libvpx: Pull from upstream

source/libvpx/vp9/common/vp9_idct.h

Index: source/libvpx/vp9/common/vp9_idct.h
===================================================================
--- source/libvpx/vp9/common/vp9_idct.h	(revision 292072)
+++ source/libvpx/vp9/common/vp9_idct.h	(working copy)
@@ -36,52 +36,69 @@
 #define dual_set_epi16(a, b) \
   _mm_set_epi16(b, b, b, b, a, a, a, a)
 
+// Note:
+// tran_low_t  is the datatype used for final transform coefficients.
+// tran_high_t is the datatype used for intermediate transform stages.
+#if CONFIG_VP9_HIGHBITDEPTH
+typedef int64_t tran_high_t;
+typedef int32_t tran_low_t;
+#else
+typedef int32_t tran_high_t;
+typedef int16_t tran_low_t;
+#endif
+
 // Constants:
 //  for (int i = 1; i< 32; ++i)
 //    printf("static const int cospi_%d_64 = %.0f;\n", i,
 //           round(16384 * cos(i*M_PI/64)));
 // Note: sin(k*Pi/64) = cos((32-k)*Pi/64)
-static const int cospi_1_64  = 16364;
-static const int cospi_2_64  = 16305;
-static const int cospi_3_64  = 16207;
-static const int cospi_4_64  = 16069;
-static const int cospi_5_64  = 15893;
-static const int cospi_6_64  = 15679;
-static const int cospi_7_64  = 15426;
-static const int cospi_8_64  = 15137;
-static const int cospi_9_64  = 14811;
-static const int cospi_10_64 = 14449;
-static const int cospi_11_64 = 14053;
-static const int cospi_12_64 = 13623;
-static const int cospi_13_64 = 13160;
-static const int cospi_14_64 = 12665;
-static const int cospi_15_64 = 12140;
-static const int cospi_16_64 = 11585;
-static const int cospi_17_64 = 11003;
-static const int cospi_18_64 = 10394;
-static const int cospi_19_64 = 9760;
-static const int cospi_20_64 = 9102;
-static const int cospi_21_64 = 8423;
-static const int cospi_22_64 = 7723;
-static const int cospi_23_64 = 7005;
-static const int cospi_24_64 = 6270;
-static const int cospi_25_64 = 5520;
-static const int cospi_26_64 = 4756;
-static const int cospi_27_64 = 3981;
-static const int cospi_28_64 = 3196;
-static const int cospi_29_64 = 2404;
-static const int cospi_30_64 = 1606;
-static const int cospi_31_64 = 804;
+static const tran_high_t cospi_1_64  = 16364;
+static const tran_high_t cospi_2_64  = 16305;
+static const tran_high_t cospi_3_64  = 16207;
+static const tran_high_t cospi_4_64  = 16069;
+static const tran_high_t cospi_5_64  = 15893;
+static const tran_high_t cospi_6_64  = 15679;
+static const tran_high_t cospi_7_64  = 15426;
+static const tran_high_t cospi_8_64  = 15137;
+static const tran_high_t cospi_9_64  = 14811;
+static const tran_high_t cospi_10_64 = 14449;
+static const tran_high_t cospi_11_64 = 14053;
+static const tran_high_t cospi_12_64 = 13623;
+static const tran_high_t cospi_13_64 = 13160;
+static const tran_high_t cospi_14_64 = 12665;
+static const tran_high_t cospi_15_64 = 12140;
+static const tran_high_t cospi_16_64 = 11585;
+static const tran_high_t cospi_17_64 = 11003;
+static const tran_high_t cospi_18_64 = 10394;
+static const tran_high_t cospi_19_64 = 9760;
+static const tran_high_t cospi_20_64 = 9102;
+static const tran_high_t cospi_21_64 = 8423;
+static const tran_high_t cospi_22_64 = 7723;
+static const tran_high_t cospi_23_64 = 7005;
+static const tran_high_t cospi_24_64 = 6270;
+static const tran_high_t cospi_25_64 = 5520;
+static const tran_high_t cospi_26_64 = 4756;
+static const tran_high_t cospi_27_64 = 3981;
+static const tran_high_t cospi_28_64 = 3196;
+static const tran_high_t cospi_29_64 = 2404;
+static const tran_high_t cospi_30_64 = 1606;
+static const tran_high_t cospi_31_64 = 804;
 
 //  16384 * sqrt(2) * sin(kPi/9) * 2 / 3
-static const int sinpi_1_9 = 5283;
-static const int sinpi_2_9 = 9929;
-static const int sinpi_3_9 = 13377;
-static const int sinpi_4_9 = 15212;
+static const tran_high_t sinpi_1_9 = 5283;
+static const tran_high_t sinpi_2_9 = 9929;
+static const tran_high_t sinpi_3_9 = 13377;
+static const tran_high_t sinpi_4_9 = 15212;
 
-static INLINE int dct_const_round_shift(int input) {
-  int rv = ROUND_POWER_OF_TWO(input, DCT_CONST_BITS);
-#if CONFIG_COEFFICIENT_RANGE_CHECKING
+static INLINE tran_low_t dct_const_round_shift(tran_high_t input) {
+  tran_high_t rv = ROUND_POWER_OF_TWO(input, DCT_CONST_BITS);
+#if CONFIG_VP9_HIGHBITDEPTH
+  // For valid highbitdepth VP9 streams, intermediate stage coefficients will
+  // stay within the ranges:
+  // - 8 bit: signed 16 bit integer
+  // - 10 bit: signed 18 bit integer
+  // - 12 bit: signed 20 bit integer
+#elif CONFIG_COEFFICIENT_RANGE_CHECKING
   // For valid VP9 input streams, intermediate stage coefficients should always
   // stay within the range of a signed 16 bit integer. Coefficients can go out
   // of this range for invalid/corrupt VP9 streams. However, strictly checking
@@ -91,32 +108,59 @@
   assert(INT16_MIN <= rv);
   assert(rv <= INT16_MAX);
 #endif
-  return (int16_t)rv;
+  return (tran_low_t)rv;
 }
 
-typedef void (*transform_1d)(const int16_t*, int16_t*);
+typedef void (*transform_1d)(const tran_low_t*, tran_low_t*);
 
 typedef struct {
   transform_1d cols, rows;  // vertical and horizontal
 } transform_2d;
 
-void vp9_iwht4x4_add(const int16_t *input, uint8_t *dest, int stride, int eob);
+#if CONFIG_VP9_HIGHBITDEPTH
+typedef void (*high_transform_1d)(const tran_low_t*, tran_low_t*, int bd);
 
-void vp9_idct4x4_add(const int16_t *input, uint8_t *dest, int stride, int eob);
-void vp9_idct8x8_add(const int16_t *input, uint8_t *dest, int stride, int eob);
-void vp9_idct16x16_add(const int16_t *input, uint8_t *dest, int stride, int
+typedef struct {
+  high_transform_1d cols, rows;  // vertical and horizontal
+} high_transform_2d;
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+void vp9_iwht4x4_add(const tran_low_t *input, uint8_t *dest, int stride,
+                     int eob);
+void vp9_idct4x4_add(const tran_low_t *input, uint8_t *dest, int stride,
+                     int eob);
+void vp9_idct8x8_add(const tran_low_t *input, uint8_t *dest, int stride,
+                     int eob);
+void vp9_idct16x16_add(const tran_low_t *input, uint8_t *dest, int stride, int
                        eob);
-void vp9_idct32x32_add(const int16_t *input, uint8_t *dest, int stride,
+void vp9_idct32x32_add(const tran_low_t *input, uint8_t *dest, int stride,
                        int eob);
 
-void vp9_iht4x4_add(TX_TYPE tx_type, const int16_t *input, uint8_t *dest,
+void vp9_iht4x4_add(TX_TYPE tx_type, const tran_low_t *input, uint8_t *dest,
                     int stride, int eob);
-void vp9_iht8x8_add(TX_TYPE tx_type, const int16_t *input, uint8_t *dest,
+void vp9_iht8x8_add(TX_TYPE tx_type, const tran_low_t *input, uint8_t *dest,
                     int stride, int eob);
-void vp9_iht16x16_add(TX_TYPE tx_type, const int16_t *input, uint8_t *dest,
+void vp9_iht16x16_add(TX_TYPE tx_type, const tran_low_t *input, uint8_t *dest,
                       int stride, int eob);
 
-
+#if CONFIG_VP9_HIGHBITDEPTH
+void vp9_high_iwht4x4_add(const tran_low_t *input, uint8_t *dest, int stride,
+                          int eob, int bd);
+void vp9_high_idct4x4_add(const tran_low_t *input, uint8_t *dest, int stride,
+                          int eob, int bd);
+void vp9_high_idct8x8_add(const tran_low_t *input, uint8_t *dest, int stride,
+                          int eob, int bd);
+void vp9_high_idct16x16_add(const tran_low_t *input, uint8_t *dest, int stride,
+                            int eob, int bd);
+void vp9_high_idct32x32_add(const tran_low_t *input, uint8_t *dest, int stride,
+                            int eob, int bd);
+void vp9_high_iht4x4_add(TX_TYPE tx_type, const tran_low_t *input,
+                         uint8_t *dest, int stride, int eob, int bd);
+void vp9_high_iht8x8_add(TX_TYPE tx_type, const tran_low_t *input,
+                         uint8_t *dest, int stride, int eob, int bd);
+void vp9_high_iht16x16_add(TX_TYPE tx_type, const tran_low_t *input,
+                           uint8_t *dest, int stride, int eob, int bd);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
 #ifdef __cplusplus
 }  // extern "C"
 #endif