From: Linfeng Zhang Date: Tue, 14 Feb 2017 18:24:51 +0000 (-0800) Subject: Add vpx_highbd_idct16x16_10_add_neon() X-Git-Tag: v1.7.0~695^2 X-Git-Url: https://granicus.if.org/sourcecode?a=commitdiff_plain;h=06200817310b6c1f214a7c3d155d10c853051e27;p=libvpx Add vpx_highbd_idct16x16_10_add_neon() BUG=webm:1301 Change-Id: If686c8144764c4162458f0bc4bb1bbf6555c48ab --- diff --git a/test/partial_idct_test.cc b/test/partial_idct_test.cc index 9390d0c57..742ee4db6 100644 --- a/test/partial_idct_test.cc +++ b/test/partial_idct_test.cc @@ -486,6 +486,15 @@ const PartialInvTxfmParam neon_partial_idct_tests[] = { make_tuple( &vpx_highbd_fdct16x16_c, &highbd_wrapper, &highbd_wrapper, TX_16X16, 38, 12, 2), + make_tuple( + &vpx_highbd_fdct16x16_c, &highbd_wrapper, + &highbd_wrapper, TX_16X16, 10, 8, 2), + make_tuple( + &vpx_highbd_fdct16x16_c, &highbd_wrapper, + &highbd_wrapper, TX_16X16, 10, 10, 2), + make_tuple( + &vpx_highbd_fdct16x16_c, &highbd_wrapper, + &highbd_wrapper, TX_16X16, 10, 12, 2), make_tuple( &vpx_highbd_fdct16x16_c, &highbd_wrapper, &highbd_wrapper, TX_16X16, 1, 8, 2), diff --git a/vpx_dsp/arm/highbd_idct16x16_add_neon.c b/vpx_dsp/arm/highbd_idct16x16_add_neon.c index 1a97bfdff..d361c8263 100644 --- a/vpx_dsp/arm/highbd_idct16x16_add_neon.c +++ b/vpx_dsp/arm/highbd_idct16x16_add_neon.c @@ -33,6 +33,19 @@ static INLINE void highbd_idct16x16_add_wrap_low_8x2(const int64x2x2_t *const t, d1->val[1] = vcombine_s32(t32[3].val[0], t32[3].val[1]); } +static INLINE void highbd_idct16x16_add_wrap_low_4x2(const int64x2x2_t *const t, + int32x4_t *const d0, + int32x4_t *const d1) { + int32x2x2_t t32[2]; + + t32[0].val[0] = vrshrn_n_s64(t[0].val[0], DCT_CONST_BITS); + t32[0].val[1] = vrshrn_n_s64(t[0].val[1], DCT_CONST_BITS); + t32[1].val[0] = vrshrn_n_s64(t[1].val[0], DCT_CONST_BITS); + t32[1].val[1] = vrshrn_n_s64(t[1].val[1], DCT_CONST_BITS); + *d0 = vcombine_s32(t32[0].val[0], t32[0].val[1]); + *d1 = vcombine_s32(t32[1].val[0], t32[1].val[1]); +} + static INLINE int32x4x2_t highbd_idct16x16_add_wrap_low_8x1(const int64x2x2_t *const t) { int32x2x2_t t32[2]; @@ -47,6 +60,14 @@ highbd_idct16x16_add_wrap_low_8x1(const int64x2x2_t *const t) { return d; } +static INLINE int32x4_t highbd_idct16x16_add_wrap_low_4x1(const int64x2x2_t t) { + int32x2x2_t t32; + + t32.val[0] = vrshrn_n_s64(t.val[0], DCT_CONST_BITS); + t32.val[1] = vrshrn_n_s64(t.val[1], DCT_CONST_BITS); + return vcombine_s32(t32.val[0], t32.val[1]); +} + static INLINE void highbd_idct_cospi_2_30(const int32x4x2_t s0, const int32x4x2_t s1, const int32x4_t cospi_2_30_10_22, @@ -336,6 +357,27 @@ static INLINE void highbd_idct_cospi_8_24_q_kernel( vget_low_s32(cospi_0_8_16_24), 1); } +static INLINE void highbd_idct_cospi_8_24_d_kernel( + const int32x4_t s0, const int32x4_t s1, const int32x4_t cospi_0_8_16_24, + int64x2x2_t *const t) { + t[0].val[0] = + vmull_lane_s32(vget_low_s32(s0), vget_high_s32(cospi_0_8_16_24), 1); + t[0].val[1] = + vmull_lane_s32(vget_high_s32(s0), vget_high_s32(cospi_0_8_16_24), 1); + t[1].val[0] = + vmull_lane_s32(vget_low_s32(s1), vget_high_s32(cospi_0_8_16_24), 1); + t[1].val[1] = + vmull_lane_s32(vget_high_s32(s1), vget_high_s32(cospi_0_8_16_24), 1); + t[0].val[0] = vmlsl_lane_s32(t[0].val[0], vget_low_s32(s1), + vget_low_s32(cospi_0_8_16_24), 1); + t[0].val[1] = vmlsl_lane_s32(t[0].val[1], vget_high_s32(s1), + vget_low_s32(cospi_0_8_16_24), 1); + t[1].val[0] = vmlal_lane_s32(t[1].val[0], vget_low_s32(s0), + vget_low_s32(cospi_0_8_16_24), 1); + t[1].val[1] = vmlal_lane_s32(t[1].val[1], vget_high_s32(s0), + vget_low_s32(cospi_0_8_16_24), 1); +} + static INLINE void highbd_idct_cospi_8_24_q(const int32x4x2_t s0, const int32x4x2_t s1, const int32x4_t cospi_0_8_16_24, @@ -347,6 +389,17 @@ static INLINE void highbd_idct_cospi_8_24_q(const int32x4x2_t s0, highbd_idct16x16_add_wrap_low_8x2(t, d0, d1); } +static INLINE void highbd_idct_cospi_8_24_d(const int32x4_t s0, + const int32x4_t s1, + const int32x4_t cospi_0_8_16_24, + int32x4_t *const d0, + int32x4_t *const d1) { + int64x2x2_t t[2]; + + highbd_idct_cospi_8_24_d_kernel(s0, s1, cospi_0_8_16_24, t); + highbd_idct16x16_add_wrap_low_4x2(t, d0, d1); +} + static INLINE void highbd_idct_cospi_8_24_neg_q(const int32x4x2_t s0, const int32x4x2_t s1, const int32x4_t cospi_0_8_16_24, @@ -362,6 +415,19 @@ static INLINE void highbd_idct_cospi_8_24_neg_q(const int32x4x2_t s0, highbd_idct16x16_add_wrap_low_8x2(t, d0, d1); } +static INLINE void highbd_idct_cospi_8_24_neg_d(const int32x4_t s0, + const int32x4_t s1, + const int32x4_t cospi_0_8_16_24, + int32x4_t *const d0, + int32x4_t *const d1) { + int64x2x2_t t[2]; + + highbd_idct_cospi_8_24_d_kernel(s0, s1, cospi_0_8_16_24, t); + t[1].val[0] = vsubq_s64(vdupq_n_s64(0), t[1].val[0]); + t[1].val[1] = vsubq_s64(vdupq_n_s64(0), t[1].val[1]); + highbd_idct16x16_add_wrap_low_4x2(t, d0, d1); +} + static INLINE void highbd_idct_cospi_16_16_q(const int32x4x2_t s0, const int32x4x2_t s1, const int32x4_t cospi_0_8_16_24, @@ -396,8 +462,30 @@ static INLINE void highbd_idct_cospi_16_16_q(const int32x4x2_t s0, highbd_idct16x16_add_wrap_low_8x2(t, d0, d1); } -static INLINE void highbd_idct16x16_add_stage7(const int32x4x2_t *const step2, - int32x4x2_t *const out) { +static INLINE void highbd_idct_cospi_16_16_d(const int32x4_t s0, + const int32x4_t s1, + const int32x4_t cospi_0_8_16_24, + int32x4_t *const d0, + int32x4_t *const d1) { + int64x2x2_t t[3]; + + t[2].val[0] = + vmull_lane_s32(vget_low_s32(s1), vget_high_s32(cospi_0_8_16_24), 0); + t[2].val[1] = + vmull_lane_s32(vget_high_s32(s1), vget_high_s32(cospi_0_8_16_24), 0); + t[0].val[0] = vmlsl_lane_s32(t[2].val[0], vget_low_s32(s0), + vget_high_s32(cospi_0_8_16_24), 0); + t[0].val[1] = vmlsl_lane_s32(t[2].val[1], vget_high_s32(s0), + vget_high_s32(cospi_0_8_16_24), 0); + t[1].val[0] = vmlal_lane_s32(t[2].val[0], vget_low_s32(s0), + vget_high_s32(cospi_0_8_16_24), 0); + t[1].val[1] = vmlal_lane_s32(t[2].val[1], vget_high_s32(s0), + vget_high_s32(cospi_0_8_16_24), 0); + highbd_idct16x16_add_wrap_low_4x2(t, d0, d1); +} + +static INLINE void highbd_idct16x16_add_stage7_dual( + const int32x4x2_t *const step2, int32x4x2_t *const out) { out[0].val[0] = vaddq_s32(step2[0].val[0], step2[15].val[0]); out[0].val[1] = vaddq_s32(step2[0].val[1], step2[15].val[1]); out[1].val[0] = vaddq_s32(step2[1].val[0], step2[14].val[0]); @@ -432,6 +520,26 @@ static INLINE void highbd_idct16x16_add_stage7(const int32x4x2_t *const step2, out[15].val[1] = vsubq_s32(step2[0].val[1], step2[15].val[1]); } +static INLINE void highbd_idct16x16_add_stage7(const int32x4_t *const step2, + int32x4_t *const out) { + out[0] = vaddq_s32(step2[0], step2[15]); + out[1] = vaddq_s32(step2[1], step2[14]); + out[2] = vaddq_s32(step2[2], step2[13]); + out[3] = vaddq_s32(step2[3], step2[12]); + out[4] = vaddq_s32(step2[4], step2[11]); + out[5] = vaddq_s32(step2[5], step2[10]); + out[6] = vaddq_s32(step2[6], step2[9]); + out[7] = vaddq_s32(step2[7], step2[8]); + out[8] = vsubq_s32(step2[7], step2[8]); + out[9] = vsubq_s32(step2[6], step2[9]); + out[10] = vsubq_s32(step2[5], step2[10]); + out[11] = vsubq_s32(step2[4], step2[11]); + out[12] = vsubq_s32(step2[3], step2[12]); + out[13] = vsubq_s32(step2[2], step2[13]); + out[14] = vsubq_s32(step2[1], step2[14]); + out[15] = vsubq_s32(step2[0], step2[15]); +} + static INLINE void highbd_idct16x16_store_pass1(const int32x4x2_t *const out, int32_t *output) { // Save the result into output @@ -745,7 +853,7 @@ static void highbd_idct16x16_256_add_half1d(const int32_t *input, step2[15] = step1[15]; // stage 7 - highbd_idct16x16_add_stage7(step2, out); + highbd_idct16x16_add_stage7_dual(step2, out); if (output) { highbd_idct16x16_store_pass1(out, output); @@ -765,6 +873,15 @@ static INLINE int32x4x2_t highbd_idct_cospi_lane0_dual(const int32x4x2_t s, return highbd_idct16x16_add_wrap_low_8x1(t); } +static INLINE int32x4_t highbd_idct_cospi_lane0(const int32x4_t s, + const int32x2_t coef) { + int64x2x2_t t; + + t.val[0] = vmull_lane_s32(vget_low_s32(s), coef, 0); + t.val[1] = vmull_lane_s32(vget_high_s32(s), coef, 0); + return highbd_idct16x16_add_wrap_low_4x1(t); +} + static INLINE int32x4x2_t highbd_idct_cospi_lane1_dual(const int32x4x2_t s, const int32x2_t coef) { int64x2x2_t t[2]; @@ -776,6 +893,15 @@ static INLINE int32x4x2_t highbd_idct_cospi_lane1_dual(const int32x4x2_t s, return highbd_idct16x16_add_wrap_low_8x1(t); } +static INLINE int32x4_t highbd_idct_cospi_lane1(const int32x4_t s, + const int32x2_t coef) { + int64x2x2_t t; + + t.val[0] = vmull_lane_s32(vget_low_s32(s), coef, 1); + t.val[1] = vmull_lane_s32(vget_high_s32(s), coef, 1); + return highbd_idct16x16_add_wrap_low_4x1(t); +} + static INLINE int32x4x2_t highbd_idct_add_dual(const int32x4x2_t s0, const int32x4x2_t s1) { int32x4x2_t t; @@ -938,9 +1064,274 @@ static void highbd_idct16x16_38_add_half1d(const int32_t *input, step2[14] = step1[14]; step2[15] = step1[15]; + // stage 7 + highbd_idct16x16_add_stage7_dual(step2, out); + + if (output) { + highbd_idct16x16_store_pass1(out, output); + } else { + highbd_idct16x16_add_store(out, dest, stride, bd); + } +} + +void highbd_idct16x16_10_add_half1d_pass1(const tran_low_t *input, + int32_t *output) { + const int32x4_t cospi_0_8_16_24 = vld1q_s32(kCospi32 + 0); + const int32x4_t cospi_4_12_20N_28 = vld1q_s32(kCospi32 + 4); + const int32x4_t cospi_2_30_10_22 = vld1q_s32(kCospi32 + 8); + const int32x4_t cospi_6_26_14_18N = vld1q_s32(kCospi32 + 12); + int32x4_t in[4], step1[16], step2[16], out[16]; + + // Load input (4x4) + in[0] = vld1q_s32(input); + input += 16; + in[1] = vld1q_s32(input); + input += 16; + in[2] = vld1q_s32(input); + input += 16; + in[3] = vld1q_s32(input); + + // Transpose + transpose_s32_4x4(&in[0], &in[1], &in[2], &in[3]); + + // stage 1 + step1[0] = in[0 / 2]; + step1[4] = in[4 / 2]; + step1[8] = in[2 / 2]; + step1[12] = in[6 / 2]; + + // stage 2 + step2[0] = step1[0]; + step2[4] = step1[4]; + step2[8] = highbd_idct_cospi_lane1(step1[8], vget_low_s32(cospi_2_30_10_22)); + step2[11] = + highbd_idct_cospi_lane1(step1[12], vget_low_s32(cospi_6_26_14_18N)); + step2[12] = + highbd_idct_cospi_lane0(step1[12], vget_low_s32(cospi_6_26_14_18N)); + step2[15] = highbd_idct_cospi_lane0(step1[8], vget_low_s32(cospi_2_30_10_22)); + + // stage 3 + step1[0] = step2[0]; + step1[4] = + highbd_idct_cospi_lane1(step2[4], vget_high_s32(cospi_4_12_20N_28)); + step1[7] = highbd_idct_cospi_lane0(step2[4], vget_low_s32(cospi_4_12_20N_28)); + step1[8] = step2[8]; + step1[9] = step2[8]; + step1[10] = step2[11]; + step1[11] = step2[11]; + step1[12] = step2[12]; + step1[13] = step2[12]; + step1[14] = step2[15]; + step1[15] = step2[15]; + + // stage 4 + step2[0] = step2[1] = + highbd_idct_cospi_lane0(step1[0], vget_high_s32(cospi_0_8_16_24)); + step2[4] = step1[4]; + step2[5] = step1[4]; + step2[6] = step1[7]; + step2[7] = step1[7]; + step2[8] = step1[8]; + highbd_idct_cospi_8_24_d(step1[14], step1[9], cospi_0_8_16_24, &step2[9], + &step2[14]); + highbd_idct_cospi_8_24_neg_d(step1[13], step1[10], cospi_0_8_16_24, + &step2[13], &step2[10]); + step2[11] = step1[11]; + step2[12] = step1[12]; + step2[15] = step1[15]; + + // stage 5 + step1[0] = step2[0]; + step1[1] = step2[1]; + step1[2] = step2[1]; + step1[3] = step2[0]; + step1[4] = step2[4]; + highbd_idct_cospi_16_16_d(step2[5], step2[6], cospi_0_8_16_24, &step1[5], + &step1[6]); + step1[7] = step2[7]; + step1[8] = vaddq_s32(step2[8], step2[11]); + step1[9] = vaddq_s32(step2[9], step2[10]); + step1[10] = vsubq_s32(step2[9], step2[10]); + step1[11] = vsubq_s32(step2[8], step2[11]); + step1[12] = vsubq_s32(step2[15], step2[12]); + step1[13] = vsubq_s32(step2[14], step2[13]); + step1[14] = vaddq_s32(step2[14], step2[13]); + step1[15] = vaddq_s32(step2[15], step2[12]); + + // stage 6 + step2[0] = vaddq_s32(step1[0], step1[7]); + step2[1] = vaddq_s32(step1[1], step1[6]); + step2[2] = vaddq_s32(step1[2], step1[5]); + step2[3] = vaddq_s32(step1[3], step1[4]); + step2[4] = vsubq_s32(step1[3], step1[4]); + step2[5] = vsubq_s32(step1[2], step1[5]); + step2[6] = vsubq_s32(step1[1], step1[6]); + step2[7] = vsubq_s32(step1[0], step1[7]); + highbd_idct_cospi_16_16_d(step1[10], step1[13], cospi_0_8_16_24, &step2[10], + &step2[13]); + highbd_idct_cospi_16_16_d(step1[11], step1[12], cospi_0_8_16_24, &step2[11], + &step2[12]); + step2[8] = step1[8]; + step2[9] = step1[9]; + step2[14] = step1[14]; + step2[15] = step1[15]; + // stage 7 highbd_idct16x16_add_stage7(step2, out); + // pass 1: save the result into output + vst1q_s32(output, out[0]); + output += 4; + vst1q_s32(output, out[1]); + output += 4; + vst1q_s32(output, out[2]); + output += 4; + vst1q_s32(output, out[3]); + output += 4; + vst1q_s32(output, out[4]); + output += 4; + vst1q_s32(output, out[5]); + output += 4; + vst1q_s32(output, out[6]); + output += 4; + vst1q_s32(output, out[7]); + output += 4; + vst1q_s32(output, out[8]); + output += 4; + vst1q_s32(output, out[9]); + output += 4; + vst1q_s32(output, out[10]); + output += 4; + vst1q_s32(output, out[11]); + output += 4; + vst1q_s32(output, out[12]); + output += 4; + vst1q_s32(output, out[13]); + output += 4; + vst1q_s32(output, out[14]); + output += 4; + vst1q_s32(output, out[15]); +} + +void highbd_idct16x16_10_add_half1d_pass2(const int32_t *input, + int32_t *const output, + uint16_t *const dest, + const int stride, const int bd) { + const int32x4_t cospi_0_8_16_24 = vld1q_s32(kCospi32 + 0); + const int32x4_t cospi_4_12_20N_28 = vld1q_s32(kCospi32 + 4); + const int32x4_t cospi_2_30_10_22 = vld1q_s32(kCospi32 + 8); + const int32x4_t cospi_6_26_14_18N = vld1q_s32(kCospi32 + 12); + int32x4x2_t in[4], step1[16], step2[16], out[16]; + + // Load input (4x8) + in[0].val[0] = vld1q_s32(input); + input += 4; + in[0].val[1] = vld1q_s32(input); + input += 4; + in[1].val[0] = vld1q_s32(input); + input += 4; + in[1].val[1] = vld1q_s32(input); + input += 4; + in[2].val[0] = vld1q_s32(input); + input += 4; + in[2].val[1] = vld1q_s32(input); + input += 4; + in[3].val[0] = vld1q_s32(input); + input += 4; + in[3].val[1] = vld1q_s32(input); + + // Transpose + transpose_s32_4x8(&in[0].val[0], &in[0].val[1], &in[1].val[0], &in[1].val[1], + &in[2].val[0], &in[2].val[1], &in[3].val[0], &in[3].val[1]); + + // stage 1 + step1[0] = in[0 / 2]; + step1[4] = in[4 / 2]; + step1[8] = in[2 / 2]; + step1[12] = in[6 / 2]; + + // stage 2 + step2[0] = step1[0]; + step2[4] = step1[4]; + step2[8] = + highbd_idct_cospi_lane1_dual(step1[8], vget_low_s32(cospi_2_30_10_22)); + step2[11] = + highbd_idct_cospi_lane1_dual(step1[12], vget_low_s32(cospi_6_26_14_18N)); + step2[12] = + highbd_idct_cospi_lane0_dual(step1[12], vget_low_s32(cospi_6_26_14_18N)); + step2[15] = + highbd_idct_cospi_lane0_dual(step1[8], vget_low_s32(cospi_2_30_10_22)); + + // stage 3 + step1[0] = step2[0]; + step1[4] = + highbd_idct_cospi_lane1_dual(step2[4], vget_high_s32(cospi_4_12_20N_28)); + step1[7] = + highbd_idct_cospi_lane0_dual(step2[4], vget_low_s32(cospi_4_12_20N_28)); + step1[8] = step2[8]; + step1[9] = step2[8]; + step1[10] = step2[11]; + step1[11] = step2[11]; + step1[12] = step2[12]; + step1[13] = step2[12]; + step1[14] = step2[15]; + step1[15] = step2[15]; + + // stage 4 + step2[0] = step2[1] = + highbd_idct_cospi_lane0_dual(step1[0], vget_high_s32(cospi_0_8_16_24)); + step2[4] = step1[4]; + step2[5] = step1[4]; + step2[6] = step1[7]; + step2[7] = step1[7]; + step2[8] = step1[8]; + highbd_idct_cospi_8_24_q(step1[14], step1[9], cospi_0_8_16_24, &step2[9], + &step2[14]); + highbd_idct_cospi_8_24_neg_q(step1[13], step1[10], cospi_0_8_16_24, + &step2[13], &step2[10]); + step2[11] = step1[11]; + step2[12] = step1[12]; + step2[15] = step1[15]; + + // stage 5 + step1[0] = step2[0]; + step1[1] = step2[1]; + step1[2] = step2[1]; + step1[3] = step2[0]; + step1[4] = step2[4]; + highbd_idct_cospi_16_16_q(step2[5], step2[6], cospi_0_8_16_24, &step1[5], + &step1[6]); + step1[7] = step2[7]; + step1[8] = highbd_idct_add_dual(step2[8], step2[11]); + step1[9] = highbd_idct_add_dual(step2[9], step2[10]); + step1[10] = highbd_idct_sub_dual(step2[9], step2[10]); + step1[11] = highbd_idct_sub_dual(step2[8], step2[11]); + step1[12] = highbd_idct_sub_dual(step2[15], step2[12]); + step1[13] = highbd_idct_sub_dual(step2[14], step2[13]); + step1[14] = highbd_idct_add_dual(step2[14], step2[13]); + step1[15] = highbd_idct_add_dual(step2[15], step2[12]); + + // stage 6 + step2[0] = highbd_idct_add_dual(step1[0], step1[7]); + step2[1] = highbd_idct_add_dual(step1[1], step1[6]); + step2[2] = highbd_idct_add_dual(step1[2], step1[5]); + step2[3] = highbd_idct_add_dual(step1[3], step1[4]); + step2[4] = highbd_idct_sub_dual(step1[3], step1[4]); + step2[5] = highbd_idct_sub_dual(step1[2], step1[5]); + step2[6] = highbd_idct_sub_dual(step1[1], step1[6]); + step2[7] = highbd_idct_sub_dual(step1[0], step1[7]); + highbd_idct_cospi_16_16_q(step1[10], step1[13], cospi_0_8_16_24, &step2[10], + &step2[13]); + highbd_idct_cospi_16_16_q(step1[11], step1[12], cospi_0_8_16_24, &step2[11], + &step2[12]); + step2[8] = step1[8]; + step2[9] = step1[9]; + step2[14] = step1[14]; + step2[15] = step1[15]; + + // stage 7 + highbd_idct16x16_add_stage7_dual(step2, out); + if (output) { highbd_idct16x16_store_pass1(out, output); } else { @@ -1026,6 +1417,42 @@ void vpx_highbd_idct16x16_38_add_neon(const tran_low_t *input, uint8_t *dest8, } } +void vpx_highbd_idct16x16_10_add_neon(const tran_low_t *input, uint8_t *dest8, + int stride, int bd) { + uint16_t *dest = CONVERT_TO_SHORTPTR(dest8); + + if (bd == 8) { + int16_t row_idct_output[4 * 16]; + + // pass 1 + // Parallel idct on the upper 8 rows + idct16x16_10_add_half1d_pass1(input, row_idct_output); + + // pass 2 + // Parallel idct to get the left 8 columns + idct16x16_10_add_half1d_pass2(row_idct_output, NULL, dest, stride, 1); + + // Parallel idct to get the right 8 columns + idct16x16_10_add_half1d_pass2(row_idct_output + 4 * 8, NULL, dest + 8, + stride, 1); + } else { + int32_t row_idct_output[4 * 16]; + + // pass 1 + // Parallel idct on the upper 8 rows + highbd_idct16x16_10_add_half1d_pass1(input, row_idct_output); + + // pass 2 + // Parallel idct to get the left 8 columns + highbd_idct16x16_10_add_half1d_pass2(row_idct_output, NULL, dest, stride, + bd); + + // Parallel idct to get the right 8 columns + highbd_idct16x16_10_add_half1d_pass2(row_idct_output + 4 * 8, NULL, + dest + 8, stride, bd); + } +} + static INLINE void highbd_idct16x16_1_add_pos_kernel(uint16_t **dest, const int stride, const int16x8_t res, diff --git a/vpx_dsp/arm/idct16x16_add_neon.c b/vpx_dsp/arm/idct16x16_add_neon.c index 30bdb4802..4259cd8ca 100644 --- a/vpx_dsp/arm/idct16x16_add_neon.c +++ b/vpx_dsp/arm/idct16x16_add_neon.c @@ -484,8 +484,7 @@ void idct16x16_38_add_half1d(const void *const input, int16_t *const output, } } -static void idct16x16_10_add_half1d_pass1(const tran_low_t *input, - int16_t *output) { +void idct16x16_10_add_half1d_pass1(const tran_low_t *input, int16_t *output) { const int16x8_t cospis0 = vld1q_s16(kCospi); const int16x8_t cospis1 = vld1q_s16(kCospi + 8); const int16x8_t cospisd0 = vaddq_s16(cospis0, cospis0); @@ -638,8 +637,9 @@ static void idct16x16_10_add_half1d_pass1(const tran_low_t *input, vst1_s16(output, out[15]); } -static void idct16x16_10_add_half1d_pass2(const int16_t *input, int16_t *output, - uint8_t *dest, const int stride) { +void idct16x16_10_add_half1d_pass2(const int16_t *input, int16_t *const output, + void *const dest, const int stride, + const int highbd_flag) { const int16x8_t cospis0 = vld1q_s16(kCospi); const int16x8_t cospis1 = vld1q_s16(kCospi + 8); const int16x8_t cospisd0 = vaddq_s16(cospis0, cospis0); @@ -751,27 +751,16 @@ static void idct16x16_10_add_half1d_pass2(const int16_t *input, int16_t *output, step2[15] = step1[15]; // stage 7 - out[0] = vaddq_s16(step2[0], step2[15]); - out[1] = vaddq_s16(step2[1], step2[14]); - out[2] = vaddq_s16(step2[2], step2[13]); - out[3] = vaddq_s16(step2[3], step2[12]); - out[4] = vaddq_s16(step2[4], step2[11]); - out[5] = vaddq_s16(step2[5], step2[10]); - out[6] = vaddq_s16(step2[6], step2[9]); - out[7] = vaddq_s16(step2[7], step2[8]); - out[8] = vsubq_s16(step2[7], step2[8]); - out[9] = vsubq_s16(step2[6], step2[9]); - out[10] = vsubq_s16(step2[5], step2[10]); - out[11] = vsubq_s16(step2[4], step2[11]); - out[12] = vsubq_s16(step2[3], step2[12]); - out[13] = vsubq_s16(step2[2], step2[13]); - out[14] = vsubq_s16(step2[1], step2[14]); - out[15] = vsubq_s16(step2[0], step2[15]); + idct16x16_add_stage7(step2, out); if (output) { idct16x16_store_pass1(out, output); } else { - idct16x16_add_store(out, dest, stride); + if (highbd_flag) { + idct16x16_add_store_bd8(out, dest, stride); + } else { + idct16x16_add_store(out, dest, stride); + } } } @@ -821,9 +810,9 @@ void vpx_idct16x16_10_add_neon(const tran_low_t *input, uint8_t *dest, // pass 2 // Parallel idct to get the left 8 columns - idct16x16_10_add_half1d_pass2(row_idct_output, NULL, dest, stride); + idct16x16_10_add_half1d_pass2(row_idct_output, NULL, dest, stride, 0); // Parallel idct to get the right 8 columns - idct16x16_10_add_half1d_pass2(row_idct_output + 4 * 8, NULL, dest + 8, - stride); + idct16x16_10_add_half1d_pass2(row_idct_output + 4 * 8, NULL, dest + 8, stride, + 0); } diff --git a/vpx_dsp/arm/idct_neon.h b/vpx_dsp/arm/idct_neon.h index 0a2052ced..fe5b603e2 100644 --- a/vpx_dsp/arm/idct_neon.h +++ b/vpx_dsp/arm/idct_neon.h @@ -752,4 +752,10 @@ void idct16x16_38_add_half1d(const void *const input, int16_t *const output, void *const dest, const int stride, const int highbd_flag); +void idct16x16_10_add_half1d_pass1(const tran_low_t *input, int16_t *output); + +void idct16x16_10_add_half1d_pass2(const int16_t *input, int16_t *const output, + void *const dest, const int stride, + const int highbd_flag); + #endif // VPX_DSP_ARM_IDCT_NEON_H_ diff --git a/vpx_dsp/vpx_dsp_rtcd_defs.pl b/vpx_dsp/vpx_dsp_rtcd_defs.pl index 070614a15..ee973b932 100644 --- a/vpx_dsp/vpx_dsp_rtcd_defs.pl +++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl @@ -737,7 +737,7 @@ if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") { $vpx_highbd_idct16x16_38_add_sse2=vpx_highbd_idct16x16_256_add_sse2; add_proto qw/void vpx_highbd_idct16x16_10_add/, "const tran_low_t *input, uint8_t *dest, int stride, int bd"; - specialize qw/vpx_highbd_idct16x16_10_add sse2/; + specialize qw/vpx_highbd_idct16x16_10_add neon sse2/; } # CONFIG_EMULATE_HARDWARE } else { # Force C versions if CONFIG_EMULATE_HARDWARE is 1