summaryrefslogtreecommitdiff
path: root/third_party/libopenjpeg20/0003-dwt-decode.patch
diff options
context:
space:
mode:
Diffstat (limited to 'third_party/libopenjpeg20/0003-dwt-decode.patch')
-rw-r--r--third_party/libopenjpeg20/0003-dwt-decode.patch103
1 files changed, 76 insertions, 27 deletions
diff --git a/third_party/libopenjpeg20/0003-dwt-decode.patch b/third_party/libopenjpeg20/0003-dwt-decode.patch
index 48e1b8956e..631219a8b3 100644
--- a/third_party/libopenjpeg20/0003-dwt-decode.patch
+++ b/third_party/libopenjpeg20/0003-dwt-decode.patch
@@ -1,26 +1,25 @@
diff --git a/third_party/libopenjpeg20/dwt.c b/third_party/libopenjpeg20/dwt.c
-index f3994ca7b..8bf23b6d6 100644
+index 83c148002..1455ee84a 100644
--- a/third_party/libopenjpeg20/dwt.c
+++ b/third_party/libopenjpeg20/dwt.c
-@@ -63,9 +63,6 @@
+@@ -62,8 +62,6 @@
+
/** @defgroup DWT DWT - Implementation of a discrete wavelet transform */
/*@{*/
-
-#define OPJ_WS(i) v->mem[(i)*2]
-#define OPJ_WD(i) v->mem[(1+(i)*2)]
--
+
#ifdef __AVX2__
/** Number of int32 values in a AVX2 register */
- #define VREG_INT_COUNT 8
-@@ -82,6 +79,7 @@
+@@ -81,6 +79,7 @@
typedef struct dwt_local {
OPJ_INT32* mem;
+ OPJ_SIZE_T mem_count;
- OPJ_INT32 dn;
- OPJ_INT32 sn;
- OPJ_INT32 cas;
-@@ -129,13 +127,15 @@ static void opj_dwt_deinterleave_v(OPJ_INT32 *a, OPJ_INT32 *b, OPJ_INT32 dn,
+ OPJ_INT32 dn; /* number of elements in high pass band */
+ OPJ_INT32 sn; /* number of elements in low pass band */
+ OPJ_INT32 cas; /* 0 = start on even coord, 1 = start on odd coord */
+@@ -132,13 +131,14 @@ static void opj_dwt_deinterleave_v(OPJ_INT32 *a, OPJ_INT32 *b, OPJ_INT32 dn,
/**
Forward 5-3 wavelet transform in 1-D
*/
@@ -28,7 +27,6 @@ index f3994ca7b..8bf23b6d6 100644
- OPJ_INT32 cas);
+static void opj_dwt_encode_1(OPJ_INT32 *a, OPJ_SIZE_T a_count, OPJ_INT32 dn,
+ OPJ_INT32 sn, OPJ_INT32 cas);
-+
/**
Forward 9-7 wavelet transform in 1-D
*/
@@ -37,16 +35,20 @@ index f3994ca7b..8bf23b6d6 100644
+static void opj_dwt_encode_1_real(OPJ_INT32 *a, OPJ_SIZE_T a_count,
+ OPJ_INT32 dn, OPJ_INT32 sn, OPJ_INT32 cas);
+
+
/**
Explicit calculation of the Quantization Stepsizes
- */
-@@ -145,10 +145,10 @@ static void opj_dwt_encode_stepsize(OPJ_INT32 stepsize, OPJ_INT32 numbps,
+@@ -149,14 +149,14 @@ static void opj_dwt_encode_stepsize(OPJ_INT32 stepsize, OPJ_INT32 numbps,
Inverse wavelet transform in 2-D.
*/
static OPJ_BOOL opj_dwt_decode_tile(opj_thread_pool_t* tp,
- opj_tcd_tilecomp_t* tilec, OPJ_UINT32 i);
+ const opj_tcd_tilecomp_t* tilec, OPJ_UINT32 i);
+ static OPJ_BOOL opj_dwt_decode_partial_tile(
+ opj_tcd_tilecomp_t* tilec,
+ OPJ_UINT32 numres);
+
-static OPJ_BOOL opj_dwt_encode_procedure(opj_tcd_tilecomp_t * tilec,
- void (*p_function)(OPJ_INT32 *, OPJ_INT32, OPJ_INT32, OPJ_INT32));
+static OPJ_BOOL opj_dwt_encode_procedure(const opj_tcd_tilecomp_t * tilec,
@@ -54,7 +56,7 @@ index f3994ca7b..8bf23b6d6 100644
static OPJ_UINT32 opj_dwt_max_resolution(opj_tcd_resolution_t* OPJ_RESTRICT r,
OPJ_UINT32 i);
-@@ -184,13 +184,20 @@ static void opj_v4dwt_decode_step2(opj_v4_t* l, opj_v4_t* w, OPJ_INT32 k,
+@@ -205,13 +205,20 @@ static void opj_v4dwt_decode_step2(opj_v4_t* l, opj_v4_t* w,
/*@}*/
@@ -81,7 +83,7 @@ index f3994ca7b..8bf23b6d6 100644
/* <summary> */
/* This table contains the norms of the 5-3 wavelets for different bands. */
-@@ -319,8 +326,8 @@ static void opj_dwt_interleave_v(const opj_dwt_t* v, OPJ_INT32 *a, OPJ_INT32 x)
+@@ -344,8 +351,8 @@ static void opj_dwt_interleave_v(const opj_dwt_t* v, OPJ_INT32 *a, OPJ_INT32 x)
/* <summary> */
/* Forward 5-3 wavelet transform in 1-D. */
/* </summary> */
@@ -92,7 +94,7 @@ index f3994ca7b..8bf23b6d6 100644
{
OPJ_INT32 i;
-@@ -351,8 +358,8 @@ static void opj_dwt_encode_1(OPJ_INT32 *a, OPJ_INT32 dn, OPJ_INT32 sn,
+@@ -376,8 +383,8 @@ static void opj_dwt_encode_1(OPJ_INT32 *a, OPJ_INT32 dn, OPJ_INT32 sn,
/* <summary> */
/* Inverse 5-3 wavelet transform in 1-D. */
/* </summary> */
@@ -103,7 +105,7 @@ index f3994ca7b..8bf23b6d6 100644
{
OPJ_INT32 i;
-@@ -381,7 +388,7 @@ static void opj_dwt_decode_1_(OPJ_INT32 *a, OPJ_INT32 dn, OPJ_INT32 sn,
+@@ -406,7 +413,7 @@ static void opj_dwt_decode_1_(OPJ_INT32 *a, OPJ_INT32 dn, OPJ_INT32 sn,
static void opj_dwt_decode_1(const opj_dwt_t *v)
{
@@ -112,7 +114,7 @@ index f3994ca7b..8bf23b6d6 100644
}
#endif /* STANDARD_SLOW_VERSION */
-@@ -1010,8 +1017,8 @@ static void opj_idwt53_v(const opj_dwt_t *dwt,
+@@ -1037,8 +1044,8 @@ static void opj_idwt53_v(const opj_dwt_t *dwt,
/* <summary> */
/* Forward 9-7 wavelet transform in 1-D. */
/* </summary> */
@@ -123,7 +125,7 @@ index f3994ca7b..8bf23b6d6 100644
{
OPJ_INT32 i;
if (!cas) {
-@@ -1079,8 +1086,8 @@ static void opj_dwt_encode_stepsize(OPJ_INT32 stepsize, OPJ_INT32 numbps,
+@@ -1106,8 +1113,8 @@ static void opj_dwt_encode_stepsize(OPJ_INT32 stepsize, OPJ_INT32 numbps,
/* <summary> */
/* Forward 5-3 wavelet transform in 2-D. */
/* </summary> */
@@ -134,15 +136,15 @@ index f3994ca7b..8bf23b6d6 100644
{
OPJ_INT32 i, j, k;
OPJ_INT32 *a = 00;
-@@ -1090,6 +1097,7 @@ static INLINE OPJ_BOOL opj_dwt_encode_procedure(opj_tcd_tilecomp_t * tilec,
+@@ -1117,6 +1124,7 @@ static INLINE OPJ_BOOL opj_dwt_encode_procedure(opj_tcd_tilecomp_t * tilec,
OPJ_INT32 rw; /* width of the resolution level computed */
OPJ_INT32 rh; /* height of the resolution level computed */
-+ size_t l_data_count;
- size_t l_data_size;
++ OPJ_SIZE_T l_data_count;
+ OPJ_SIZE_T l_data_size;
opj_tcd_resolution_t * l_cur_res = 0;
-@@ -1102,13 +1110,13 @@ static INLINE OPJ_BOOL opj_dwt_encode_procedure(opj_tcd_tilecomp_t * tilec,
+@@ -1129,13 +1137,13 @@ static INLINE OPJ_BOOL opj_dwt_encode_procedure(opj_tcd_tilecomp_t * tilec,
l_cur_res = tilec->resolutions + l;
l_last_res = l_cur_res - 1;
@@ -159,7 +161,7 @@ index f3994ca7b..8bf23b6d6 100644
bj = (OPJ_INT32*)opj_malloc(l_data_size);
/* l_data_size is equal to 0 when numresolutions == 1 but bj is not used */
/* in that case, so do not error out */
-@@ -1140,7 +1148,7 @@ static INLINE OPJ_BOOL opj_dwt_encode_procedure(opj_tcd_tilecomp_t * tilec,
+@@ -1167,7 +1175,7 @@ static INLINE OPJ_BOOL opj_dwt_encode_procedure(opj_tcd_tilecomp_t * tilec,
bj[k] = aj[k * w];
}
@@ -168,7 +170,7 @@ index f3994ca7b..8bf23b6d6 100644
opj_dwt_deinterleave_v(bj, aj, dn, sn, w, cas_col);
}
-@@ -1153,7 +1161,7 @@ static INLINE OPJ_BOOL opj_dwt_encode_procedure(opj_tcd_tilecomp_t * tilec,
+@@ -1180,7 +1188,7 @@ static INLINE OPJ_BOOL opj_dwt_encode_procedure(opj_tcd_tilecomp_t * tilec,
for (k = 0; k < rw; k++) {
bj[k] = aj[k];
}
@@ -177,7 +179,7 @@ index f3994ca7b..8bf23b6d6 100644
opj_dwt_deinterleave_h(bj, aj, dn, sn, cas_row);
}
-@@ -1332,7 +1340,7 @@ static void opj_dwt_decode_v_func(void* user_data, opj_tls_t* tls)
+@@ -1379,7 +1387,7 @@ static void opj_dwt_decode_v_func(void* user_data, opj_tls_t* tls)
/* Inverse wavelet transform in 2-D. */
/* </summary> */
static OPJ_BOOL opj_dwt_decode_tile(opj_thread_pool_t* tp,
@@ -186,7 +188,7 @@ index f3994ca7b..8bf23b6d6 100644
{
opj_dwt_t h;
opj_dwt_t v;
-@@ -1352,22 +1360,23 @@ static OPJ_BOOL opj_dwt_decode_tile(opj_thread_pool_t* tp,
+@@ -1401,22 +1409,23 @@ static OPJ_BOOL opj_dwt_decode_tile(opj_thread_pool_t* tp,
return OPJ_TRUE;
}
num_threads = opj_thread_pool_get_thread_count(tp);
@@ -213,3 +215,50 @@ index f3994ca7b..8bf23b6d6 100644
v.mem = h.mem;
while (--numres) {
+@@ -1594,7 +1603,8 @@ static void opj_dwt_interleave_partial_v(OPJ_INT32 *dest,
+ OPJ_UNUSED(ret);
+ }
+
+-static void opj_dwt_decode_partial_1(OPJ_INT32 *a, OPJ_INT32 dn, OPJ_INT32 sn,
++static void opj_dwt_decode_partial_1(OPJ_INT32 *a, OPJ_SIZE_T a_count,
++ OPJ_INT32 dn, OPJ_INT32 sn,
+ OPJ_INT32 cas,
+ OPJ_INT32 win_l_x0,
+ OPJ_INT32 win_l_x1,
+@@ -1974,16 +1984,16 @@ static OPJ_BOOL opj_dwt_decode_partial_tile(
+ opj_sparse_array_int32_free(sa);
+ return OPJ_TRUE;
+ }
+- h_mem_size = opj_dwt_max_resolution(tr, numres);
++ h.mem_count = opj_dwt_max_resolution(tr, numres);
+ /* overflow check */
+ /* in vertical pass, we process 4 columns at a time */
+- if (h_mem_size > (SIZE_MAX / (4 * sizeof(OPJ_INT32)))) {
++ if (h.mem_count > (SIZE_MAX / (4 * sizeof(OPJ_INT32)))) {
+ /* FIXME event manager error callback */
+ opj_sparse_array_int32_free(sa);
+ return OPJ_FALSE;
+ }
+
+- h_mem_size *= 4 * sizeof(OPJ_INT32);
++ h_mem_size = h.mem_count * 4 * sizeof(OPJ_INT32);
+ h.mem = (OPJ_INT32*)opj_aligned_32_malloc(h_mem_size);
+ if (! h.mem) {
+ /* FIXME event manager error callback */
+@@ -1991,6 +2001,7 @@ static OPJ_BOOL opj_dwt_decode_partial_tile(
+ return OPJ_FALSE;
+ }
+
++ v.mem_count = h.mem_count;
+ v.mem = h.mem;
+
+ for (resno = 1; resno < numres; resno ++) {
+@@ -2101,7 +2112,7 @@ static OPJ_BOOL opj_dwt_decode_partial_tile(
+ win_ll_x1,
+ win_hl_x0,
+ win_hl_x1);
+- opj_dwt_decode_partial_1(h.mem, h.dn, h.sn, h.cas,
++ opj_dwt_decode_partial_1(h.mem, h.mem_count, h.dn, h.sn, h.cas,
+ (OPJ_INT32)win_ll_x0,
+ (OPJ_INT32)win_ll_x1,
+ (OPJ_INT32)win_hl_x0,
generated by cgit v1.2.3 (git 2.25.1) at 2024-06-02 07:59:57 +0000