ref: 45048dc9dc750da16a8c4df365f10dac4a55966d
parent: 48c0e13286020d6ce8a810197b6736ae4c1ee2dc
author: Linfeng Zhang <linfengz@google.com>
date: Mon May 22 12:04:05 EDT 2017
Update vpx_highbd_idct4x4_16_add_sse2() BUG=webm:1412 Change-Id: I26e4b34ae9bc1ae80c24f56d740d737a95f1ab84
--- a/vpx_dsp/x86/highbd_idct4x4_add_sse2.c
+++ b/vpx_dsp/x86/highbd_idct4x4_add_sse2.c
@@ -14,117 +14,227 @@
#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));
+static INLINE __m128i dct_const_round_shift_4_sse2(const __m128i in0,
+ const __m128i in1) {+ const __m128i t0 = _mm_unpacklo_epi32(in0, in1); // 0, 1
+ const __m128i t1 = _mm_unpackhi_epi32(in0, in1); // 2, 3
+ const __m128i t2 = _mm_unpacklo_epi64(t0, t1); // 0, 1, 2, 3
+ return dct_const_round_shift_sse2(t2);
+}
- // Pack to 16 bits
- inptr[0] = _mm_packs_epi32(inptr[0], inptr[1]);
- inptr[1] = _mm_packs_epi32(inptr[2], inptr[3]);
+static INLINE __m128i wraplow_16bit_sse2(const __m128i in0, const __m128i in1,
+ const __m128i rounding) {+ __m128i temp[2];
+ temp[0] = _mm_add_epi32(in0, rounding);
+ temp[1] = _mm_add_epi32(in1, rounding);
+ temp[0] = _mm_srai_epi32(temp[0], 4);
+ temp[1] = _mm_srai_epi32(temp[1], 4);
+ return _mm_packs_epi32(temp[0], temp[1]);
+}
- 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);
+static INLINE void highbd_idct4_small_sse2(__m128i *const io) {+ const __m128i cospi_p16_p16 = _mm_setr_epi32(cospi_16_64, 0, cospi_16_64, 0);
+ const __m128i cospi_p08_p08 = _mm_setr_epi32(cospi_8_64, 0, cospi_8_64, 0);
+ const __m128i cospi_p24_p24 = _mm_setr_epi32(cospi_24_64, 0, cospi_24_64, 0);
+ __m128i temp1[4], temp2[4], step[4];
- if (!test) {- // Do the row transform
- idct4_sse2(inptr);
+ transpose_32bit_4x4(&io[0], &io[1], &io[2], &io[3]);
- // 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);
+ // Note: There is no 32-bit signed multiply SIMD instruction in SSE2.
+ // _mm_mul_epu32() is used which can only guarantee the lower 32-bit
+ // (signed) result is meaningful, which is enough in this function.
- if (test) {- transpose_16bit_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;
- }
- }
+ // stage 1
+ temp1[0] = _mm_add_epi32(io[0], io[2]); // input[0] + input[2]
+ temp2[0] = _mm_sub_epi32(io[0], io[2]); // input[0] - input[2]
+ temp1[1] = _mm_srli_si128(temp1[0], 4); // 1, 3
+ temp2[1] = _mm_srli_si128(temp2[0], 4); // 1, 3
+ temp1[0] = _mm_mul_epu32(temp1[0], cospi_p16_p16); // ([0] + [2])*cospi_16_64
+ temp1[1] = _mm_mul_epu32(temp1[1], cospi_p16_p16); // ([0] + [2])*cospi_16_64
+ temp2[0] = _mm_mul_epu32(temp2[0], cospi_p16_p16); // ([0] - [2])*cospi_16_64
+ temp2[1] = _mm_mul_epu32(temp2[1], cospi_p16_p16); // ([0] - [2])*cospi_16_64
+ step[0] = dct_const_round_shift_4_sse2(temp1[0], temp1[1]);
+ step[1] = dct_const_round_shift_4_sse2(temp2[0], temp2[1]);
- if (optimised_cols) {- idct4_sse2(inptr);
+ temp1[3] = _mm_srli_si128(io[1], 4);
+ temp2[3] = _mm_srli_si128(io[3], 4);
+ temp1[0] = _mm_mul_epu32(io[1], cospi_p24_p24); // input[1] * cospi_24_64
+ temp1[1] = _mm_mul_epu32(temp1[3], cospi_p24_p24); // input[1] * cospi_24_64
+ temp2[0] = _mm_mul_epu32(io[1], cospi_p08_p08); // input[1] * cospi_8_64
+ temp2[1] = _mm_mul_epu32(temp1[3], cospi_p08_p08); // input[1] * cospi_8_64
+ temp1[2] = _mm_mul_epu32(io[3], cospi_p08_p08); // input[3] * cospi_8_64
+ temp1[3] = _mm_mul_epu32(temp2[3], cospi_p08_p08); // input[3] * cospi_8_64
+ temp2[2] = _mm_mul_epu32(io[3], cospi_p24_p24); // input[3] * cospi_24_64
+ temp2[3] = _mm_mul_epu32(temp2[3], cospi_p24_p24); // input[3] * cospi_24_64
+ temp1[0] = _mm_sub_epi64(temp1[0], temp1[2]); // [1]*cospi_24 - [3]*cospi_8
+ temp1[1] = _mm_sub_epi64(temp1[1], temp1[3]); // [1]*cospi_24 - [3]*cospi_8
+ temp2[0] = _mm_add_epi64(temp2[0], temp2[2]); // [1]*cospi_8 + [3]*cospi_24
+ temp2[1] = _mm_add_epi64(temp2[1], temp2[3]); // [1]*cospi_8 + [3]*cospi_24
+ step[2] = dct_const_round_shift_4_sse2(temp1[0], temp1[1]);
+ step[3] = dct_const_round_shift_4_sse2(temp2[0], temp2[1]);
- // Final round and shift
- inptr[0] = _mm_add_epi16(inptr[0], eight);
- inptr[1] = _mm_add_epi16(inptr[1], eight);
+ // stage 2
+ io[0] = _mm_add_epi32(step[0], step[3]); // step[0] + step[3]
+ io[1] = _mm_add_epi32(step[1], step[2]); // step[1] + step[2]
+ io[2] = _mm_sub_epi32(step[1], step[2]); // step[1] - step[2]
+ io[3] = _mm_sub_epi32(step[0], step[3]); // step[0] - step[3]
+}
- inptr[0] = _mm_srai_epi16(inptr[0], 4);
- inptr[1] = _mm_srai_epi16(inptr[1], 4);
+static INLINE void abs_extend_64bit_sse2(const __m128i in,
+ __m128i *const out /*out[2]*/,
+ __m128i *const sign /*sign[2]*/) {+ sign[0] = _mm_srai_epi32(in, 31);
+ out[0] = _mm_xor_si128(in, sign[0]);
+ out[0] = _mm_sub_epi32(out[0], sign[0]);
+ sign[1] = _mm_unpackhi_epi32(sign[0], sign[0]); // 64-bit sign of 2, 3
+ sign[0] = _mm_unpacklo_epi32(sign[0], sign[0]); // 64-bit sign of 0, 1
+ out[1] = _mm_unpackhi_epi32(out[0], out[0]); // 2, 3
+ out[0] = _mm_unpacklo_epi32(out[0], out[0]); // 0, 1
+}
- // 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);
- }
+static INLINE __m128i multiply_apply_sign_sse2(const __m128i in,
+ const __m128i sign,
+ const __m128i cospi) {+ __m128i out = _mm_mul_epu32(in, cospi);
+ out = _mm_xor_si128(out, sign);
+ return _mm_sub_epi64(out, sign);
+}
+
+static INLINE __m128i dct_const_round_shift_64bit_sse2(const __m128i in) {+ const __m128i t = _mm_add_epi64(
+ in,
+ _mm_setr_epi32(DCT_CONST_ROUNDING << 2, 0, DCT_CONST_ROUNDING << 2, 0));
+ return _mm_srli_si128(t, 2);
+}
+
+static INLINE __m128i pack_4_sse2(const __m128i in0, const __m128i in1) {+ const __m128i t0 = _mm_unpacklo_epi32(in0, in1); // 0, 2
+ const __m128i t1 = _mm_unpackhi_epi32(in0, in1); // 1, 3
+ return _mm_unpacklo_epi32(t0, t1); // 0, 1, 2, 3
+}
+
+static INLINE void highbd_idct4_large_sse2(__m128i *const io) {+ const __m128i cospi_p16_p16 =
+ _mm_setr_epi32(cospi_16_64 << 2, 0, cospi_16_64 << 2, 0);
+ const __m128i cospi_p08_p08 =
+ _mm_setr_epi32(cospi_8_64 << 2, 0, cospi_8_64 << 2, 0);
+ const __m128i cospi_p24_p24 =
+ _mm_setr_epi32(cospi_24_64 << 2, 0, cospi_24_64 << 2, 0);
+ __m128i temp1[4], temp2[4], step[4], sign1[4], sign2[4];
+
+ transpose_32bit_4x4(&io[0], &io[1], &io[2], &io[3]);
+
+ // stage 1
+ temp1[0] = _mm_add_epi32(io[0], io[2]); // input[0] + input[2]
+ temp2[0] = _mm_sub_epi32(io[0], io[2]); // input[0] - input[2]
+ abs_extend_64bit_sse2(temp1[0], temp1, sign1);
+ abs_extend_64bit_sse2(temp2[0], temp2, sign2);
+ temp1[0] = multiply_apply_sign_sse2(temp1[0], sign1[0], cospi_p16_p16);
+ temp1[1] = multiply_apply_sign_sse2(temp1[1], sign1[1], cospi_p16_p16);
+ temp2[0] = multiply_apply_sign_sse2(temp2[0], sign2[0], cospi_p16_p16);
+ temp2[1] = multiply_apply_sign_sse2(temp2[1], sign2[1], cospi_p16_p16);
+ temp1[0] = dct_const_round_shift_64bit_sse2(temp1[0]);
+ temp1[1] = dct_const_round_shift_64bit_sse2(temp1[1]);
+ temp2[0] = dct_const_round_shift_64bit_sse2(temp2[0]);
+ temp2[1] = dct_const_round_shift_64bit_sse2(temp2[1]);
+ step[0] = pack_4_sse2(temp1[0], temp1[1]);
+ step[1] = pack_4_sse2(temp2[0], temp2[1]);
+
+ abs_extend_64bit_sse2(io[1], temp1, sign1);
+ abs_extend_64bit_sse2(io[3], temp2, sign2);
+ temp1[2] = multiply_apply_sign_sse2(temp1[0], sign1[0], cospi_p08_p08);
+ temp1[3] = multiply_apply_sign_sse2(temp1[1], sign1[1], cospi_p08_p08);
+ temp1[0] = multiply_apply_sign_sse2(temp1[0], sign1[0], cospi_p24_p24);
+ temp1[1] = multiply_apply_sign_sse2(temp1[1], sign1[1], cospi_p24_p24);
+ temp2[2] = multiply_apply_sign_sse2(temp2[0], sign2[0], cospi_p24_p24);
+ temp2[3] = multiply_apply_sign_sse2(temp2[1], sign2[1], cospi_p24_p24);
+ temp2[0] = multiply_apply_sign_sse2(temp2[0], sign2[0], cospi_p08_p08);
+ temp2[1] = multiply_apply_sign_sse2(temp2[1], sign2[1], cospi_p08_p08);
+ temp1[0] = _mm_sub_epi64(temp1[0], temp2[0]); // [1]*cospi_24 - [3]*cospi_8
+ temp1[1] = _mm_sub_epi64(temp1[1], temp2[1]); // [1]*cospi_24 - [3]*cospi_8
+ temp2[0] = _mm_add_epi64(temp1[2], temp2[2]); // [1]*cospi_8 + [3]*cospi_24
+ temp2[1] = _mm_add_epi64(temp1[3], temp2[3]); // [1]*cospi_8 + [3]*cospi_24
+ temp1[0] = dct_const_round_shift_64bit_sse2(temp1[0]);
+ temp1[1] = dct_const_round_shift_64bit_sse2(temp1[1]);
+ temp2[0] = dct_const_round_shift_64bit_sse2(temp2[0]);
+ temp2[1] = dct_const_round_shift_64bit_sse2(temp2[1]);
+ step[2] = pack_4_sse2(temp1[0], temp1[1]);
+ step[3] = pack_4_sse2(temp2[0], temp2[1]);
+
+ // stage 2
+ io[0] = _mm_add_epi32(step[0], step[3]); // step[0] + step[3]
+ io[1] = _mm_add_epi32(step[1], step[2]); // step[1] + step[2]
+ io[2] = _mm_sub_epi32(step[1], step[2]); // step[1] - step[2]
+ io[3] = _mm_sub_epi32(step[0], step[3]); // step[0] - step[3]
+}
+
+void vpx_highbd_idct4x4_16_add_sse2(const tran_low_t *input, uint16_t *dest,
+ int stride, int bd) {+ int16_t max = 0, min = 0;
+ __m128i io[4], io_short[2];
+
+ io[0] = _mm_load_si128((const __m128i *)(input + 0));
+ io[1] = _mm_load_si128((const __m128i *)(input + 4));
+ io[2] = _mm_load_si128((const __m128i *)(input + 8));
+ io[3] = _mm_load_si128((const __m128i *)(input + 12));
+
+ io_short[0] = _mm_packs_epi32(io[0], io[1]);
+ io_short[1] = _mm_packs_epi32(io[2], io[3]);
+
+ if (bd != 8) {+ __m128i max_input, min_input;
+
+ max_input = _mm_max_epi16(io_short[0], io_short[1]);
+ min_input = _mm_min_epi16(io_short[0], io_short[1]);
+ max_input = _mm_max_epi16(max_input, _mm_srli_si128(max_input, 8));
+ min_input = _mm_min_epi16(min_input, _mm_srli_si128(min_input, 8));
+ max_input = _mm_max_epi16(max_input, _mm_srli_si128(max_input, 4));
+ min_input = _mm_min_epi16(min_input, _mm_srli_si128(min_input, 4));
+ max_input = _mm_max_epi16(max_input, _mm_srli_si128(max_input, 2));
+ min_input = _mm_min_epi16(min_input, _mm_srli_si128(min_input, 2));
+ max = _mm_extract_epi16(max_input, 0);
+ min = _mm_extract_epi16(min_input, 0);
+ }
+
+ if (bd == 8 || (max < 4096 && min >= -4096)) {+ idct4_sse2(io_short);
+ idct4_sse2(io_short);
+ io_short[0] = _mm_add_epi16(io_short[0], _mm_set1_epi16(8));
+ io_short[1] = _mm_add_epi16(io_short[1], _mm_set1_epi16(8));
+ io[0] = _mm_srai_epi16(io_short[0], 4);
+ io[1] = _mm_srai_epi16(io_short[1], 4);
} 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);
- }
+ if (max < 32767 && min > -32768) {+ highbd_idct4_small_sse2(io);
+ highbd_idct4_small_sse2(io);
+ } else {+ highbd_idct4_large_sse2(io);
+ highbd_idct4_large_sse2(io);
}
+ io[0] = wraplow_16bit_sse2(io[0], io[1], _mm_set1_epi32(8));
+ io[1] = wraplow_16bit_sse2(io[2], io[3], _mm_set1_epi32(8));
+ }
+
+ // 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, io[0]), bd);
+ d2 = clamp_high_sse2(_mm_adds_epi16(d2, io[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);
}
}