqf
+addf sumf, if, qf
+sqrtf rootf, sumf
+mulf rootf, rootf, scalar
+convfl rootl, rootf
+convlw dst, rootl
diff --git a/volk/orc/volk_16sc_magnitude_32f_aligned16_orc_impl.orc b/volk/orc/volk_16sc_magnitude_32f_aligned16_orc_impl.orc
new file mode 100644
index 000000000..66fef7d2e
--- /dev/null
+++ b/volk/orc/volk_16sc_magnitude_32f_aligned16_orc_impl.orc
@@ -0,0 +1,25 @@
+.function volk_16ic_magnitude_32f_a_orc_impl
+.source 4 src
+.dest 4 dst
+.floatparam 4 scalar
+.temp 4 reall
+.temp 4 imagl
+.temp 2 reals
+.temp 2 imags
+.temp 4 realf
+.temp 4 imagf
+.temp 4 sumf
+
+
+
+splitlw reals, imags, src
+convswl reall, reals
+convswl imagl, imags
+convlf realf, reall
+convlf imagf, imagl
+divf realf, realf, scalar
+divf imagf, imagf, scalar
+mulf realf, realf, realf
+mulf imagf, imagf, imagf
+addf sumf, realf, imagf
+sqrtf dst, sumf
diff --git a/volk/orc/volk_16u_byteswap_a_orc_impl.orc b/volk/orc/volk_16u_byteswap_a_orc_impl.orc
new file mode 100644
index 000000000..b96ba84af
--- /dev/null
+++ b/volk/orc/volk_16u_byteswap_a_orc_impl.orc
@@ -0,0 +1,3 @@
+.function volk_16u_byteswap_a_orc_impl
+.dest 2 dst
+swapw dst, dst
diff --git a/volk/orc/volk_32f_s32f_multiply_32f_a_orc_impl.orc b/volk/orc/volk_32f_s32f_multiply_32f_a_orc_impl.orc
new file mode 100644
index 000000000..ea23fc045
--- /dev/null
+++ b/volk/orc/volk_32f_s32f_multiply_32f_a_orc_impl.orc
@@ -0,0 +1,5 @@
+.function volk_32f_s32f_multiply_32f_a_orc_impl
+.dest 4 dst
+.source 4 src1
+.floatparam 4 scalar
+mulf dst, src1, scalar
diff --git a/volk/orc/volk_32f_s32f_normalize_a_orc_impl.orc b/volk/orc/volk_32f_s32f_normalize_a_orc_impl.orc
new file mode 100644
index 000000000..986fdf665
--- /dev/null
+++ b/volk/orc/volk_32f_s32f_normalize_a_orc_impl.orc
@@ -0,0 +1,5 @@
+.function volk_32f_s32f_normalize_a_orc_impl
+.source 4 src1
+.floatparam 4 invscalar
+.dest 4 dst
+mulf dst, src1, invscalar
diff --git a/volk/orc/volk_32f_sqrt_32f_a_orc_impl.orc b/volk/orc/volk_32f_sqrt_32f_a_orc_impl.orc
new file mode 100644
index 000000000..f339b1122
--- /dev/null
+++ b/volk/orc/volk_32f_sqrt_32f_a_orc_impl.orc
@@ -0,0 +1,4 @@
+.function volk_32f_sqrt_32f_a_orc_impl
+.source 4 src
+.dest 4 dst
+sqrtf dst, src
diff --git a/volk/orc/volk_32f_x2_add_32f_a_orc_impl.orc b/volk/orc/volk_32f_x2_add_32f_a_orc_impl.orc
new file mode 100644
index 000000000..450cc6a9e
--- /dev/null
+++ b/volk/orc/volk_32f_x2_add_32f_a_orc_impl.orc
@@ -0,0 +1,5 @@
+.function volk_32f_x2_add_32f_a_orc_impl
+.dest 4 dst
+.source 4 src1
+.source 4 src2
+addf dst, src1, src2
diff --git a/volk/orc/volk_32f_x2_divide_32f_a_orc_impl.orc b/volk/orc/volk_32f_x2_divide_32f_a_orc_impl.orc
new file mode 100644
index 000000000..ee3b61b82
--- /dev/null
+++ b/volk/orc/volk_32f_x2_divide_32f_a_orc_impl.orc
@@ -0,0 +1,5 @@
+.function volk_32f_x2_divide_32f_a_orc_impl
+.dest 4 dst
+.source 4 src1
+.source 4 src2
+divf dst, src1, src2
diff --git a/volk/orc/volk_32f_x2_dot_prod_32f_a_orc_impl.orc b/volk/orc/volk_32f_x2_dot_prod_32f_a_orc_impl.orc
new file mode 100644
index 000000000..b367f3091
--- /dev/null
+++ b/volk/orc/volk_32f_x2_dot_prod_32f_a_orc_impl.orc
@@ -0,0 +1,6 @@
+.function volk_32f_x2_dot_prod_32f_a_orc_impl
+.source 4 src1
+.source 4 src2
+.dest 4 dst
+.accumulator 4 accum
+addf dst, src1, src2
diff --git a/volk/orc/volk_32f_x2_max_32f_a_orc_impl.orc b/volk/orc/volk_32f_x2_max_32f_a_orc_impl.orc
new file mode 100644
index 000000000..725201633
--- /dev/null
+++ b/volk/orc/volk_32f_x2_max_32f_a_orc_impl.orc
@@ -0,0 +1,5 @@
+.function volk_32f_x2_max_32f_a_orc_impl
+.dest 4 dst
+.source 4 src1
+.source 4 src2
+maxf dst, src1, src2
diff --git a/volk/orc/volk_32f_x2_min_32f_a_orc_impl.orc b/volk/orc/volk_32f_x2_min_32f_a_orc_impl.orc
new file mode 100644
index 000000000..a71ed8250
--- /dev/null
+++ b/volk/orc/volk_32f_x2_min_32f_a_orc_impl.orc
@@ -0,0 +1,5 @@
+.function volk_32f_x2_min_32f_a_orc_impl
+.dest 4 dst
+.source 4 src1
+.source 4 src2
+minf dst, src1, src2
diff --git a/volk/orc/volk_32f_x2_multiply_32f_a_orc_impl.orc b/volk/orc/volk_32f_x2_multiply_32f_a_orc_impl.orc
new file mode 100644
index 000000000..c17d539fd
--- /dev/null
+++ b/volk/orc/volk_32f_x2_multiply_32f_a_orc_impl.orc
@@ -0,0 +1,5 @@
+.function volk_32f_x2_multiply_32f_a_orc_impl
+.dest 4 dst
+.source 4 src1
+.source 4 src2
+mulf dst, src1, src2
diff --git a/volk/orc/volk_32f_x2_subtract_32f_a_orc_impl.orc b/volk/orc/volk_32f_x2_subtract_32f_a_orc_impl.orc
new file mode 100644
index 000000000..b3b0f256e
--- /dev/null
+++ b/volk/orc/volk_32f_x2_subtract_32f_a_orc_impl.orc
@@ -0,0 +1,5 @@
+.function volk_32f_x2_subtract_32f_a_orc_impl
+.dest 4 dst
+.source 4 src1
+.source 4 src2
+subf dst, src1, src2
diff --git a/volk/orc/volk_32fc_32f_multiply_32fc_a_orc_impl.orc b/volk/orc/volk_32fc_32f_multiply_32fc_a_orc_impl.orc
new file mode 100644
index 000000000..aa82699f5
--- /dev/null
+++ b/volk/orc/volk_32fc_32f_multiply_32fc_a_orc_impl.orc
@@ -0,0 +1,7 @@
+.function volk_32fc_32f_multiply_32fc_a_orc_impl
+.source 8 src1
+.source 4 src2
+.dest 8 dst
+.temp 8 tmp
+mergelq tmp, src2, src2
+x2 mulf dst, src1, tmp
diff --git a/volk/orc/volk_32fc_magnitude_32f_a_orc_impl.orc b/volk/orc/volk_32fc_magnitude_32f_a_orc_impl.orc
new file mode 100644
index 000000000..032ab2b1b
--- /dev/null
+++ b/volk/orc/volk_32fc_magnitude_32f_a_orc_impl.orc
@@ -0,0 +1,13 @@
+.function volk_32fc_magnitude_32f_a_orc_impl
+.source 8 src
+.dest 4 dst
+.temp 8 iqf
+.temp 8 prodiqf
+.temp 4 qf
+.temp 4 if
+.temp 4 sumf
+
+x2 mulf prodiqf, src, src
+splitql qf, if, prodiqf
+addf sumf, if, qf
+sqrtf dst, sumf
diff --git a/volk/orc/volk_32fc_s32f_magnitude_16i_a_orc_impl.orc b/volk/orc/volk_32fc_s32f_magnitude_16i_a_orc_impl.orc
new file mode 100644
index 000000000..d3bf78935
--- /dev/null
+++ b/volk/orc/volk_32fc_s32f_magnitude_16i_a_orc_impl.orc
@@ -0,0 +1,23 @@
+.function volk_32fc_s32f_magnitude_16i_a_orc_impl
+.source 8 src
+.dest 2 dst
+.floatparam 4 scalar
+.temp 8 iqf
+.temp 8 prodiqf
+.temp 4 qf
+.temp 4 if
+.temp 4 sumf
+.temp 4 rootf
+.temp 4 rootl
+#.temp 4 maskl
+
+x2 mulf prodiqf, src, src
+splitql qf, if, prodiqf
+addf sumf, if, qf
+sqrtf rootf, sumf
+mulf rootf, rootf, scalar
+#cmpltf maskl, 32768.0, rootf
+#andl maskl, maskl, 0x80000000
+#orl rootf, rootf, maskl
+convfl rootl, rootf
+convsuslw dst, rootl
diff --git a/volk/orc/volk_32fc_s32fc_multiply_32fc_a_orc_impl.orc b/volk/orc/volk_32fc_s32fc_multiply_32fc_a_orc_impl.orc
new file mode 100644
index 000000000..2577e034f
--- /dev/null
+++ b/volk/orc/volk_32fc_s32fc_multiply_32fc_a_orc_impl.orc
@@ -0,0 +1,18 @@
+.function volk_32fc_s32fc_multiply_32fc_a_orc_impl
+.source 8 src1
+.floatparam 8 scalar
+.dest 8 dst
+.temp 8 iqprod
+.temp 4 real
+.temp 4 imag
+.temp 4 ac
+.temp 4 bd
+.temp 8 swapped
+x2 mulf iqprod, src1, scalar
+splitql bd, ac, iqprod
+subf real, ac, bd
+swaplq swapped, src1
+x2 mulf iqprod, swapped, scalar
+splitql bd, ac, iqprod
+addf imag, ac, bd
+mergelq dst, real, imag
diff --git a/volk/orc/volk_32fc_x2_multiply_32fc_a_orc_impl.orc b/volk/orc/volk_32fc_x2_multiply_32fc_a_orc_impl.orc
new file mode 100644
index 000000000..cb8a12d81
--- /dev/null
+++ b/volk/orc/volk_32fc_x2_multiply_32fc_a_orc_impl.orc
@@ -0,0 +1,18 @@
+.function volk_32fc_x2_multiply_32fc_a_orc_impl
+.source 8 src1
+.source 8 src2
+.dest 8 dst
+.temp 8 iqprod
+.temp 4 real
+.temp 4 imag
+.temp 4 ac
+.temp 4 bd
+.temp 8 swapped
+x2 mulf iqprod, src1, src2
+splitql bd, ac, iqprod
+subf real, ac, bd
+swaplq swapped, src1
+x2 mulf iqprod, swapped, src2
+splitql bd, ac, iqprod
+addf imag, ac, bd
+mergelq dst, real, imag
diff --git a/volk/orc/volk_32i_x2_and_32i_a_orc_impl.orc b/volk/orc/volk_32i_x2_and_32i_a_orc_impl.orc
new file mode 100644
index 000000000..1845e4654
--- /dev/null
+++ b/volk/orc/volk_32i_x2_and_32i_a_orc_impl.orc
@@ -0,0 +1,5 @@
+.function volk_32i_x2_and_32i_a_orc_impl
+.dest 4 dst
+.source 4 src1
+.source 4 src2
+andl dst, src1, src2
diff --git a/volk/orc/volk_32i_x2_or_32i_a_orc_impl.orc b/volk/orc/volk_32i_x2_or_32i_a_orc_impl.orc
new file mode 100644
index 000000000..004663f42
--- /dev/null
+++ b/volk/orc/volk_32i_x2_or_32i_a_orc_impl.orc
@@ -0,0 +1,5 @@
+.function volk_32i_x2_or_32i_a_orc_impl
+.dest 4 dst
+.source 4 src1
+.source 4 src2
+orl dst, src1, src2
diff --git a/volk/orc/volk_8i_convert_16i_a_orc_impl.orc b/volk/orc/volk_8i_convert_16i_a_orc_impl.orc
new file mode 100644
index 000000000..17198bf1e
--- /dev/null
+++ b/volk/orc/volk_8i_convert_16i_a_orc_impl.orc
@@ -0,0 +1,6 @@
+.function volk_8i_convert_16i_a_orc_impl
+.source 1 src
+.dest 2 dst
+.temp 2 tmp
+convsbw tmp, src
+shlw dst, tmp, 8
diff --git a/volk/orc/volk_8i_s32f_convert_32f_a_orc_impl.orc b/volk/orc/volk_8i_s32f_convert_32f_a_orc_impl.orc
new file mode 100644
index 000000000..ad54fb1e1
--- /dev/null
+++ b/volk/orc/volk_8i_s32f_convert_32f_a_orc_impl.orc
@@ -0,0 +1,11 @@
+.function volk_8i_s32f_convert_32f_a_orc_impl
+.source 1 src
+.dest 4 dst
+.floatparam 4 scalar
+.temp 4 flsrc
+.temp 4 lsrc
+.temp 2 ssrc
+convsbw ssrc, src
+convswl lsrc, ssrc
+convlf flsrc, lsrc
+mulf dst, flsrc, scalar