convert: Add CHDR converters

The difference between the _chdr converters and the _item32_ converters
is that the former do not require item32 boundaries, they do not require
endianness swapping, and they don't use IQ swapping either.

This is possible because the FPGA will do byte-swapping.

1 files changed, 72 insertions, 1 deletions

diff --git a/host/lib/convert/gen_convert_general.py b/host/lib/convert/gen_convert_general.py
index c4a8c9725..1b1106d4c 100644
--- a/host/lib/convert/gen_convert_general.py
+++ b/host/lib/convert/gen_convert_general.py
@@ -5,6 +5,9 @@
 #
 # SPDX-License-Identifier: GPL-3.0-or-later
 #
+"""
+Auto-Generator for generic converters
+"""
 
 TMPL_HEADER = """
 <%
@@ -16,6 +19,7 @@ TMPL_HEADER = """
 
 #include "convert_common.hpp"
 #include <uhd/utils/byteswap.hpp>
+#include <algorithm>
 
 using namespace uhd::convert;
 
@@ -29,6 +33,48 @@ DECLARE_CONVERTER(item32, 1, item32, 1, PRIORITY_GENERAL) {
 }
 """
 
+TMPL_CONV_CHDR_SC16_TO_FP = """
+DECLARE_CONVERTER({fctype}, 1, sc16_chdr, 1, PRIORITY_GENERAL) {{
+    // Note: We convert I and Q separately, because there's no optimized
+    // constructor to create a complex<{fptype}> from a complex<int16_t>. This
+    // means we need to multiply nsamps by 2
+    const {fptype}* input = reinterpret_cast<const {fptype}*>(inputs[0]);
+    int16_t* output = reinterpret_cast<int16_t*>(outputs[0]);
+
+    for (size_t i = 0; i < nsamps * 2; i += 2) {{
+        output[i]   = static_cast<int16_t>(input[i]   * {fptype}(scale_factor));
+        output[i+1] = static_cast<int16_t>(input[i+1] * {fptype}(scale_factor));
+    }}
+}}
+
+DECLARE_CONVERTER(sc16_chdr, 1, {fctype}, 1, PRIORITY_GENERAL) {{
+    // Note: We convert I and Q separately, because there's no optimized
+    // constructor to create a complex<{fptype}> from a complex<int16_t>. This
+    // means we need to multiply nsamps by 2
+    const int16_t* input = reinterpret_cast<const int16_t*>(inputs[0]);
+    {fptype}* output = reinterpret_cast<{fptype}*>(outputs[0]);
+
+    for (size_t i = 0; i < nsamps * 2; i += 2) {{
+        output[i]   = static_cast<{fptype}>(input[i])   * {fptype}(scale_factor);
+        output[i+1] = static_cast<{fptype}>(input[i+1]) * {fptype}(scale_factor);
+    }}
+}}
+"""
+
+# For CHDR converters, all converters where the input and output type are the
+# same can be done by a memcpy, because we can do the endianness adaptation in
+# the FPGA.
+TMPL_CONV_CHDR_MEMCPY = """
+DECLARE_CONVERTER({in_type}, 1, {out_type}, 1, PRIORITY_GENERAL) {{
+    const {ptr_type} *input = reinterpret_cast<const {ptr_type} *>(inputs[0]);
+    {ptr_type}* output = reinterpret_cast<{ptr_type}*>(outputs[0]);
+
+    // Benchmark shows that copy_n can be significantly slower in some cases
+    //std::copy_n(input, nsamps, output);
+    memcpy(output, input, sizeof({ptr_type}) * nsamps);
+}}
+"""
+
 # Some 32-bit types converters are also defined in convert_item32.cpp to
 # take care of quirks such as I/Q ordering on the wire etc.
 TMPL_CONV_ITEM32 = """
@@ -177,10 +223,35 @@ def parse_tmpl(_tmpl_text, **kwargs):
     return Template(_tmpl_text).render(**kwargs)
 
 if __name__ == '__main__':
-    import sys, os
+    import sys
+    import os
     file = os.path.basename(__file__)
     output = parse_tmpl(TMPL_HEADER, file=file)
 
+    for fctype, fptype in (
+            ('fc32', 'float'),
+            ('fc64', 'double'),
+        ):
+        output += TMPL_CONV_CHDR_SC16_TO_FP.format(
+            fctype=fctype, fptype=fptype)
+
+    ## Generate CHDR converters
+    # These guys don't have to worry about endianness
+    for uhd_type, ptr_type in (
+            ('u8', 'uint8_t'),
+            ('s8', 'int8_t'),
+            ('s16', 'int16_t'),
+            ('f32', 'float'),
+            ('sc8', 'std::complex<int8_t>'),
+            ('sc16', 'std::complex<int16_t>'),
+            ('fc32', 'std::complex<float>'),
+            ('fc64', 'std::complex<double>'),
+    ):
+        for in_type, out_type in ((uhd_type + '_chdr', uhd_type), (uhd_type, uhd_type + '_chdr')):
+            output += TMPL_CONV_CHDR_MEMCPY.format(
+                in_type=in_type,
+                out_type=out_type,
+                ptr_type=ptr_type)
     ## Generate all data types that are exactly
     ## item32 or multiples thereof:
     for end in ('be', 'le'):