ref: 56aa0da405b072447bf8165650983fdb36d272a5
parent: 0c2a14f9e24fda448161bbaf13878b202ea57f1f
author: Parag Salasakar <img.mips1@gmail.com>
date: Thu Jul 30 06:56:40 EDT 2015
mips msa vp8 quantize optimization average improvement ~2x-3x Change-Id: I6fc37191bf9cb5a67e1af9787d0d27659c17bdba
--- a/test/quantize_test.cc
+++ b/test/quantize_test.cc
@@ -192,4 +192,12 @@
::testing::Values(make_tuple(&vp8_fast_quantize_b_neon,
&vp8_fast_quantize_b_c)));
#endif // HAVE_NEON
+
+#if HAVE_MSA
+INSTANTIATE_TEST_CASE_P(
+ MSA, QuantizeTest,
+ ::testing::Values(
+ 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
} // namespace
--- a/vp8/common/mips/msa/vp8_macros_msa.h
+++ b/vp8/common/mips/msa/vp8_macros_msa.h
@@ -553,6 +553,20 @@
}
#define VSHF_B3_SB(...) VSHF_B3(v16i8, __VA_ARGS__)
+/* Description : Shuffle halfword vector elements as per mask vector
+ Arguments : Inputs - in0, in1, in2, in3, mask0, mask1
+ Outputs - out0, out1
+ Return Type - as per RTYPE
+ Details : halfword elements from 'in0' & 'in1' are copied selectively to
+ 'out0' as per control vector 'mask0'
+*/
+#define VSHF_H2(RTYPE, in0, in1, in2, in3, mask0, mask1, out0, out1) \
+{ \+ out0 = (RTYPE)__msa_vshf_h((v8i16)mask0, (v8i16)in1, (v8i16)in0); \
+ out1 = (RTYPE)__msa_vshf_h((v8i16)mask1, (v8i16)in3, (v8i16)in2); \
+}
+#define VSHF_H2_SH(...) VSHF_H2(v8i16, __VA_ARGS__)
+
/* Description : Dot product of byte vector elements
Arguments : Inputs - mult0, mult1, cnst0, cnst1
Outputs - out0, out1
@@ -604,6 +618,31 @@
}
#define DOTP_SB4_SH(...) DOTP_SB4(v8i16, __VA_ARGS__)
+/* Description : Dot product of halfword vector elements
+ Arguments : Inputs - mult0, mult1, cnst0, cnst1
+ Outputs - out0, out1
+ Return Type - as per RTYPE
+ Details : Signed halfword elements from 'mult0' are multiplied with
+ signed halfword elements from 'cnst0' producing a result
+ twice the size of input i.e. signed word.
+ The multiplication result of adjacent odd-even elements
+ are added together and written to the 'out0' vector
+*/
+#define DOTP_SH2(RTYPE, mult0, mult1, cnst0, cnst1, out0, out1) \
+{ \+ out0 = (RTYPE)__msa_dotp_s_w((v8i16)mult0, (v8i16)cnst0); \
+ out1 = (RTYPE)__msa_dotp_s_w((v8i16)mult1, (v8i16)cnst1); \
+}
+
+#define DOTP_SH4(RTYPE, mult0, mult1, mult2, mult3, \
+ cnst0, cnst1, cnst2, cnst3, \
+ out0, out1, out2, out3) \
+{ \+ DOTP_SH2(RTYPE, mult0, mult1, cnst0, cnst1, out0, out1); \
+ DOTP_SH2(RTYPE, mult2, mult3, cnst2, cnst3, out2, out3); \
+}
+#define DOTP_SH4_SW(...) DOTP_SH4(v4i32, __VA_ARGS__)
+
/* Description : Dot product & addition of byte vector elements
Arguments : Inputs - mult0, mult1, cnst0, cnst1
Outputs - out0, out1
@@ -1307,6 +1346,18 @@
{ \ADD2(in0, in1, in2, in3, out0, out1); \
ADD2(in4, in5, in6, in7, out2, out3); \
+}
+
+/* Description : Subtraction of 2 pairs of vectors
+ Arguments : Inputs - in0, in1, in2, in3
+ Outputs - out0, out1
+ Details : Each element in 'in1' is subtracted from 'in0' and result is
+ written to 'out0'.
+*/
+#define SUB2(in0, in1, in2, in3, out0, out1) \
+{ \+ out0 = in0 - in1; \
+ out1 = in2 - in3; \
}
/* Description : Sign extend halfword elements from right half of the vector
--- a/vp8/common/rtcd_defs.pl
+++ b/vp8/common/rtcd_defs.pl
@@ -286,10 +286,10 @@
# Quantizer
#
add_proto qw/void vp8_regular_quantize_b/, "struct block *, struct blockd *";
-specialize qw/vp8_regular_quantize_b sse2 sse4_1/;
+specialize qw/vp8_regular_quantize_b sse2 sse4_1 msa/;
add_proto qw/void vp8_fast_quantize_b/, "struct block *, struct blockd *";
-specialize qw/vp8_fast_quantize_b sse2 ssse3 neon/;
+specialize qw/vp8_fast_quantize_b sse2 ssse3 neon msa/;
#
# Block subtraction
--- /dev/null
+++ b/vp8/encoder/mips/msa/quantize_msa.c
@@ -1,0 +1,246 @@
+/*
+ * Copyright (c) 2015 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 "./vp8_rtcd.h"
+#include "vp8/common/mips/msa/vp8_macros_msa.h"
+#include "vp8/encoder/block.h"
+
+static int8_t fast_quantize_b_msa(int16_t *coeff_ptr, int16_t *zbin,
+ int16_t *round, int16_t *quant,
+ int16_t *de_quant, int16_t *q_coeff,
+ int16_t *dq_coeff)
+{+ int32_t cnt, eob;
+ v16i8 inv_zig_zag = { 0, 1, 5, 6, 2, 4, 7, 12,+ 3, 8, 11, 13, 9, 10, 14, 15 };
+ v8i16 round0, round1;
+ v8i16 sign_z0, sign_z1;
+ v8i16 q_coeff0, q_coeff1;
+ v8i16 x0, x1, de_quant0, de_quant1;
+ v8i16 coeff0, coeff1, z0, z1;
+ v8i16 quant0, quant1, quant2, quant3;
+ v8i16 zero = { 0 };+ v8i16 inv_zig_zag0, inv_zig_zag1;
+ v8i16 zigzag_mask0 = { 0, 1, 4, 8, 5, 2, 3, 6 };+ v8i16 zigzag_mask1 = { 9, 12, 13, 10, 7, 11, 14, 15 };+ v8i16 temp0_h, temp1_h, temp2_h, temp3_h;
+ v4i32 temp0_w, temp1_w, temp2_w, temp3_w;
+
+ ILVRL_B2_SH(zero, inv_zig_zag, inv_zig_zag0, inv_zig_zag1);
+ eob = -1;
+ LD_SH2(coeff_ptr, 8, coeff0, coeff1);
+ VSHF_H2_SH(coeff0, coeff1, coeff0, coeff1, zigzag_mask0, zigzag_mask1,
+ z0, z1);
+ LD_SH2(round, 8, coeff0, coeff1);
+ VSHF_H2_SH(coeff0, coeff1, coeff0, coeff1, zigzag_mask0, zigzag_mask1,
+ round0, round1);
+ LD_SH2(quant, 8, coeff0, coeff1);
+ VSHF_H2_SH(coeff0, coeff1, coeff0, coeff1, zigzag_mask0, zigzag_mask1,
+ quant0, quant2);
+ sign_z0 = z0 >> 15;
+ sign_z1 = z1 >> 15;
+ x0 = __msa_add_a_h(z0, zero);
+ x1 = __msa_add_a_h(z1, zero);
+ ILVL_H2_SH(quant0, quant0, quant2, quant2, quant1, quant3);
+ ILVR_H2_SH(quant0, quant0, quant2, quant2, quant0, quant2);
+ ILVL_H2_SH(round0, x0, round1, x1, temp1_h, temp3_h);
+ ILVR_H2_SH(round0, x0, round1, x1, temp0_h, temp2_h);
+ DOTP_SH4_SW(temp0_h, temp1_h, temp2_h, temp3_h, quant0, quant1, quant2,
+ quant3, temp0_w, temp1_w, temp2_w, temp3_w);
+ SRA_4V(temp0_w, temp1_w, temp2_w, temp3_w, 16);
+ PCKEV_H2_SH(temp1_w, temp0_w, temp3_w, temp2_w, x0, x1);
+ x0 = x0 ^ sign_z0;
+ x1 = x1 ^ sign_z1;
+ SUB2(x0, sign_z0, x1, sign_z1, x0, x1);
+ VSHF_H2_SH(x0, x1, x0, x1, inv_zig_zag0, inv_zig_zag1, q_coeff0, q_coeff1);
+ ST_SH2(q_coeff0, q_coeff1, q_coeff, 8);
+ LD_SH2(de_quant, 8, de_quant0, de_quant1);
+ q_coeff0 *= de_quant0;
+ q_coeff1 *= de_quant1;
+ ST_SH2(q_coeff0, q_coeff1, dq_coeff, 8);
+
+ for (cnt = 0; cnt < 16; ++cnt)
+ {+ if ((cnt <= 7) && (x1[7 - cnt] != 0))
+ {+ eob = (15 - cnt);
+ break;
+ }
+
+ if ((cnt > 7) && (x0[7 - (cnt - 8)] != 0))
+ {+ eob = (7 - (cnt - 8));
+ break;
+ }
+ }
+
+ return (int8_t)(eob + 1);
+}
+
+static int8_t exact_regular_quantize_b_msa(int16_t *zbin_boost,
+ int16_t *coeff_ptr,
+ int16_t *zbin,
+ int16_t *round,
+ int16_t *quant,
+ int16_t *quant_shift,
+ int16_t *de_quant,
+ int16_t zbin_oq_in,
+ int16_t *q_coeff,
+ int16_t *dq_coeff)
+{+ int32_t cnt, eob;
+ int16_t *boost_temp = zbin_boost;
+ v16i8 inv_zig_zag = { 0, 1, 5, 6, 2, 4, 7, 12,+ 3, 8, 11, 13, 9, 10, 14, 15 };
+ v8i16 round0, round1;
+ v8i16 sign_z0, sign_z1;
+ v8i16 q_coeff0, q_coeff1;
+ v8i16 z_bin0, z_bin1, zbin_o_q;
+ v8i16 x0, x1, sign_x0, sign_x1, de_quant0, de_quant1;
+ v8i16 coeff0, coeff1, z0, z1;
+ v8i16 quant0, quant1, quant2, quant3;
+ v8i16 zero = { 0 };+ v8i16 inv_zig_zag0, inv_zig_zag1;
+ v8i16 zigzag_mask0 = { 0, 1, 4, 8, 5, 2, 3, 6 };+ v8i16 zigzag_mask1 = { 9, 12, 13, 10, 7, 11, 14, 15 };+ v8i16 temp0_h, temp1_h, temp2_h, temp3_h;
+ v4i32 temp0_w, temp1_w, temp2_w, temp3_w;
+
+ ILVRL_B2_SH(zero, inv_zig_zag, inv_zig_zag0, inv_zig_zag1);
+ zbin_o_q = __msa_fill_h(zbin_oq_in);
+ eob = -1;
+ LD_SH2(coeff_ptr, 8, coeff0, coeff1);
+ VSHF_H2_SH(coeff0, coeff1, coeff0, coeff1, zigzag_mask0, zigzag_mask1,
+ z0, z1);
+ LD_SH2(round, 8, coeff0, coeff1);
+ VSHF_H2_SH(coeff0, coeff1, coeff0, coeff1, zigzag_mask0, zigzag_mask1,
+ round0, round1);
+ LD_SH2(quant, 8, coeff0, coeff1);
+ VSHF_H2_SH(coeff0, coeff1, coeff0, coeff1, zigzag_mask0, zigzag_mask1,
+ quant0, quant2);
+ LD_SH2(zbin, 8, coeff0, coeff1);
+ VSHF_H2_SH(coeff0, coeff1, coeff0, coeff1, zigzag_mask0, zigzag_mask1,
+ z_bin0, z_bin1);
+ sign_z0 = z0 >> 15;
+ sign_z1 = z1 >> 15;
+ x0 = __msa_add_a_h(z0, zero);
+ x1 = __msa_add_a_h(z1, zero);
+ SUB2(x0, z_bin0, x1, z_bin1, z_bin0, z_bin1);
+ SUB2(z_bin0, zbin_o_q, z_bin1, zbin_o_q, z_bin0, z_bin1);
+ ILVL_H2_SH(quant0, quant0, quant2, quant2, quant1, quant3);
+ ILVR_H2_SH(quant0, quant0, quant2, quant2, quant0, quant2);
+ ILVL_H2_SH(round0, x0, round1, x1, temp1_h, temp3_h);
+ ILVR_H2_SH(round0, x0, round1, x1, temp0_h, temp2_h);
+ DOTP_SH4_SW(temp0_h, temp1_h, temp2_h, temp3_h, quant0, quant1, quant2,
+ quant3, temp0_w, temp1_w, temp2_w, temp3_w);
+ SRA_4V(temp0_w, temp1_w, temp2_w, temp3_w, 16);
+ PCKEV_H2_SH(temp1_w, temp0_w, temp3_w, temp2_w, temp0_h, temp2_h);
+ LD_SH2(quant_shift, 8, coeff0, coeff1);
+ VSHF_H2_SH(coeff0, coeff1, coeff0, coeff1, zigzag_mask0, zigzag_mask1,
+ quant0, quant2);
+ ILVL_H2_SH(quant0, quant0, quant2, quant2, quant1, quant3);
+ ILVR_H2_SH(quant0, quant0, quant2, quant2, quant0, quant2);
+ ADD2(x0, round0, x1, round1, x0, x1);
+ ILVL_H2_SH(temp0_h, x0, temp2_h, x1, temp1_h, temp3_h);
+ ILVR_H2_SH(temp0_h, x0, temp2_h, x1, temp0_h, temp2_h);
+ DOTP_SH4_SW(temp0_h, temp1_h, temp2_h, temp3_h, quant0, quant1, quant2,
+ quant3, temp0_w, temp1_w, temp2_w, temp3_w);
+ SRA_4V(temp0_w, temp1_w, temp2_w, temp3_w, 16);
+ PCKEV_H2_SH(temp1_w, temp0_w, temp3_w, temp2_w, x0, x1);
+ sign_x0 = x0 ^ sign_z0;
+ sign_x1 = x1 ^ sign_z1;
+ SUB2(sign_x0, sign_z0, sign_x1, sign_z1, sign_x0, sign_x1);
+ for (cnt = 0; cnt < 16; ++cnt)
+ {+ if (cnt <= 7)
+ {+ if (boost_temp[0] <= z_bin0[cnt])
+ {+ if (x0[cnt])
+ {+ eob = cnt;
+ boost_temp = zbin_boost;
+ }
+ else
+ {+ boost_temp++;
+ }
+ }
+ else
+ {+ sign_x0[cnt] = 0;
+ boost_temp++;
+ }
+ }
+ else
+ {+ if (boost_temp[0] <= z_bin1[cnt - 8])
+ {+ if (x1[cnt - 8])
+ {+ eob = cnt;
+ boost_temp = zbin_boost;
+ }
+ else
+ {+ boost_temp++;
+ }
+ }
+ else
+ {+ sign_x1[cnt - 8] = 0;
+ boost_temp++;
+ }
+ }
+ }
+
+ VSHF_H2_SH(sign_x0, sign_x1, sign_x0, sign_x1, inv_zig_zag0, inv_zig_zag1,
+ q_coeff0, q_coeff1);
+ ST_SH2(q_coeff0, q_coeff1, q_coeff, 8);
+ LD_SH2(de_quant, 8, de_quant0, de_quant1);
+ MUL2(de_quant0, q_coeff0, de_quant1, q_coeff1, de_quant0, de_quant1);
+ ST_SH2(de_quant0, de_quant1, dq_coeff, 8);
+
+ return (int8_t)(eob + 1);
+}
+
+void vp8_fast_quantize_b_msa(BLOCK *b, BLOCKD *d)
+{+ int16_t *coeff_ptr = b->coeff;
+ int16_t *zbin_ptr = b->zbin;
+ int16_t *round_ptr = b->round;
+ int16_t *quant_ptr = b->quant_fast;
+ int16_t *qcoeff_ptr = d->qcoeff;
+ int16_t *dqcoeff_ptr = d->dqcoeff;
+ int16_t *dequant_ptr = d->dequant;
+
+ *d->eob = fast_quantize_b_msa(coeff_ptr, zbin_ptr, round_ptr, quant_ptr,
+ dequant_ptr, qcoeff_ptr, dqcoeff_ptr);
+}
+
+void vp8_regular_quantize_b_msa(BLOCK *b, BLOCKD *d)
+{+ int16_t *zbin_boost_ptr = b->zrun_zbin_boost;
+ int16_t *coeff_ptr = b->coeff;
+ int16_t *zbin_ptr = b->zbin;
+ int16_t *round_ptr = b->round;
+ int16_t *quant_ptr = b->quant;
+ int16_t *quant_shift_ptr = b->quant_shift;
+ int16_t *qcoeff_ptr = d->qcoeff;
+ int16_t *dqcoeff_ptr = d->dqcoeff;
+ int16_t *dequant_ptr = d->dequant;
+ int16_t zbin_oq_value = b->zbin_extra;
+
+ *d->eob = exact_regular_quantize_b_msa(zbin_boost_ptr, coeff_ptr,
+ zbin_ptr, round_ptr,
+ quant_ptr, quant_shift_ptr,
+ dequant_ptr, zbin_oq_value,
+ qcoeff_ptr, dqcoeff_ptr);
+}
--- a/vp8/vp8cx.mk
+++ b/vp8/vp8cx.mk
@@ -104,5 +104,6 @@
endif
VP8_CX_SRCS-$(HAVE_MSA) += encoder/mips/msa/dct_msa.c
+VP8_CX_SRCS-$(HAVE_MSA) += encoder/mips/msa/quantize_msa.c
VP8_CX_SRCS-yes := $(filter-out $(VP8_CX_SRCS_REMOVE-yes),$(VP8_CX_SRCS-yes))