l_src += 2;
} /* b[i*x]=a[2*i+cas]; */
- l_dest = b + sn * x;
+ l_dest = b + (OPJ_SIZE_T)sn * (OPJ_SIZE_T)x;
l_src = a + 1 - cas;
i = dn;
bi += 2;
ai += x;
}
- ai = a + (v->sn * x);
+ ai = a + (v->sn * (OPJ_SIZE_T)x);
bi = v->mem + 1 - v->cas;
i = v->dn ;
while (i--) {
void opj_idwt53_v_final_memcpy(OPJ_INT32* tiledp_col,
const OPJ_INT32* tmp,
OPJ_INT32 len,
- OPJ_INT32 stride)
+ OPJ_SIZE_T stride)
{
OPJ_INT32 i;
for (i = 0; i < len; ++i) {
PARALLEL_COLS_53 * sizeof(OPJ_INT32))
would do but would be a tiny bit slower.
We can take here advantage of our knowledge of alignment */
- STOREU(&tiledp_col[i * stride + 0],
+ STOREU(&tiledp_col[(OPJ_SIZE_T)i * stride + 0],
LOAD(&tmp[PARALLEL_COLS_53 * i + 0]));
- STOREU(&tiledp_col[i * stride + VREG_INT_COUNT],
+ STOREU(&tiledp_col[(OPJ_SIZE_T)i * stride + VREG_INT_COUNT],
LOAD(&tmp[PARALLEL_COLS_53 * i + VREG_INT_COUNT]));
}
}
const OPJ_INT32 sn,
const OPJ_INT32 len,
OPJ_INT32* tiledp_col,
- const OPJ_INT32 stride)
+ const OPJ_SIZE_T stride)
{
const OPJ_INT32* in_even = &tiledp_col[0];
- const OPJ_INT32* in_odd = &tiledp_col[sn * stride];
+ const OPJ_INT32* in_odd = &tiledp_col[(OPJ_SIZE_T)sn * stride];
- OPJ_INT32 i, j;
+ OPJ_INT32 i;
+ OPJ_SIZE_T j;
VREG d1c_0, d1n_0, s1n_0, s0c_0, s0n_0;
VREG d1c_1, d1n_1, s1n_1, s0c_1, s0n_1;
const VREG two = LOAD_CST(2);
/* Note: loads of input even/odd values must be done in a unaligned */
/* fashion. But stores in tmp can be done with aligned store, since */
/* the temporary buffer is properly aligned */
- assert((size_t)tmp % (sizeof(OPJ_INT32) * VREG_INT_COUNT) == 0);
+ assert((OPJ_SIZE_T)tmp % (sizeof(OPJ_INT32) * VREG_INT_COUNT) == 0);
s1n_0 = LOADU(in_even + 0);
s1n_1 = LOADU(in_even + VREG_INT_COUNT);
if (len & 1) {
VREG tmp_len_minus_1;
- s1n_0 = LOADU(in_even + ((len - 1) / 2) * stride);
+ s1n_0 = LOADU(in_even + (OPJ_SIZE_T)((len - 1) / 2) * stride);
/* tmp_len_minus_1 = s1n - ((d1n + 1) >> 1); */
tmp_len_minus_1 = SUB(s1n_0, SAR(ADD3(d1n_0, d1n_0, two), 2));
STORE(tmp + PARALLEL_COLS_53 * (len - 1), tmp_len_minus_1);
STORE(tmp + PARALLEL_COLS_53 * (len - 2),
ADD(d1n_0, SAR(ADD(s0n_0, tmp_len_minus_1), 1)));
- s1n_1 = LOADU(in_even + ((len - 1) / 2) * stride + VREG_INT_COUNT);
+ s1n_1 = LOADU(in_even + (OPJ_SIZE_T)((len - 1) / 2) * stride + VREG_INT_COUNT);
/* tmp_len_minus_1 = s1n - ((d1n + 1) >> 1); */
tmp_len_minus_1 = SUB(s1n_1, SAR(ADD3(d1n_1, d1n_1, two), 2));
STORE(tmp + PARALLEL_COLS_53 * (len - 1) + VREG_INT_COUNT,
const OPJ_INT32 sn,
const OPJ_INT32 len,
OPJ_INT32* tiledp_col,
- const OPJ_INT32 stride)
+ const OPJ_SIZE_T stride)
{
- OPJ_INT32 i, j;
+ OPJ_INT32 i;
+ OPJ_SIZE_T j;
VREG s1_0, s2_0, dc_0, dn_0;
VREG s1_1, s2_1, dc_1, dn_1;
const VREG two = LOAD_CST(2);
- const OPJ_INT32* in_even = &tiledp_col[sn * stride];
+ const OPJ_INT32* in_even = &tiledp_col[(OPJ_SIZE_T)sn * stride];
const OPJ_INT32* in_odd = &tiledp_col[0];
assert(len > 2);
/* Note: loads of input even/odd values must be done in a unaligned */
/* fashion. But stores in tmp can be done with aligned store, since */
/* the temporary buffer is properly aligned */
- assert((size_t)tmp % (sizeof(OPJ_INT32) * VREG_INT_COUNT) == 0);
+ assert((OPJ_SIZE_T)tmp % (sizeof(OPJ_INT32) * VREG_INT_COUNT) == 0);
s1_0 = LOADU(in_even + stride);
/* in_odd[0] - ((in_even[0] + s1 + 2) >> 2); */
if (!(len & 1)) {
/*dn = in_odd[(len / 2 - 1) * stride] - ((s1 + 1) >> 1); */
- dn_0 = SUB(LOADU(in_odd + (len / 2 - 1) * stride),
+ dn_0 = SUB(LOADU(in_odd + (OPJ_SIZE_T)(len / 2 - 1) * stride),
SAR(ADD3(s1_0, s1_0, two), 2));
- dn_1 = SUB(LOADU(in_odd + (len / 2 - 1) * stride + VREG_INT_COUNT),
+ dn_1 = SUB(LOADU(in_odd + (OPJ_SIZE_T)(len / 2 - 1) * stride + VREG_INT_COUNT),
SAR(ADD3(s1_1, s1_1, two), 2));
/* tmp[len - 2] = s1 + ((dn + dc) >> 1); */
const OPJ_INT32 sn,
const OPJ_INT32 len,
OPJ_INT32* tiledp_col,
- const OPJ_INT32 stride)
+ const OPJ_SIZE_T stride)
{
OPJ_INT32 i, j;
OPJ_INT32 d1c, d1n, s1n, s0c, s0n;
/* accesses and explicit interleaving. */
s1n = tiledp_col[0];
- d1n = tiledp_col[sn * stride];
+ d1n = tiledp_col[(OPJ_SIZE_T)sn * stride];
s0n = s1n - ((d1n + 1) >> 1);
for (i = 0, j = 0; i < (len - 3); i += 2, j++) {
d1c = d1n;
s0c = s0n;
- s1n = tiledp_col[(j + 1) * stride];
- d1n = tiledp_col[(sn + j + 1) * stride];
+ s1n = tiledp_col[(OPJ_SIZE_T)(j + 1) * stride];
+ d1n = tiledp_col[(OPJ_SIZE_T)(sn + j + 1) * stride];
s0n = s1n - ((d1c + d1n + 2) >> 2);
if (len & 1) {
tmp[len - 1] =
- tiledp_col[((len - 1) / 2) * stride] -
+ tiledp_col[(OPJ_SIZE_T)((len - 1) / 2) * stride] -
((d1n + 1) >> 1);
tmp[len - 2] = d1n + ((s0n + tmp[len - 1]) >> 1);
} else {
}
for (i = 0; i < len; ++i) {
- tiledp_col[i * stride] = tmp[i];
+ tiledp_col[(OPJ_SIZE_T)i * stride] = tmp[i];
}
}
const OPJ_INT32 sn,
const OPJ_INT32 len,
OPJ_INT32* tiledp_col,
- const OPJ_INT32 stride)
+ const OPJ_SIZE_T stride)
{
OPJ_INT32 i, j;
OPJ_INT32 s1, s2, dc, dn;
- const OPJ_INT32* in_even = &tiledp_col[sn * stride];
+ const OPJ_INT32* in_even = &tiledp_col[(OPJ_SIZE_T)sn * stride];
const OPJ_INT32* in_odd = &tiledp_col[0];
assert(len > 2);
tmp[0] = in_even[0] + dc;
for (i = 1, j = 1; i < (len - 2 - !(len & 1)); i += 2, j++) {
- s2 = in_even[(j + 1) * stride];
+ s2 = in_even[(OPJ_SIZE_T)(j + 1) * stride];
- dn = in_odd[j * stride] - ((s1 + s2 + 2) >> 2);
+ dn = in_odd[(OPJ_SIZE_T)j * stride] - ((s1 + s2 + 2) >> 2);
tmp[i ] = dc;
tmp[i + 1] = s1 + ((dn + dc) >> 1);
}
tmp[i] = dc;
if (!(len & 1)) {
- dn = in_odd[(len / 2 - 1) * stride] - ((s1 + 1) >> 1);
+ dn = in_odd[(OPJ_SIZE_T)(len / 2 - 1) * stride] - ((s1 + 1) >> 1);
tmp[len - 2] = s1 + ((dn + dc) >> 1);
tmp[len - 1] = dn;
} else {
}
for (i = 0; i < len; ++i) {
- tiledp_col[i * stride] = tmp[i];
+ tiledp_col[(OPJ_SIZE_T)i * stride] = tmp[i];
}
}
#endif /* !defined(STANDARD_SLOW_VERSION) */
/* Performs interleave, inverse wavelet transform and copy back to buffer */
static void opj_idwt53_v(const opj_dwt_t *dwt,
OPJ_INT32* tiledp_col,
- OPJ_INT32 stride,
+ OPJ_SIZE_T stride,
OPJ_INT32 nb_cols)
{
#ifdef STANDARD_SLOW_VERSION
OPJ_INT32* out = dwt->mem;
for (c = 0; c < nb_cols; c++, tiledp_col++) {
OPJ_INT32 i;
- const OPJ_INT32* in_even = &tiledp_col[sn * stride];
+ const OPJ_INT32* in_even = &tiledp_col[(OPJ_SIZE_T)sn * stride];
const OPJ_INT32* in_odd = &tiledp_col[0];
out[1] = in_odd[0] - ((in_even[0] + 1) >> 1);
out[0] = in_even[0] + out[1];
for (i = 0; i < len; ++i) {
- tiledp_col[i * stride] = out[i];
+ tiledp_col[(OPJ_SIZE_T)i * stride] = out[i];
}
}
OPJ_INT32 rw; /* width of the resolution level computed */
OPJ_INT32 rh; /* height of the resolution level computed */
- size_t l_data_size;
+ OPJ_SIZE_T l_data_size;
opj_tcd_resolution_t * l_cur_res = 0;
opj_tcd_resolution_t * l_last_res = 0;
job = (opj_dwd_decode_v_job_t*)user_data;
for (j = job->min_j; j + PARALLEL_COLS_53 <= job->max_j;
j += PARALLEL_COLS_53) {
- opj_idwt53_v(&job->v, &job->tiledp[j], (OPJ_INT32)job->w,
+ opj_idwt53_v(&job->v, &job->tiledp[j], (OPJ_SIZE_T)job->w,
PARALLEL_COLS_53);
}
if (j < job->max_j)
- opj_idwt53_v(&job->v, &job->tiledp[j], (OPJ_INT32)job->w,
+ opj_idwt53_v(&job->v, &job->tiledp[j], (OPJ_SIZE_T)job->w,
(OPJ_INT32)(job->max_j - j));
opj_aligned_free(job->v.mem);
OPJ_UINT32 w = (OPJ_UINT32)(tilec->resolutions[tilec->minimum_num_resolutions -
1].x1 -
tilec->resolutions[tilec->minimum_num_resolutions - 1].x0);
- size_t h_mem_size;
+ OPJ_SIZE_T h_mem_size;
int num_threads;
if (numres == 1U) {
if (num_threads <= 1 || rh <= 1) {
for (j = 0; j < rh; ++j) {
- opj_idwt53_h(&h, &tiledp[j * w]);
+ opj_idwt53_h(&h, &tiledp[(OPJ_SIZE_T)j * w]);
}
} else {
OPJ_UINT32 num_jobs = (OPJ_UINT32)num_threads;
if (num_threads <= 1 || rw <= 1) {
for (j = 0; j + PARALLEL_COLS_53 <= rw;
j += PARALLEL_COLS_53) {
- opj_idwt53_v(&v, &tiledp[j], (OPJ_INT32)w, PARALLEL_COLS_53);
+ opj_idwt53_v(&v, &tiledp[j], (OPJ_SIZE_T)w, PARALLEL_COLS_53);
}
if (j < rw) {
- opj_idwt53_v(&v, &tiledp[j], (OPJ_INT32)w, (OPJ_INT32)(rw - j));
+ opj_idwt53_v(&v, &tiledp[j], (OPJ_SIZE_T)w, (OPJ_INT32)(rw - j));
}
} else {
OPJ_UINT32 num_jobs = (OPJ_UINT32)num_threads;
dest + 1 - cas + 2 * win_h_x0,
2, 0, OPJ_TRUE);
assert(ret);
+ OPJ_UNUSED(ret);
}
OPJ_INT32 cas,
opj_sparse_array_int32_t* sa,
OPJ_UINT32 sa_col,
+ OPJ_UINT32 nb_cols,
OPJ_UINT32 sn,
OPJ_UINT32 win_l_y0,
OPJ_UINT32 win_l_y1,
OPJ_BOOL ret;
ret = opj_sparse_array_int32_read(sa,
sa_col, win_l_y0,
- sa_col + 1, win_l_y1,
- dest + cas + 2 * win_l_y0,
- 0, 2, OPJ_TRUE);
+ sa_col + nb_cols, win_l_y1,
+ dest + cas * 4 + 2 * 4 * win_l_y0,
+ 1, 2 * 4, OPJ_TRUE);
assert(ret);
ret = opj_sparse_array_int32_read(sa,
sa_col, sn + win_h_y0,
- sa_col + 1, sn + win_h_y1,
- dest + 1 - cas + 2 * win_h_y0,
- 0, 2, OPJ_TRUE);
+ sa_col + nb_cols, sn + win_h_y1,
+ dest + (1 - cas) * 4 + 2 * 4 * win_h_y0,
+ 1, 2 * 4, OPJ_TRUE);
assert(ret);
+ OPJ_UNUSED(ret);
}
static void opj_dwt_decode_partial_1(OPJ_INT32 *a, OPJ_INT32 dn, OPJ_INT32 sn,
if (!cas) {
if ((dn > 0) || (sn > 1)) { /* NEW : CASE ONE ELEMENT */
- for (i = win_l_x0; i < win_l_x1; i++) {
+
+ /* Naive version is :
+ for (i = win_l_x0; i < i_max; i++) {
OPJ_S(i) -= (OPJ_D_(i - 1) + OPJ_D_(i) + 2) >> 2;
}
for (i = win_h_x0; i < win_h_x1; i++) {
OPJ_D(i) += (OPJ_S_(i) + OPJ_S_(i + 1)) >> 1;
}
+ but the compiler doesn't manage to unroll it to avoid bound
+ checking in OPJ_S_ and OPJ_D_ macros
+ */
+
+ i = win_l_x0;
+ if (i < win_l_x1) {
+ OPJ_INT32 i_max;
+
+ /* Left-most case */
+ OPJ_S(i) -= (OPJ_D_(i - 1) + OPJ_D_(i) + 2) >> 2;
+ i ++;
+
+ i_max = win_l_x1;
+ if (i_max > dn) {
+ i_max = dn;
+ }
+ for (; i < i_max; i++) {
+ /* No bound checking */
+ OPJ_S(i) -= (OPJ_D(i - 1) + OPJ_D(i) + 2) >> 2;
+ }
+ for (; i < win_l_x1; i++) {
+ /* Right-most case */
+ OPJ_S(i) -= (OPJ_D_(i - 1) + OPJ_D_(i) + 2) >> 2;
+ }
+ }
+
+ i = win_h_x0;
+ if (i < win_h_x1) {
+ OPJ_INT32 i_max = win_h_x1;
+ if (i_max >= sn) {
+ i_max = sn - 1;
+ }
+ for (; i < i_max; i++) {
+ /* No bound checking */
+ OPJ_D(i) += (OPJ_S(i) + OPJ_S(i + 1)) >> 1;
+ }
+ for (; i < win_h_x1; i++) {
+ /* Right-most case */
+ OPJ_D(i) += (OPJ_S_(i) + OPJ_S_(i + 1)) >> 1;
+ }
+ }
}
} else {
if (!sn && dn == 1) { /* NEW : CASE ONE ELEMENT */
}
}
+#define OPJ_S_off(i,off) a[(OPJ_UINT32)(i)*2*4+off]
+#define OPJ_D_off(i,off) a[(1+(OPJ_UINT32)(i)*2)*4+off]
+#define OPJ_S__off(i,off) ((i)<0?OPJ_S_off(0,off):((i)>=sn?OPJ_S_off(sn-1,off):OPJ_S_off(i,off)))
+#define OPJ_D__off(i,off) ((i)<0?OPJ_D_off(0,off):((i)>=dn?OPJ_D_off(dn-1,off):OPJ_D_off(i,off)))
+#define OPJ_SS__off(i,off) ((i)<0?OPJ_S_off(0,off):((i)>=dn?OPJ_S_off(dn-1,off):OPJ_S_off(i,off)))
+#define OPJ_DD__off(i,off) ((i)<0?OPJ_D_off(0,off):((i)>=sn?OPJ_D_off(sn-1,off):OPJ_D_off(i,off)))
+
+static void opj_dwt_decode_partial_1_parallel(OPJ_INT32 *a,
+ OPJ_UINT32 nb_cols,
+ OPJ_INT32 dn, OPJ_INT32 sn,
+ OPJ_INT32 cas,
+ OPJ_INT32 win_l_x0,
+ OPJ_INT32 win_l_x1,
+ OPJ_INT32 win_h_x0,
+ OPJ_INT32 win_h_x1)
+{
+ OPJ_INT32 i;
+ OPJ_UINT32 off;
+
+ (void)nb_cols;
+
+ if (!cas) {
+ if ((dn > 0) || (sn > 1)) { /* NEW : CASE ONE ELEMENT */
+
+ /* Naive version is :
+ for (i = win_l_x0; i < i_max; i++) {
+ OPJ_S(i) -= (OPJ_D_(i - 1) + OPJ_D_(i) + 2) >> 2;
+ }
+ for (i = win_h_x0; i < win_h_x1; i++) {
+ OPJ_D(i) += (OPJ_S_(i) + OPJ_S_(i + 1)) >> 1;
+ }
+ but the compiler doesn't manage to unroll it to avoid bound
+ checking in OPJ_S_ and OPJ_D_ macros
+ */
+
+ i = win_l_x0;
+ if (i < win_l_x1) {
+ OPJ_INT32 i_max;
+
+ /* Left-most case */
+ for (off = 0; off < 4; off++) {
+ OPJ_S_off(i, off) -= (OPJ_D__off(i - 1, off) + OPJ_D__off(i, off) + 2) >> 2;
+ }
+ i ++;
+
+ i_max = win_l_x1;
+ if (i_max > dn) {
+ i_max = dn;
+ }
+
+#ifdef __SSE2__
+ if (i + 1 < i_max) {
+ const __m128i two = _mm_set1_epi32(2);
+ __m128i Dm1 = _mm_load_si128((__m128i * const)(a + 4 + (i - 1) * 8));
+ for (; i + 1 < i_max; i += 2) {
+ /* No bound checking */
+ __m128i S = _mm_load_si128((__m128i * const)(a + i * 8));
+ __m128i D = _mm_load_si128((__m128i * const)(a + 4 + i * 8));
+ __m128i S1 = _mm_load_si128((__m128i * const)(a + (i + 1) * 8));
+ __m128i D1 = _mm_load_si128((__m128i * const)(a + 4 + (i + 1) * 8));
+ S = _mm_sub_epi32(S,
+ _mm_srai_epi32(_mm_add_epi32(_mm_add_epi32(Dm1, D), two), 2));
+ S1 = _mm_sub_epi32(S1,
+ _mm_srai_epi32(_mm_add_epi32(_mm_add_epi32(D, D1), two), 2));
+ _mm_store_si128((__m128i*)(a + i * 8), S);
+ _mm_store_si128((__m128i*)(a + (i + 1) * 8), S1);
+ Dm1 = D1;
+ }
+ }
+#endif
+
+ for (; i < i_max; i++) {
+ /* No bound checking */
+ for (off = 0; off < 4; off++) {
+ OPJ_S_off(i, off) -= (OPJ_D_off(i - 1, off) + OPJ_D_off(i, off) + 2) >> 2;
+ }
+ }
+ for (; i < win_l_x1; i++) {
+ /* Right-most case */
+ for (off = 0; off < 4; off++) {
+ OPJ_S_off(i, off) -= (OPJ_D__off(i - 1, off) + OPJ_D__off(i, off) + 2) >> 2;
+ }
+ }
+ }
+
+ i = win_h_x0;
+ if (i < win_h_x1) {
+ OPJ_INT32 i_max = win_h_x1;
+ if (i_max >= sn) {
+ i_max = sn - 1;
+ }
+
+#ifdef __SSE2__
+ if (i + 1 < i_max) {
+ __m128i S = _mm_load_si128((__m128i * const)(a + i * 8));
+ for (; i + 1 < i_max; i += 2) {
+ /* No bound checking */
+ __m128i D = _mm_load_si128((__m128i * const)(a + 4 + i * 8));
+ __m128i S1 = _mm_load_si128((__m128i * const)(a + (i + 1) * 8));
+ __m128i D1 = _mm_load_si128((__m128i * const)(a + 4 + (i + 1) * 8));
+ __m128i S2 = _mm_load_si128((__m128i * const)(a + (i + 2) * 8));
+ D = _mm_add_epi32(D, _mm_srai_epi32(_mm_add_epi32(S, S1), 1));
+ D1 = _mm_add_epi32(D1, _mm_srai_epi32(_mm_add_epi32(S1, S2), 1));
+ _mm_store_si128((__m128i*)(a + 4 + i * 8), D);
+ _mm_store_si128((__m128i*)(a + 4 + (i + 1) * 8), D1);
+ S = S2;
+ }
+ }
+#endif
+
+ for (; i < i_max; i++) {
+ /* No bound checking */
+ for (off = 0; off < 4; off++) {
+ OPJ_D_off(i, off) += (OPJ_S_off(i, off) + OPJ_S_off(i + 1, off)) >> 1;
+ }
+ }
+ for (; i < win_h_x1; i++) {
+ /* Right-most case */
+ for (off = 0; off < 4; off++) {
+ OPJ_D_off(i, off) += (OPJ_S__off(i, off) + OPJ_S__off(i + 1, off)) >> 1;
+ }
+ }
+ }
+ }
+ } else {
+ if (!sn && dn == 1) { /* NEW : CASE ONE ELEMENT */
+ for (off = 0; off < 4; off++) {
+ OPJ_S_off(0, off) /= 2;
+ }
+ } else {
+ for (i = win_l_x0; i < win_l_x1; i++) {
+ for (off = 0; off < 4; off++) {
+ OPJ_D_off(i, off) -= (OPJ_SS__off(i, off) + OPJ_SS__off(i + 1, off) + 2) >> 2;
+ }
+ }
+ for (i = win_h_x0; i < win_h_x1; i++) {
+ for (off = 0; off < 4; off++) {
+ OPJ_S_off(i, off) += (OPJ_DD__off(i, off) + OPJ_DD__off(i - 1, off)) >> 1;
+ }
+ }
+ }
+ }
+}
+
static void opj_dwt_get_band_coordinates(opj_tcd_tilecomp_t* tilec,
OPJ_UINT32 resno,
OPJ_UINT32 bandno,
OPJ_UINT32 rh = (OPJ_UINT32)(tr->y1 -
tr->y0); /* height of the resolution level computed */
- size_t h_mem_size;
+ OPJ_SIZE_T h_mem_size;
/* Compute the intersection of the area of interest, expressed in tile coordinates */
/* with the tile coordinates */
OPJ_UINT32 win_tcx1 = tilec->win_x1;
OPJ_UINT32 win_tcy1 = tilec->win_y1;
+ if (tr_max->x0 == tr_max->x1 || tr_max->y0 == tr_max->y1) {
+ return OPJ_TRUE;
+ }
+
sa = opj_dwt_init_sparse_array(tilec, numres);
+ if (sa == NULL) {
+ return OPJ_FALSE;
+ }
if (numres == 1U) {
OPJ_BOOL ret = opj_sparse_array_int32_read(sa,
1, tr_max->win_x1 - tr_max->win_x0,
OPJ_TRUE);
assert(ret);
+ OPJ_UNUSED(ret);
opj_sparse_array_int32_free(sa);
return OPJ_TRUE;
}
h_mem_size = opj_dwt_max_resolution(tr, numres);
/* overflow check */
- if (h_mem_size > (SIZE_MAX / sizeof(OPJ_INT32))) {
+ /* in vertical pass, we process 4 columns at a time */
+ if (h_mem_size > (SIZE_MAX / (4 * sizeof(OPJ_INT32)))) {
/* FIXME event manager error callback */
opj_sparse_array_int32_free(sa);
return OPJ_FALSE;
}
- h_mem_size *= sizeof(OPJ_INT32);
+ h_mem_size *= 4 * sizeof(OPJ_INT32);
h.mem = (OPJ_INT32*)opj_aligned_32_malloc(h_mem_size);
if (! h.mem) {
/* FIXME event manager error callback */
}
}
- for (i = win_tr_x0; i < win_tr_x1; ++i) {
+ for (i = win_tr_x0; i < win_tr_x1;) {
+ OPJ_UINT32 nb_cols = opj_uint_min(4U, win_tr_x1 - i);
opj_dwt_interleave_partial_v(v.mem,
v.cas,
sa,
i,
+ nb_cols,
(OPJ_UINT32)v.sn,
win_ll_y0,
win_ll_y1,
win_lh_y0,
win_lh_y1);
- opj_dwt_decode_partial_1(v.mem, v.dn, v.sn, v.cas,
- (OPJ_INT32)win_ll_y0,
- (OPJ_INT32)win_ll_y1,
- (OPJ_INT32)win_lh_y0,
- (OPJ_INT32)win_lh_y1);
+ opj_dwt_decode_partial_1_parallel(v.mem, nb_cols, v.dn, v.sn, v.cas,
+ (OPJ_INT32)win_ll_y0,
+ (OPJ_INT32)win_ll_y1,
+ (OPJ_INT32)win_lh_y0,
+ (OPJ_INT32)win_lh_y1);
if (!opj_sparse_array_int32_write(sa,
i, win_tr_y0,
- i + 1, win_tr_y1,
- v.mem + win_tr_y0,
- 0, 1, OPJ_TRUE)) {
+ i + nb_cols, win_tr_y1,
+ v.mem + 4 * win_tr_y0,
+ 1, 4, OPJ_TRUE)) {
/* FIXME event manager error callback */
opj_sparse_array_int32_free(sa);
opj_aligned_free(h.mem);
return OPJ_FALSE;
}
+
+ i += nb_cols;
}
}
opj_aligned_free(h.mem);
1, tr_max->win_x1 - tr_max->win_x0,
OPJ_TRUE);
assert(ret);
+ OPJ_UNUSED(ret);
}
opj_sparse_array_int32_free(sa);
return OPJ_TRUE;
OPJ_UINT32 x1 = dwt->win_l_x1;
for (k = 0; k < 2; ++k) {
- if (remaining_height >= 4 && ((size_t) a & 0x0f) == 0 &&
- ((size_t) bi & 0x0f) == 0 && (width & 0x0f) == 0) {
+ if (remaining_height >= 4 && ((OPJ_SIZE_T) a & 0x0f) == 0 &&
+ ((OPJ_SIZE_T) bi & 0x0f) == 0 && (width & 0x0f) == 0) {
/* Fast code path */
for (i = x0; i < x1; ++i) {
OPJ_UINT32 j = i;
(OPJ_INT32*)(dwt->wavelet + 1 - dwt->cas + 2 * dwt->win_h_x0) + i,
8, 0, OPJ_TRUE);
assert(ret);
+ OPJ_UNUSED(ret);
}
}
OPJ_UINT32 i;
for (i = dwt->win_l_x0; i < dwt->win_l_x1; ++i) {
- memcpy(&bi[i * 2], &a[i * width], (size_t)nb_elts_read * sizeof(OPJ_FLOAT32));
+ memcpy(&bi[i * 2], &a[i * (OPJ_SIZE_T)width],
+ (OPJ_SIZE_T)nb_elts_read * sizeof(OPJ_FLOAT32));
}
- a += (OPJ_UINT32)dwt->sn * width;
+ a += (OPJ_UINT32)dwt->sn * (OPJ_SIZE_T)width;
bi = dwt->wavelet + 1 - dwt->cas;
for (i = dwt->win_h_x0; i < dwt->win_h_x1; ++i) {
- memcpy(&bi[i * 2], &a[i * width], (size_t)nb_elts_read * sizeof(OPJ_FLOAT32));
+ memcpy(&bi[i * 2], &a[i * (OPJ_SIZE_T)width],
+ (OPJ_SIZE_T)nb_elts_read * sizeof(OPJ_FLOAT32));
}
}
OPJ_UINT32 sa_col,
OPJ_UINT32 nb_elts_read)
{
- OPJ_UINT32 i;
- for (i = 0; i < nb_elts_read; i++) {
- OPJ_BOOL ret;
- ret = opj_sparse_array_int32_read(sa,
- sa_col + i, dwt->win_l_x0,
- sa_col + i + 1, dwt->win_l_x1,
- (OPJ_INT32*)(dwt->wavelet + dwt->cas + 2 * dwt->win_l_x0) + i,
- 0, 8, OPJ_TRUE);
- assert(ret);
- ret = opj_sparse_array_int32_read(sa,
- sa_col + i, (OPJ_UINT32)dwt->sn + dwt->win_h_x0,
- sa_col + i + 1, (OPJ_UINT32)dwt->sn + dwt->win_h_x1,
- (OPJ_INT32*)(dwt->wavelet + 1 - dwt->cas + 2 * dwt->win_h_x0) + i,
- 0, 8, OPJ_TRUE);
- assert(ret);
- }
+ OPJ_BOOL ret;
+ ret = opj_sparse_array_int32_read(sa,
+ sa_col, dwt->win_l_x0,
+ sa_col + nb_elts_read, dwt->win_l_x1,
+ (OPJ_INT32*)(dwt->wavelet + dwt->cas + 2 * dwt->win_l_x0),
+ 1, 8, OPJ_TRUE);
+ assert(ret);
+ ret = opj_sparse_array_int32_read(sa,
+ sa_col, (OPJ_UINT32)dwt->sn + dwt->win_h_x0,
+ sa_col + nb_elts_read, (OPJ_UINT32)dwt->sn + dwt->win_h_x1,
+ (OPJ_INT32*)(dwt->wavelet + 1 - dwt->cas + 2 * dwt->win_h_x0),
+ 1, 8, OPJ_TRUE);
+ assert(ret);
+ OPJ_UNUSED(ret);
}
#ifdef __SSE__
__m128* OPJ_RESTRICT vw = (__m128*) w;
OPJ_UINT32 i;
/* 4x unrolled loop */
- for (i = start; i + 3 < end; i += 4) {
- vw[2 * i] = _mm_mul_ps(vw[2 * i], c);
- vw[2 * i + 2] = _mm_mul_ps(vw[2 * i + 2], c);
- vw[2 * i + 4] = _mm_mul_ps(vw[2 * i + 4], c);
- vw[2 * i + 6] = _mm_mul_ps(vw[2 * i + 6], c);
- }
- for (; i < end; ++i) {
- vw[2 * i] = _mm_mul_ps(vw[2 * i], c);
+ vw += 2 * start;
+ for (i = start; i + 3 < end; i += 4, vw += 8) {
+ __m128 xmm0 = _mm_mul_ps(vw[0], c);
+ __m128 xmm2 = _mm_mul_ps(vw[2], c);
+ __m128 xmm4 = _mm_mul_ps(vw[4], c);
+ __m128 xmm6 = _mm_mul_ps(vw[6], c);
+ vw[0] = xmm0;
+ vw[2] = xmm2;
+ vw[4] = xmm4;
+ vw[6] = xmm6;
+ }
+ for (; i < end; ++i, vw += 2) {
+ vw[0] = _mm_mul_ps(vw[0], c);
}
}
vw += start * 2;
tmp1 = vw[-3];
}
- for (i = start; i < imax; ++i) {
+
+ i = start;
+
+ /* 4x loop unrolling */
+ for (; i + 3 < imax; i += 4) {
+ __m128 tmp4, tmp5, tmp6, tmp7, tmp8, tmp9;
+ tmp2 = vw[-1];
+ tmp3 = vw[ 0];
+ tmp4 = vw[ 1];
+ tmp5 = vw[ 2];
+ tmp6 = vw[ 3];
+ tmp7 = vw[ 4];
+ tmp8 = vw[ 5];
+ tmp9 = vw[ 6];
+ vw[-1] = _mm_add_ps(tmp2, _mm_mul_ps(_mm_add_ps(tmp1, tmp3), c));
+ vw[ 1] = _mm_add_ps(tmp4, _mm_mul_ps(_mm_add_ps(tmp3, tmp5), c));
+ vw[ 3] = _mm_add_ps(tmp6, _mm_mul_ps(_mm_add_ps(tmp5, tmp7), c));
+ vw[ 5] = _mm_add_ps(tmp8, _mm_mul_ps(_mm_add_ps(tmp7, tmp9), c));
+ tmp1 = tmp9;
+ vw += 8;
+ }
+
+ for (; i < imax; ++i) {
tmp2 = vw[-1];
tmp3 = vw[ 0];
vw[-1] = _mm_add_ps(tmp2, _mm_mul_ps(_mm_add_ps(tmp1, tmp3), c));
1].x1 -
tilec->resolutions[tilec->minimum_num_resolutions - 1].x0);
- size_t l_data_size;
+ OPJ_SIZE_T l_data_size;
l_data_size = opj_dwt_max_resolution(res, numres);
/* overflow check */
for (k = 0; k < rw; k++) {
aj[k ] = h.wavelet[k].f[0];
- aj[k + w ] = h.wavelet[k].f[1];
- aj[k + w * 2] = h.wavelet[k].f[2];
- aj[k + w * 3] = h.wavelet[k].f[3];
+ aj[k + (OPJ_SIZE_T)w ] = h.wavelet[k].f[1];
+ aj[k + (OPJ_SIZE_T)w * 2] = h.wavelet[k].f[2];
+ aj[k + (OPJ_SIZE_T)w * 3] = h.wavelet[k].f[3];
}
aj += w * 4;
for (k = 0; k < rw; k++) {
switch (rh - j) {
case 3:
- aj[k + w * 2] = h.wavelet[k].f[2];
+ aj[k + (OPJ_SIZE_T)w * 2] = h.wavelet[k].f[2];
/* FALLTHRU */
case 2:
- aj[k + w ] = h.wavelet[k].f[1];
+ aj[k + (OPJ_SIZE_T)w ] = h.wavelet[k].f[1];
/* FALLTHRU */
case 1:
aj[k] = h.wavelet[k].f[0];
opj_v4dwt_decode(&v);
for (k = 0; k < rh; ++k) {
- memcpy(&aj[k * w], &v.wavelet[k], 4 * sizeof(OPJ_FLOAT32));
+ memcpy(&aj[k * (OPJ_SIZE_T)w], &v.wavelet[k], 4 * sizeof(OPJ_FLOAT32));
}
aj += 4;
}
opj_v4dwt_decode(&v);
for (k = 0; k < rh; ++k) {
- memcpy(&aj[k * w], &v.wavelet[k], (size_t)j * sizeof(OPJ_FLOAT32));
+ memcpy(&aj[k * (OPJ_SIZE_T)w], &v.wavelet[k],
+ (OPJ_SIZE_T)j * sizeof(OPJ_FLOAT32));
}
}
}
OPJ_UINT32 rh = (OPJ_UINT32)(tr->y1 -
tr->y0); /* height of the resolution level computed */
- size_t l_data_size;
+ OPJ_SIZE_T l_data_size;
/* Compute the intersection of the area of interest, expressed in tile coordinates */
/* with the tile coordinates */
OPJ_UINT32 win_tcx1 = tilec->win_x1;
OPJ_UINT32 win_tcy1 = tilec->win_y1;
+ if (tr_max->x0 == tr_max->x1 || tr_max->y0 == tr_max->y1) {
+ return OPJ_TRUE;
+ }
+
sa = opj_dwt_init_sparse_array(tilec, numres);
+ if (sa == NULL) {
+ return OPJ_FALSE;
+ }
if (numres == 1U) {
OPJ_BOOL ret = opj_sparse_array_int32_read(sa,
1, tr_max->win_x1 - tr_max->win_x0,
OPJ_TRUE);
assert(ret);
+ OPJ_UNUSED(ret);
opj_sparse_array_int32_free(sa);
return OPJ_TRUE;
}
if ((j + 3 >= win_ll_y0 && j < win_ll_y1) ||
(j + 3 >= win_lh_y0 + (OPJ_UINT32)v.sn &&
j < win_lh_y1 + (OPJ_UINT32)v.sn)) {
- OPJ_UINT32 k;
opj_v4dwt_interleave_partial_h(&h, sa, j, opj_uint_min(4U, rh - j));
opj_v4dwt_decode(&h);
- for (k = 0; k < 4; k++) {
- if (!opj_sparse_array_int32_write(sa,
- win_tr_x0, j + k,
- win_tr_x1, j + k + 1,
- (OPJ_INT32*)&h.wavelet[win_tr_x0].f[k],
- 4, 0, OPJ_TRUE)) {
- /* FIXME event manager error callback */
- opj_sparse_array_int32_free(sa);
- opj_aligned_free(h.wavelet);
- return OPJ_FALSE;
- }
+ if (!opj_sparse_array_int32_write(sa,
+ win_tr_x0, j,
+ win_tr_x1, j + 4,
+ (OPJ_INT32*)&h.wavelet[win_tr_x0].f[0],
+ 4, 1, OPJ_TRUE)) {
+ /* FIXME event manager error callback */
+ opj_sparse_array_int32_free(sa);
+ opj_aligned_free(h.wavelet);
+ return OPJ_FALSE;
}
}
}
((j + 3 >= win_ll_y0 && j < win_ll_y1) ||
(j + 3 >= win_lh_y0 + (OPJ_UINT32)v.sn &&
j < win_lh_y1 + (OPJ_UINT32)v.sn))) {
- OPJ_UINT32 k;
opj_v4dwt_interleave_partial_h(&h, sa, j, rh - j);
opj_v4dwt_decode(&h);
- for (k = 0; k < rh - j; k++) {
- if (!opj_sparse_array_int32_write(sa,
- win_tr_x0, j + k,
- win_tr_x1, j + k + 1,
- (OPJ_INT32*)&h.wavelet[win_tr_x0].f[k],
- 4, 0, OPJ_TRUE)) {
- /* FIXME event manager error callback */
- opj_sparse_array_int32_free(sa);
- opj_aligned_free(h.wavelet);
- return OPJ_FALSE;
- }
+ if (!opj_sparse_array_int32_write(sa,
+ win_tr_x0, j,
+ win_tr_x1, rh,
+ (OPJ_INT32*)&h.wavelet[win_tr_x0].f[0],
+ 4, 1, OPJ_TRUE)) {
+ /* FIXME event manager error callback */
+ opj_sparse_array_int32_free(sa);
+ opj_aligned_free(h.wavelet);
+ return OPJ_FALSE;
}
}
v.win_h_x1 = win_lh_y1;
for (j = win_tr_x0; j < win_tr_x1; j += 4) {
OPJ_UINT32 nb_elts = opj_uint_min(4U, win_tr_x1 - j);
- OPJ_UINT32 k;
opj_v4dwt_interleave_partial_v(&v, sa, j, nb_elts);
opj_v4dwt_decode(&v);
- for (k = 0; k < nb_elts; k++) {
- if (!opj_sparse_array_int32_write(sa,
- j + k, win_tr_y0,
- j + k + 1, win_tr_y1,
- (OPJ_INT32*)&h.wavelet[win_tr_y0].f[k],
- 0, 4, OPJ_TRUE)) {
- /* FIXME event manager error callback */
- opj_sparse_array_int32_free(sa);
- opj_aligned_free(h.wavelet);
- return OPJ_FALSE;
- }
+ if (!opj_sparse_array_int32_write(sa,
+ j, win_tr_y0,
+ j + nb_elts, win_tr_y1,
+ (OPJ_INT32*)&h.wavelet[win_tr_y0].f[0],
+ 1, 4, OPJ_TRUE)) {
+ /* FIXME event manager error callback */
+ opj_sparse_array_int32_free(sa);
+ opj_aligned_free(h.wavelet);
+ return OPJ_FALSE;
}
}
}
1, tr_max->win_x1 - tr_max->win_x0,
OPJ_TRUE);
assert(ret);
+ OPJ_UNUSED(ret);
}
opj_sparse_array_int32_free(sa);