uhd: various performance tweaks

2 files changed, 41 insertions, 24 deletions

diff --git a/host/lib/transport/gen_vrt_if_packet.py b/host/lib/transport/gen_vrt_if_packet.py
index dbe026ba3..bf740ffa9 100755
--- a/host/lib/transport/gen_vrt_if_packet.py
+++ b/host/lib/transport/gen_vrt_if_packet.py
@@ -1,6 +1,6 @@
 #!/usr/bin/env python
 #
-# Copyright 2010 Ettus Research LLC
+# Copyright 2010-2011 Ettus Research LLC
 #
 # This program is free software: you can redistribute it and/or modify
 # it under the terms of the GNU General Public License as published by
@@ -35,6 +35,7 @@ TMPL_TEXT = """
 \#include <uhd/utils/byteswap.hpp>
 \#include <boost/detail/endian.hpp>
 \#include <stdexcept>
+\#include <vector>
 
 //define the endian macros to convert integers
 \#ifdef BOOST_BIG_ENDIAN
@@ -48,18 +49,26 @@ TMPL_TEXT = """
 using namespace uhd;
 using namespace uhd::transport;
 
-########################################################################
-#def gen_code($XE_MACRO, $suffix)
-########################################################################
+typedef size_t pred_type;
+typedef std::vector<pred_type> pred_table_type;
+#define pred_table_index(hdr) ((hdr >> 20) & 0x1ff)
+
+static pred_table_type get_pred_unpack_table(void){
+    pred_table_type table(1 << 9, 0); //only 9 bits useful here (20-28)
+    for (size_t i = 0; i < table.size(); i++){
+        boost::uint32_t vrt_hdr_word = i << 20;
+        if(vrt_hdr_word & $hex(0x1 << 28)) table[i] |= $hex($sid_p);
+        if(vrt_hdr_word & $hex(0x1 << 27)) table[i] |= $hex($cid_p);
+        if(vrt_hdr_word & $hex(0x3 << 22)) table[i] |= $hex($tsi_p);
+        if(vrt_hdr_word & $hex(0x3 << 20)) table[i] |= $hex($tsf_p);
+        if(vrt_hdr_word & $hex(0x1 << 26)) table[i] |= $hex($tlr_p);
+    }
+    return table;
+}
 
 ########################################################################
-## setup predicates
+#def gen_code($XE_MACRO, $suffix)
 ########################################################################
-#set $sid_p = 0b00001
-#set $cid_p = 0b00010
-#set $tsi_p = 0b00100
-#set $tsf_p = 0b01000
-#set $tlr_p = 0b10000
 
 void vrt::if_hdr_pack_$(suffix)(
     boost::uint32_t *packet_buff,
@@ -67,7 +76,7 @@ void vrt::if_hdr_pack_$(suffix)(
 ){
     boost::uint32_t vrt_hdr_flags = 0;
 
-    boost::uint8_t pred = 0;
+    pred_type pred = 0;
     if (if_packet_info.has_sid) pred |= $hex($sid_p);
     if (if_packet_info.has_cid) pred |= $hex($cid_p);
     if (if_packet_info.has_tsi) pred |= $hex($tsi_p);
@@ -159,12 +168,8 @@ void vrt::if_hdr_unpack_$(suffix)(
     //if_packet_info.sob = bool(vrt_hdr_word & $hex(0x1 << 25)); //not implemented
     //if_packet_info.eob = bool(vrt_hdr_word & $hex(0x1 << 24)); //not implemented
 
-    boost::uint8_t pred = 0;
-    if(vrt_hdr_word & $hex(0x1 << 28)) pred |= $hex($sid_p);
-    if(vrt_hdr_word & $hex(0x1 << 27)) pred |= $hex($cid_p);
-    if(vrt_hdr_word & $hex(0x3 << 22)) pred |= $hex($tsi_p);
-    if(vrt_hdr_word & $hex(0x3 << 20)) pred |= $hex($tsf_p);
-    if(vrt_hdr_word & $hex(0x1 << 26)) pred |= $hex($tlr_p);
+    static const pred_table_type pred_unpack_table(get_pred_unpack_table());
+    const pred_type pred = pred_unpack_table[pred_table_index(vrt_hdr_word)];
 
     switch(pred){
     #for $pred in range(2**5)
@@ -239,4 +244,12 @@ def parse_tmpl(_tmpl_text, **kwargs):
 
 if __name__ == '__main__':
     import sys
-    open(sys.argv[1], 'w').write(parse_tmpl(TMPL_TEXT, file=__file__))
+    open(sys.argv[1], 'w').write(parse_tmpl(
+        TMPL_TEXT,
+        file=__file__,
+        sid_p = 0b00001,
+        cid_p = 0b00010,
+        tsi_p = 0b00100,
+        tsf_p = 0b01000,
+        tlr_p = 0b10000,
+    ))
diff --git a/host/lib/transport/vrt_packet_handler.hpp b/host/lib/transport/vrt_packet_handler.hpp
index c535edd04..2e005b87a 100644
--- a/host/lib/transport/vrt_packet_handler.hpp
+++ b/host/lib/transport/vrt_packet_handler.hpp
@@ -301,7 +301,12 @@ template <typename T> UHD_INLINE T get_context_code(
         //init the expected seq number
         size_t next_packet_seq;
 
-        send_state(void) : next_packet_seq(0){
+        managed_send_buffs_t managed_buffs;
+
+        send_state(size_t width = 1):
+            next_packet_seq(0),
+            managed_buffs(width)
+        {
             /* NOP */
         }
     };
@@ -326,9 +331,8 @@ template <typename T> UHD_INLINE T get_context_code(
         if_packet_info.num_payload_words32 = (num_samps*chans_per_otw_buff*OTW_BYTES_PER_SAMP)/sizeof(boost::uint32_t);
         if_packet_info.packet_count = state.next_packet_seq;
 
-        //get send buffers for each channel
-        managed_send_buffs_t send_buffs(buffs.size()/chans_per_otw_buff);
-        if (not get_send_buffs(send_buffs)) return 0;
+        //get send buffers for each otw channel
+        if (not get_send_buffs(state.managed_buffs)) return 0;
 
         std::vector<const void *> io_buffs(chans_per_otw_buff);
         for (size_t i = 0; i < buffs.size(); i+=chans_per_otw_buff){
@@ -336,7 +340,7 @@ template <typename T> UHD_INLINE T get_context_code(
             for (size_t j = 0; j < chans_per_otw_buff; j++){
                 io_buffs[j] = reinterpret_cast<const boost::uint8_t *>(buffs[i+j]) + offset_bytes;
             }
-            boost::uint32_t *otw_mem = send_buffs[i]->cast<boost::uint32_t *>() + vrt_header_offset_words32;
+            boost::uint32_t *otw_mem = state.managed_buffs[i]->cast<boost::uint32_t *>() + vrt_header_offset_words32;
 
             //pack metadata into a vrt header
             vrt_packer(otw_mem, if_packet_info);
@@ -348,7 +352,7 @@ template <typename T> UHD_INLINE T get_context_code(
 
             //commit the samples to the zero-copy interface
             size_t num_bytes_total = (vrt_header_offset_words32+if_packet_info.num_packet_words32)*sizeof(boost::uint32_t);
-            send_buffs[i]->commit(num_bytes_total);
+            state.managed_buffs[i]->commit(num_bytes_total);
         }
         state.next_packet_seq++; //increment sequence after commits
         return num_samps;