/*@}*/
-/**
-Virtual function type for wavelet transform in 1-D
-*/
-typedef void (*DWT1DFN)(const opj_dwt_t* v);
-
/** @name Local static functions */
/*@{*/
OPJ_INT32 * OPJ_RESTRICT b,
OPJ_INT32 dn,
OPJ_INT32 sn, OPJ_INT32 cas);
-/**
-Forward lazy transform (vertical)
-*/
-static void opj_dwt_deinterleave_v(const OPJ_INT32 * OPJ_RESTRICT a,
- OPJ_INT32 * OPJ_RESTRICT b,
- OPJ_INT32 dn,
- OPJ_INT32 sn, OPJ_UINT32 x, OPJ_INT32 cas);
-/**
-Forward 5-3 wavelet transform in 1-D
-*/
-static void opj_dwt_encode_1(void *a, OPJ_INT32 dn, OPJ_INT32 sn,
- OPJ_INT32 cas);
+
/**
Forward 9-7 wavelet transform in 1-D
*/
opj_tcd_tilecomp_t* tilec,
OPJ_UINT32 numres);
+/* Forward transform, for the vertical pass, processing cols columns */
+/* where cols <= NB_ELTS_V8 */
/* Where void* is a OPJ_INT32* for 5x3 and OPJ_FLOAT32* for 9x7 */
-typedef void (*opj_encode_one_row_fnptr_type)(void *, OPJ_INT32, OPJ_INT32,
- OPJ_INT32);
+typedef void (*opj_encode_and_deinterleave_v_fnptr_type)(
+ void *array,
+ void *tmp,
+ OPJ_UINT32 height,
+ OPJ_BOOL even,
+ OPJ_UINT32 stride_width,
+ OPJ_UINT32 cols);
+/* Where void* is a OPJ_INT32* for 5x3 and OPJ_FLOAT32* for 9x7 */
typedef void (*opj_encode_and_deinterleave_h_one_row_fnptr_type)(
void *row,
void *tmp,
static OPJ_BOOL opj_dwt_encode_procedure(opj_thread_pool_t* tp,
opj_tcd_tilecomp_t * tilec,
- opj_encode_one_row_fnptr_type p_function,
+ opj_encode_and_deinterleave_v_fnptr_type p_encode_and_deinterleave_v,
opj_encode_and_deinterleave_h_one_row_fnptr_type
p_encode_and_deinterleave_h_one_row);
}
}
-/* <summary> */
-/* Forward lazy transform (vertical). */
-/* </summary> */
-static void opj_dwt_deinterleave_v(const OPJ_INT32 * OPJ_RESTRICT a,
- OPJ_INT32 * OPJ_RESTRICT b,
- OPJ_INT32 dn,
- OPJ_INT32 sn, OPJ_UINT32 x, OPJ_INT32 cas)
-{
- OPJ_INT32 i = sn;
- OPJ_INT32 * OPJ_RESTRICT l_dest = b;
- const OPJ_INT32 * OPJ_RESTRICT l_src = a + cas;
-
- while (i--) {
- *l_dest = *l_src;
- l_dest += x;
- l_src += 2;
- } /* b[i*x]=a[2*i+cas]; */
-
- l_dest = b + (OPJ_SIZE_T)sn * (OPJ_SIZE_T)x;
- l_src = a + 1 - cas;
-
- i = dn;
- while (i--) {
- *l_dest = *l_src;
- l_dest += x;
- l_src += 2;
- } /*b[(sn+i)*x]=a[(2*i+1-cas)];*/
-}
-
#ifdef STANDARD_SLOW_VERSION
/* <summary> */
/* Inverse lazy transform (horizontal). */
#endif /* STANDARD_SLOW_VERSION */
-/* <summary> */
-/* Forward 5-3 wavelet transform in 1-D. */
-/* </summary> */
-static void opj_dwt_encode_1(void *aIn, OPJ_INT32 dn, OPJ_INT32 sn,
- OPJ_INT32 cas)
-{
- OPJ_INT32 i;
- OPJ_INT32* a = (OPJ_INT32*)aIn;
-
- if (!cas) {
- if ((dn > 0) || (sn > 1)) { /* NEW : CASE ONE ELEMENT */
- for (i = 0; i < dn; i++) {
- OPJ_D(i) -= (OPJ_S_(i) + OPJ_S_(i + 1)) >> 1;
- }
- for (i = 0; i < sn; i++) {
- OPJ_S(i) += (OPJ_D_(i - 1) + OPJ_D_(i) + 2) >> 2;
- }
- }
- } else {
- if (!sn && dn == 1) { /* NEW : CASE ONE ELEMENT */
- OPJ_S(0) *= 2;
- } else {
- for (i = 0; i < dn; i++) {
- OPJ_S(i) -= (OPJ_DD_(i) + OPJ_DD_(i - 1)) >> 1;
- }
- for (i = 0; i < sn; i++) {
- OPJ_D(i) += (OPJ_SS_(i) + OPJ_SS_(i + 1) + 2) >> 2;
- }
- }
- }
-}
-
#ifdef STANDARD_SLOW_VERSION
/* <summary> */
/* Inverse 5-3 wavelet transform in 1-D. */
s0n = s1n - ((d1c + d1n + 2) >> 2);
tmp[i ] = s0c;
- tmp[i + 1] = d1c + ((s0c + s0n) >> 1);
+ tmp[i + 1] = opj_int_add_no_overflow(d1c, opj_int_add_no_overflow(s0c,
+ s0n) >> 1);
}
tmp[i] = s0n;
dn = in_odd[j] - ((s1 + s2 + 2) >> 2);
tmp[i ] = dc;
- tmp[i + 1] = s1 + ((dn + dc) >> 1);
+ tmp[i + 1] = opj_int_add_no_overflow(s1, opj_int_add_no_overflow(dn, dc) >> 1);
dc = dn;
s1 = s2;
#if (defined(__SSE2__) || defined(__AVX2__)) && !defined(STANDARD_SLOW_VERSION)
-/* Conveniency macros to improve the readabilty of the formulas */
+/* Conveniency macros to improve the readability of the formulas */
#if __AVX2__
#define VREG __m256i
#define LOAD_CST(x) _mm256_set1_epi32(x)
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);
+ s0n = opj_int_sub_no_overflow(s1n,
+ opj_int_add_no_overflow(opj_int_add_no_overflow(d1c, d1n), 2) >> 2);
tmp[i ] = s0c;
tmp[i + 1] = d1c + ((s0c + s0n) >> 1);
#endif
}
+#if 0
static void opj_dwt_encode_step1(OPJ_FLOAT32* fw,
- OPJ_UINT32 start,
OPJ_UINT32 end,
const OPJ_FLOAT32 c)
{
- OPJ_UINT32 i;
- for (i = start; i < end; ++i) {
- fw[i * 2] *= c;
+ OPJ_UINT32 i = 0;
+ for (; i < end; ++i) {
+ fw[0] *= c;
+ fw += 2;
}
}
+#else
+static void opj_dwt_encode_step1_combined(OPJ_FLOAT32* fw,
+ OPJ_UINT32 iters_c1,
+ OPJ_UINT32 iters_c2,
+ const OPJ_FLOAT32 c1,
+ const OPJ_FLOAT32 c2)
+{
+ OPJ_UINT32 i = 0;
+ const OPJ_UINT32 iters_common = opj_uint_min(iters_c1, iters_c2);
+ assert((((OPJ_SIZE_T)fw) & 0xf) == 0);
+ assert(opj_int_abs((OPJ_INT32)iters_c1 - (OPJ_INT32)iters_c2) <= 1);
+ for (; i + 3 < iters_common; i += 4) {
+#ifdef __SSE__
+ const __m128 vcst = _mm_set_ps(c2, c1, c2, c1);
+ *(__m128*)fw = _mm_mul_ps(*(__m128*)fw, vcst);
+ *(__m128*)(fw + 4) = _mm_mul_ps(*(__m128*)(fw + 4), vcst);
+#else
+ fw[0] *= c1;
+ fw[1] *= c2;
+ fw[2] *= c1;
+ fw[3] *= c2;
+ fw[4] *= c1;
+ fw[5] *= c2;
+ fw[6] *= c1;
+ fw[7] *= c2;
+#endif
+ fw += 8;
+ }
+ for (; i < iters_common; i++) {
+ fw[0] *= c1;
+ fw[1] *= c2;
+ fw += 2;
+ }
+ if (i < iters_c1) {
+ fw[0] *= c1;
+ } else if (i < iters_c2) {
+ fw[1] *= c2;
+ }
+}
+
+#endif
+
static void opj_dwt_encode_step2(OPJ_FLOAT32* fl, OPJ_FLOAT32* fw,
- OPJ_UINT32 start,
OPJ_UINT32 end,
OPJ_UINT32 m,
OPJ_FLOAT32 c)
{
OPJ_UINT32 i;
OPJ_UINT32 imax = opj_uint_min(end, m);
- if (start > 0) {
- fw += 2 * start;
- fl = fw - 2;
- }
- for (i = start; i < imax; ++i) {
+ if (imax > 0) {
fw[-1] += (fl[0] + fw[0]) * c;
- fl = fw;
fw += 2;
+ i = 1;
+ for (; i + 3 < imax; i += 4) {
+ fw[-1] += (fw[-2] + fw[0]) * c;
+ fw[1] += (fw[0] + fw[2]) * c;
+ fw[3] += (fw[2] + fw[4]) * c;
+ fw[5] += (fw[4] + fw[6]) * c;
+ fw += 8;
+ }
+ for (; i < imax; ++i) {
+ fw[-1] += (fw[-2] + fw[0]) * c;
+ fw += 2;
+ }
}
if (m < end) {
assert(m + 1 == end);
- fw[-1] += (2 * fl[0]) * c;
+ fw[-1] += (2 * fw[-2]) * c;
}
}
{
OPJ_FLOAT32* w = (OPJ_FLOAT32*)aIn;
OPJ_INT32 a, b;
+ assert(dn + sn > 1);
if (cas == 0) {
- if (!((dn > 0) || (sn > 1))) {
- return;
- }
a = 0;
b = 1;
} else {
- if (!((sn > 0) || (dn > 1))) {
- return;
- }
a = 1;
b = 0;
}
opj_dwt_encode_step2(w + a, w + b + 1,
- 0, (OPJ_UINT32)dn,
+ (OPJ_UINT32)dn,
(OPJ_UINT32)opj_int_min(dn, sn - b),
opj_dwt_alpha);
opj_dwt_encode_step2(w + b, w + a + 1,
- 0, (OPJ_UINT32)sn,
+ (OPJ_UINT32)sn,
(OPJ_UINT32)opj_int_min(sn, dn - a),
opj_dwt_beta);
opj_dwt_encode_step2(w + a, w + b + 1,
- 0, (OPJ_UINT32)dn,
+ (OPJ_UINT32)dn,
(OPJ_UINT32)opj_int_min(dn, sn - b),
opj_dwt_gamma);
opj_dwt_encode_step2(w + b, w + a + 1,
- 0, (OPJ_UINT32)sn,
+ (OPJ_UINT32)sn,
(OPJ_UINT32)opj_int_min(sn, dn - a),
opj_dwt_delta);
- opj_dwt_encode_step1(w + b, 0, (OPJ_UINT32)dn,
+#if 0
+ opj_dwt_encode_step1(w + b, (OPJ_UINT32)dn,
opj_K);
- opj_dwt_encode_step1(w + a, 0, (OPJ_UINT32)sn,
+ opj_dwt_encode_step1(w + a, (OPJ_UINT32)sn,
opj_invK);
+#else
+ if (a == 0) {
+ opj_dwt_encode_step1_combined(w,
+ (OPJ_UINT32)sn,
+ (OPJ_UINT32)dn,
+ opj_invK,
+ opj_K);
+ } else {
+ opj_dwt_encode_step1_combined(w,
+ (OPJ_UINT32)dn,
+ (OPJ_UINT32)sn,
+ opj_K,
+ opj_invK);
+ }
+#endif
}
static void opj_dwt_encode_stepsize(OPJ_INT32 stepsize, OPJ_INT32 numbps,
OPJ_INT32* OPJ_RESTRICT tmp = (OPJ_INT32*)tmpIn;
const OPJ_INT32 sn = (OPJ_INT32)((width + (even ? 1 : 0)) >> 1);
const OPJ_INT32 dn = (OPJ_INT32)(width - (OPJ_UINT32)sn);
- memcpy(tmp, row, width * sizeof(OPJ_INT32));
- opj_dwt_encode_1(tmp, dn, sn, even ? 0 : 1);
- opj_dwt_deinterleave_h(tmp, row, dn, sn, even ? 0 : 1);
+
+ if (even) {
+ if (width > 1) {
+ OPJ_INT32 i;
+ for (i = 0; i < sn - 1; i++) {
+ tmp[sn + i] = row[2 * i + 1] - ((row[(i) * 2] + row[(i + 1) * 2]) >> 1);
+ }
+ if ((width % 2) == 0) {
+ tmp[sn + i] = row[2 * i + 1] - row[(i) * 2];
+ }
+ row[0] += (tmp[sn] + tmp[sn] + 2) >> 2;
+ for (i = 1; i < dn; i++) {
+ row[i] = row[2 * i] + ((tmp[sn + (i - 1)] + tmp[sn + i] + 2) >> 2);
+ }
+ if ((width % 2) == 1) {
+ row[i] = row[2 * i] + ((tmp[sn + (i - 1)] + tmp[sn + (i - 1)] + 2) >> 2);
+ }
+ memcpy(row + sn, tmp + sn, (OPJ_SIZE_T)dn * sizeof(OPJ_INT32));
+ }
+ } else {
+ if (width == 1) {
+ row[0] *= 2;
+ } else {
+ OPJ_INT32 i;
+ tmp[sn + 0] = row[0] - row[1];
+ for (i = 1; i < sn; i++) {
+ tmp[sn + i] = row[2 * i] - ((row[2 * i + 1] + row[2 * (i - 1) + 1]) >> 1);
+ }
+ if ((width % 2) == 1) {
+ tmp[sn + i] = row[2 * i] - row[2 * (i - 1) + 1];
+ }
+
+ for (i = 0; i < dn - 1; i++) {
+ row[i] = row[2 * i + 1] + ((tmp[sn + i] + tmp[sn + i + 1] + 2) >> 2);
+ }
+ if ((width % 2) == 0) {
+ row[i] = row[2 * i + 1] + ((tmp[sn + i] + tmp[sn + i] + 2) >> 2);
+ }
+ memcpy(row + sn, tmp + sn, (OPJ_SIZE_T)dn * sizeof(OPJ_INT32));
+ }
+ }
}
/** Process one line for the horizontal pass of the 9x7 forward transform */
OPJ_FLOAT32* OPJ_RESTRICT tmp = (OPJ_FLOAT32*)tmpIn;
const OPJ_INT32 sn = (OPJ_INT32)((width + (even ? 1 : 0)) >> 1);
const OPJ_INT32 dn = (OPJ_INT32)(width - (OPJ_UINT32)sn);
+ if (width == 1) {
+ return;
+ }
memcpy(tmp, row, width * sizeof(OPJ_FLOAT32));
opj_dwt_encode_1_real(tmp, dn, sn, even ? 0 : 1);
opj_dwt_deinterleave_h((OPJ_INT32 * OPJ_RESTRICT)tmp,
OPJ_INT32 * OPJ_RESTRICT tiledp;
OPJ_UINT32 min_j;
OPJ_UINT32 max_j;
- opj_encode_one_row_fnptr_type p_function;
+ opj_encode_and_deinterleave_v_fnptr_type p_encode_and_deinterleave_v;
} opj_dwt_encode_v_job_t;
static void opj_dwt_encode_v_func(void* user_data, opj_tls_t* tls)
(void)tls;
job = (opj_dwt_encode_v_job_t*)user_data;
- for (j = job->min_j; j < job->max_j; j++) {
- OPJ_INT32* OPJ_RESTRICT aj = job->tiledp + j;
+ for (j = job->min_j; j + NB_ELTS_V8 - 1 < job->max_j; j += NB_ELTS_V8) {
+ (*job->p_encode_and_deinterleave_v)(job->tiledp + j,
+ job->v.mem,
+ job->rh,
+ job->v.cas == 0,
+ job->w,
+ NB_ELTS_V8);
+ }
+ if (j < job->max_j) {
+ (*job->p_encode_and_deinterleave_v)(job->tiledp + j,
+ job->v.mem,
+ job->rh,
+ job->v.cas == 0,
+ job->w,
+ job->max_j - j);
+ }
+
+ opj_aligned_free(job->v.mem);
+ opj_free(job);
+}
+
+/** Fetch up to cols <= NB_ELTS_V8 for each line, and put them in tmpOut */
+/* that has a NB_ELTS_V8 interleave factor. */
+static void opj_dwt_fetch_cols_vertical_pass(const void *arrayIn,
+ void *tmpOut,
+ OPJ_UINT32 height,
+ OPJ_UINT32 stride_width,
+ OPJ_UINT32 cols)
+{
+ const OPJ_INT32* OPJ_RESTRICT array = (const OPJ_INT32 * OPJ_RESTRICT)arrayIn;
+ OPJ_INT32* OPJ_RESTRICT tmp = (OPJ_INT32 * OPJ_RESTRICT)tmpOut;
+ if (cols == NB_ELTS_V8) {
OPJ_UINT32 k;
- for (k = 0; k < job->rh; ++k) {
- job->v.mem[k] = aj[k * job->w];
+ for (k = 0; k < height; ++k) {
+ memcpy(tmp + NB_ELTS_V8 * k,
+ array + k * stride_width,
+ NB_ELTS_V8 * sizeof(OPJ_INT32));
}
+ } else {
+ OPJ_UINT32 k;
+ for (k = 0; k < height; ++k) {
+ OPJ_UINT32 c;
+ for (c = 0; c < cols; c++) {
+ tmp[NB_ELTS_V8 * k + c] = array[c + k * stride_width];
+ }
+ for (; c < NB_ELTS_V8; c++) {
+ tmp[NB_ELTS_V8 * k + c] = 0;
+ }
+ }
+ }
+}
- (*job->p_function)(job->v.mem, job->v.dn, job->v.sn, job->v.cas);
+/* Deinterleave result of forward transform, where cols <= NB_ELTS_V8 */
+/* and src contains NB_ELTS_V8 consecutive values for up to NB_ELTS_V8 */
+/* columns. */
+static INLINE void opj_dwt_deinterleave_v_cols(
+ const OPJ_INT32 * OPJ_RESTRICT src,
+ OPJ_INT32 * OPJ_RESTRICT dst,
+ OPJ_INT32 dn,
+ OPJ_INT32 sn,
+ OPJ_UINT32 stride_width,
+ OPJ_INT32 cas,
+ OPJ_UINT32 cols)
+{
+ OPJ_INT32 k;
+ OPJ_INT32 i = sn;
+ OPJ_INT32 * OPJ_RESTRICT l_dest = dst;
+ const OPJ_INT32 * OPJ_RESTRICT l_src = src + cas * NB_ELTS_V8;
+ OPJ_UINT32 c;
+
+ for (k = 0; k < 2; k++) {
+ while (i--) {
+ if (cols == NB_ELTS_V8) {
+ memcpy(l_dest, l_src, NB_ELTS_V8 * sizeof(OPJ_INT32));
+ } else {
+ c = 0;
+ switch (cols) {
+ case 7:
+ l_dest[c] = l_src[c];
+ c++; /* fallthru */
+ case 6:
+ l_dest[c] = l_src[c];
+ c++; /* fallthru */
+ case 5:
+ l_dest[c] = l_src[c];
+ c++; /* fallthru */
+ case 4:
+ l_dest[c] = l_src[c];
+ c++; /* fallthru */
+ case 3:
+ l_dest[c] = l_src[c];
+ c++; /* fallthru */
+ case 2:
+ l_dest[c] = l_src[c];
+ c++; /* fallthru */
+ default:
+ l_dest[c] = l_src[c];
+ break;
+ }
+ }
+ l_dest += stride_width;
+ l_src += 2 * NB_ELTS_V8;
+ }
- opj_dwt_deinterleave_v(job->v.mem, aj, job->v.dn, job->v.sn, job->w,
- job->v.cas);
+ l_dest = dst + (OPJ_SIZE_T)sn * (OPJ_SIZE_T)stride_width;
+ l_src = src + (1 - cas) * NB_ELTS_V8;
+ i = dn;
}
+}
- opj_aligned_free(job->v.mem);
- opj_free(job);
+
+/* Forward 5-3 transform, for the vertical pass, processing cols columns */
+/* where cols <= NB_ELTS_V8 */
+static void opj_dwt_encode_and_deinterleave_v(
+ void *arrayIn,
+ void *tmpIn,
+ OPJ_UINT32 height,
+ OPJ_BOOL even,
+ OPJ_UINT32 stride_width,
+ OPJ_UINT32 cols)
+{
+ OPJ_INT32* OPJ_RESTRICT array = (OPJ_INT32 * OPJ_RESTRICT)arrayIn;
+ OPJ_INT32* OPJ_RESTRICT tmp = (OPJ_INT32 * OPJ_RESTRICT)tmpIn;
+ const OPJ_UINT32 sn = (height + (even ? 1 : 0)) >> 1;
+ const OPJ_UINT32 dn = height - sn;
+
+ opj_dwt_fetch_cols_vertical_pass(arrayIn, tmpIn, height, stride_width, cols);
+
+#define OPJ_Sc(i) tmp[(i)*2* NB_ELTS_V8 + c]
+#define OPJ_Dc(i) tmp[((1+(i)*2))* NB_ELTS_V8 + c]
+
+#ifdef __SSE2__
+ if (height == 1) {
+ if (!even) {
+ OPJ_UINT32 c;
+ for (c = 0; c < NB_ELTS_V8; c++) {
+ tmp[c] *= 2;
+ }
+ }
+ } else if (even) {
+ OPJ_UINT32 c;
+ OPJ_UINT32 i;
+ i = 0;
+ if (i + 1 < sn) {
+ __m128i xmm_Si_0 = *(const __m128i*)(tmp + 4 * 0);
+ __m128i xmm_Si_1 = *(const __m128i*)(tmp + 4 * 1);
+ for (; i + 1 < sn; i++) {
+ __m128i xmm_Sip1_0 = *(const __m128i*)(tmp +
+ (i + 1) * 2 * NB_ELTS_V8 + 4 * 0);
+ __m128i xmm_Sip1_1 = *(const __m128i*)(tmp +
+ (i + 1) * 2 * NB_ELTS_V8 + 4 * 1);
+ __m128i xmm_Di_0 = *(const __m128i*)(tmp +
+ (1 + i * 2) * NB_ELTS_V8 + 4 * 0);
+ __m128i xmm_Di_1 = *(const __m128i*)(tmp +
+ (1 + i * 2) * NB_ELTS_V8 + 4 * 1);
+ xmm_Di_0 = _mm_sub_epi32(xmm_Di_0,
+ _mm_srai_epi32(_mm_add_epi32(xmm_Si_0, xmm_Sip1_0), 1));
+ xmm_Di_1 = _mm_sub_epi32(xmm_Di_1,
+ _mm_srai_epi32(_mm_add_epi32(xmm_Si_1, xmm_Sip1_1), 1));
+ *(__m128i*)(tmp + (1 + i * 2) * NB_ELTS_V8 + 4 * 0) = xmm_Di_0;
+ *(__m128i*)(tmp + (1 + i * 2) * NB_ELTS_V8 + 4 * 1) = xmm_Di_1;
+ xmm_Si_0 = xmm_Sip1_0;
+ xmm_Si_1 = xmm_Sip1_1;
+ }
+ }
+ if (((height) % 2) == 0) {
+ for (c = 0; c < NB_ELTS_V8; c++) {
+ OPJ_Dc(i) -= OPJ_Sc(i);
+ }
+ }
+ for (c = 0; c < NB_ELTS_V8; c++) {
+ OPJ_Sc(0) += (OPJ_Dc(0) + OPJ_Dc(0) + 2) >> 2;
+ }
+ i = 1;
+ if (i < dn) {
+ __m128i xmm_Dim1_0 = *(const __m128i*)(tmp + (1 +
+ (i - 1) * 2) * NB_ELTS_V8 + 4 * 0);
+ __m128i xmm_Dim1_1 = *(const __m128i*)(tmp + (1 +
+ (i - 1) * 2) * NB_ELTS_V8 + 4 * 1);
+ const __m128i xmm_two = _mm_set1_epi32(2);
+ for (; i < dn; i++) {
+ __m128i xmm_Di_0 = *(const __m128i*)(tmp +
+ (1 + i * 2) * NB_ELTS_V8 + 4 * 0);
+ __m128i xmm_Di_1 = *(const __m128i*)(tmp +
+ (1 + i * 2) * NB_ELTS_V8 + 4 * 1);
+ __m128i xmm_Si_0 = *(const __m128i*)(tmp +
+ (i * 2) * NB_ELTS_V8 + 4 * 0);
+ __m128i xmm_Si_1 = *(const __m128i*)(tmp +
+ (i * 2) * NB_ELTS_V8 + 4 * 1);
+ xmm_Si_0 = _mm_add_epi32(xmm_Si_0,
+ _mm_srai_epi32(_mm_add_epi32(_mm_add_epi32(xmm_Dim1_0, xmm_Di_0), xmm_two), 2));
+ xmm_Si_1 = _mm_add_epi32(xmm_Si_1,
+ _mm_srai_epi32(_mm_add_epi32(_mm_add_epi32(xmm_Dim1_1, xmm_Di_1), xmm_two), 2));
+ *(__m128i*)(tmp + (i * 2) * NB_ELTS_V8 + 4 * 0) = xmm_Si_0;
+ *(__m128i*)(tmp + (i * 2) * NB_ELTS_V8 + 4 * 1) = xmm_Si_1;
+ xmm_Dim1_0 = xmm_Di_0;
+ xmm_Dim1_1 = xmm_Di_1;
+ }
+ }
+ if (((height) % 2) == 1) {
+ for (c = 0; c < NB_ELTS_V8; c++) {
+ OPJ_Sc(i) += (OPJ_Dc(i - 1) + OPJ_Dc(i - 1) + 2) >> 2;
+ }
+ }
+ } else {
+ OPJ_UINT32 c;
+ OPJ_UINT32 i;
+ for (c = 0; c < NB_ELTS_V8; c++) {
+ OPJ_Sc(0) -= OPJ_Dc(0);
+ }
+ i = 1;
+ if (i < sn) {
+ __m128i xmm_Dim1_0 = *(const __m128i*)(tmp + (1 +
+ (i - 1) * 2) * NB_ELTS_V8 + 4 * 0);
+ __m128i xmm_Dim1_1 = *(const __m128i*)(tmp + (1 +
+ (i - 1) * 2) * NB_ELTS_V8 + 4 * 1);
+ for (; i < sn; i++) {
+ __m128i xmm_Di_0 = *(const __m128i*)(tmp +
+ (1 + i * 2) * NB_ELTS_V8 + 4 * 0);
+ __m128i xmm_Di_1 = *(const __m128i*)(tmp +
+ (1 + i * 2) * NB_ELTS_V8 + 4 * 1);
+ __m128i xmm_Si_0 = *(const __m128i*)(tmp +
+ (i * 2) * NB_ELTS_V8 + 4 * 0);
+ __m128i xmm_Si_1 = *(const __m128i*)(tmp +
+ (i * 2) * NB_ELTS_V8 + 4 * 1);
+ xmm_Si_0 = _mm_sub_epi32(xmm_Si_0,
+ _mm_srai_epi32(_mm_add_epi32(xmm_Di_0, xmm_Dim1_0), 1));
+ xmm_Si_1 = _mm_sub_epi32(xmm_Si_1,
+ _mm_srai_epi32(_mm_add_epi32(xmm_Di_1, xmm_Dim1_1), 1));
+ *(__m128i*)(tmp + (i * 2) * NB_ELTS_V8 + 4 * 0) = xmm_Si_0;
+ *(__m128i*)(tmp + (i * 2) * NB_ELTS_V8 + 4 * 1) = xmm_Si_1;
+ xmm_Dim1_0 = xmm_Di_0;
+ xmm_Dim1_1 = xmm_Di_1;
+ }
+ }
+ if (((height) % 2) == 1) {
+ for (c = 0; c < NB_ELTS_V8; c++) {
+ OPJ_Sc(i) -= OPJ_Dc(i - 1);
+ }
+ }
+ i = 0;
+ if (i + 1 < dn) {
+ __m128i xmm_Si_0 = *((const __m128i*)(tmp + 4 * 0));
+ __m128i xmm_Si_1 = *((const __m128i*)(tmp + 4 * 1));
+ const __m128i xmm_two = _mm_set1_epi32(2);
+ for (; i + 1 < dn; i++) {
+ __m128i xmm_Sip1_0 = *(const __m128i*)(tmp +
+ (i + 1) * 2 * NB_ELTS_V8 + 4 * 0);
+ __m128i xmm_Sip1_1 = *(const __m128i*)(tmp +
+ (i + 1) * 2 * NB_ELTS_V8 + 4 * 1);
+ __m128i xmm_Di_0 = *(const __m128i*)(tmp +
+ (1 + i * 2) * NB_ELTS_V8 + 4 * 0);
+ __m128i xmm_Di_1 = *(const __m128i*)(tmp +
+ (1 + i * 2) * NB_ELTS_V8 + 4 * 1);
+ xmm_Di_0 = _mm_add_epi32(xmm_Di_0,
+ _mm_srai_epi32(_mm_add_epi32(_mm_add_epi32(xmm_Si_0, xmm_Sip1_0), xmm_two), 2));
+ xmm_Di_1 = _mm_add_epi32(xmm_Di_1,
+ _mm_srai_epi32(_mm_add_epi32(_mm_add_epi32(xmm_Si_1, xmm_Sip1_1), xmm_two), 2));
+ *(__m128i*)(tmp + (1 + i * 2) * NB_ELTS_V8 + 4 * 0) = xmm_Di_0;
+ *(__m128i*)(tmp + (1 + i * 2) * NB_ELTS_V8 + 4 * 1) = xmm_Di_1;
+ xmm_Si_0 = xmm_Sip1_0;
+ xmm_Si_1 = xmm_Sip1_1;
+ }
+ }
+ if (((height) % 2) == 0) {
+ for (c = 0; c < NB_ELTS_V8; c++) {
+ OPJ_Dc(i) += (OPJ_Sc(i) + OPJ_Sc(i) + 2) >> 2;
+ }
+ }
+ }
+#else
+ if (even) {
+ OPJ_UINT32 c;
+ if (height > 1) {
+ OPJ_UINT32 i;
+ for (i = 0; i + 1 < sn; i++) {
+ for (c = 0; c < NB_ELTS_V8; c++) {
+ OPJ_Dc(i) -= (OPJ_Sc(i) + OPJ_Sc(i + 1)) >> 1;
+ }
+ }
+ if (((height) % 2) == 0) {
+ for (c = 0; c < NB_ELTS_V8; c++) {
+ OPJ_Dc(i) -= OPJ_Sc(i);
+ }
+ }
+ for (c = 0; c < NB_ELTS_V8; c++) {
+ OPJ_Sc(0) += (OPJ_Dc(0) + OPJ_Dc(0) + 2) >> 2;
+ }
+ for (i = 1; i < dn; i++) {
+ for (c = 0; c < NB_ELTS_V8; c++) {
+ OPJ_Sc(i) += (OPJ_Dc(i - 1) + OPJ_Dc(i) + 2) >> 2;
+ }
+ }
+ if (((height) % 2) == 1) {
+ for (c = 0; c < NB_ELTS_V8; c++) {
+ OPJ_Sc(i) += (OPJ_Dc(i - 1) + OPJ_Dc(i - 1) + 2) >> 2;
+ }
+ }
+ }
+ } else {
+ OPJ_UINT32 c;
+ if (height == 1) {
+ for (c = 0; c < NB_ELTS_V8; c++) {
+ OPJ_Sc(0) *= 2;
+ }
+ } else {
+ OPJ_UINT32 i;
+ for (c = 0; c < NB_ELTS_V8; c++) {
+ OPJ_Sc(0) -= OPJ_Dc(0);
+ }
+ for (i = 1; i < sn; i++) {
+ for (c = 0; c < NB_ELTS_V8; c++) {
+ OPJ_Sc(i) -= (OPJ_Dc(i) + OPJ_Dc(i - 1)) >> 1;
+ }
+ }
+ if (((height) % 2) == 1) {
+ for (c = 0; c < NB_ELTS_V8; c++) {
+ OPJ_Sc(i) -= OPJ_Dc(i - 1);
+ }
+ }
+ for (i = 0; i + 1 < dn; i++) {
+ for (c = 0; c < NB_ELTS_V8; c++) {
+ OPJ_Dc(i) += (OPJ_Sc(i) + OPJ_Sc(i + 1) + 2) >> 2;
+ }
+ }
+ if (((height) % 2) == 0) {
+ for (c = 0; c < NB_ELTS_V8; c++) {
+ OPJ_Dc(i) += (OPJ_Sc(i) + OPJ_Sc(i) + 2) >> 2;
+ }
+ }
+ }
+ }
+#endif
+
+ if (cols == NB_ELTS_V8) {
+ opj_dwt_deinterleave_v_cols(tmp, array, (OPJ_INT32)dn, (OPJ_INT32)sn,
+ stride_width, even ? 0 : 1, NB_ELTS_V8);
+ } else {
+ opj_dwt_deinterleave_v_cols(tmp, array, (OPJ_INT32)dn, (OPJ_INT32)sn,
+ stride_width, even ? 0 : 1, cols);
+ }
}
+static void opj_v8dwt_encode_step1(OPJ_FLOAT32* fw,
+ OPJ_UINT32 end,
+ const OPJ_FLOAT32 cst)
+{
+ OPJ_UINT32 i;
+#ifdef __SSE__
+ __m128* vw = (__m128*) fw;
+ const __m128 vcst = _mm_set1_ps(cst);
+ for (i = 0; i < end; ++i) {
+ vw[0] = _mm_mul_ps(vw[0], vcst);
+ vw[1] = _mm_mul_ps(vw[1], vcst);
+ vw += 2 * (NB_ELTS_V8 * sizeof(OPJ_FLOAT32) / sizeof(__m128));
+ }
+#else
+ OPJ_UINT32 c;
+ for (i = 0; i < end; ++i) {
+ for (c = 0; c < NB_ELTS_V8; c++) {
+ fw[i * 2 * NB_ELTS_V8 + c] *= cst;
+ }
+ }
+#endif
+}
+
+static void opj_v8dwt_encode_step2(OPJ_FLOAT32* fl, OPJ_FLOAT32* fw,
+ OPJ_UINT32 end,
+ OPJ_UINT32 m,
+ OPJ_FLOAT32 cst)
+{
+ OPJ_UINT32 i;
+ OPJ_UINT32 imax = opj_uint_min(end, m);
+#ifdef __SSE__
+ __m128* vw = (__m128*) fw;
+ __m128 vcst = _mm_set1_ps(cst);
+ if (imax > 0) {
+ __m128* vl = (__m128*) fl;
+ vw[-2] = _mm_add_ps(vw[-2], _mm_mul_ps(_mm_add_ps(vl[0], vw[0]), vcst));
+ vw[-1] = _mm_add_ps(vw[-1], _mm_mul_ps(_mm_add_ps(vl[1], vw[1]), vcst));
+ vw += 2 * (NB_ELTS_V8 * sizeof(OPJ_FLOAT32) / sizeof(__m128));
+ i = 1;
+
+ for (; i < imax; ++i) {
+ vw[-2] = _mm_add_ps(vw[-2], _mm_mul_ps(_mm_add_ps(vw[-4], vw[0]), vcst));
+ vw[-1] = _mm_add_ps(vw[-1], _mm_mul_ps(_mm_add_ps(vw[-3], vw[1]), vcst));
+ vw += 2 * (NB_ELTS_V8 * sizeof(OPJ_FLOAT32) / sizeof(__m128));
+ }
+ }
+ if (m < end) {
+ assert(m + 1 == end);
+ vcst = _mm_add_ps(vcst, vcst);
+ vw[-2] = _mm_add_ps(vw[-2], _mm_mul_ps(vw[-4], vcst));
+ vw[-1] = _mm_add_ps(vw[-1], _mm_mul_ps(vw[-3], vcst));
+ }
+#else
+ OPJ_INT32 c;
+ if (imax > 0) {
+ for (c = 0; c < NB_ELTS_V8; c++) {
+ fw[-1 * NB_ELTS_V8 + c] += (fl[0 * NB_ELTS_V8 + c] + fw[0 * NB_ELTS_V8 + c]) *
+ cst;
+ }
+ fw += 2 * NB_ELTS_V8;
+ i = 1;
+ for (; i < imax; ++i) {
+ for (c = 0; c < NB_ELTS_V8; c++) {
+ fw[-1 * NB_ELTS_V8 + c] += (fw[-2 * NB_ELTS_V8 + c] + fw[0 * NB_ELTS_V8 + c]) *
+ cst;
+ }
+ fw += 2 * NB_ELTS_V8;
+ }
+ }
+ if (m < end) {
+ assert(m + 1 == end);
+ for (c = 0; c < NB_ELTS_V8; c++) {
+ fw[-1 * NB_ELTS_V8 + c] += (2 * fw[-2 * NB_ELTS_V8 + c]) * cst;
+ }
+ }
+#endif
+}
+
+/* Forward 9-7 transform, for the vertical pass, processing cols columns */
+/* where cols <= NB_ELTS_V8 */
+static void opj_dwt_encode_and_deinterleave_v_real(
+ void *arrayIn,
+ void *tmpIn,
+ OPJ_UINT32 height,
+ OPJ_BOOL even,
+ OPJ_UINT32 stride_width,
+ OPJ_UINT32 cols)
+{
+ OPJ_FLOAT32* OPJ_RESTRICT array = (OPJ_FLOAT32 * OPJ_RESTRICT)arrayIn;
+ OPJ_FLOAT32* OPJ_RESTRICT tmp = (OPJ_FLOAT32 * OPJ_RESTRICT)tmpIn;
+ const OPJ_INT32 sn = (OPJ_INT32)((height + (even ? 1 : 0)) >> 1);
+ const OPJ_INT32 dn = (OPJ_INT32)(height - (OPJ_UINT32)sn);
+ OPJ_INT32 a, b;
+
+ if (height == 1) {
+ return;
+ }
+
+ opj_dwt_fetch_cols_vertical_pass(arrayIn, tmpIn, height, stride_width, cols);
+
+ if (even) {
+ a = 0;
+ b = 1;
+ } else {
+ a = 1;
+ b = 0;
+ }
+ opj_v8dwt_encode_step2(tmp + a * NB_ELTS_V8,
+ tmp + (b + 1) * NB_ELTS_V8,
+ (OPJ_UINT32)dn,
+ (OPJ_UINT32)opj_int_min(dn, sn - b),
+ opj_dwt_alpha);
+ opj_v8dwt_encode_step2(tmp + b * NB_ELTS_V8,
+ tmp + (a + 1) * NB_ELTS_V8,
+ (OPJ_UINT32)sn,
+ (OPJ_UINT32)opj_int_min(sn, dn - a),
+ opj_dwt_beta);
+ opj_v8dwt_encode_step2(tmp + a * NB_ELTS_V8,
+ tmp + (b + 1) * NB_ELTS_V8,
+ (OPJ_UINT32)dn,
+ (OPJ_UINT32)opj_int_min(dn, sn - b),
+ opj_dwt_gamma);
+ opj_v8dwt_encode_step2(tmp + b * NB_ELTS_V8,
+ tmp + (a + 1) * NB_ELTS_V8,
+ (OPJ_UINT32)sn,
+ (OPJ_UINT32)opj_int_min(sn, dn - a),
+ opj_dwt_delta);
+ opj_v8dwt_encode_step1(tmp + b * NB_ELTS_V8, (OPJ_UINT32)dn,
+ opj_K);
+ opj_v8dwt_encode_step1(tmp + a * NB_ELTS_V8, (OPJ_UINT32)sn,
+ opj_invK);
+
+
+ if (cols == NB_ELTS_V8) {
+ opj_dwt_deinterleave_v_cols((OPJ_INT32*)tmp,
+ (OPJ_INT32*)array,
+ (OPJ_INT32)dn, (OPJ_INT32)sn,
+ stride_width, even ? 0 : 1, NB_ELTS_V8);
+ } else {
+ opj_dwt_deinterleave_v_cols((OPJ_INT32*)tmp,
+ (OPJ_INT32*)array,
+ (OPJ_INT32)dn, (OPJ_INT32)sn,
+ stride_width, even ? 0 : 1, cols);
+ }
+}
+
+
/* <summary> */
/* Forward 5-3 wavelet transform in 2-D. */
/* </summary> */
static INLINE OPJ_BOOL opj_dwt_encode_procedure(opj_thread_pool_t* tp,
opj_tcd_tilecomp_t * tilec,
- opj_encode_one_row_fnptr_type p_function,
+ opj_encode_and_deinterleave_v_fnptr_type p_encode_and_deinterleave_v,
opj_encode_and_deinterleave_h_one_row_fnptr_type
p_encode_and_deinterleave_h_one_row)
{
l_data_size = opj_dwt_max_resolution(tilec->resolutions, tilec->numresolutions);
/* overflow check */
- if (l_data_size > (SIZE_MAX / sizeof(OPJ_INT32))) {
+ if (l_data_size > (SIZE_MAX / (NB_ELTS_V8 * sizeof(OPJ_INT32)))) {
/* FIXME event manager error callback */
return OPJ_FALSE;
}
- l_data_size *= sizeof(OPJ_INT32);
+ l_data_size *= NB_ELTS_V8 * sizeof(OPJ_INT32);
bj = (OPJ_INT32*)opj_aligned_32_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 */
dn = (OPJ_INT32)(rh - rh1);
/* Perform vertical pass */
- if (num_threads <= 1 || rw <= 1) {
- for (j = 0; j < rw; ++j) {
- OPJ_INT32* OPJ_RESTRICT aj = tiledp + j;
- OPJ_UINT32 k;
- for (k = 0; k < rh; ++k) {
- bj[k] = aj[k * w];
- }
-
- (*p_function)(bj, dn, sn, cas_col);
-
- opj_dwt_deinterleave_v(bj, aj, dn, sn, w, cas_col);
+ if (num_threads <= 1 || rw < 2 * NB_ELTS_V8) {
+ for (j = 0; j + NB_ELTS_V8 - 1 < rw; j += NB_ELTS_V8) {
+ p_encode_and_deinterleave_v(tiledp + j,
+ bj,
+ rh,
+ cas_col == 0,
+ w,
+ NB_ELTS_V8);
+ }
+ if (j < rw) {
+ p_encode_and_deinterleave_v(tiledp + j,
+ bj,
+ rh,
+ cas_col == 0,
+ w,
+ rw - j);
}
} else {
OPJ_UINT32 num_jobs = (OPJ_UINT32)num_threads;
if (rw < num_jobs) {
num_jobs = rw;
}
- step_j = (rw / num_jobs);
+ step_j = ((rw / num_jobs) / NB_ELTS_V8) * NB_ELTS_V8;
for (j = 0; j < num_jobs; j++) {
opj_dwt_encode_v_job_t* job;
job->w = w;
job->tiledp = tiledp;
job->min_j = j * step_j;
- job->max_j = (j + 1U) * step_j; /* this can overflow */
- if (j == (num_jobs - 1U)) { /* this will take care of the overflow */
- job->max_j = rw;
- }
- job->p_function = p_function;
+ job->max_j = (j + 1 == num_jobs) ? rw : (j + 1) * step_j;
+ job->p_encode_and_deinterleave_v = p_encode_and_deinterleave_v;
opj_thread_pool_submit_job(tp, opj_dwt_encode_v_func, job);
}
opj_thread_pool_wait_completion(tp, 0);
opj_tcd_tilecomp_t * tilec)
{
return opj_dwt_encode_procedure(p_tcd->thread_pool, tilec,
- opj_dwt_encode_1,
+ opj_dwt_encode_and_deinterleave_v,
opj_dwt_encode_and_deinterleave_h_one_row);
}
opj_tcd_tilecomp_t * tilec)
{
return opj_dwt_encode_procedure(p_tcd->thread_pool, tilec,
- opj_dwt_encode_1_real,
+ opj_dwt_encode_and_deinterleave_v_real,
opj_dwt_encode_and_deinterleave_h_one_row_real);
}
if (tccp->qntsty == J2K_CCP_QNTSTY_NOQNT) {
stepsize = 1.0;
} else {
- OPJ_FLOAT64 norm = opj_dwt_norms_real[orient][level];
+ OPJ_FLOAT64 norm = opj_dwt_getnorm_real(level, orient);
stepsize = (1 << (gain)) / norm;
}
opj_dwt_encode_stepsize((OPJ_INT32) floor(stepsize * 8192.0),
OPJ_S(0) /= 2;
} else {
for (i = win_l_x0; i < win_l_x1; i++) {
- OPJ_D(i) -= (OPJ_SS_(i) + OPJ_SS_(i + 1) + 2) >> 2;
+ OPJ_D(i) = opj_int_sub_no_overflow(OPJ_D(i),
+ opj_int_add_no_overflow(opj_int_add_no_overflow(OPJ_SS_(i), OPJ_SS_(i + 1)),
+ 2) >> 2);
}
for (i = win_h_x0; i < win_h_x1; i++) {
- OPJ_S(i) += (OPJ_DD_(i) + OPJ_DD_(i - 1)) >> 1;
+ OPJ_S(i) = opj_int_add_no_overflow(OPJ_S(i),
+ opj_int_add_no_overflow(OPJ_DD_(i), OPJ_DD_(i - 1)) >> 1);
}
}
}
} 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;
+ OPJ_D_off(i, off) = opj_int_sub_no_overflow(
+ OPJ_D_off(i, off),
+ opj_int_add_no_overflow(
+ opj_int_add_no_overflow(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;
+ OPJ_S_off(i, off) = opj_int_add_no_overflow(
+ OPJ_S_off(i, off),
+ opj_int_add_no_overflow(OPJ_DD__off(i, off), OPJ_DD__off(i - 1, off)) >> 1);
}
}
}