partial fdct neon: add 32x32_1

Always return an int32_t. Since it needs to be moved to a register for
shifting, this doesn't really penalize the smaller transforms.

The values could potentially be summed and shifted in place.

BUG=webm:1424

Change-Id: Id5beb35d79c7574ebd99285fc4182788cf2bb972

2 files changed, 33 insertions, 10 deletions

diff --git a/vpx_dsp/arm/fdct_partial_neon.c b/vpx_dsp/arm/fdct_partial_neon.c
index 4e1a6dfda..3c4b292d7 100644
--- a/vpx_dsp/arm/fdct_partial_neon.c
+++ b/vpx_dsp/arm/fdct_partial_neon.c
@@ -14,16 +14,12 @@
 #include "./vpx_config.h"
 #include "vpx_dsp/arm/mem_neon.h"
 
-static INLINE tran_low_t sum_int16x8(const int16x8_t a) {
+static INLINE int32_t sum_int16x8(const int16x8_t a) {
   const int32x4_t b = vpaddlq_s16(a);
   const int64x2_t c = vpaddlq_s32(b);
   const int32x2_t d = vadd_s32(vreinterpret_s32_s64(vget_low_s64(c)),
                                vreinterpret_s32_s64(vget_high_s64(c)));
-#if CONFIG_VP9_HIGHBITDEPTH
   return vget_lane_s32(d, 0);
-#else
-  return vget_lane_s16(vreinterpret_s16_s32(d), 0);
-#endif
 }
 
 void vpx_fdct4x4_1_neon(const int16_t *input, tran_low_t *output, int stride) {
@@ -44,7 +40,7 @@ void vpx_fdct4x4_1_neon(const int16_t *input, tran_low_t *output, int stride) {
 
   c = vaddq_s16(b0, b1);
 
-  output[0] = sum_int16x8(c) << 1;
+  output[0] = (tran_low_t)(sum_int16x8(c) << 1);
   output[1] = 0;
 }
 
@@ -56,7 +52,7 @@ void vpx_fdct8x8_1_neon(const int16_t *input, tran_low_t *output, int stride) {
     sum = vaddq_s16(sum, input_00);
   }
 
-  output[0] = sum_int16x8(sum);
+  output[0] = (tran_low_t)sum_int16x8(sum);
   output[1] = 0;
 }
 
@@ -74,6 +70,33 @@ void vpx_fdct16x16_1_neon(const int16_t *input, tran_low_t *output,
     right = vaddq_s16(right, b);
   }
 
-  output[0] = (sum_int16x8(left) + sum_int16x8(right)) >> 1;
+  output[0] = (tran_low_t)((sum_int16x8(left) + sum_int16x8(right)) >> 1);
+  output[1] = 0;
+}
+
+void vpx_fdct32x32_1_neon(const int16_t *input, tran_low_t *output,
+                          int stride) {
+  int r;
+  int16x8_t a0 = vld1q_s16(input);
+  int16x8_t a1 = vld1q_s16(input + 8);
+  int16x8_t a2 = vld1q_s16(input + 16);
+  int16x8_t a3 = vld1q_s16(input + 24);
+  input += stride;
+  for (r = 1; r < 32; ++r) {
+    const int16x8_t b0 = vld1q_s16(input);
+    const int16x8_t b1 = vld1q_s16(input + 8);
+    const int16x8_t b2 = vld1q_s16(input + 16);
+    const int16x8_t b3 = vld1q_s16(input + 24);
+    input += stride;
+    a0 = vaddq_s16(a0, b0);
+    a1 = vaddq_s16(a1, b1);
+    a2 = vaddq_s16(a2, b2);
+    a3 = vaddq_s16(a3, b3);
+  }
+
+  // TODO(johannkoenig): sum and shift the values in neon registers.
+  output[0] = (tran_low_t)(
+      (sum_int16x8(a0) + sum_int16x8(a1) + sum_int16x8(a2) + sum_int16x8(a3)) >>
+      3);
   output[1] = 0;
 }
diff --git a/vpx_dsp/vpx_dsp_rtcd_defs.pl b/vpx_dsp/vpx_dsp_rtcd_defs.pl
index 734f6e1e2..bb83021d7 100644
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -508,7 +508,7 @@ if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
   specialize qw/vpx_fdct32x32_rd sse2/;
 
   add_proto qw/void vpx_fdct32x32_1/, "const int16_t *input, tran_low_t *output, int stride";
-  specialize qw/vpx_fdct32x32_1 sse2/;
+  specialize qw/vpx_fdct32x32_1 sse2 neon/;
 
   add_proto qw/void vpx_highbd_fdct4x4/, "const int16_t *input, tran_low_t *output, int stride";
   specialize qw/vpx_highbd_fdct4x4 sse2/;
@@ -558,7 +558,7 @@ if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
   specialize qw/vpx_fdct32x32_rd sse2 avx2 msa/;
 
   add_proto qw/void vpx_fdct32x32_1/, "const int16_t *input, tran_low_t *output, int stride";
-  specialize qw/vpx_fdct32x32_1 sse2 msa/;
+  specialize qw/vpx_fdct32x32_1 sse2 neon msa/;
 }  # CONFIG_VP9_HIGHBITDEPTH
 }  # CONFIG_VP9_ENCODER