Add unittest for encoder row based multi-threading
Added row-mt/row_mt as a command-line/init-time API parameter
in the encoder. Currently, encoding with row-mt=1 has only tile
level multi-threading. Support for row level multi-threading of
the encoding stage will be added eventually.
Change-Id: I67450dda05d1a1c32131d0998dd6546e147b681d
diff --git a/aom/aomcx.h b/aom/aomcx.h
index 07e9287..0d6ce7d 100644
--- a/aom/aomcx.h
+++ b/aom/aomcx.h
@@ -282,6 +282,12 @@
*/
AV1E_SET_LOSSLESS = AV1E_SET_GF_CBR_BOOST_PCT + 2,
+ /** control function to enable the row based multi-threading of encoder. A
+ * value that is equal to 1 indicates that row based multi-threading is
+ * enabled.
+ */
+ AV1E_SET_ROW_MT,
+
/*!\brief Codec control function to set number of tile columns.
*
* In encoding and decoding, AV1 allows an input image frame be partitioned
@@ -989,6 +995,9 @@
AOM_CTRL_USE_TYPE(AOME_SET_CQ_LEVEL, unsigned int)
#define AOM_CTRL_AOME_SET_CQ_LEVEL
+AOM_CTRL_USE_TYPE(AV1E_SET_ROW_MT, int)
+#define AOM_CTRL_AV1E_SET_ROW_MT
+
AOM_CTRL_USE_TYPE(AV1E_SET_TILE_COLUMNS, int)
#define AOM_CTRL_AV1E_SET_TILE_COLUMNS
AOM_CTRL_USE_TYPE(AV1E_SET_TILE_ROWS, int)
diff --git a/apps/aomenc.c b/apps/aomenc.c
index 7abed80..7518e3a 100644
--- a/apps/aomenc.c
+++ b/apps/aomenc.c
@@ -414,6 +414,9 @@
static const arg_def_t single_tile_decoding =
ARG_DEF(NULL, "single-tile-decoding", 1,
"Single tile decoding (0: off (default), 1: on)");
+static const arg_def_t rowmtarg =
+ ARG_DEF(NULL, "row-mt", 1,
+ "Enable row based multi-threading (0: off (default), 1: on)");
static const arg_def_t tile_cols =
ARG_DEF(NULL, "tile-columns", 1, "Number of tile columns to use, log2");
static const arg_def_t tile_rows =
@@ -621,6 +624,7 @@
&sharpness,
&static_thresh,
&single_tile_decoding,
+ &rowmtarg,
&tile_cols,
&tile_rows,
&arnr_maxframes,
@@ -676,6 +680,7 @@
AOME_SET_SHARPNESS,
AOME_SET_STATIC_THRESHOLD,
AV1E_SET_SINGLE_TILE_DECODING,
+ AV1E_SET_ROW_MT,
AV1E_SET_TILE_COLUMNS,
AV1E_SET_TILE_ROWS,
AOME_SET_ARNR_MAXFRAMES,
diff --git a/av1/av1_cx_iface.c b/av1/av1_cx_iface.c
index 0dd7c5e..12831f4 100644
--- a/av1/av1_cx_iface.c
+++ b/av1/av1_cx_iface.c
@@ -36,6 +36,7 @@
unsigned int noise_sensitivity;
unsigned int sharpness;
unsigned int static_thresh;
+ unsigned int row_mt;
unsigned int tile_columns; // log2 number of tile columns
unsigned int tile_rows; // log2 number of tile rows
unsigned int arnr_max_frames;
@@ -107,6 +108,7 @@
0, // noise_sensitivity
0, // sharpness
0, // static_thresh
+ 0, // row_mt
0, // tile_columns
0, // tile_rows
7, // arnr_max_frames
@@ -286,6 +288,8 @@
RANGE_CHECK_HI(cfg, large_scale_tile, 1);
RANGE_CHECK_HI(extra_cfg, single_tile_decoding, 1);
+ RANGE_CHECK_HI(extra_cfg, row_mt, 1);
+
RANGE_CHECK_HI(extra_cfg, tile_columns, 6);
RANGE_CHECK_HI(extra_cfg, tile_rows, 6);
@@ -631,6 +635,8 @@
oxcf->superblock_size = AOM_SUPERBLOCK_SIZE_64X64;
}
+ oxcf->row_mt = extra_cfg->row_mt;
+
oxcf->tile_columns = extra_cfg->tile_columns;
oxcf->tile_rows = extra_cfg->tile_rows;
@@ -794,6 +800,13 @@
return update_extra_cfg(ctx, &extra_cfg);
}
+static aom_codec_err_t ctrl_set_row_mt(aom_codec_alg_priv_t *ctx,
+ va_list args) {
+ struct av1_extracfg extra_cfg = ctx->extra_cfg;
+ extra_cfg.row_mt = CAST(AV1E_SET_ROW_MT, args);
+ return update_extra_cfg(ctx, &extra_cfg);
+}
+
static aom_codec_err_t ctrl_set_tile_columns(aom_codec_alg_priv_t *ctx,
va_list args) {
struct av1_extracfg extra_cfg = ctx->extra_cfg;
@@ -1673,6 +1686,7 @@
{ AOME_SET_ENABLEAUTOBWDREF, ctrl_set_enable_auto_bwd_ref },
{ AOME_SET_SHARPNESS, ctrl_set_sharpness },
{ AOME_SET_STATIC_THRESHOLD, ctrl_set_static_thresh },
+ { AV1E_SET_ROW_MT, ctrl_set_row_mt },
{ AV1E_SET_TILE_COLUMNS, ctrl_set_tile_columns },
{ AV1E_SET_TILE_ROWS, ctrl_set_tile_rows },
{ AOME_SET_ARNR_MAXFRAMES, ctrl_set_arnr_max_frames },
diff --git a/av1/encoder/encodeframe.c b/av1/encoder/encodeframe.c
index fd87ac1..3f28a52 100644
--- a/av1/encoder/encodeframe.c
+++ b/av1/encoder/encodeframe.c
@@ -5172,7 +5172,9 @@
}
#endif
- if (AOMMIN(cpi->oxcf.max_threads, cm->tile_cols * cm->tile_rows) > 1)
+ if (cpi->row_mt && (cpi->oxcf.max_threads > 1))
+ av1_encode_tiles_mt(cpi);
+ else if (AOMMIN(cpi->oxcf.max_threads, cm->tile_cols * cm->tile_rows) > 1)
av1_encode_tiles_mt(cpi);
else
encode_tiles(cpi);
diff --git a/av1/encoder/encoder.c b/av1/encoder/encoder.c
index 13fba0b..0be7699 100644
--- a/av1/encoder/encoder.c
+++ b/av1/encoder/encoder.c
@@ -2321,6 +2321,7 @@
cpi->oxcf = *oxcf;
cpi->common.options = oxcf->cfg;
+ cpi->row_mt = oxcf->row_mt;
x->e_mbd.bd = (int)seq_params->bit_depth;
x->e_mbd.global_motion = cm->global_motion;
diff --git a/av1/encoder/encoder.h b/av1/encoder/encoder.h
index 4fa6d66..621e0e3 100644
--- a/av1/encoder/encoder.h
+++ b/av1/encoder/encoder.h
@@ -248,6 +248,7 @@
int min_gf_interval;
int max_gf_interval;
+ int row_mt;
int tile_columns;
int tile_rows;
int tile_width_count;
@@ -539,6 +540,7 @@
int previous_index;
int cur_poc; // DebugInfo
+ unsigned int row_mt;
int scaled_ref_idx[REF_FRAMES];
int ref_fb_idx[REF_FRAMES];
int refresh_fb_idx; // ref frame buffer index to refresh
diff --git a/test/ethread_test.cc b/test/ethread_test.cc
index dd9fc2f..341c179 100644
--- a/test/ethread_test.cc
+++ b/test/ethread_test.cc
@@ -72,6 +72,7 @@
if (!encoder_initialized_) {
SetTileSize(encoder);
encoder->Control(AOME_SET_CPUUSED, set_cpu_used_);
+ encoder->Control(AV1E_SET_ROW_MT, row_mt_);
if (encoding_mode_ != ::libaom_test::kRealTime) {
encoder->Control(AOME_SET_ENABLEAUTOALTREF, 1);
encoder->Control(AOME_SET_ARNR_MAXFRAMES, 7);
@@ -119,6 +120,7 @@
cfg_.rc_target_bitrate = 1000;
// Encode using single thread.
+ row_mt_ = 0;
cfg_.g_threads = 1;
init_flags_ = AOM_CODEC_USE_PSNR;
ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
@@ -149,6 +151,54 @@
ASSERT_EQ(single_thr_size_enc, multi_thr_size_enc);
ASSERT_EQ(single_thr_md5_enc, multi_thr_md5_enc);
ASSERT_EQ(single_thr_md5_dec, multi_thr_md5_dec);
+
+ // Encode using multiple threads row-mt enabled.
+ row_mt_ = 1;
+ cfg_.g_threads = 2;
+ ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+ std::vector<size_t> multi_thr2_row_mt_size_enc;
+ std::vector<std::string> multi_thr2_row_mt_md5_enc;
+ std::vector<std::string> multi_thr2_row_mt_md5_dec;
+ multi_thr2_row_mt_size_enc = size_enc_;
+ multi_thr2_row_mt_md5_enc = md5_enc_;
+ multi_thr2_row_mt_md5_dec = md5_dec_;
+ size_enc_.clear();
+ md5_enc_.clear();
+ md5_dec_.clear();
+
+ cfg_.g_threads = 3;
+ ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+ std::vector<size_t> multi_thr3_row_mt_size_enc;
+ std::vector<std::string> multi_thr3_row_mt_md5_enc;
+ std::vector<std::string> multi_thr3_row_mt_md5_dec;
+ multi_thr3_row_mt_size_enc = size_enc_;
+ multi_thr3_row_mt_md5_enc = md5_enc_;
+ multi_thr3_row_mt_md5_dec = md5_dec_;
+ size_enc_.clear();
+ md5_enc_.clear();
+ md5_dec_.clear();
+
+ // Check that the vectors are equal.
+ ASSERT_EQ(multi_thr3_row_mt_size_enc, multi_thr2_row_mt_size_enc);
+ ASSERT_EQ(multi_thr3_row_mt_md5_enc, multi_thr2_row_mt_md5_enc);
+ ASSERT_EQ(multi_thr3_row_mt_md5_dec, multi_thr2_row_mt_md5_dec);
+
+ cfg_.g_threads = 4;
+ ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+ std::vector<size_t> multi_thr4_row_mt_size_enc;
+ std::vector<std::string> multi_thr4_row_mt_md5_enc;
+ std::vector<std::string> multi_thr4_row_mt_md5_dec;
+ multi_thr4_row_mt_size_enc = size_enc_;
+ multi_thr4_row_mt_md5_enc = md5_enc_;
+ multi_thr4_row_mt_md5_dec = md5_dec_;
+ size_enc_.clear();
+ md5_enc_.clear();
+ md5_dec_.clear();
+
+ // Check that the vectors are equal.
+ ASSERT_EQ(multi_thr4_row_mt_size_enc, multi_thr2_row_mt_size_enc);
+ ASSERT_EQ(multi_thr4_row_mt_md5_enc, multi_thr2_row_mt_md5_enc);
+ ASSERT_EQ(multi_thr4_row_mt_md5_dec, multi_thr2_row_mt_md5_dec);
}
bool encoder_initialized_;
@@ -156,6 +206,7 @@
int set_cpu_used_;
int tile_cols_;
int tile_rows_;
+ int row_mt_;
::libaom_test::Decoder *decoder_;
std::vector<size_t> size_enc_;
std::vector<std::string> md5_enc_;
@@ -180,13 +231,14 @@
AV1_INSTANTIATE_TEST_CASE(AVxEncoderThreadTest,
::testing::Values(::libaom_test::kTwoPassGood,
::libaom_test::kOnePassGood),
- ::testing::Range(2, 4), ::testing::Values(1, 2),
- ::testing::Values(0, 1));
+ ::testing::Values(1, 3), ::testing::Values(0, 1, 2),
+ ::testing::Values(0, 1, 2));
AV1_INSTANTIATE_TEST_CASE(AVxEncoderThreadTestLarge,
::testing::Values(::libaom_test::kTwoPassGood,
::libaom_test::kOnePassGood),
- ::testing::Range(0, 2), ::testing::Values(0, 1, 2, 6),
+ ::testing::Values(0, 2),
+ ::testing::Values(0, 1, 2, 6),
::testing::Values(0, 1, 2, 6));
class AVxEncoderThreadLSTest : public AVxEncoderThreadTest {
@@ -215,11 +267,12 @@
AV1_INSTANTIATE_TEST_CASE(AVxEncoderThreadLSTest,
::testing::Values(::libaom_test::kTwoPassGood,
::libaom_test::kOnePassGood),
- ::testing::Range(2, 4), ::testing::Values(6),
+ ::testing::Values(1, 3), ::testing::Values(0, 6),
::testing::Values(0, 6));
+
AV1_INSTANTIATE_TEST_CASE(AVxEncoderThreadLSTestLarge,
::testing::Values(::libaom_test::kTwoPassGood,
::libaom_test::kOnePassGood),
- ::testing::Range(0, 2), ::testing::Values(6),
+ ::testing::Range(0, 2), ::testing::Values(0, 6),
::testing::Values(0, 6));
} // namespace
