T1 encoder: speed-up by aggressive inlining and more cache friendly data organization

~ 9% speed improvement seen on 10980x10980 uint16 image, T36JTT_20160914T074612_B02.tif
opj_compress time from 17.2s to 15.8s

5 files changed, 460 insertions, 392 deletions

diff --git a/src/lib/openjp2/mqc.c b/src/lib/openjp2/mqc.c
index 6299b171..4cbfabd0 100644
--- a/src/lib/openjp2/mqc.c
+++ b/src/lib/openjp2/mqc.c
@@ -47,27 +47,6 @@
 /*@{*/
 
 /**
-Output a byte, doing bit-stuffing if necessary.
-After a 0xff byte, the next byte must be smaller than 0x90.
-@param mqc MQC handle
-*/
-static void opj_mqc_byteout(opj_mqc_t *mqc);
-/**
-Renormalize mqc->a and mqc->c while encoding, so that mqc->a stays between 0x8000 and 0x10000
-@param mqc MQC handle
-*/
-static void opj_mqc_renorme(opj_mqc_t *mqc);
-/**
-Encode the most probable symbol
-@param mqc MQC handle
-*/
-static void opj_mqc_codemps(opj_mqc_t *mqc);
-/**
-Encode the most least symbol
-@param mqc MQC handle
-*/
-static void opj_mqc_codelps(opj_mqc_t *mqc);
-/**
 Fill mqc->c with 1's for flushing
 @param mqc MQC handle
 */
@@ -182,80 +161,6 @@ static const opj_mqc_state_t mqc_states[47 * 2] = {
 ==========================================================
 */
 
-static void opj_mqc_byteout(opj_mqc_t *mqc)
-{
-    /* bp is initialized to start - 1 in opj_mqc_init_enc() */
-    /* but this is safe, see opj_tcd_code_block_enc_allocate_data() */
-    assert(mqc->bp >= mqc->start - 1);
-    if (*mqc->bp == 0xff) {
-        mqc->bp++;
-        *mqc->bp = (OPJ_BYTE)(mqc->c >> 20);
-        mqc->c &= 0xfffff;
-        mqc->ct = 7;
-    } else {
-        if ((mqc->c & 0x8000000) == 0) {
-            mqc->bp++;
-            *mqc->bp = (OPJ_BYTE)(mqc->c >> 19);
-            mqc->c &= 0x7ffff;
-            mqc->ct = 8;
-        } else {
-            (*mqc->bp)++;
-            if (*mqc->bp == 0xff) {
-                mqc->c &= 0x7ffffff;
-                mqc->bp++;
-                *mqc->bp = (OPJ_BYTE)(mqc->c >> 20);
-                mqc->c &= 0xfffff;
-                mqc->ct = 7;
-            } else {
-                mqc->bp++;
-                *mqc->bp = (OPJ_BYTE)(mqc->c >> 19);
-                mqc->c &= 0x7ffff;
-                mqc->ct = 8;
-            }
-        }
-    }
-}
-
-static void opj_mqc_renorme(opj_mqc_t *mqc)
-{
-    do {
-        mqc->a <<= 1;
-        mqc->c <<= 1;
-        mqc->ct--;
-        if (mqc->ct == 0) {
-            opj_mqc_byteout(mqc);
-        }
-    } while ((mqc->a & 0x8000) == 0);
-}
-
-static void opj_mqc_codemps(opj_mqc_t *mqc)
-{
-    mqc->a -= (*mqc->curctx)->qeval;
-    if ((mqc->a & 0x8000) == 0) {
-        if (mqc->a < (*mqc->curctx)->qeval) {
-            mqc->a = (*mqc->curctx)->qeval;
-        } else {
-            mqc->c += (*mqc->curctx)->qeval;
-        }
-        *mqc->curctx = (*mqc->curctx)->nmps;
-        opj_mqc_renorme(mqc);
-    } else {
-        mqc->c += (*mqc->curctx)->qeval;
-    }
-}
-
-static void opj_mqc_codelps(opj_mqc_t *mqc)
-{
-    mqc->a -= (*mqc->curctx)->qeval;
-    if (mqc->a < (*mqc->curctx)->qeval) {
-        mqc->c += (*mqc->curctx)->qeval;
-    } else {
-        mqc->a = (*mqc->curctx)->qeval;
-    }
-    *mqc->curctx = (*mqc->curctx)->nlps;
-    opj_mqc_renorme(mqc);
-}
-
 static void opj_mqc_setbits(opj_mqc_t *mqc)
 {
     OPJ_UINT32 tempc = mqc->c + mqc->a;
@@ -303,14 +208,6 @@ void opj_mqc_init_enc(opj_mqc_t *mqc, OPJ_BYTE *bp)
     mqc->end_of_byte_stream_counter = 0;
 }
 
-void opj_mqc_encode(opj_mqc_t *mqc, OPJ_UINT32 d)
-{
-    if ((*mqc->curctx)->mps == d) {
-        opj_mqc_codemps(mqc);
-    } else {
-        opj_mqc_codelps(mqc);
-    }
-}
 
 void opj_mqc_flush(opj_mqc_t *mqc)
 {
@@ -329,8 +226,6 @@ void opj_mqc_flush(opj_mqc_t *mqc)
     }
 }
 
-#define BYPASS_CT_INIT  0xDEADBEEF
-
 void opj_mqc_bypass_init_enc(opj_mqc_t *mqc)
 {
     /* This function is normally called after at least one opj_mqc_flush() */
@@ -475,6 +370,43 @@ void opj_mqc_erterm_enc(opj_mqc_t *mqc)
     }
 }
 
+static INLINE void opj_mqc_renorme(opj_mqc_t *mqc)
+{
+    opj_mqc_renorme_macro(mqc, mqc->a, mqc->c, mqc->ct);
+}
+
+/**
+Encode the most probable symbol
+@param mqc MQC handle
+*/
+static INLINE void opj_mqc_codemps(opj_mqc_t *mqc)
+{
+    opj_mqc_codemps_macro(mqc, mqc->curctx, mqc->a, mqc->c, mqc->ct);
+}
+
+/**
+Encode the most least symbol
+@param mqc MQC handle
+*/
+static INLINE void opj_mqc_codelps(opj_mqc_t *mqc)
+{
+    opj_mqc_codelps_macro(mqc, mqc->curctx, mqc->a, mqc->c, mqc->ct);
+}
+
+/**
+Encode a symbol using the MQ-coder
+@param mqc MQC handle
+@param d The symbol to be encoded (0 or 1)
+*/
+static INLINE void opj_mqc_encode(opj_mqc_t *mqc, OPJ_UINT32 d)
+{
+    if ((*mqc->curctx)->mps == d) {
+        opj_mqc_codemps(mqc);
+    } else {
+        opj_mqc_codelps(mqc);
+    }
+}
+
 void opj_mqc_segmark_enc(opj_mqc_t *mqc)
 {
     OPJ_UINT32 i;
@@ -557,4 +489,36 @@ void opj_mqc_setstate(opj_mqc_t *mqc, OPJ_UINT32 ctxno, OPJ_UINT32 msb,
     mqc->ctxs[ctxno] = &mqc_states[msb + (OPJ_UINT32)(prob << 1)];
 }
 
-
+void opj_mqc_byteout(opj_mqc_t *mqc)
+{
+    /* bp is initialized to start - 1 in opj_mqc_init_enc() */
+    /* but this is safe, see opj_tcd_code_block_enc_allocate_data() */
+    assert(mqc->bp >= mqc->start - 1);
+    if (*mqc->bp == 0xff) {
+        mqc->bp++;
+        *mqc->bp = (OPJ_BYTE)(mqc->c >> 20);
+        mqc->c &= 0xfffff;
+        mqc->ct = 7;
+    } else {
+        if ((mqc->c & 0x8000000) == 0) {
+            mqc->bp++;
+            *mqc->bp = (OPJ_BYTE)(mqc->c >> 19);
+            mqc->c &= 0x7ffff;
+            mqc->ct = 8;
+        } else {
+            (*mqc->bp)++;
+            if (*mqc->bp == 0xff) {
+                mqc->c &= 0x7ffffff;
+                mqc->bp++;
+                *mqc->bp = (OPJ_BYTE)(mqc->c >> 20);
+                mqc->c &= 0xfffff;
+                mqc->ct = 7;
+            } else {
+                mqc->bp++;
+                *mqc->bp = (OPJ_BYTE)(mqc->c >> 19);
+                mqc->c &= 0x7ffff;
+                mqc->ct = 8;
+            }
+        }
+    }
+}
\ No newline at end of file
diff --git a/src/lib/openjp2/mqc.h b/src/lib/openjp2/mqc.h
index 69a2a79d..9850fed0 100644
--- a/src/lib/openjp2/mqc.h
+++ b/src/lib/openjp2/mqc.h
@@ -96,6 +96,8 @@ typedef struct opj_mqc {
     OPJ_BYTE backup[OPJ_COMMON_CBLK_DATA_EXTRA];
 } opj_mqc_t;
 
+#define BYPASS_CT_INIT  0xDEADBEEF
+
 #include "mqc_inl.h"
 
 /** @name Exported functions */
@@ -135,12 +137,7 @@ Set the current context used for coding/decoding
 @param ctxno Number that identifies the context
 */
 #define opj_mqc_setcurctx(mqc, ctxno)   (mqc)->curctx = &(mqc)->ctxs[(OPJ_UINT32)(ctxno)]
-/**
-Encode a symbol using the MQ-coder
-@param mqc MQC handle
-@param d The symbol to be encoded (0 or 1)
-*/
-void opj_mqc_encode(opj_mqc_t *mqc, OPJ_UINT32 d);
+
 /**
 Flush the encoder, so that all remaining data is written
 @param mqc MQC handle
diff --git a/src/lib/openjp2/mqc_inl.h b/src/lib/openjp2/mqc_inl.h
index 310a3287..0031b94b 100644
--- a/src/lib/openjp2/mqc_inl.h
+++ b/src/lib/openjp2/mqc_inl.h
@@ -156,13 +156,13 @@ static INLINE OPJ_UINT32 opj_mqc_raw_decode(opj_mqc_t *mqc)
     } \
 }
 
-#define DOWNLOAD_MQC_VARIABLES(mqc, curctx, c, a, ct) \
+#define DOWNLOAD_MQC_VARIABLES(mqc, curctx, a, c, ct) \
         register const opj_mqc_state_t **curctx = mqc->curctx; \
         register OPJ_UINT32 c = mqc->c; \
         register OPJ_UINT32 a = mqc->a; \
         register OPJ_UINT32 ct = mqc->ct
 
-#define UPLOAD_MQC_VARIABLES(mqc, curctx, c, a, ct) \
+#define UPLOAD_MQC_VARIABLES(mqc, curctx, a, c, ct) \
         mqc->curctx = curctx; \
         mqc->c = c; \
         mqc->a = a; \
@@ -193,4 +193,90 @@ Decode a symbol
 #define opj_mqc_decode(d, mqc) \
     opj_mqc_decode_macro(d, mqc, mqc->curctx, mqc->a, mqc->c, mqc->ct)
 
+/**
+Output a byte, doing bit-stuffing if necessary.
+After a 0xff byte, the next byte must be smaller than 0x90.
+@param mqc MQC handle
+*/
+void opj_mqc_byteout(opj_mqc_t *mqc);
+
+/**
+Renormalize mqc->a and mqc->c while encoding, so that mqc->a stays between 0x8000 and 0x10000
+@param mqc MQC handle
+@param a_ value of mqc->a
+@param c_ value of mqc->c_
+@param ct_ value of mqc->ct_
+*/
+#define opj_mqc_renorme_macro(mqc, a_, c_, ct_) \
+{ \
+    do { \
+        a_ <<= 1; \
+        c_ <<= 1; \
+        ct_--; \
+        if (ct_ == 0) { \
+            mqc->c = c_; \
+            opj_mqc_byteout(mqc); \
+            c_ = mqc->c; \
+            ct_ = mqc->ct; \
+        } \
+    } while( (a_ & 0x8000) == 0); \
+}
+
+#define opj_mqc_codemps_macro(mqc, curctx, a, c, ct) \
+{ \
+    a -= (*curctx)->qeval; \
+    if ((a & 0x8000) == 0) { \
+        if (a < (*curctx)->qeval) { \
+            a = (*curctx)->qeval; \
+        } else { \
+            c += (*curctx)->qeval; \
+        } \
+        *curctx = (*curctx)->nmps; \
+        opj_mqc_renorme_macro(mqc, a, c, ct); \
+    } else { \
+        c += (*curctx)->qeval; \
+    } \
+}
+
+#define opj_mqc_codelps_macro(mqc, curctx, a, c, ct) \
+{ \
+    a -= (*curctx)->qeval; \
+    if (a < (*curctx)->qeval) { \
+        c += (*curctx)->qeval; \
+    } else { \
+        a = (*curctx)->qeval; \
+    } \
+    *curctx = (*curctx)->nlps; \
+    opj_mqc_renorme_macro(mqc, a, c, ct); \
+}
+
+#define opj_mqc_encode_macro(mqc, curctx, a, c, ct, d) \
+{ \
+    if ((*curctx)->mps == (d)) { \
+        opj_mqc_codemps_macro(mqc, curctx, a, c, ct); \
+    } else { \
+        opj_mqc_codelps_macro(mqc, curctx, a, c, ct); \
+    } \
+}
+
+
+#define opj_mqc_bypass_enc_macro(mqc, c, ct, d) \
+{\
+    if (ct == BYPASS_CT_INIT) {\
+        ct = 8;\
+    }\
+    ct--;\
+    c = c + ((d) << ct);\
+    if (ct == 0) {\
+        *mqc->bp = (OPJ_BYTE)c;\
+        ct = 8;\
+        /* If the previous byte was 0xff, make sure that the next msb is 0 */ \
+        if (*mqc->bp == 0xff) {\
+            ct = 7;\
+        }\
+        mqc->bp++;\
+        c = 0;\
+    }\
+}
+
 #endif /* OPJ_MQC_INL_H */
diff --git a/src/lib/openjp2/t1.c b/src/lib/openjp2/t1.c
index 937f420a..92030b21 100644
--- a/src/lib/openjp2/t1.c
+++ b/src/lib/openjp2/t1.c
@@ -61,6 +61,13 @@
 
 #define opj_t1_setcurctx(curctx, ctxno)  curctx = &(mqc)->ctxs[(OPJ_UINT32)(ctxno)]
 
+/* Macros to deal with signed integer with just MSB bit set for
+ * negative values (smr = signed magnitude representation) */
+#define opj_smr_abs(x)  (((OPJ_UINT32)(x)) & 0x7FFFFFFFU)
+#define opj_smr_sign(x) (((OPJ_UINT32)(x)) >> 31)
+#define opj_to_smr(x)   ((x) >= 0 ? (OPJ_UINT32)(x) : ((OPJ_UINT32)(-x) | 0x80000000U))
+
+
 /** @name Local static functions */
 /*@{*/
 
@@ -329,61 +336,53 @@ static INLINE void opj_t1_update_flags(opj_flag_t *flagsp, OPJ_UINT32 ci,
 /**
 Encode significant pass
 */
-static INLINE void opj_t1_enc_sigpass_step(opj_t1_t *t1,
-        opj_flag_t *flagsp,
-        OPJ_INT32 *datap,
-        OPJ_INT32 bpno,
-        OPJ_INT32 one,
-        OPJ_INT32 *nmsedec,
-        OPJ_BYTE type,
-        OPJ_UINT32 ci,
-        OPJ_UINT32 vsc)
-{
-    OPJ_UINT32 v;
-
-    opj_mqc_t *mqc = &(t1->mqc);   /* MQC component */
-
-    OPJ_UINT32 const flags = *flagsp;
-
-    if ((flags & ((T1_SIGMA_THIS | T1_PI_THIS) << (ci * 3U))) == 0U &&
-            (flags & (T1_SIGMA_NEIGHBOURS << (ci * 3U))) != 0U) {
-        OPJ_UINT32 ctxt1 = opj_t1_getctxno_zc(mqc, flags >> (ci * 3U));
-        v = (opj_int_abs(*datap) & one) ? 1 : 0;
-#ifdef DEBUG_ENC_SIG
-        fprintf(stderr, "   ctxt1=%d\n", ctxt1);
-#endif
-        opj_mqc_setcurctx(mqc, ctxt1);
-        if (type == T1_TYPE_RAW) {  /* BYPASS/LAZY MODE */
-            opj_mqc_bypass_enc(mqc, v);
-        } else {
-            opj_mqc_encode(mqc, v);
-        }
-        if (v) {
-            OPJ_UINT32 lu = opj_t1_getctxtno_sc_or_spb_index(
-                                *flagsp,
-                                flagsp[-1], flagsp[1],
-                                ci);
-            OPJ_UINT32 ctxt2 = opj_t1_getctxno_sc(lu);
-            v = *datap < 0 ? 1U : 0U;
-            *nmsedec += opj_t1_getnmsedec_sig((OPJ_UINT32)opj_int_abs(*datap),
-                                              (OPJ_UINT32)bpno);
-#ifdef DEBUG_ENC_SIG
-            fprintf(stderr, "   ctxt2=%d\n", ctxt2);
-#endif
-            opj_mqc_setcurctx(mqc, ctxt2);
-            if (type == T1_TYPE_RAW) {  /* BYPASS/LAZY MODE */
-                opj_mqc_bypass_enc(mqc, v);
-            } else {
-                OPJ_UINT32 spb = opj_t1_getspb(lu);
-#ifdef DEBUG_ENC_SIG
-                fprintf(stderr, "   spb=%d\n", spb);
-#endif
-                opj_mqc_encode(mqc, v ^ spb);
-            }
-            opj_t1_update_flags(flagsp, ci, v, t1->w + 2, vsc);
-        }
-        *flagsp |= T1_PI_THIS << (ci * 3U);
-    }
+#define opj_t1_enc_sigpass_step_macro(mqc, curctx, a, c, ct, flagspIn, datapIn, bpno, one, nmsedec, type, ciIn, vscIn) \
+{ \
+    OPJ_UINT32 v; \
+    const OPJ_UINT32 ci = (ciIn); \
+    const OPJ_UINT32 vsc = (vscIn); \
+    const OPJ_INT32* l_datap = (datapIn); \
+    opj_flag_t* flagsp = (flagspIn); \
+    OPJ_UINT32 const flags = *flagsp; \
+    if ((flags & ((T1_SIGMA_THIS | T1_PI_THIS) << (ci * 3U))) == 0U && \
+            (flags & (T1_SIGMA_NEIGHBOURS << (ci * 3U))) != 0U) { \
+        OPJ_UINT32 ctxt1 = opj_t1_getctxno_zc(mqc, flags >> (ci * 3U)); \
+        v = (opj_smr_abs(*l_datap) & (OPJ_UINT32)one) ? 1 : 0; \
+/* #ifdef DEBUG_ENC_SIG */ \
+/*        fprintf(stderr, "   ctxt1=%d\n", ctxt1); */ \
+/* #endif */ \
+        opj_t1_setcurctx(curctx, ctxt1); \
+        if (type == T1_TYPE_RAW) {  /* BYPASS/LAZY MODE */ \
+            opj_mqc_bypass_enc_macro(mqc, c, ct, v); \
+        } else { \
+            opj_mqc_encode_macro(mqc, curctx, a, c, ct, v); \
+        } \
+        if (v) { \
+            OPJ_UINT32 lu = opj_t1_getctxtno_sc_or_spb_index( \
+                                *flagsp, \
+                                flagsp[-1], flagsp[1], \
+                                ci); \
+            OPJ_UINT32 ctxt2 = opj_t1_getctxno_sc(lu); \
+            v = opj_smr_sign(*l_datap); \
+            *nmsedec += opj_t1_getnmsedec_sig(opj_smr_abs(*l_datap), \
+                                              (OPJ_UINT32)bpno); \
+/* #ifdef DEBUG_ENC_SIG */ \
+/*            fprintf(stderr, "   ctxt2=%d\n", ctxt2); */ \
+/* #endif */ \
+            opj_t1_setcurctx(curctx, ctxt2); \
+            if (type == T1_TYPE_RAW) {  /* BYPASS/LAZY MODE */ \
+                opj_mqc_bypass_enc_macro(mqc, c, ct, v); \
+            } else { \
+                OPJ_UINT32 spb = opj_t1_getspb(lu); \
+/* #ifdef DEBUG_ENC_SIG */ \
+/*                fprintf(stderr, "   spb=%d\n", spb); */ \
+/* #endif */ \
+                opj_mqc_encode_macro(mqc, curctx, a, c, ct, v ^ spb); \
+            } \
+            opj_t1_update_flags(flagsp, ci, v, t1->w + 2, vsc); \
+        } \
+        *flagsp |= T1_PI_THIS << (ci * 3U); \
+    } \
 }
 
 static INLINE void opj_t1_dec_sigpass_step_raw(
@@ -464,63 +463,64 @@ static void opj_t1_enc_sigpass(opj_t1_t *t1,
     OPJ_INT32 const one = 1 << (bpno + T1_NMSEDEC_FRACBITS);
     opj_flag_t* f = &T1_FLAGS(0, 0);
     OPJ_UINT32 const extra = 2;
+    opj_mqc_t* mqc = &(t1->mqc);
+    DOWNLOAD_MQC_VARIABLES(mqc, curctx, a, c, ct);
+    const OPJ_INT32* datap = t1->data;
 
     *nmsedec = 0;
 #ifdef DEBUG_ENC_SIG
     fprintf(stderr, "enc_sigpass: bpno=%d\n", bpno);
 #endif
-    for (k = 0; k < (t1->h & ~3U); k += 4) {
+    for (k = 0; k < (t1->h & ~3U); k += 4, f += extra) {
+        const OPJ_UINT32 w = t1->w;
 #ifdef DEBUG_ENC_SIG
         fprintf(stderr, " k=%d\n", k);
 #endif
-        for (i = 0; i < t1->w; ++i) {
+        for (i = 0; i < w; ++i, ++f, datap += 4) {
 #ifdef DEBUG_ENC_SIG
             fprintf(stderr, " i=%d\n", i);
 #endif
             if (*f == 0U) {
                 /* Nothing to do for any of the 4 data points */
-                f++;
                 continue;
             }
-            opj_t1_enc_sigpass_step(
-                t1,
+            opj_t1_enc_sigpass_step_macro(
+                mqc, curctx, a, c, ct,
                 f,
-                &t1->data[((k + 0) * t1->data_stride) + i],
+                &datap[0],
                 bpno,
                 one,
                 nmsedec,
                 type,
                 0, cblksty & J2K_CCP_CBLKSTY_VSC);
-            opj_t1_enc_sigpass_step(
-                t1,
+            opj_t1_enc_sigpass_step_macro(
+                mqc, curctx, a, c, ct,
                 f,
-                &t1->data[((k + 1) * t1->data_stride) + i],
+                &datap[1],
                 bpno,
                 one,
                 nmsedec,
                 type,
                 1, 0);
-            opj_t1_enc_sigpass_step(
-                t1,
+            opj_t1_enc_sigpass_step_macro(
+                mqc, curctx, a, c, ct,
                 f,
-                &t1->data[((k + 2) * t1->data_stride) + i],
+                &datap[2],
                 bpno,
                 one,
                 nmsedec,
                 type,
                 2, 0);
-            opj_t1_enc_sigpass_step(
-                t1,
+            opj_t1_enc_sigpass_step_macro(
+                mqc, curctx, a, c, ct,
                 f,
-                &t1->data[((k + 3) * t1->data_stride) + i],
+                &datap[3],
                 bpno,
                 one,
                 nmsedec,
                 type,
                 3, 0);
-            ++f;
         }
-        f += extra;
     }
 
     if (k < t1->h) {
@@ -528,20 +528,20 @@ static void opj_t1_enc_sigpass(opj_t1_t *t1,
 #ifdef DEBUG_ENC_SIG
         fprintf(stderr, " k=%d\n", k);
 #endif
-        for (i = 0; i < t1->w; ++i) {
+        for (i = 0; i < t1->w; ++i, ++f) {
 #ifdef DEBUG_ENC_SIG
             fprintf(stderr, " i=%d\n", i);
 #endif
             if (*f == 0U) {
                 /* Nothing to do for any of the 4 data points */
-                f++;
+                datap += (t1->h - k);
                 continue;
             }
-            for (j = k; j < t1->h; ++j) {
-                opj_t1_enc_sigpass_step(
-                    t1,
+            for (j = k; j < t1->h; ++j, ++datap) {
+                opj_t1_enc_sigpass_step_macro(
+                    mqc, curctx, a, c, ct,
                     f,
-                    &t1->data[(j * t1->data_stride) + i],
+                    &datap[0],
                     bpno,
                     one,
                     nmsedec,
@@ -549,9 +549,10 @@ static void opj_t1_enc_sigpass(opj_t1_t *t1,
                     j - k,
                     (j == k && (cblksty & J2K_CCP_CBLKSTY_VSC) != 0));
             }
-            ++f;
         }
     }
+
+    UPLOAD_MQC_VARIABLES(mqc, curctx, a, c, ct);
 }
 
 static void opj_t1_dec_sigpass_raw(
@@ -626,7 +627,7 @@ static void opj_t1_dec_sigpass_raw(
         register opj_flag_t *flagsp = &t1->flags[(flags_stride) + 1]; \
         const OPJ_UINT32 l_w = w; \
         opj_mqc_t* mqc = &(t1->mqc); \
-        DOWNLOAD_MQC_VARIABLES(mqc, curctx, c, a, ct); \
+        DOWNLOAD_MQC_VARIABLES(mqc, curctx, a, c, ct); \
         register OPJ_UINT32 v; \
         one = 1 << bpno; \
         half = one >> 1; \
@@ -651,7 +652,7 @@ static void opj_t1_dec_sigpass_raw(
                         } \
                 } \
         } \
-        UPLOAD_MQC_VARIABLES(mqc, curctx, c, a, ct); \
+        UPLOAD_MQC_VARIABLES(mqc, curctx, a, c, ct); \
         if( k < h ) { \
             for (i = 0; i < l_w; ++i, ++data, ++flagsp) { \
                 for (j = 0; j < h - k; ++j) { \
@@ -715,38 +716,27 @@ static void opj_t1_dec_sigpass_mqc(
 /**
 Encode refinement pass step
 */
-static INLINE void opj_t1_enc_refpass_step(opj_t1_t *t1,
-        opj_flag_t *flagsp,
-        OPJ_INT32 *datap,
-        OPJ_INT32 bpno,
-        OPJ_INT32 one,
-        OPJ_INT32 *nmsedec,
-        OPJ_BYTE type,
-        OPJ_UINT32 ci)
-{
-    OPJ_UINT32 v;
-
-    opj_mqc_t *mqc = &(t1->mqc);   /* MQC component */
-
-    OPJ_UINT32 const shift_flags =
-        (*flagsp >> (ci * 3U));
-
-    if ((shift_flags & (T1_SIGMA_THIS | T1_PI_THIS)) == T1_SIGMA_THIS) {
-        OPJ_UINT32 ctxt = opj_t1_getctxno_mag(shift_flags);
-        *nmsedec += opj_t1_getnmsedec_ref((OPJ_UINT32)opj_int_abs(*datap),
-                                          (OPJ_UINT32)bpno);
-        v = (opj_int_abs(*datap) & one) ? 1 : 0;
-#ifdef DEBUG_ENC_REF
-        fprintf(stderr, "  ctxt=%d\n", ctxt);
-#endif
-        opj_mqc_setcurctx(mqc, ctxt);
-        if (type == T1_TYPE_RAW) {  /* BYPASS/LAZY MODE */
-            opj_mqc_bypass_enc(mqc, v);
-        } else {
-            opj_mqc_encode(mqc, v);
-        }
-        *flagsp |= T1_MU_THIS << (ci * 3U);
-    }
+#define opj_t1_enc_refpass_step_macro(mqc, curctx, a, c, ct, flags, flagsUpdated, datap, bpno, one, nmsedec, type, ci) \
+{\
+    OPJ_UINT32 v; \
+    if ((flags & ((T1_SIGMA_THIS | T1_PI_THIS) << ((ci) * 3U))) == (T1_SIGMA_THIS << ((ci) * 3U))) { \
+        const OPJ_UINT32 shift_flags = (flags >> ((ci) * 3U)); \
+        OPJ_UINT32 ctxt = opj_t1_getctxno_mag(shift_flags); \
+        OPJ_UINT32 abs_data = opj_smr_abs(*datap); \
+        *nmsedec += opj_t1_getnmsedec_ref(abs_data, \
+                                          (OPJ_UINT32)bpno); \
+        v = ((OPJ_INT32)abs_data & one) ? 1 : 0; \
+/* #ifdef DEBUG_ENC_REF */ \
+/*        fprintf(stderr, "  ctxt=%d\n", ctxt); */ \
+/* #endif */ \
+        opj_t1_setcurctx(curctx, ctxt); \
+        if (type == T1_TYPE_RAW) {  /* BYPASS/LAZY MODE */ \
+            opj_mqc_bypass_enc_macro(mqc, c, ct, v); \
+        } else { \
+            opj_mqc_encode_macro(mqc, curctx, a, c, ct, v); \
+        } \
+        flagsUpdated |= T1_MU_THIS << ((ci) * 3U); \
+    } \
 }
 
 
@@ -807,100 +797,104 @@ static void opj_t1_enc_refpass(
     const OPJ_INT32 one = 1 << (bpno + T1_NMSEDEC_FRACBITS);
     opj_flag_t* f = &T1_FLAGS(0, 0);
     const OPJ_UINT32 extra = 2U;
+    opj_mqc_t* mqc = &(t1->mqc);
+    DOWNLOAD_MQC_VARIABLES(mqc, curctx, a, c, ct);
+    const OPJ_INT32* datap = t1->data;
 
     *nmsedec = 0;
 #ifdef DEBUG_ENC_REF
     fprintf(stderr, "enc_refpass: bpno=%d\n", bpno);
 #endif
-    for (k = 0; k < (t1->h & ~3U); k += 4) {
+    for (k = 0; k < (t1->h & ~3U); k += 4, f += extra) {
 #ifdef DEBUG_ENC_REF
         fprintf(stderr, " k=%d\n", k);
 #endif
-        for (i = 0; i < t1->w; ++i) {
+        for (i = 0; i < t1->w; ++i, f++, datap += 4) {
+            const OPJ_UINT32 flags = *f;
+            OPJ_UINT32 flagsUpdated = flags;
 #ifdef DEBUG_ENC_REF
             fprintf(stderr, " i=%d\n", i);
 #endif
-            if ((*f & (T1_SIGMA_4 | T1_SIGMA_7 | T1_SIGMA_10 | T1_SIGMA_13)) == 0) {
+            if ((flags & (T1_SIGMA_4 | T1_SIGMA_7 | T1_SIGMA_10 | T1_SIGMA_13)) == 0) {
                 /* none significant */
-                f++;
                 continue;
             }
-            if ((*f & (T1_PI_0 | T1_PI_1 | T1_PI_2 | T1_PI_3)) ==
+            if ((flags & (T1_PI_0 | T1_PI_1 | T1_PI_2 | T1_PI_3)) ==
                     (T1_PI_0 | T1_PI_1 | T1_PI_2 | T1_PI_3)) {
                 /* all processed by sigpass */
-                f++;
                 continue;
             }
 
-            opj_t1_enc_refpass_step(
-                t1,
-                f,
-                &t1->data[((k + 0) * t1->data_stride) + i],
+            opj_t1_enc_refpass_step_macro(
+                mqc, curctx, a, c, ct,
+                flags, flagsUpdated,
+                &datap[0],
                 bpno,
                 one,
                 nmsedec,
                 type,
                 0);
-            opj_t1_enc_refpass_step(
-                t1,
-                f,
-                &t1->data[((k + 1) * t1->data_stride) + i],
+            opj_t1_enc_refpass_step_macro(
+                mqc, curctx, a, c, ct,
+                flags, flagsUpdated,
+                &datap[1],
                 bpno,
                 one,
                 nmsedec,
                 type,
                 1);
-            opj_t1_enc_refpass_step(
-                t1,
-                f,
-                &t1->data[((k + 2) * t1->data_stride) + i],
+            opj_t1_enc_refpass_step_macro(
+                mqc, curctx, a, c, ct,
+                flags, flagsUpdated,
+                &datap[2],
                 bpno,
                 one,
                 nmsedec,
                 type,
                 2);
-            opj_t1_enc_refpass_step(
-                t1,
-                f,
-                &t1->data[((k + 3) * t1->data_stride) + i],
+            opj_t1_enc_refpass_step_macro(
+                mqc, curctx, a, c, ct,
+                flags, flagsUpdated,
+                &datap[3],
                 bpno,
                 one,
                 nmsedec,
                 type,
                 3);
-            ++f;
+            *f = flagsUpdated;
         }
-        f += extra;
     }
 
     if (k < t1->h) {
         OPJ_UINT32 j;
+        const OPJ_UINT32 remaining_lines = t1->h - k;
 #ifdef DEBUG_ENC_REF
         fprintf(stderr, " k=%d\n", k);
 #endif
-        for (i = 0; i < t1->w; ++i) {
+        for (i = 0; i < t1->w; ++i, ++f) {
 #ifdef DEBUG_ENC_REF
             fprintf(stderr, " i=%d\n", i);
 #endif
             if ((*f & (T1_SIGMA_4 | T1_SIGMA_7 | T1_SIGMA_10 | T1_SIGMA_13)) == 0) {
                 /* none significant */
-                f++;
+                datap += remaining_lines;
                 continue;
             }
-            for (j = k; j < t1->h; ++j) {
-                opj_t1_enc_refpass_step(
-                    t1,
-                    f,
-                    &t1->data[(j * t1->data_stride) + i],
+            for (j = 0; j < remaining_lines; ++j, datap ++) {
+                opj_t1_enc_refpass_step_macro(
+                    mqc, curctx, a, c, ct,
+                    *f, *f,
+                    &datap[0],
                     bpno,
                     one,
                     nmsedec,
                     type,
-                    j - k);
+                    j);
             }
-            ++f;
         }
     }
+
+    UPLOAD_MQC_VARIABLES(mqc, curctx, a, c, ct);
 }
 
 
@@ -968,7 +962,7 @@ static void opj_t1_dec_refpass_raw(
         register opj_flag_t *flagsp = &t1->flags[flags_stride + 1]; \
         const OPJ_UINT32 l_w = w; \
         opj_mqc_t* mqc = &(t1->mqc); \
-        DOWNLOAD_MQC_VARIABLES(mqc, curctx, c, a, ct); \
+        DOWNLOAD_MQC_VARIABLES(mqc, curctx, a, c, ct); \
         register OPJ_UINT32 v; \
         one = 1 << bpno; \
         poshalf = one >> 1; \
@@ -992,7 +986,7 @@ static void opj_t1_dec_refpass_raw(
                         } \
                 } \
         } \
-        UPLOAD_MQC_VARIABLES(mqc, curctx, c, a, ct); \
+        UPLOAD_MQC_VARIABLES(mqc, curctx, a, c, ct); \
         if( k < h ) { \
             for (i = 0; i < l_w; ++i, ++data, ++flagsp) { \
                 for (j = 0; j < h - k; ++j) { \
@@ -1030,86 +1024,71 @@ static void opj_t1_dec_refpass_mqc(
 /**
 Encode clean-up pass step
 */
-static void opj_t1_enc_clnpass_step(
-    opj_t1_t *t1,
-    opj_flag_t *flagsp,
-    OPJ_INT32 *datap,
-    OPJ_INT32 bpno,
-    OPJ_INT32 one,
-    OPJ_INT32 *nmsedec,
-    OPJ_UINT32 agg,
-    OPJ_UINT32 runlen,
-    OPJ_UINT32 lim,
-    OPJ_UINT32 cblksty)
-{
-    OPJ_UINT32 v;
-    OPJ_UINT32 ci;
-    opj_mqc_t *mqc = &(t1->mqc);   /* MQC component */
-
-    const OPJ_UINT32 check = (T1_SIGMA_4 | T1_SIGMA_7 | T1_SIGMA_10 | T1_SIGMA_13 |
-                              T1_PI_0 | T1_PI_1 | T1_PI_2 | T1_PI_3);
-
-    if ((*flagsp & check) == check) {
-        if (runlen == 0) {
-            *flagsp &= ~(T1_PI_0 | T1_PI_1 | T1_PI_2 | T1_PI_3);
-        } else if (runlen == 1) {
-            *flagsp &= ~(T1_PI_1 | T1_PI_2 | T1_PI_3);
-        } else if (runlen == 2) {
-            *flagsp &= ~(T1_PI_2 | T1_PI_3);
-        } else if (runlen == 3) {
-            *flagsp &= ~(T1_PI_3);
-        }
-        return;
-    }
-
-    for (ci = runlen; ci < lim; ++ci) {
-        OPJ_UINT32 vsc;
-        opj_flag_t flags;
-        OPJ_UINT32 ctxt1;
-
-        flags = *flagsp;
-
-        if ((agg != 0) && (ci == runlen)) {
-            goto LABEL_PARTIAL;
-        }
-
-        if (!(flags & ((T1_SIGMA_THIS | T1_PI_THIS) << (ci * 3U)))) {
-            ctxt1 = opj_t1_getctxno_zc(mqc, flags >> (ci * 3U));
-#ifdef DEBUG_ENC_CLN
-            printf("   ctxt1=%d\n", ctxt1);
-#endif
-            opj_mqc_setcurctx(mqc, ctxt1);
-            v = (opj_int_abs(*datap) & one) ? 1 : 0;
-            opj_mqc_encode(mqc, v);
-            if (v) {
-                OPJ_UINT32 ctxt2, spb;
-                OPJ_UINT32 lu;
-LABEL_PARTIAL:
-                lu = opj_t1_getctxtno_sc_or_spb_index(
-                         *flagsp,
-                         flagsp[-1], flagsp[1],
-                         ci);
-                *nmsedec += opj_t1_getnmsedec_sig((OPJ_UINT32)opj_int_abs(*datap),
-                                                  (OPJ_UINT32)bpno);
-                ctxt2 = opj_t1_getctxno_sc(lu);
-#ifdef DEBUG_ENC_CLN
-                printf("   ctxt2=%d\n", ctxt2);
-#endif
-                opj_mqc_setcurctx(mqc, ctxt2);
-
-                v = *datap < 0 ? 1U : 0U;
-                spb = opj_t1_getspb(lu);
-#ifdef DEBUG_ENC_CLN
-                printf("   spb=%d\n", spb);
-#endif
-                opj_mqc_encode(mqc, v ^ spb);
-                vsc = ((cblksty & J2K_CCP_CBLKSTY_VSC) && (ci == 0)) ? 1 : 0;
-                opj_t1_update_flags(flagsp, ci, v, t1->w + 2U, vsc);
-            }
-        }
-        *flagsp &= ~(T1_PI_THIS << (3U * ci));
-        datap += t1->data_stride;
-    }
+#define opj_t1_enc_clnpass_step_macro(mqc, curctx, a, c, ct, flagspIn, datapIn, bpno, one, nmsedec, agg, runlen, lim, cblksty) \
+{ \
+    OPJ_UINT32 v; \
+    OPJ_UINT32 ci; \
+    opj_flag_t* const flagsp = (flagspIn); \
+    const OPJ_INT32* l_datap = (datapIn); \
+    const OPJ_UINT32 check = (T1_SIGMA_4 | T1_SIGMA_7 | T1_SIGMA_10 | T1_SIGMA_13 | \
+                              T1_PI_0 | T1_PI_1 | T1_PI_2 | T1_PI_3); \
+ \
+    if ((*flagsp & check) == check) { \
+        if (runlen == 0) { \
+            *flagsp &= ~(T1_PI_0 | T1_PI_1 | T1_PI_2 | T1_PI_3); \
+        } else if (runlen == 1) { \
+            *flagsp &= ~(T1_PI_1 | T1_PI_2 | T1_PI_3); \
+        } else if (runlen == 2) { \
+            *flagsp &= ~(T1_PI_2 | T1_PI_3); \
+        } else if (runlen == 3) { \
+            *flagsp &= ~(T1_PI_3); \
+        } \
+    } \
+    else \
+    for (ci = runlen; ci < lim; ++ci) { \
+        OPJ_BOOL goto_PARTIAL = OPJ_FALSE; \
+        if ((agg != 0) && (ci == runlen)) { \
+            goto_PARTIAL = OPJ_TRUE; \
+        } \
+        else if (!(*flagsp & ((T1_SIGMA_THIS | T1_PI_THIS) << (ci * 3U)))) { \
+            OPJ_UINT32 ctxt1 = opj_t1_getctxno_zc(mqc, *flagsp >> (ci * 3U)); \
+/* #ifdef DEBUG_ENC_CLN */ \
+/*            printf("   ctxt1=%d\n", ctxt1); */ \
+/* #endif */ \
+            opj_t1_setcurctx(curctx, ctxt1); \
+            v = (opj_smr_abs(*l_datap) & (OPJ_UINT32)one) ? 1 : 0; \
+            opj_mqc_encode_macro(mqc, curctx, a, c, ct, v); \
+            if (v) { \
+                goto_PARTIAL = OPJ_TRUE; \
+            } \
+        } \
+        if( goto_PARTIAL ) { \
+            OPJ_UINT32 vsc; \
+            OPJ_UINT32 ctxt2, spb; \
+            OPJ_UINT32 lu = opj_t1_getctxtno_sc_or_spb_index( \
+                        *flagsp, \
+                        flagsp[-1], flagsp[1], \
+                        ci); \
+            *nmsedec += opj_t1_getnmsedec_sig(opj_smr_abs(*l_datap), \
+                                                (OPJ_UINT32)bpno); \
+            ctxt2 = opj_t1_getctxno_sc(lu); \
+/* #ifdef DEBUG_ENC_CLN */ \
+/*           printf("   ctxt2=%d\n", ctxt2); */ \
+/* #endif */ \
+            opj_t1_setcurctx(curctx, ctxt2); \
+ \
+            v = opj_smr_sign(*l_datap); \
+            spb = opj_t1_getspb(lu); \
+/* #ifdef DEBUG_ENC_CLN */ \
+/*           printf("   spb=%d\n", spb); */\
+/* #endif */ \
+            opj_mqc_encode_macro(mqc, curctx, a, c, ct, v ^ spb); \
+            vsc = ((cblksty & J2K_CCP_CBLKSTY_VSC) && (ci == 0)) ? 1 : 0; \
+            opj_t1_update_flags(flagsp, ci, v, t1->w + 2U, vsc); \
+        } \
+        *flagsp &= ~(T1_PI_THIS << (3U * ci)); \
+        l_datap ++; \
+    } \
 }
 
 #define opj_t1_dec_clnpass_step_macro(check_flags, partial, \
@@ -1165,47 +1144,50 @@ static void opj_t1_enc_clnpass(
 {
     OPJ_UINT32 i, k;
     const OPJ_INT32 one = 1 << (bpno + T1_NMSEDEC_FRACBITS);
-    OPJ_UINT32 agg, runlen;
-
-    opj_mqc_t *mqc = &(t1->mqc);   /* MQC component */
+    opj_mqc_t* mqc = &(t1->mqc);
+    DOWNLOAD_MQC_VARIABLES(mqc, curctx, a, c, ct);
+    const OPJ_INT32* datap = t1->data;
+    opj_flag_t *f = &T1_FLAGS(0, 0);
+    const OPJ_UINT32 extra = 2U;
 
     *nmsedec = 0;
 #ifdef DEBUG_ENC_CLN
     printf("enc_clnpass: bpno=%d\n", bpno);
 #endif
-    for (k = 0; k < (t1->h & ~3U); k += 4) {
+    for (k = 0; k < (t1->h & ~3U); k += 4, f += extra) {
 #ifdef DEBUG_ENC_CLN
         printf(" k=%d\n", k);
 #endif
-        for (i = 0; i < t1->w; ++i) {
+        for (i = 0; i < t1->w; ++i, f++) {
+            OPJ_UINT32 agg, runlen;
 #ifdef DEBUG_ENC_CLN
             printf("  i=%d\n", i);
 #endif
-            agg = !(T1_FLAGS(i, k));
+            agg = !*f;
 #ifdef DEBUG_ENC_CLN
             printf("   agg=%d\n", agg);
 #endif
             if (agg) {
-                for (runlen = 0; runlen < 4; ++runlen) {
-                    if (opj_int_abs(t1->data[((k + runlen)*t1->data_stride) + i]) & one) {
+                for (runlen = 0; runlen < 4; ++runlen, ++datap) {
+                    if (opj_smr_abs(*datap) & (OPJ_UINT32)one) {
                         break;
                     }
                 }
-                opj_mqc_setcurctx(mqc, T1_CTXNO_AGG);
-                opj_mqc_encode(mqc, runlen != 4);
+                opj_t1_setcurctx(curctx, T1_CTXNO_AGG);
+                opj_mqc_encode_macro(mqc, curctx, a, c, ct, runlen != 4);
                 if (runlen == 4) {
                     continue;
                 }
-                opj_mqc_setcurctx(mqc, T1_CTXNO_UNI);
-                opj_mqc_encode(mqc, runlen >> 1);
-                opj_mqc_encode(mqc, runlen & 1);
+                opj_t1_setcurctx(curctx, T1_CTXNO_UNI);
+                opj_mqc_encode_macro(mqc, curctx, a, c, ct, runlen >> 1);
+                opj_mqc_encode_macro(mqc, curctx, a, c, ct, runlen & 1);
             } else {
                 runlen = 0;
             }
-            opj_t1_enc_clnpass_step(
-                t1,
-                &T1_FLAGS(i, k),
-                &t1->data[((k + runlen) * t1->data_stride) + i],
+            opj_t1_enc_clnpass_step_macro(
+                mqc, curctx, a, c, ct,
+                f,
+                datap,
                 bpno,
                 one,
                 nmsedec,
@@ -1213,23 +1195,24 @@ static void opj_t1_enc_clnpass(
                 runlen,
                 4U,
                 cblksty);
+            datap += 4 - runlen;
         }
     }
     if (k < t1->h) {
-        agg = 0;
-        runlen = 0;
+        const OPJ_UINT32 agg = 0;
+        const OPJ_UINT32 runlen = 0;
 #ifdef DEBUG_ENC_CLN
         printf(" k=%d\n", k);
 #endif
-        for (i = 0; i < t1->w; ++i) {
+        for (i = 0; i < t1->w; ++i, f++) {
 #ifdef DEBUG_ENC_CLN
             printf("  i=%d\n", i);
             printf("   agg=%d\n", agg);
 #endif
-            opj_t1_enc_clnpass_step(
-                t1,
-                &T1_FLAGS(i, k),
-                &t1->data[((k + runlen) * t1->data_stride) + i],
+            opj_t1_enc_clnpass_step_macro(
+                mqc, curctx, a, c, ct,
+                f,
+                datap,
                 bpno,
                 one,
                 nmsedec,
@@ -1237,8 +1220,11 @@ static void opj_t1_enc_clnpass(
                 runlen,
                 t1->h - k,
                 cblksty);
+            datap += t1->h - k;
         }
     }
+
+    UPLOAD_MQC_VARIABLES(mqc, curctx, a, c, ct);
 }
 
 #define opj_t1_dec_clnpass_internal(t1, bpno, vsc, w, h, flags_stride) \
@@ -1250,7 +1236,7 @@ static void opj_t1_enc_clnpass(
     opj_mqc_t* mqc = &(t1->mqc); \
     register OPJ_INT32 *data = t1->data; \
     register opj_flag_t *flagsp = &t1->flags[flags_stride + 1]; \
-    DOWNLOAD_MQC_VARIABLES(mqc, curctx, c, a, ct); \
+    DOWNLOAD_MQC_VARIABLES(mqc, curctx, a, c, ct); \
     register OPJ_UINT32 v; \
     one = 1 << bpno; \
     half = one >> 1; \
@@ -1319,7 +1305,7 @@ static void opj_t1_enc_clnpass(
             *flagsp = flags & ~(T1_PI_0 | T1_PI_1 | T1_PI_2 | T1_PI_3); \
         } \
     } \
-    UPLOAD_MQC_VARIABLES(mqc, curctx, c, a, ct); \
+    UPLOAD_MQC_VARIABLES(mqc, curctx, a, c, ct); \
     if( k < h ) { \
         for (i = 0; i < l_w; ++i, ++flagsp, ++data) { \
             for (j = 0; j < h - k; ++j) { \
@@ -1427,7 +1413,7 @@ static OPJ_FLOAT64 opj_t1_getwmsedec(
         w2 = opj_dwt_getnorm(level, orient);
     } else {    /* if (qmfbid == 0) */
         const OPJ_INT32 log2_gain = (orient == 0) ? 0 :
-                                                (orient == 3) ? 2 : 1;
+                                    (orient == 3) ? 2 : 1;
         w2 = opj_dwt_getnorm_real(level, orient);
         /* Not sure this is right. But preserves past behaviour */
         stepsize /= (1 << log2_gain);
@@ -1454,7 +1440,7 @@ static OPJ_BOOL opj_t1_allocate_buffers(
     assert(w * h <= 4096);
 
     /* encoder uses tile buffer, so no need to allocate */
-    if (!t1->encoder) {
+    {
         OPJ_UINT32 datasize = w * h;
 
         if (datasize > t1->datasize) {
@@ -1564,8 +1550,7 @@ void opj_t1_destroy(opj_t1_t *p_t1)
         return;
     }
 
-    /* encoder uses tile buffer, so no need to free */
-    if (!p_t1->encoder && p_t1->data) {
+    if (p_t1->data) {
         opj_aligned_free(p_t1->data);
         p_t1->data = 00;
     }
@@ -2140,8 +2125,7 @@ static void opj_t1_clbl_encode_processor(void* user_data, opj_tls_t* tls)
     OPJ_INT32* OPJ_RESTRICT tiledp;
     OPJ_UINT32 cblk_w;
     OPJ_UINT32 cblk_h;
-    OPJ_UINT32 i, j, tileLineAdvance;
-    OPJ_SIZE_T tileIndex = 0;
+    OPJ_UINT32 i, j;
 
     OPJ_INT32 x = cblk->x0 - band->x0;
     OPJ_INT32 y = cblk->y0 - band->y0;
@@ -2177,11 +2161,9 @@ static void opj_t1_clbl_encode_processor(void* user_data, opj_tls_t* tls)
 
     cblk_w = t1->w;
     cblk_h = t1->h;
-    tileLineAdvance = tile_w - cblk_w;
 
     tiledp = &tilec->data[(OPJ_SIZE_T)y * tile_w + (OPJ_SIZE_T)x];
-    t1->data = tiledp;
-    t1->data_stride = tile_w;
+
     if (tccp->qmfbid == 1) {
         /* Do multiplication on unsigned type, even if the
             * underlying type is signed, to avoid potential
@@ -2192,22 +2174,52 @@ static void opj_t1_clbl_encode_processor(void* user_data, opj_tls_t* tls)
             * Fixes https://github.com/uclouvain/openjpeg/issues/1053
             */
         OPJ_UINT32* OPJ_RESTRICT tiledp_u = (OPJ_UINT32*) tiledp;
-        for (j = 0; j < cblk_h; ++j) {
+        OPJ_UINT32* OPJ_RESTRICT t1data = (OPJ_UINT32*) t1->data;
+        /* Change from "natural" order to "zigzag" order of T1 passes */
+        for (j = 0; j < (cblk_h & ~3U); j += 4) {
             for (i = 0; i < cblk_w; ++i) {
-                tiledp_u[tileIndex] <<= T1_NMSEDEC_FRACBITS;
-                tileIndex++;
+                t1data[0] = tiledp_u[(j + 0) * tile_w + i] << T1_NMSEDEC_FRACBITS;
+                t1data[1] = tiledp_u[(j + 1) * tile_w + i] << T1_NMSEDEC_FRACBITS;
+                t1data[2] = tiledp_u[(j + 2) * tile_w + i] << T1_NMSEDEC_FRACBITS;
+                t1data[3] = tiledp_u[(j + 3) * tile_w + i] << T1_NMSEDEC_FRACBITS;
+                t1data += 4;
+            }
+        }
+        if (j < cblk_h) {
+            for (i = 0; i < cblk_w; ++i) {
+                OPJ_UINT32 k;
+                for (k = j; k < cblk_h; k++) {
+                    t1data[0] = tiledp_u[k * tile_w + i] << T1_NMSEDEC_FRACBITS;
+                    t1data ++;
+                }
             }
-            tileIndex += tileLineAdvance;
         }
     } else {        /* if (tccp->qmfbid == 0) */
-        for (j = 0; j < cblk_h; ++j) {
+        OPJ_FLOAT32* OPJ_RESTRICT tiledp_f = (OPJ_FLOAT32*) tiledp;
+        OPJ_INT32* OPJ_RESTRICT t1data = t1->data;
+        /* Change from "natural" order to "zigzag" order of T1 passes */
+        for (j = 0; j < (cblk_h & ~3U); j += 4) {
+            for (i = 0; i < cblk_w; ++i) {
+                t1data[0] = (OPJ_INT32)opj_lrintf((tiledp_f[(j + 0) * tile_w + i] /
+                                                   band->stepsize) * (1 << T1_NMSEDEC_FRACBITS));
+                t1data[1] = (OPJ_INT32)opj_lrintf((tiledp_f[(j + 1) * tile_w + i] /
+                                                   band->stepsize) * (1 << T1_NMSEDEC_FRACBITS));
+                t1data[2] = (OPJ_INT32)opj_lrintf((tiledp_f[(j + 2) * tile_w + i] /
+                                                   band->stepsize) * (1 << T1_NMSEDEC_FRACBITS));
+                t1data[3] = (OPJ_INT32)opj_lrintf((tiledp_f[(j + 3) * tile_w + i] /
+                                                   band->stepsize) * (1 << T1_NMSEDEC_FRACBITS));
+                t1data += 4;
+            }
+        }
+        if (j < cblk_h) {
             for (i = 0; i < cblk_w; ++i) {
-                OPJ_FLOAT32 tmp = ((OPJ_FLOAT32*)tiledp)[tileIndex];
-                tiledp[tileIndex] = (OPJ_INT32)opj_lrintf((tmp / band->stepsize) *
-                                    (1 << T1_NMSEDEC_FRACBITS));
-                tileIndex++;
+                OPJ_UINT32 k;
+                for (k = j; k < cblk_h; k++) {
+                    t1data[0] = (OPJ_INT32)opj_lrintf((tiledp_f[k * tile_w + i] / band->stepsize)
+                                                      * (1 << T1_NMSEDEC_FRACBITS));
+                    t1data ++;
+                }
             }
-            tileIndex += tileLineAdvance;
         }
     }
 
@@ -2363,6 +2375,7 @@ static OPJ_FLOAT64 opj_t1_encode_cblk(opj_t1_t *t1,
     OPJ_UINT32 i, j;
     OPJ_BYTE type = T1_TYPE_MQ;
     OPJ_FLOAT64 tempwmsedec;
+    OPJ_INT32* datap;
 
 #ifdef EXTRA_DEBUG
     printf("encode_cblk(x=%d,y=%d,x1=%d,y1=%d,orient=%d,compno=%d,level=%d\n",
@@ -2372,10 +2385,19 @@ static OPJ_FLOAT64 opj_t1_encode_cblk(opj_t1_t *t1,
     mqc->lut_ctxno_zc_orient = lut_ctxno_zc + (orient << 9);
 
     max = 0;
-    for (i = 0; i < t1->w; ++i) {
-        for (j = 0; j < t1->h; ++j) {
-            OPJ_INT32 tmp = abs(t1->data[i + j * t1->data_stride]);
-            max = opj_int_max(max, tmp);
+    datap = t1->data;
+    for (j = 0; j < t1->h; ++j) {
+        const OPJ_UINT32 w = t1->w;
+        for (i = 0; i < w; ++i, ++datap) {
+            OPJ_INT32 tmp = *datap;
+            if (tmp < 0) {
+                OPJ_UINT32 tmp_unsigned;
+                max = opj_int_max(max, -tmp);
+                tmp_unsigned = opj_to_smr(tmp);
+                memcpy(datap, &tmp_unsigned, sizeof(OPJ_INT32));
+            } else {
+                max = opj_int_max(max, tmp);
+            }
         }
     }
 
diff --git a/src/lib/openjp2/t1.h b/src/lib/openjp2/t1.h
index bc8a8111..81ad0d00 100644
--- a/src/lib/openjp2/t1.h
+++ b/src/lib/openjp2/t1.h
@@ -198,7 +198,6 @@ typedef struct opj_t1 {
     OPJ_UINT32 h;
     OPJ_UINT32 datasize;
     OPJ_UINT32 flagssize;
-    OPJ_UINT32 data_stride;
     OPJ_BOOL   encoder;
 
     /* Thre 3 variables below are only used by the decoder */