shithub: libvpx

--- a/vpx_dsp/arm/idct32x32_135_add_neon.c

+++ b/vpx_dsp/arm/idct32x32_135_add_neon.c

@@ -16,6 +16,50 @@

 #include "vpx_dsp/arm/transpose_neon.h"

 #include "vpx_dsp/txfm_common.h"

+static INLINE void load_8x8_s16(const tran_low_t *input, int16x8_t *const in0,

+                                int16x8_t *const in1, int16x8_t *const in2,

+                                int16x8_t *const in3, int16x8_t *const in4,

+                                int16x8_t *const in5, int16x8_t *const in6,

+                                int16x8_t *const in7) {

+  *in0 = load_tran_low_to_s16q(input);

+  input += 32;

+  *in1 = load_tran_low_to_s16q(input);

+  input += 32;

+  *in2 = load_tran_low_to_s16q(input);

+  input += 32;

+  *in3 = load_tran_low_to_s16q(input);

+  input += 32;

+  *in4 = load_tran_low_to_s16q(input);

+  input += 32;

+  *in5 = load_tran_low_to_s16q(input);

+  input += 32;

+  *in6 = load_tran_low_to_s16q(input);

+  input += 32;

+  *in7 = load_tran_low_to_s16q(input);

+}

+static INLINE void load_4x8_s16(const tran_low_t *input, int16x4_t *const in0,

+                                int16x4_t *const in1, int16x4_t *const in2,

+                                int16x4_t *const in3, int16x4_t *const in4,

+                                int16x4_t *const in5, int16x4_t *const in6,

+                                int16x4_t *const in7) {

+  *in0 = load_tran_low_to_s16d(input);

+  input += 32;

+  *in1 = load_tran_low_to_s16d(input);

+  input += 32;

+  *in2 = load_tran_low_to_s16d(input);

+  input += 32;

+  *in3 = load_tran_low_to_s16d(input);

+  input += 32;

+  *in4 = load_tran_low_to_s16d(input);

+  input += 32;

+  *in5 = load_tran_low_to_s16d(input);

+  input += 32;

+  *in6 = load_tran_low_to_s16d(input);

+  input += 32;

+  *in7 = load_tran_low_to_s16d(input);

+}

 // Only for the first pass of the  _135_ variant. Since it only uses values from

 // the top left 16x16 it can safely assume all the remaining values are 0 and

 // skip an awful lot of calculations. In fact, only the first 12 columns make

@@ -43,7 +87,7 @@

 // 13  84  93 103 110 125

 // 14  98 106 115 127

 // 15 117 128

-static void idct32_12_neon(const int16_t *input, int16_t *output) {

+static void idct32_12_neon(const tran_low_t *input, int16_t *output) {

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

   int16x4_t tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;

   int16x8_t in8, in9, in10, in11;

@@ -66,27 +110,11 @@

       s7_11, s7_12, s7_13, s7_14, s7_15, s7_20, s7_21, s7_22, s7_23, s7_24,

       s7_25, s7_26, s7_27;

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

-                             &in6, &in7);

+  load_8x8_s16(input, &in0, &in1, &in2, &in3, &in4, &in5, &in6, &in7);

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

-  input += 8;

-  tmp0 = vld1_s16(input);

-  input += 32;

-  tmp1 = vld1_s16(input);

-  input += 32;

-  tmp2 = vld1_s16(input);

-  input += 32;

-  tmp3 = vld1_s16(input);

-  input += 32;

-  tmp4 = vld1_s16(input);

-  input += 32;

-  tmp5 = vld1_s16(input);

-  input += 32;

-  tmp6 = vld1_s16(input);

-  input += 32;

-  tmp7 = vld1_s16(input);

+  load_4x8_s16(input + 8, &tmp0, &tmp1, &tmp2, &tmp3, &tmp4, &tmp5, &tmp6,

+               &tmp7);

   transpose_s16_4x8(tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7, &in8, &in9,

                     &in10, &in11);

@@ -669,7 +697,7 @@

                        output + (24 * stride), stride);

-void vpx_idct32x32_135_add_neon(const int16_t *input, uint8_t *dest,

+void vpx_idct32x32_135_add_neon(const tran_low_t *input, uint8_t *dest,

                                 int stride) {

   int i;

   int16_t temp[32 * 16];

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

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

@@ -46,21 +46,21 @@

       s2_31;

   int16x8_t s3_24, s3_25, s3_26, s3_27;

-  in0 = load_tran_low_to_s16(input);

+  in0 = load_tran_low_to_s16q(input);

   input += 32;

-  in1 = load_tran_low_to_s16(input);

+  in1 = load_tran_low_to_s16q(input);

   input += 32;

-  in2 = load_tran_low_to_s16(input);

+  in2 = load_tran_low_to_s16q(input);

   input += 32;

-  in3 = load_tran_low_to_s16(input);

+  in3 = load_tran_low_to_s16q(input);

   input += 32;

-  in4 = load_tran_low_to_s16(input);

+  in4 = load_tran_low_to_s16q(input);

   input += 32;

-  in5 = load_tran_low_to_s16(input);

+  in5 = load_tran_low_to_s16q(input);

   input += 32;

-  in6 = load_tran_low_to_s16(input);

+  in6 = load_tran_low_to_s16q(input);

   input += 32;

-  in7 = load_tran_low_to_s16(input);

+  in7 = load_tran_low_to_s16q(input);

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

   // stage 1

--- a/vpx_dsp/arm/idct4x4_add_neon.c

+++ b/vpx_dsp/arm/idct4x4_add_neon.c

@@ -62,8 +62,8 @@

   assert(!(dest_stride % sizeof(uint32_t)));

   // Rows

-  a0 = load_tran_low_to_s16(input);

-  a1 = load_tran_low_to_s16(input + 8);

+  a0 = load_tran_low_to_s16q(input);

+  a1 = load_tran_low_to_s16q(input + 8);

   idct4x4_16_kernel(cospis, &a0, &a1);

   // Columns

--- a/vpx_dsp/arm/idct8x8_add_neon.c

+++ b/vpx_dsp/arm/idct8x8_add_neon.c

@@ -174,14 +174,14 @@

   int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16, q13s16, q14s16, q15s16;

   uint16x8_t q8u16, q9u16, q10u16, q11u16;

-  q8s16 = load_tran_low_to_s16(input);

-  q9s16 = load_tran_low_to_s16(input + 8);

-  q10s16 = load_tran_low_to_s16(input + 16);

-  q11s16 = load_tran_low_to_s16(input + 24);

-  q12s16 = load_tran_low_to_s16(input + 32);

-  q13s16 = load_tran_low_to_s16(input + 40);

-  q14s16 = load_tran_low_to_s16(input + 48);

-  q15s16 = load_tran_low_to_s16(input + 56);

+  q8s16 = load_tran_low_to_s16q(input);

+  q9s16 = load_tran_low_to_s16q(input + 8);

+  q10s16 = load_tran_low_to_s16q(input + 16);

+  q11s16 = load_tran_low_to_s16q(input + 24);

+  q12s16 = load_tran_low_to_s16q(input + 32);

+  q13s16 = load_tran_low_to_s16q(input + 40);

+  q14s16 = load_tran_low_to_s16q(input + 48);

+  q15s16 = load_tran_low_to_s16q(input + 56);

   transpose_s16_8x8(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,

                     &q15s16);

@@ -280,14 +280,14 @@

   uint16x8_t q8u16, q9u16, q10u16, q11u16;

   int32x4_t q9s32, q10s32, q11s32, q12s32;

-  q8s16 = load_tran_low_to_s16(input);

-  q9s16 = load_tran_low_to_s16(input + 8);

-  q10s16 = load_tran_low_to_s16(input + 16);

-  q11s16 = load_tran_low_to_s16(input + 24);

-  q12s16 = load_tran_low_to_s16(input + 32);

-  q13s16 = load_tran_low_to_s16(input + 40);

-  q14s16 = load_tran_low_to_s16(input + 48);

-  q15s16 = load_tran_low_to_s16(input + 56);

+  q8s16 = load_tran_low_to_s16q(input);

+  q9s16 = load_tran_low_to_s16q(input + 8);

+  q10s16 = load_tran_low_to_s16q(input + 16);

+  q11s16 = load_tran_low_to_s16q(input + 24);

+  q12s16 = load_tran_low_to_s16q(input + 32);

+  q13s16 = load_tran_low_to_s16q(input + 40);

+  q14s16 = load_tran_low_to_s16q(input + 48);

+  q15s16 = load_tran_low_to_s16q(input + 56);

   transpose_s16_8x8(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,

                     &q15s16);

--- a/vpx_dsp/arm/idct_neon.h

+++ b/vpx_dsp/arm/idct_neon.h

@@ -18,9 +18,9 @@

 #include "vpx_dsp/vpx_dsp_common.h"

 //------------------------------------------------------------------------------

+// Helper functions used to load tran_low_t into int16, narrowing if necessary.

-// Helper function used to load tran_low_t into int16, narrowing if necessary.

-static INLINE int16x8_t load_tran_low_to_s16(const tran_low_t *buf) {

+static INLINE int16x8_t load_tran_low_to_s16q(const tran_low_t *buf) {

 #if CONFIG_VP9_HIGHBITDEPTH

   const int32x4_t v0 = vld1q_s32(buf);

   const int32x4_t v1 = vld1q_s32(buf + 4);

@@ -31,6 +31,17 @@

   return vld1q_s16(buf);

 #endif

+static INLINE int16x4_t load_tran_low_to_s16d(const tran_low_t *buf) {

+#if CONFIG_VP9_HIGHBITDEPTH

+  const int32x4_t v0 = vld1q_s32(buf);

+  return vmovn_s32(v0);

+#else

+  return vld1_s16(buf);

+#endif

+}

+//------------------------------------------------------------------------------

 // Multiply a by a_const. Saturate, shift and narrow by 14.

 static INLINE int16x8_t multiply_shift_and_narrow_s16(const int16x8_t a,

--- 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_135_add_neon.c

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

 DSP_SRCS-$(HAVE_MSA)   += mips/idct4x4_msa.c

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

 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

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

 endif  # CONFIG_VP9

--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl

+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl

@@ -698,7 +698,7 @@

     specialize qw/vpx_idct32x32_1024_add sse2/, "$ssse3_x86_64";

     add_proto qw/void vpx_idct32x32_135_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";

-    specialize qw/vpx_idct32x32_135_add sse2/, "$ssse3_x86_64";

+    specialize qw/vpx_idct32x32_135_add neon sse2/, "$ssse3_x86_64";

     # Need to add 135 eob idct32x32 implementations.

     $vpx_idct32x32_135_add_sse2=vpx_idct32x32_1024_add_sse2;