From 2231669a83d01cfcea2b2c838c82c8ab830f45aa Mon Sep 17 00:00:00 2001 From: Linfeng Zhang Date: Wed, 3 May 2017 15:43:02 -0700 Subject: [PATCH] Split dsp/x86/inv_txfm_sse2.c Spin out highbd idct functions. BUG=webm:1412 Change-Id: I0cfe4117c00039b6778c59c022eee79ad089a2af --- vpx_dsp/vpx_dsp.mk | 8 + vpx_dsp/x86/highbd_idct16x16_add_sse2.c | 244 +++++++++++++ vpx_dsp/x86/highbd_idct32x32_add_sse2.c | 41 +++ vpx_dsp/x86/highbd_idct4x4_add_sse2.c | 129 +++++++ vpx_dsp/x86/highbd_idct8x8_add_sse2.c | 216 ++++++++++++ vpx_dsp/x86/highbd_inv_txfm_sse2.h | 33 ++ vpx_dsp/x86/inv_txfm_sse2.c | 595 +------------------------------- vpx_dsp/x86/transpose_sse2.h | 26 ++ 8 files changed, 698 insertions(+), 594 deletions(-) create mode 100644 vpx_dsp/x86/highbd_idct16x16_add_sse2.c create mode 100644 vpx_dsp/x86/highbd_idct32x32_add_sse2.c create mode 100644 vpx_dsp/x86/highbd_idct4x4_add_sse2.c create mode 100644 vpx_dsp/x86/highbd_idct8x8_add_sse2.c create mode 100644 vpx_dsp/x86/highbd_inv_txfm_sse2.h create mode 100644 vpx_dsp/x86/transpose_sse2.h diff --git a/vpx_dsp/vpx_dsp.mk b/vpx_dsp/vpx_dsp.mk index d1c9917..6daa583 100644 --- a/vpx_dsp/vpx_dsp.mk +++ b/vpx_dsp/vpx_dsp.mk @@ -231,6 +231,11 @@ DSP_SRCS-$(HAVE_NEON) += arm/highbd_idct32x32_add_neon.c DSP_SRCS-$(HAVE_NEON) += arm/highbd_idct32x32_34_add_neon.c DSP_SRCS-$(HAVE_NEON) += arm/highbd_idct32x32_135_add_neon.c DSP_SRCS-$(HAVE_NEON) += arm/highbd_idct32x32_1024_add_neon.c +DSP_SRCS-$(HAVE_SSE2) += x86/highbd_inv_txfm_sse2.h +DSP_SRCS-$(HAVE_SSE2) += x86/highbd_idct4x4_add_sse2.c +DSP_SRCS-$(HAVE_SSE2) += x86/highbd_idct8x8_add_sse2.c +DSP_SRCS-$(HAVE_SSE2) += x86/highbd_idct16x16_add_sse2.c +DSP_SRCS-$(HAVE_SSE2) += x86/highbd_idct32x32_add_sse2.c endif # !CONFIG_VP9_HIGHBITDEPTH ifeq ($(HAVE_NEON_ASM),yes) @@ -351,6 +356,9 @@ DSP_SRCS-$(HAVE_VSX) += ppc/types_vsx.h DSP_SRCS-$(HAVE_VSX) += ppc/transpose_vsx.h DSP_SRCS-$(HAVE_VSX) += ppc/bitdepth_conversion_vsx.h +# X86 utilities +DSP_SRCS-$(HAVE_SSE2) += x86/transpose_sse2.h + DSP_SRCS-no += $(DSP_SRCS_REMOVE-yes) DSP_SRCS-yes += vpx_dsp_rtcd.c diff --git a/vpx_dsp/x86/highbd_idct16x16_add_sse2.c b/vpx_dsp/x86/highbd_idct16x16_add_sse2.c new file mode 100644 index 0000000..f16e4d0 --- /dev/null +++ b/vpx_dsp/x86/highbd_idct16x16_add_sse2.c @@ -0,0 +1,244 @@ +/* + * Copyright (c) 2015 The WebM project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "./vpx_dsp_rtcd.h" +#include "vpx_dsp/x86/highbd_inv_txfm_sse2.h" +#include "vpx_dsp/x86/inv_txfm_sse2.h" +#include "vpx_dsp/x86/transpose_sse2.h" +#include "vpx_dsp/x86/txfm_common_sse2.h" + +void vpx_highbd_idct16x16_256_add_sse2(const tran_low_t *input, uint16_t *dest, + int stride, int bd) { + tran_low_t out[16 * 16]; + tran_low_t *outptr = out; + int i, j, test; + __m128i inptr[32]; + __m128i min_input, max_input, temp1, temp2, sign_bits; + const __m128i zero = _mm_set1_epi16(0); + const __m128i rounding = _mm_set1_epi16(32); + const __m128i max = _mm_set1_epi16(3155); + const __m128i min = _mm_set1_epi16(-3155); + int optimised_cols = 0; + + // Load input into __m128i & pack to 16 bits + for (i = 0; i < 16; i++) { + temp1 = _mm_loadu_si128((const __m128i *)(input + 16 * i)); + temp2 = _mm_loadu_si128((const __m128i *)(input + 16 * i + 4)); + inptr[i] = _mm_packs_epi32(temp1, temp2); + temp1 = _mm_loadu_si128((const __m128i *)(input + 16 * i + 8)); + temp2 = _mm_loadu_si128((const __m128i *)(input + 16 * i + 12)); + inptr[i + 16] = _mm_packs_epi32(temp1, temp2); + } + + // Find the min & max for the row transform + max_input = _mm_max_epi16(inptr[0], inptr[1]); + min_input = _mm_min_epi16(inptr[0], inptr[1]); + for (i = 2; i < 32; i++) { + max_input = _mm_max_epi16(max_input, inptr[i]); + min_input = _mm_min_epi16(min_input, inptr[i]); + } + max_input = _mm_cmpgt_epi16(max_input, max); + min_input = _mm_cmplt_epi16(min_input, min); + temp1 = _mm_or_si128(max_input, min_input); + test = _mm_movemask_epi8(temp1); + + if (!test) { + // Do the row transform + idct16_sse2(inptr, inptr + 16); + + // Find the min & max for the column transform + max_input = _mm_max_epi16(inptr[0], inptr[1]); + min_input = _mm_min_epi16(inptr[0], inptr[1]); + for (i = 2; i < 32; i++) { + max_input = _mm_max_epi16(max_input, inptr[i]); + min_input = _mm_min_epi16(min_input, inptr[i]); + } + max_input = _mm_cmpgt_epi16(max_input, max); + min_input = _mm_cmplt_epi16(min_input, min); + temp1 = _mm_or_si128(max_input, min_input); + test = _mm_movemask_epi8(temp1); + + if (test) { + array_transpose_16x16(inptr, inptr + 16); + for (i = 0; i < 16; i++) { + sign_bits = _mm_cmplt_epi16(inptr[i], zero); + temp1 = _mm_unpacklo_epi16(inptr[i], sign_bits); + temp2 = _mm_unpackhi_epi16(inptr[i], sign_bits); + _mm_storeu_si128((__m128i *)(outptr + 4 * (i * 4)), temp1); + _mm_storeu_si128((__m128i *)(outptr + 4 * (i * 4 + 1)), temp2); + sign_bits = _mm_cmplt_epi16(inptr[i + 16], zero); + temp1 = _mm_unpacklo_epi16(inptr[i + 16], sign_bits); + temp2 = _mm_unpackhi_epi16(inptr[i + 16], sign_bits); + _mm_storeu_si128((__m128i *)(outptr + 4 * (i * 4 + 2)), temp1); + _mm_storeu_si128((__m128i *)(outptr + 4 * (i * 4 + 3)), temp2); + } + } else { + // Set to use the optimised transform for the column + optimised_cols = 1; + } + } else { + // Run the un-optimised row transform + for (i = 0; i < 16; ++i) { + vpx_highbd_idct16_c(input, outptr, bd); + input += 16; + outptr += 16; + } + } + + if (optimised_cols) { + idct16_sse2(inptr, inptr + 16); + + // Final round & shift and Reconstruction and Store + { + __m128i d[2]; + for (i = 0; i < 16; i++) { + inptr[i] = _mm_add_epi16(inptr[i], rounding); + inptr[i + 16] = _mm_add_epi16(inptr[i + 16], rounding); + d[0] = _mm_loadu_si128((const __m128i *)(dest + stride * i)); + d[1] = _mm_loadu_si128((const __m128i *)(dest + stride * i + 8)); + inptr[i] = _mm_srai_epi16(inptr[i], 6); + inptr[i + 16] = _mm_srai_epi16(inptr[i + 16], 6); + d[0] = clamp_high_sse2(_mm_add_epi16(d[0], inptr[i]), bd); + d[1] = clamp_high_sse2(_mm_add_epi16(d[1], inptr[i + 16]), bd); + // Store + _mm_storeu_si128((__m128i *)(dest + stride * i), d[0]); + _mm_storeu_si128((__m128i *)(dest + stride * i + 8), d[1]); + } + } + } else { + // Run the un-optimised column transform + tran_low_t temp_in[16], temp_out[16]; + for (i = 0; i < 16; ++i) { + for (j = 0; j < 16; ++j) temp_in[j] = out[j * 16 + i]; + vpx_highbd_idct16_c(temp_in, temp_out, bd); + for (j = 0; j < 16; ++j) { + dest[j * stride + i] = highbd_clip_pixel_add( + dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 6), bd); + } + } + } +} + +void vpx_highbd_idct16x16_10_add_sse2(const tran_low_t *input, uint16_t *dest, + int stride, int bd) { + tran_low_t out[16 * 16] = { 0 }; + tran_low_t *outptr = out; + int i, j, test; + __m128i inptr[32]; + __m128i min_input, max_input, temp1, temp2, sign_bits; + const __m128i zero = _mm_set1_epi16(0); + const __m128i rounding = _mm_set1_epi16(32); + const __m128i max = _mm_set1_epi16(3155); + const __m128i min = _mm_set1_epi16(-3155); + int optimised_cols = 0; + + // Load input into __m128i & pack to 16 bits + for (i = 0; i < 16; i++) { + temp1 = _mm_loadu_si128((const __m128i *)(input + 16 * i)); + temp2 = _mm_loadu_si128((const __m128i *)(input + 16 * i + 4)); + inptr[i] = _mm_packs_epi32(temp1, temp2); + temp1 = _mm_loadu_si128((const __m128i *)(input + 16 * i + 8)); + temp2 = _mm_loadu_si128((const __m128i *)(input + 16 * i + 12)); + inptr[i + 16] = _mm_packs_epi32(temp1, temp2); + } + + // Find the min & max for the row transform + // Since all non-zero dct coefficients are in upper-left 4x4 area, + // we only need to consider first 4 rows here. + max_input = _mm_max_epi16(inptr[0], inptr[1]); + min_input = _mm_min_epi16(inptr[0], inptr[1]); + for (i = 2; i < 4; i++) { + max_input = _mm_max_epi16(max_input, inptr[i]); + min_input = _mm_min_epi16(min_input, inptr[i]); + } + max_input = _mm_cmpgt_epi16(max_input, max); + min_input = _mm_cmplt_epi16(min_input, min); + temp1 = _mm_or_si128(max_input, min_input); + test = _mm_movemask_epi8(temp1); + + if (!test) { + // Do the row transform (N.B. This transposes inptr) + idct16_sse2(inptr, inptr + 16); + + // Find the min & max for the column transform + // N.B. Only first 4 cols contain non-zero coeffs + max_input = _mm_max_epi16(inptr[0], inptr[1]); + min_input = _mm_min_epi16(inptr[0], inptr[1]); + for (i = 2; i < 16; i++) { + max_input = _mm_max_epi16(max_input, inptr[i]); + min_input = _mm_min_epi16(min_input, inptr[i]); + } + max_input = _mm_cmpgt_epi16(max_input, max); + min_input = _mm_cmplt_epi16(min_input, min); + temp1 = _mm_or_si128(max_input, min_input); + test = _mm_movemask_epi8(temp1); + + if (test) { + // Use fact only first 4 rows contain non-zero coeffs + array_transpose_8x8(inptr, inptr); + array_transpose_8x8(inptr + 8, inptr + 16); + for (i = 0; i < 4; i++) { + sign_bits = _mm_cmplt_epi16(inptr[i], zero); + temp1 = _mm_unpacklo_epi16(inptr[i], sign_bits); + temp2 = _mm_unpackhi_epi16(inptr[i], sign_bits); + _mm_storeu_si128((__m128i *)(outptr + 4 * (i * 4)), temp1); + _mm_storeu_si128((__m128i *)(outptr + 4 * (i * 4 + 1)), temp2); + sign_bits = _mm_cmplt_epi16(inptr[i + 16], zero); + temp1 = _mm_unpacklo_epi16(inptr[i + 16], sign_bits); + temp2 = _mm_unpackhi_epi16(inptr[i + 16], sign_bits); + _mm_storeu_si128((__m128i *)(outptr + 4 * (i * 4 + 2)), temp1); + _mm_storeu_si128((__m128i *)(outptr + 4 * (i * 4 + 3)), temp2); + } + } else { + // Set to use the optimised transform for the column + optimised_cols = 1; + } + } else { + // Run the un-optimised row transform + for (i = 0; i < 4; ++i) { + vpx_highbd_idct16_c(input, outptr, bd); + input += 16; + outptr += 16; + } + } + + if (optimised_cols) { + idct16_sse2(inptr, inptr + 16); + + // Final round & shift and Reconstruction and Store + { + __m128i d[2]; + for (i = 0; i < 16; i++) { + inptr[i] = _mm_add_epi16(inptr[i], rounding); + inptr[i + 16] = _mm_add_epi16(inptr[i + 16], rounding); + d[0] = _mm_loadu_si128((const __m128i *)(dest + stride * i)); + d[1] = _mm_loadu_si128((const __m128i *)(dest + stride * i + 8)); + inptr[i] = _mm_srai_epi16(inptr[i], 6); + inptr[i + 16] = _mm_srai_epi16(inptr[i + 16], 6); + d[0] = clamp_high_sse2(_mm_add_epi16(d[0], inptr[i]), bd); + d[1] = clamp_high_sse2(_mm_add_epi16(d[1], inptr[i + 16]), bd); + // Store + _mm_storeu_si128((__m128i *)(dest + stride * i), d[0]); + _mm_storeu_si128((__m128i *)(dest + stride * i + 8), d[1]); + } + } + } else { + // Run the un-optimised column transform + tran_low_t temp_in[16], temp_out[16]; + for (i = 0; i < 16; ++i) { + for (j = 0; j < 16; ++j) temp_in[j] = out[j * 16 + i]; + vpx_highbd_idct16_c(temp_in, temp_out, bd); + for (j = 0; j < 16; ++j) { + dest[j * stride + i] = highbd_clip_pixel_add( + dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 6), bd); + } + } + } +} diff --git a/vpx_dsp/x86/highbd_idct32x32_add_sse2.c b/vpx_dsp/x86/highbd_idct32x32_add_sse2.c new file mode 100644 index 0000000..bc9debf --- /dev/null +++ b/vpx_dsp/x86/highbd_idct32x32_add_sse2.c @@ -0,0 +1,41 @@ +/* + * Copyright (c) 2015 The WebM project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "./vpx_dsp_rtcd.h" +#include "vpx_dsp/x86/inv_txfm_sse2.h" +#include "vpx_dsp/x86/transpose_sse2.h" +#include "vpx_dsp/x86/txfm_common_sse2.h" + +void vpx_highbd_idct32x32_1_add_sse2(const tran_low_t *input, uint16_t *dest, + int stride, int bd) { + __m128i dc_value, d; + const __m128i zero = _mm_setzero_si128(); + const __m128i one = _mm_set1_epi16(1); + const __m128i max = _mm_sub_epi16(_mm_slli_epi16(one, bd), one); + int a, i, j; + tran_low_t out; + + out = HIGHBD_WRAPLOW(dct_const_round_shift(input[0] * cospi_16_64), bd); + out = HIGHBD_WRAPLOW(dct_const_round_shift(out * cospi_16_64), bd); + a = ROUND_POWER_OF_TWO(out, 6); + + d = _mm_set1_epi32(a); + dc_value = _mm_packs_epi32(d, d); + for (i = 0; i < 32; ++i) { + for (j = 0; j < 4; ++j) { + d = _mm_loadu_si128((const __m128i *)(&dest[j * 8])); + d = _mm_adds_epi16(d, dc_value); + d = _mm_max_epi16(d, zero); + d = _mm_min_epi16(d, max); + _mm_storeu_si128((__m128i *)(&dest[j * 8]), d); + } + dest += stride; + } +} diff --git a/vpx_dsp/x86/highbd_idct4x4_add_sse2.c b/vpx_dsp/x86/highbd_idct4x4_add_sse2.c new file mode 100644 index 0000000..f3a6979 --- /dev/null +++ b/vpx_dsp/x86/highbd_idct4x4_add_sse2.c @@ -0,0 +1,129 @@ +/* + * Copyright (c) 2015 The WebM project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "./vpx_dsp_rtcd.h" +#include "vpx_dsp/x86/highbd_inv_txfm_sse2.h" +#include "vpx_dsp/x86/inv_txfm_sse2.h" +#include "vpx_dsp/x86/transpose_sse2.h" +#include "vpx_dsp/x86/txfm_common_sse2.h" + +void vpx_highbd_idct4x4_16_add_sse2(const tran_low_t *input, uint16_t *dest, + int stride, int bd) { + tran_low_t out[4 * 4]; + tran_low_t *outptr = out; + int i, j; + __m128i inptr[4]; + __m128i sign_bits[2]; + __m128i temp_mm, min_input, max_input; + int test; + int optimised_cols = 0; + const __m128i zero = _mm_set1_epi16(0); + const __m128i eight = _mm_set1_epi16(8); + const __m128i max = _mm_set1_epi16(12043); + const __m128i min = _mm_set1_epi16(-12043); + // Load input into __m128i + inptr[0] = _mm_loadu_si128((const __m128i *)input); + inptr[1] = _mm_loadu_si128((const __m128i *)(input + 4)); + inptr[2] = _mm_loadu_si128((const __m128i *)(input + 8)); + inptr[3] = _mm_loadu_si128((const __m128i *)(input + 12)); + + // Pack to 16 bits + inptr[0] = _mm_packs_epi32(inptr[0], inptr[1]); + inptr[1] = _mm_packs_epi32(inptr[2], inptr[3]); + + max_input = _mm_max_epi16(inptr[0], inptr[1]); + min_input = _mm_min_epi16(inptr[0], inptr[1]); + max_input = _mm_cmpgt_epi16(max_input, max); + min_input = _mm_cmplt_epi16(min_input, min); + temp_mm = _mm_or_si128(max_input, min_input); + test = _mm_movemask_epi8(temp_mm); + + if (!test) { + // Do the row transform + idct4_sse2(inptr); + + // Check the min & max values + max_input = _mm_max_epi16(inptr[0], inptr[1]); + min_input = _mm_min_epi16(inptr[0], inptr[1]); + max_input = _mm_cmpgt_epi16(max_input, max); + min_input = _mm_cmplt_epi16(min_input, min); + temp_mm = _mm_or_si128(max_input, min_input); + test = _mm_movemask_epi8(temp_mm); + + if (test) { + transpose_4x4(inptr); + sign_bits[0] = _mm_cmplt_epi16(inptr[0], zero); + sign_bits[1] = _mm_cmplt_epi16(inptr[1], zero); + inptr[3] = _mm_unpackhi_epi16(inptr[1], sign_bits[1]); + inptr[2] = _mm_unpacklo_epi16(inptr[1], sign_bits[1]); + inptr[1] = _mm_unpackhi_epi16(inptr[0], sign_bits[0]); + inptr[0] = _mm_unpacklo_epi16(inptr[0], sign_bits[0]); + _mm_storeu_si128((__m128i *)outptr, inptr[0]); + _mm_storeu_si128((__m128i *)(outptr + 4), inptr[1]); + _mm_storeu_si128((__m128i *)(outptr + 8), inptr[2]); + _mm_storeu_si128((__m128i *)(outptr + 12), inptr[3]); + } else { + // Set to use the optimised transform for the column + optimised_cols = 1; + } + } else { + // Run the un-optimised row transform + for (i = 0; i < 4; ++i) { + vpx_highbd_idct4_c(input, outptr, bd); + input += 4; + outptr += 4; + } + } + + if (optimised_cols) { + idct4_sse2(inptr); + + // Final round and shift + inptr[0] = _mm_add_epi16(inptr[0], eight); + inptr[1] = _mm_add_epi16(inptr[1], eight); + + inptr[0] = _mm_srai_epi16(inptr[0], 4); + inptr[1] = _mm_srai_epi16(inptr[1], 4); + + // Reconstruction and Store + { + __m128i d0 = _mm_loadl_epi64((const __m128i *)dest); + __m128i d2 = _mm_loadl_epi64((const __m128i *)(dest + stride * 2)); + d0 = _mm_unpacklo_epi64( + d0, _mm_loadl_epi64((const __m128i *)(dest + stride))); + d2 = _mm_unpacklo_epi64( + d2, _mm_loadl_epi64((const __m128i *)(dest + stride * 3))); + d0 = clamp_high_sse2(_mm_adds_epi16(d0, inptr[0]), bd); + d2 = clamp_high_sse2(_mm_adds_epi16(d2, inptr[1]), bd); + // store input0 + _mm_storel_epi64((__m128i *)dest, d0); + // store input1 + d0 = _mm_srli_si128(d0, 8); + _mm_storel_epi64((__m128i *)(dest + stride), d0); + // store input2 + _mm_storel_epi64((__m128i *)(dest + stride * 2), d2); + // store input3 + d2 = _mm_srli_si128(d2, 8); + _mm_storel_epi64((__m128i *)(dest + stride * 3), d2); + } + } else { + // Run the un-optimised column transform + tran_low_t temp_in[4], temp_out[4]; + // Columns + for (i = 0; i < 4; ++i) { + for (j = 0; j < 4; ++j) temp_in[j] = out[j * 4 + i]; + vpx_highbd_idct4_c(temp_in, temp_out, bd); + for (j = 0; j < 4; ++j) { + dest[j * stride + i] = highbd_clip_pixel_add( + dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 4), bd); + } + } + } +} diff --git a/vpx_dsp/x86/highbd_idct8x8_add_sse2.c b/vpx_dsp/x86/highbd_idct8x8_add_sse2.c new file mode 100644 index 0000000..6a2e180 --- /dev/null +++ b/vpx_dsp/x86/highbd_idct8x8_add_sse2.c @@ -0,0 +1,216 @@ +/* + * Copyright (c) 2015 The WebM project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "./vpx_dsp_rtcd.h" +#include "vpx_dsp/x86/highbd_inv_txfm_sse2.h" +#include "vpx_dsp/x86/inv_txfm_sse2.h" +#include "vpx_dsp/x86/transpose_sse2.h" +#include "vpx_dsp/x86/txfm_common_sse2.h" + +void vpx_highbd_idct8x8_64_add_sse2(const tran_low_t *input, uint16_t *dest, + int stride, int bd) { + tran_low_t out[8 * 8]; + tran_low_t *outptr = out; + int i, j, test; + __m128i inptr[8]; + __m128i min_input, max_input, temp1, temp2, sign_bits; + const __m128i zero = _mm_set1_epi16(0); + const __m128i sixteen = _mm_set1_epi16(16); + const __m128i max = _mm_set1_epi16(6201); + const __m128i min = _mm_set1_epi16(-6201); + int optimised_cols = 0; + + // Load input into __m128i & pack to 16 bits + for (i = 0; i < 8; i++) { + temp1 = _mm_loadu_si128((const __m128i *)(input + 8 * i)); + temp2 = _mm_loadu_si128((const __m128i *)(input + 8 * i + 4)); + inptr[i] = _mm_packs_epi32(temp1, temp2); + } + + // Find the min & max for the row transform + max_input = _mm_max_epi16(inptr[0], inptr[1]); + min_input = _mm_min_epi16(inptr[0], inptr[1]); + for (i = 2; i < 8; i++) { + max_input = _mm_max_epi16(max_input, inptr[i]); + min_input = _mm_min_epi16(min_input, inptr[i]); + } + max_input = _mm_cmpgt_epi16(max_input, max); + min_input = _mm_cmplt_epi16(min_input, min); + temp1 = _mm_or_si128(max_input, min_input); + test = _mm_movemask_epi8(temp1); + + if (!test) { + // Do the row transform + idct8_sse2(inptr); + + // Find the min & max for the column transform + max_input = _mm_max_epi16(inptr[0], inptr[1]); + min_input = _mm_min_epi16(inptr[0], inptr[1]); + for (i = 2; i < 8; i++) { + max_input = _mm_max_epi16(max_input, inptr[i]); + min_input = _mm_min_epi16(min_input, inptr[i]); + } + max_input = _mm_cmpgt_epi16(max_input, max); + min_input = _mm_cmplt_epi16(min_input, min); + temp1 = _mm_or_si128(max_input, min_input); + test = _mm_movemask_epi8(temp1); + + if (test) { + array_transpose_8x8(inptr, inptr); + for (i = 0; i < 8; i++) { + sign_bits = _mm_cmplt_epi16(inptr[i], zero); + temp1 = _mm_unpackhi_epi16(inptr[i], sign_bits); + temp2 = _mm_unpacklo_epi16(inptr[i], sign_bits); + _mm_storeu_si128((__m128i *)(outptr + 4 * (2 * i + 1)), temp1); + _mm_storeu_si128((__m128i *)(outptr + 4 * (2 * i)), temp2); + } + } else { + // Set to use the optimised transform for the column + optimised_cols = 1; + } + } else { + // Run the un-optimised row transform + for (i = 0; i < 8; ++i) { + vpx_highbd_idct8_c(input, outptr, bd); + input += 8; + outptr += 8; + } + } + + if (optimised_cols) { + idct8_sse2(inptr); + + // Final round & shift and Reconstruction and Store + { + __m128i d[8]; + for (i = 0; i < 8; i++) { + inptr[i] = _mm_add_epi16(inptr[i], sixteen); + d[i] = _mm_loadu_si128((const __m128i *)(dest + stride * i)); + inptr[i] = _mm_srai_epi16(inptr[i], 5); + d[i] = clamp_high_sse2(_mm_adds_epi16(d[i], inptr[i]), bd); + // Store + _mm_storeu_si128((__m128i *)(dest + stride * i), d[i]); + } + } + } else { + // Run the un-optimised column transform + tran_low_t temp_in[8], temp_out[8]; + for (i = 0; i < 8; ++i) { + for (j = 0; j < 8; ++j) temp_in[j] = out[j * 8 + i]; + vpx_highbd_idct8_c(temp_in, temp_out, bd); + for (j = 0; j < 8; ++j) { + dest[j * stride + i] = highbd_clip_pixel_add( + dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 5), bd); + } + } + } +} + +void vpx_highbd_idct8x8_12_add_sse2(const tran_low_t *input, uint16_t *dest, + int stride, int bd) { + tran_low_t out[8 * 8] = { 0 }; + tran_low_t *outptr = out; + int i, j, test; + __m128i inptr[8]; + __m128i min_input, max_input, temp1, temp2, sign_bits; + const __m128i zero = _mm_set1_epi16(0); + const __m128i sixteen = _mm_set1_epi16(16); + const __m128i max = _mm_set1_epi16(6201); + const __m128i min = _mm_set1_epi16(-6201); + int optimised_cols = 0; + + // Load input into __m128i & pack to 16 bits + for (i = 0; i < 8; i++) { + temp1 = _mm_loadu_si128((const __m128i *)(input + 8 * i)); + temp2 = _mm_loadu_si128((const __m128i *)(input + 8 * i + 4)); + inptr[i] = _mm_packs_epi32(temp1, temp2); + } + + // Find the min & max for the row transform + // only first 4 row has non-zero coefs + max_input = _mm_max_epi16(inptr[0], inptr[1]); + min_input = _mm_min_epi16(inptr[0], inptr[1]); + for (i = 2; i < 4; i++) { + max_input = _mm_max_epi16(max_input, inptr[i]); + min_input = _mm_min_epi16(min_input, inptr[i]); + } + max_input = _mm_cmpgt_epi16(max_input, max); + min_input = _mm_cmplt_epi16(min_input, min); + temp1 = _mm_or_si128(max_input, min_input); + test = _mm_movemask_epi8(temp1); + + if (!test) { + // Do the row transform + idct8_sse2(inptr); + + // Find the min & max for the column transform + // N.B. Only first 4 cols contain non-zero coeffs + max_input = _mm_max_epi16(inptr[0], inptr[1]); + min_input = _mm_min_epi16(inptr[0], inptr[1]); + for (i = 2; i < 8; i++) { + max_input = _mm_max_epi16(max_input, inptr[i]); + min_input = _mm_min_epi16(min_input, inptr[i]); + } + max_input = _mm_cmpgt_epi16(max_input, max); + min_input = _mm_cmplt_epi16(min_input, min); + temp1 = _mm_or_si128(max_input, min_input); + test = _mm_movemask_epi8(temp1); + + if (test) { + // Use fact only first 4 rows contain non-zero coeffs + array_transpose_4X8(inptr, inptr); + for (i = 0; i < 4; i++) { + sign_bits = _mm_cmplt_epi16(inptr[i], zero); + temp1 = _mm_unpackhi_epi16(inptr[i], sign_bits); + temp2 = _mm_unpacklo_epi16(inptr[i], sign_bits); + _mm_storeu_si128((__m128i *)(outptr + 4 * (2 * i + 1)), temp1); + _mm_storeu_si128((__m128i *)(outptr + 4 * (2 * i)), temp2); + } + } else { + // Set to use the optimised transform for the column + optimised_cols = 1; + } + } else { + // Run the un-optimised row transform + for (i = 0; i < 4; ++i) { + vpx_highbd_idct8_c(input, outptr, bd); + input += 8; + outptr += 8; + } + } + + if (optimised_cols) { + idct8_sse2(inptr); + + // Final round & shift and Reconstruction and Store + { + __m128i d[8]; + for (i = 0; i < 8; i++) { + inptr[i] = _mm_add_epi16(inptr[i], sixteen); + d[i] = _mm_loadu_si128((const __m128i *)(dest + stride * i)); + inptr[i] = _mm_srai_epi16(inptr[i], 5); + d[i] = clamp_high_sse2(_mm_adds_epi16(d[i], inptr[i]), bd); + // Store + _mm_storeu_si128((__m128i *)(dest + stride * i), d[i]); + } + } + } else { + // Run the un-optimised column transform + tran_low_t temp_in[8], temp_out[8]; + for (i = 0; i < 8; ++i) { + for (j = 0; j < 8; ++j) temp_in[j] = out[j * 8 + i]; + vpx_highbd_idct8_c(temp_in, temp_out, bd); + for (j = 0; j < 8; ++j) { + dest[j * stride + i] = highbd_clip_pixel_add( + dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 5), bd); + } + } + } +} diff --git a/vpx_dsp/x86/highbd_inv_txfm_sse2.h b/vpx_dsp/x86/highbd_inv_txfm_sse2.h new file mode 100644 index 0000000..774cce1 --- /dev/null +++ b/vpx_dsp/x86/highbd_inv_txfm_sse2.h @@ -0,0 +1,33 @@ +/* + * Copyright (c) 2015 The WebM project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#ifndef VPX_DSP_X86_HIGHBD_INV_TXFM_SSE2_H_ +#define VPX_DSP_X86_HIGHBD_INV_TXFM_SSE2_H_ + +#include // SSE2 +#include "./vpx_config.h" +#include "vpx/vpx_integer.h" +#include "vpx_dsp/inv_txfm.h" +#include "vpx_dsp/x86/txfm_common_sse2.h" + +static INLINE __m128i clamp_high_sse2(__m128i value, int bd) { + __m128i ubounded, retval; + const __m128i zero = _mm_set1_epi16(0); + const __m128i one = _mm_set1_epi16(1); + const __m128i max = _mm_sub_epi16(_mm_slli_epi16(one, bd), one); + ubounded = _mm_cmpgt_epi16(value, max); + retval = _mm_andnot_si128(ubounded, value); + ubounded = _mm_and_si128(ubounded, max); + retval = _mm_or_si128(retval, ubounded); + retval = _mm_and_si128(retval, _mm_cmpgt_epi16(retval, zero)); + return retval; +} + +#endif // VPX_DSP_X86_HIGHBD_INV_TXFM_SSE2_H_ diff --git a/vpx_dsp/x86/inv_txfm_sse2.c b/vpx_dsp/x86/inv_txfm_sse2.c index 81e09dc..4dd4dd8 100644 --- a/vpx_dsp/x86/inv_txfm_sse2.c +++ b/vpx_dsp/x86/inv_txfm_sse2.c @@ -10,6 +10,7 @@ #include "./vpx_dsp_rtcd.h" #include "vpx_dsp/x86/inv_txfm_sse2.h" +#include "vpx_dsp/x86/transpose_sse2.h" #include "vpx_dsp/x86/txfm_common_sse2.h" #define RECON_AND_STORE4X4(dest, in_x) \ @@ -170,14 +171,6 @@ void vpx_idct4x4_1_add_sse2(const tran_low_t *input, uint8_t *dest, RECON_AND_STORE4X4(dest + 3 * stride, dc_value); } -static INLINE void transpose_4x4(__m128i *res) { - const __m128i tr0_0 = _mm_unpacklo_epi16(res[0], res[1]); - const __m128i tr0_1 = _mm_unpackhi_epi16(res[0], res[1]); - - res[0] = _mm_unpacklo_epi16(tr0_0, tr0_1); - res[1] = _mm_unpackhi_epi16(tr0_0, tr0_1); -} - void idct4_sse2(__m128i *in) { const __m128i k__cospi_p16_p16 = pair_set_epi16(cospi_16_64, cospi_16_64); const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64); @@ -3349,589 +3342,3 @@ void vpx_idct32x32_1_add_sse2(const tran_low_t *input, uint8_t *dest, RECON_AND_STORE(dest + 24 + j * stride, dc_value); } } - -#if CONFIG_VP9_HIGHBITDEPTH -static INLINE __m128i clamp_high_sse2(__m128i value, int bd) { - __m128i ubounded, retval; - const __m128i zero = _mm_set1_epi16(0); - const __m128i one = _mm_set1_epi16(1); - const __m128i max = _mm_sub_epi16(_mm_slli_epi16(one, bd), one); - ubounded = _mm_cmpgt_epi16(value, max); - retval = _mm_andnot_si128(ubounded, value); - ubounded = _mm_and_si128(ubounded, max); - retval = _mm_or_si128(retval, ubounded); - retval = _mm_and_si128(retval, _mm_cmpgt_epi16(retval, zero)); - return retval; -} - -void vpx_highbd_idct4x4_16_add_sse2(const tran_low_t *input, uint16_t *dest, - int stride, int bd) { - tran_low_t out[4 * 4]; - tran_low_t *outptr = out; - int i, j; - __m128i inptr[4]; - __m128i sign_bits[2]; - __m128i temp_mm, min_input, max_input; - int test; - int optimised_cols = 0; - const __m128i zero = _mm_set1_epi16(0); - const __m128i eight = _mm_set1_epi16(8); - const __m128i max = _mm_set1_epi16(12043); - const __m128i min = _mm_set1_epi16(-12043); - // Load input into __m128i - inptr[0] = _mm_loadu_si128((const __m128i *)input); - inptr[1] = _mm_loadu_si128((const __m128i *)(input + 4)); - inptr[2] = _mm_loadu_si128((const __m128i *)(input + 8)); - inptr[3] = _mm_loadu_si128((const __m128i *)(input + 12)); - - // Pack to 16 bits - inptr[0] = _mm_packs_epi32(inptr[0], inptr[1]); - inptr[1] = _mm_packs_epi32(inptr[2], inptr[3]); - - max_input = _mm_max_epi16(inptr[0], inptr[1]); - min_input = _mm_min_epi16(inptr[0], inptr[1]); - max_input = _mm_cmpgt_epi16(max_input, max); - min_input = _mm_cmplt_epi16(min_input, min); - temp_mm = _mm_or_si128(max_input, min_input); - test = _mm_movemask_epi8(temp_mm); - - if (!test) { - // Do the row transform - idct4_sse2(inptr); - - // Check the min & max values - max_input = _mm_max_epi16(inptr[0], inptr[1]); - min_input = _mm_min_epi16(inptr[0], inptr[1]); - max_input = _mm_cmpgt_epi16(max_input, max); - min_input = _mm_cmplt_epi16(min_input, min); - temp_mm = _mm_or_si128(max_input, min_input); - test = _mm_movemask_epi8(temp_mm); - - if (test) { - transpose_4x4(inptr); - sign_bits[0] = _mm_cmplt_epi16(inptr[0], zero); - sign_bits[1] = _mm_cmplt_epi16(inptr[1], zero); - inptr[3] = _mm_unpackhi_epi16(inptr[1], sign_bits[1]); - inptr[2] = _mm_unpacklo_epi16(inptr[1], sign_bits[1]); - inptr[1] = _mm_unpackhi_epi16(inptr[0], sign_bits[0]); - inptr[0] = _mm_unpacklo_epi16(inptr[0], sign_bits[0]); - _mm_storeu_si128((__m128i *)outptr, inptr[0]); - _mm_storeu_si128((__m128i *)(outptr + 4), inptr[1]); - _mm_storeu_si128((__m128i *)(outptr + 8), inptr[2]); - _mm_storeu_si128((__m128i *)(outptr + 12), inptr[3]); - } else { - // Set to use the optimised transform for the column - optimised_cols = 1; - } - } else { - // Run the un-optimised row transform - for (i = 0; i < 4; ++i) { - vpx_highbd_idct4_c(input, outptr, bd); - input += 4; - outptr += 4; - } - } - - if (optimised_cols) { - idct4_sse2(inptr); - - // Final round and shift - inptr[0] = _mm_add_epi16(inptr[0], eight); - inptr[1] = _mm_add_epi16(inptr[1], eight); - - inptr[0] = _mm_srai_epi16(inptr[0], 4); - inptr[1] = _mm_srai_epi16(inptr[1], 4); - - // Reconstruction and Store - { - __m128i d0 = _mm_loadl_epi64((const __m128i *)dest); - __m128i d2 = _mm_loadl_epi64((const __m128i *)(dest + stride * 2)); - d0 = _mm_unpacklo_epi64( - d0, _mm_loadl_epi64((const __m128i *)(dest + stride))); - d2 = _mm_unpacklo_epi64( - d2, _mm_loadl_epi64((const __m128i *)(dest + stride * 3))); - d0 = clamp_high_sse2(_mm_adds_epi16(d0, inptr[0]), bd); - d2 = clamp_high_sse2(_mm_adds_epi16(d2, inptr[1]), bd); - // store input0 - _mm_storel_epi64((__m128i *)dest, d0); - // store input1 - d0 = _mm_srli_si128(d0, 8); - _mm_storel_epi64((__m128i *)(dest + stride), d0); - // store input2 - _mm_storel_epi64((__m128i *)(dest + stride * 2), d2); - // store input3 - d2 = _mm_srli_si128(d2, 8); - _mm_storel_epi64((__m128i *)(dest + stride * 3), d2); - } - } else { - // Run the un-optimised column transform - tran_low_t temp_in[4], temp_out[4]; - // Columns - for (i = 0; i < 4; ++i) { - for (j = 0; j < 4; ++j) temp_in[j] = out[j * 4 + i]; - vpx_highbd_idct4_c(temp_in, temp_out, bd); - for (j = 0; j < 4; ++j) { - dest[j * stride + i] = highbd_clip_pixel_add( - dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 4), bd); - } - } - } -} - -void vpx_highbd_idct8x8_64_add_sse2(const tran_low_t *input, uint16_t *dest, - int stride, int bd) { - tran_low_t out[8 * 8]; - tran_low_t *outptr = out; - int i, j, test; - __m128i inptr[8]; - __m128i min_input, max_input, temp1, temp2, sign_bits; - const __m128i zero = _mm_set1_epi16(0); - const __m128i sixteen = _mm_set1_epi16(16); - const __m128i max = _mm_set1_epi16(6201); - const __m128i min = _mm_set1_epi16(-6201); - int optimised_cols = 0; - - // Load input into __m128i & pack to 16 bits - for (i = 0; i < 8; i++) { - temp1 = _mm_loadu_si128((const __m128i *)(input + 8 * i)); - temp2 = _mm_loadu_si128((const __m128i *)(input + 8 * i + 4)); - inptr[i] = _mm_packs_epi32(temp1, temp2); - } - - // Find the min & max for the row transform - max_input = _mm_max_epi16(inptr[0], inptr[1]); - min_input = _mm_min_epi16(inptr[0], inptr[1]); - for (i = 2; i < 8; i++) { - max_input = _mm_max_epi16(max_input, inptr[i]); - min_input = _mm_min_epi16(min_input, inptr[i]); - } - max_input = _mm_cmpgt_epi16(max_input, max); - min_input = _mm_cmplt_epi16(min_input, min); - temp1 = _mm_or_si128(max_input, min_input); - test = _mm_movemask_epi8(temp1); - - if (!test) { - // Do the row transform - idct8_sse2(inptr); - - // Find the min & max for the column transform - max_input = _mm_max_epi16(inptr[0], inptr[1]); - min_input = _mm_min_epi16(inptr[0], inptr[1]); - for (i = 2; i < 8; i++) { - max_input = _mm_max_epi16(max_input, inptr[i]); - min_input = _mm_min_epi16(min_input, inptr[i]); - } - max_input = _mm_cmpgt_epi16(max_input, max); - min_input = _mm_cmplt_epi16(min_input, min); - temp1 = _mm_or_si128(max_input, min_input); - test = _mm_movemask_epi8(temp1); - - if (test) { - array_transpose_8x8(inptr, inptr); - for (i = 0; i < 8; i++) { - sign_bits = _mm_cmplt_epi16(inptr[i], zero); - temp1 = _mm_unpackhi_epi16(inptr[i], sign_bits); - temp2 = _mm_unpacklo_epi16(inptr[i], sign_bits); - _mm_storeu_si128((__m128i *)(outptr + 4 * (2 * i + 1)), temp1); - _mm_storeu_si128((__m128i *)(outptr + 4 * (2 * i)), temp2); - } - } else { - // Set to use the optimised transform for the column - optimised_cols = 1; - } - } else { - // Run the un-optimised row transform - for (i = 0; i < 8; ++i) { - vpx_highbd_idct8_c(input, outptr, bd); - input += 8; - outptr += 8; - } - } - - if (optimised_cols) { - idct8_sse2(inptr); - - // Final round & shift and Reconstruction and Store - { - __m128i d[8]; - for (i = 0; i < 8; i++) { - inptr[i] = _mm_add_epi16(inptr[i], sixteen); - d[i] = _mm_loadu_si128((const __m128i *)(dest + stride * i)); - inptr[i] = _mm_srai_epi16(inptr[i], 5); - d[i] = clamp_high_sse2(_mm_adds_epi16(d[i], inptr[i]), bd); - // Store - _mm_storeu_si128((__m128i *)(dest + stride * i), d[i]); - } - } - } else { - // Run the un-optimised column transform - tran_low_t temp_in[8], temp_out[8]; - for (i = 0; i < 8; ++i) { - for (j = 0; j < 8; ++j) temp_in[j] = out[j * 8 + i]; - vpx_highbd_idct8_c(temp_in, temp_out, bd); - for (j = 0; j < 8; ++j) { - dest[j * stride + i] = highbd_clip_pixel_add( - dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 5), bd); - } - } - } -} - -void vpx_highbd_idct8x8_12_add_sse2(const tran_low_t *input, uint16_t *dest, - int stride, int bd) { - tran_low_t out[8 * 8] = { 0 }; - tran_low_t *outptr = out; - int i, j, test; - __m128i inptr[8]; - __m128i min_input, max_input, temp1, temp2, sign_bits; - const __m128i zero = _mm_set1_epi16(0); - const __m128i sixteen = _mm_set1_epi16(16); - const __m128i max = _mm_set1_epi16(6201); - const __m128i min = _mm_set1_epi16(-6201); - int optimised_cols = 0; - - // Load input into __m128i & pack to 16 bits - for (i = 0; i < 8; i++) { - temp1 = _mm_loadu_si128((const __m128i *)(input + 8 * i)); - temp2 = _mm_loadu_si128((const __m128i *)(input + 8 * i + 4)); - inptr[i] = _mm_packs_epi32(temp1, temp2); - } - - // Find the min & max for the row transform - // only first 4 row has non-zero coefs - max_input = _mm_max_epi16(inptr[0], inptr[1]); - min_input = _mm_min_epi16(inptr[0], inptr[1]); - for (i = 2; i < 4; i++) { - max_input = _mm_max_epi16(max_input, inptr[i]); - min_input = _mm_min_epi16(min_input, inptr[i]); - } - max_input = _mm_cmpgt_epi16(max_input, max); - min_input = _mm_cmplt_epi16(min_input, min); - temp1 = _mm_or_si128(max_input, min_input); - test = _mm_movemask_epi8(temp1); - - if (!test) { - // Do the row transform - idct8_sse2(inptr); - - // Find the min & max for the column transform - // N.B. Only first 4 cols contain non-zero coeffs - max_input = _mm_max_epi16(inptr[0], inptr[1]); - min_input = _mm_min_epi16(inptr[0], inptr[1]); - for (i = 2; i < 8; i++) { - max_input = _mm_max_epi16(max_input, inptr[i]); - min_input = _mm_min_epi16(min_input, inptr[i]); - } - max_input = _mm_cmpgt_epi16(max_input, max); - min_input = _mm_cmplt_epi16(min_input, min); - temp1 = _mm_or_si128(max_input, min_input); - test = _mm_movemask_epi8(temp1); - - if (test) { - // Use fact only first 4 rows contain non-zero coeffs - array_transpose_4X8(inptr, inptr); - for (i = 0; i < 4; i++) { - sign_bits = _mm_cmplt_epi16(inptr[i], zero); - temp1 = _mm_unpackhi_epi16(inptr[i], sign_bits); - temp2 = _mm_unpacklo_epi16(inptr[i], sign_bits); - _mm_storeu_si128((__m128i *)(outptr + 4 * (2 * i + 1)), temp1); - _mm_storeu_si128((__m128i *)(outptr + 4 * (2 * i)), temp2); - } - } else { - // Set to use the optimised transform for the column - optimised_cols = 1; - } - } else { - // Run the un-optimised row transform - for (i = 0; i < 4; ++i) { - vpx_highbd_idct8_c(input, outptr, bd); - input += 8; - outptr += 8; - } - } - - if (optimised_cols) { - idct8_sse2(inptr); - - // Final round & shift and Reconstruction and Store - { - __m128i d[8]; - for (i = 0; i < 8; i++) { - inptr[i] = _mm_add_epi16(inptr[i], sixteen); - d[i] = _mm_loadu_si128((const __m128i *)(dest + stride * i)); - inptr[i] = _mm_srai_epi16(inptr[i], 5); - d[i] = clamp_high_sse2(_mm_adds_epi16(d[i], inptr[i]), bd); - // Store - _mm_storeu_si128((__m128i *)(dest + stride * i), d[i]); - } - } - } else { - // Run the un-optimised column transform - tran_low_t temp_in[8], temp_out[8]; - for (i = 0; i < 8; ++i) { - for (j = 0; j < 8; ++j) temp_in[j] = out[j * 8 + i]; - vpx_highbd_idct8_c(temp_in, temp_out, bd); - for (j = 0; j < 8; ++j) { - dest[j * stride + i] = highbd_clip_pixel_add( - dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 5), bd); - } - } - } -} - -void vpx_highbd_idct16x16_256_add_sse2(const tran_low_t *input, uint16_t *dest, - int stride, int bd) { - tran_low_t out[16 * 16]; - tran_low_t *outptr = out; - int i, j, test; - __m128i inptr[32]; - __m128i min_input, max_input, temp1, temp2, sign_bits; - const __m128i zero = _mm_set1_epi16(0); - const __m128i rounding = _mm_set1_epi16(32); - const __m128i max = _mm_set1_epi16(3155); - const __m128i min = _mm_set1_epi16(-3155); - int optimised_cols = 0; - - // Load input into __m128i & pack to 16 bits - for (i = 0; i < 16; i++) { - temp1 = _mm_loadu_si128((const __m128i *)(input + 16 * i)); - temp2 = _mm_loadu_si128((const __m128i *)(input + 16 * i + 4)); - inptr[i] = _mm_packs_epi32(temp1, temp2); - temp1 = _mm_loadu_si128((const __m128i *)(input + 16 * i + 8)); - temp2 = _mm_loadu_si128((const __m128i *)(input + 16 * i + 12)); - inptr[i + 16] = _mm_packs_epi32(temp1, temp2); - } - - // Find the min & max for the row transform - max_input = _mm_max_epi16(inptr[0], inptr[1]); - min_input = _mm_min_epi16(inptr[0], inptr[1]); - for (i = 2; i < 32; i++) { - max_input = _mm_max_epi16(max_input, inptr[i]); - min_input = _mm_min_epi16(min_input, inptr[i]); - } - max_input = _mm_cmpgt_epi16(max_input, max); - min_input = _mm_cmplt_epi16(min_input, min); - temp1 = _mm_or_si128(max_input, min_input); - test = _mm_movemask_epi8(temp1); - - if (!test) { - // Do the row transform - idct16_sse2(inptr, inptr + 16); - - // Find the min & max for the column transform - max_input = _mm_max_epi16(inptr[0], inptr[1]); - min_input = _mm_min_epi16(inptr[0], inptr[1]); - for (i = 2; i < 32; i++) { - max_input = _mm_max_epi16(max_input, inptr[i]); - min_input = _mm_min_epi16(min_input, inptr[i]); - } - max_input = _mm_cmpgt_epi16(max_input, max); - min_input = _mm_cmplt_epi16(min_input, min); - temp1 = _mm_or_si128(max_input, min_input); - test = _mm_movemask_epi8(temp1); - - if (test) { - array_transpose_16x16(inptr, inptr + 16); - for (i = 0; i < 16; i++) { - sign_bits = _mm_cmplt_epi16(inptr[i], zero); - temp1 = _mm_unpacklo_epi16(inptr[i], sign_bits); - temp2 = _mm_unpackhi_epi16(inptr[i], sign_bits); - _mm_storeu_si128((__m128i *)(outptr + 4 * (i * 4)), temp1); - _mm_storeu_si128((__m128i *)(outptr + 4 * (i * 4 + 1)), temp2); - sign_bits = _mm_cmplt_epi16(inptr[i + 16], zero); - temp1 = _mm_unpacklo_epi16(inptr[i + 16], sign_bits); - temp2 = _mm_unpackhi_epi16(inptr[i + 16], sign_bits); - _mm_storeu_si128((__m128i *)(outptr + 4 * (i * 4 + 2)), temp1); - _mm_storeu_si128((__m128i *)(outptr + 4 * (i * 4 + 3)), temp2); - } - } else { - // Set to use the optimised transform for the column - optimised_cols = 1; - } - } else { - // Run the un-optimised row transform - for (i = 0; i < 16; ++i) { - vpx_highbd_idct16_c(input, outptr, bd); - input += 16; - outptr += 16; - } - } - - if (optimised_cols) { - idct16_sse2(inptr, inptr + 16); - - // Final round & shift and Reconstruction and Store - { - __m128i d[2]; - for (i = 0; i < 16; i++) { - inptr[i] = _mm_add_epi16(inptr[i], rounding); - inptr[i + 16] = _mm_add_epi16(inptr[i + 16], rounding); - d[0] = _mm_loadu_si128((const __m128i *)(dest + stride * i)); - d[1] = _mm_loadu_si128((const __m128i *)(dest + stride * i + 8)); - inptr[i] = _mm_srai_epi16(inptr[i], 6); - inptr[i + 16] = _mm_srai_epi16(inptr[i + 16], 6); - d[0] = clamp_high_sse2(_mm_add_epi16(d[0], inptr[i]), bd); - d[1] = clamp_high_sse2(_mm_add_epi16(d[1], inptr[i + 16]), bd); - // Store - _mm_storeu_si128((__m128i *)(dest + stride * i), d[0]); - _mm_storeu_si128((__m128i *)(dest + stride * i + 8), d[1]); - } - } - } else { - // Run the un-optimised column transform - tran_low_t temp_in[16], temp_out[16]; - for (i = 0; i < 16; ++i) { - for (j = 0; j < 16; ++j) temp_in[j] = out[j * 16 + i]; - vpx_highbd_idct16_c(temp_in, temp_out, bd); - for (j = 0; j < 16; ++j) { - dest[j * stride + i] = highbd_clip_pixel_add( - dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 6), bd); - } - } - } -} - -void vpx_highbd_idct16x16_10_add_sse2(const tran_low_t *input, uint16_t *dest, - int stride, int bd) { - tran_low_t out[16 * 16] = { 0 }; - tran_low_t *outptr = out; - int i, j, test; - __m128i inptr[32]; - __m128i min_input, max_input, temp1, temp2, sign_bits; - const __m128i zero = _mm_set1_epi16(0); - const __m128i rounding = _mm_set1_epi16(32); - const __m128i max = _mm_set1_epi16(3155); - const __m128i min = _mm_set1_epi16(-3155); - int optimised_cols = 0; - - // Load input into __m128i & pack to 16 bits - for (i = 0; i < 16; i++) { - temp1 = _mm_loadu_si128((const __m128i *)(input + 16 * i)); - temp2 = _mm_loadu_si128((const __m128i *)(input + 16 * i + 4)); - inptr[i] = _mm_packs_epi32(temp1, temp2); - temp1 = _mm_loadu_si128((const __m128i *)(input + 16 * i + 8)); - temp2 = _mm_loadu_si128((const __m128i *)(input + 16 * i + 12)); - inptr[i + 16] = _mm_packs_epi32(temp1, temp2); - } - - // Find the min & max for the row transform - // Since all non-zero dct coefficients are in upper-left 4x4 area, - // we only need to consider first 4 rows here. - max_input = _mm_max_epi16(inptr[0], inptr[1]); - min_input = _mm_min_epi16(inptr[0], inptr[1]); - for (i = 2; i < 4; i++) { - max_input = _mm_max_epi16(max_input, inptr[i]); - min_input = _mm_min_epi16(min_input, inptr[i]); - } - max_input = _mm_cmpgt_epi16(max_input, max); - min_input = _mm_cmplt_epi16(min_input, min); - temp1 = _mm_or_si128(max_input, min_input); - test = _mm_movemask_epi8(temp1); - - if (!test) { - // Do the row transform (N.B. This transposes inptr) - idct16_sse2(inptr, inptr + 16); - - // Find the min & max for the column transform - // N.B. Only first 4 cols contain non-zero coeffs - max_input = _mm_max_epi16(inptr[0], inptr[1]); - min_input = _mm_min_epi16(inptr[0], inptr[1]); - for (i = 2; i < 16; i++) { - max_input = _mm_max_epi16(max_input, inptr[i]); - min_input = _mm_min_epi16(min_input, inptr[i]); - } - max_input = _mm_cmpgt_epi16(max_input, max); - min_input = _mm_cmplt_epi16(min_input, min); - temp1 = _mm_or_si128(max_input, min_input); - test = _mm_movemask_epi8(temp1); - - if (test) { - // Use fact only first 4 rows contain non-zero coeffs - array_transpose_8x8(inptr, inptr); - array_transpose_8x8(inptr + 8, inptr + 16); - for (i = 0; i < 4; i++) { - sign_bits = _mm_cmplt_epi16(inptr[i], zero); - temp1 = _mm_unpacklo_epi16(inptr[i], sign_bits); - temp2 = _mm_unpackhi_epi16(inptr[i], sign_bits); - _mm_storeu_si128((__m128i *)(outptr + 4 * (i * 4)), temp1); - _mm_storeu_si128((__m128i *)(outptr + 4 * (i * 4 + 1)), temp2); - sign_bits = _mm_cmplt_epi16(inptr[i + 16], zero); - temp1 = _mm_unpacklo_epi16(inptr[i + 16], sign_bits); - temp2 = _mm_unpackhi_epi16(inptr[i + 16], sign_bits); - _mm_storeu_si128((__m128i *)(outptr + 4 * (i * 4 + 2)), temp1); - _mm_storeu_si128((__m128i *)(outptr + 4 * (i * 4 + 3)), temp2); - } - } else { - // Set to use the optimised transform for the column - optimised_cols = 1; - } - } else { - // Run the un-optimised row transform - for (i = 0; i < 4; ++i) { - vpx_highbd_idct16_c(input, outptr, bd); - input += 16; - outptr += 16; - } - } - - if (optimised_cols) { - idct16_sse2(inptr, inptr + 16); - - // Final round & shift and Reconstruction and Store - { - __m128i d[2]; - for (i = 0; i < 16; i++) { - inptr[i] = _mm_add_epi16(inptr[i], rounding); - inptr[i + 16] = _mm_add_epi16(inptr[i + 16], rounding); - d[0] = _mm_loadu_si128((const __m128i *)(dest + stride * i)); - d[1] = _mm_loadu_si128((const __m128i *)(dest + stride * i + 8)); - inptr[i] = _mm_srai_epi16(inptr[i], 6); - inptr[i + 16] = _mm_srai_epi16(inptr[i + 16], 6); - d[0] = clamp_high_sse2(_mm_add_epi16(d[0], inptr[i]), bd); - d[1] = clamp_high_sse2(_mm_add_epi16(d[1], inptr[i + 16]), bd); - // Store - _mm_storeu_si128((__m128i *)(dest + stride * i), d[0]); - _mm_storeu_si128((__m128i *)(dest + stride * i + 8), d[1]); - } - } - } else { - // Run the un-optimised column transform - tran_low_t temp_in[16], temp_out[16]; - for (i = 0; i < 16; ++i) { - for (j = 0; j < 16; ++j) temp_in[j] = out[j * 16 + i]; - vpx_highbd_idct16_c(temp_in, temp_out, bd); - for (j = 0; j < 16; ++j) { - dest[j * stride + i] = highbd_clip_pixel_add( - dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 6), bd); - } - } - } -} - -void vpx_highbd_idct32x32_1_add_sse2(const tran_low_t *input, uint16_t *dest, - int stride, int bd) { - __m128i dc_value, d; - const __m128i zero = _mm_setzero_si128(); - const __m128i one = _mm_set1_epi16(1); - const __m128i max = _mm_sub_epi16(_mm_slli_epi16(one, bd), one); - int a, i, j; - tran_low_t out; - - out = HIGHBD_WRAPLOW(dct_const_round_shift(input[0] * cospi_16_64), bd); - out = HIGHBD_WRAPLOW(dct_const_round_shift(out * cospi_16_64), bd); - a = ROUND_POWER_OF_TWO(out, 6); - - d = _mm_set1_epi32(a); - dc_value = _mm_packs_epi32(d, d); - for (i = 0; i < 32; ++i) { - for (j = 0; j < 4; ++j) { - d = _mm_loadu_si128((const __m128i *)(&dest[j * 8])); - d = _mm_adds_epi16(d, dc_value); - d = _mm_max_epi16(d, zero); - d = _mm_min_epi16(d, max); - _mm_storeu_si128((__m128i *)(&dest[j * 8]), d); - } - dest += stride; - } -} -#endif // CONFIG_VP9_HIGHBITDEPTH diff --git a/vpx_dsp/x86/transpose_sse2.h b/vpx_dsp/x86/transpose_sse2.h new file mode 100644 index 0000000..7292723 --- /dev/null +++ b/vpx_dsp/x86/transpose_sse2.h @@ -0,0 +1,26 @@ +/* + * Copyright (c) 2015 The WebM project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#ifndef VPX_DSP_X86_TRANSPOSE_SSE2_H_ +#define VPX_DSP_X86_TRANSPOSE_SSE2_H_ + +#include "./vpx_dsp_rtcd.h" +#include "vpx_dsp/x86/inv_txfm_sse2.h" +#include "vpx_dsp/x86/txfm_common_sse2.h" + +static INLINE void transpose_4x4(__m128i *res) { + const __m128i tr0_0 = _mm_unpacklo_epi16(res[0], res[1]); + const __m128i tr0_1 = _mm_unpackhi_epi16(res[0], res[1]); + + res[0] = _mm_unpacklo_epi16(tr0_0, tr0_1); + res[1] = _mm_unpackhi_epi16(tr0_0, tr0_1); +} + +#endif // VPX_DSP_X86_TRANSPOSE_SSE2_H_ -- 2.7.4