Skip to content

GitLab

A
aom-rav1e
Commit 89b4b315 authored by Parag Salasakar's avatar Parag Salasakar
Browse files
Options

mips msa vp9 fdct 16x16 optimization 

average improvement ~4x-6x

Change-Id: Id3b2243e5b3c7844c90c4231a5e75fa69911362c
parent a4bb5f2a
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 1032 additions and 15 deletions
test/dct16x16_test.cc
View file @ 89b4b315
......@@ -933,14 +933,15 @@ INSTANTIATE_TEST_CASE_P(
INSTANTIATE_TEST_CASE_P(
MSA, Trans16x16DCT,
::testing::Values(
make_tuple(&vp9_fdct16x16_c,
make_tuple(&vp9_fdct16x16_msa,
&vp9_idct16x16_256_add_msa, 0, VPX_BITS_8)));
INSTANTIATE_TEST_CASE_P(
MSA, Trans16x16HT,
::testing::Values(
make_tuple(&vp9_fht16x16_c, &vp9_iht16x16_256_add_msa, 0, VPX_BITS_8),
make_tuple(&vp9_fht16x16_c, &vp9_iht16x16_256_add_msa, 1, VPX_BITS_8),
make_tuple(&vp9_fht16x16_c, &vp9_iht16x16_256_add_msa, 2, VPX_BITS_8),
make_tuple(&vp9_fht16x16_c, &vp9_iht16x16_256_add_msa, 3, VPX_BITS_8)));
make_tuple(&vp9_fht16x16_msa, &vp9_iht16x16_256_add_msa, 0, VPX_BITS_8),
make_tuple(&vp9_fht16x16_msa, &vp9_iht16x16_256_add_msa, 1, VPX_BITS_8),
make_tuple(&vp9_fht16x16_msa, &vp9_iht16x16_256_add_msa, 2, VPX_BITS_8),
make_tuple(&vp9_fht16x16_msa, &vp9_iht16x16_256_add_msa, 3,
VPX_BITS_8)));
#endif // HAVE_MSA && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
} // namespace
vp9/common/mips/msa/vp9_macros_msa.h
View file @ 89b4b315
......@@ -380,6 +380,17 @@
out3 = (v8i16)__msa_ilvl_d((v2i64)out2, (v2i64)out2); \
}
/* Description : Load 2 vectors of signed word elements with stride
Arguments : Inputs - psrc (source pointer to load from)
- stride
Outputs - out0, out1
Return Type - signed word
*/
#define LD_SW2(psrc, stride, out0, out1) { \
out0 = LD_SW((psrc)); \
out1 = LD_SW((psrc) + stride); \
}
/* Description : Store vectors of 16 byte elements with stride
Arguments : Inputs - in0, in1, stride
Outputs - pdst (destination pointer to store to)
......@@ -777,6 +788,24 @@
CLIP_SH2_0_255(in2, in3); \
}
/* Description : Addition of 4 signed word elements
4 signed word elements of input vector are added together and
the resulting integer sum is returned
Arguments : Inputs - in (signed word vector)
Outputs - sum_m (i32 sum)
Return Type - signed word
*/
#define HADD_SW_S32(in) ({ \
v2i64 res0_m, res1_m; \
int32_t sum_m; \
\
res0_m = __msa_hadd_s_d((v4i32)in, (v4i32)in); \
res1_m = __msa_splati_d(res0_m, 1); \
res0_m = res0_m + res1_m; \
sum_m = __msa_copy_s_w((v4i32)res0_m, 0); \
sum_m; \
})
/* Description : Horizontal addition of unsigned byte vector elements
Arguments : Inputs - in0, in1
Outputs - out0, out1
......@@ -1073,8 +1102,8 @@
Outputs - in0, in1, in2, in3 (in place)
Return Type - unsigned halfword
Details : Each unsigned halfword element from 'in0' is saturated to the
value generated with (sat_val+1) bit range
Results are in placed to original vectors
value generated with (sat_val+1) bit range.
The results are in placed to original vectors
*/
#define SAT_UH2(RTYPE, in0, in1, sat_val) { \
in0 = (RTYPE)__msa_sat_u_h((v8u16)in0, sat_val); \
......@@ -1096,7 +1125,7 @@
Return Type - unsigned halfword
Details : Each unsigned halfword element from 'in0' is saturated to the
value generated with (sat_val+1) bit range
Results are in placed to original vectors
The results are in placed to original vectors
*/
#define SAT_SH2(RTYPE, in0, in1, sat_val) { \
in0 = (RTYPE)__msa_sat_s_h((v8i16)in0, sat_val); \
......@@ -1216,10 +1245,10 @@
Outputs - in0, in1 (in-place)
Return Type - as per RTYPE
Details : Each unsigned byte element from input vector 'in0' is
logically xor'ed with 128 and result is in-place stored in
logically xor'ed with 128 and the result is in-place stored in
'in0' vector
Each unsigned byte element from input vector 'in1' is
logically xor'ed with 128 and result is in-place stored in
logically xor'ed with 128 and the result is in-place stored in
'in1' vector
Similar for other pairs
*/
......@@ -1271,13 +1300,28 @@
}
#define ADDS_SH4_SH(...) ADDS_SH4(v8i16, __VA_ARGS__)
/* Description : Shift left all elements of vector (generic for all data types)
Arguments : Inputs - in0, in1, in2, in3, shift
Outputs - in0, in1, in2, in3 (in place)
Return Type - as per input vector RTYPE
Details : Each element of vector 'in0' is left shifted by 'shift' and
the result is in place written to 'in0'
Similar for other pairs
*/
#define SLLI_4V(in0, in1, in2, in3, shift) { \
in0 = in0 << shift; \
in1 = in1 << shift; \
in2 = in2 << shift; \
in3 = in3 << shift; \
}
/* Description : Arithmetic shift right all elements of vector
(generic for all data types)
Arguments : Inputs - in0, in1, in2, in3, shift
Outputs - in0, in1, in2, in3 (in place)
Return Type - as per input vector RTYPE
Details : Each element of vector 'in0' is right shifted by 'shift' and
result is in place written to 'in0'
the result is in place written to 'in0'
Here, 'shift' is GP variable passed in
Similar for other pairs
*/
......@@ -1427,6 +1471,34 @@
out7 = in0 - in7; \
}
/* Description : Butterfly of 16 input vectors
Arguments : Inputs - in0 ... in15
Outputs - out0 .. out15
Details : Butterfly operation
*/
#define BUTTERFLY_16(in0, in1, in2, in3, in4, in5, in6, in7, \
in8, in9, in10, in11, in12, in13, in14, in15, \
out0, out1, out2, out3, out4, out5, out6, out7, \
out8, out9, out10, out11, out12, out13, out14, out15) { \
out0 = in0 + in15; \
out1 = in1 + in14; \
out2 = in2 + in13; \
out3 = in3 + in12; \
out4 = in4 + in11; \
out5 = in5 + in10; \
out6 = in6 + in9; \
out7 = in7 + in8; \
\
out8 = in7 - in8; \
out9 = in6 - in9; \
out10 = in5 - in10; \
out11 = in4 - in11; \
out12 = in3 - in12; \
out13 = in2 - in13; \
out14 = in1 - in14; \
out15 = in0 - in15; \
}
/* Description : Transposes input 8x8 byte block
Arguments : Inputs - in0, in1, in2, in3, in4, in5, in6, in7
(input 8x8 byte block)
......@@ -1640,7 +1712,7 @@
Outputs - out_m
Return Type - unsigned byte
Details : Signed byte even elements from 'in0' and 'in1' are packed
together in one vector and the resulted vector is xor'ed with
together in one vector and the resulting vector is xor'ed with
128 to shift the range from signed to unsigned byte
*/
#define PCKEV_XORI128_UB(in0, in1) ({ \
......
vp9/common/vp9_rtcd_defs.pl
View file @ 89b4b315
......@@ -1029,7 +1029,7 @@ if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
specialize qw/vp9_fht8x8 sse2/;
add_proto qw/void vp9_fht16x16/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
specialize qw/vp9_fht16x16 sse2/;
specialize qw/vp9_fht16x16 sse2 msa/;
add_proto qw/void vp9_fwht4x4/, "const int16_t *input, tran_low_t *output, int stride";
specialize qw/vp9_fwht4x4/, "$mmx_x86inc";
......@@ -1047,10 +1047,10 @@ if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
specialize qw/vp9_fdct8x8 sse2 neon/, "$ssse3_x86_64";
add_proto qw/void vp9_fdct16x16_1/, "const int16_t *input, tran_low_t *output, int stride";
specialize qw/vp9_fdct16x16_1 sse2/;
specialize qw/vp9_fdct16x16_1 sse2 msa/;
add_proto qw/void vp9_fdct16x16/, "const int16_t *input, tran_low_t *output, int stride";
specialize qw/vp9_fdct16x16 sse2/;
specialize qw/vp9_fdct16x16 sse2 msa/;
add_proto qw/void vp9_fdct32x32_1/, "const int16_t *input, tran_low_t *output, int stride";
specialize qw/vp9_fdct32x32_1 sse2/;
......
vp9/encoder/mips/msa/vp9_fdct16x16_msa.c 0 → 100644
View file @ 89b4b315
This diff is collapsed.
vp9/encoder/mips/msa/vp9_fdct_msa.h 0 → 100644
View file @ 89b4b315
This diff is collapsed.
vp9/vp9cx.mk
View file @ 89b4b315
......@@ -152,4 +152,7 @@ VP9_CX_SRCS-$(HAVE_NEON) += encoder/arm/neon/vp9_quantize_neon.c
VP9_CX_SRCS-$(HAVE_NEON) += encoder/arm/neon/vp9_subtract_neon.c
VP9_CX_SRCS-$(HAVE_NEON) += encoder/arm/neon/vp9_variance_neon.c
VP9_CX_SRCS-$(HAVE_MSA) += encoder/mips/msa/vp9_fdct16x16_msa.c
VP9_CX_SRCS-$(HAVE_MSA) += encoder/mips/msa/vp9_fdct_msa.h
VP9_CX_SRCS-yes := $(filter-out $(VP9_CX_SRCS_REMOVE-yes),$(VP9_CX_SRCS-yes))
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment