hal: fixes external harq size request function, initializes private vector sizes

Author: ofontbach

include/srsran/hal/dpdk/bbdev/bbdev_acc.h

@@ -78,8 +78,8 @@ class bbdev_acc
   unsigned get_nof_fft_cores() const { return nof_fft_lcores; }
 
   /// Returns the size of the (external) HARQ buffer size embedded in the hardware-accelerator.
-  /// \return HARQ buffer size (in bytes).
-  units::bytes get_harq_buff_size() const { return units::bytes(1024 * info.drv.harq_buffer_size); }
+  /// \return HARQ buffer size in bytes. Note that 64 bits are used to enable sizes >= 4GB.
+  uint64_t get_harq_buff_size_bytes() const { return static_cast<uint64_t>(info.drv.harq_buffer_size) * 1024; }
 
   /// Returns the size of each mbuf used to exchange unencoded and unrate-matched messages with the accelerator.
   /// \return Unencoded and unrate-matched mbuf size (in bytes).

include/srsran/hal/phy/upper/channel_processors/pusch/ext_harq_buffer_context_repository.h

@@ -14,14 +14,15 @@
 #pragma once
 
 #include "srsran/support/srsran_assert.h"
+#include "srsran/support/units.h"
 #include <vector>
 
 namespace srsran {
 namespace hal {
 
 /// Fixed CB-offset increment used in the accelerator's HARQ memory.
 /// Note that it is assumed that the HARQ memory is organized in N slots of HARQ_INCR bytes.
-static constexpr unsigned HARQ_INCR = 32768;
+static constexpr units::bytes HARQ_INCR{32768};
 
 /// External HARQ buffer context.
 struct ext_harq_buffer_context_entry {
@@ -39,10 +40,10 @@ class ext_harq_buffer_context_repository
   /// \param[in] nof_codeblocks_    Indicates the number of codeblocks to store in the repository.
   /// \param[in] ext_harq_buff_size Size of the external HARQ buffer (in bytes).
   /// \param[in] debug_mode_        Requests to implement the debug mode (meant for unittesting).
-  explicit ext_harq_buffer_context_repository(unsigned nof_codeblocks_, unsigned ext_harq_buff_size, bool debug_mode_) :
+  explicit ext_harq_buffer_context_repository(unsigned nof_codeblocks_, uint64_t ext_harq_buff_size, bool debug_mode_) :
     nof_codeblocks(nof_codeblocks_)
   {
-    unsigned requested_size = nof_codeblocks_ * HARQ_INCR;
+    uint64_t requested_size = static_cast<uint64_t>(nof_codeblocks_) * static_cast<uint64_t>(HARQ_INCR.value());
     srsran_assert(requested_size <= ext_harq_buff_size,
                   "Requested size ({} bytes) for {} codeblocks exceeds external HARQ buffer capacity ({} bytes).",
                   requested_size,

include/srsran/hal/phy/upper/channel_processors/pusch/ext_harq_buffer_context_repository_factory.h

@@ -18,7 +18,7 @@ namespace hal {
 /// Returns a ext_harq_buffer_context_repository instance on success, otherwise returns nullptr.
 std::shared_ptr<ext_harq_buffer_context_repository>
 create_ext_harq_buffer_context_repository(unsigned nof_codeblocks,
-                                          unsigned ext_harq_buff_size,
+                                          uint64_t ext_harq_buff_size,
                                           bool     debug_mode = false);
 
 } // namespace hal

lib/hal/dpdk/bbdev/ldpc/bbdev_ldpc_decoder.cpp

@@ -131,7 +131,7 @@ bool dpdk::set_ldpc_dec_bbdev_data(::rte_bbdev_dec_op&   dec_op,
   dec_op.ldpc_dec.input.length += cw_len;
 
   // Harq offset for the current CB (based on its unique absolute ID).
-  unsigned harq_offset = srsran::hal::HARQ_INCR * absolute_cb_id;
+  unsigned harq_offset = srsran::hal::HARQ_INCR.value() * absolute_cb_id;
 
   // Input HARQ combining data is only needed in case of a retransmission.
   if (!new_data) {

lib/hal/phy/upper/channel_processors/pusch/ext_harq_buffer_context_repository_factory.cpp

@@ -14,7 +14,9 @@ using namespace srsran;
 using namespace hal;
 
 std::shared_ptr<ext_harq_buffer_context_repository>
-srsran::hal::create_ext_harq_buffer_context_repository(unsigned nof_rnti, unsigned nof_harq_id, bool debug_mode)
+srsran::hal::create_ext_harq_buffer_context_repository(unsigned nof_codeblocks,
+                                                       uint64_t ext_harq_buff_size,
+                                                       bool     debug_mode)
 {
-  return std::make_shared<ext_harq_buffer_context_repository>(nof_rnti, nof_harq_id, debug_mode);
+  return std::make_shared<ext_harq_buffer_context_repository>(nof_codeblocks, ext_harq_buff_size, debug_mode);
 }

lib/phy/upper/channel_processors/pdsch_encoder_hw_impl.cpp

@@ -120,8 +120,10 @@ void pdsch_encoder_hw_impl::encode(span<uint8_t>        codeword,
         srsran_assert(offset + rm_length <= codeword.size(), "Wrong codeword length.");
 
         codeblock = span<uint8_t>(codeword).subspan(offset, rm_length);
+        codeblock_packed.resize(rm_length);
       } else {
         codeblock = codeword;
+        codeblock_packed.resize(codeword.size());
       }
 
       // Make sure at least one operation is dequeued.

lib/phy/upper/channel_processors/pusch/pusch_decoder_hw_impl.h

@@ -73,6 +73,7 @@ class pusch_decoder_hw_impl : public pusch_decoder, private pusch_decoder_buffer
     srsran_assert(crc_set.crc24B->get_generator_poly() == crc_generator_poly::CRC24B,
                   "Not a CRC generator of type CRC24B.");
     srsran_assert(decoder, "Invalid hardware-accelerated PUSCH decoder.");
+    absolute_cb_ids.resize(MAX_NOF_SEGMENTS);
   }
 
   // See interface for the documentation.

tests/benchmarks/phy/upper/channel_processors/pusch/pusch_decoder_hwacc_benchmark.cpp

@@ -206,7 +206,7 @@ static std::shared_ptr<hal::hw_accelerator_pusch_dec_factory> create_hw_accelera
 
   // Interfacing to a shared external HARQ buffer context repository.
   unsigned nof_cbs                   = MAX_NOF_SEGMENTS;
-  unsigned acc100_ext_harq_buff_size = bbdev_accelerator->get_harq_buff_size().value();
+  uint64_t acc100_ext_harq_buff_size = bbdev_accelerator->get_harq_buff_size_bytes();
   std::shared_ptr<ext_harq_buffer_context_repository> harq_buffer_context =
       create_ext_harq_buffer_context_repository(nof_cbs, acc100_ext_harq_buff_size, test_harq);
   TESTASSERT(harq_buffer_context);

tests/benchmarks/phy/upper/channel_processors/pusch/pusch_processor_benchmark.cpp

@@ -489,7 +489,7 @@ static std::shared_ptr<hal::hw_accelerator_pusch_dec_factory> create_hw_accelera
 
   // Interfacing to a shared external HARQ buffer context repository.
   unsigned nof_cbs                   = MAX_NOF_SEGMENTS;
-  unsigned acc100_ext_harq_buff_size = bbdev_accelerator->get_harq_buff_size().value();
+  uint64_t acc100_ext_harq_buff_size = bbdev_accelerator->get_harq_buff_size_bytes();
   std::shared_ptr<ext_harq_buffer_context_repository> harq_buffer_context =
       create_ext_harq_buffer_context_repository(nof_cbs, acc100_ext_harq_buff_size, false);
   TESTASSERT(harq_buffer_context);

tests/unittests/phy/upper/channel_processors/pusch/pusch_decoder_vectortest.cpp

@@ -187,7 +187,7 @@ static std::shared_ptr<hal::hw_accelerator_pusch_dec_factory> create_hw_accelera
 
   // Interfacing to a shared external HARQ buffer context repository.
   unsigned nof_cbs                   = MAX_NOF_SEGMENTS;
-  unsigned acc100_ext_harq_buff_size = bbdev_accelerator->get_harq_buff_size().value();
+  uint64_t acc100_ext_harq_buff_size = bbdev_accelerator->get_harq_buff_size_bytes();
   std::shared_ptr<ext_harq_buffer_context_repository> harq_buffer_context =
       create_ext_harq_buffer_context_repository(nof_cbs, acc100_ext_harq_buff_size, test_harq);
   TESTASSERT(harq_buffer_context);