Update vp9_iht4x4_16_add_neon()

Change-Id: Ica8dbe5f8167e5d370d89d233c598b70bba123b7

test/dct_test.cc

@@ -752,6 +752,29 @@ INSTANTIATE_TEST_CASE_P(C, TransHT, ::testing::ValuesIn(c_ht_tests));
 
 #if !CONFIG_EMULATE_HARDWARE
 
+#if HAVE_NEON
+INSTANTIATE_TEST_CASE_P(
+    NEON, TransHT,
+    ::testing::Values(
+        make_tuple(&vp9_fht8x8_c, &iht_wrapper<vp9_iht8x8_64_add_neon>, 8, 0,
+                   VPX_BITS_8, 1),
+        make_tuple(&vp9_fht8x8_c, &iht_wrapper<vp9_iht8x8_64_add_neon>, 8, 1,
+                   VPX_BITS_8, 1),
+        make_tuple(&vp9_fht8x8_c, &iht_wrapper<vp9_iht8x8_64_add_neon>, 8, 2,
+                   VPX_BITS_8, 1),
+        make_tuple(&vp9_fht8x8_c, &iht_wrapper<vp9_iht8x8_64_add_neon>, 8, 3,
+                   VPX_BITS_8, 1),
+
+        make_tuple(&vp9_fht4x4_c, &iht_wrapper<vp9_iht4x4_16_add_neon>, 4, 0,
+                   VPX_BITS_8, 1),
+        make_tuple(&vp9_fht4x4_c, &iht_wrapper<vp9_iht4x4_16_add_neon>, 4, 1,
+                   VPX_BITS_8, 1),
+        make_tuple(&vp9_fht4x4_c, &iht_wrapper<vp9_iht4x4_16_add_neon>, 4, 2,
+                   VPX_BITS_8, 1),
+        make_tuple(&vp9_fht4x4_c, &iht_wrapper<vp9_iht4x4_16_add_neon>, 4, 3,
+                   VPX_BITS_8, 1)));
+#endif  // HAVE_NEON
+
 #if HAVE_SSE2
 INSTANTIATE_TEST_CASE_P(
     SSE2, TransHT,

vp9/common/arm/neon/vp9_iht4x4_add_neon.c

@@ -14,206 +14,103 @@
 #include "./vp9_rtcd.h"
 #include "./vpx_config.h"
 #include "vp9/common/vp9_common.h"
+#include "vpx_dsp/arm/idct_neon.h"
+#include "vpx_dsp/arm/mem_neon.h"
 #include "vpx_dsp/txfm_common.h"
 
-static INLINE void TRANSPOSE4X4(int16x8_t *q8s16, int16x8_t *q9s16) {
+static INLINE void iadst4(int16x8_t *const io) {
-  int32x4_t q8s32, q9s32;
+  const int32x4_t c3 = vdupq_n_s32(sinpi_3_9);
-  int16x4x2_t d0x2s16, d1x2s16;
+  int16x4_t c[5], x[4];
-  int32x4x2_t q0x2s32;
+  int32x4_t s[8], output[4];
 
-  d0x2s16 = vtrn_s16(vget_low_s16(*q8s16), vget_high_s16(*q8s16));
+  c[1] = vdup_n_s16(sinpi_1_9);
-  d1x2s16 = vtrn_s16(vget_low_s16(*q9s16), vget_high_s16(*q9s16));
+  c[2] = vdup_n_s16(sinpi_2_9);
+  c[3] = vdup_n_s16(sinpi_3_9);
+  c[4] = vdup_n_s16(sinpi_4_9);
 
-  q8s32 = vreinterpretq_s32_s16(vcombine_s16(d0x2s16.val[0], d0x2s16.val[1]));
+  x[0] = vget_low_s16(io[0]);
-  q9s32 = vreinterpretq_s32_s16(vcombine_s16(d1x2s16.val[0], d1x2s16.val[1]));
+  x[1] = vget_low_s16(io[1]);
-  q0x2s32 = vtrnq_s32(q8s32, q9s32);
+  x[2] = vget_high_s16(io[0]);
+  x[3] = vget_high_s16(io[1]);
 
-  *q8s16 = vreinterpretq_s16_s32(q0x2s32.val[0]);
+  s[0] = vmull_s16(c[1], x[0]);
-  *q9s16 = vreinterpretq_s16_s32(q0x2s32.val[1]);
+  s[1] = vmull_s16(c[2], x[0]);
-}
+  s[2] = vmull_s16(c[3], x[1]);
+  s[3] = vmull_s16(c[4], x[2]);
+  s[4] = vmull_s16(c[1], x[2]);
+  s[5] = vmull_s16(c[2], x[3]);
+  s[6] = vmull_s16(c[4], x[3]);
+  s[7] = vaddl_s16(x[0], x[3]);
+  s[7] = vsubw_s16(s[7], x[2]);
 
-static INLINE void GENERATE_COSINE_CONSTANTS(int16x4_t *d0s16, int16x4_t *d1s16,
+  s[0] = vaddq_s32(s[0], s[3]);
-                                             int16x4_t *d2s16) {
+  s[0] = vaddq_s32(s[0], s[5]);
-  *d0s16 = vdup_n_s16(cospi_8_64);
+  s[1] = vsubq_s32(s[1], s[4]);
-  *d1s16 = vdup_n_s16(cospi_16_64);
+  s[1] = vsubq_s32(s[1], s[6]);
-  *d2s16 = vdup_n_s16(cospi_24_64);
+  s[3] = s[2];
-}
+  s[2] = vmulq_s32(c3, s[7]);
 
-static INLINE void GENERATE_SINE_CONSTANTS(int16x4_t *d3s16, int16x4_t *d4s16,
+  output[0] = vaddq_s32(s[0], s[3]);
-                                           int16x4_t *d5s16, int16x8_t *q3s16) {
+  output[1] = vaddq_s32(s[1], s[3]);
-  *d3s16 = vdup_n_s16(sinpi_1_9);
+  output[2] = s[2];
-  *d4s16 = vdup_n_s16(sinpi_2_9);
+  output[3] = vaddq_s32(s[0], s[1]);
-  *q3s16 = vdupq_n_s16(sinpi_3_9);
+  output[3] = vsubq_s32(output[3], s[3]);
-  *d5s16 = vdup_n_s16(sinpi_4_9);
+  dct_const_round_shift_low_8_dual(output, &io[0], &io[1]);
-}
-
-static INLINE void IDCT4x4_1D(int16x4_t *d0s16, int16x4_t *d1s16,
-                              int16x4_t *d2s16, int16x8_t *q8s16,
-                              int16x8_t *q9s16) {
-  int16x4_t d16s16, d17s16, d18s16, d19s16, d23s16, d24s16;
-  int16x4_t d26s16, d27s16, d28s16, d29s16;
-  int32x4_t q10s32, q13s32, q14s32, q15s32;
-  int16x8_t q13s16, q14s16;
-
-  d16s16 = vget_low_s16(*q8s16);
-  d17s16 = vget_high_s16(*q8s16);
-  d18s16 = vget_low_s16(*q9s16);
-  d19s16 = vget_high_s16(*q9s16);
-
-  d23s16 = vadd_s16(d16s16, d18s16);
-  d24s16 = vsub_s16(d16s16, d18s16);
-
-  q15s32 = vmull_s16(d17s16, *d2s16);
-  q10s32 = vmull_s16(d17s16, *d0s16);
-  q13s32 = vmull_s16(d23s16, *d1s16);
-  q14s32 = vmull_s16(d24s16, *d1s16);
-  q15s32 = vmlsl_s16(q15s32, d19s16, *d0s16);
-  q10s32 = vmlal_s16(q10s32, d19s16, *d2s16);
-
-  d26s16 = vrshrn_n_s32(q13s32, 14);
-  d27s16 = vrshrn_n_s32(q14s32, 14);
-  d29s16 = vrshrn_n_s32(q15s32, 14);
-  d28s16 = vrshrn_n_s32(q10s32, 14);
-
-  q13s16 = vcombine_s16(d26s16, d27s16);
-  q14s16 = vcombine_s16(d28s16, d29s16);
-  *q8s16 = vaddq_s16(q13s16, q14s16);
-  *q9s16 = vsubq_s16(q13s16, q14s16);
-  *q9s16 = vcombine_s16(vget_high_s16(*q9s16), vget_low_s16(*q9s16));  // vswp
-}
-
-static INLINE void IADST4x4_1D(int16x4_t *d3s16, int16x4_t *d4s16,
-                               int16x4_t *d5s16, int16x8_t *q3s16,
-                               int16x8_t *q8s16, int16x8_t *q9s16) {
-  int16x4_t d6s16, d16s16, d17s16, d18s16, d19s16;
-  int32x4_t q8s32, q9s32, q10s32, q11s32, q12s32, q13s32, q14s32, q15s32;
-
-  d6s16 = vget_low_s16(*q3s16);
-
-  d16s16 = vget_low_s16(*q8s16);
-  d17s16 = vget_high_s16(*q8s16);
-  d18s16 = vget_low_s16(*q9s16);
-  d19s16 = vget_high_s16(*q9s16);
-
-  q10s32 = vmull_s16(*d3s16, d16s16);
-  q11s32 = vmull_s16(*d4s16, d16s16);
-  q12s32 = vmull_s16(d6s16, d17s16);
-  q13s32 = vmull_s16(*d5s16, d18s16);
-  q14s32 = vmull_s16(*d3s16, d18s16);
-  q15s32 = vmovl_s16(d16s16);
-  q15s32 = vaddw_s16(q15s32, d19s16);
-  q8s32 = vmull_s16(*d4s16, d19s16);
-  q15s32 = vsubw_s16(q15s32, d18s16);
-  q9s32 = vmull_s16(*d5s16, d19s16);
-
-  q10s32 = vaddq_s32(q10s32, q13s32);
-  q10s32 = vaddq_s32(q10s32, q8s32);
-  q11s32 = vsubq_s32(q11s32, q14s32);
-  q8s32 = vdupq_n_s32(sinpi_3_9);
-  q11s32 = vsubq_s32(q11s32, q9s32);
-  q15s32 = vmulq_s32(q15s32, q8s32);
-
-  q13s32 = vaddq_s32(q10s32, q12s32);
-  q10s32 = vaddq_s32(q10s32, q11s32);
-  q14s32 = vaddq_s32(q11s32, q12s32);
-  q10s32 = vsubq_s32(q10s32, q12s32);
-
-  d16s16 = vrshrn_n_s32(q13s32, 14);
-  d17s16 = vrshrn_n_s32(q14s32, 14);
-  d18s16 = vrshrn_n_s32(q15s32, 14);
-  d19s16 = vrshrn_n_s32(q10s32, 14);
-
-  *q8s16 = vcombine_s16(d16s16, d17s16);
-  *q9s16 = vcombine_s16(d18s16, d19s16);
 }
 
 void vp9_iht4x4_16_add_neon(const tran_low_t *input, uint8_t *dest, int stride,
                             int tx_type) {
-  uint8x8_t d26u8, d27u8;
+  int16x8_t a[2];
-  int16x4_t d0s16, d1s16, d2s16, d3s16, d4s16, d5s16;
+  uint8x8_t s[2], d[2];
-  uint32x2_t d26u32, d27u32;
+  uint16x8_t sum[2];
-  int16x8_t q3s16, q8s16, q9s16;
-  uint16x8_t q8u16, q9u16;
 
-  d26u32 = d27u32 = vdup_n_u32(0);
+  assert(!((intptr_t)dest % sizeof(uint32_t)));
+  assert(!(stride % sizeof(uint32_t)));
 
-  q8s16 = vld1q_s16(input);
+  a[0] = load_tran_low_to_s16q(input);
-  q9s16 = vld1q_s16(input + 8);
+  a[1] = load_tran_low_to_s16q(input + 8);
-
+  transpose_s16_4x4q(&a[0], &a[1]);
-  TRANSPOSE4X4(&q8s16, &q9s16);
 
   switch (tx_type) {
-    case 0:  // idct_idct is not supported. Fall back to C
+    case 0:  // DCT_DCT
-      vp9_iht4x4_16_add_c(input, dest, stride, tx_type);
+      idct4x4_16_kernel_bd8(a);
-      return;
+      a[1] = vcombine_s16(vget_high_s16(a[1]), vget_low_s16(a[1]));
-    case 1:  // iadst_idct
+      transpose_s16_4x4q(&a[0], &a[1]);
-      // generate constants
+      idct4x4_16_kernel_bd8(a);
-      GENERATE_COSINE_CONSTANTS(&d0s16, &d1s16, &d2s16);
+      a[1] = vcombine_s16(vget_high_s16(a[1]), vget_low_s16(a[1]));
-      GENERATE_SINE_CONSTANTS(&d3s16, &d4s16, &d5s16, &q3s16);
-
-      // first transform rows
-      IDCT4x4_1D(&d0s16, &d1s16, &d2s16, &q8s16, &q9s16);
-
-      // transpose the matrix
-      TRANSPOSE4X4(&q8s16, &q9s16);
-
-      // then transform columns
-      IADST4x4_1D(&d3s16, &d4s16, &d5s16, &q3s16, &q8s16, &q9s16);
       break;
-    case 2:  // idct_iadst
-      // generate constantsyy
-      GENERATE_COSINE_CONSTANTS(&d0s16, &d1s16, &d2s16);
-      GENERATE_SINE_CONSTANTS(&d3s16, &d4s16, &d5s16, &q3s16);
 
-      // first transform rows
+    case 1:  // ADST_DCT
-      IADST4x4_1D(&d3s16, &d4s16, &d5s16, &q3s16, &q8s16, &q9s16);
+      idct4x4_16_kernel_bd8(a);
-
+      a[1] = vcombine_s16(vget_high_s16(a[1]), vget_low_s16(a[1]));
-      // transpose the matrix
+      transpose_s16_4x4q(&a[0], &a[1]);
-      TRANSPOSE4X4(&q8s16, &q9s16);
+      iadst4(a);
-
-      // then transform columns
-      IDCT4x4_1D(&d0s16, &d1s16, &d2s16, &q8s16, &q9s16);
       break;
-    case 3:  // iadst_iadst
-      // generate constants
-      GENERATE_SINE_CONSTANTS(&d3s16, &d4s16, &d5s16, &q3s16);
 
-      // first transform rows
+    case 2:  // DCT_ADST
-      IADST4x4_1D(&d3s16, &d4s16, &d5s16, &q3s16, &q8s16, &q9s16);
+      iadst4(a);
-
+      transpose_s16_4x4q(&a[0], &a[1]);
-      // transpose the matrix
+      idct4x4_16_kernel_bd8(a);
-      TRANSPOSE4X4(&q8s16, &q9s16);
+      a[1] = vcombine_s16(vget_high_s16(a[1]), vget_low_s16(a[1]));
-
-      // then transform columns
-      IADST4x4_1D(&d3s16, &d4s16, &d5s16, &q3s16, &q8s16, &q9s16);
       break;
-    default:  // iadst_idct
+
-      assert(0);
+    case 3:  // ADST_ADST
+      iadst4(a);
+      transpose_s16_4x4q(&a[0], &a[1]);
+      iadst4(a);
       break;
+
+    default: assert(0); break;
   }
 
-  q8s16 = vrshrq_n_s16(q8s16, 4);
+  a[0] = vrshrq_n_s16(a[0], 4);
-  q9s16 = vrshrq_n_s16(q9s16, 4);
+  a[1] = vrshrq_n_s16(a[1], 4);
-
+  s[0] = load_u8(dest, stride);
-  d26u32 = vld1_lane_u32((const uint32_t *)dest, d26u32, 0);
+  s[1] = load_u8(dest + 2 * stride, stride);
-  dest += stride;
+  sum[0] = vaddw_u8(vreinterpretq_u16_s16(a[0]), s[0]);
-  d26u32 = vld1_lane_u32((const uint32_t *)dest, d26u32, 1);
+  sum[1] = vaddw_u8(vreinterpretq_u16_s16(a[1]), s[1]);
-  dest += stride;
+  d[0] = vqmovun_s16(vreinterpretq_s16_u16(sum[0]));
-  d27u32 = vld1_lane_u32((const uint32_t *)dest, d27u32, 0);
+  d[1] = vqmovun_s16(vreinterpretq_s16_u16(sum[1]));
-  dest += stride;
+  store_u8(dest, stride, d[0]);
-  d27u32 = vld1_lane_u32((const uint32_t *)dest, d27u32, 1);
+  store_u8(dest + 2 * stride, stride, d[1]);
-
-  q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16), vreinterpret_u8_u32(d26u32));
-  q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16), vreinterpret_u8_u32(d27u32));
-
-  d26u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
-  d27u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
-
-  vst1_lane_u32((uint32_t *)dest, vreinterpret_u32_u8(d27u8), 1);
-  dest -= stride;
-  vst1_lane_u32((uint32_t *)dest, vreinterpret_u32_u8(d27u8), 0);
-  dest -= stride;
-  vst1_lane_u32((uint32_t *)dest, vreinterpret_u32_u8(d26u8), 1);
-  dest -= stride;
-  vst1_lane_u32((uint32_t *)dest, vreinterpret_u32_u8(d26u8), 0);
 }

vp9/common/arm/neon/vp9_iht8x8_add_neon.c

@@ -14,6 +14,7 @@
 #include "./vp9_rtcd.h"
 #include "./vpx_config.h"
 #include "vp9/common/vp9_common.h"
+#include "vpx_dsp/arm/mem_neon.h"
 #include "vpx_dsp/arm/transpose_neon.h"
 
 static int16_t cospi_2_64 = 16305;
@@ -415,14 +416,14 @@ void vp9_iht8x8_64_add_neon(const tran_low_t *input, uint8_t *dest, int stride,
   int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16, q13s16, q14s16, q15s16;
   uint16x8_t q8u16, q9u16, q10u16, q11u16;
 
-  q8s16 = vld1q_s16(input);
+  q8s16 = load_tran_low_to_s16q(input);
-  q9s16 = vld1q_s16(input + 8);
+  q9s16 = load_tran_low_to_s16q(input + 8);
-  q10s16 = vld1q_s16(input + 8 * 2);
+  q10s16 = load_tran_low_to_s16q(input + 8 * 2);
-  q11s16 = vld1q_s16(input + 8 * 3);
+  q11s16 = load_tran_low_to_s16q(input + 8 * 3);
-  q12s16 = vld1q_s16(input + 8 * 4);
+  q12s16 = load_tran_low_to_s16q(input + 8 * 4);
-  q13s16 = vld1q_s16(input + 8 * 5);
+  q13s16 = load_tran_low_to_s16q(input + 8 * 5);
-  q14s16 = vld1q_s16(input + 8 * 6);
+  q14s16 = load_tran_low_to_s16q(input + 8 * 6);
-  q15s16 = vld1q_s16(input + 8 * 7);
+  q15s16 = load_tran_low_to_s16q(input + 8 * 7);
 
   transpose_s16_8x8(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
                     &q15s16);

vp9/common/vp9_rtcd_defs.pl

@@ -67,13 +67,13 @@ add_proto qw/void vp9_iht16x16_256_add/, "const tran_low_t *input, uint8_t *outp
 if (vpx_config("CONFIG_EMULATE_HARDWARE") ne "yes") {
   # Note that there are more specializations appended when
   # CONFIG_VP9_HIGHBITDEPTH is off.
-  specialize qw/vp9_iht4x4_16_add sse2/;
+  specialize qw/vp9_iht4x4_16_add neon sse2/;
-  specialize qw/vp9_iht8x8_64_add sse2/;
+  specialize qw/vp9_iht8x8_64_add neon sse2/;
   specialize qw/vp9_iht16x16_256_add sse2/;
   if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") ne "yes") {
     # Note that these specializations are appended to the above ones.
-    specialize qw/vp9_iht4x4_16_add neon dspr2 msa/;
+    specialize qw/vp9_iht4x4_16_add dspr2 msa/;
-    specialize qw/vp9_iht8x8_64_add neon dspr2 msa/;
+    specialize qw/vp9_iht8x8_64_add dspr2 msa/;
     specialize qw/vp9_iht16x16_256_add dspr2 msa/;
   }
 }

vp9/vp9_common.mk

@@ -64,10 +64,12 @@ VP9_COMMON_SRCS-$(CONFIG_VP9_POSTPROC) += common/vp9_postproc.c
 VP9_COMMON_SRCS-$(CONFIG_VP9_POSTPROC) += common/vp9_mfqe.h
 VP9_COMMON_SRCS-$(CONFIG_VP9_POSTPROC) += common/vp9_mfqe.c
 
-VP9_COMMON_SRCS-$(HAVE_MSA)  += common/mips/msa/vp9_idct4x4_msa.c
+VP9_COMMON_SRCS-$(HAVE_MSA)   += common/mips/msa/vp9_idct4x4_msa.c
-VP9_COMMON_SRCS-$(HAVE_MSA)  += common/mips/msa/vp9_idct8x8_msa.c
+VP9_COMMON_SRCS-$(HAVE_MSA)   += common/mips/msa/vp9_idct8x8_msa.c
-VP9_COMMON_SRCS-$(HAVE_MSA)  += common/mips/msa/vp9_idct16x16_msa.c
+VP9_COMMON_SRCS-$(HAVE_MSA)   += common/mips/msa/vp9_idct16x16_msa.c
-VP9_COMMON_SRCS-$(HAVE_SSE2) += common/x86/vp9_idct_intrin_sse2.c
+VP9_COMMON_SRCS-$(HAVE_SSE2)  += common/x86/vp9_idct_intrin_sse2.c
+VP9_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/vp9_iht4x4_add_neon.c
+VP9_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/vp9_iht8x8_add_neon.c
 
 ifeq ($(CONFIG_VP9_POSTPROC),yes)
 VP9_COMMON_SRCS-$(HAVE_MSA)  += common/mips/msa/vp9_mfqe_msa.c
@@ -78,8 +80,6 @@ ifneq ($(CONFIG_VP9_HIGHBITDEPTH),yes)
 VP9_COMMON_SRCS-$(HAVE_DSPR2) += common/mips/dspr2/vp9_itrans4_dspr2.c
 VP9_COMMON_SRCS-$(HAVE_DSPR2) += common/mips/dspr2/vp9_itrans8_dspr2.c
 VP9_COMMON_SRCS-$(HAVE_DSPR2) += common/mips/dspr2/vp9_itrans16_dspr2.c
-VP9_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/vp9_iht4x4_add_neon.c
-VP9_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/vp9_iht8x8_add_neon.c
 else
 VP9_COMMON_SRCS-$(HAVE_SSE4_1) += common/x86/vp9_highbd_iht4x4_add_sse4.c
 VP9_COMMON_SRCS-$(HAVE_SSE4_1) += common/x86/vp9_highbd_iht8x8_add_sse4.c