X-Git-Url: https://granicus.if.org/sourcecode?a=blobdiff_plain;f=vpx_dsp%2Farm%2Fidct8x8_add_neon.c;h=1121ade27961486ceb57a7d7c46664e57e5ec106;hb=ff42e04f9cb60e63ca3fe12ac497f27c68555e1f;hp=159a6ec9891de7fd9277c06fef91aa94df951ed4;hpb=cb5a2ac920a8a85daf87f8f9b96067fbf74a69ff;p=libvpx

diff --git a/vpx_dsp/arm/idct8x8_add_neon.c b/vpx_dsp/arm/idct8x8_add_neon.c
index 159a6ec98..1121ade27 100644
--- a/vpx_dsp/arm/idct8x8_add_neon.c
+++ b/vpx_dsp/arm/idct8x8_add_neon.c
@@ -13,434 +13,115 @@
 #include "./vpx_config.h"
 #include "./vpx_dsp_rtcd.h"
 #include "vpx_dsp/arm/idct_neon.h"
+#include "vpx_dsp/arm/mem_neon.h"
 #include "vpx_dsp/arm/transpose_neon.h"
 #include "vpx_dsp/txfm_common.h"
 
-static INLINE void IDCT8x8_1D(int16x8_t *q8s16, int16x8_t *q9s16,
-                              int16x8_t *q10s16, int16x8_t *q11s16,
-                              int16x8_t *q12s16, int16x8_t *q13s16,
-                              int16x8_t *q14s16, int16x8_t *q15s16) {
-  int16x4_t d0s16, d1s16, d2s16, d3s16;
-  int16x4_t d8s16, d9s16, d10s16, d11s16, d12s16, d13s16, d14s16, d15s16;
-  int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16, d22s16, d23s16;
-  int16x4_t d24s16, d25s16, d26s16, d27s16, d28s16, d29s16, d30s16, d31s16;
-  int16x8_t q0s16, q1s16, q2s16, q3s16, q4s16, q5s16, q6s16, q7s16;
-  int32x4_t q2s32, q3s32, q5s32, q6s32, q8s32, q9s32;
-  int32x4_t q10s32, q11s32, q12s32, q13s32, q15s32;
-
-  d0s16 = vdup_n_s16((int16_t)cospi_28_64);
-  d1s16 = vdup_n_s16((int16_t)cospi_4_64);
-  d2s16 = vdup_n_s16((int16_t)cospi_12_64);
-  d3s16 = vdup_n_s16((int16_t)cospi_20_64);
-
-  d16s16 = vget_low_s16(*q8s16);
-  d17s16 = vget_high_s16(*q8s16);
-  d18s16 = vget_low_s16(*q9s16);
-  d19s16 = vget_high_s16(*q9s16);
-  d20s16 = vget_low_s16(*q10s16);
-  d21s16 = vget_high_s16(*q10s16);
-  d22s16 = vget_low_s16(*q11s16);
-  d23s16 = vget_high_s16(*q11s16);
-  d24s16 = vget_low_s16(*q12s16);
-  d25s16 = vget_high_s16(*q12s16);
-  d26s16 = vget_low_s16(*q13s16);
-  d27s16 = vget_high_s16(*q13s16);
-  d28s16 = vget_low_s16(*q14s16);
-  d29s16 = vget_high_s16(*q14s16);
-  d30s16 = vget_low_s16(*q15s16);
-  d31s16 = vget_high_s16(*q15s16);
-
-  q2s32 = vmull_s16(d18s16, d0s16);
-  q3s32 = vmull_s16(d19s16, d0s16);
-  q5s32 = vmull_s16(d26s16, d2s16);
-  q6s32 = vmull_s16(d27s16, d2s16);
-
-  q2s32 = vmlsl_s16(q2s32, d30s16, d1s16);
-  q3s32 = vmlsl_s16(q3s32, d31s16, d1s16);
-  q5s32 = vmlsl_s16(q5s32, d22s16, d3s16);
-  q6s32 = vmlsl_s16(q6s32, d23s16, d3s16);
-
-  d8s16 = vqrshrn_n_s32(q2s32, 14);
-  d9s16 = vqrshrn_n_s32(q3s32, 14);
-  d10s16 = vqrshrn_n_s32(q5s32, 14);
-  d11s16 = vqrshrn_n_s32(q6s32, 14);
-  q4s16 = vcombine_s16(d8s16, d9s16);
-  q5s16 = vcombine_s16(d10s16, d11s16);
-
-  q2s32 = vmull_s16(d18s16, d1s16);
-  q3s32 = vmull_s16(d19s16, d1s16);
-  q9s32 = vmull_s16(d26s16, d3s16);
-  q13s32 = vmull_s16(d27s16, d3s16);
-
-  q2s32 = vmlal_s16(q2s32, d30s16, d0s16);
-  q3s32 = vmlal_s16(q3s32, d31s16, d0s16);
-  q9s32 = vmlal_s16(q9s32, d22s16, d2s16);
-  q13s32 = vmlal_s16(q13s32, d23s16, d2s16);
-
-  d14s16 = vqrshrn_n_s32(q2s32, 14);
-  d15s16 = vqrshrn_n_s32(q3s32, 14);
-  d12s16 = vqrshrn_n_s32(q9s32, 14);
-  d13s16 = vqrshrn_n_s32(q13s32, 14);
-  q6s16 = vcombine_s16(d12s16, d13s16);
-  q7s16 = vcombine_s16(d14s16, d15s16);
-
-  d0s16 = vdup_n_s16((int16_t)cospi_16_64);
-
-  q2s32 = vmull_s16(d16s16, d0s16);
-  q3s32 = vmull_s16(d17s16, d0s16);
-  q13s32 = vmull_s16(d16s16, d0s16);
-  q15s32 = vmull_s16(d17s16, d0s16);
-
-  q2s32 = vmlal_s16(q2s32, d24s16, d0s16);
-  q3s32 = vmlal_s16(q3s32, d25s16, d0s16);
-  q13s32 = vmlsl_s16(q13s32, d24s16, d0s16);
-  q15s32 = vmlsl_s16(q15s32, d25s16, d0s16);
-
-  d0s16 = vdup_n_s16((int16_t)cospi_24_64);
-  d1s16 = vdup_n_s16((int16_t)cospi_8_64);
-
-  d18s16 = vqrshrn_n_s32(q2s32, 14);
-  d19s16 = vqrshrn_n_s32(q3s32, 14);
-  d22s16 = vqrshrn_n_s32(q13s32, 14);
-  d23s16 = vqrshrn_n_s32(q15s32, 14);
-  *q9s16 = vcombine_s16(d18s16, d19s16);
-  *q11s16 = vcombine_s16(d22s16, d23s16);
-
-  q2s32 = vmull_s16(d20s16, d0s16);
-  q3s32 = vmull_s16(d21s16, d0s16);
-  q8s32 = vmull_s16(d20s16, d1s16);
-  q12s32 = vmull_s16(d21s16, d1s16);
-
-  q2s32 = vmlsl_s16(q2s32, d28s16, d1s16);
-  q3s32 = vmlsl_s16(q3s32, d29s16, d1s16);
-  q8s32 = vmlal_s16(q8s32, d28s16, d0s16);
-  q12s32 = vmlal_s16(q12s32, d29s16, d0s16);
-
-  d26s16 = vqrshrn_n_s32(q2s32, 14);
-  d27s16 = vqrshrn_n_s32(q3s32, 14);
-  d30s16 = vqrshrn_n_s32(q8s32, 14);
-  d31s16 = vqrshrn_n_s32(q12s32, 14);
-  *q13s16 = vcombine_s16(d26s16, d27s16);
-  *q15s16 = vcombine_s16(d30s16, d31s16);
-
-  q0s16 = vaddq_s16(*q9s16, *q15s16);
-  q1s16 = vaddq_s16(*q11s16, *q13s16);
-  q2s16 = vsubq_s16(*q11s16, *q13s16);
-  q3s16 = vsubq_s16(*q9s16, *q15s16);
-
-  *q13s16 = vsubq_s16(q4s16, q5s16);
-  q4s16 = vaddq_s16(q4s16, q5s16);
-  *q14s16 = vsubq_s16(q7s16, q6s16);
-  q7s16 = vaddq_s16(q7s16, q6s16);
-  d26s16 = vget_low_s16(*q13s16);
-  d27s16 = vget_high_s16(*q13s16);
-  d28s16 = vget_low_s16(*q14s16);
-  d29s16 = vget_high_s16(*q14s16);
-
-  d16s16 = vdup_n_s16((int16_t)cospi_16_64);
-
-  q9s32 = vmull_s16(d28s16, d16s16);
-  q10s32 = vmull_s16(d29s16, d16s16);
-  q11s32 = vmull_s16(d28s16, d16s16);
-  q12s32 = vmull_s16(d29s16, d16s16);
-
-  q9s32 = vmlsl_s16(q9s32, d26s16, d16s16);
-  q10s32 = vmlsl_s16(q10s32, d27s16, d16s16);
-  q11s32 = vmlal_s16(q11s32, d26s16, d16s16);
-  q12s32 = vmlal_s16(q12s32, d27s16, d16s16);
-
-  d10s16 = vqrshrn_n_s32(q9s32, 14);
-  d11s16 = vqrshrn_n_s32(q10s32, 14);
-  d12s16 = vqrshrn_n_s32(q11s32, 14);
-  d13s16 = vqrshrn_n_s32(q12s32, 14);
-  q5s16 = vcombine_s16(d10s16, d11s16);
-  q6s16 = vcombine_s16(d12s16, d13s16);
-
-  *q8s16 = vaddq_s16(q0s16, q7s16);
-  *q9s16 = vaddq_s16(q1s16, q6s16);
-  *q10s16 = vaddq_s16(q2s16, q5s16);
-  *q11s16 = vaddq_s16(q3s16, q4s16);
-  *q12s16 = vsubq_s16(q3s16, q4s16);
-  *q13s16 = vsubq_s16(q2s16, q5s16);
-  *q14s16 = vsubq_s16(q1s16, q6s16);
-  *q15s16 = vsubq_s16(q0s16, q7s16);
+static INLINE void add8x8(int16x8_t a0, int16x8_t a1, int16x8_t a2,
+                          int16x8_t a3, int16x8_t a4, int16x8_t a5,
+                          int16x8_t a6, int16x8_t a7, uint8_t *dest,
+                          const int stride) {
+  const uint8_t *dst = dest;
+  uint8x8_t d0, d1, d2, d3, d4, d5, d6, d7;
+  uint16x8_t d0_u16, d1_u16, d2_u16, d3_u16, d4_u16, d5_u16, d6_u16, d7_u16;
+
+  a0 = vrshrq_n_s16(a0, 5);
+  a1 = vrshrq_n_s16(a1, 5);
+  a2 = vrshrq_n_s16(a2, 5);
+  a3 = vrshrq_n_s16(a3, 5);
+  a4 = vrshrq_n_s16(a4, 5);
+  a5 = vrshrq_n_s16(a5, 5);
+  a6 = vrshrq_n_s16(a6, 5);
+  a7 = vrshrq_n_s16(a7, 5);
+
+  d0 = vld1_u8(dst);
+  dst += stride;
+  d1 = vld1_u8(dst);
+  dst += stride;
+  d2 = vld1_u8(dst);
+  dst += stride;
+  d3 = vld1_u8(dst);
+  dst += stride;
+  d4 = vld1_u8(dst);
+  dst += stride;
+  d5 = vld1_u8(dst);
+  dst += stride;
+  d6 = vld1_u8(dst);
+  dst += stride;
+  d7 = vld1_u8(dst);
+
+  d0_u16 = vaddw_u8(vreinterpretq_u16_s16(a0), d0);
+  d1_u16 = vaddw_u8(vreinterpretq_u16_s16(a1), d1);
+  d2_u16 = vaddw_u8(vreinterpretq_u16_s16(a2), d2);
+  d3_u16 = vaddw_u8(vreinterpretq_u16_s16(a3), d3);
+  d4_u16 = vaddw_u8(vreinterpretq_u16_s16(a4), d4);
+  d5_u16 = vaddw_u8(vreinterpretq_u16_s16(a5), d5);
+  d6_u16 = vaddw_u8(vreinterpretq_u16_s16(a6), d6);
+  d7_u16 = vaddw_u8(vreinterpretq_u16_s16(a7), d7);
+
+  d0 = vqmovun_s16(vreinterpretq_s16_u16(d0_u16));
+  d1 = vqmovun_s16(vreinterpretq_s16_u16(d1_u16));
+  d2 = vqmovun_s16(vreinterpretq_s16_u16(d2_u16));
+  d3 = vqmovun_s16(vreinterpretq_s16_u16(d3_u16));
+  d4 = vqmovun_s16(vreinterpretq_s16_u16(d4_u16));
+  d5 = vqmovun_s16(vreinterpretq_s16_u16(d5_u16));
+  d6 = vqmovun_s16(vreinterpretq_s16_u16(d6_u16));
+  d7 = vqmovun_s16(vreinterpretq_s16_u16(d7_u16));
+
+  vst1_u8(dest, d0);
+  dest += stride;
+  vst1_u8(dest, d1);
+  dest += stride;
+  vst1_u8(dest, d2);
+  dest += stride;
+  vst1_u8(dest, d3);
+  dest += stride;
+  vst1_u8(dest, d4);
+  dest += stride;
+  vst1_u8(dest, d5);
+  dest += stride;
+  vst1_u8(dest, d6);
+  dest += stride;
+  vst1_u8(dest, d7);
 }
 
 void vpx_idct8x8_64_add_neon(const tran_low_t *input, uint8_t *dest,
-                             int dest_stride) {
-  uint8_t *d1, *d2;
-  uint8x8_t d0u8, d1u8, d2u8, d3u8;
-  uint64x1_t d0u64, d1u64, d2u64, d3u64;
-  int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16, q13s16, q14s16, q15s16;
-  uint16x8_t q8u16, q9u16, q10u16, q11u16;
-
-  q8s16 = load_tran_low_to_s16(input);
-  q9s16 = load_tran_low_to_s16(input + 8);
-  q10s16 = load_tran_low_to_s16(input + 16);
-  q11s16 = load_tran_low_to_s16(input + 24);
-  q12s16 = load_tran_low_to_s16(input + 32);
-  q13s16 = load_tran_low_to_s16(input + 40);
-  q14s16 = load_tran_low_to_s16(input + 48);
-  q15s16 = load_tran_low_to_s16(input + 56);
-
-  transpose_s16_8x8(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
-                    &q15s16);
-
-  IDCT8x8_1D(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
-             &q15s16);
-
-  transpose_s16_8x8(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
-                    &q15s16);
-
-  IDCT8x8_1D(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
-             &q15s16);
-
-  q8s16 = vrshrq_n_s16(q8s16, 5);
-  q9s16 = vrshrq_n_s16(q9s16, 5);
-  q10s16 = vrshrq_n_s16(q10s16, 5);
-  q11s16 = vrshrq_n_s16(q11s16, 5);
-  q12s16 = vrshrq_n_s16(q12s16, 5);
-  q13s16 = vrshrq_n_s16(q13s16, 5);
-  q14s16 = vrshrq_n_s16(q14s16, 5);
-  q15s16 = vrshrq_n_s16(q15s16, 5);
-
-  d1 = d2 = dest;
-
-  d0u64 = vld1_u64((uint64_t *)d1);
-  d1 += dest_stride;
-  d1u64 = vld1_u64((uint64_t *)d1);
-  d1 += dest_stride;
-  d2u64 = vld1_u64((uint64_t *)d1);
-  d1 += dest_stride;
-  d3u64 = vld1_u64((uint64_t *)d1);
-  d1 += dest_stride;
-
-  q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16), vreinterpret_u8_u64(d0u64));
-  q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16), vreinterpret_u8_u64(d1u64));
-  q10u16 = vaddw_u8(vreinterpretq_u16_s16(q10s16), vreinterpret_u8_u64(d2u64));
-  q11u16 = vaddw_u8(vreinterpretq_u16_s16(q11s16), vreinterpret_u8_u64(d3u64));
-
-  d0u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
-  d1u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
-  d2u8 = vqmovun_s16(vreinterpretq_s16_u16(q10u16));
-  d3u8 = vqmovun_s16(vreinterpretq_s16_u16(q11u16));
-
-  vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d0u8));
-  d2 += dest_stride;
-  vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d1u8));
-  d2 += dest_stride;
-  vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d2u8));
-  d2 += dest_stride;
-  vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d3u8));
-  d2 += dest_stride;
-
-  q8s16 = q12s16;
-  q9s16 = q13s16;
-  q10s16 = q14s16;
-  q11s16 = q15s16;
-
-  d0u64 = vld1_u64((uint64_t *)d1);
-  d1 += dest_stride;
-  d1u64 = vld1_u64((uint64_t *)d1);
-  d1 += dest_stride;
-  d2u64 = vld1_u64((uint64_t *)d1);
-  d1 += dest_stride;
-  d3u64 = vld1_u64((uint64_t *)d1);
-  d1 += dest_stride;
-
-  q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16), vreinterpret_u8_u64(d0u64));
-  q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16), vreinterpret_u8_u64(d1u64));
-  q10u16 = vaddw_u8(vreinterpretq_u16_s16(q10s16), vreinterpret_u8_u64(d2u64));
-  q11u16 = vaddw_u8(vreinterpretq_u16_s16(q11s16), vreinterpret_u8_u64(d3u64));
-
-  d0u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
-  d1u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
-  d2u8 = vqmovun_s16(vreinterpretq_s16_u16(q10u16));
-  d3u8 = vqmovun_s16(vreinterpretq_s16_u16(q11u16));
-
-  vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d0u8));
-  d2 += dest_stride;
-  vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d1u8));
-  d2 += dest_stride;
-  vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d2u8));
-  d2 += dest_stride;
-  vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d3u8));
-  d2 += dest_stride;
+                             int stride) {
+  const int16x8_t cospis = vld1q_s16(kCospi);
+  const int16x4_t cospis0 = vget_low_s16(cospis);   // cospi 0, 8, 16, 24
+  const int16x4_t cospis1 = vget_high_s16(cospis);  // cospi 4, 12, 20, 28
+  int16x8_t a0 = load_tran_low_to_s16q(input);
+  int16x8_t a1 = load_tran_low_to_s16q(input + 8);
+  int16x8_t a2 = load_tran_low_to_s16q(input + 16);
+  int16x8_t a3 = load_tran_low_to_s16q(input + 24);
+  int16x8_t a4 = load_tran_low_to_s16q(input + 32);
+  int16x8_t a5 = load_tran_low_to_s16q(input + 40);
+  int16x8_t a6 = load_tran_low_to_s16q(input + 48);
+  int16x8_t a7 = load_tran_low_to_s16q(input + 56);
+
+  idct8x8_64_1d_bd8(cospis0, cospis1, &a0, &a1, &a2, &a3, &a4, &a5, &a6, &a7);
+  idct8x8_64_1d_bd8(cospis0, cospis1, &a0, &a1, &a2, &a3, &a4, &a5, &a6, &a7);
+  add8x8(a0, a1, a2, a3, a4, a5, a6, a7, dest, stride);
 }
 
 void vpx_idct8x8_12_add_neon(const tran_low_t *input, uint8_t *dest,
-                             int dest_stride) {
-  uint8_t *d1, *d2;
-  uint8x8_t d0u8, d1u8, d2u8, d3u8;
-  int16x4_t d10s16, d11s16, d12s16, d13s16, d16s16;
-  int16x4_t d26s16, d27s16, d28s16, d29s16;
-  uint64x1_t d0u64, d1u64, d2u64, d3u64;
-  int16x8_t q0s16, q1s16, q2s16, q3s16, q4s16, q5s16, q6s16, q7s16;
-  int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16, q13s16, q14s16, q15s16;
-  uint16x8_t q8u16, q9u16, q10u16, q11u16;
-  int32x4_t q9s32, q10s32, q11s32, q12s32;
-
-  q8s16 = load_tran_low_to_s16(input);
-  q9s16 = load_tran_low_to_s16(input + 8);
-  q10s16 = load_tran_low_to_s16(input + 16);
-  q11s16 = load_tran_low_to_s16(input + 24);
-  q12s16 = load_tran_low_to_s16(input + 32);
-  q13s16 = load_tran_low_to_s16(input + 40);
-  q14s16 = load_tran_low_to_s16(input + 48);
-  q15s16 = load_tran_low_to_s16(input + 56);
-
-  transpose_s16_8x8(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
-                    &q15s16);
-
-  // First transform rows
-  // stage 1
-  q0s16 = vdupq_n_s16((int16_t)cospi_28_64 * 2);
-  q1s16 = vdupq_n_s16((int16_t)cospi_4_64 * 2);
-
-  q4s16 = vqrdmulhq_s16(q9s16, q0s16);
-
-  q0s16 = vdupq_n_s16(-(int16_t)cospi_20_64 * 2);
-
-  q7s16 = vqrdmulhq_s16(q9s16, q1s16);
-
-  q1s16 = vdupq_n_s16((int16_t)cospi_12_64 * 2);
-
-  q5s16 = vqrdmulhq_s16(q11s16, q0s16);
-
-  q0s16 = vdupq_n_s16((int16_t)cospi_16_64 * 2);
-
-  q6s16 = vqrdmulhq_s16(q11s16, q1s16);
-
-  // stage 2 & stage 3 - even half
-  q1s16 = vdupq_n_s16((int16_t)cospi_24_64 * 2);
-
-  q9s16 = vqrdmulhq_s16(q8s16, q0s16);
-
-  q0s16 = vdupq_n_s16((int16_t)cospi_8_64 * 2);
-
-  q13s16 = vqrdmulhq_s16(q10s16, q1s16);
-
-  q15s16 = vqrdmulhq_s16(q10s16, q0s16);
-
-  // stage 3 -odd half
-  q0s16 = vaddq_s16(q9s16, q15s16);
-  q1s16 = vaddq_s16(q9s16, q13s16);
-  q2s16 = vsubq_s16(q9s16, q13s16);
-  q3s16 = vsubq_s16(q9s16, q15s16);
-
-  // stage 2 - odd half
-  q13s16 = vsubq_s16(q4s16, q5s16);
-  q4s16 = vaddq_s16(q4s16, q5s16);
-  q14s16 = vsubq_s16(q7s16, q6s16);
-  q7s16 = vaddq_s16(q7s16, q6s16);
-  d26s16 = vget_low_s16(q13s16);
-  d27s16 = vget_high_s16(q13s16);
-  d28s16 = vget_low_s16(q14s16);
-  d29s16 = vget_high_s16(q14s16);
-
-  d16s16 = vdup_n_s16((int16_t)cospi_16_64);
-  q9s32 = vmull_s16(d28s16, d16s16);
-  q10s32 = vmull_s16(d29s16, d16s16);
-  q11s32 = vmull_s16(d28s16, d16s16);
-  q12s32 = vmull_s16(d29s16, d16s16);
-
-  q9s32 = vmlsl_s16(q9s32, d26s16, d16s16);
-  q10s32 = vmlsl_s16(q10s32, d27s16, d16s16);
-  q11s32 = vmlal_s16(q11s32, d26s16, d16s16);
-  q12s32 = vmlal_s16(q12s32, d27s16, d16s16);
-
-  d10s16 = vqrshrn_n_s32(q9s32, 14);
-  d11s16 = vqrshrn_n_s32(q10s32, 14);
-  d12s16 = vqrshrn_n_s32(q11s32, 14);
-  d13s16 = vqrshrn_n_s32(q12s32, 14);
-  q5s16 = vcombine_s16(d10s16, d11s16);
-  q6s16 = vcombine_s16(d12s16, d13s16);
-
-  // stage 4
-  q8s16 = vaddq_s16(q0s16, q7s16);
-  q9s16 = vaddq_s16(q1s16, q6s16);
-  q10s16 = vaddq_s16(q2s16, q5s16);
-  q11s16 = vaddq_s16(q3s16, q4s16);
-  q12s16 = vsubq_s16(q3s16, q4s16);
-  q13s16 = vsubq_s16(q2s16, q5s16);
-  q14s16 = vsubq_s16(q1s16, q6s16);
-  q15s16 = vsubq_s16(q0s16, q7s16);
-
-  transpose_s16_8x8(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
-                    &q15s16);
-
-  IDCT8x8_1D(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
-             &q15s16);
-
-  q8s16 = vrshrq_n_s16(q8s16, 5);
-  q9s16 = vrshrq_n_s16(q9s16, 5);
-  q10s16 = vrshrq_n_s16(q10s16, 5);
-  q11s16 = vrshrq_n_s16(q11s16, 5);
-  q12s16 = vrshrq_n_s16(q12s16, 5);
-  q13s16 = vrshrq_n_s16(q13s16, 5);
-  q14s16 = vrshrq_n_s16(q14s16, 5);
-  q15s16 = vrshrq_n_s16(q15s16, 5);
-
-  d1 = d2 = dest;
-
-  d0u64 = vld1_u64((uint64_t *)d1);
-  d1 += dest_stride;
-  d1u64 = vld1_u64((uint64_t *)d1);
-  d1 += dest_stride;
-  d2u64 = vld1_u64((uint64_t *)d1);
-  d1 += dest_stride;
-  d3u64 = vld1_u64((uint64_t *)d1);
-  d1 += dest_stride;
-
-  q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16), vreinterpret_u8_u64(d0u64));
-  q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16), vreinterpret_u8_u64(d1u64));
-  q10u16 = vaddw_u8(vreinterpretq_u16_s16(q10s16), vreinterpret_u8_u64(d2u64));
-  q11u16 = vaddw_u8(vreinterpretq_u16_s16(q11s16), vreinterpret_u8_u64(d3u64));
-
-  d0u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
-  d1u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
-  d2u8 = vqmovun_s16(vreinterpretq_s16_u16(q10u16));
-  d3u8 = vqmovun_s16(vreinterpretq_s16_u16(q11u16));
-
-  vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d0u8));
-  d2 += dest_stride;
-  vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d1u8));
-  d2 += dest_stride;
-  vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d2u8));
-  d2 += dest_stride;
-  vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d3u8));
-  d2 += dest_stride;
-
-  q8s16 = q12s16;
-  q9s16 = q13s16;
-  q10s16 = q14s16;
-  q11s16 = q15s16;
-
-  d0u64 = vld1_u64((uint64_t *)d1);
-  d1 += dest_stride;
-  d1u64 = vld1_u64((uint64_t *)d1);
-  d1 += dest_stride;
-  d2u64 = vld1_u64((uint64_t *)d1);
-  d1 += dest_stride;
-  d3u64 = vld1_u64((uint64_t *)d1);
-  d1 += dest_stride;
-
-  q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16), vreinterpret_u8_u64(d0u64));
-  q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16), vreinterpret_u8_u64(d1u64));
-  q10u16 = vaddw_u8(vreinterpretq_u16_s16(q10s16), vreinterpret_u8_u64(d2u64));
-  q11u16 = vaddw_u8(vreinterpretq_u16_s16(q11s16), vreinterpret_u8_u64(d3u64));
-
-  d0u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
-  d1u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
-  d2u8 = vqmovun_s16(vreinterpretq_s16_u16(q10u16));
-  d3u8 = vqmovun_s16(vreinterpretq_s16_u16(q11u16));
-
-  vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d0u8));
-  d2 += dest_stride;
-  vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d1u8));
-  d2 += dest_stride;
-  vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d2u8));
-  d2 += dest_stride;
-  vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d3u8));
-  d2 += dest_stride;
+                             int stride) {
+  const int16x8_t cospis = vld1q_s16(kCospi);
+  const int16x8_t cospisd = vaddq_s16(cospis, cospis);
+  const int16x4_t cospis0 = vget_low_s16(cospis);     // cospi 0, 8, 16, 24
+  const int16x4_t cospisd0 = vget_low_s16(cospisd);   // doubled 0, 8, 16, 24
+  const int16x4_t cospisd1 = vget_high_s16(cospisd);  // doubled 4, 12, 20, 28
+  int16x4_t a0, a1, a2, a3, a4, a5, a6, a7;
+  int16x8_t b0, b1, b2, b3, b4, b5, b6, b7;
+
+  a0 = load_tran_low_to_s16d(input);
+  a1 = load_tran_low_to_s16d(input + 8);
+  a2 = load_tran_low_to_s16d(input + 16);
+  a3 = load_tran_low_to_s16d(input + 24);
+
+  idct8x8_12_pass1_bd8(cospis0, cospisd0, cospisd1, &a0, &a1, &a2, &a3, &a4,
+                       &a5, &a6, &a7);
+  idct8x8_12_pass2_bd8(cospis0, cospisd0, cospisd1, a0, a1, a2, a3, a4, a5, a6,
+                       a7, &b0, &b1, &b2, &b3, &b4, &b5, &b6, &b7);
+  add8x8(b0, b1, b2, b3, b4, b5, b6, b7, dest, stride);
 }