neon 4 byte helper functions

When data is guaranteed to be aligned, use helper functions which
assert that requirement.

Change-Id: Ic4b188593aea0799d5bd8eda64f9858a1592a2a3

2 files changed, 33 insertions, 11 deletions

diff --git a/vpx_dsp/arm/idct4x4_add_neon.c b/vpx_dsp/arm/idct4x4_add_neon.c
index e44ba6e75..673a36840 100644
--- a/vpx_dsp/arm/idct4x4_add_neon.c
+++ b/vpx_dsp/arm/idct4x4_add_neon.c
@@ -20,7 +20,7 @@ void vpx_idct4x4_16_add_neon(const tran_low_t *input, uint8_t *dest,
                              int stride) {
   const uint8_t *dst = dest;
   const int16x4_t cospis = vld1_s16(kCospi);
-  uint32x2_t dest01_u32 = vdup_n_u32(0);
+  uint8x8_t dest01_u8;
   uint32x2_t dest32_u32 = vdup_n_u32(0);
   int16x8_t a0, a1;
   uint8x8_t d01, d32;
@@ -40,25 +40,22 @@ void vpx_idct4x4_16_add_neon(const tran_low_t *input, uint8_t *dest,
   a0 = vrshrq_n_s16(a0, 4);
   a1 = vrshrq_n_s16(a1, 4);
 
-  dest01_u32 = vld1_lane_u32((const uint32_t *)dst, dest01_u32, 0);
-  dst += stride;
-  dest01_u32 = vld1_lane_u32((const uint32_t *)dst, dest01_u32, 1);
-  dst += stride;
+  dest01_u8 = load_u8(dst, stride);
+  dst += 2 * stride;
+  // The elements are loaded in reverse order.
   dest32_u32 = vld1_lane_u32((const uint32_t *)dst, dest32_u32, 1);
   dst += stride;
   dest32_u32 = vld1_lane_u32((const uint32_t *)dst, dest32_u32, 0);
 
-  d01_u16 =
-      vaddw_u8(vreinterpretq_u16_s16(a0), vreinterpret_u8_u32(dest01_u32));
+  d01_u16 = vaddw_u8(vreinterpretq_u16_s16(a0), dest01_u8);
   d32_u16 =
       vaddw_u8(vreinterpretq_u16_s16(a1), vreinterpret_u8_u32(dest32_u32));
   d01 = vqmovun_s16(vreinterpretq_s16_u16(d01_u16));
   d32 = vqmovun_s16(vreinterpretq_s16_u16(d32_u16));
 
-  vst1_lane_u32((uint32_t *)dest, vreinterpret_u32_u8(d01), 0);
-  dest += stride;
-  vst1_lane_u32((uint32_t *)dest, vreinterpret_u32_u8(d01), 1);
-  dest += stride;
+  store_u8(dest, stride, d01);
+  dest += 2 * stride;
+  // The elements are stored in reverse order.
   vst1_lane_u32((uint32_t *)dest, vreinterpret_u32_u8(d32), 1);
   dest += stride;
   vst1_lane_u32((uint32_t *)dest, vreinterpret_u32_u8(d32), 0);
diff --git a/vpx_dsp/arm/mem_neon.h b/vpx_dsp/arm/mem_neon.h
index ef6e9decd..ba5c3d513 100644
--- a/vpx_dsp/arm/mem_neon.h
+++ b/vpx_dsp/arm/mem_neon.h
@@ -68,4 +68,29 @@ static INLINE void store_s16q_to_tran_low(tran_low_t *buf, const int16x8_t a) {
   vst1q_s16(buf, a);
 #endif
 }
+
+// Load 2 sets of 4 bytes when alignment is guaranteed.
+static INLINE uint8x8_t load_u8(const uint8_t *buf, int stride) {
+  uint32x2_t a = vdup_n_u32(0);
+
+  assert(!((intptr_t)buf % sizeof(uint32_t)));
+  assert(!(stride % sizeof(uint32_t)));
+
+  a = vld1_lane_u32((const uint32_t *)buf, a, 0);
+  buf += stride;
+  a = vld1_lane_u32((const uint32_t *)buf, a, 1);
+  return vreinterpret_u8_u32(a);
+}
+
+// Store 2 sets of 4 bytes when alignment is guaranteed.
+static INLINE void store_u8(uint8_t *buf, int stride, const uint8x8_t a) {
+  uint32x2_t a_u32 = vreinterpret_u32_u8(a);
+
+  assert(!((intptr_t)buf % sizeof(uint32_t)));
+  assert(!(stride % sizeof(uint32_t)));
+
+  vst1_lane_u32((uint32_t *)buf, a_u32, 0);
+  buf += stride;
+  vst1_lane_u32((uint32_t *)buf, a_u32, 1);
+}
 #endif  // VPX_DSP_ARM_MEM_NEON_H_