audio: ffmpeg_dec: add libavfilter graph support (afftdn, alimiter)

src/audio/ffmpeg_dec/CMakeLists.txt

@@ -24,6 +24,7 @@ else()
   else()
     add_local_sources(sof ffmpeg_dec-ffmpeg.c)
     add_local_sources(sof ffmpeg_dec-convert.c)
+    add_local_sources(sof ffmpeg-convert-hifi.c)
     if(CONFIG_FFMPEG_ENC)
       add_local_sources(sof ffmpeg-encode.c)
     endif()
@@ -33,6 +34,9 @@ else()
     if(CONFIG_FFMPEG_ENC_AAC)
       add_local_sources(sof ffmpeg-enc-aac.c)
     endif()
+    if(CONFIG_FFMPEG_FILTER)
+      add_local_sources(sof ffmpeg-filter.c)
+    endif()
     # Scalar soft-float for the SOF glue that touches float. core-isa.h reports
     # XCHAL_HAVE_FP=1 so LLVM would lower the FLT/FLTP -> Q1.31 sample conversion
     # in ffmpeg_dec_emit_frame() to hardware FP0 ops (lsi/mul.s/trunc.s), which
@@ -47,7 +51,9 @@ else()
     set(_ffmpeg_dec_sources
         ${CMAKE_CURRENT_SOURCE_DIR}/ffmpeg_dec.c
         ${CMAKE_CURRENT_SOURCE_DIR}/ffmpeg_dec-ffmpeg.c
-        ${CMAKE_CURRENT_SOURCE_DIR}/ffmpeg_dec-convert.c)
+        ${CMAKE_CURRENT_SOURCE_DIR}/ffmpeg_dec-convert.c
+        ${CMAKE_CURRENT_SOURCE_DIR}/ffmpeg-convert-hifi.c
+        ${CMAKE_CURRENT_SOURCE_DIR}/ffmpeg_dec-builtin-libc.c)
     if(CONFIG_FFMPEG_ENC)
       list(APPEND _ffmpeg_dec_sources ${CMAKE_CURRENT_SOURCE_DIR}/ffmpeg-encode.c)
     endif()
@@ -57,6 +63,9 @@ else()
     if(CONFIG_FFMPEG_ENC_AAC)
       list(APPEND _ffmpeg_dec_sources ${CMAKE_CURRENT_SOURCE_DIR}/ffmpeg-enc-aac.c)
     endif()
+    if(CONFIG_FFMPEG_FILTER)
+      list(APPEND _ffmpeg_dec_sources ${CMAKE_CURRENT_SOURCE_DIR}/ffmpeg-filter.c)
+    endif()
     if(CMAKE_C_COMPILER_ID MATCHES "Clang")
         set_source_files_properties(
             ${_ffmpeg_dec_sources}
@@ -103,6 +112,9 @@ else()
       # stubs in ffmpeg_dec-builtin-libc.c). FLAC is integer-only so it never
       # calls libm at runtime, but the references still resolve to real kernels
       # now, and lossy decoders (AAC/Opus/Vorbis) need them to decode correctly.
+      if(CONFIG_FFMPEG_BUILD_AVFILTER)
+        target_link_libraries(SOF INTERFACE avfilter)
+      endif()
       target_link_libraries(SOF INTERFACE avcodec swresample avutil)
       # Encode mode: libshine MP3 encoder archive (built by ffmpeg.cmake
       # section 3b when FFMPEG_ENC_MP3 is set). Linked AFTER avcodec,

src/audio/ffmpeg_dec/Kconfig

@@ -77,11 +77,65 @@ config FFMPEG_ENC
 	bool
 	default y if FFMPEG_ENC_MP3 || FFMPEG_ENC_AAC
 
-# Selected automatically when a decoder that needs IEEE float math is
+comment "FFmpeg audio filters to build (libavfilter)"
+	depends on IPC_MAJOR_4
+
+config FFMPEG_FILTER_AFFTDN
+	bool "FFT noise reduction (afftdn)"
+	depends on IPC_MAJOR_4
+	help
+	  FFT-based audio denoiser (libavfilter af_afftdn). Builds libavfilter
+	  with abuffer, abuffersink, aformat and afftdn, plus the avfilter-graph
+	  runtime path (ffmpeg-filter.c).
+
+config FFMPEG_FILTER_ALIMITER
+	bool "Lookahead peak limiter (alimiter)"
+	depends on IPC_MAJOR_4
+	help
+	  Lookahead brickwall peak limiter (libavfilter af_alimiter). Prevents
+	  digital clipping and provides speaker protection. Builds libavfilter
+	  with abuffer, abuffersink, aformat and alimiter, plus the avfilter-graph
+	  runtime path (ffmpeg-filter.c).
+
+# Turns on libavfilter in the FFmpeg cross-build when any filter is selected.
+config FFMPEG_BUILD_AVFILTER
+	bool
+	default y if FFMPEG_FILTER_AFFTDN || FFMPEG_FILTER_ALIMITER
+
+# Gates compilation of the runtime avfilter-graph path (ffmpeg-filter.c).
+config FFMPEG_FILTER
+	bool
+	default y if FFMPEG_BUILD_AVFILTER
+
+choice
+	prompt "Default FFmpeg audio filter"
+	depends on FFMPEG_FILTER
+	default FFMPEG_FILTER_DEFAULT_AFFTDN if FFMPEG_FILTER_AFFTDN
+	default FFMPEG_FILTER_DEFAULT_ALIMITER if FFMPEG_FILTER_ALIMITER
+
+config FFMPEG_FILTER_DEFAULT_AFFTDN
+	bool "FFT noise reduction (afftdn)"
+	depends on FFMPEG_FILTER_AFFTDN
+	help
+	  Select afftdn (FFT noise reduction) as the default libavfilter audio
+	  filter when no explicit filter profile or runtime configuration is
+	  supplied.
+
+config FFMPEG_FILTER_DEFAULT_ALIMITER
+	bool "Lookahead peak limiter (alimiter)"
+	depends on FFMPEG_FILTER_ALIMITER
+	help
+	  Select alimiter (lookahead brickwall peak limiter) as the default
+	  libavfilter audio filter when no explicit filter profile or runtime
+	  configuration is supplied.
+
+endchoice
+
+# Selected automatically when a decoder or filter that needs IEEE float math is
 # built. Gates compilation of the fast single-precision libm (fastmathf.c).
 config FFMPEG_DEC_FLOAT_MATH
 	bool
-	default y if FFMPEG_DEC_AAC || FFMPEG_DEC_OPUS || FFMPEG_DEC_MP3
+	default y if FFMPEG_DEC_AAC || FFMPEG_DEC_OPUS || FFMPEG_DEC_MP3 || FFMPEG_BUILD_AVFILTER
 
 config FFMPEG_DEC_COLD_SPLIT
 	bool "Place cold FFmpeg code (av_cold init/setup) in DRAM [EXPERIMENTAL]"

src/audio/ffmpeg_dec/ffmpeg-convert-hifi.c

@@ -0,0 +1,158 @@
+// SPDX-License-Identifier: BSD-3-Clause
+//
+// Copyright(c) 2026 Intel Corporation.
+//
+// Xtensa HiFi3 / HiFi4 / HiFi5 VFPU SIMD routines for FFmpeg PCM format
+// conversion and interleaving.
+//
+// Each helper processes the SIMD-aligned prefix of @nb_samples and returns the
+// number of samples per channel converted (0 when the target core or compiler
+// does not provide the required TIE intrinsics). The caller in
+// ffmpeg_dec-convert.c finishes any remaining tail samples in scalar C.
+
+#include <stdint.h>
+#include "ffmpeg_dec-convert.h"
+
+#if defined(__XTENSA__) && defined(__has_include)
+#if __has_include(<xtensa/config/core-isa.h>)
+#include <xtensa/config/core-isa.h>
+#endif
+
+#if defined(__has_builtin) && __has_builtin(__builtin_xtensa_trunc_sx2) && \
+	__has_builtin(__builtin_xtensa_float_sx2) && \
+	defined(XCHAL_HAVE_HIFI4_VFPU) && XCHAL_HAVE_HIFI4_VFPU && \
+	__has_include(<xtensahifiintrin.h>)
+#include <xtensahifiintrin.h>
+#define FFMPEG_DEC_CONV_VFPU 1
+#endif
+
+#if defined(XCHAL_HAVE_HIFI3) && XCHAL_HAVE_HIFI3 && __has_include(<xtensa/tie/xt_hifi3.h>)
+#include <xtensa/tie/xt_hifi3.h>
+#define FFMPEG_DEC_CONV_HIFI3 1
+#endif
+#endif
+
+int ffmpeg_hifi_fltp_to_s32(const float *const *src, int32_t *dst,
+			    int channels, int nb_samples)
+{
+#if defined(FFMPEG_DEC_CONV_VFPU)
+	int i = 0;
+
+	if (channels == 1) {
+		const float *c0 = src[0];
+
+		for (; i + 2 <= nb_samples; i += 2) {
+			int32_t tmp[2] __attribute__((aligned(8)));
+			ae_int32x2 f = AE_L32X2_I((const ae_int32x2 *)&c0[i], 0);
+			ae_int32x2 q = XT_TRUNC_SX2((ae_xtfloatx2)f, 31);
+
+			AE_S32X2_I(q, (ae_int32x2 *)tmp, 0);
+			dst[i + 0] = tmp[0];
+			dst[i + 1] = tmp[1];
+		}
+		return i;
+	}
+
+	if (channels == 2) {
+		const float *c0 = src[0];
+		const float *c1 = src[1];
+
+		for (; i + 2 <= nb_samples; i += 2) {
+			ae_int32x2 f0 = AE_L32X2_I((const ae_int32x2 *)&c0[i], 0);
+			ae_int32x2 f1 = AE_L32X2_I((const ae_int32x2 *)&c1[i], 0);
+			ae_int32x2 q0 = XT_TRUNC_SX2((ae_xtfloatx2)f0, 31);
+			ae_int32x2 q1 = XT_TRUNC_SX2((ae_xtfloatx2)f1, 31);
+			int32_t tmp0[2] __attribute__((aligned(8)));
+			int32_t tmp1[2] __attribute__((aligned(8)));
+
+			AE_S32X2_I(q0, tmp0, 0);
+			AE_S32X2_I(q1, tmp1, 0);
+
+			dst[2 * i + 0] = tmp0[0];
+			dst[2 * i + 1] = tmp1[0];
+			dst[2 * i + 2] = tmp0[1];
+			dst[2 * i + 3] = tmp1[1];
+		}
+		return i;
+	}
+
+	for (int c = 0; c < channels; c++) {
+		const float *pl = src[c];
+		int32_t *dc = dst + c;
+
+		for (i = 0; i + 2 <= nb_samples; i += 2) {
+			int32_t tmp[2] __attribute__((aligned(8)));
+			ae_int32x2 f = AE_L32X2_I((const ae_int32x2 *)&pl[i], 0);
+			ae_int32x2 q = XT_TRUNC_SX2((ae_xtfloatx2)f, 31);
+
+			AE_S32X2_I(q, (ae_int32x2 *)tmp, 0);
+			dc[i * channels] = tmp[0];
+			dc[(i + 1) * channels] = tmp[1];
+		}
+	}
+	return nb_samples & ~1;
+#else
+	(void)src;
+	(void)dst;
+	(void)channels;
+	(void)nb_samples;
+	return 0;
+#endif
+}
+
+int ffmpeg_hifi_s32_to_fltp(const int32_t *src, float *const *dst,
+			    int channels, int nb_samples)
+{
+#if defined(FFMPEG_DEC_CONV_VFPU)
+	int i = 0;
+
+	for (int c = 0; c < channels; c++) {
+		float *pl = dst[c];
+		const int32_t *sc = src + c;
+
+		for (i = 0; i + 2 <= nb_samples; i += 2) {
+			int32_t tmp[2] __attribute__((aligned(8)));
+			ae_int32x2 p;
+			ae_xtfloatx2 r;
+
+			tmp[0] = sc[i * channels];
+			tmp[1] = sc[(i + 1) * channels];
+			p = AE_L32X2_I((const ae_int32x2 *)tmp, 0);
+			r = XT_FLOAT_SX2(p, 31);
+			AE_S32X2_I((ae_int32x2)r, (ae_int32x2 *)&pl[i], 0);
+		}
+	}
+	return nb_samples & ~1;
+#else
+	(void)src;
+	(void)dst;
+	(void)channels;
+	(void)nb_samples;
+	return 0;
+#endif
+}
+
+int ffmpeg_hifi_s32p_to_s32_stereo(const int32_t *c0, const int32_t *c1,
+				   int32_t *dst, int nb_samples)
+{
+#if defined(FFMPEG_DEC_CONV_HIFI3)
+	int i = 0;
+
+	for (; i + 2 <= nb_samples; i += 2) {
+		ae_int32x2 v0 = AE_L32X2_I((const ae_int32x2 *)&c0[i], 0);
+		ae_int32x2 v1 = AE_L32X2_I((const ae_int32x2 *)&c1[i], 0);
+		ae_int32x2 pair0 = AE_SEL32_LL(v0, v1);
+		ae_int32x2 pair1 = AE_SEL32_HH(v0, v1);
+
+		AE_S32X2_I(pair0, (ae_int32x2 *)&dst[2 * i + 0], 0);
+		AE_S32X2_I(pair1, (ae_int32x2 *)&dst[2 * i + 2], 0);
+	}
+	return i;
+#else
+	(void)c0;
+	(void)c1;
+	(void)dst;
+	(void)nb_samples;
+	return 0;
+#endif
+}