Merge pull request 'WavPack decoding and seeking speed improvements' (#342) from ayumi/tangara-fw:wavpack into main

Reviewed-on: https://codeberg.org/cool-tech-zone/tangara-fw/pulls/342

5 files changed, 46 insertions, 25 deletions

diff --git a/lib/komihash/include/komihash.h b/lib/komihash/include/komihash.h
index 7a72fda8..4f6b0f25 100644
--- a/lib/komihash/include/komihash.h
+++ b/lib/komihash/include/komihash.h
@@ -35,6 +35,8 @@
 #include <stdint.h>
 #include <string.h>
 
+#include "esp_attr.h"
+
 // Macros that apply byte-swapping.
 
 #if defined( __GNUC__ ) || defined( __clang__ )
@@ -142,7 +144,7 @@
  * @return Endianness-corrected 32-bit value from memory.
  */
 
-static inline uint32_t kh_lu32ec( const uint8_t* const p )
+static IRAM_ATTR inline uint32_t kh_lu32ec( const uint8_t* const p )
 {
 	uint32_t v;
 	memcpy( &v, p, 4 );
@@ -160,7 +162,7 @@ static inline uint32_t kh_lu32ec( const uint8_t* const p )
  * @return Endianness-corrected 64-bit value from memory.
  */
 
-static inline uint64_t kh_lu64ec( const uint8_t* const p )
+static IRAM_ATTR inline uint64_t kh_lu64ec( const uint8_t* const p )
 {
 	uint64_t v;
 	memcpy( &v, p, 8 );
@@ -179,7 +181,7 @@ static inline uint64_t kh_lu64ec( const uint8_t* const p )
  * @return Final byte-padded value from the message.
  */
 
-static inline uint64_t kh_lpu64ec_l3( const uint8_t* const Msg,
+static IRAM_ATTR inline uint64_t kh_lpu64ec_l3( const uint8_t* const Msg,
 	const size_t MsgLen )
 {
 	const int ml8 = -(int) ( MsgLen * 8 );
@@ -211,7 +213,7 @@ static inline uint64_t kh_lpu64ec_l3( const uint8_t* const Msg,
  * @return Final byte-padded value from the message.
  */
 
-static inline uint64_t kh_lpu64ec_nz( const uint8_t* const Msg,
+static IRAM_ATTR inline uint64_t kh_lpu64ec_nz( const uint8_t* const Msg,
 	const size_t MsgLen )
 {
 	const int ml8 = -(int) ( MsgLen * 8 );
@@ -252,7 +254,7 @@ static inline uint64_t kh_lpu64ec_nz( const uint8_t* const Msg,
  * @return Final byte-padded value from the message.
  */
 
-static inline uint64_t kh_lpu64ec_l4( const uint8_t* const Msg,
+static IRAM_ATTR inline uint64_t kh_lpu64ec_l4( const uint8_t* const Msg,
 	const size_t MsgLen )
 {
 	const int ml8 = -(int) ( MsgLen * 8 );
@@ -280,7 +282,7 @@ static inline uint64_t kh_lpu64ec_l4( const uint8_t* const Msg,
 	 * @param[out] rh The higher half of the 128-bit result.
 	 */
 
-	static inline void kh_m128( const uint64_t m1, const uint64_t m2,
+	static IRAM_ATTR inline void kh_m128( const uint64_t m1, const uint64_t m2,
 		uint64_t* const rl, uint64_t* const rh )
 	{
 		const __uint128_t r = (__uint128_t) m1 * m2;
@@ -293,7 +295,7 @@ static inline uint64_t kh_lpu64ec_l4( const uint8_t* const Msg,
 
 	#include <intrin.h>
 
-	static inline void kh_m128( const uint64_t m1, const uint64_t m2,
+	static IRAM_ATTR inline void kh_m128( const uint64_t m1, const uint64_t m2,
 		uint64_t* const rl, uint64_t* const rh )
 	{
 		*rl = _umul128( m1, m2, rh );
@@ -305,12 +307,12 @@ static inline uint64_t kh_lpu64ec_l4( const uint8_t* const Msg,
 	// from https://go.dev/src/runtime/softfloat64.go
 	// Licensed under BSD-style license.
 
-	static inline uint64_t kh__emulu( const uint32_t x, const uint32_t y )
+	static IRAM_ATTR inline uint64_t kh__emulu( const uint32_t x, const uint32_t y )
 	{
 		return( x * (uint64_t) y );
 	}
 
-	static inline void kh_m128( const uint64_t u, const uint64_t v,
+	static IRAM_ATTR inline void kh_m128( const uint64_t u, const uint64_t v,
 		uint64_t* const rl, uint64_t* const rh )
 	{
 		*rl = u * v;
@@ -406,7 +408,7 @@ static inline uint64_t kh_lpu64ec_l4( const uint8_t* const Msg,
  * @return 64-bit hash value.
  */
 
-static inline uint64_t komihash_epi( const uint8_t* Msg, size_t MsgLen,
+static IRAM_ATTR inline uint64_t komihash_epi( const uint8_t* Msg, size_t MsgLen,
 	uint64_t Seed1, uint64_t Seed5 )
 {
 	uint64_t r1h, r2h;
@@ -463,7 +465,7 @@ static inline uint64_t komihash_epi( const uint8_t* Msg, size_t MsgLen,
  * @return 64-bit hash of the input data.
  */
 
-static inline uint64_t komihash( const void* const Msg0, size_t MsgLen,
+static IRAM_ATTR inline uint64_t komihash( const void* const Msg0, size_t MsgLen,
 	const uint64_t UseSeed )
 {
 	const uint8_t* Msg = (const uint8_t*) Msg0;
@@ -571,7 +573,7 @@ static inline uint64_t komihash( const void* const Msg0, size_t MsgLen,
  * @return The next uniformly-random 64-bit value.
  */
 
-static inline uint64_t komirand( uint64_t* const Seed1, uint64_t* const Seed2 )
+static IRAM_ATTR inline uint64_t komirand( uint64_t* const Seed1, uint64_t* const Seed2 )
 {
 	uint64_t rh;
 
@@ -615,7 +617,7 @@ typedef struct {
  * little-endian systems.
  */
 
-static inline void komihash_stream_init( komihash_stream_t* const ctx,
+static IRAM_ATTR inline void komihash_stream_init( komihash_stream_t* const ctx,
 	const uint64_t UseSeed )
 {
 	ctx -> Seed[ 0 ] = UseSeed;
@@ -633,7 +635,7 @@ static inline void komihash_stream_init( komihash_stream_t* const ctx,
  * @param MsgLen Message's length, in bytes, can be zero.
  */
 
-static inline void komihash_stream_update( komihash_stream_t* const ctx,
+static IRAM_ATTR inline void komihash_stream_update( komihash_stream_t* const ctx,
 	const void* const Msg0, size_t MsgLen )
 {
 	const uint8_t* Msg = (const uint8_t*) Msg0;
@@ -772,7 +774,7 @@ static inline void komihash_stream_update( komihash_stream_t* const ctx,
  * @return 64-bit hash value.
  */
 
-static inline uint64_t komihash_stream_final( komihash_stream_t* const ctx )
+static IRAM_ATTR inline uint64_t komihash_stream_final( komihash_stream_t* const ctx )
 {
 	const uint8_t* Msg = ctx -> Buf;
 	size_t MsgLen = ctx -> BufFill;
@@ -817,7 +819,7 @@ static inline uint64_t komihash_stream_final( komihash_stream_t* const ctx )
  * @return 64-bit hash value.
  */
 
-static inline uint64_t komihash_stream_oneshot( const void* const Msg,
+static IRAM_ATTR inline uint64_t komihash_stream_oneshot( const void* const Msg,
 	const size_t MsgLen, const uint64_t UseSeed )
 {
 	komihash_stream_t ctx;
diff --git a/lib/wavpack/CMakeLists.txt b/lib/wavpack/CMakeLists.txt
index 98fcda95..f2f24933 100644
--- a/lib/wavpack/CMakeLists.txt
+++ b/lib/wavpack/CMakeLists.txt
@@ -1,4 +1,5 @@
 idf_component_register(
   SRCS bits.c float.c wputils.c metadata.c unpack.c words.c
   INCLUDE_DIRS .
+  LDFRAGMENTS wavpack.lf
 )
diff --git a/lib/wavpack/wavpack.lf b/lib/wavpack/wavpack.lf
new file mode 100644
index 00000000..686acad8
--- /dev/null
+++ b/lib/wavpack/wavpack.lf
@@ -0,0 +1,4 @@
+[mapping:libwavpack]
+archive: libwavpack.a
+entries:
+	* (noflash)
diff --git a/src/codecs/sample.cpp b/src/codecs/sample.cpp
index 63d14203..faf8b0f9 100644
--- a/src/codecs/sample.cpp
+++ b/src/codecs/sample.cpp
@@ -13,13 +13,17 @@
 
 namespace sample {
 
-static uint64_t sSeed1{0};
-static uint64_t sSeed2{0};
-
 auto shiftWithDither(int64_t src, uint_fast8_t bits) -> Sample {
   // FIXME: Use a better dither.
-  int16_t noise = static_cast<int16_t>(komirand(&sSeed1, &sSeed2) & 1);
-  return (src >> bits) ^ noise;
+  static uint64_t sSeed1{0};
+  static uint64_t sSeed2{0};
+  static uint64_t noise;
+  static uint_fast8_t pos = 0;
+  if (pos++ % 64 == 0)
+    noise = komirand(&sSeed1, &sSeed2);
+  else
+    noise >>= 1;
+  return (src >> bits) ^ (noise & 1);
 }
 
 }  // namespace sample
diff --git a/src/codecs/wavpack.cpp b/src/codecs/wavpack.cpp
index 21865785..709302e6 100644
--- a/src/codecs/wavpack.cpp
+++ b/src/codecs/wavpack.cpp
@@ -73,7 +73,7 @@ auto WavPackDecoder::OpenStream(std::shared_ptr<IStream> input, uint32_t offset)
         );
   const auto rate = WavpackGetSampleRate(&wavpack_);
   if (offset && total && input_.get()->CanSeek()) {
-    const uint32_t want = offset * rate - 1;
+    const uint32_t want = offset * rate;
     if (total < want) {
       ESP_LOGE(kTag, "seeking: offset points beyond the end of the file");
       return cpp::fail(Error::kInternalError);
@@ -96,7 +96,11 @@ auto WavPackDecoder::OpenStream(std::shared_ptr<IStream> input, uint32_t offset)
       }
       const uint32_t blockIndex = loadLe32(header + 16);
       const uint32_t blockSamples = loadLe32(header + 20);
-      if (want >= blockIndex && want <= blockIndex + blockSamples) {
+      if (want >= blockIndex && want == blockIndex + blockSamples) {
+        input_->SeekTo(size - 24, IStream::SeekFrom::kCurrentPosition);
+        target = 0;
+        break;
+      } else if (want >= blockIndex && want < blockIndex + blockSamples) {
         input_->SeekTo(-32, IStream::SeekFrom::kCurrentPosition);
         target = want - blockIndex;
         break;
@@ -150,8 +154,14 @@ auto WavPackDecoder::DecodeTo(std::span<sample::Sample> output)
     ESP_LOGE(kTag, "CRC error");
     return cpp::fail(Error::kMalformedData);
   }
-  for (size_t i = 0; i < samples; i++)
-    output[i] = sample::FromSigned(buf_[i], bitdepth_);
+  if (bitdepth_ == 16)
+    for (size_t i = 0; i < samples; i++)
+      output[i] = buf_[i];
+  else if (bitdepth_ > 16)
+    for (size_t i = 0; i < samples; i++)
+      output[i] = sample::shiftWithDither(buf_[i], bitdepth_ - 16);
+  else for (size_t i = 0; i < samples; i++)
+    output[i] = buf_[i] << (16 - bitdepth_);
   return OutputInfo{
       .samples_written = samples,
       .is_stream_finished = samples == 0,