diff --git a/src/cdf.rs b/src/cdf.rs
index 9619a35e2..d3634bc2e 100644
--- a/src/cdf.rs
+++ b/src/cdf.rs
@@ -6,8 +6,6 @@ use ::libc;
 extern "C" {
     fn memcpy(_: *mut libc::c_void, _: *const libc::c_void, _: libc::c_ulong) -> *mut libc::c_void;
     fn memset(_: *mut libc::c_void, _: libc::c_int, _: libc::c_ulong) -> *mut libc::c_void;
-    fn dav1d_ref_create_using_pool(pool: *mut Dav1dMemPool, size: size_t) -> *mut Dav1dRef;
-    fn dav1d_ref_dec(r#ref: *mut *mut Dav1dRef);
 }
 
 use crate::src::tables::dav1d_partition_type_count;
@@ -17,6 +15,8 @@ use crate::include::stdatomic::atomic_uint;
 
 use crate::include::dav1d::common::Dav1dDataProps;
 use crate::include::dav1d::data::Dav1dData;
+use crate::src::r#ref::dav1d_ref_create_using_pool;
+use crate::src::r#ref::dav1d_ref_dec;
 use crate::src::r#ref::Dav1dRef;
 #[derive(Copy, Clone)]
 #[repr(C)]
@@ -5402,8 +5402,8 @@ static av1_default_coef_cdf: [CdfCoefContext; 4] = [
         ])),
     },
 ];
-#[no_mangle]
-pub unsafe extern "C" fn dav1d_cdf_thread_update(
+
+pub unsafe fn dav1d_cdf_thread_update(
     hdr: *const Dav1dFrameHeader,
     dst: *mut CdfContext,
     src: *const CdfContext,
@@ -6136,6 +6136,7 @@ pub unsafe extern "C" fn dav1d_cdf_thread_update(
         k_17 += 1;
     }
 }
+
 #[inline]
 unsafe extern "C" fn get_qcat_idx(q: libc::c_int) -> libc::c_int {
     if q <= 20 {
@@ -6149,16 +6150,13 @@ unsafe extern "C" fn get_qcat_idx(q: libc::c_int) -> libc::c_int {
     }
     return 3 as libc::c_int;
 }
-#[no_mangle]
-pub unsafe extern "C" fn dav1d_cdf_thread_init_static(
-    cdf: *mut CdfThreadContext,
-    qidx: libc::c_int,
-) {
+
+pub unsafe fn dav1d_cdf_thread_init_static(cdf: *mut CdfThreadContext, qidx: libc::c_int) {
     (*cdf).r#ref = 0 as *mut Dav1dRef;
     (*cdf).data.qcat = get_qcat_idx(qidx) as libc::c_uint;
 }
-#[no_mangle]
-pub unsafe extern "C" fn dav1d_cdf_thread_copy(dst: *mut CdfContext, src: *const CdfThreadContext) {
+
+pub unsafe fn dav1d_cdf_thread_copy(dst: *mut CdfContext, src: *const CdfThreadContext) {
     if !((*src).r#ref).is_null() {
         memcpy(
             dst as *mut libc::c_void,
@@ -6189,6 +6187,7 @@ pub unsafe extern "C" fn dav1d_cdf_thread_copy(dst: *mut CdfContext, src: *const
         (*dst).mv.comp[0] = (*dst).mv.comp[1];
     };
 }
+
 #[no_mangle]
 pub unsafe extern "C" fn dav1d_cdf_thread_alloc(
     c: *mut Dav1dContext,
@@ -6209,18 +6208,15 @@ pub unsafe extern "C" fn dav1d_cdf_thread_alloc(
     }
     return 0 as libc::c_int;
 }
-#[no_mangle]
-pub unsafe extern "C" fn dav1d_cdf_thread_ref(
-    dst: *mut CdfThreadContext,
-    src: *mut CdfThreadContext,
-) {
+
+pub unsafe fn dav1d_cdf_thread_ref(dst: *mut CdfThreadContext, src: *mut CdfThreadContext) {
     *dst = *src;
     if !((*src).r#ref).is_null() {
         dav1d_ref_inc((*src).r#ref);
     }
 }
-#[no_mangle]
-pub unsafe extern "C" fn dav1d_cdf_thread_unref(cdf: *mut CdfThreadContext) {
+
+pub unsafe fn dav1d_cdf_thread_unref(cdf: *mut CdfThreadContext) {
     memset(
         &mut (*cdf).data as *mut CdfThreadContext_data as *mut libc::c_void,
         0 as libc::c_int,
diff --git a/src/cpu.rs b/src/cpu.rs
index 87c519ce6..5195aa6ac 100644
--- a/src/cpu.rs
+++ b/src/cpu.rs
@@ -46,7 +46,6 @@ pub fn dav1d_get_cpu_flags() -> libc::c_uint {
     flags
 }
 
-#[no_mangle]
 #[cold]
 pub unsafe fn dav1d_init_cpu() {
     #[cfg(feature = "asm")]
diff --git a/src/data.rs b/src/data.rs
index d651a32e8..343ff3eed 100644
--- a/src/data.rs
+++ b/src/data.rs
@@ -5,25 +5,18 @@ use ::libc;
 extern "C" {
     fn memset(_: *mut libc::c_void, _: libc::c_int, _: libc::c_ulong) -> *mut libc::c_void;
     fn fprintf(_: *mut libc::FILE, _: *const libc::c_char, _: ...) -> libc::c_int;
-    fn dav1d_ref_dec(r#ref: *mut *mut Dav1dRef);
-    fn dav1d_ref_wrap(
-        ptr: *const uint8_t,
-        free_callback: Option<unsafe extern "C" fn(*const uint8_t, *mut libc::c_void) -> ()>,
-        user_data: *mut libc::c_void,
-    ) -> *mut Dav1dRef;
-    fn dav1d_ref_create(size: size_t) -> *mut Dav1dRef;
 }
 
 use crate::src::r#ref::Dav1dRef;
 
 use crate::include::dav1d::common::Dav1dDataProps;
 use crate::include::dav1d::data::Dav1dData;
+use crate::src::r#ref::dav1d_ref_create;
+use crate::src::r#ref::dav1d_ref_dec;
 use crate::src::r#ref::dav1d_ref_inc;
-#[no_mangle]
-pub unsafe extern "C" fn dav1d_data_create_internal(
-    buf: *mut Dav1dData,
-    sz: size_t,
-) -> *mut uint8_t {
+use crate::src::r#ref::dav1d_ref_wrap;
+
+pub unsafe fn dav1d_data_create_internal(buf: *mut Dav1dData, sz: size_t) -> *mut uint8_t {
     if buf.is_null() {
         fprintf(
             stderr,
@@ -49,8 +42,8 @@ pub unsafe extern "C" fn dav1d_data_create_internal(
     (*buf).m.size = sz;
     return (*(*buf).r#ref).data as *mut uint8_t;
 }
-#[no_mangle]
-pub unsafe extern "C" fn dav1d_data_wrap_internal(
+
+pub unsafe fn dav1d_data_wrap_internal(
     buf: *mut Dav1dData,
     ptr: *const uint8_t,
     sz: size_t,
@@ -103,8 +96,8 @@ pub unsafe extern "C" fn dav1d_data_wrap_internal(
     (*buf).m.size = sz;
     return 0 as libc::c_int;
 }
-#[no_mangle]
-pub unsafe extern "C" fn dav1d_data_wrap_user_data_internal(
+
+pub unsafe fn dav1d_data_wrap_user_data_internal(
     buf: *mut Dav1dData,
     user_data: *const uint8_t,
     free_callback: Option<unsafe extern "C" fn(*const uint8_t, *mut libc::c_void) -> ()>,
@@ -141,8 +134,8 @@ pub unsafe extern "C" fn dav1d_data_wrap_user_data_internal(
     (*buf).m.user_data.data = user_data;
     return 0 as libc::c_int;
 }
-#[no_mangle]
-pub unsafe extern "C" fn dav1d_data_ref(dst: *mut Dav1dData, src: *const Dav1dData) {
+
+pub unsafe fn dav1d_data_ref(dst: *mut Dav1dData, src: *const Dav1dData) {
     if dst.is_null() {
         fprintf(
             stderr,
@@ -192,11 +185,8 @@ pub unsafe extern "C" fn dav1d_data_ref(dst: *mut Dav1dData, src: *const Dav1dDa
     }
     *dst = *src;
 }
-#[no_mangle]
-pub unsafe extern "C" fn dav1d_data_props_copy(
-    dst: *mut Dav1dDataProps,
-    src: *const Dav1dDataProps,
-) {
+
+pub unsafe fn dav1d_data_props_copy(dst: *mut Dav1dDataProps, src: *const Dav1dDataProps) {
     if dst.is_null() {
         unreachable!();
     }
@@ -209,8 +199,8 @@ pub unsafe extern "C" fn dav1d_data_props_copy(
         dav1d_ref_inc((*dst).user_data.r#ref);
     }
 }
-#[no_mangle]
-pub unsafe extern "C" fn dav1d_data_props_set_defaults(props: *mut Dav1dDataProps) {
+
+pub unsafe fn dav1d_data_props_set_defaults(props: *mut Dav1dDataProps) {
     if props.is_null() {
         unreachable!();
     }
@@ -222,8 +212,8 @@ pub unsafe extern "C" fn dav1d_data_props_set_defaults(props: *mut Dav1dDataProp
     (*props).timestamp = i64::MIN;
     (*props).offset = -1;
 }
-#[no_mangle]
-pub unsafe extern "C" fn dav1d_data_props_unref_internal(props: *mut Dav1dDataProps) {
+
+pub unsafe fn dav1d_data_props_unref_internal(props: *mut Dav1dDataProps) {
     if props.is_null() {
         fprintf(
             stderr,
@@ -240,8 +230,8 @@ pub unsafe extern "C" fn dav1d_data_props_unref_internal(props: *mut Dav1dDataPr
     dav1d_data_props_set_defaults(props);
     dav1d_ref_dec(&mut user_data_ref);
 }
-#[no_mangle]
-pub unsafe extern "C" fn dav1d_data_unref_internal(buf: *mut Dav1dData) {
+
+pub unsafe fn dav1d_data_unref_internal(buf: *mut Dav1dData) {
     if buf.is_null() {
         fprintf(
             stderr,
diff --git a/src/decode.rs b/src/decode.rs
index be28af77e..68cc675fe 100644
--- a/src/decode.rs
+++ b/src/decode.rs
@@ -9,7 +9,8 @@ use crate::include::dav1d::headers::{Dav1dTxfmMode, DAV1D_MAX_SEGMENTS};
 use crate::include::stddef::*;
 use crate::include::stdint::*;
 use crate::src::align::Align16;
-use crate::src::cdf::{CdfContext, CdfMvComponent, CdfMvContext};
+use crate::src::cdf::CdfMvComponent;
+use crate::src::cdf::CdfMvContext;
 use crate::src::ctx::CaseSet;
 use crate::src::looprestoration::dav1d_loop_restoration_dsp_init;
 
@@ -36,54 +37,11 @@ extern "C" {
     fn pthread_cond_signal(__cond: *mut pthread_cond_t) -> libc::c_int;
     fn pthread_cond_wait(__cond: *mut pthread_cond_t, __mutex: *mut pthread_mutex_t)
         -> libc::c_int;
-    fn dav1d_ref_create_using_pool(pool: *mut Dav1dMemPool, size: size_t) -> *mut Dav1dRef;
-    fn dav1d_ref_dec(r#ref: *mut *mut Dav1dRef);
-    fn dav1d_cdf_thread_init_static(cdf: *mut CdfThreadContext, qidx: libc::c_int);
     fn dav1d_cdf_thread_alloc(
         c: *mut Dav1dContext,
         cdf: *mut CdfThreadContext,
         have_frame_mt: libc::c_int,
     ) -> libc::c_int;
-    fn dav1d_cdf_thread_copy(dst: *mut CdfContext, src: *const CdfThreadContext);
-    fn dav1d_cdf_thread_ref(dst: *mut CdfThreadContext, src: *mut CdfThreadContext);
-    fn dav1d_cdf_thread_unref(cdf: *mut CdfThreadContext);
-    fn dav1d_cdf_thread_update(
-        hdr: *const Dav1dFrameHeader,
-        dst: *mut CdfContext,
-        src: *const CdfContext,
-    );
-    fn dav1d_data_props_copy(dst: *mut Dav1dDataProps, src: *const Dav1dDataProps);
-    fn dav1d_data_unref_internal(buf: *mut Dav1dData);
-    fn dav1d_refmvs_init_frame(
-        rf: *mut refmvs_frame,
-        seq_hdr: *const Dav1dSequenceHeader,
-        frm_hdr: *const Dav1dFrameHeader,
-        ref_poc: *const libc::c_uint,
-        rp: *mut refmvs_temporal_block,
-        ref_ref_poc: *const [libc::c_uint; 7],
-        rp_ref: *const *mut refmvs_temporal_block,
-        n_tile_threads: libc::c_int,
-        n_frame_threads: libc::c_int,
-    ) -> libc::c_int;
-    fn dav1d_refmvs_save_tmvs(
-        dsp: *const Dav1dRefmvsDSPContext,
-        rt: *const refmvs_tile,
-        col_start8: libc::c_int,
-        col_end8: libc::c_int,
-        row_start8: libc::c_int,
-        row_end8: libc::c_int,
-    );
-    fn dav1d_refmvs_tile_sbrow_init(
-        rt: *mut refmvs_tile,
-        rf: *const refmvs_frame,
-        tile_col_start4: libc::c_int,
-        tile_col_end4: libc::c_int,
-        tile_row_start4: libc::c_int,
-        tile_row_end4: libc::c_int,
-        sby: libc::c_int,
-        tile_row_idx: libc::c_int,
-        pass: libc::c_int,
-    );
     #[cfg(feature = "bitdepth_8")]
     fn dav1d_film_grain_dsp_init_8bpc(c: *mut Dav1dFilmGrainDSPContext);
     #[cfg(feature = "bitdepth_16")]
@@ -115,11 +73,6 @@ extern "C" {
         w: libc::c_int,
         src: *const Dav1dPicture,
     ) -> libc::c_int;
-    fn dav1d_picture_ref(dst: *mut Dav1dPicture, src: *const Dav1dPicture);
-    fn dav1d_thread_picture_ref(dst: *mut Dav1dThreadPicture, src: *const Dav1dThreadPicture);
-    fn dav1d_thread_picture_unref(p: *mut Dav1dThreadPicture);
-    fn dav1d_picture_unref_internal(p: *mut Dav1dPicture);
-    fn dav1d_picture_get_event_flags(p: *const Dav1dThreadPicture) -> Dav1dEventFlags;
     #[cfg(feature = "bitdepth_8")]
     fn dav1d_recon_b_intra_8bpc(
         t: *mut Dav1dTaskContext,
@@ -202,6 +155,9 @@ use crate::src::msac::dav1d_msac_decode_symbol_adapt8;
 use crate::src::msac::dav1d_msac_init;
 use crate::src::qm::dav1d_qm_tbl;
 use crate::src::refmvs::dav1d_refmvs_find;
+use crate::src::refmvs::dav1d_refmvs_init_frame;
+use crate::src::refmvs::dav1d_refmvs_save_tmvs;
+use crate::src::refmvs::dav1d_refmvs_tile_sbrow_init;
 use crate::src::tables::dav1d_al_part_ctx;
 use crate::src::tables::dav1d_block_dimensions;
 use crate::src::tables::dav1d_block_sizes;
@@ -223,6 +179,8 @@ use crate::include::dav1d::common::Dav1dDataProps;
 use crate::include::dav1d::data::Dav1dData;
 use crate::include::stdatomic::atomic_int;
 use crate::include::stdatomic::atomic_uint;
+use crate::src::data::dav1d_data_props_copy;
+use crate::src::data::dav1d_data_unref_internal;
 use crate::src::r#ref::Dav1dRef;
 #[derive(Copy, Clone)]
 #[repr(C)]
@@ -808,11 +766,21 @@ pub type generate_grain_uv_fn = Option<
 >;
 pub type generate_grain_y_fn =
     Option<unsafe extern "C" fn(*mut [entry; 82], *const Dav1dFilmGrainData) -> ()>;
+use crate::src::cdf::dav1d_cdf_thread_copy;
+use crate::src::cdf::dav1d_cdf_thread_init_static;
+use crate::src::cdf::dav1d_cdf_thread_ref;
+use crate::src::cdf::dav1d_cdf_thread_unref;
+use crate::src::cdf::dav1d_cdf_thread_update;
 use crate::src::cdf::CdfThreadContext;
 
 use crate::src::internal::Dav1dContext_frame_thread;
 use crate::src::internal::Dav1dContext_refs;
 use crate::src::internal::Dav1dTileGroup;
+use crate::src::picture::dav1d_picture_get_event_flags;
+use crate::src::picture::dav1d_picture_ref;
+use crate::src::picture::dav1d_picture_unref_internal;
+use crate::src::picture::dav1d_thread_picture_ref;
+use crate::src::picture::dav1d_thread_picture_unref;
 use crate::src::picture::Dav1dThreadPicture;
 pub type backup_ipred_edge_fn = Option<unsafe extern "C" fn(*mut Dav1dTaskContext) -> ()>;
 pub type filter_sbrow_fn = Option<unsafe extern "C" fn(*mut Dav1dFrameContext, libc::c_int) -> ()>;
@@ -901,6 +869,8 @@ use crate::src::mem::dav1d_alloc_aligned;
 use crate::src::mem::dav1d_free_aligned;
 use crate::src::mem::dav1d_freep_aligned;
 use crate::src::mem::freep;
+use crate::src::r#ref::dav1d_ref_create_using_pool;
+use crate::src::r#ref::dav1d_ref_dec;
 use crate::src::r#ref::dav1d_ref_inc;
 
 use crate::src::tables::cfl_allowed_mask;
diff --git a/src/lib.rs b/src/lib.rs
index fc61f0c3a..263f659b3 100644
--- a/src/lib.rs
+++ b/src/lib.rs
@@ -17,7 +17,6 @@ extern "C" {
     ) -> libc::c_int;
     fn abort() -> !;
     fn fprintf(_: *mut libc::FILE, _: *const libc::c_char, _: ...) -> libc::c_int;
-    fn dav1d_init_cpu();
     fn dav1d_num_logical_processors(c: *mut Dav1dContext) -> libc::c_int;
     #[cfg(feature = "bitdepth_16")]
     fn dav1d_apply_grain_16bpc(
@@ -31,41 +30,14 @@ extern "C" {
         out: *mut Dav1dPicture,
         in_0: *const Dav1dPicture,
     );
-    fn dav1d_data_props_unref_internal(props: *mut Dav1dDataProps);
-    fn dav1d_picture_unref_internal(p: *mut Dav1dPicture);
-    fn dav1d_data_create_internal(buf: *mut Dav1dData, sz: size_t) -> *mut uint8_t;
-    fn dav1d_data_props_copy(dst: *mut Dav1dDataProps, src: *const Dav1dDataProps);
-    fn dav1d_data_wrap_internal(
-        buf: *mut Dav1dData,
-        ptr: *const uint8_t,
-        sz: size_t,
-        free_callback: Option<unsafe extern "C" fn(*const uint8_t, *mut libc::c_void) -> ()>,
-        user_data: *mut libc::c_void,
-    ) -> libc::c_int;
-    fn dav1d_thread_picture_ref(dst: *mut Dav1dThreadPicture, src: *const Dav1dThreadPicture);
-    fn dav1d_picture_get_event_flags(p: *const Dav1dThreadPicture) -> Dav1dEventFlags;
-    fn dav1d_picture_move_ref(dst: *mut Dav1dPicture, src: *mut Dav1dPicture);
-    fn dav1d_data_wrap_user_data_internal(
-        buf: *mut Dav1dData,
-        user_data: *const uint8_t,
-        free_callback: Option<unsafe extern "C" fn(*const uint8_t, *mut libc::c_void) -> ()>,
-        cookie: *mut libc::c_void,
-    ) -> libc::c_int;
-    fn dav1d_picture_ref(dst: *mut Dav1dPicture, src: *const Dav1dPicture);
-    fn dav1d_data_unref_internal(buf: *mut Dav1dData);
     fn dav1d_picture_alloc_copy(
         c: *mut Dav1dContext,
         dst: *mut Dav1dPicture,
         w: libc::c_int,
         src: *const Dav1dPicture,
     ) -> libc::c_int;
-    fn dav1d_data_ref(dst: *mut Dav1dData, src: *const Dav1dData);
-    fn dav1d_thread_picture_move_ref(dst: *mut Dav1dThreadPicture, src: *mut Dav1dThreadPicture);
     fn pthread_attr_init(__attr: *mut pthread_attr_t) -> libc::c_int;
     fn __sysconf(__name: libc::c_int) -> libc::c_long;
-    fn dav1d_data_props_set_defaults(props: *mut Dav1dDataProps);
-    fn dav1d_mem_pool_init(pool: *mut *mut Dav1dMemPool) -> libc::c_int;
-    fn dav1d_refmvs_init(rf: *mut refmvs_frame);
     fn pthread_create(
         __newthread: *mut pthread_t,
         __attr: *const pthread_attr_t,
@@ -74,14 +46,7 @@ extern "C" {
     ) -> libc::c_int;
     fn dav1d_refmvs_dsp_init(dsp: *mut Dav1dRefmvsDSPContext);
     fn pthread_join(__th: pthread_t, __thread_return: *mut *mut libc::c_void) -> libc::c_int;
-    fn dav1d_refmvs_clear(rf: *mut refmvs_frame);
-    fn dav1d_cdf_thread_unref(cdf: *mut CdfThreadContext);
-    fn dav1d_thread_picture_unref(p: *mut Dav1dThreadPicture);
-    fn dav1d_ref_dec(r#ref: *mut *mut Dav1dRef);
-    fn dav1d_mem_pool_end(pool: *mut Dav1dMemPool);
     fn pthread_attr_destroy(__attr: *mut pthread_attr_t) -> libc::c_int;
-    fn dav1d_default_picture_alloc(p: *mut Dav1dPicture, cookie: *mut libc::c_void) -> libc::c_int;
-    fn dav1d_default_picture_release(p: *mut Dav1dPicture, cookie: *mut libc::c_void);
     fn pthread_once(
         __once_control: *mut pthread_once_t,
         __init_routine: Option<unsafe extern "C" fn() -> ()>,
@@ -102,11 +67,6 @@ extern "C" {
         -> libc::c_int;
     fn pthread_cond_broadcast(__cond: *mut pthread_cond_t) -> libc::c_int;
     fn pthread_cond_destroy(__cond: *mut pthread_cond_t) -> libc::c_int;
-    fn dav1d_log_default_callback(
-        cookie: *mut libc::c_void,
-        format: *const libc::c_char,
-        ap: ::core::ffi::VaList,
-    );
     fn dav1d_log(c: *mut Dav1dContext, format: *const libc::c_char, _: ...);
     fn dav1d_parse_obus(
         c: *mut Dav1dContext,
@@ -118,7 +78,6 @@ extern "C" {
         out: *mut Dav1dPicture,
         in_0: *const Dav1dPicture,
     );
-    fn dav1d_worker_task(data: *mut libc::c_void) -> *mut libc::c_void;
     fn dav1d_decode_frame_exit(f: *mut Dav1dFrameContext, retval: libc::c_int);
     fn dav1d_init_wedge_masks();
     fn dav1d_init_interintra_masks();
@@ -126,7 +85,19 @@ extern "C" {
 use crate::include::dav1d::common::Dav1dDataProps;
 use crate::include::dav1d::common::Dav1dUserData;
 use crate::include::stdatomic::atomic_int;
+use crate::src::cpu::dav1d_init_cpu;
+use crate::src::data::dav1d_data_create_internal;
+use crate::src::data::dav1d_data_props_copy;
+use crate::src::data::dav1d_data_props_set_defaults;
+use crate::src::data::dav1d_data_props_unref_internal;
+use crate::src::data::dav1d_data_ref;
+use crate::src::data::dav1d_data_unref_internal;
+use crate::src::data::dav1d_data_wrap_internal;
+use crate::src::data::dav1d_data_wrap_user_data_internal;
+use crate::src::log::dav1d_log_default_callback;
+use crate::src::r#ref::dav1d_ref_dec;
 use crate::src::r#ref::Dav1dRef;
+use crate::src::thread_task::dav1d_worker_task;
 
 use crate::include::dav1d::headers::Dav1dWarpedMotionParams;
 
@@ -216,6 +187,8 @@ use crate::include::dav1d::dav1d::DAV1D_INLOOPFILTER_NONE;
 use crate::src::internal::Dav1dContext_intra_edge;
 use crate::src::intra_edge::EdgeFlags;
 
+use crate::src::refmvs::dav1d_refmvs_clear;
+use crate::src::refmvs::dav1d_refmvs_init;
 use crate::src::refmvs::Dav1dRefmvsDSPContext;
 #[derive(Copy, Clone)]
 #[repr(C)]
@@ -523,12 +496,22 @@ pub type generate_grain_uv_fn = Option<
 pub type generate_grain_y_fn =
     Option<unsafe extern "C" fn(*mut [entry; 82], *const Dav1dFilmGrainData) -> ()>;
 use crate::include::stdatomic::atomic_uint;
+use crate::src::cdf::dav1d_cdf_thread_unref;
 use crate::src::cdf::CdfThreadContext;
 
 use crate::src::internal::Dav1dContext_frame_thread;
 use crate::src::internal::Dav1dContext_refs;
 use crate::src::internal::Dav1dTileGroup;
 use crate::src::internal::TaskThreadData;
+use crate::src::picture::dav1d_default_picture_alloc;
+use crate::src::picture::dav1d_default_picture_release;
+use crate::src::picture::dav1d_picture_get_event_flags;
+use crate::src::picture::dav1d_picture_move_ref;
+use crate::src::picture::dav1d_picture_ref;
+use crate::src::picture::dav1d_picture_unref_internal;
+use crate::src::picture::dav1d_thread_picture_move_ref;
+use crate::src::picture::dav1d_thread_picture_ref;
+use crate::src::picture::dav1d_thread_picture_unref;
 use crate::src::picture::Dav1dThreadPicture;
 use libc::pthread_cond_t;
 #[derive(Copy, Clone)]
@@ -696,6 +679,8 @@ use crate::include::common::intops::umin;
 use crate::src::mem::dav1d_alloc_aligned;
 use crate::src::mem::dav1d_free_aligned;
 use crate::src::mem::dav1d_freep_aligned;
+use crate::src::mem::dav1d_mem_pool_end;
+use crate::src::mem::dav1d_mem_pool_init;
 use crate::src::mem::freep;
 #[cold]
 unsafe extern "C" fn init_internal() {
diff --git a/src/log.rs b/src/log.rs
index 36036982e..1b7b38fc3 100644
--- a/src/log.rs
+++ b/src/log.rs
@@ -538,7 +538,7 @@ pub type recon_b_intra_fn = Option<
     unsafe extern "C" fn(*mut Dav1dTaskContext, BlockSize, EdgeFlags, *const Av1Block) -> (),
 >;
 use crate::src::internal::ScalableMotionParams;
-#[no_mangle]
+
 #[cold]
 pub unsafe extern "C" fn dav1d_log_default_callback(
     _cookie: *mut libc::c_void,
diff --git a/src/mem.rs b/src/mem.rs
index 0771dde28..74b892c32 100644
--- a/src/mem.rs
+++ b/src/mem.rs
@@ -79,11 +79,8 @@ unsafe extern "C" fn mem_pool_destroy(pool: *mut Dav1dMemPool) {
     pthread_mutex_destroy(&mut (*pool).lock);
     free(pool as *mut libc::c_void);
 }
-#[no_mangle]
-pub unsafe extern "C" fn dav1d_mem_pool_push(
-    pool: *mut Dav1dMemPool,
-    buf: *mut Dav1dMemPoolBuffer,
-) {
+
+pub unsafe fn dav1d_mem_pool_push(pool: *mut Dav1dMemPool, buf: *mut Dav1dMemPoolBuffer) {
     pthread_mutex_lock(&mut (*pool).lock);
     (*pool).ref_cnt -= 1;
     let ref_cnt = (*pool).ref_cnt;
@@ -102,11 +99,8 @@ pub unsafe extern "C" fn dav1d_mem_pool_push(
         }
     };
 }
-#[no_mangle]
-pub unsafe extern "C" fn dav1d_mem_pool_pop(
-    pool: *mut Dav1dMemPool,
-    size: size_t,
-) -> *mut Dav1dMemPoolBuffer {
+
+pub unsafe fn dav1d_mem_pool_pop(pool: *mut Dav1dMemPool, size: size_t) -> *mut Dav1dMemPoolBuffer {
     if size & ::core::mem::size_of::<*mut libc::c_void>().wrapping_sub(1) != 0 {
         unreachable!();
     }
@@ -152,9 +146,9 @@ pub unsafe extern "C" fn dav1d_mem_pool_pop(
     }
     return buf;
 }
-#[no_mangle]
+
 #[cold]
-pub unsafe extern "C" fn dav1d_mem_pool_init(ppool: *mut *mut Dav1dMemPool) -> libc::c_int {
+pub unsafe fn dav1d_mem_pool_init(ppool: *mut *mut Dav1dMemPool) -> libc::c_int {
     let pool: *mut Dav1dMemPool =
         malloc(::core::mem::size_of::<Dav1dMemPool>() as libc::c_ulong) as *mut Dav1dMemPool;
     if !pool.is_null() {
@@ -170,9 +164,9 @@ pub unsafe extern "C" fn dav1d_mem_pool_init(ppool: *mut *mut Dav1dMemPool) -> l
     *ppool = 0 as *mut Dav1dMemPool;
     return -(12 as libc::c_int);
 }
-#[no_mangle]
+
 #[cold]
-pub unsafe extern "C" fn dav1d_mem_pool_end(pool: *mut Dav1dMemPool) {
+pub unsafe fn dav1d_mem_pool_end(pool: *mut Dav1dMemPool) {
     if !pool.is_null() {
         pthread_mutex_lock(&mut (*pool).lock);
         let mut buf: *mut Dav1dMemPoolBuffer = (*pool).buf;
diff --git a/src/obu.rs b/src/obu.rs
index 523a5f391..c18f18521 100644
--- a/src/obu.rs
+++ b/src/obu.rs
@@ -11,17 +11,6 @@ extern "C" {
     fn pthread_mutex_unlock(__mutex: *mut pthread_mutex_t) -> libc::c_int;
     fn pthread_cond_wait(__cond: *mut pthread_cond_t, __mutex: *mut pthread_mutex_t)
         -> libc::c_int;
-    fn dav1d_ref_create(size: size_t) -> *mut Dav1dRef;
-    fn dav1d_ref_create_using_pool(pool: *mut Dav1dMemPool, size: size_t) -> *mut Dav1dRef;
-    fn dav1d_ref_dec(r#ref: *mut *mut Dav1dRef);
-    fn dav1d_cdf_thread_ref(dst: *mut CdfThreadContext, src: *mut CdfThreadContext);
-    fn dav1d_cdf_thread_unref(cdf: *mut CdfThreadContext);
-    fn dav1d_data_ref(dst: *mut Dav1dData, src: *const Dav1dData);
-    fn dav1d_data_props_copy(dst: *mut Dav1dDataProps, src: *const Dav1dDataProps);
-    fn dav1d_data_unref_internal(buf: *mut Dav1dData);
-    fn dav1d_thread_picture_ref(dst: *mut Dav1dThreadPicture, src: *const Dav1dThreadPicture);
-    fn dav1d_thread_picture_unref(p: *mut Dav1dThreadPicture);
-    fn dav1d_picture_get_event_flags(p: *const Dav1dThreadPicture) -> Dav1dEventFlags;
     fn dav1d_init_get_bits(c: *mut GetBits, data: *const uint8_t, sz: size_t);
     fn dav1d_get_bit(c: *mut GetBits) -> libc::c_uint;
     fn dav1d_get_bits(c: *mut GetBits, n: libc::c_int) -> libc::c_uint;
@@ -39,6 +28,12 @@ use crate::src::tables::dav1d_default_wm_params;
 use crate::include::dav1d::common::Dav1dDataProps;
 use crate::include::dav1d::data::Dav1dData;
 use crate::include::stdatomic::atomic_int;
+use crate::src::data::dav1d_data_props_copy;
+use crate::src::data::dav1d_data_ref;
+use crate::src::data::dav1d_data_unref_internal;
+use crate::src::r#ref::dav1d_ref_create;
+use crate::src::r#ref::dav1d_ref_create_using_pool;
+use crate::src::r#ref::dav1d_ref_dec;
 use crate::src::r#ref::Dav1dRef;
 
 use crate::include::stdatomic::atomic_uint;
@@ -615,11 +610,16 @@ pub type generate_grain_uv_fn = Option<
 >;
 pub type generate_grain_y_fn =
     Option<unsafe extern "C" fn(*mut [entry; 82], *const Dav1dFilmGrainData) -> ()>;
+use crate::src::cdf::dav1d_cdf_thread_ref;
+use crate::src::cdf::dav1d_cdf_thread_unref;
 use crate::src::cdf::CdfThreadContext;
 
 use crate::src::internal::Dav1dContext_frame_thread;
 use crate::src::internal::Dav1dContext_refs;
 use crate::src::internal::Dav1dTileGroup;
+use crate::src::picture::dav1d_picture_get_event_flags;
+use crate::src::picture::dav1d_thread_picture_ref;
+use crate::src::picture::dav1d_thread_picture_unref;
 use crate::src::picture::Dav1dThreadPicture;
 pub type backup_ipred_edge_fn = Option<unsafe extern "C" fn(*mut Dav1dTaskContext) -> ()>;
 pub type filter_sbrow_fn = Option<unsafe extern "C" fn(*mut Dav1dFrameContext, libc::c_int) -> ()>;
diff --git a/src/picture.rs b/src/picture.rs
index d8dc867cc..127fcaa9d 100644
--- a/src/picture.rs
+++ b/src/picture.rs
@@ -9,22 +9,16 @@ extern "C" {
     fn free(_: *mut libc::c_void);
     fn memset(_: *mut libc::c_void, _: libc::c_int, _: size_t) -> *mut libc::c_void;
     fn strerror(_: libc::c_int) -> *mut libc::c_char;
-    fn dav1d_mem_pool_push(pool: *mut Dav1dMemPool, buf: *mut Dav1dMemPoolBuffer);
-    fn dav1d_mem_pool_pop(pool: *mut Dav1dMemPool, size: size_t) -> *mut Dav1dMemPoolBuffer;
-    fn dav1d_ref_wrap(
-        ptr: *const uint8_t,
-        free_callback: Option<unsafe extern "C" fn(*const uint8_t, *mut libc::c_void) -> ()>,
-        user_data: *mut libc::c_void,
-    ) -> *mut Dav1dRef;
-    fn dav1d_ref_dec(r#ref: *mut *mut Dav1dRef);
-    fn dav1d_data_props_copy(dst: *mut Dav1dDataProps, src: *const Dav1dDataProps);
-    fn dav1d_data_props_set_defaults(props: *mut Dav1dDataProps);
     fn dav1d_log(c: *mut Dav1dContext, format: *const libc::c_char, _: ...);
 }
 use crate::include::dav1d::common::Dav1dDataProps;
 use crate::include::dav1d::data::Dav1dData;
 use crate::include::stdatomic::atomic_int;
 use crate::include::stdatomic::atomic_uint;
+use crate::src::data::dav1d_data_props_copy;
+use crate::src::data::dav1d_data_props_set_defaults;
+use crate::src::r#ref::dav1d_ref_dec;
+use crate::src::r#ref::dav1d_ref_wrap;
 use crate::src::r#ref::Dav1dRef;
 #[derive(Copy, Clone)]
 #[repr(C)]
@@ -229,6 +223,8 @@ use crate::include::dav1d::dav1d::Dav1dEventFlags;
 use crate::include::dav1d::dav1d::Dav1dLogger;
 use crate::include::dav1d::dav1d::DAV1D_EVENT_FLAG_NEW_OP_PARAMS_INFO;
 use crate::include::dav1d::dav1d::DAV1D_EVENT_FLAG_NEW_SEQUENCE;
+use crate::src::mem::dav1d_mem_pool_pop;
+use crate::src::mem::dav1d_mem_pool_push;
 use crate::src::mem::Dav1dMemPool;
 use crate::src::mem::Dav1dMemPoolBuffer;
 pub type PictureFlags = libc::c_uint;
@@ -578,7 +574,7 @@ pub struct pic_ctx_context {
     pub extra_ptr: *mut libc::c_void,
 }
 use crate::src::r#ref::dav1d_ref_inc;
-#[no_mangle]
+
 pub unsafe extern "C" fn dav1d_default_picture_alloc(
     p: *mut Dav1dPicture,
     cookie: *mut libc::c_void,
@@ -637,7 +633,7 @@ pub unsafe extern "C" fn dav1d_default_picture_alloc(
     }) as *mut libc::c_void;
     return 0 as libc::c_int;
 }
-#[no_mangle]
+
 pub unsafe extern "C" fn dav1d_default_picture_release(
     p: *mut Dav1dPicture,
     cookie: *mut libc::c_void,
@@ -842,8 +838,8 @@ pub unsafe extern "C" fn dav1d_picture_alloc_copy(
     );
     return res;
 }
-#[no_mangle]
-pub unsafe extern "C" fn dav1d_picture_ref(dst: *mut Dav1dPicture, src: *const Dav1dPicture) {
+
+pub unsafe fn dav1d_picture_ref(dst: *mut Dav1dPicture, src: *const Dav1dPicture) {
     if dst.is_null() {
         fprintf(
             stderr,
@@ -908,8 +904,8 @@ pub unsafe extern "C" fn dav1d_picture_ref(dst: *mut Dav1dPicture, src: *const D
     }
     *dst = *src;
 }
-#[no_mangle]
-pub unsafe extern "C" fn dav1d_picture_move_ref(dst: *mut Dav1dPicture, src: *mut Dav1dPicture) {
+
+pub unsafe fn dav1d_picture_move_ref(dst: *mut Dav1dPicture, src: *mut Dav1dPicture) {
     if dst.is_null() {
         fprintf(
             stderr,
@@ -968,8 +964,8 @@ pub unsafe extern "C" fn dav1d_picture_move_ref(dst: *mut Dav1dPicture, src: *mu
         ::core::mem::size_of::<Dav1dPicture>(),
     );
 }
-#[no_mangle]
-pub unsafe extern "C" fn dav1d_thread_picture_ref(
+
+pub unsafe fn dav1d_thread_picture_ref(
     dst: *mut Dav1dThreadPicture,
     src: *const Dav1dThreadPicture,
 ) {
@@ -979,8 +975,8 @@ pub unsafe extern "C" fn dav1d_thread_picture_ref(
     (*dst).progress = (*src).progress;
     (*dst).flags = (*src).flags;
 }
-#[no_mangle]
-pub unsafe extern "C" fn dav1d_thread_picture_move_ref(
+
+pub unsafe fn dav1d_thread_picture_move_ref(
     dst: *mut Dav1dThreadPicture,
     src: *mut Dav1dThreadPicture,
 ) {
@@ -995,8 +991,8 @@ pub unsafe extern "C" fn dav1d_thread_picture_move_ref(
         ::core::mem::size_of::<Dav1dThreadPicture>(),
     );
 }
-#[no_mangle]
-pub unsafe extern "C" fn dav1d_picture_unref_internal(p: *mut Dav1dPicture) {
+
+pub unsafe fn dav1d_picture_unref_internal(p: *mut Dav1dPicture) {
     if p.is_null() {
         fprintf(
             stderr,
@@ -1038,15 +1034,13 @@ pub unsafe extern "C" fn dav1d_picture_unref_internal(p: *mut Dav1dPicture) {
     );
     dav1d_data_props_set_defaults(&mut (*p).m);
 }
-#[no_mangle]
-pub unsafe extern "C" fn dav1d_thread_picture_unref(p: *mut Dav1dThreadPicture) {
+
+pub unsafe fn dav1d_thread_picture_unref(p: *mut Dav1dThreadPicture) {
     dav1d_picture_unref_internal(&mut (*p).p);
     (*p).progress = 0 as *mut atomic_uint;
 }
-#[no_mangle]
-pub unsafe extern "C" fn dav1d_picture_get_event_flags(
-    p: *const Dav1dThreadPicture,
-) -> Dav1dEventFlags {
+
+pub unsafe fn dav1d_picture_get_event_flags(p: *const Dav1dThreadPicture) -> Dav1dEventFlags {
     if (*p).flags as u64 == 0 {
         return 0 as Dav1dEventFlags;
     }
diff --git a/src/ref.rs b/src/ref.rs
index c5fd3991c..576c46116 100644
--- a/src/ref.rs
+++ b/src/ref.rs
@@ -9,8 +9,6 @@ extern "C" {
         __alignment: size_t,
         __size: size_t,
     ) -> libc::c_int;
-    fn dav1d_mem_pool_push(pool: *mut Dav1dMemPool, buf: *mut Dav1dMemPoolBuffer);
-    fn dav1d_mem_pool_pop(pool: *mut Dav1dMemPool, size: size_t) -> *mut Dav1dMemPoolBuffer;
 }
 #[derive(Copy, Clone)]
 #[repr(C)]
@@ -26,6 +24,8 @@ use crate::include::stdatomic::atomic_int;
 
 use crate::src::mem::dav1d_alloc_aligned;
 use crate::src::mem::dav1d_free_aligned;
+use crate::src::mem::dav1d_mem_pool_pop;
+use crate::src::mem::dav1d_mem_pool_push;
 use crate::src::mem::Dav1dMemPool;
 use crate::src::mem::Dav1dMemPoolBuffer;
 
@@ -40,8 +40,8 @@ unsafe extern "C" fn default_free_callback(data: *const uint8_t, user_data: *mut
     }
     dav1d_free_aligned(user_data);
 }
-#[no_mangle]
-pub unsafe extern "C" fn dav1d_ref_create(mut size: size_t) -> *mut Dav1dRef {
+
+pub unsafe fn dav1d_ref_create(mut size: size_t) -> *mut Dav1dRef {
     size = size
         .wrapping_add(::core::mem::size_of::<*mut libc::c_void>())
         .wrapping_sub(1)
@@ -70,8 +70,8 @@ unsafe extern "C" fn pool_free_callback(data: *const uint8_t, user_data: *mut li
         user_data as *mut Dav1dMemPoolBuffer,
     );
 }
-#[no_mangle]
-pub unsafe extern "C" fn dav1d_ref_create_using_pool(
+
+pub unsafe fn dav1d_ref_create_using_pool(
     pool: *mut Dav1dMemPool,
     mut size: size_t,
 ) -> *mut Dav1dRef {
@@ -95,8 +95,8 @@ pub unsafe extern "C" fn dav1d_ref_create_using_pool(
     (*res).user_data = buf as *mut libc::c_void;
     return res;
 }
-#[no_mangle]
-pub unsafe extern "C" fn dav1d_ref_wrap(
+
+pub unsafe fn dav1d_ref_wrap(
     ptr: *const uint8_t,
     mut free_callback: Option<unsafe extern "C" fn(*const uint8_t, *mut libc::c_void) -> ()>,
     user_data: *mut libc::c_void,
@@ -114,8 +114,8 @@ pub unsafe extern "C" fn dav1d_ref_wrap(
     (*res).user_data = user_data;
     return res;
 }
-#[no_mangle]
-pub unsafe extern "C" fn dav1d_ref_dec(pref: *mut *mut Dav1dRef) {
+
+pub unsafe fn dav1d_ref_dec(pref: *mut *mut Dav1dRef) {
     if pref.is_null() {
         unreachable!();
     }
diff --git a/src/refmvs.rs b/src/refmvs.rs
index f23d3689b..0df8b9c7d 100644
--- a/src/refmvs.rs
+++ b/src/refmvs.rs
@@ -1124,8 +1124,7 @@ pub unsafe fn dav1d_refmvs_find(
 
 // cache the current tile/sbrow (or frame/sbrow)'s projectable motion vectors
 // into buffers for use in future frame's temporal MV prediction
-#[no_mangle]
-pub unsafe extern "C" fn dav1d_refmvs_save_tmvs(
+pub unsafe fn dav1d_refmvs_save_tmvs(
     dsp: *const Dav1dRefmvsDSPContext,
     rt: *const refmvs_tile,
     col_start8: libc::c_int,
@@ -1158,8 +1157,7 @@ pub unsafe extern "C" fn dav1d_refmvs_save_tmvs(
     );
 }
 
-#[no_mangle]
-pub unsafe extern "C" fn dav1d_refmvs_tile_sbrow_init(
+pub unsafe fn dav1d_refmvs_tile_sbrow_init(
     rt: *mut refmvs_tile,
     rf: *const refmvs_frame,
     tile_col_start4: libc::c_int,
@@ -1405,8 +1403,8 @@ pub unsafe extern "C" fn save_tmvs_c(
         y += 1;
     }
 }
-#[no_mangle]
-pub unsafe extern "C" fn dav1d_refmvs_init_frame(
+
+pub unsafe fn dav1d_refmvs_init_frame(
     rf: *mut refmvs_frame,
     seq_hdr: *const Dav1dSequenceHeader,
     frm_hdr: *const Dav1dFrameHeader,
@@ -1582,15 +1580,15 @@ pub unsafe extern "C" fn dav1d_refmvs_init_frame(
     (*rf).use_ref_frame_mvs = ((*rf).n_mfmvs > 0) as libc::c_int;
     return 0 as libc::c_int;
 }
-#[no_mangle]
-pub unsafe extern "C" fn dav1d_refmvs_init(rf: *mut refmvs_frame) {
+
+pub unsafe fn dav1d_refmvs_init(rf: *mut refmvs_frame) {
     (*rf).r = 0 as *mut refmvs_block;
     (*rf).r_stride = 0 as libc::c_int as ptrdiff_t;
     (*rf).rp_proj = 0 as *mut refmvs_temporal_block;
     (*rf).rp_stride = 0 as libc::c_int as ptrdiff_t;
 }
-#[no_mangle]
-pub unsafe extern "C" fn dav1d_refmvs_clear(rf: *mut refmvs_frame) {
+
+pub unsafe fn dav1d_refmvs_clear(rf: *mut refmvs_frame) {
     if !((*rf).r).is_null() {
         dav1d_freep_aligned(&mut (*rf).r as *mut *mut refmvs_block as *mut libc::c_void);
     }
diff --git a/src/thread_task.rs b/src/thread_task.rs
index 0e81e3c92..23ea27cd5 100644
--- a/src/thread_task.rs
+++ b/src/thread_task.rs
@@ -1,6 +1,5 @@
 use crate::include::stddef::*;
 use crate::include::stdint::*;
-use crate::src::cdf::CdfContext;
 use ::libc;
 use cfg_if::cfg_if;
 extern "C" {
@@ -19,11 +18,6 @@ extern "C" {
             fn pthread_setname_np(name: *const libc::c_char);
         }
     }
-    fn dav1d_cdf_thread_update(
-        hdr: *const Dav1dFrameHeader,
-        dst: *mut CdfContext,
-        src: *const CdfContext,
-    );
     fn dav1d_decode_frame_init(f: *mut Dav1dFrameContext) -> libc::c_int;
     fn dav1d_decode_frame_init_cdf(f: *mut Dav1dFrameContext) -> libc::c_int;
     fn dav1d_decode_tile_sbrow(t: *mut Dav1dTaskContext) -> libc::c_int;
@@ -604,6 +598,7 @@ pub type generate_grain_uv_fn = Option<
 >;
 pub type generate_grain_y_fn =
     Option<unsafe extern "C" fn(*mut [entry; 82], *const Dav1dFilmGrainData) -> ()>;
+use crate::src::cdf::dav1d_cdf_thread_update;
 use crate::src::cdf::CdfThreadContext;
 
 use crate::src::internal::Dav1dContext_frame_thread;
@@ -1506,7 +1501,7 @@ unsafe extern "C" fn delayed_fg_task(c: *const Dav1dContext, ttd: *mut TaskThrea
         pthread_cond_signal(&mut (*ttd).delayed_fg.cond);
     }
 }
-#[no_mangle]
+
 pub unsafe extern "C" fn dav1d_worker_task(mut data: *mut libc::c_void) -> *mut libc::c_void {
     let mut flush = 0;
     let mut error_0 = 0;
diff --git a/tools/dav1d_cli_parse.rs b/tools/dav1d_cli_parse.rs
index 3524223c3..2513432c2 100644
--- a/tools/dav1d_cli_parse.rs
+++ b/tools/dav1d_cli_parse.rs
@@ -26,7 +26,6 @@ extern "C" {
     fn memset(_: *mut libc::c_void, _: libc::c_int, _: libc::c_ulong) -> *mut libc::c_void;
     fn dav1d_version() -> *const libc::c_char;
     fn dav1d_default_settings(s: *mut Dav1dSettings);
-    fn dav1d_set_cpu_flags_mask(mask: libc::c_uint);
 }
 #[derive(Copy, Clone)]
 #[repr(C)]
@@ -50,6 +49,7 @@ use rav1d::include::dav1d::dav1d::DAV1D_INLOOPFILTER_DEBLOCK;
 use rav1d::include::dav1d::dav1d::DAV1D_INLOOPFILTER_NONE;
 use rav1d::include::dav1d::dav1d::DAV1D_INLOOPFILTER_RESTORATION;
 use rav1d::include::dav1d::picture::Dav1dPicAllocator;
+use rav1d::src::cpu::dav1d_set_cpu_flags_mask;
 #[derive(Copy, Clone)]
 #[repr(C)]
 pub struct Dav1dSettings {