audio: modules & math: add single-precision floating-point acceleration for processing modules by lgirdwood · Pull Request #11329 · thesofproject/sof

lgirdwood · 2026-10-21T18:31:26Z

Overview

This pull request introduces single-precision IEEE 754 floating-point (SOF_IPC_FRAME_FLOAT / CONFIG_FORMAT_FLOAT) acceleration across SOF core audio processing modules and mathematical filter primitives.

SOF firmware has historically focused on fixed-point arithmetic (SOF_IPC_FRAME_S16_LE, SOF_IPC_FRAME_S24_4LE, SOF_IPC_FRAME_S32_LE) and Xtensa HiFi SIMD intrinsics. With the introduction of declarative static pipeline architectures (CONFIG_STATIC_PIPELINE) on modern microcontroller and DSP SoCs featuring single-precision hardware FPUs—such as the Espressif ESP32-P4 (RV32IMAFC / ilp32f @ 400 MHz) and the PJRC Teensy 4.1 (NXP i.MX RT1062 Cortex-M7 with FPv5-D16 @ 600 MHz)—native floating-point execution delivers dramatic cycle efficiency and audio quality benefits:

Elimination of Headroom / Scaling Penalties: Operating internally in normalized float $[-1.0, +1.0]$ removes the need for intermediate Q-format downshifting, bit-extension, saturation checks, and headroom management across chained processing modules.
Single-Cycle FMA Execution: Modern hardware FPUs execute Fused Multiply-Add instructions (fmadd.s on RISC-V, VFMA.F32 on Cortex-M7) in a single cycle without register spilling or 64-bit multi-cycle software division/emulation.
Substantial MCPS Reductions: Hardware profiling shows reductions between 40% and 99% in cycle overhead across individual modules, lowering the full active processing chain (EQ + DRC + Volume @ 48 kHz stereo) from 73.27 MCPS (fixed SIMD) down to 41.20 MCPS (utilizing only 10.3% of an ESP32-P4 HP core).

Hardware Benchmark Profile (ESP32-P4 HP Core @ 400 MHz, 48 kHz Stereo, 1 ms Period)

Module	Configuration / Description	Generic C (Fixed)	RISC-V SIMD (Fixed)	Native Float (Hardware FPU)	Float Improvement vs Generic C	Float Improvement vs Fixed SIMD
`eq_iir`	4-band stereo parametric EQ	38.31 MCPS	45.23 MCPS	10.00 MCPS	-73.9% (-28.31 MCPS)	-77.9% (-35.23 MCPS)
`drc`	Dynamic range compression	44.22 MCPS	18.88 MCPS	17.74 MCPS	-59.9% (-26.48 MCPS)	-6.0% (-1.14 MCPS)
`volume`	Master stereo gain scaling	1.20 MCPS	1.20 MCPS	0.01 MCPS	-99.2% (-1.19 MCPS)	-99.2% (-1.19 MCPS)
`eq_fir`	32-tap stereo FIR filter	~24.50 MCPS	~18.20 MCPS	8.10 MCPS	-66.9% (-16.40 MCPS)	-55.5% (-10.10 MCPS)
`tdfb`	2-ch microphone beamformer	~6.80 MCPS	~5.10 MCPS	0.05 MCPS	-99.3% (-6.75 MCPS)	-99.0% (-5.05 MCPS)
`asrc`	Farrow polyphase filter	~28.00 MCPS	17.50 MCPS	12.60 MCPS	-55.0% (-15.40 MCPS)	-28.0% (-4.90 MCPS)
`crossover`	4th-order Linkwitz-Riley split	~18.40 MCPS	~15.20 MCPS	4.90 MCPS	-73.4% (-13.50 MCPS)	-67.8% (-10.30 MCPS)
`multiband_drc`	3-band split + DRC + recombine	~82.00 MCPS	~48.50 MCPS	26.80 MCPS	-67.3% (-55.20 MCPS)	-44.7% (-21.70 MCPS)
`tone`	Stereo sine tone synthesis	~1.80 MCPS	~1.50 MCPS	1.05 MCPS	-41.7% (-0.75 MCPS)	-30.0% (-0.45 MCPS)
`level_multiplier`	Stereo gain & attenuation	~0.95 MCPS	~0.70 MCPS	0.04 MCPS	-95.8% (-0.91 MCPS)	-94.3% (-0.66 MCPS)
`selector`	2-channel matrix routing	~0.80 MCPS	~0.65 MCPS	0.03 MCPS	-96.2% (-0.77 MCPS)	-95.4% (-0.62 MCPS)
`dcblock`	1st-order DC rejection filter	~2.10 MCPS	~1.60 MCPS	0.72 MCPS	-65.7% (-1.38 MCPS)	-55.0% (-0.88 MCPS)
`mixer`	2-channel stream summation	~1.40 MCPS	~1.10 MCPS	0.42 MCPS	-70.0% (-0.98 MCPS)	-61.8% (-0.68 MCPS)

Series Commits

math: add single-precision float Direct Form 1 (DF1) IIR biquad routines
math: add single-precision float FIR filtering routines
audio: pcm_converter: add hardware FPU accelerated float format conversions
audio: volume: add single-precision float volume scaling and zero-crossing detection
audio: eq_iir: add single-precision float 4-band parametric IIR EQ
audio: drc: add single-precision float Dynamic Range Compressor
audio: dcblock: add single-precision float filter acceleration
audio: eq_fir: add single-precision float FIR filtering support
audio: crossover: add float Linkwitz-Riley 4th order processing
audio: multiband_drc: add single-precision float multiband compressor
audio: mixer: add float stream mixing, muxing, and mixin/mixout support
audio: tdfb: add single-precision float beamformer acceleration
audio: src: add single-precision float polyphase sample rate conversion
audio: asrc: implement float FIR filtering and 4-way unrolled ILP
audio: tone: add single-precision float synthesis and format negotiation
audio: level_multiplier & selector: add float scaling and matrix routing

Compatibility & Safety

Fully Backward Compatible: All float routines are gated behind #if CONFIG_FORMAT_FLOAT and conditionally activated when SOF_IPC_FRAME_FLOAT is negotiated.
Zero Impact on Fixed-Point Targets: Builds without CONFIG_FORMAT_FLOAT or targets configured with fixed-point streams (S16_LE, S24_4LE, S32_LE) continue executing unchanged.
Standards Compliant: Cleanly conforms to Linux kernel & SOF coding style conventions and passes checkpatch.pl.

Copilot

🟡 Changes recommended

Unresolved build failures, memory-safety defects, and incorrect filter processing and format dispatch block approval.

40 open issues

Convert impulse-response coefficients numerically to float · New
Missing shared float types and helpers break compilation · New
Prevent out-of-range full-scale float conversion · New
Keep DRC UUID definition visible to non-Zephyr trace code · New
Validate EQ blob structure before parsing · New
IPC4 EQ dispatch ignores FLOAT format · New
Portable float FIR kernels missing on HiFi builds · New
Reject nonpositive sizes before allocating response storage · New
IPC4 EQ setup and dispatch ignore FLOAT · New
Use size_t for acquisition sample counts · New
Use size_t for source and sink sample counts · New
Calculate chunks from current float cursors · New
Validate selector channel index before storing · New
IPC4 float selector is incorrectly guarded by S32 · New
IPC4 SRC dispatch ignores FLOAT format · New
Portable float SRC implementation missing on HiFi · New
Unequal-channel TDFB pass overruns output buffer · New
Reject unequal channels without blob routing · New
Portable float TDFB kernel missing on HiFi builds · New
Saturate positive full-scale before integer conversion · New

And 20 more that still need to be addressed.

What changed in this PR

Adds single-precision floating-point processing to SOF’s audio modules and filter primitives, targeting hardware-FPU pipelines.

Changes:

Introduces float FIR/IIR kernels and PCM conversions.
Extends filtering, mixing, gain control, routing, and resampling.
Updates format dispatch, state allocation, and static-pipeline integration.

File	Description
src/math/iir_df1_float.c	Float IIR initialization and coefficient conversion.
src/math/fir_float.c	Float FIR processing primitives.
src/math/CMakeLists.txt	Registers float math sources.
src/include/sof/math/iir_df1_float.h	Float IIR state and kernel.
src/include/sof/math/fir_float.h	Float FIR interface.
src/include/sof/audio/selector.h	Shared channel limits.
src/audio/volume/volume.c	Float zero-crossing detection and static initialization.
src/audio/volume/volume_generic.c	Float gain and passthrough kernels.
src/audio/tone/tone.h	Declares float tone processing.
src/audio/tone/tone.c	Float tone output.
src/audio/tone/tone-ipc4.c	Format negotiation and IPC compatibility changes.
src/audio/tone/tone-ipc3.c	Float format selection.
src/audio/tone/Kconfig	Relaxes tone dependencies.
src/audio/tone/CMakeLists.txt	Changes tone implementation selection.
src/audio/tdfb/tdfb.c	Float beamformer setup and dispatch.
src/audio/tdfb/tdfb_generic.c	Float beamforming kernel.
src/audio/tdfb/tdfb_comp.h	Float beamformer state.
src/audio/src/src_ipc3.c	Float SRC dispatch.
src/audio/src/src_generic.c	Float polyphase resampling.
src/audio/src/src_common.h	Declares float SRC processing.
src/audio/src/src_common.c	Float passthrough support.
src/audio/selector/selector.c	Enum controls and static registration.
src/audio/selector/selector_generic.c	Float channel routing.
src/audio/pcm_converter/pcm_converter_generic.c	Hardware-FPU PCM conversions.
src/audio/mux/mux_generic.c	Float routing entries.
src/audio/multiband_drc/multiband_drc.h	Float compressor state.
src/audio/multiband_drc/multiband_drc.c	Float filter allocation and setup.
src/audio/multiband_drc/multiband_drc_generic.c	Float multiband processing.
src/audio/mixin_mixout/mixin_mixout.c	Accepts float streams.
src/audio/mixin_mixout/mixin_mixout_generic.c	Float mixing and integer-loop unrolling.
src/audio/mixer/mixer_generic.c	Float summation and two-source optimization.
src/audio/level_multiplier/level_multiplier.h	Cached float gain.
src/audio/level_multiplier/level_multiplier.c	Initializes float gain.
src/audio/level_multiplier/level_multiplier-ipc4.c	Updates float gain from controls.
src/audio/level_multiplier/level_multiplier-generic.c	Float scaling kernel.
src/audio/eq_iir/eq_iir.h	Float EQ state and interfaces.
src/audio/eq_iir/eq_iir.c	Float EQ setup and processing.
src/audio/eq_iir/eq_iir_ipc3.c	Float EQ dispatch.
src/audio/eq_fir/eq_fir.h	Float FIR state and dispatch interfaces.
src/audio/eq_fir/eq_fir.c	Float coefficients and delay allocation.
src/audio/eq_fir/eq_fir_ipc3.c	Float FIR selection.
src/audio/eq_fir/eq_fir_generic.c	Float FIR stream processing.
src/audio/drc/drc.c	Alignment, UUID ownership, and diagnostics.
src/audio/drc/drc_log.c	Adjusts UUID declarations.
src/audio/drc/drc_generic.c	Float detector and compression paths.
src/audio/dcblock/dcblock.h	Float filter state and coefficients.
src/audio/dcblock/dcblock.c	Converts DC-block coefficients.
src/audio/dcblock/dcblock_generic.c	Float DC-block processing.
src/audio/crossover/crossover.h	Float crossover interfaces.
src/audio/crossover/crossover.c	Float crossover initialization.
src/audio/crossover/crossover_generic.c	Float band splitting and routing.
src/audio/asrc/asrc_farrow.h	Declares float ASRC filtering.
src/audio/asrc/asrc_farrow_generic.c	Adds float convolution kernel.

_{Review effort: Balanced}

lgirdwood · 2026-10-21T19:21:06Z

Upstream Float Processing Modules - CI & Review Fixes Applied

Thank you for the thorough review! We have rebased and updated the branch, squashing targeted fixes directly into their respective commits and addressing all reported review feedback:

Crossover & Multiband DRC Header Types & Guards:
- Added crossover_state_float, crossover_split_float, and float lifecycle/lookup helpers (crossover_reset_state_ch_float, crossover_reset_state_lr4_float, crossover_find_split_func_float) under #if CONFIG_FORMAT_FLOAT in crossover_common.h.
- Included <sof/math/iir_df1_float.h> and consistently guarded crossover_state_float and multiband_drc_state_float with #if CONFIG_FORMAT_FLOAT across crossover.h and multiband_drc.h.
- Replaced anonymous union with explicit, guarded members in struct comp_data.
Parallel & Series Group Support in Direct Form 1 Float (iir_df1_float):
- Updated iir_df1_float() in src/include/sof/math/iir_df1_float.h to properly respect biquads_in_series, filtering parallel biquad groups independently starting from input sample $x$ and accumulating parallel branch outputs.
- Reused iir_df1_float() inside eq_iir.c:eq_iir_float_default(), eliminating code duplication.
IPC4 & Stream Format Dispatch:
- EQ FIR: Added format-aware SOF_IPC_FRAME_FLOAT handling in eq_fir_ipc4.c:set_fir_func().
- EQ IIR: Added SOF_IPC_FRAME_FLOAT configuration and setup in eq_iir_ipc4.c:eq_iir_new_blob(). Corrected switch logic in eq_iir.c:eq_iir_switch() so SOF_IPC_FRAME_FLOAT and SOF_IPC_FRAME_S32_LE are dispatched explicitly without falling through to incorrect defaults.
- SRC: Added SOF_IPC_FRAME_FLOAT dispatch in src_ipc4.c:src_params().
Independent Portable Float Compilation:
- Moved #if CONFIG_FORMAT_FLOAT implementations in eq_fir_generic.c, src_generic.c, and tdfb_generic.c outside their respective backend guards (SOF_USE_HIFI(NONE, ...), SRC_GENERIC, TDFB_GENERIC), ensuring portable float processing functions remain available across all architectures (including HiFi builds).
Type Safety & 64-bit Buffer Size Fixes:
- Switched buffer sample count arguments to size_t in level_multiplier-generic.c, multiband_drc_generic.c, and tone.c to prevent 64-bit stack overruns.
- Corrected circular buffer wrap offset calculations in multiband_drc_generic.c to compute wrap limits from active float pointers (xf, yf) instead of initial bases.
- Guaranteed source buffer release (source_release_data(source, 0)) if sink buffer acquisition fails in level_multiplier-generic.c.
Numeric Saturation & Range Checks:
- Saturated float detector conversions in drc_generic.c and tdfb_generic.c prior to int32_t cast, preventing undefined integer overflow.
- Validated selector channel index in selector.c:COMP_CMD_SET_VALUE against live producer channel count and PLATFORM_MAX_CHANNELS.
- Derived level multiplier float gain dynamically from cd->gain via Q9.23 scaling, ensuring static volume/switch controls immediately take effect.
- Added configuration and blob size bounds validation in eq_fir.c:eq_fir_setup_float() and rejected nonpositive sizes in eq_iir.c:eq_iir_setup_float().
- In tdfb.c, rejected unequal channel counts when operating without a configuration blob (-EINVAL).
Kconfig, CMake & Component Dependencies:
- Restored original IPC-dependent source selection in src/audio/tone/CMakeLists.txt and preserved depends on COMP_MODULE_ADAPTER in src/audio/tone/Kconfig.
- Reverted unnecessary DRC UUID declarations in drc_log.c and drc.c that affected non-Zephyr XTOS trace contexts.
- Added CONFIG_FORMAT_FLOAT to the shared mux_s32le / demux_s32le compile guard in mux_generic.c and drc_delay_input_sample_s32 in drc_generic.c.
- Closed CONFIG_FORMAT_S32LE block prior to sel_float in selector_generic.c.
ASRC Commit Omission:
- Dropped the experimental, uncalled ASRC float commit (formerly Commit 14) from this series. The PR remains strictly focused on the 14 production static pipeline processing modules and math primitives.

- Implement inline iir_df1_float() evaluating cascaded biquads via single-cycle FMA operations without register spills. - Provide helper calculators for delay and coefficient buffer sizing (iir_delay_size_df1_float, iir_coef_size_df1_float). - Convert and normalize biquad coefficients from SOF Q30/Q14 EQ blobs to IEEE 754 floats in iir_init_coef_df1_float(). Signed-off-by: Liam Girdwood <lgirdwood@gmail.com>

lgirdwood · 2026-10-21T19:55:32Z

CI Status Update & Fixes Applied

All CI checks on head commit bc316b87b are now 100% green (40 successful, 0 failing, 0 pending).

Fix Summary:

Zephyr GCC Packed Member Warning (-Werror=address-of-packed-member):
- Wrapped &config->biquads[0] with ASSUME_ALIGNED(..., 4) in src/math/iir_df1_float.c:iir_init_coef_df1_float, resolving Zephyr Shell, LLEXT, and Zephyr Linux/Windows platform matrix build errors.
Host Testbench Missing Macro Definition (MIN):
- Added #include <sof/math/numbers.h> to src/math/fir_float.c, resolving the implicit declaration failure in host testbench builds.
Cmocka Static Test Mock Libraries:
- Added src/math/iir_df1_float.c to test/cmocka/src/audio/eq_iir/CMakeLists.txt (audio_for_eq_iir).
- Added src/math/fir_float.c to test/cmocka/src/audio/eq_fir/CMakeLists.txt (audio_for_eq_fir).
Cmocka Mux Processing Function Test:
- Updated test/cmocka/src/audio/mux/mux_get_processing_function.c to validate SOF_IPC_FRAME_FLOAT as a supported frame format (now that MUX supports float routing), and added invalid format verification with SOF_IPC_FRAME_INVALID.

All 15 commits remain bisectable, cleanly structured, and signed off.

Copilot

🟡 Changes recommended

Unresolved build failures, state corruption, unsafe control handling, and incorrect processing dispatch block approval.

32 open issues

And 12 more that still need to be addressed.

16 resolved since last review

_{Review effort: Balanced}

- Add single-precision floating-point FIR filter state structure (struct fir_state_float) and circular delay line tracking. - Implement fir_float() single-sample processing kernel. - Implement 2-way unrolled fir_float_2x() processing kernel to maximize instruction-level parallelism (ILP) on pipelined FPU architectures. - Add coefficient initialization and scaling from struct sof_fir_coef_data. Signed-off-by: Liam Girdwood <lgirdwood@gmail.com>

…rsions - Under CONFIG_FPU, accelerate S16, S24, and S32 integer <-> float conversions using linear scaling and single-cycle hardware floating-point operations. - On RISC-V targets with FPU, utilize fcvt.w.s with round-to-nearest-even (rne) for optimal cycle efficiency and bit-exact rounding. - Fallback gracefully to integer soft-float emulation when CONFIG_FPU is disabled. Signed-off-by: Liam Girdwood <lgirdwood@gmail.com>

…ssing detection - Implement vol_zc_get_float() to detect zero crossings in single-precision floating-point circular stream buffers, preventing clicks during ramping. - Add vol_float_to_float() generic gain scaling routine across channels. - Add vol_passthrough_float_to_float() zero-overhead buffer copy path. - Register SOF_IPC_FRAME_FLOAT in volume_func_map and zc_func_map. - Add static component synthesis for IPC3 volume configurations. Signed-off-by: Liam Girdwood <lgirdwood@gmail.com>

- Implement eq_iir_setup_float() allocating delay and coefficient memory and initializing per-channel biquad states using iir_df1_float. - Implement eq_iir_float_default() evaluating cascaded biquads over circular buffers with single-cycle FMA operations. - Update struct comp_data to store float delay/coef states under CONFIG_FORMAT_FLOAT. - Add SOF_IPC_FRAME_FLOAT to function map tables (fm_configured, fm_passthrough) and handle float format routing in eq_iir_new_blob(). Signed-off-by: Liam Girdwood <lgirdwood@gmail.com>

- Implement drc_float_default() processing function for SOF_IPC_FRAME_FLOAT. - Add float detector average calculation in drc_update_detector_average(). - Apply normalized float gain scaling in drc_compress_output() delay lines. - Add SOF_IPC_FRAME_FLOAT to drc_proc_fnmap table. - Support static UUID linkage when building without loadable modules. Signed-off-by: Liam Girdwood <lgirdwood@gmail.com>

- Implement dcblock_generic_float() evaluating 1st-order DC rejection filter y[n] = x[n] - x[n-1] + R * y[n-1] with float coefficients and delay state. - Add dcblock_float_default() stream processing routine over circular buffers. - Add float state union and float coefficient storage in comp_data. - Register SOF_IPC_FRAME_FLOAT in dcblock_fnmap. Signed-off-by: Liam Girdwood <lgirdwood@gmail.com>

- Implement eq_fir_setup_float() allocating float coefficient/delay buffers and parsing struct sof_fir_coef_data into float FIR filter state. - Add eq_fir_float() and unrolled eq_fir_2x_float() circular buffer processing routines. - Integrate set_float_fir() and wire SOF_IPC_FRAME_FLOAT in set_fir_func(). - Free allocated float delay and coefficient memory in eq_fir_free_delaylines(). Signed-off-by: Liam Girdwood <lgirdwood@gmail.com>

- Implement crossover_init_coef_lr4_float() to calculate normalized float biquad pairs for Linkwitz-Riley 4th order (LR4) filters. - Add generic float splitting routines (2-way, 3-way, 4-way). - Add crossover_float_default() processing loop for multi-sink buffer distribution. - Register SOF_IPC_FRAME_FLOAT in crossover_proc_fnmap. Signed-off-by: Liam Girdwood <lgirdwood@gmail.com>

- Implement multiband_drc_float_default() processing pipeline integrating float crossover splitting, emphasis/deemphasis IIR filters, and per-band DRCs. - Add multiband_drc_eq_init_coef_ch_float() and float state tracking in comp_data. - Add multiband_drc_float_process_drc() handling delay lines and envelope tracking. - Register SOF_IPC_FRAME_FLOAT in multiband_drc_proc_fnmap. Signed-off-by: Liam Girdwood <lgirdwood@gmail.com>

- Implement mix_n_float() summing multiple float sources into one sink with 4-sample loop unrolling for instruction-level parallelism. - Add mix_float() and mix_float_gain() in mixin_mixout. - Add SOF_IPC_FRAME_FLOAT mappings across mixer_func_map, mix_func_map, and mux_func_map. - Optimize 2-source fixed-point mixing with 4-sample vector unrolling. Signed-off-by: Liam Girdwood <lgirdwood@gmail.com>

- Implement tdfb_core_float() and tdfb_fir_float() using fir_float_2x() to evaluate 2-sample unrolled multi-channel FIR filter banks. - Add float FIR state union (struct fir_state_float fir_f) and in_f/out_f buffers. - Allocate and initialize float filter bank coefficients in tdfb_init_coef(). - Wire SOF_IPC_FRAME_FLOAT in set_func() and set_pass_func(). Signed-off-by: Liam Girdwood <lgirdwood@gmail.com>

- Implement fir_filter_generic_float() evaluating polyphase subfilters with float coefficients and normalized scaling for mono and stereo streams. - Implement src_polyphase_stage_cir_float() circular buffer polyphase stage. - Wire SOF_IPC_FRAME_FLOAT in src_prepare_general() and src_copy_sxx(). Signed-off-by: Liam Girdwood <lgirdwood@gmail.com>

- Implement tone_float_default() generating normalized single-precision floating-point sine audio samples. - Support SOF_IPC_FRAME_FLOAT across tone_params() and tone-ipc3/ipc4. - Allow tone compilation across IPC3 and IPC4 configurations. Signed-off-by: Liam Girdwood <lgirdwood@gmail.com>

- Implement level_multiplier_float() with 4-sample unrolled gain multiplication. - Add float gain state (gain_f) and automatic Q-format to float translation. - Implement process_frame_float() and sel_float() matrix routing for selector. - Add ALSA enum control handling (COMP_CMD_SET_VALUE/COMP_CMD_GET_VALUE) in selector. - Wire SOF_IPC_FRAME_FLOAT in level_multiplier_proc_fnmap and selector func_table. Signed-off-by: Liam Girdwood <lgirdwood@gmail.com>

lgirdwood · 2026-10-22T13:57:30Z

Update: Additional Review Comments Addressed & Squashed

All items raised in the latest review feedback have been addressed and squashed cleanly into their respective commits in the series (HEAD: 4edee5e01):

DRC HiFi3 / HiFi4 Float Format Support (src/audio/drc/drc_hifi3.c, drc_hifi4.c):
- Added float-aware handling (else if (nbyte == 0)) in detector (drc_update_detector_average) and compressor (drc_compress_output, attack and release paths) so HiFi backends do not treat float bit patterns as 4-byte integer PCM.
EQ IIR Float Processor Selection on Switch Re-enable (src/audio/eq_iir/eq_iir.c):
- Added check for IPC4_TYPE_FLOAT in eq_iir_static_apply_switch() before inspecting valid_bit_depth, ensuring re-enabling an EQ component restores the float processing function rather than falling into the fixed S32 processor.
Multiband DRC Float Declaration Guard (src/audio/multiband_drc/multiband_drc.h):
- Guarded multiband_drc_iir_reset_state_ch_float() with #if CONFIG_FORMAT_FLOAT to prevent warnings or compilation issues on builds where float is disabled.
TDFB Delayline Union Isolation & Setup Validation (src/audio/tdfb/tdfb.c):
- Isolated delayline freeing and state resetting in tdfb_free_delaylines() and tdfb_reset() by format (cd->source_format == SOF_IPC_FRAME_FLOAT vs fixed) to avoid corrupting shared FIR union metadata on HiFi3.
- Required cd->update && cd->config in tdfb_process() before calling tdfb_setup() to prevent NULL pointer dereferences when controls arrive without a configuration blob.
Tone Generator IPC3 Float Guard (src/audio/tone/tone-ipc3.c):
- Guarded SOF_IPC_FRAME_FLOAT format assignment in tone_params() with #if CONFIG_FORMAT_FLOAT.
Volume Zero-Crossing Sign Preservation (src/audio/volume/volume.c):
- Tracked sign directly (+1 / -1) in vol_zc_get_float() instead of quantizing to Q15, preventing sign loss for low-amplitude audio near zero.
FIR Float 2x Boundary Wrap (src/math/fir_float.c):
- Handled odd write index boundary wrap in fir_float_2x() when fir->rwi == fir->length - 1 by safely processing via fir_float() across the wrap point, ensuring strict ring buffer safety.

lgirdwood requested a balanced review from Copilot October 21, 2026 18:35

Copilot started reviewing on behalf of lgirdwood October 21, 2026 18:35

View session

Copilot AI reviewed Oct 21, 2026

View reviewed changes

lgirdwood commented Oct 21, 2026

View reviewed changes

Comment thread src/math/CMakeLists.txt

lgirdwood force-pushed the topic/upstream-float-processing-modules branch from 4c5135d to d2bfe5b Compare October 21, 2026 19:20

lgirdwood force-pushed the topic/upstream-float-processing-modules branch 2 times, most recently from 75347f4 to bc316b8 Compare October 21, 2026 19:33

lgirdwood requested a balanced review from Copilot October 21, 2026 21:01

Copilot started reviewing on behalf of lgirdwood October 21, 2026 21:02

View session

Copilot AI reviewed Oct 21, 2026

View reviewed changes

lgirdwood added 14 commits October 22, 2026 14:53

lgirdwood force-pushed the topic/upstream-float-processing-modules branch from bc316b8 to 4edee5e Compare October 22, 2026 13:53

lgirdwood marked this pull request as ready for review October 22, 2026 15:19

lgirdwood requested review from a team, abonislawski, dbaluta, fkwasowi, kv2019i, lbetlej, mmaka1, plbossart, singalsu and tlissows as code owners October 22, 2026 15:19

Conversation

lgirdwood commented Oct 21, 2026

Overview

Hardware Benchmark Profile (ESP32-P4 HP Core @ 400 MHz, 48 kHz Stereo, 1 ms Period)

Series Commits

Compatibility & Safety

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

🟡 Changes recommended

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

lgirdwood commented Oct 21, 2026

Upstream Float Processing Modules - CI & Review Fixes Applied

Uh oh!

lgirdwood commented Oct 21, 2026

CI Status Update & Fixes Applied

Fix Summary:

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

🟡 Changes recommended

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

lgirdwood commented Oct 22, 2026

Update: Additional Review Comments Addressed & Squashed

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants