tap · tap · Jul 21, 2026 · Jul 20, 2026 · Jul 21, 2026 · Jul 21, 2026
diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
@@ -36,6 +36,8 @@ jobs:
             werror: OFF
     steps:
       - uses: actions/checkout@df4cb1c069e1874edd31b4311f1884172cec0e10 # v6
+        with:
+          submodules: recursive
 
       - name: Configure
         env:
@@ -57,6 +59,8 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - uses: actions/checkout@df4cb1c069e1874edd31b4311f1884172cec0e10 # v6
+        with:
+          submodules: recursive
 
       - name: Configure
         env:
@@ -89,6 +93,8 @@ jobs:
     timeout-minutes: 45
     steps:
       - uses: actions/checkout@df4cb1c069e1874edd31b4311f1884172cec0e10 # v6
+        with:
+          submodules: recursive
 
       - name: Install toolchain and QEMU
         run: >
@@ -144,6 +150,8 @@ jobs:
       TOOLCHAIN: clang+llvm-19.1.5-cross-hexagon-unknown-linux-musl
     steps:
       - uses: actions/checkout@df4cb1c069e1874edd31b4311f1884172cec0e10 # v6
+        with:
+          submodules: recursive
 
       - name: Install QEMU
         run: >
@@ -194,6 +202,8 @@ jobs:
     timeout-minutes: 15
     steps:
       - uses: actions/checkout@df4cb1c069e1874edd31b4311f1884172cec0e10 # v6
+        with:
+          submodules: recursive
 
       - name: Configure
         run: >
@@ -219,14 +229,16 @@ jobs:
     timeout-minutes: 10
     steps:
       - uses: actions/checkout@df4cb1c069e1874edd31b4311f1884172cec0e10 # v6
+        with:
+          submodules: recursive
       - name: Compile both forms and diff the fingerprint
         run: |
           set -e
           # Ooura is C: compile with gcc so g++ does not name-mangle rdft/rdft_f.
-          gcc -O2 -c third_party/ooura/fftsg.c -o /tmp/fftsg.o
-          gcc -O2 -c third_party/ooura/fftsg_float.c -o /tmp/fftsg_float.o
+          gcc -O2 -c submodules/dsptap/third_party/ooura/fftsg.c -o /tmp/fftsg.o
+          gcc -O2 -c submodules/dsptap/third_party/ooura/fftsg_float.c -o /tmp/fftsg_float.o
           for bl in 0 1; do
-            g++ -std=c++20 -O2 -DMUTAP_SUPPRESSOR_BRANCHLESS=$bl -Iinclude \
+            g++ -std=c++20 -O2 -DMUTAP_SUPPRESSOR_BRANCHLESS=$bl -Iinclude -Isubmodules/dsptap/include \
               tests/branchless_parity_check.cpp /tmp/fftsg.o /tmp/fftsg_float.o -o /tmp/parity_$bl
             /tmp/parity_$bl | tee /tmp/out_$bl
           done
@@ -260,6 +272,8 @@ jobs:
       TOOLCHAIN: clang+llvm-19.1.5-cross-hexagon-unknown-linux-musl
     steps:
       - uses: actions/checkout@df4cb1c069e1874edd31b4311f1884172cec0e10 # v6
+        with:
+          submodules: recursive
 
       - name: Install toolchains and QEMU
         run: >
@@ -369,6 +383,8 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - uses: actions/checkout@df4cb1c069e1874edd31b4311f1884172cec0e10 # v6
+        with:
+          submodules: recursive
       # Run the TapHouse pre-commit hook — the exact same pinned clang-format
       # developers get from `pre-commit install` (.pre-commit-config.yaml). CI
       # and local therefore share one clang-format version by construction, and

diff --git a/.github/workflows/style.yml b/.github/workflows/style.yml
@@ -7,15 +7,17 @@ on: [push, pull_request]
 
 jobs:
   drift:
-    uses: tap/taphouse/.github/workflows/drift-check.yml@v3
+    uses: tap/taphouse/.github/workflows/drift-check.yml@v4
     with:
-      ref: v3
+      ref: v4
 
   clang-tidy:
     runs-on: ubuntu-latest
     timeout-minutes: 30
     steps:
       - uses: actions/checkout@v4
+        with:
+          submodules: recursive
       - name: Install tools
         run: sudo apt-get update && sudo apt-get install -y clang-tidy-18 cmake python3
       - name: Configure (compile database)

diff --git a/.gitmodules b/.gitmodules
@@ -0,0 +1,4 @@
+[submodule "submodules/dsptap"]
+	path = submodules/dsptap
+	url = https://github.com/tap/dsptap
+	branch = main
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -2,16 +2,15 @@ cmake_minimum_required(VERSION 3.24)
 project(MuTap VERSION 0.1.0 LANGUAGES C CXX)
 
 # ------------------------------------------------------------------------------
-# Ooura FFT (vendored) — provides the extern "C" rdft/rdft_f used by
-# include/mutap/fft.h. PIC so MODULE/SHARED consumers (host applications,
-# Max externals) can link it.
+# Shared real FFT (DspTap) — the Ooura wrapper with optional CMSIS-DSP Helium /
+# Apple vDSP float32 backends, formerly vendored here (include/mutap/fft.h +
+# third_party/{ooura,cmsis-dsp}). Now consumed from the tap/dsptap submodule as
+# tap::dsp, which carries the same numeric contract and selects the per-target
+# backend itself (CMSIS on the bare-metal M55, vDSP on Apple, Ooura elsewhere).
+# Override with -DTAP_DSP_FFT_CMSIS=OFF / -DTAP_DSP_FFT_ACCELERATE=OFF.
 # ------------------------------------------------------------------------------
 
-add_library(mutap_fft STATIC third_party/ooura/fftsg.c third_party/ooura/fftsg_float.c)
-add_library(MuTap::fft ALIAS mutap_fft)
-set_target_properties(mutap_fft PROPERTIES
-    POSITION_INDEPENDENT_CODE ON
-    OUTPUT_NAME mutap_fft)
+add_subdirectory(submodules/dsptap)
 
 # ------------------------------------------------------------------------------
 # The library proper (header-only)
@@ -23,80 +22,7 @@ target_include_directories(mutap INTERFACE
     $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/include>
     $<INSTALL_INTERFACE:include>)
 target_compile_features(mutap INTERFACE cxx_std_20)
-target_link_libraries(mutap INTERFACE mutap_fft)
-
-# ------------------------------------------------------------------------------
-# Helium (Cortex-M55 MVE) FFT backend. When ON, the float32 real FFT in
-# include/mutap/fft.h routes through the vendored CMSIS-DSP subset
-# (third_party/cmsis-dsp, Apache-2.0) instead of Ooura — measured ~3x fewer
-# instructions per transform, ~43% off the whole chain on the M55
-# (docs/optimization.md). Double is untouched (always Ooura), so the
-# double-precision ITU certification is unaffected; on the float32 embedded
-# path CMSIS agrees with Ooura to single-precision rounding (not bit-identical).
-#
-# Default ON for the bare-metal M55 embedded profile (the deployment target),
-# OFF everywhere else — desktop, the Max/C-ABI host builds (incl. Apple
-# Silicon arm64, which is not this Generic+arm profile), and Hexagon all keep
-# Ooura. Force the Ooura path on the M55 with -DMUTAP_FFT_CMSIS=OFF.
-# ------------------------------------------------------------------------------
-if(CMAKE_SYSTEM_NAME STREQUAL "Generic" AND CMAKE_SYSTEM_PROCESSOR MATCHES "arm|ARM")
-    set(_mutap_fft_cmsis_default ON) # bare-metal M55 (cmake/arm-cortex-m55-mps3.cmake)
-else()
-    set(_mutap_fft_cmsis_default OFF)
-endif()
-option(MUTAP_FFT_CMSIS "Route the float32 FFT through the vendored CMSIS-DSP Helium backend (Arm/Helium only)" ${_mutap_fft_cmsis_default})
-if(MUTAP_FFT_CMSIS)
-    if(NOT CMAKE_SYSTEM_PROCESSOR MATCHES "arm|ARM")
-        message(FATAL_ERROR
-            "MUTAP_FFT_CMSIS is an Arm/Helium backend; configure with the "
-            "cortex-m55 toolchain (cmake/arm-cortex-m55-mps3.cmake).")
-    endif()
-    file(GLOB _mutap_cmsis_src CONFIGURE_DEPENDS
-        ${CMAKE_CURRENT_SOURCE_DIR}/third_party/cmsis-dsp/Source/*/*.c)
-    target_sources(mutap_fft PRIVATE ${_mutap_cmsis_src})
-    # PRIVATE for the vendored .c compiles; the same dirs are exposed on the
-    # INTERFACE below so consumers' fft.h can resolve "arm_math.h".
-    set(_mutap_cmsis_inc
-        ${CMAKE_CURRENT_SOURCE_DIR}/third_party/cmsis-dsp/Include
-        ${CMAKE_CURRENT_SOURCE_DIR}/third_party/cmsis-dsp/PrivateInclude
-        ${CMAKE_CURRENT_SOURCE_DIR}/third_party/cmsis-dsp/cmsis-core/Include)
-    target_include_directories(mutap_fft PRIVATE ${_mutap_cmsis_inc})
-    target_include_directories(mutap INTERFACE
-        $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/third_party/cmsis-dsp/Include>
-        $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/third_party/cmsis-dsp/PrivateInclude>
-        $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/third_party/cmsis-dsp/cmsis-core/Include>)
-    target_compile_definitions(mutap INTERFACE MUTAP_FFT_CMSIS)
-    message(STATUS "MuTap: float32 FFT backend = CMSIS-DSP Helium")
-endif()
-
-# ------------------------------------------------------------------------------
-# Apple vDSP (Accelerate) FFT backend. When ON, the float32 real FFT in
-# include/mutap/fft.h routes through Apple's vDSP instead of Ooura — measured
-# ~3x faster per transform than clang-autovectorized Ooura on Apple Silicon
-# (docs/optimization.md), which is what the mutap.aec~ Max external runs.
-# Double is untouched (always Ooura), so the double-precision ITU certification
-# is unaffected; on the float32 path vDSP agrees with Ooura to single-precision
-# rounding (not bit-identical). Default ON for Apple, OFF everywhere else; it is
-# mutually exclusive with the CMSIS backend. Force Ooura with
-# -DMUTAP_FFT_ACCELERATE=OFF.
-# ------------------------------------------------------------------------------
-if(APPLE AND NOT MUTAP_FFT_CMSIS)
-    set(_mutap_fft_accel_default ON)
-else()
-    set(_mutap_fft_accel_default OFF)
-endif()
-option(MUTAP_FFT_ACCELERATE "Route the float32 FFT through Apple's vDSP (Accelerate) backend (Apple only)" ${_mutap_fft_accel_default})
-if(MUTAP_FFT_ACCELERATE)
-    if(NOT APPLE)
-        message(FATAL_ERROR "MUTAP_FFT_ACCELERATE requires Apple's Accelerate framework (macOS).")
-    endif()
-    if(MUTAP_FFT_CMSIS)
-        message(FATAL_ERROR "MUTAP_FFT_ACCELERATE and MUTAP_FFT_CMSIS are mutually exclusive.")
-    endif()
-    target_compile_definitions(mutap INTERFACE MUTAP_FFT_ACCELERATE)
-    target_link_libraries(mutap INTERFACE "-framework Accelerate")
-    message(STATUS "MuTap: float32 FFT backend = Apple vDSP (Accelerate)")
-endif()
+target_link_libraries(mutap INTERFACE tap::dsp)
 
 # ------------------------------------------------------------------------------
 # Options

diff --git a/bench/bench_aec.cpp b/bench/bench_aec.cpp
@@ -69,8 +69,8 @@ namespace {
     };
 
     template <typename S>
-    typename mutap::aec_chain<S>::config preset_of(const geometry& g) {
-        return mutap::aec_chain_preset<S>(g.block, g.partitions, g.fs);
+    typename tap::mu::aec_chain<S>::config preset_of(const geometry& g) {
+        return tap::mu::aec_chain_preset<S>(g.block, g.partitions, g.fs);
     }
 
     template <typename Proc, typename S>
@@ -82,9 +82,9 @@ namespace {
 
     template <typename S>
     void bench_fdkf(benchmark::State& state, geometry g) {
-        mutap::partitioned_fdkf<S> core(preset_of<S>(g).canceller);
-        corpus<S>                  c(g);
-        std::vector<S>             e(g.block);
+        tap::mu::partitioned_fdkf<S> core(preset_of<S>(g).canceller);
+        corpus<S>                    c(g);
+        std::vector<S>               e(g.block);
         warm(core, c, e);
         size_t i = 2000;
         for (auto _ : state) {
@@ -99,9 +99,9 @@ namespace {
     void bench_suppressor(benchmark::State& state, geometry g) {
         auto pf       = preset_of<S>(g).postfilter;
         pf.block_size = g.block;
-        mutap::residual_suppressor<S> sup(pf);
-        corpus<S>                     c(g);
-        std::vector<S>                e(g.block);
+        tap::mu::residual_suppressor<S> sup(pf);
+        corpus<S>                       c(g);
+        std::vector<S>                  e(g.block);
         // mid = the corpus mic, yhat = the corpus echo estimate stand-in
         for (size_t i = 0; i < 2000; ++i) {
             sup.process_block(c.yb(i), c.xb(i), e.data());
@@ -121,9 +121,9 @@ namespace {
         auto sc       = cfg.canceller;
         sc.partitions = cfg.shadow_partitions;
         sc.transition = cfg.shadow_transition;
-        mutap::partitioned_fdkf<S> shadow(sc);
-        corpus<S>                  c(g);
-        std::vector<S>             e(g.block);
+        tap::mu::partitioned_fdkf<S> shadow(sc);
+        corpus<S>                    c(g);
+        std::vector<S>               e(g.block);
         warm(shadow, c, e);
         size_t i = 2000;
         for (auto _ : state) {
@@ -136,9 +136,9 @@ namespace {
 
     template <typename S>
     void bench_chain(benchmark::State& state, geometry g) {
-        mutap::aec_chain<S> chain(preset_of<S>(g));
-        corpus<S>           c(g);
-        std::vector<S>      e(g.block);
+        tap::mu::aec_chain<S> chain(preset_of<S>(g));
+        corpus<S>             c(g);
+        std::vector<S>        e(g.block);
         warm(chain, c, e);
         size_t i = 2000;
         for (auto _ : state) {

diff --git a/bench/icount/icount_main.cpp b/bench/icount/icount_main.cpp
@@ -87,7 +87,7 @@ namespace {
     };
 
     auto preset() {
-        return mutap::aec_chain_preset<float>(k_geo.block, k_geo.partitions, k_geo.fs);
+        return tap::mu::aec_chain_preset<float>(k_geo.block, k_geo.partitions, k_geo.fs);
     }
 
     // Accumulate a checksum over the timed output so nothing is dead code.
@@ -105,7 +105,7 @@ namespace {
         double             sink = 0.0;
 
 #if MUTAP_SC_LAYER == 0
-        mutap::partitioned_fdkf<float> core(preset().canceller);
+        tap::mu::partitioned_fdkf<float> core(preset().canceller);
         for (std::size_t i = 0; i < k_warm; ++i) {
             core.process_block(c.xb(i), c.yb(i), e.data());
         }
@@ -116,7 +116,7 @@ namespace {
 #elif MUTAP_SC_LAYER == 1
         auto pf       = preset().postfilter;
         pf.block_size = k_geo.block;
-        mutap::residual_suppressor<float> sup(pf);
+        tap::mu::residual_suppressor<float> sup(pf);
         for (std::size_t i = 0; i < k_warm; ++i) {
             sup.process_block(c.yb(i), c.xb(i), e.data());
         }
@@ -129,7 +129,7 @@ namespace {
         auto sc       = cfg.canceller;
         sc.partitions = cfg.shadow_partitions;
         sc.transition = cfg.shadow_transition;
-        mutap::partitioned_fdkf<float> shadow(sc);
+        tap::mu::partitioned_fdkf<float> shadow(sc);
         for (std::size_t i = 0; i < k_warm; ++i) {
             shadow.process_block(c.xb(i), c.yb(i), e.data());
         }
@@ -138,7 +138,7 @@ namespace {
             sink += sum_out(e);
         }
 #else
-        mutap::aec_chain<float> chain(preset());
+        tap::mu::aec_chain<float> chain(preset());
         for (std::size_t i = 0; i < k_warm; ++i) {
             chain.process_block(c.xb(i), c.yb(i), e.data());
         }

diff --git a/include/mutap/fd_kalman.h b/include/mutap/fd_kalman.h
@@ -15,7 +15,7 @@
 #include "mutap/fdaf.h"
 #include "mutap/fft.h"
 
-namespace mutap {
+namespace tap::mu {
 
     /// Partitioned-block frequency-domain Kalman filter (after Enzner & Vary
     /// 2006; Kuech, Mabande & Enzner 2014; PEM-wrapped for feedback
@@ -549,4 +549,4 @@ namespace mutap {
         bool                   m_adapt    = true;
     };
 
-} // namespace mutap
+} // namespace tap::mu
diff --git a/include/mutap/fdaf.h b/include/mutap/fdaf.h
@@ -13,7 +13,7 @@
 
 #include "mutap/fft.h"
 
-namespace mutap {
+namespace tap::mu {
 
     namespace detail {
         /// Packed-spectrum multiply-accumulate: y += a * b, in the Ooura real-FFT
@@ -448,4 +448,4 @@ namespace mutap {
         Sample                 m_ee_weight = Sample(0); ///< 1 - a^n debiasing weight for m_s_ee
     };
 
-} // namespace mutap
+} // namespace tap::mu