ppc: add vp9_iht4x4_16_add_vsx

Change-Id: Id584d8f65fdda51b8680f41424074b4b0c979622

diff --git a/test/dct_test.cc b/test/dct_test.cc
index e8ad0cd5dc306a531a9164c8b5d8f788ccb717ec..0d13e6b1465e5a1c4aa683aa07469ba9ce1be49a 100644 (file)
--- a/test/dct_test.cc
+++ b/test/dct_test.cc
@@ -683,6 +683,17 @@ INSTANTIATE_TEST_CASE_P(
                                          VPX_BITS_12)));
 #endif  // HAVE_SSE4_1 && CONFIG_VP9_HIGHBITDEPTH
 
+#if HAVE_VSX && !CONFIG_EMULATE_HARDWARE && !CONFIG_VP9_HIGHBITDEPTH
+static const FuncInfo ht_vsx_func_info[1] = {
+  { &vp9_fht4x4_c, &iht_wrapper<vp9_iht4x4_16_add_vsx>, 4, 1 }
+};
+
+INSTANTIATE_TEST_CASE_P(VSX, TransHT,
+                        ::testing::Combine(::testing::Range(0, 1),
+                                           ::testing::Values(ht_vsx_func_info),
+                                           ::testing::Range(0, 4),
+                                           ::testing::Values(VPX_BITS_8)));
+#endif  // HAVE_VSX
 #endif  // !CONFIG_EMULATE_HARDWARE
 
 /* -------------------------------------------------------------------------- */

diff --git a/vp9/common/ppc/vp9_idct_vsx.c b/vp9/common/ppc/vp9_idct_vsx.c
new file mode 100644 (file)
index 0000000..a7b66e7
--- /dev/null
+++ b/vp9/common/ppc/vp9_idct_vsx.c
@@ -0,0 +1,47 @@
+/*
+ *  Copyright (c) 2018 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 <assert.h>
+
+#include "vpx_dsp/vpx_dsp_common.h"
+#include "vpx_dsp/ppc/inv_txfm_vsx.h"
+#include "vpx_dsp/ppc/bitdepth_conversion_vsx.h"
+
+#include "vp9/common/vp9_enums.h"
+
+void vp9_iht4x4_16_add_vsx(const tran_low_t *input, uint8_t *dest, int stride,
+                           int tx_type) {
+  int16x8_t in[2], out[2];
+
+  in[0] = load_tran_low(0, input);
+  in[1] = load_tran_low(8 * sizeof(*input), input);
+
+  switch (tx_type) {
+    case DCT_DCT:
+      vpx_idct4_vsx(in, out);
+      vpx_idct4_vsx(out, in);
+      break;
+    case ADST_DCT:
+      vpx_idct4_vsx(in, out);
+      vp9_iadst4_vsx(out, in);
+      break;
+    case DCT_ADST:
+      vp9_iadst4_vsx(in, out);
+      vpx_idct4_vsx(out, in);
+      break;
+    default:
+      assert(tx_type == ADST_ADST);
+      vp9_iadst4_vsx(in, out);
+      vp9_iadst4_vsx(out, in);
+      break;
+  }
+
+  vpx_round_store4x4_vsx(in, out, dest, stride);
+}

diff --git a/vp9/common/vp9_rtcd_defs.pl b/vp9/common/vp9_rtcd_defs.pl
index 20c9ed64174d47a5af9b65f7bfb473bf4628d7c6..c5e99764a2bbbba64f43dbf12ad07cbf4fc50db6 100644 (file)
--- a/vp9/common/vp9_rtcd_defs.pl
+++ b/vp9/common/vp9_rtcd_defs.pl
@@ -67,7 +67,7 @@ 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 neon sse2/;
+  specialize qw/vp9_iht4x4_16_add neon sse2 vsx/;
   specialize qw/vp9_iht8x8_64_add neon sse2/;
   specialize qw/vp9_iht16x16_256_add neon sse2/;
   if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") ne "yes") {

diff --git a/vp9/vp9_common.mk b/vp9/vp9_common.mk
index d40d3c445d94c98dc30ecacdb7e29c13feb18137..7ca4004b0e4fd28dbf50289cf405ae8a98a1f991 100644 (file)
--- a/vp9/vp9_common.mk
+++ b/vp9/vp9_common.mk
@@ -68,6 +68,7 @@ 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-$(HAVE_MSA)   += common/mips/msa/vp9_idct16x16_msa.c
 VP9_COMMON_SRCS-$(HAVE_SSE2)  += common/x86/vp9_idct_intrin_sse2.c
+VP9_COMMON_SRCS-$(HAVE_VSX)   += common/ppc/vp9_idct_vsx.c
 VP9_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/vp9_iht4x4_add_neon.c
 VP9_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/vp9_iht8x8_add_neon.c
 VP9_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/vp9_iht16x16_add_neon.c

diff --git a/vpx_dsp/ppc/inv_txfm_vsx.c b/vpx_dsp/ppc/inv_txfm_vsx.c
index 6603b85acbad560ee3a6c71f76dd6e548b23c713..89b5d87bfe4c6ff10cfac39dfe32f906b6108782 100644 (file)
--- a/vpx_dsp/ppc/inv_txfm_vsx.c
+++ b/vpx_dsp/ppc/inv_txfm_vsx.c
@@ -14,6 +14,7 @@
 
 #include "vpx_dsp/ppc/bitdepth_conversion_vsx.h"
 #include "vpx_dsp/ppc/types_vsx.h"
+#include "vpx_dsp/ppc/inv_txfm_vsx.h"
 
 #include "./vpx_dsp_rtcd.h"
 #include "vpx_dsp/inv_txfm.h"
@@ -76,8 +77,20 @@ static int16x8_t cospi29_v = { 2404, 2404, 2404, 2404, 2404, 2404, 2404, 2404 };
 static int16x8_t cospi30_v = { 1606, 1606, 1606, 1606, 1606, 1606, 1606, 1606 };
 static int16x8_t cospi31_v = { 804, 804, 804, 804, 804, 804, 804, 804 };
 
-static uint8x16_t mask1 = { 0x0,  0x1,  0x2,  0x3,  0x4,  0x5,  0x6,  0x7,
-                            0x10, 0x11, 0x12, 0x13, 0x14, 0x15, 0x16, 0x17 };
+static const int16x8_t sinpi_1_9_v = { 5283, 5283, 5283, 5283,
+                                       5283, 5283, 5283, 5283 };
+static const int16x8_t sinpi_2_9_v = { 9929, 9929, 9929, 9929,
+                                       9929, 9929, 9929, 9929 };
+static const int16x8_t sinpi_3_9_v = { 13377, 13377, 13377, 13377,
+                                       13377, 13377, 13377, 13377 };
+static const int16x8_t sinpi_4_9_v = { 15212, 15212, 15212, 15212,
+                                       15212, 15212, 15212, 15212 };
+
+static const uint8x16_t mask1 = {
+  0x0,  0x1,  0x2,  0x3,  0x4,  0x5,  0x6,  0x7,
+  0x10, 0x11, 0x12, 0x13, 0x14, 0x15, 0x16, 0x17
+};
+
 #define ROUND_SHIFT_INIT                                               \
   const int32x4_t shift = vec_sl(vec_splat_s32(1), vec_splat_u32(13)); \
   const uint32x4_t shift14 = vec_splat_u32(14);
@@ -118,17 +131,9 @@ static uint8x16_t mask1 = { 0x0,  0x1,  0x2,  0x3,  0x4,  0x5,  0x6,  0x7,
   for (i = 0; i < 4; i++)                             \
     for (j = 0; j < 4; j++) dest[j * stride + i] = tmp_dest[j * 4 + i];
 
-void vpx_idct4x4_16_add_vsx(const tran_low_t *input, uint8_t *dest,
+void vpx_round_store4x4_vsx(int16x8_t *in, int16x8_t *out, uint8_t *dest,
                             int stride) {
   int i, j;
-  int32x4_t temp1, temp2, temp3, temp4;
-  int16x8_t step0, step1, tmp16_0, tmp16_1, t_out0, t_out1;
-  uint8x16_t mask0 = { 0x8, 0x9, 0xA, 0xB, 0xC, 0xD, 0xE, 0xF,
-                       0x0, 0x1, 0x2, 0x3, 0x4, 0x5, 0x6, 0x7 };
-  int16x8_t v0 = load_tran_low(0, input);
-  int16x8_t v1 = load_tran_low(8 * sizeof(*input), input);
-  int16x8_t t0 = vec_mergeh(v0, v1);
-  int16x8_t t1 = vec_mergel(v0, v1);
   uint8x16_t dest0 = vec_vsx_ld(0, dest);
   uint8x16_t dest1 = vec_vsx_ld(stride, dest);
   uint8x16_t dest2 = vec_vsx_ld(2 * stride, dest);
@@ -138,27 +143,45 @@ void vpx_idct4x4_16_add_vsx(const tran_low_t *input, uint8_t *dest,
   int16x8_t d_u1 = (int16x8_t)vec_mergeh(dest1, zerov);
   int16x8_t d_u2 = (int16x8_t)vec_mergeh(dest2, zerov);
   int16x8_t d_u3 = (int16x8_t)vec_mergeh(dest3, zerov);
-
+  int16x8_t tmp16_0, tmp16_1;
   uint8x16_t output_v;
   uint8_t tmp_dest[16];
-  ROUND_SHIFT_INIT
   PIXEL_ADD_INIT;
 
-  v0 = vec_mergeh(t0, t1);
-  v1 = vec_mergel(t0, t1);
+  PIXEL_ADD4(out[0], in[0]);
+  PIXEL_ADD4(out[1], in[1]);
 
-  IDCT4(v0, v1, t_out0, t_out1);
-  // transpose
-  t0 = vec_mergeh(t_out0, t_out1);
-  t1 = vec_mergel(t_out0, t_out1);
-  v0 = vec_mergeh(t0, t1);
-  v1 = vec_mergel(t0, t1);
-  IDCT4(v0, v1, t_out0, t_out1);
+  PACK_STORE(out[0], out[1]);
+}
 
-  PIXEL_ADD4(v0, t_out0);
-  PIXEL_ADD4(v1, t_out1);
+void vpx_idct4_vsx(int16x8_t *in, int16x8_t *out) {
+  int32x4_t temp1, temp2, temp3, temp4;
+  int16x8_t step0, step1, tmp16_0;
+  uint8x16_t mask0 = { 0x8, 0x9, 0xA, 0xB, 0xC, 0xD, 0xE, 0xF,
+                       0x0, 0x1, 0x2, 0x3, 0x4, 0x5, 0x6, 0x7 };
+  int16x8_t t0 = vec_mergeh(in[0], in[1]);
+  int16x8_t t1 = vec_mergel(in[0], in[1]);
+  ROUND_SHIFT_INIT
 
-  PACK_STORE(v0, v1);
+  in[0] = vec_mergeh(t0, t1);
+  in[1] = vec_mergel(t0, t1);
+
+  IDCT4(in[0], in[1], out[0], out[1]);
+}
+
+void vpx_idct4x4_16_add_vsx(const tran_low_t *input, uint8_t *dest,
+                            int stride) {
+  int16x8_t in[2], out[2];
+
+  in[0] = load_tran_low(0, input);
+  in[1] = load_tran_low(8 * sizeof(*input), input);
+  // Rows
+  vpx_idct4_vsx(in, out);
+
+  // Columns
+  vpx_idct4_vsx(out, in);
+
+  vpx_round_store4x4_vsx(in, out, dest, stride);
 }
 
 #define TRANSPOSE8x8(in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, out2, \
@@ -1130,3 +1153,50 @@ void vpx_iwht4x4_16_add_vsx(const tran_low_t *input, uint8_t *dest,
 
   PACK_STORE(v_a, v_c);
 }
+
+void vp9_iadst4_vsx(int16x8_t *in, int16x8_t *out) {
+  int16x8_t sinpi_1_3_v, sinpi_4_2_v, sinpi_2_3_v, sinpi_1_4_v, sinpi_12_n3_v;
+  int32x4_t v_v[5], u_v[4];
+  int32x4_t zerov = vec_splat_s32(0);
+  int16x8_t tmp0, tmp1;
+  int16x8_t zero16v = vec_splat_s16(0);
+  uint32x4_t shift16 = vec_sl(vec_splat_u32(8), vec_splat_u32(1));
+  ROUND_SHIFT_INIT;
+
+  sinpi_1_3_v = vec_mergel(sinpi_1_9_v, sinpi_3_9_v);
+  sinpi_4_2_v = vec_mergel(sinpi_4_9_v, sinpi_2_9_v);
+  sinpi_2_3_v = vec_mergel(sinpi_2_9_v, sinpi_3_9_v);
+  sinpi_1_4_v = vec_mergel(sinpi_1_9_v, sinpi_4_9_v);
+  sinpi_12_n3_v = vec_mergel(vec_add(sinpi_1_9_v, sinpi_2_9_v),
+                             vec_sub(zero16v, sinpi_3_9_v));
+
+  tmp0 = (int16x8_t)vec_mergeh((int32x4_t)in[0], (int32x4_t)in[1]);
+  tmp1 = (int16x8_t)vec_mergel((int32x4_t)in[0], (int32x4_t)in[1]);
+  in[0] = (int16x8_t)vec_mergeh((int32x4_t)tmp0, (int32x4_t)tmp1);
+  in[1] = (int16x8_t)vec_mergel((int32x4_t)tmp0, (int32x4_t)tmp1);
+
+  v_v[0] = vec_msum(in[0], sinpi_1_3_v, zerov);
+  v_v[1] = vec_msum(in[1], sinpi_4_2_v, zerov);
+  v_v[2] = vec_msum(in[0], sinpi_2_3_v, zerov);
+  v_v[3] = vec_msum(in[1], sinpi_1_4_v, zerov);
+  v_v[4] = vec_msum(in[0], sinpi_12_n3_v, zerov);
+
+  in[0] = vec_sub(in[0], in[1]);
+  in[1] = (int16x8_t)vec_sra((int32x4_t)in[1], shift16);
+  in[0] = vec_add(in[0], in[1]);
+  in[0] = (int16x8_t)vec_sl((int32x4_t)in[0], shift16);
+
+  u_v[0] = vec_add(v_v[0], v_v[1]);
+  u_v[1] = vec_sub(v_v[2], v_v[3]);
+  u_v[2] = vec_msum(in[0], sinpi_1_3_v, zerov);
+  u_v[3] = vec_sub(v_v[1], v_v[3]);
+  u_v[3] = vec_add(u_v[3], v_v[4]);
+
+  DCT_CONST_ROUND_SHIFT(u_v[0]);
+  DCT_CONST_ROUND_SHIFT(u_v[1]);
+  DCT_CONST_ROUND_SHIFT(u_v[2]);
+  DCT_CONST_ROUND_SHIFT(u_v[3]);
+
+  out[0] = vec_packs(u_v[0], u_v[1]);
+  out[1] = vec_packs(u_v[2], u_v[3]);
+}

diff --git a/vpx_dsp/ppc/inv_txfm_vsx.h b/vpx_dsp/ppc/inv_txfm_vsx.h
new file mode 100644 (file)
index 0000000..1b8c71d
--- /dev/null
+++ b/vpx_dsp/ppc/inv_txfm_vsx.h
@@ -0,0 +1,6 @@
+#include "vpx_dsp/ppc/types_vsx.h"
+
+void vpx_round_store4x4_vsx(int16x8_t *in, int16x8_t *out, uint8_t *dest,
+                            int stride);
+void vpx_idct4_vsx(int16x8_t *in, int16x8_t *out);
+void vp9_iadst4_vsx(int16x8_t *in, int16x8_t *out);