ref: fdd1186f9775cec72bcaaf98739b4d5f2e08b0ac
parent: b6e686b1ea0cb5d609de85836b829557660e7d55
author: James Zern <jzern@google.com>
date: Tue Sep 20 16:22:08 EDT 2016
vpx_idct32x32_34_add_sse2: rm unneeded transposes this change is neutral to mildly positive across various x86-64 platforms Change-Id: I28fb5ae598fc1317b7a42c9a846ac5d57d104784
--- a/vpx_dsp/x86/inv_txfm_sse2.c
+++ b/vpx_dsp/x86/inv_txfm_sse2.c
@@ -3066,17 +3066,7 @@
in[6] = load_input_data(input + 192);
in[7] = load_input_data(input + 224);
- for (i = 8; i < 32; ++i) {
- in[i] = _mm_setzero_si128();
- }
-
array_transpose_8x8(in, in);
- // TODO(hkuang): Following transposes are unnecessary. But remove them will
- // lead to performance drop on some devices.
- array_transpose_8x8(in + 8, in + 8);
- array_transpose_8x8(in + 16, in + 16);
- array_transpose_8x8(in + 24, in + 24);
-
IDCT32_34
// 1_D: Store 32 intermediate results for each 8x32 block.