From 9d0251f7a6751a0d998da4fad81a35a043608275 Mon Sep 17 00:00:00 2001 From: HiFiPhile Date: Tue, 19 Sep 2023 15:42:52 +0200 Subject: Fix UAC interleaved copy. --- src/class/audio/audio_device.c | 113 +++++++++++++++++++---------------------- 1 file changed, 51 insertions(+), 62 deletions(-) (limited to 'src/class') diff --git a/src/class/audio/audio_device.c b/src/class/audio/audio_device.c index 1a7ce7870..0e4220f4f 100644 --- a/src/class/audio/audio_device.c +++ b/src/class/audio/audio_device.c @@ -631,68 +631,59 @@ static bool audiod_rx_done_cb(uint8_t rhport, audiod_function_t* audio, uint16_t // Decoding according to 2.3.1.5 Audio Streams // Helper function -static inline uint8_t * audiod_interleaved_copy_bytes_fast_decode(uint16_t const nBytesToCopy, void * dst, uint8_t * dst_end, uint8_t * src, uint8_t const n_ff_used) +static inline uint8_t * audiod_interleaved_copy_bytes_fast_decode(uint16_t const nBytesToCopy, uint8_t * dst, uint8_t * dst_end, uint8_t * src, uint8_t const n_ff_used) { - // This function is an optimized version of - // while((uint8_t *)dst < dst_end) - // { - // memcpy(dst, src, nBytesToCopy); - // dst = (uint8_t *)dst + nBytesToCopy; - // src += nBytesToCopy * n_ff_used; - // } - - // Optimize for fast half word copies - typedef struct{ - uint16_t val; - } __attribute((__packed__)) unaligned_uint16_t; - - // Optimize for fast word copies - typedef struct{ - uint32_t val; - } __attribute((__packed__)) unaligned_uint32_t; + // Due to one FIFO contains 2 channels, data always aligned to (nBytesToCopy * 2) switch (nBytesToCopy) { case 1: while((uint8_t *)dst < dst_end) { - *(uint8_t *)dst++ = *src; - src += n_ff_used; + *(uint16_t*)dst = *(uint16_t*)src; + src += 2; + dst += 2; + src += 2 * (n_ff_used - 1); } break; case 2: while((uint8_t *)dst < dst_end) { - *(unaligned_uint16_t*)dst = *(unaligned_uint16_t*)src; - dst += 2; - src += 2 * n_ff_used; + *(uint32_t*)dst = *(uint32_t*)src; + src += 4; + dst += 4; + src += 4 * (n_ff_used - 1); } break; case 3: while((uint8_t *)dst < dst_end) { - // memcpy(dst, src, 3); - // dst = (uint8_t *)dst + 3; - // src += 3 * n_ff_used; - - // TODO: Is there a faster way to copy 3 bytes? - *(uint8_t *)dst++ = *src++; - *(uint8_t *)dst++ = *src++; - *(uint8_t *)dst++ = *src++; - - src += 3 * (n_ff_used - 1); + *(uint16_t*)dst = *(uint16_t*)src; + src += 2; + dst += 2; + *(uint16_t*)dst = *(uint16_t*)src; + src += 2; + dst += 2; + *(uint16_t*)dst = *(uint16_t*)src; + src += 2; + dst += 2; + src += 6 * (n_ff_used - 1); } break; case 4: while((uint8_t *)dst < dst_end) { - *(unaligned_uint32_t*)dst = *(unaligned_uint32_t*)src; + *(uint32_t*)dst++ = *(uint32_t*)src++; + src += 4; + dst += 4; + *(uint32_t*)dst++ = *(uint32_t*)src++; + src += 4; dst += 4; - src += 4 * n_ff_used; + src += 8 * (n_ff_used - 1); } break; } @@ -946,57 +937,55 @@ range [-1, +1) // Helper function static inline uint8_t * audiod_interleaved_copy_bytes_fast_encode(uint16_t const nBytesToCopy, uint8_t * src, uint8_t * src_end, uint8_t * dst, uint8_t const n_ff_used) { - // Optimize for fast half word copies - typedef struct{ - uint16_t val; - } __attribute((__packed__)) unaligned_uint16_t; - - // Optimize for fast word copies - typedef struct{ - uint32_t val; - } __attribute((__packed__)) unaligned_uint32_t; - + // Due to one FIFO contains 2 channels, data always aligned to (nBytesToCopy * 2) switch (nBytesToCopy) { case 1: while(src < src_end) { - *dst = *src++; - dst += n_ff_used; + *(uint16_t*)dst = *(uint16_t*)src; + src += 2; + dst += 2; + dst += 2 * (n_ff_used - 1); } break; case 2: while(src < src_end) { - *(unaligned_uint16_t*)dst = *(unaligned_uint16_t*)src; - src += 2; - dst += 2 * n_ff_used; + *(uint32_t*)dst = *(uint32_t*)src; + src += 4; + dst += 4; + dst += 4 * (n_ff_used - 1); } break; case 3: while(src < src_end) { - // memcpy(dst, src, 3); - // src = (uint8_t *)src + 3; - // dst += 3 * n_ff_used; - - // TODO: Is there a faster way to copy 3 bytes? - *dst++ = *src++; - *dst++ = *src++; - *dst++ = *src++; - - dst += 3 * (n_ff_used - 1); + *(uint16_t*)dst = *(uint16_t*)src; + src += 2; + dst += 2; + *(uint16_t*)dst = *(uint16_t*)src; + src += 2; + dst += 2; + *(uint16_t*)dst = *(uint16_t*)src; + src += 2; + dst += 2; + dst += 6 * (n_ff_used - 1); } break; case 4: while(src < src_end) { - *(unaligned_uint32_t*)dst = *(unaligned_uint32_t*)src; + *(uint32_t*)dst++ = *(uint32_t*)src++; + src += 4; + dst += 4; + *(uint32_t*)dst++ = *(uint32_t*)src++; src += 4; - dst += 4 * n_ff_used; + dst += 4; + dst += 8 * (n_ff_used - 1); } break; } -- cgit v1.3.1 From 6be7f354c26e09a87d1b856b0e237ff82be5963f Mon Sep 17 00:00:00 2001 From: HiFiPhile Date: Tue, 19 Sep 2023 16:32:49 +0200 Subject: Fix cast-align. --- src/class/audio/audio_device.c | 196 +++++++++++++++++++---------------------- 1 file changed, 90 insertions(+), 106 deletions(-) (limited to 'src/class') diff --git a/src/class/audio/audio_device.c b/src/class/audio/audio_device.c index 0e4220f4f..46fd6dfa9 100644 --- a/src/class/audio/audio_device.c +++ b/src/class/audio/audio_device.c @@ -631,64 +631,55 @@ static bool audiod_rx_done_cb(uint8_t rhport, audiod_function_t* audio, uint16_t // Decoding according to 2.3.1.5 Audio Streams // Helper function -static inline uint8_t * audiod_interleaved_copy_bytes_fast_decode(uint16_t const nBytesToCopy, uint8_t * dst, uint8_t * dst_end, uint8_t * src, uint8_t const n_ff_used) +static inline void * audiod_interleaved_copy_bytes_fast_decode(uint16_t const nBytesToCopy, void * dst, const void * dst_end, void * src, uint8_t const n_ff_used) { - // Due to one FIFO contains 2 channels, data always aligned to (nBytesToCopy * 2) - - switch (nBytesToCopy) + uint16_t * dst16 = dst; + uint16_t * src16 = src; + const uint16_t * dst_end16 = dst_end; + uint32_t * dst32 = dst; + uint32_t * src32 = src; + const uint32_t * dst_end32 = dst_end; + + if (nBytesToCopy == 1) { - case 1: - while((uint8_t *)dst < dst_end) - { - *(uint16_t*)dst = *(uint16_t*)src; - src += 2; - dst += 2; - src += 2 * (n_ff_used - 1); - } - break; - - case 2: - while((uint8_t *)dst < dst_end) - { - *(uint32_t*)dst = *(uint32_t*)src; - src += 4; - dst += 4; - src += 4 * (n_ff_used - 1); - } - break; - - case 3: - while((uint8_t *)dst < dst_end) - { - *(uint16_t*)dst = *(uint16_t*)src; - src += 2; - dst += 2; - *(uint16_t*)dst = *(uint16_t*)src; - src += 2; - dst += 2; - *(uint16_t*)dst = *(uint16_t*)src; - src += 2; - dst += 2; - src += 6 * (n_ff_used - 1); - } - break; - - case 4: - while((uint8_t *)dst < dst_end) - { - *(uint32_t*)dst++ = *(uint32_t*)src++; - src += 4; - dst += 4; - *(uint32_t*)dst++ = *(uint32_t*)src++; - src += 4; - dst += 4; - src += 8 * (n_ff_used - 1); - } - break; + while(dst16 < dst_end16) + { + *dst16++ = *src16++; + src16 += n_ff_used - 1; + } + return src16; + } + else if (nBytesToCopy == 2) + { + while(dst32 < dst_end32) + { + *dst32++ = *src32++; + src32 += n_ff_used - 1; + } + return src32; + } + else if (nBytesToCopy == 3) + { + while(dst16 < dst_end16) + { + *dst16++ = *src16++; + *dst16++ = *src16++; + *dst16++ = *src16++; + src16 += 3 * (n_ff_used - 1); + } + return src16; + } + else // nBytesToCopy == 4 + { + while(dst32 < dst_end32) + { + *dst32++ = *src32++; + *dst32++ = *src32++; + src32 += 2 * (n_ff_used - 1); + } + return src32; } - - return src; } static bool audiod_decode_type_I_pcm(uint8_t rhport, audiod_function_t* audio, uint16_t n_bytes_received) @@ -935,62 +926,55 @@ range [-1, +1) * */ // Helper function -static inline uint8_t * audiod_interleaved_copy_bytes_fast_encode(uint16_t const nBytesToCopy, uint8_t * src, uint8_t * src_end, uint8_t * dst, uint8_t const n_ff_used) +static inline void * audiod_interleaved_copy_bytes_fast_encode(uint16_t const nBytesToCopy, void * src, const void * src_end, void * dst, uint8_t const n_ff_used) { // Due to one FIFO contains 2 channels, data always aligned to (nBytesToCopy * 2) - switch (nBytesToCopy) + uint16_t * dst16 = dst; + uint16_t * src16 = src; + const uint16_t * src_end16 = src_end; + uint32_t * dst32 = dst; + uint32_t * src32 = src; + const uint32_t * src_end32 = src_end; + + if (nBytesToCopy == 1) { - case 1: - while(src < src_end) - { - *(uint16_t*)dst = *(uint16_t*)src; - src += 2; - dst += 2; - dst += 2 * (n_ff_used - 1); - } - break; - - case 2: - while(src < src_end) - { - *(uint32_t*)dst = *(uint32_t*)src; - src += 4; - dst += 4; - dst += 4 * (n_ff_used - 1); - } - break; - - case 3: - while(src < src_end) - { - *(uint16_t*)dst = *(uint16_t*)src; - src += 2; - dst += 2; - *(uint16_t*)dst = *(uint16_t*)src; - src += 2; - dst += 2; - *(uint16_t*)dst = *(uint16_t*)src; - src += 2; - dst += 2; - dst += 6 * (n_ff_used - 1); - } - break; - - case 4: - while(src < src_end) - { - *(uint32_t*)dst++ = *(uint32_t*)src++; - src += 4; - dst += 4; - *(uint32_t*)dst++ = *(uint32_t*)src++; - src += 4; - dst += 4; - dst += 8 * (n_ff_used - 1); - } - break; + while(src16 < src_end16) + { + *dst16++ = *src16++; + dst16 += n_ff_used - 1; + } + return dst16; + } + else if (nBytesToCopy == 2) + { + while(src32 < src_end32) + { + *dst32++ = *src32++; + dst32 += n_ff_used - 1; + } + return dst32; + } + else if (nBytesToCopy == 3) + { + while(src16 < src_end16) + { + *dst16++ = *src16++; + *dst16++ = *src16++; + *dst16++ = *src16++; + dst16 += 3 * (n_ff_used - 1); + } + return dst16; + } + else // nBytesToCopy == 4 + { + while(src32 < src_end32) + { + *dst32++ = *src32++; + *dst32++ = *src32++; + dst32 += 2 * (n_ff_used - 1); + } + return dst32; } - - return dst; } static uint16_t audiod_encode_type_I_pcm(uint8_t rhport, audiod_function_t* audio) -- cgit v1.3.1 From 8b2cb63da88c0f790fee8d514971153569993b05 Mon Sep 17 00:00:00 2001 From: HiFiPhile Date: Tue, 10 Oct 2023 18:22:20 +0200 Subject: Fix naming nBytesPerSample. --- src/class/audio/audio_device.c | 24 ++++++++++++------------ 1 file changed, 12 insertions(+), 12 deletions(-) (limited to 'src/class') diff --git a/src/class/audio/audio_device.c b/src/class/audio/audio_device.c index 46fd6dfa9..5d3772a9d 100644 --- a/src/class/audio/audio_device.c +++ b/src/class/audio/audio_device.c @@ -631,9 +631,9 @@ static bool audiod_rx_done_cb(uint8_t rhport, audiod_function_t* audio, uint16_t // Decoding according to 2.3.1.5 Audio Streams // Helper function -static inline void * audiod_interleaved_copy_bytes_fast_decode(uint16_t const nBytesToCopy, void * dst, const void * dst_end, void * src, uint8_t const n_ff_used) +static inline void * audiod_interleaved_copy_bytes_fast_decode(uint16_t const nBytesPerSample, void * dst, const void * dst_end, void * src, uint8_t const n_ff_used) { - // Due to one FIFO contains 2 channels, data always aligned to (nBytesToCopy * 2) + // Due to one FIFO contains 2 channels, data always aligned to (nBytesPerSample * 2) uint16_t * dst16 = dst; uint16_t * src16 = src; const uint16_t * dst_end16 = dst_end; @@ -641,7 +641,7 @@ static inline void * audiod_interleaved_copy_bytes_fast_decode(uint16_t const nB uint32_t * src32 = src; const uint32_t * dst_end32 = dst_end; - if (nBytesToCopy == 1) + if (nBytesPerSample == 1) { while(dst16 < dst_end16) { @@ -650,7 +650,7 @@ static inline void * audiod_interleaved_copy_bytes_fast_decode(uint16_t const nB } return src16; } - else if (nBytesToCopy == 2) + else if (nBytesPerSample == 2) { while(dst32 < dst_end32) { @@ -659,7 +659,7 @@ static inline void * audiod_interleaved_copy_bytes_fast_decode(uint16_t const nB } return src32; } - else if (nBytesToCopy == 3) + else if (nBytesPerSample == 3) { while(dst16 < dst_end16) { @@ -670,7 +670,7 @@ static inline void * audiod_interleaved_copy_bytes_fast_decode(uint16_t const nB } return src16; } - else // nBytesToCopy == 4 + else // nBytesPerSample == 4 { while(dst32 < dst_end32) { @@ -926,9 +926,9 @@ range [-1, +1) * */ // Helper function -static inline void * audiod_interleaved_copy_bytes_fast_encode(uint16_t const nBytesToCopy, void * src, const void * src_end, void * dst, uint8_t const n_ff_used) +static inline void * audiod_interleaved_copy_bytes_fast_encode(uint16_t const nBytesPerSample, void * src, const void * src_end, void * dst, uint8_t const n_ff_used) { - // Due to one FIFO contains 2 channels, data always aligned to (nBytesToCopy * 2) + // Due to one FIFO contains 2 channels, data always aligned to (nBytesPerSample * 2) uint16_t * dst16 = dst; uint16_t * src16 = src; const uint16_t * src_end16 = src_end; @@ -936,7 +936,7 @@ static inline void * audiod_interleaved_copy_bytes_fast_encode(uint16_t const nB uint32_t * src32 = src; const uint32_t * src_end32 = src_end; - if (nBytesToCopy == 1) + if (nBytesPerSample == 1) { while(src16 < src_end16) { @@ -945,7 +945,7 @@ static inline void * audiod_interleaved_copy_bytes_fast_encode(uint16_t const nB } return dst16; } - else if (nBytesToCopy == 2) + else if (nBytesPerSample == 2) { while(src32 < src_end32) { @@ -954,7 +954,7 @@ static inline void * audiod_interleaved_copy_bytes_fast_encode(uint16_t const nB } return dst32; } - else if (nBytesToCopy == 3) + else if (nBytesPerSample == 3) { while(src16 < src_end16) { @@ -965,7 +965,7 @@ static inline void * audiod_interleaved_copy_bytes_fast_encode(uint16_t const nB } return dst16; } - else // nBytesToCopy == 4 + else // nBytesPerSample == 4 { while(src32 < src_end32) { -- cgit v1.3.1