ref: c757d6dde49553345ed3aee576eb8086032482fb
parent: d8c277030ce6285fe37be232c6bc5dc9f96110e0
author: Alexandra Hájková <[email protected]>
date: Mon Jun 12 05:35:50 EDT 2017
ppc: Add vpx_idct8x8_64_add_vsx Change-Id: I4ed1312f365509e0595dcc09890ecb050f6f2069
--- a/test/fdct8x8_test.cc
+++ b/test/fdct8x8_test.cc
@@ -756,4 +756,11 @@
make_tuple(&vp9_fht8x8_msa, &vp9_iht8x8_64_add_msa, 2, VPX_BITS_8),
make_tuple(&vp9_fht8x8_msa, &vp9_iht8x8_64_add_msa, 3, VPX_BITS_8)));
#endif // HAVE_MSA && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
+
+#if HAVE_VSX && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
+INSTANTIATE_TEST_CASE_P(VSX, FwdTrans8x8DCT,
+ ::testing::Values(make_tuple(&vpx_fdct8x8_c,
+ &vpx_idct8x8_64_add_vsx, 0,
+ VPX_BITS_8)));
+#endif // HAVE_VSX && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
} // namespace
--- a/vpx_dsp/ppc/inv_txfm_vsx.c
+++ b/vpx_dsp/ppc/inv_txfm_vsx.c
@@ -17,11 +17,17 @@
#include "./vpx_dsp_rtcd.h"
#include "vpx_dsp/inv_txfm.h"
+static int16x8_t cospi4_v = { 16069, 16069, 16069, 16069,+ 16069, 16069, 16069, 16069 };
static int16x8_t cospi8_v = { 15137, 15137, 15137, 15137,15137, 15137, 15137, 15137 };
+static int16x8_t cospi12_v = { 13623, 13623, 13623, 13623,+ 13623, 13623, 13623, 13623 };
static int16x8_t cospi16_v = { 11585, 11585, 11585, 11585,11585, 11585, 11585, 11585 };
+static int16x8_t cospi20_v = { 9102, 9102, 9102, 9102, 9102, 9102, 9102, 9102 }; static int16x8_t cospi24_v = { 6270, 6270, 6270, 6270, 6270, 6270, 6270, 6270 };+static int16x8_t cospi28_v = { 3196, 3196, 3196, 3196, 3196, 3196, 3196, 3196 };#define ROUND_SHIFT_INIT \
const int32x4_t shift = vec_sl(vec_splat_s32(1), vec_splat_u32(13)); \
@@ -101,4 +107,171 @@
vec_vsx_st(output_v, 0, tmp_dest);
for (int i = 0; i < 4; i++)
for (int j = 0; j < 4; j++) dest[j * stride + i] = tmp_dest[j * 4 + i];
+}
+
+#define TRANSPOSE8x8(in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, out2, \
+ out3, out4, out5, out6, out7) \
+ out0 = vec_mergeh(in0, in1); \
+ out1 = vec_mergel(in0, in1); \
+ out2 = vec_mergeh(in2, in3); \
+ out3 = vec_mergel(in2, in3); \
+ out4 = vec_mergeh(in4, in5); \
+ out5 = vec_mergel(in4, in5); \
+ out6 = vec_mergeh(in6, in7); \
+ out7 = vec_mergel(in6, in7); \
+ in0 = (int16x8_t)vec_mergeh((int32x4_t)out0, (int32x4_t)out2); \
+ in1 = (int16x8_t)vec_mergel((int32x4_t)out0, (int32x4_t)out2); \
+ in2 = (int16x8_t)vec_mergeh((int32x4_t)out1, (int32x4_t)out3); \
+ in3 = (int16x8_t)vec_mergel((int32x4_t)out1, (int32x4_t)out3); \
+ in4 = (int16x8_t)vec_mergeh((int32x4_t)out4, (int32x4_t)out6); \
+ in5 = (int16x8_t)vec_mergel((int32x4_t)out4, (int32x4_t)out6); \
+ in6 = (int16x8_t)vec_mergeh((int32x4_t)out5, (int32x4_t)out7); \
+ in7 = (int16x8_t)vec_mergel((int32x4_t)out5, (int32x4_t)out7); \
+ out0 = vec_perm(in0, in4, tr8_mask0); \
+ out1 = vec_perm(in0, in4, tr8_mask1); \
+ out2 = vec_perm(in1, in5, tr8_mask0); \
+ out3 = vec_perm(in1, in5, tr8_mask1); \
+ out4 = vec_perm(in2, in6, tr8_mask0); \
+ out5 = vec_perm(in2, in6, tr8_mask1); \
+ out6 = vec_perm(in3, in7, tr8_mask0); \
+ out7 = vec_perm(in3, in7, tr8_mask1);
+
+#define STEP8_0(inpt0, inpt1, outpt0, outpt1, cospi0, cospi1) \
+ tmp16_0 = vec_mergeh(inpt0, inpt1); \
+ tmp16_1 = vec_mergel(inpt0, inpt1); \
+ temp10 = vec_sub(vec_mule(tmp16_0, cospi0), vec_mulo(tmp16_0, cospi1)); \
+ temp11 = vec_sub(vec_mule(tmp16_1, cospi0), vec_mulo(tmp16_1, cospi1)); \
+ DCT_CONST_ROUND_SHIFT(temp10); \
+ DCT_CONST_ROUND_SHIFT(temp11); \
+ outpt0 = vec_packs(temp10, temp11); \
+ temp10 = vec_add(vec_mule(tmp16_0, cospi1), vec_mulo(tmp16_0, cospi0)); \
+ temp11 = vec_add(vec_mule(tmp16_1, cospi1), vec_mulo(tmp16_1, cospi0)); \
+ DCT_CONST_ROUND_SHIFT(temp10); \
+ DCT_CONST_ROUND_SHIFT(temp11); \
+ outpt1 = vec_packs(temp10, temp11);
+
+#define STEP8_1(inpt0, inpt1, outpt0, outpt1, cospi) \
+ tmp16_2 = vec_sub(inpt0, inpt1); \
+ tmp16_3 = vec_add(inpt0, inpt1); \
+ tmp16_0 = vec_mergeh(tmp16_2, tmp16_3); \
+ tmp16_1 = vec_mergel(tmp16_2, tmp16_3); \
+ temp10 = vec_mule(tmp16_0, cospi); \
+ temp11 = vec_mule(tmp16_1, cospi); \
+ DCT_CONST_ROUND_SHIFT(temp10); \
+ DCT_CONST_ROUND_SHIFT(temp11); \
+ outpt0 = vec_packs(temp10, temp11); \
+ temp10 = vec_mulo(tmp16_0, cospi); \
+ temp11 = vec_mulo(tmp16_1, cospi); \
+ DCT_CONST_ROUND_SHIFT(temp10); \
+ DCT_CONST_ROUND_SHIFT(temp11); \
+ outpt1 = vec_packs(temp10, temp11);
+
+#define IDCT8(in0, in1, in2, in3, in4, in5, in6, in7) \
+ /* stage 1 */ \
+ step0 = in0; \
+ step2 = in4; \
+ step1 = in2; \
+ step3 = in6; \
+ \
+ STEP8_0(in1, in7, step4, step7, cospi28_v, cospi4_v); \
+ STEP8_0(in5, in3, step5, step6, cospi12_v, cospi20_v); \
+ \
+ /* stage 2 */ \
+ STEP8_1(step0, step2, in1, in0, cospi16_v); \
+ STEP8_0(step1, step3, in2, in3, cospi24_v, cospi8_v); \
+ in4 = vec_add(step4, step5); \
+ in5 = vec_sub(step4, step5); \
+ in6 = vec_sub(step7, step6); \
+ in7 = vec_add(step6, step7); \
+ \
+ /* stage 3 */ \
+ step0 = vec_add(in0, in3); \
+ step1 = vec_add(in1, in2); \
+ step2 = vec_sub(in1, in2); \
+ step3 = vec_sub(in0, in3); \
+ step4 = in4; \
+ STEP8_1(in6, in5, step5, step6, cospi16_v); \
+ step7 = in7; \
+ \
+ /* stage 4 */ \
+ in0 = vec_add(step0, step7); \
+ in1 = vec_add(step1, step6); \
+ in2 = vec_add(step2, step5); \
+ in3 = vec_add(step3, step4); \
+ in4 = vec_sub(step3, step4); \
+ in5 = vec_sub(step2, step5); \
+ in6 = vec_sub(step1, step6); \
+ in7 = vec_sub(step0, step7);
+
+#define PIXEL_ADD8(in, out) \
+ out = vec_add(vec_sra(vec_add(in, add), shift5), out);
+
+void vpx_idct8x8_64_add_vsx(const tran_low_t *input, uint8_t *dest,
+ int stride) {+ int32x4_t temp10, temp11;
+ int16x8_t step0, step1, step2, step3, step4, step5, step6, step7;
+ uint8x16_t tr8_mask0 = { 0x0, 0x1, 0x2, 0x3, 0x4, 0x5, 0x6, 0x7,+ 0x10, 0x11, 0x12, 0x13, 0x14, 0x15, 0x16, 0x17 };
+ uint8x16_t tr8_mask1 = { 0x8, 0x9, 0xA, 0xB, 0xC, 0xD, 0xE, 0xF,+ 0x18, 0x19, 0x1A, 0x1B, 0x1C, 0x1D, 0x1E, 0x1F };
+ int16x8_t tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7, tmp16_0, tmp16_1,
+ tmp16_2, tmp16_3;
+ int16x8_t src0 = vec_vsx_ld(0, input);
+ int16x8_t src1 = vec_vsx_ld(16, input);
+ int16x8_t src2 = vec_vsx_ld(2 * 16, input);
+ int16x8_t src3 = vec_vsx_ld(3 * 16, input);
+ int16x8_t src4 = vec_vsx_ld(4 * 16, input);
+ int16x8_t src5 = vec_vsx_ld(5 * 16, input);
+ int16x8_t src6 = vec_vsx_ld(6 * 16, input);
+ int16x8_t src7 = vec_vsx_ld(7 * 16, input);
+ 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);
+ uint8x16_t dest3 = vec_vsx_ld(3 * stride, dest);
+ uint8x16_t dest4 = vec_vsx_ld(4 * stride, dest);
+ uint8x16_t dest5 = vec_vsx_ld(5 * stride, dest);
+ uint8x16_t dest6 = vec_vsx_ld(6 * stride, dest);
+ uint8x16_t dest7 = vec_vsx_ld(7 * stride, dest);
+ uint8x16_t zerov = vec_splat_u8(0);
+ int16x8_t d_u0 = (int16x8_t)vec_mergeh(dest0, zerov);
+ 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 d_u4 = (int16x8_t)vec_mergeh(dest4, zerov);
+ int16x8_t d_u5 = (int16x8_t)vec_mergeh(dest5, zerov);
+ int16x8_t d_u6 = (int16x8_t)vec_mergeh(dest6, zerov);
+ int16x8_t d_u7 = (int16x8_t)vec_mergeh(dest7, zerov);
+ int16x8_t add = vec_sl(vec_splat_s16(8), vec_splat_u16(1));
+ uint16x8_t shift5 = vec_splat_u16(5);
+ uint8x16_t output0, output1, output2, output3;
+ ROUND_SHIFT_INIT;
+
+ TRANSPOSE8x8(src0, src1, src2, src3, src4, src5, src6, src7, tmp0, tmp1, tmp2,
+ tmp3, tmp4, tmp5, tmp6, tmp7);
+
+ IDCT8(tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7);
+ TRANSPOSE8x8(tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7, src0, src1, src2,
+ src3, src4, src5, src6, src7);
+ IDCT8(src0, src1, src2, src3, src4, src5, src6, src7);
+ PIXEL_ADD8(src0, d_u0);
+ PIXEL_ADD8(src1, d_u1);
+ PIXEL_ADD8(src2, d_u2);
+ PIXEL_ADD8(src3, d_u3);
+ PIXEL_ADD8(src4, d_u4);
+ PIXEL_ADD8(src5, d_u5);
+ PIXEL_ADD8(src6, d_u6);
+ PIXEL_ADD8(src7, d_u7);
+ output0 = vec_packsu(d_u0, d_u1);
+ output1 = vec_packsu(d_u2, d_u3);
+ output2 = vec_packsu(d_u4, d_u5);
+ output3 = vec_packsu(d_u6, d_u7);
+
+ vec_vsx_st(xxpermdi(output0, dest0, 1), 0, dest);
+ vec_vsx_st(xxpermdi(output0, dest1, 3), stride, dest);
+ vec_vsx_st(xxpermdi(output1, dest2, 1), 2 * stride, dest);
+ vec_vsx_st(xxpermdi(output1, dest3, 3), 3 * stride, dest);
+ vec_vsx_st(xxpermdi(output2, dest4, 1), 4 * stride, dest);
+ vec_vsx_st(xxpermdi(output2, dest5, 3), 5 * stride, dest);
+ vec_vsx_st(xxpermdi(output3, dest6, 1), 6 * stride, dest);
+ vec_vsx_st(xxpermdi(output3, dest7, 3), 7 * stride, dest);
}
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -586,7 +586,7 @@
# Note that there are more specializations appended when CONFIG_VP9_HIGHBITDEPTH is off.
specialize qw/vpx_idct4x4_16_add neon sse2 vsx/;
specialize qw/vpx_idct4x4_1_add neon sse2/;
- specialize qw/vpx_idct8x8_64_add neon sse2/;
+ specialize qw/vpx_idct8x8_64_add neon sse2 vsx/;
specialize qw/vpx_idct8x8_12_add neon sse2 ssse3/;
specialize qw/vpx_idct8x8_1_add neon sse2/;
specialize qw/vpx_idct16x16_256_add neon sse2/;