[FFmpeg-cvslog] lavu/tx: add parity revtab generator version

[Lynne]

ffmpeg | branch: master | Lynne <dev at lynne.ee> | Sat Apr 10 03:52:31 2021 +0200| [ff71671d88ef23e1a539e1a99eadd766c27f3ed3] | committer: Lynne

lavu/tx: add parity revtab generator version

This will be used for SIMD support.

> http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=ff71671d88ef23e1a539e1a99eadd766c27f3ed3
---

 libavutil/tx.c      | 49 +++++++++++++++++++++++++++++++++++++++++++++++++
 libavutil/tx_priv.h | 31 +++++++++++++++++++++++++++++++
 2 files changed, 80 insertions(+)

diff --git a/libavutil/tx.c b/libavutil/tx.c
index 05d4de30cc..6d0e854084 100644
--- a/libavutil/tx.c
+++ b/libavutil/tx.c
@@ -158,6 +158,55 @@ int ff_tx_gen_ptwo_inplace_revtab_idx(AVTXContext *s)
     return 0;
 }
 
+static void parity_revtab_generator(int *revtab, int n, int inv, int offset,
+                                    int is_dual, int dual_high, int len,
+                                    int basis, int dual_stride)
+{
+    len >>= 1;
+
+    if (len <= basis) {
+        int k1, k2, *even, *odd, stride;
+
+        is_dual = is_dual && dual_stride;
+        dual_high = is_dual & dual_high;
+        stride = is_dual ? FFMIN(dual_stride, len) : 0;
+
+        even = &revtab[offset + dual_high*(stride - 2*len)];
+        odd  = &even[len + (is_dual && !dual_high)*len + dual_high*len];
+
+        for (int i = 0; i < len; i++) {
+            k1 = -split_radix_permutation(offset + i*2 + 0, n, inv) & (n - 1);
+            k2 = -split_radix_permutation(offset + i*2 + 1, n, inv) & (n - 1);
+            *even++ = k1;
+            *odd++  = k2;
+            if (stride && !((i + 1) % stride)) {
+                even += stride;
+                odd  += stride;
+            }
+        }
+
+        return;
+    }
+
+    parity_revtab_generator(revtab, n, inv, offset,
+                            0, 0, len >> 0, basis, dual_stride);
+    parity_revtab_generator(revtab, n, inv, offset + (len >> 0),
+                            1, 0, len >> 1, basis, dual_stride);
+    parity_revtab_generator(revtab, n, inv, offset + (len >> 0) + (len >> 1),
+                            1, 1, len >> 1, basis, dual_stride);
+}
+
+void ff_tx_gen_split_radix_parity_revtab(int *revtab, int len, int inv,
+                                         int basis, int dual_stride)
+{
+    basis >>= 1;
+    if (len < basis)
+        return;
+    av_assert0(!dual_stride || !(dual_stride & (dual_stride - 1)));
+    av_assert0(dual_stride <= basis);
+    parity_revtab_generator(revtab, len, inv, 0, 0, 0, len, basis, dual_stride);
+}
+
 av_cold void av_tx_uninit(AVTXContext **ctx)
 {
     if (!(*ctx))
diff --git a/libavutil/tx_priv.h b/libavutil/tx_priv.h
index 1d4245e71b..b889f6d3b4 100644
--- a/libavutil/tx_priv.h
+++ b/libavutil/tx_priv.h
@@ -149,6 +149,37 @@ int ff_tx_gen_ptwo_revtab(AVTXContext *s, int invert_lookup);
  */
 int ff_tx_gen_ptwo_inplace_revtab_idx(AVTXContext *s);
 
+/*
+ * This generates a parity-based revtab of length len and direction inv.
+ *
+ * Parity means even and odd complex numbers will be split, e.g. the even
+ * coefficients will come first, after which the odd coefficients will be
+ * placed. For example, a 4-point transform's coefficients after reordering:
+ * z[0].re, z[0].im, z[2].re, z[2].im, z[1].re, z[1].im, z[3].re, z[3].im
+ *
+ * The basis argument is the length of the largest non-composite transform
+ * supported, and also implies that the basis/2 transform is supported as well,
+ * as the split-radix algorithm requires it to be.
+ *
+ * The dual_stride argument indicates that both the basis, as well as the
+ * basis/2 transforms support doing two transforms at once, and the coefficients
+ * will be interleaved between each pair in a split-radix like so (stride == 2):
+ * tx1[0], tx1[2], tx2[0], tx2[2], tx1[1], tx1[3], tx2[1], tx2[3]
+ * A non-zero number switches this on, with the value indicating the stride
+ * (how many values of 1 transform to put first before switching to the other).
+ * Must be a power of two or 0. Must be less than the basis.
+ * Value will be clipped to the transform size, so for a basis of 16 and a
+ * dual_stride of 8, dual 8-point transforms will be laid out as if dual_stride
+ * was set to 4.
+ * Usually you'll set this to half the complex numbers that fit in a single
+ * register or 0. This allows to reuse SSE functions as dual-transform
+ * functions in AVX mode.
+ *
+ * If length is smaller than basis/2 this function will not do anything.
+ */
+void ff_tx_gen_split_radix_parity_revtab(int *revtab, int len, int inv,
+                                         int basis, int dual_stride);
+
 /* Templated init functions */
 int ff_tx_init_mdct_fft_float(AVTXContext *s, av_tx_fn *tx,
                               enum AVTXType type, int inv, int len,