summaryrefslogtreecommitdiffstats
path: root/host/lib
diff options
context:
space:
mode:
Diffstat (limited to 'host/lib')
-rw-r--r--host/lib/convert/convert_with_sse2.cpp204
-rw-r--r--host/lib/transport/libusb1_zero_copy.cpp5
-rw-r--r--host/lib/usrp/usrp2/io_impl.cpp5
3 files changed, 128 insertions, 86 deletions
diff --git a/host/lib/convert/convert_with_sse2.cpp b/host/lib/convert/convert_with_sse2.cpp
index 52beea24a..9772028dc 100644
--- a/host/lib/convert/convert_with_sse2.cpp
+++ b/host/lib/convert/convert_with_sse2.cpp
@@ -25,25 +25,37 @@ DECLARE_CONVERTER(convert_fc32_1_to_item32_1_nswap, PRIORITY_CUSTOM){
const fc32_t *input = reinterpret_cast<const fc32_t *>(inputs[0]);
item32_t *output = reinterpret_cast<item32_t *>(outputs[0]);
- __m128 scalar = _mm_set_ps1(float(scale_factor));
-
- //convert blocks of samples with intrinsics
- size_t i = 0; for (; i < (nsamps & ~0x3); i+=4){
- //load from input
- __m128 tmplo = _mm_loadu_ps(reinterpret_cast<const float *>(input+i+0));
- __m128 tmphi = _mm_loadu_ps(reinterpret_cast<const float *>(input+i+2));
-
- //convert and scale
- __m128i tmpilo = _mm_cvtps_epi32(_mm_mul_ps(tmplo, scalar));
- __m128i tmpihi = _mm_cvtps_epi32(_mm_mul_ps(tmphi, scalar));
-
- //pack + swap 16-bit pairs
- __m128i tmpi = _mm_packs_epi32(tmpilo, tmpihi);
- tmpi = _mm_shufflelo_epi16(tmpi, _MM_SHUFFLE(2, 3, 0, 1));
- tmpi = _mm_shufflehi_epi16(tmpi, _MM_SHUFFLE(2, 3, 0, 1));
-
- //store to output
- _mm_storeu_si128(reinterpret_cast<__m128i *>(output+i), tmpi);
+ const __m128 scalar = _mm_set_ps1(float(scale_factor));
+
+ #define convert_fc32_1_to_item32_1_nswap_guts(_al_) \
+ for (; i < (nsamps & ~0x3); i+=4){ \
+ /* load from input */ \
+ __m128 tmplo = _mm_load ## _al_ ## _ps(reinterpret_cast<const float *>(input+i+0)); \
+ __m128 tmphi = _mm_load ## _al_ ## _ps(reinterpret_cast<const float *>(input+i+2)); \
+ \
+ /* convert and scale */ \
+ __m128i tmpilo = _mm_cvtps_epi32(_mm_mul_ps(tmplo, scalar)); \
+ __m128i tmpihi = _mm_cvtps_epi32(_mm_mul_ps(tmphi, scalar)); \
+ \
+ /* pack + swap 16-bit pairs */ \
+ __m128i tmpi = _mm_packs_epi32(tmpilo, tmpihi); \
+ tmpi = _mm_shufflelo_epi16(tmpi, _MM_SHUFFLE(2, 3, 0, 1)); \
+ tmpi = _mm_shufflehi_epi16(tmpi, _MM_SHUFFLE(2, 3, 0, 1)); \
+ \
+ /* store to output */ \
+ _mm_storeu_si128(reinterpret_cast<__m128i *>(output+i), tmpi); \
+ } \
+
+ size_t i = 0;
+
+ //dispatch according to alignment
+ switch (size_t(input) & 0xf){
+ case 0x8:
+ output[i] = fc32_to_item32(input[i], float(scale_factor)); i++;
+ case 0x0:
+ convert_fc32_1_to_item32_1_nswap_guts()
+ break;
+ default: convert_fc32_1_to_item32_1_nswap_guts(u)
}
//convert remainder
@@ -56,24 +68,36 @@ DECLARE_CONVERTER(convert_fc32_1_to_item32_1_bswap, PRIORITY_CUSTOM){
const fc32_t *input = reinterpret_cast<const fc32_t *>(inputs[0]);
item32_t *output = reinterpret_cast<item32_t *>(outputs[0]);
- __m128 scalar = _mm_set_ps1(float(scale_factor));
-
- //convert blocks of samples with intrinsics
- size_t i = 0; for (; i < (nsamps & ~0x3); i+=4){
- //load from input
- __m128 tmplo = _mm_loadu_ps(reinterpret_cast<const float *>(input+i+0));
- __m128 tmphi = _mm_loadu_ps(reinterpret_cast<const float *>(input+i+2));
-
- //convert and scale
- __m128i tmpilo = _mm_cvtps_epi32(_mm_mul_ps(tmplo, scalar));
- __m128i tmpihi = _mm_cvtps_epi32(_mm_mul_ps(tmphi, scalar));
-
- //pack + byteswap -> byteswap 16 bit words
- __m128i tmpi = _mm_packs_epi32(tmpilo, tmpihi);
- tmpi = _mm_or_si128(_mm_srli_epi16(tmpi, 8), _mm_slli_epi16(tmpi, 8));
-
- //store to output
- _mm_storeu_si128(reinterpret_cast<__m128i *>(output+i), tmpi);
+ const __m128 scalar = _mm_set_ps1(float(scale_factor));
+
+ #define convert_fc32_1_to_item32_1_bswap_guts(_al_) \
+ for (; i < (nsamps & ~0x3); i+=4){ \
+ /* load from input */ \
+ __m128 tmplo = _mm_load ## _al_ ## _ps(reinterpret_cast<const float *>(input+i+0)); \
+ __m128 tmphi = _mm_load ## _al_ ## _ps(reinterpret_cast<const float *>(input+i+2)); \
+ \
+ /* convert and scale */ \
+ __m128i tmpilo = _mm_cvtps_epi32(_mm_mul_ps(tmplo, scalar)); \
+ __m128i tmpihi = _mm_cvtps_epi32(_mm_mul_ps(tmphi, scalar)); \
+ \
+ /* pack + byteswap -> byteswap 16 bit words */ \
+ __m128i tmpi = _mm_packs_epi32(tmpilo, tmpihi); \
+ tmpi = _mm_or_si128(_mm_srli_epi16(tmpi, 8), _mm_slli_epi16(tmpi, 8)); \
+ \
+ /* store to output */ \
+ _mm_storeu_si128(reinterpret_cast<__m128i *>(output+i), tmpi); \
+ } \
+
+ size_t i = 0;
+
+ //dispatch according to alignment
+ switch (size_t(input) & 0xf){
+ case 0x8:
+ output[i] = uhd::byteswap(fc32_to_item32(input[i], float(scale_factor))); i++;
+ case 0x0:
+ convert_fc32_1_to_item32_1_bswap_guts()
+ break;
+ default: convert_fc32_1_to_item32_1_bswap_guts(u)
}
//convert remainder
@@ -86,27 +110,39 @@ DECLARE_CONVERTER(convert_item32_1_to_fc32_1_nswap, PRIORITY_CUSTOM){
const item32_t *input = reinterpret_cast<const item32_t *>(inputs[0]);
fc32_t *output = reinterpret_cast<fc32_t *>(outputs[0]);
- __m128 scalar = _mm_set_ps1(float(scale_factor)/(1 << 16));
- __m128i zeroi = _mm_setzero_si128();
-
- //convert blocks of samples with intrinsics
- size_t i = 0; for (; i < (nsamps & ~0x3); i+=4){
- //load from input
- __m128i tmpi = _mm_loadu_si128(reinterpret_cast<const __m128i *>(input+i));
-
- //unpack + swap 16-bit pairs
- tmpi = _mm_shufflelo_epi16(tmpi, _MM_SHUFFLE(2, 3, 0, 1));
- tmpi = _mm_shufflehi_epi16(tmpi, _MM_SHUFFLE(2, 3, 0, 1));
- __m128i tmpilo = _mm_unpacklo_epi16(zeroi, tmpi); //value in upper 16 bits
- __m128i tmpihi = _mm_unpackhi_epi16(zeroi, tmpi);
-
- //convert and scale
- __m128 tmplo = _mm_mul_ps(_mm_cvtepi32_ps(tmpilo), scalar);
- __m128 tmphi = _mm_mul_ps(_mm_cvtepi32_ps(tmpihi), scalar);
-
- //store to output
- _mm_storeu_ps(reinterpret_cast<float *>(output+i+0), tmplo);
- _mm_storeu_ps(reinterpret_cast<float *>(output+i+2), tmphi);
+ const __m128 scalar = _mm_set_ps1(float(scale_factor)/(1 << 16));
+ const __m128i zeroi = _mm_setzero_si128();
+
+ #define convert_item32_1_to_fc32_1_nswap_guts(_al_) \
+ for (; i < (nsamps & ~0x3); i+=4){ \
+ /* load from input */ \
+ __m128i tmpi = _mm_loadu_si128(reinterpret_cast<const __m128i *>(input+i)); \
+ \
+ /* unpack + swap 16-bit pairs */ \
+ tmpi = _mm_shufflelo_epi16(tmpi, _MM_SHUFFLE(2, 3, 0, 1)); \
+ tmpi = _mm_shufflehi_epi16(tmpi, _MM_SHUFFLE(2, 3, 0, 1)); \
+ __m128i tmpilo = _mm_unpacklo_epi16(zeroi, tmpi); /* value in upper 16 bits */ \
+ __m128i tmpihi = _mm_unpackhi_epi16(zeroi, tmpi); \
+ \
+ /* convert and scale */ \
+ __m128 tmplo = _mm_mul_ps(_mm_cvtepi32_ps(tmpilo), scalar); \
+ __m128 tmphi = _mm_mul_ps(_mm_cvtepi32_ps(tmpihi), scalar); \
+ \
+ /* store to output */ \
+ _mm_store ## _al_ ## _ps(reinterpret_cast<float *>(output+i+0), tmplo); \
+ _mm_store ## _al_ ## _ps(reinterpret_cast<float *>(output+i+2), tmphi); \
+ } \
+
+ size_t i = 0;
+
+ //dispatch according to alignment
+ switch (size_t(output) & 0xf){
+ case 0x8:
+ output[i] = item32_to_fc32(input[i], float(scale_factor)); i++;
+ case 0x0:
+ convert_item32_1_to_fc32_1_nswap_guts()
+ break;
+ default: convert_item32_1_to_fc32_1_nswap_guts(u)
}
//convert remainder
@@ -119,26 +155,38 @@ DECLARE_CONVERTER(convert_item32_1_to_fc32_1_bswap, PRIORITY_CUSTOM){
const item32_t *input = reinterpret_cast<const item32_t *>(inputs[0]);
fc32_t *output = reinterpret_cast<fc32_t *>(outputs[0]);
- __m128 scalar = _mm_set_ps1(float(scale_factor)/(1 << 16));
- __m128i zeroi = _mm_setzero_si128();
-
- //convert blocks of samples with intrinsics
- size_t i = 0; for (; i < (nsamps & ~0x3); i+=4){
- //load from input
- __m128i tmpi = _mm_loadu_si128(reinterpret_cast<const __m128i *>(input+i));
-
- //byteswap + unpack -> byteswap 16 bit words
- tmpi = _mm_or_si128(_mm_srli_epi16(tmpi, 8), _mm_slli_epi16(tmpi, 8));
- __m128i tmpilo = _mm_unpacklo_epi16(zeroi, tmpi); //value in upper 16 bits
- __m128i tmpihi = _mm_unpackhi_epi16(zeroi, tmpi);
-
- //convert and scale
- __m128 tmplo = _mm_mul_ps(_mm_cvtepi32_ps(tmpilo), scalar);
- __m128 tmphi = _mm_mul_ps(_mm_cvtepi32_ps(tmpihi), scalar);
-
- //store to output
- _mm_storeu_ps(reinterpret_cast<float *>(output+i+0), tmplo);
- _mm_storeu_ps(reinterpret_cast<float *>(output+i+2), tmphi);
+ const __m128 scalar = _mm_set_ps1(float(scale_factor)/(1 << 16));
+ const __m128i zeroi = _mm_setzero_si128();
+
+ #define convert_item32_1_to_fc32_1_bswap_guts(_al_) \
+ for (; i < (nsamps & ~0x3); i+=4){ \
+ /* load from input */ \
+ __m128i tmpi = _mm_loadu_si128(reinterpret_cast<const __m128i *>(input+i)); \
+ \
+ /* byteswap + unpack -> byteswap 16 bit words */ \
+ tmpi = _mm_or_si128(_mm_srli_epi16(tmpi, 8), _mm_slli_epi16(tmpi, 8)); \
+ __m128i tmpilo = _mm_unpacklo_epi16(zeroi, tmpi); /* value in upper 16 bits */ \
+ __m128i tmpihi = _mm_unpackhi_epi16(zeroi, tmpi); \
+ \
+ /* convert and scale */ \
+ __m128 tmplo = _mm_mul_ps(_mm_cvtepi32_ps(tmpilo), scalar); \
+ __m128 tmphi = _mm_mul_ps(_mm_cvtepi32_ps(tmpihi), scalar); \
+ \
+ /* store to output */ \
+ _mm_store ## _al_ ## _ps(reinterpret_cast<float *>(output+i+0), tmplo); \
+ _mm_store ## _al_ ## _ps(reinterpret_cast<float *>(output+i+2), tmphi); \
+ } \
+
+ size_t i = 0;
+
+ //dispatch according to alignment
+ switch (size_t(output) & 0xf){
+ case 0x8:
+ output[i] = item32_to_fc32(uhd::byteswap(input[i]), float(scale_factor)); i++;
+ case 0x0:
+ convert_item32_1_to_fc32_1_bswap_guts()
+ break;
+ default: convert_item32_1_to_fc32_1_bswap_guts(u)
}
//convert remainder
diff --git a/host/lib/transport/libusb1_zero_copy.cpp b/host/lib/transport/libusb1_zero_copy.cpp
index 19a7a3742..f781f890d 100644
--- a/host/lib/transport/libusb1_zero_copy.cpp
+++ b/host/lib/transport/libusb1_zero_copy.cpp
@@ -222,7 +222,6 @@ public:
}
}
//shutdown the threads
- _threads_running = false;
_thread_group.interrupt_all();
_thread_group.join_all();
}
@@ -277,15 +276,13 @@ private:
//! event handler threads
boost::thread_group _thread_group;
- bool _threads_running;
void run_event_loop(boost::barrier &spawn_barrier){
- _threads_running = true;
spawn_barrier.wait();
set_thread_priority_safe();
libusb_context *context = libusb::session::get_global_session()->get_context();
try{
- while(_threads_running){
+ while (not boost::this_thread::interruption_requested()){
timeval tv;
tv.tv_sec = 0;
tv.tv_usec = 100000; //100ms
diff --git a/host/lib/usrp/usrp2/io_impl.cpp b/host/lib/usrp/usrp2/io_impl.cpp
index ffe9a88e7..df452942c 100644
--- a/host/lib/usrp/usrp2/io_impl.cpp
+++ b/host/lib/usrp/usrp2/io_impl.cpp
@@ -146,7 +146,6 @@ struct usrp2_impl::io_impl{
}
~io_impl(void){
- recv_pirate_crew_raiding = false;
recv_pirate_crew.interrupt_all();
recv_pirate_crew.join_all();
}
@@ -185,7 +184,6 @@ struct usrp2_impl::io_impl{
//methods and variables for the pirate crew
void recv_pirate_loop(boost::barrier &, usrp2_mboard_impl::sptr, zero_copy_if::sptr, size_t);
boost::thread_group recv_pirate_crew;
- bool recv_pirate_crew_raiding;
bounded_buffer<async_metadata_t> async_msg_fifo;
};
@@ -201,14 +199,13 @@ void usrp2_impl::io_impl::recv_pirate_loop(
zero_copy_if::sptr err_xport,
size_t index
){
- recv_pirate_crew_raiding = true;
spawn_barrier.wait();
set_thread_priority_safe();
//store a reference to the flow control monitor (offset by max dsps)
flow_control_monitor &fc_mon = *(this->fc_mons[index*usrp2_mboard_impl::MAX_NUM_DSPS]);
- while(recv_pirate_crew_raiding){
+ while (not boost::this_thread::interruption_requested()){
managed_recv_buffer::sptr buff = err_xport->get_recv_buff();
if (not buff.get()) continue; //ignore timeout/error buffers
generated by cgit v1.2.3 (git 2.39.1) at 2024-11-17 06:41:51 +0000