shithub: libvpx

Info  •  Files  •  Log  •  Branches

enable vpx_idct32x32_34_add_neon in hbd builds

replace load_and_transpose_s16_8x8() in idct32_6_neon() with a separate
load_tran_low_to_s16() and transpose_s16_8x8(). the combined function is
used in idct32_8_neon() where the input is the correctly sized output
from the earlier stage.

BUG=webm:1294

Change-Id: I4257c4b3a421b2cf5d13651f966eee0680ef98a9

--- a/test/partial_idct_test.cc

+++ b/test/partial_idct_test.cc

@@ -280,6 +280,8 @@

     NEON, PartialIDctTest,

     ::testing::Values(make_tuple(&vpx_fdct32x32_c, &vpx_idct32x32_1024_add_c,

                                  &vpx_idct32x32_1_add_neon, TX_32X32, 1),

+                      make_tuple(&vpx_fdct32x32_c, &vpx_idct32x32_1024_add_c,

+                                 &vpx_idct32x32_34_add_neon, TX_32X32, 34),

                       make_tuple(&vpx_fdct16x16_c, &vpx_idct16x16_256_add_c,

                                  &vpx_idct16x16_1_add_neon, TX_16X16, 1),

                       make_tuple(&vpx_fdct8x8_c, &vpx_idct8x8_64_add_c,

--- a/vpx_dsp/arm/idct32x32_34_add_neon.c

+++ b/vpx_dsp/arm/idct32x32_34_add_neon.c

@@ -34,7 +34,7 @@

 // 5 13 20 26

 // 6 21 27 33

 // 7 24 32

-static void idct32_6_neon(const int16_t *input, int16_t *output) {

+static void idct32_6_neon(const tran_low_t *input, int16_t *output) {

   int16x8_t in0, in1, in2, in3, in4, in5, in6, in7;

   int16x8_t s1_0, s1_1, s1_2, s1_3, s1_4, s1_5, s1_6, s1_7, s1_8, s1_9, s1_10,

       s1_11, s1_12, s1_13, s1_14, s1_15, s1_16, s1_17, s1_18, s1_19, s1_20,

@@ -46,8 +46,22 @@

       s2_31;

   int16x8_t s3_24, s3_25, s3_26, s3_27;

-  load_and_transpose_s16_8x8(input, 32, &in0, &in1, &in2, &in3, &in4, &in5,

-                             &in6, &in7);

+  in0 = load_tran_low_to_s16(input);

+  input += 32;

+  in1 = load_tran_low_to_s16(input);

+  input += 32;

+  in2 = load_tran_low_to_s16(input);

+  input += 32;

+  in3 = load_tran_low_to_s16(input);

+  input += 32;

+  in4 = load_tran_low_to_s16(input);

+  input += 32;

+  in5 = load_tran_low_to_s16(input);

+  input += 32;

+  in6 = load_tran_low_to_s16(input);

+  input += 32;

+  in7 = load_tran_low_to_s16(input);

+  transpose_s16_8x8(&in0, &in1, &in2, &in3, &in4, &in5, &in6, &in7);

   // stage 1

   // input[1] * cospi_31_64 - input[31] * cospi_1_64 (but input[31] == 0)

@@ -503,7 +517,7 @@

                        output + (24 * stride), stride);

 }

-void vpx_idct32x32_34_add_neon(const int16_t *input, uint8_t *dest,

+void vpx_idct32x32_34_add_neon(const tran_low_t *input, uint8_t *dest,

                                int stride) {

   int i;

   int16_t temp[32 * 8];

--- a/vpx_dsp/vpx_dsp.mk

+++ b/vpx_dsp/vpx_dsp.mk

@@ -213,7 +213,6 @@

 endif  # HAVE_NEON_ASM

 DSP_SRCS-$(HAVE_NEON)  += arm/idct16x16_neon.c

 DSP_SRCS-$(HAVE_NEON)  += arm/idct32x32_add_neon.c

-DSP_SRCS-$(HAVE_NEON)  += arm/idct32x32_34_add_neon.c

 DSP_SRCS-$(HAVE_NEON)  += arm/idct32x32_135_add_neon.c

 DSP_SRCS-$(HAVE_MSA)   += mips/inv_txfm_msa.h

@@ -246,6 +245,7 @@

 endif  # HAVE_NEON_ASM

 DSP_SRCS-$(HAVE_NEON) += arm/idct_neon.h

 DSP_SRCS-$(HAVE_NEON) += arm/idct32x32_1_add_neon.c

+DSP_SRCS-$(HAVE_NEON) += arm/idct32x32_34_add_neon.c

 endif  # CONFIG_VP9

--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl

+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl

@@ -688,7 +688,7 @@

     $vpx_idct32x32_135_add_sse2=vpx_idct32x32_1024_add_sse2;

     add_proto qw/void vpx_idct32x32_34_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";

-    specialize qw/vpx_idct32x32_34_add sse2/, "$ssse3_x86_64";

+    specialize qw/vpx_idct32x32_34_add neon sse2/, "$ssse3_x86_64";

     add_proto qw/void vpx_idct32x32_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";

     specialize qw/vpx_idct32x32_1_add neon sse2/;