Use nanors for optimized Reed-Solomon FEC decoding (#125)

src/RtpAudioQueue.c

@@ -55,8 +55,7 @@ void RtpaInitializeQueue(PRTP_AUDIO_QUEUE queue) {
     // works correctly. This is possible because the data and FEC shard count is
     // constant and known in advance.
     const unsigned char parity[] = { 0x77, 0x40, 0x38, 0x0e, 0xc7, 0xa7, 0x0d, 0x6c };
-    memcpy(&queue->rs->m[16], parity, sizeof(parity));
-    memcpy(queue->rs->parity, parity, sizeof(parity));
+    memcpy(queue->rs->p, parity, sizeof(parity));
 }
 
 static void validateFecBlockState(PRTP_AUDIO_QUEUE queue) {
@@ -444,7 +443,7 @@ static bool completeFecBlock(PRTP_AUDIO_QUEUE queue, PRTPA_FEC_BLOCK block) {
     memset(block->dataPackets[dropIndex], 0, sizeof(RTP_PACKET) + block->blockSize);
 #endif
 
-    int res = reed_solomon_reconstruct(queue->rs, shards, block->marks, RTPA_TOTAL_SHARDS, block->blockSize);
+    int res = reed_solomon_decode(queue->rs, shards, block->marks, RTPA_TOTAL_SHARDS, block->blockSize);
     if (res != 0) {
         // We should always have enough data to recover the entire block since we checked above.
         LC_ASSERT(res == 0);

src/RtpAudioQueue.h

@@ -2,7 +2,14 @@
 
 #include "Video.h"
 
-#include "rs.h"
+#include "rswrapper.h"
+
+typedef struct _reed_solomon {
+    int ds;
+    int ps;
+    int ts;
+    uint8_t p[];
+} reed_solomon;
 
 // Maximum time to wait for an OOS data/FEC shard
 // after the entire FEC block should have been received

src/RtpVideoQueue.c

@@ -1,5 +1,5 @@
 #include "Limelight-internal.h"
-#include "rs.h"
+#include "rswrapper.h"
 
 #if defined(LC_DEBUG) && !defined(LC_FUZZING)
 // This enables FEC validation mode with a synthetic drop
@@ -328,7 +328,7 @@ static int reconstructFrame(PRTP_VIDEO_QUEUE queue) {
         }
     }
 
-    ret = reed_solomon_reconstruct(rs, packets, marks, totalPackets, receiveSize);
+    ret = reed_solomon_decode(rs, packets, marks, totalPackets, receiveSize);
 
     // We should always provide enough parity to recover the missing data successfully.
     // If this fails, something is probably wrong with our FEC state.

src/rswrapper.c

@@ -0,0 +1,191 @@
+/**
+ * @file src/rswrapper.c
+ * @brief Wrappers for nanors vectorization with different ISA options
+ */
+
+// _FORTIY_SOURCE can cause some versions of GCC to try to inline
+// memset() with incompatible target options when compiling rs.c
+#ifdef _FORTIFY_SOURCE
+  #undef _FORTIFY_SOURCE
+#endif
+
+// The assert() function is decorated with __cold on macOS which
+// is incompatible with Clang's target multiversioning feature
+#ifndef NDEBUG
+  #define NDEBUG
+#endif
+
+#define DECORATE_FUNC_I(a, b) a##b
+#define DECORATE_FUNC(a, b) DECORATE_FUNC_I(a, b)
+
+// Append an ISA suffix to the public RS API
+#define reed_solomon_init DECORATE_FUNC(reed_solomon_init, ISA_SUFFIX)
+#define reed_solomon_new DECORATE_FUNC(reed_solomon_new, ISA_SUFFIX)
+#define reed_solomon_new_static DECORATE_FUNC(reed_solomon_new_static, ISA_SUFFIX)
+#define reed_solomon_release DECORATE_FUNC(reed_solomon_release, ISA_SUFFIX)
+#define reed_solomon_decode DECORATE_FUNC(reed_solomon_decode, ISA_SUFFIX)
+#define reed_solomon_encode DECORATE_FUNC(reed_solomon_encode, ISA_SUFFIX)
+
+// Append an ISA suffix to internal functions to prevent multiple definition errors
+#define obl_axpy_ref DECORATE_FUNC(obl_axpy_ref, ISA_SUFFIX)
+#define obl_scal_ref DECORATE_FUNC(obl_scal_ref, ISA_SUFFIX)
+#define obl_axpyb32_ref DECORATE_FUNC(obl_axpyb32_ref, ISA_SUFFIX)
+#define obl_axpy DECORATE_FUNC(obl_axpy, ISA_SUFFIX)
+#define obl_scal DECORATE_FUNC(obl_scal, ISA_SUFFIX)
+#define obl_swap DECORATE_FUNC(obl_swap, ISA_SUFFIX)
+#define obl_axpyb32 DECORATE_FUNC(obl_axpyb32, ISA_SUFFIX)
+#define axpy DECORATE_FUNC(axpy, ISA_SUFFIX)
+#define scal DECORATE_FUNC(scal, ISA_SUFFIX)
+#define gemm DECORATE_FUNC(gemm, ISA_SUFFIX)
+#define invert_mat DECORATE_FUNC(invert_mat, ISA_SUFFIX)
+
+#if defined(__x86_64__) || defined(__i386__) || (defined(_MSC_VER) && (defined(_M_IX86) || defined(_M_AMD64)))
+
+  // Compile a variant for SSSE3
+  #if defined(__clang__)
+    #pragma clang attribute push(__attribute__((target("ssse3"))), apply_to = function)
+  #elif __GNUC__
+    #pragma GCC push_options
+    #pragma GCC target("ssse3")
+  #endif
+  #define ISA_SUFFIX _ssse3
+  #define OBLAS_SSE3
+  #include "../nanors/rs.c"
+  #undef OBLAS_SSE3
+  #undef ISA_SUFFIX
+  #if defined(__clang__)
+    #pragma clang attribute pop
+  #elif __GNUC__
+    #pragma GCC pop_options
+  #endif
+
+  // Compile a variant for AVX2
+  #if defined(__clang__)
+    #pragma clang attribute push(__attribute__((target("avx2"))), apply_to = function)
+  #elif __GNUC__
+    #pragma GCC push_options
+    #pragma GCC target("avx2")
+  #endif
+  #define ISA_SUFFIX _avx2
+  #define OBLAS_AVX2
+  #include "../nanors/rs.c"
+  #undef OBLAS_AVX2
+  #undef ISA_SUFFIX
+  #if defined(__clang__)
+    #pragma clang attribute pop
+  #elif __GNUC__
+    #pragma GCC pop_options
+  #endif
+
+  // Compile a variant for AVX512BW
+  #if defined(__clang__)
+    #pragma clang attribute push(__attribute__((target("avx512f,avx512bw"))), apply_to = function)
+  #elif __GNUC__
+    #pragma GCC push_options
+    #pragma GCC target("avx512f,avx512bw")
+  #endif
+  #define ISA_SUFFIX _avx512
+  #define OBLAS_AVX512
+  #include "../nanors/rs.c"
+  #undef OBLAS_AVX512
+  #undef ISA_SUFFIX
+  #if defined(__clang__)
+    #pragma clang attribute pop
+  #elif __GNUC__
+    #pragma GCC pop_options
+  #endif
+
+#endif
+
+// Compile a default variant
+#define ISA_SUFFIX _def
+#include "../nanors/deps/obl/autoshim.h"
+#include "../nanors/rs.c"
+#undef ISA_SUFFIX
+
+#undef reed_solomon_init
+#undef reed_solomon_new
+#undef reed_solomon_new_static
+#undef reed_solomon_release
+#undef reed_solomon_decode
+#undef reed_solomon_encode
+
+#include "rswrapper.h"
+
+reed_solomon_new_t reed_solomon_new_fn;
+reed_solomon_release_t reed_solomon_release_fn;
+reed_solomon_encode_t reed_solomon_encode_fn;
+reed_solomon_decode_t reed_solomon_decode_fn;
+
+#if defined(_MSC_VER) && (defined(_M_IX86) || defined(_M_AMD64))
+
+  #if defined(_M_AMD64)
+  // For some reason this is needed to avoid a "C1189 No target architecture" error from winnt.h
+  # define _AMD64_
+  #endif
+  #include <processthreadsapi.h>
+  BOOL _msc_supports_ssse3(void)   { return IsProcessorFeaturePresent(PF_SSSE3_INSTRUCTIONS_AVAILABLE); }
+  BOOL _msc_supports_avx2(void)    { return IsProcessorFeaturePresent(PF_AVX2_INSTRUCTIONS_AVAILABLE); }
+  BOOL _msc_supports_avx512f(void) { return IsProcessorFeaturePresent(PF_AVX512F_INSTRUCTIONS_AVAILABLE); }
+
+#endif
+
+/**
+ * @brief This initializes the RS function pointers to the best vectorized version available.
+ * @details The streaming code will directly invoke these function pointers during encoding.
+ */
+void reed_solomon_init(void) {
+#if defined(_MSC_VER) && (defined(_M_IX86) || defined(_M_AMD64))
+  // Visual Studio
+  if (_msc_supports_avx512f()) {
+    reed_solomon_new_fn = reed_solomon_new_avx512;
+    reed_solomon_release_fn = reed_solomon_release_avx512;
+    reed_solomon_encode_fn = reed_solomon_encode_avx512;
+    reed_solomon_decode_fn = reed_solomon_decode_avx512;
+    reed_solomon_init_avx512();
+  } else if (_msc_supports_avx2()) {
+    reed_solomon_new_fn = reed_solomon_new_avx2;
+    reed_solomon_release_fn = reed_solomon_release_avx2;
+    reed_solomon_encode_fn = reed_solomon_encode_avx2;
+    reed_solomon_decode_fn = reed_solomon_decode_avx2;
+    reed_solomon_init_avx2();
+  } else if (_msc_supports_ssse3()) {
+    reed_solomon_new_fn = reed_solomon_new_ssse3;
+    reed_solomon_release_fn = reed_solomon_release_ssse3;
+    reed_solomon_encode_fn = reed_solomon_encode_ssse3;
+    reed_solomon_decode_fn = reed_solomon_decode_ssse3;
+    reed_solomon_init_ssse3();
+  } else
+
+#elif defined(__x86_64__)
+  // gcc & clang
+  if (__builtin_cpu_supports("avx512f") && __builtin_cpu_supports("avx512bw")) {
+    reed_solomon_new_fn = reed_solomon_new_avx512;
+    reed_solomon_release_fn = reed_solomon_release_avx512;
+    reed_solomon_encode_fn = reed_solomon_encode_avx512;
+    reed_solomon_decode_fn = reed_solomon_decode_avx512;
+    reed_solomon_init_avx512();
+  } else if (__builtin_cpu_supports("avx2")) {
+    reed_solomon_new_fn = reed_solomon_new_avx2;
+    reed_solomon_release_fn = reed_solomon_release_avx2;
+    reed_solomon_encode_fn = reed_solomon_encode_avx2;
+    reed_solomon_decode_fn = reed_solomon_decode_avx2;
+    reed_solomon_init_avx2();
+  } else if (__builtin_cpu_supports("ssse3")) {
+    reed_solomon_new_fn = reed_solomon_new_ssse3;
+    reed_solomon_release_fn = reed_solomon_release_ssse3;
+    reed_solomon_encode_fn = reed_solomon_encode_ssse3;
+    reed_solomon_decode_fn = reed_solomon_decode_ssse3;
+    reed_solomon_init_ssse3();
+  } else
+
+#endif
+  //
+  {
+    reed_solomon_new_fn = reed_solomon_new_def;
+    reed_solomon_release_fn = reed_solomon_release_def;
+    reed_solomon_encode_fn = reed_solomon_encode_def;
+    reed_solomon_decode_fn = reed_solomon_decode_def;
+    reed_solomon_init_def();
+  }
+}

src/rswrapper.h

@@ -0,0 +1,32 @@
+/**
+ * @file src/rswrapper.h
+ * @brief Wrappers for nanors vectorization
+ * @details This is a drop-in replacement for nanors rs.h
+ */
+#pragma once
+
+// standard includes
+#include <stdint.h>
+
+typedef struct _reed_solomon reed_solomon;
+
+typedef reed_solomon *(*reed_solomon_new_t)(int data_shards, int parity_shards);
+typedef void (*reed_solomon_release_t)(reed_solomon *rs);
+typedef int (*reed_solomon_encode_t)(reed_solomon *rs, uint8_t **shards, int nr_shards, int bs);
+typedef int (*reed_solomon_decode_t)(reed_solomon *rs, uint8_t **shards, uint8_t *marks, int nr_shards, int bs);
+
+extern reed_solomon_new_t reed_solomon_new_fn;
+extern reed_solomon_release_t reed_solomon_release_fn;
+extern reed_solomon_encode_t reed_solomon_encode_fn;
+extern reed_solomon_decode_t reed_solomon_decode_fn;
+
+#define reed_solomon_new reed_solomon_new_fn
+#define reed_solomon_release reed_solomon_release_fn
+#define reed_solomon_encode reed_solomon_encode_fn
+#define reed_solomon_decode reed_solomon_decode_fn
+
+/**
+ * @brief This initializes the RS function pointers to the best vectorized version available.
+ * @details The streaming code will directly invoke these function pointers during encoding.
+ */
+void reed_solomon_init(void);