diff options
author | Shiyou Yin <yinshiyou-hf@loongson.cn> | 2017-10-12 00:33:17 +0000 |
---|---|---|
committer | Gerrit Code Review <noreply-gerritcodereview@google.com> | 2017-10-12 00:33:17 +0000 |
commit | bc4098a8e969f1ab3281a801a4ba3b78d788cec3 (patch) | |
tree | c8461ee0c1381fdaa24bbaf73be176fc78682427 | |
parent | 72c69e14ad7b5375fed113c025dca70d20fcaf0a (diff) | |
parent | e8ed2bb76258f5d0d2def256b43ed6938b6d6a45 (diff) | |
download | libvpx-bc4098a8e969f1ab3281a801a4ba3b78d788cec3.tar libvpx-bc4098a8e969f1ab3281a801a4ba3b78d788cec3.tar.gz libvpx-bc4098a8e969f1ab3281a801a4ba3b78d788cec3.tar.bz2 libvpx-bc4098a8e969f1ab3281a801a4ba3b78d788cec3.zip |
Merge "vp8: [loongson] optimize quantize with mmi"
-rw-r--r-- | test/quantize_test.cc | 8 | ||||
-rw-r--r-- | vp8/common/rtcd_defs.pl | 4 | ||||
-rw-r--r-- | vp8/encoder/mips/mmi/vp8_quantize_mmi.c | 237 | ||||
-rw-r--r-- | vp8/vp8cx.mk | 2 |
4 files changed, 249 insertions, 2 deletions
diff --git a/test/quantize_test.cc b/test/quantize_test.cc index 69da8994c..40bb2642e 100644 --- a/test/quantize_test.cc +++ b/test/quantize_test.cc @@ -200,4 +200,12 @@ INSTANTIATE_TEST_CASE_P( make_tuple(&vp8_fast_quantize_b_msa, &vp8_fast_quantize_b_c), make_tuple(&vp8_regular_quantize_b_msa, &vp8_regular_quantize_b_c))); #endif // HAVE_MSA + +#if HAVE_MMI +INSTANTIATE_TEST_CASE_P( + MMI, QuantizeTest, + ::testing::Values( + make_tuple(&vp8_fast_quantize_b_mmi, &vp8_fast_quantize_b_c), + make_tuple(&vp8_regular_quantize_b_mmi, &vp8_regular_quantize_b_c))); +#endif // HAVE_MMI } // namespace diff --git a/vp8/common/rtcd_defs.pl b/vp8/common/rtcd_defs.pl index 7a04ef17f..3bcfdc0d6 100644 --- a/vp8/common/rtcd_defs.pl +++ b/vp8/common/rtcd_defs.pl @@ -188,10 +188,10 @@ specialize qw/vp8_short_walsh4x4 sse2 neon msa/; # Quantizer # add_proto qw/void vp8_regular_quantize_b/, "struct block *, struct blockd *"; -specialize qw/vp8_regular_quantize_b sse2 sse4_1 msa/; +specialize qw/vp8_regular_quantize_b sse2 sse4_1 msa mmi/; add_proto qw/void vp8_fast_quantize_b/, "struct block *, struct blockd *"; -specialize qw/vp8_fast_quantize_b sse2 ssse3 neon msa/; +specialize qw/vp8_fast_quantize_b sse2 ssse3 neon msa mmi/; # # Block subtraction diff --git a/vp8/encoder/mips/mmi/vp8_quantize_mmi.c b/vp8/encoder/mips/mmi/vp8_quantize_mmi.c new file mode 100644 index 000000000..22b12bbab --- /dev/null +++ b/vp8/encoder/mips/mmi/vp8_quantize_mmi.c @@ -0,0 +1,237 @@ +/* + * 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 "vpx_mem/vpx_mem.h" +#include "vpx_ports/asmdefs_mmi.h" +#include "vp8/encoder/onyx_int.h" +#include "vp8/encoder/quantize.h" +#include "vp8/common/quant_common.h" + +#define REGULAR_SELECT_EOB(i, rc) \ + z = coeff_ptr[rc]; \ + sz = (z >> 31); \ + x = (z ^ sz) - sz; \ + if (x >= (zbin_ptr[rc] + *(zbin_boost_ptr++) + zbin_oq_value)) { \ + x += round_ptr[rc]; \ + y = ((((x * quant_ptr[rc]) >> 16) + x) * quant_shift_ptr[rc]) >> 16; \ + x = (y ^ sz) - sz; \ + qcoeff_ptr[rc] = x; \ + dqcoeff_ptr[rc] = x * dequant_ptr[rc]; \ + if (y) { \ + eob = i; \ + zbin_boost_ptr = b->zrun_zbin_boost; \ + } \ + } + +void vp8_fast_quantize_b_mmi(BLOCK *b, BLOCKD *d) { + const int16_t *coeff_ptr = b->coeff; + const int16_t *round_ptr = b->round; + const int16_t *quant_ptr = b->quant_fast; + int16_t *qcoeff_ptr = d->qcoeff; + int16_t *dqcoeff_ptr = d->dqcoeff; + const int16_t *dequant_ptr = d->dequant; + const int16_t *inv_zig_zag = vp8_default_inv_zig_zag; + + double ftmp[13]; + uint64_t tmp[1]; + DECLARE_ALIGNED(8, const uint64_t, ones) = { 0xffffffffffffffffULL }; + int eob = 0; + + __asm__ volatile( + // loop 0 ~ 7 + "xor %[ftmp0], %[ftmp0], %[ftmp0] \n\t" + "gsldlc1 %[ftmp1], 0x07(%[coeff_ptr]) \n\t" + "gsldrc1 %[ftmp1], 0x00(%[coeff_ptr]) \n\t" + "li %[tmp0], 0x0f \n\t" + "mtc1 %[tmp0], %[ftmp9] \n\t" + "gsldlc1 %[ftmp2], 0x0f(%[coeff_ptr]) \n\t" + "gsldrc1 %[ftmp2], 0x08(%[coeff_ptr]) \n\t" + + "psrah %[ftmp3], %[ftmp1], %[ftmp9] \n\t" + "xor %[ftmp1], %[ftmp3], %[ftmp1] \n\t" + "psubh %[ftmp1], %[ftmp1], %[ftmp3] \n\t" + "psrah %[ftmp4], %[ftmp2], %[ftmp9] \n\t" + "xor %[ftmp2], %[ftmp4], %[ftmp2] \n\t" + "psubh %[ftmp2], %[ftmp2], %[ftmp4] \n\t" + + "gsldlc1 %[ftmp5], 0x07(%[round_ptr]) \n\t" + "gsldrc1 %[ftmp5], 0x00(%[round_ptr]) \n\t" + "gsldlc1 %[ftmp6], 0x0f(%[round_ptr]) \n\t" + "gsldrc1 %[ftmp6], 0x08(%[round_ptr]) \n\t" + "paddh %[ftmp5], %[ftmp5], %[ftmp1] \n\t" + "paddh %[ftmp6], %[ftmp6], %[ftmp2] \n\t" + "gsldlc1 %[ftmp7], 0x07(%[quant_ptr]) \n\t" + "gsldrc1 %[ftmp7], 0x00(%[quant_ptr]) \n\t" + "gsldlc1 %[ftmp8], 0x0f(%[quant_ptr]) \n\t" + "gsldrc1 %[ftmp8], 0x08(%[quant_ptr]) \n\t" + "pmulhuh %[ftmp5], %[ftmp5], %[ftmp7] \n\t" + "pmulhuh %[ftmp6], %[ftmp6], %[ftmp8] \n\t" + + "xor %[ftmp7], %[ftmp5], %[ftmp3] \n\t" + "xor %[ftmp8], %[ftmp6], %[ftmp4] \n\t" + "psubh %[ftmp7], %[ftmp7], %[ftmp3] \n\t" + "psubh %[ftmp8], %[ftmp8], %[ftmp4] \n\t" + "gssdlc1 %[ftmp7], 0x07(%[qcoeff_ptr]) \n\t" + "gssdrc1 %[ftmp7], 0x00(%[qcoeff_ptr]) \n\t" + "gssdlc1 %[ftmp8], 0x0f(%[qcoeff_ptr]) \n\t" + "gssdrc1 %[ftmp8], 0x08(%[qcoeff_ptr]) \n\t" + + "gsldlc1 %[ftmp1], 0x07(%[inv_zig_zag]) \n\t" + "gsldrc1 %[ftmp1], 0x00(%[inv_zig_zag]) \n\t" + "gsldlc1 %[ftmp2], 0x0f(%[inv_zig_zag]) \n\t" + "gsldrc1 %[ftmp2], 0x08(%[inv_zig_zag]) \n\t" + "pcmpeqh %[ftmp5], %[ftmp5], %[ftmp0] \n\t" + "pcmpeqh %[ftmp6], %[ftmp6], %[ftmp0] \n\t" + "xor %[ftmp5], %[ftmp5], %[ones] \n\t" + "xor %[ftmp6], %[ftmp6], %[ones] \n\t" + "and %[ftmp5], %[ftmp5], %[ftmp1] \n\t" + "and %[ftmp6], %[ftmp6], %[ftmp2] \n\t" + "pmaxsh %[ftmp10], %[ftmp5], %[ftmp6] \n\t" + + "gsldlc1 %[ftmp5], 0x07(%[dequant_ptr]) \n\t" + "gsldrc1 %[ftmp5], 0x00(%[dequant_ptr]) \n\t" + "gsldlc1 %[ftmp6], 0x0f(%[dequant_ptr]) \n\t" + "gsldrc1 %[ftmp6], 0x08(%[dequant_ptr]) \n\t" + "pmullh %[ftmp5], %[ftmp5], %[ftmp7] \n\t" + "pmullh %[ftmp6], %[ftmp6], %[ftmp8] \n\t" + "gssdlc1 %[ftmp5], 0x07(%[dqcoeff_ptr]) \n\t" + "gssdrc1 %[ftmp5], 0x00(%[dqcoeff_ptr]) \n\t" + "gssdlc1 %[ftmp6], 0x0f(%[dqcoeff_ptr]) \n\t" + "gssdrc1 %[ftmp6], 0x08(%[dqcoeff_ptr]) \n\t" + + // loop 8 ~ 15 + "gsldlc1 %[ftmp1], 0x17(%[coeff_ptr]) \n\t" + "gsldrc1 %[ftmp1], 0x10(%[coeff_ptr]) \n\t" + "gsldlc1 %[ftmp2], 0x1f(%[coeff_ptr]) \n\t" + "gsldrc1 %[ftmp2], 0x18(%[coeff_ptr]) \n\t" + + "psrah %[ftmp3], %[ftmp1], %[ftmp9] \n\t" + "xor %[ftmp1], %[ftmp3], %[ftmp1] \n\t" + "psubh %[ftmp1], %[ftmp1], %[ftmp3] \n\t" + "psrah %[ftmp4], %[ftmp2], %[ftmp9] \n\t" + "xor %[ftmp2], %[ftmp4], %[ftmp2] \n\t" + "psubh %[ftmp2], %[ftmp2], %[ftmp4] \n\t" + + "gsldlc1 %[ftmp5], 0x17(%[round_ptr]) \n\t" + "gsldrc1 %[ftmp5], 0x10(%[round_ptr]) \n\t" + "gsldlc1 %[ftmp6], 0x1f(%[round_ptr]) \n\t" + "gsldrc1 %[ftmp6], 0x18(%[round_ptr]) \n\t" + "paddh %[ftmp5], %[ftmp5], %[ftmp1] \n\t" + "paddh %[ftmp6], %[ftmp6], %[ftmp2] \n\t" + "gsldlc1 %[ftmp7], 0x17(%[quant_ptr]) \n\t" + "gsldrc1 %[ftmp7], 0x10(%[quant_ptr]) \n\t" + "gsldlc1 %[ftmp8], 0x1f(%[quant_ptr]) \n\t" + "gsldrc1 %[ftmp8], 0x18(%[quant_ptr]) \n\t" + "pmulhuh %[ftmp5], %[ftmp5], %[ftmp7] \n\t" + "pmulhuh %[ftmp6], %[ftmp6], %[ftmp8] \n\t" + + "xor %[ftmp7], %[ftmp5], %[ftmp3] \n\t" + "xor %[ftmp8], %[ftmp6], %[ftmp4] \n\t" + "psubh %[ftmp7], %[ftmp7], %[ftmp3] \n\t" + "psubh %[ftmp8], %[ftmp8], %[ftmp4] \n\t" + "gssdlc1 %[ftmp7], 0x17(%[qcoeff_ptr]) \n\t" + "gssdrc1 %[ftmp7], 0x10(%[qcoeff_ptr]) \n\t" + "gssdlc1 %[ftmp8], 0x1f(%[qcoeff_ptr]) \n\t" + "gssdrc1 %[ftmp8], 0x18(%[qcoeff_ptr]) \n\t" + + "gsldlc1 %[ftmp1], 0x17(%[inv_zig_zag]) \n\t" + "gsldrc1 %[ftmp1], 0x10(%[inv_zig_zag]) \n\t" + "gsldlc1 %[ftmp2], 0x1f(%[inv_zig_zag]) \n\t" + "gsldrc1 %[ftmp2], 0x18(%[inv_zig_zag]) \n\t" + "pcmpeqh %[ftmp5], %[ftmp5], %[ftmp0] \n\t" + "pcmpeqh %[ftmp6], %[ftmp6], %[ftmp0] \n\t" + "xor %[ftmp5], %[ftmp5], %[ones] \n\t" + "xor %[ftmp6], %[ftmp6], %[ones] \n\t" + "and %[ftmp5], %[ftmp5], %[ftmp1] \n\t" + "and %[ftmp6], %[ftmp6], %[ftmp2] \n\t" + "pmaxsh %[ftmp11], %[ftmp5], %[ftmp6] \n\t" + + "gsldlc1 %[ftmp5], 0x17(%[dequant_ptr]) \n\t" + "gsldrc1 %[ftmp5], 0x10(%[dequant_ptr]) \n\t" + "gsldlc1 %[ftmp6], 0x1f(%[dequant_ptr]) \n\t" + "gsldrc1 %[ftmp6], 0x18(%[dequant_ptr]) \n\t" + "pmullh %[ftmp5], %[ftmp5], %[ftmp7] \n\t" + "pmullh %[ftmp6], %[ftmp6], %[ftmp8] \n\t" + "gssdlc1 %[ftmp5], 0x17(%[dqcoeff_ptr]) \n\t" + "gssdrc1 %[ftmp5], 0x10(%[dqcoeff_ptr]) \n\t" + "gssdlc1 %[ftmp6], 0x1f(%[dqcoeff_ptr]) \n\t" + "gssdrc1 %[ftmp6], 0x18(%[dqcoeff_ptr]) \n\t" + + "li %[tmp0], 0x10 \n\t" + "mtc1 %[tmp0], %[ftmp9] \n\t" + + "pmaxsh %[ftmp10], %[ftmp10], %[ftmp11] \n\t" + "psrlw %[ftmp11], %[ftmp10], %[ftmp9] \n\t" + "pmaxsh %[ftmp10], %[ftmp10], %[ftmp11] \n\t" + "li %[tmp0], 0xaa \n\t" + "mtc1 %[tmp0], %[ftmp9] \n\t" + "pshufh %[ftmp11], %[ftmp10], %[ftmp9] \n\t" + "pmaxsh %[ftmp10], %[ftmp10], %[ftmp11] \n\t" + "li %[tmp0], 0xffff \n\t" + "mtc1 %[tmp0], %[ftmp9] \n\t" + "and %[ftmp10], %[ftmp10], %[ftmp9] \n\t" + "gssdlc1 %[ftmp10], 0x07(%[eob]) \n\t" + "gssdrc1 %[ftmp10], 0x00(%[eob]) \n\t" + : [ftmp0] "=&f"(ftmp[0]), [ftmp1] "=&f"(ftmp[1]), [ftmp2] "=&f"(ftmp[2]), + [ftmp3] "=&f"(ftmp[3]), [ftmp4] "=&f"(ftmp[4]), [ftmp5] "=&f"(ftmp[5]), + [ftmp6] "=&f"(ftmp[6]), [ftmp7] "=&f"(ftmp[7]), [ftmp8] "=&f"(ftmp[8]), + [ftmp9] "=&f"(ftmp[9]), [ftmp10] "=&f"(ftmp[10]), + [ftmp11] "=&f"(ftmp[11]), [ftmp12] "=&f"(ftmp[12]), [tmp0] "=&r"(tmp[0]) + : [coeff_ptr] "r"((mips_reg)coeff_ptr), + [qcoeff_ptr] "r"((mips_reg)qcoeff_ptr), + [dequant_ptr] "r"((mips_reg)dequant_ptr), + [round_ptr] "r"((mips_reg)round_ptr), + [quant_ptr] "r"((mips_reg)quant_ptr), + [dqcoeff_ptr] "r"((mips_reg)dqcoeff_ptr), + [inv_zig_zag] "r"((mips_reg)inv_zig_zag), [eob] "r"((mips_reg)&eob), + [ones] "f"(ones) + : "memory"); + + *d->eob = eob; +} + +void vp8_regular_quantize_b_mmi(BLOCK *b, BLOCKD *d) { + int eob; + int x, y, z, sz; + const int16_t *zbin_boost_ptr = b->zrun_zbin_boost; + const int16_t *coeff_ptr = b->coeff; + const int16_t *zbin_ptr = b->zbin; + const int16_t *round_ptr = b->round; + const int16_t *quant_ptr = b->quant; + const int16_t *quant_shift_ptr = b->quant_shift; + int16_t *qcoeff_ptr = d->qcoeff; + int16_t *dqcoeff_ptr = d->dqcoeff; + const int16_t *dequant_ptr = d->dequant; + const int16_t zbin_oq_value = b->zbin_extra; + + memset(qcoeff_ptr, 0, 32); + memset(dqcoeff_ptr, 0, 32); + + eob = -1; + + REGULAR_SELECT_EOB(0, 0); + REGULAR_SELECT_EOB(1, 1); + REGULAR_SELECT_EOB(2, 4); + REGULAR_SELECT_EOB(3, 8); + REGULAR_SELECT_EOB(4, 5); + REGULAR_SELECT_EOB(5, 2); + REGULAR_SELECT_EOB(6, 3); + REGULAR_SELECT_EOB(7, 6); + REGULAR_SELECT_EOB(8, 9); + REGULAR_SELECT_EOB(9, 12); + REGULAR_SELECT_EOB(10, 13); + REGULAR_SELECT_EOB(11, 10); + REGULAR_SELECT_EOB(12, 7); + REGULAR_SELECT_EOB(13, 11); + REGULAR_SELECT_EOB(14, 14); + REGULAR_SELECT_EOB(15, 15); + + *d->eob = (char)(eob + 1); +} diff --git a/vp8/vp8cx.mk b/vp8/vp8cx.mk index 9111a2257..23d65d416 100644 --- a/vp8/vp8cx.mk +++ b/vp8/vp8cx.mk @@ -110,6 +110,8 @@ VP8_CX_SRCS-$(HAVE_MSA) += encoder/mips/msa/encodeopt_msa.c VP8_CX_SRCS-$(HAVE_MSA) += encoder/mips/msa/quantize_msa.c VP8_CX_SRCS-$(HAVE_MSA) += encoder/mips/msa/temporal_filter_msa.c +VP8_CX_SRCS-$(HAVE_MMI) += encoder/mips/mmi/vp8_quantize_mmi.c + ifeq ($(CONFIG_TEMPORAL_DENOISING),yes) VP8_CX_SRCS-$(HAVE_MSA) += encoder/mips/msa/denoising_msa.c endif |