SSE3-optimised av1_nn_predict
I have developed a SIMD-optimised neural network implementation using
SSE3. I have also added functional equivalence tests between this and
the original implementation. I added aom_clear_system_state() to a few
places where FPU operations are used after av1_nn_predict.
Speed-ups over the original C implementation for various network shapes:
10x64x16: 1.72x
12x12x1: 2.72x
12x24x1: 2.35x
12x32x1: 3.34x
18x24x4: 0.94x
18x32x4: 0.93x
4x16x1: 2.01x
8x16x1: 1.89x
8x16x4: 2.02x
8x24x1: 2.77x
8x32x1: 2.98x
8x64x1: 3.76x
9x32x3: 1.08x
4x8x4: 1.66x
A few awkwardly-shaped networks are slightly slower: these could be
padded to more convenient sizes to use the SIMD kernels.
I also wrote an AVX/AVX2 implementation but on these relatively small
networks it was barely faster than the SSE3 code.
Change-Id: I6a72be12cb7df8cf946578c3e01b21a439377d45
diff --git a/av1/common/av1_rtcd_defs.pl b/av1/common/av1_rtcd_defs.pl
index e38f897..c167c65 100755
--- a/av1/common/av1_rtcd_defs.pl
+++ b/av1/common/av1_rtcd_defs.pl
@@ -33,6 +33,8 @@
struct aom_variance_vtable;
struct search_site_config;
struct yv12_buffer_config;
+struct NN_CONFIG;
+typedef struct NN_CONFIG NN_CONFIG;
/* Function pointers return by CfL functions */
typedef void (*cfl_subsample_lbd_fn)(const uint8_t *input, int input_stride,
@@ -308,6 +310,9 @@
add_proto qw/void av1_get_horver_correlation_full/, " const int16_t *diff, int stride, int w, int h, float *hcorr, float *vcorr";
specialize qw/av1_get_horver_correlation_full sse4_1 avx2/;
+
+ add_proto qw/void av1_nn_predict/, " const float *input_nodes, const NN_CONFIG *const nn_config, float *const output";
+ specialize qw/av1_nn_predict sse3/;
}
# end encoder functions
