Add 32x32 Hadamard transform

Add 32x32 Hadamard transform in C implementation. Replace the
forward 32x32 2D-DCT in tpl model with Hadamard transform. This
would reduce the overhead encoding time due to running tpl model
by ~3x.

Change-Id: I1c743dab786b818d89f14928cc3998d056830aa9

diff --git a/vp9/encoder/vp9_encoder.c b/vp9/encoder/vp9_encoder.c
index b17ea9f..71bdb6e 100644 (file)
--- a/vp9/encoder/vp9_encoder.c
+++ b/vp9/encoder/vp9_encoder.c
@@ -5822,7 +5822,7 @@ void mc_flow_dispenser(VP9_COMP *cpi, GF_PICTURE *gf_picture, int frame_idx) {
         vpx_subtract_block(bh, bw, src_diff, bw, src, src_stride, dst,
                            dst_stride);
 
-        vpx_fdct32x32(src_diff, coeff, bw);
+        vpx_hadamard_32x32(src_diff, bw, coeff);
 
         intra_cost = vpx_satd(coeff, pix_num);
 
@@ -5879,7 +5879,7 @@ void mc_flow_dispenser(VP9_COMP *cpi, GF_PICTURE *gf_picture, int frame_idx) {
                            this_frame->y_buffer + mb_y_offset,
                            this_frame->y_stride, &predictor[0], bw);
 #endif
-        vpx_fdct32x32(src_diff, coeff, bw);
+        vpx_hadamard_32x32(src_diff, bw, coeff);
 
         inter_cost = vpx_satd(coeff, pix_num);
 

diff --git a/vpx_dsp/avg.c b/vpx_dsp/avg.c
index a7ac6d9..0936e91 100644 (file)
--- a/vpx_dsp/avg.c
+++ b/vpx_dsp/avg.c
@@ -123,6 +123,37 @@ void vpx_hadamard_16x16_c(const int16_t *src_diff, ptrdiff_t src_stride,
   }
 }
 
+void vpx_hadamard_32x32_c(const int16_t *src_diff, ptrdiff_t src_stride,
+                          tran_low_t *coeff) {
+  int idx;
+  for (idx = 0; idx < 4; ++idx) {
+    // src_diff: 9 bit, dynamic range [-255, 255]
+    const int16_t *src_ptr =
+        src_diff + (idx >> 1) * 16 * src_stride + (idx & 0x01) * 16;
+    vpx_hadamard_16x16(src_ptr, src_stride, coeff + idx * 256);
+  }
+
+  // coeff: 15 bit, dynamic range [-16320, 16320]
+  for (idx = 0; idx < 256; ++idx) {
+    tran_low_t a0 = coeff[0];
+    tran_low_t a1 = coeff[256];
+    tran_low_t a2 = coeff[512];
+    tran_low_t a3 = coeff[768];
+
+    tran_low_t b0 = (a0 + a1) >> 2;  // (a0 + a1): 16 bit, [-32640, 32640]
+    tran_low_t b1 = (a0 - a1) >> 2;  // b0-b3: 15 bit, dynamic range
+    tran_low_t b2 = (a2 + a3) >> 2;  // [-16320, 16320]
+    tran_low_t b3 = (a2 - a3) >> 2;
+
+    coeff[0] = b0 + b2;  // 16 bit, [-32640, 32640]
+    coeff[256] = b1 + b3;
+    coeff[512] = b0 - b2;
+    coeff[768] = b1 - b3;
+
+    ++coeff;
+  }
+}
+
 // coeff: 16 bits, dynamic range [-32640, 32640].
 // length: value range {16, 64, 256, 1024}.
 int vpx_satd_c(const tran_low_t *coeff, int length) {

diff --git a/vpx_dsp/vpx_dsp_rtcd_defs.pl b/vpx_dsp/vpx_dsp_rtcd_defs.pl
index 45acaf3..b662c70 100644 (file)
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -782,6 +782,9 @@ if (vpx_config("CONFIG_VP9_ENCODER") eq "yes") {
     add_proto qw/void vpx_hadamard_16x16/, "const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff";
     specialize qw/vpx_hadamard_16x16 avx2 sse2 neon vsx/;
 
+    add_proto qw/void vpx_hadamard_32x32/, "const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff";
+    specialize qw/vpx_hadamard_32x32/;
+
     add_proto qw/int vpx_satd/, "const tran_low_t *coeff, int length";
     specialize qw/vpx_satd avx2 sse2 neon/;
   } else {
@@ -791,6 +794,9 @@ if (vpx_config("CONFIG_VP9_ENCODER") eq "yes") {
     add_proto qw/void vpx_hadamard_16x16/, "const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff";
     specialize qw/vpx_hadamard_16x16 avx2 sse2 neon msa vsx/;
 
+    add_proto qw/void vpx_hadamard_32x32/, "const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff";
+    specialize qw/vpx_hadamard_32x32/;
+
     add_proto qw/int vpx_satd/, "const int16_t *coeff, int length";
     specialize qw/vpx_satd avx2 sse2 neon msa/;
   }