neon fdct: 4x4 implementation

Approximately twice as fast as C implementation.

BUG=webm:1424

Change-Id: I3c0307fb08ddc23df42545cd089a78e2ed5c9d3f

diff --git a/test/fdct4x4_test.cc b/test/fdct4x4_test.cc
index bd2327520bce00cdc61d0e7d3f443d670e76c4f8..aa90bfa181a63dcac7d966af4bbf97f52c8d0312 100644 (file)
--- a/test/fdct4x4_test.cc
+++ b/test/fdct4x4_test.cc
@@ -440,7 +440,7 @@ INSTANTIATE_TEST_CASE_P(C, Trans4x4WHT,
 
 #if HAVE_NEON && !CONFIG_EMULATE_HARDWARE
 INSTANTIATE_TEST_CASE_P(NEON, Trans4x4DCT,
-                        ::testing::Values(make_tuple(&vpx_fdct4x4_c,
+                        ::testing::Values(make_tuple(&vpx_fdct4x4_neon,
                                                      &vpx_idct4x4_16_add_neon,
                                                      0, VPX_BITS_8)));
 #if !CONFIG_VP9_HIGHBITDEPTH

diff --git a/vpx_dsp/arm/fdct_neon.c b/vpx_dsp/arm/fdct_neon.c
new file mode 100644 (file)
index 0000000..fe78f3f
--- /dev/null
+++ b/vpx_dsp/arm/fdct_neon.c
@@ -0,0 +1,92 @@
+/*
+ *  Copyright (c) 2017 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+#include "./vpx_config.h"
+#include "vpx_dsp/txfm_common.h"
+#include "vpx_dsp/vpx_dsp_common.h"
+#include "vpx_dsp/arm/idct_neon.h"
+#include "vpx_dsp/arm/mem_neon.h"
+#include "vpx_dsp/arm/transpose_neon.h"
+
+void vpx_fdct4x4_neon(const int16_t *input, tran_low_t *final_output,
+                      int stride) {
+  int i;
+  // input[M * stride] * 16
+  int16x4_t input_0 = vshl_n_s16(vld1_s16(input + 0 * stride), 4);
+  int16x4_t input_1 = vshl_n_s16(vld1_s16(input + 1 * stride), 4);
+  int16x4_t input_2 = vshl_n_s16(vld1_s16(input + 2 * stride), 4);
+  int16x4_t input_3 = vshl_n_s16(vld1_s16(input + 3 * stride), 4);
+
+  // If the very first value != 0, then add 1.
+  if (input[0] != 0) {
+    const int16x4_t one = vreinterpret_s16_s64(vdup_n_s64(1));
+    input_0 = vadd_s16(input_0, one);
+  }
+
+  for (i = 0; i < 2; ++i) {
+    const int16x8_t input_01 = vcombine_s16(input_0, input_1);
+    const int16x8_t input_32 = vcombine_s16(input_3, input_2);
+
+    // in_0 +/- in_3, in_1 +/- in_2
+    const int16x8_t s_01 = vaddq_s16(input_01, input_32);
+    const int16x8_t s_32 = vsubq_s16(input_01, input_32);
+
+    // step_0 +/- step_1, step_2 +/- step_3
+    const int16x4_t s_0 = vget_low_s16(s_01);
+    const int16x4_t s_1 = vget_high_s16(s_01);
+    const int16x4_t s_2 = vget_high_s16(s_32);
+    const int16x4_t s_3 = vget_low_s16(s_32);
+
+    // (s_0 +/- s_1) * cospi_16_64
+    // Must expand all elements to s32. See 'needs32' comment in fwd_txfm.c.
+    const int32x4_t s_0_p_s_1 = vaddl_s16(s_0, s_1);
+    const int32x4_t s_0_m_s_1 = vsubl_s16(s_0, s_1);
+    const int32x4_t temp1 = vmulq_n_s32(s_0_p_s_1, (int16_t)cospi_16_64);
+    const int32x4_t temp2 = vmulq_n_s32(s_0_m_s_1, (int16_t)cospi_16_64);
+
+    // fdct_round_shift
+    int16x4_t out_0 = vrshrn_n_s32(temp1, DCT_CONST_BITS);
+    int16x4_t out_2 = vrshrn_n_s32(temp2, DCT_CONST_BITS);
+
+    // s_3 * cospi_8_64 + s_2 * cospi_24_64
+    // s_3 * cospi_24_64 - s_2 * cospi_8_64
+    const int32x4_t s_3_cospi_8_64 = vmull_n_s16(s_3, (int16_t)cospi_8_64);
+    const int32x4_t s_3_cospi_24_64 = vmull_n_s16(s_3, (int16_t)cospi_24_64);
+
+    const int32x4_t temp3 =
+        vmlal_n_s16(s_3_cospi_8_64, s_2, (int16_t)cospi_24_64);
+    const int32x4_t temp4 =
+        vmlsl_n_s16(s_3_cospi_24_64, s_2, (int16_t)cospi_8_64);
+
+    // fdct_round_shift
+    int16x4_t out_1 = vrshrn_n_s32(temp3, DCT_CONST_BITS);
+    int16x4_t out_3 = vrshrn_n_s32(temp4, DCT_CONST_BITS);
+
+    transpose_s16_4x4d(&out_0, &out_1, &out_2, &out_3);
+
+    input_0 = out_0;
+    input_1 = out_1;
+    input_2 = out_2;
+    input_3 = out_3;
+  }
+
+  {
+    // Not quite a rounding shift. Only add 1 despite shifting by 2.
+    const int16x8_t one = vdupq_n_s16(1);
+    int16x8_t out_01 = vcombine_s16(input_0, input_1);
+    int16x8_t out_23 = vcombine_s16(input_2, input_3);
+    out_01 = vshrq_n_s16(vaddq_s16(out_01, one), 2);
+    out_23 = vshrq_n_s16(vaddq_s16(out_23, one), 2);
+    store_s16q_to_tran_low(final_output + 0 * 8, out_01);
+    store_s16q_to_tran_low(final_output + 1 * 8, out_23);
+  }
+}

diff --git a/vpx_dsp/vpx_dsp.mk b/vpx_dsp/vpx_dsp.mk
index afa14accc2ac083d9e82fe9533702dcfb51dbc1d..6ac7182abdef048014ff8a0f685476d3451ef88a 100644 (file)
--- a/vpx_dsp/vpx_dsp.mk
+++ b/vpx_dsp/vpx_dsp.mk
@@ -193,6 +193,7 @@ DSP_SRCS-$(HAVE_SSSE3)  += x86/fwd_txfm_ssse3_x86_64.asm
 endif
 DSP_SRCS-$(HAVE_AVX2)   += x86/fwd_txfm_avx2.c
 DSP_SRCS-$(HAVE_AVX2)   += x86/fwd_dct32x32_impl_avx2.h
+DSP_SRCS-$(HAVE_NEON)   += arm/fdct_neon.c
 DSP_SRCS-$(HAVE_NEON)   += arm/fwd_txfm_neon.c
 DSP_SRCS-$(HAVE_MSA)    += mips/fwd_txfm_msa.h
 DSP_SRCS-$(HAVE_MSA)    += mips/fwd_txfm_msa.c

diff --git a/vpx_dsp/vpx_dsp_rtcd_defs.pl b/vpx_dsp/vpx_dsp_rtcd_defs.pl
index 5acbf2d68e153a3dd1b1f71bc02f7cba3a72b537..410055077c5fa3691cbca6ca7fa440eb16e21be7 100644 (file)
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -484,7 +484,7 @@ if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
 if (vpx_config("CONFIG_VP9_ENCODER") eq "yes") {
 if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
   add_proto qw/void vpx_fdct4x4/, "const int16_t *input, tran_low_t *output, int stride";
-  specialize qw/vpx_fdct4x4 sse2/;
+  specialize qw/vpx_fdct4x4 neon sse2/;
 
   add_proto qw/void vpx_fdct4x4_1/, "const int16_t *input, tran_low_t *output, int stride";
   specialize qw/vpx_fdct4x4_1 sse2/;
@@ -532,7 +532,7 @@ if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
   add_proto qw/void vpx_highbd_fdct32x32_1/, "const int16_t *input, tran_low_t *output, int stride";
 } else {
   add_proto qw/void vpx_fdct4x4/, "const int16_t *input, tran_low_t *output, int stride";
-  specialize qw/vpx_fdct4x4 sse2 msa/;
+  specialize qw/vpx_fdct4x4 neon sse2 msa/;
 
   add_proto qw/void vpx_fdct4x4_1/, "const int16_t *input, tran_low_t *output, int stride";
   specialize qw/vpx_fdct4x4_1 sse2/;