From c7e4917e974e30f292ca83f744b8c8bee417b461 Mon Sep 17 00:00:00 2001 From: Linfeng Zhang Date: Thu, 15 Jun 2017 11:29:07 -0700 Subject: [PATCH] Clean 8x8 idct x86 optimization Create load_buffer_8x8() and write_buffer_8x8(). Change-Id: Ib26dd515d734a5402971c91de336ab481b213fdf --- vpx_dsp/x86/inv_txfm_sse2.c | 74 +++----------------------------------------- vpx_dsp/x86/inv_txfm_sse2.h | 48 +++++++++++++++++++++++++++- vpx_dsp/x86/inv_txfm_ssse3.c | 29 +---------------- 3 files changed, 52 insertions(+), 99 deletions(-) diff --git a/vpx_dsp/x86/inv_txfm_sse2.c b/vpx_dsp/x86/inv_txfm_sse2.c index 32f1b63..9e77bdb 100644 --- a/vpx_dsp/x86/inv_txfm_sse2.c +++ b/vpx_dsp/x86/inv_txfm_sse2.c @@ -219,57 +219,18 @@ static INLINE void idct8(const __m128i *const in, __m128i *const out) { void vpx_idct8x8_64_add_sse2(const tran_low_t *input, uint8_t *dest, int stride) { - const __m128i final_rounding = _mm_set1_epi16(1 << 4); - __m128i in[8]; int i; // Load input data. - in[0] = load_input_data(input); - in[1] = load_input_data(input + 8 * 1); - in[2] = load_input_data(input + 8 * 2); - in[3] = load_input_data(input + 8 * 3); - in[4] = load_input_data(input + 8 * 4); - in[5] = load_input_data(input + 8 * 5); - in[6] = load_input_data(input + 8 * 6); - in[7] = load_input_data(input + 8 * 7); + load_buffer_8x8(input, in); // 2-D for (i = 0; i < 2; i++) { - // 8x8 Transpose is copied from vpx_fdct8x8_sse2() - transpose_16bit_8x8(in, in); - - // 4-stage 1D idct8x8 - idct8(in, in); + idct8_sse2(in); } - // Final rounding and shift - in[0] = _mm_adds_epi16(in[0], final_rounding); - in[1] = _mm_adds_epi16(in[1], final_rounding); - in[2] = _mm_adds_epi16(in[2], final_rounding); - in[3] = _mm_adds_epi16(in[3], final_rounding); - in[4] = _mm_adds_epi16(in[4], final_rounding); - in[5] = _mm_adds_epi16(in[5], final_rounding); - in[6] = _mm_adds_epi16(in[6], final_rounding); - in[7] = _mm_adds_epi16(in[7], final_rounding); - - in[0] = _mm_srai_epi16(in[0], 5); - in[1] = _mm_srai_epi16(in[1], 5); - in[2] = _mm_srai_epi16(in[2], 5); - in[3] = _mm_srai_epi16(in[3], 5); - in[4] = _mm_srai_epi16(in[4], 5); - in[5] = _mm_srai_epi16(in[5], 5); - in[6] = _mm_srai_epi16(in[6], 5); - in[7] = _mm_srai_epi16(in[7], 5); - - recon_and_store(dest + 0 * stride, in[0]); - recon_and_store(dest + 1 * stride, in[1]); - recon_and_store(dest + 2 * stride, in[2]); - recon_and_store(dest + 3 * stride, in[3]); - recon_and_store(dest + 4 * stride, in[4]); - recon_and_store(dest + 5 * stride, in[5]); - recon_and_store(dest + 6 * stride, in[6]); - recon_and_store(dest + 7 * stride, in[7]); + write_buffer_8x8(in, dest, stride); } void vpx_idct8x8_1_add_sse2(const tran_low_t *input, uint8_t *dest, @@ -505,7 +466,6 @@ void iadst8_sse2(__m128i *in) { void vpx_idct8x8_12_add_sse2(const tran_low_t *input, uint8_t *dest, int stride) { const __m128i zero = _mm_setzero_si128(); - const __m128i final_rounding = _mm_set1_epi16(1 << 4); const __m128i stg1_0 = pair_set_epi16(cospi_28_64, -cospi_4_64); const __m128i stg1_1 = pair_set_epi16(cospi_4_64, cospi_28_64); const __m128i stg1_2 = pair_set_epi16(-cospi_20_64, cospi_12_64); @@ -575,33 +535,7 @@ void vpx_idct8x8_12_add_sse2(const tran_low_t *input, uint8_t *dest, in[4] = in[5] = in[6] = in[7] = zero; idct8(in, in); - // Final rounding and shift - in[0] = _mm_adds_epi16(in[0], final_rounding); - in[1] = _mm_adds_epi16(in[1], final_rounding); - in[2] = _mm_adds_epi16(in[2], final_rounding); - in[3] = _mm_adds_epi16(in[3], final_rounding); - in[4] = _mm_adds_epi16(in[4], final_rounding); - in[5] = _mm_adds_epi16(in[5], final_rounding); - in[6] = _mm_adds_epi16(in[6], final_rounding); - in[7] = _mm_adds_epi16(in[7], final_rounding); - - in[0] = _mm_srai_epi16(in[0], 5); - in[1] = _mm_srai_epi16(in[1], 5); - in[2] = _mm_srai_epi16(in[2], 5); - in[3] = _mm_srai_epi16(in[3], 5); - in[4] = _mm_srai_epi16(in[4], 5); - in[5] = _mm_srai_epi16(in[5], 5); - in[6] = _mm_srai_epi16(in[6], 5); - in[7] = _mm_srai_epi16(in[7], 5); - - recon_and_store(dest + 0 * stride, in[0]); - recon_and_store(dest + 1 * stride, in[1]); - recon_and_store(dest + 2 * stride, in[2]); - recon_and_store(dest + 3 * stride, in[3]); - recon_and_store(dest + 4 * stride, in[4]); - recon_and_store(dest + 5 * stride, in[5]); - recon_and_store(dest + 6 * stride, in[6]); - recon_and_store(dest + 7 * stride, in[7]); + write_buffer_8x8(in, dest, stride); } #define IDCT16 \ diff --git a/vpx_dsp/x86/inv_txfm_sse2.h b/vpx_dsp/x86/inv_txfm_sse2.h index a739fd1..acaf861 100644 --- a/vpx_dsp/x86/inv_txfm_sse2.h +++ b/vpx_dsp/x86/inv_txfm_sse2.h @@ -99,6 +99,18 @@ static INLINE __m128i load_input_data(const tran_low_t *data) { #endif } +static INLINE void load_buffer_8x8(const tran_low_t *const input, + __m128i *const in) { + in[0] = load_input_data(input + 0 * 8); + in[1] = load_input_data(input + 1 * 8); + in[2] = load_input_data(input + 2 * 8); + in[3] = load_input_data(input + 3 * 8); + in[4] = load_input_data(input + 4 * 8); + in[5] = load_input_data(input + 5 * 8); + in[6] = load_input_data(input + 6 * 8); + in[7] = load_input_data(input + 7 * 8); +} + static INLINE void load_buffer_8x16(const tran_low_t *const input, __m128i *const in) { in[0] = load_input_data(input + 0 * 16); @@ -129,7 +141,41 @@ static INLINE void recon_and_store(uint8_t *const dest, const __m128i in_x) { _mm_storel_epi64((__m128i *)(dest), d0); } -static INLINE void write_buffer_8x16(uint8_t *dest, __m128i *in, int stride) { +static INLINE void write_buffer_8x8(const __m128i *const in, + uint8_t *const dest, const int stride) { + const __m128i final_rounding = _mm_set1_epi16(1 << 4); + __m128i t[8]; + // Final rounding and shift + t[0] = _mm_adds_epi16(in[0], final_rounding); + t[1] = _mm_adds_epi16(in[1], final_rounding); + t[2] = _mm_adds_epi16(in[2], final_rounding); + t[3] = _mm_adds_epi16(in[3], final_rounding); + t[4] = _mm_adds_epi16(in[4], final_rounding); + t[5] = _mm_adds_epi16(in[5], final_rounding); + t[6] = _mm_adds_epi16(in[6], final_rounding); + t[7] = _mm_adds_epi16(in[7], final_rounding); + + t[0] = _mm_srai_epi16(t[0], 5); + t[1] = _mm_srai_epi16(t[1], 5); + t[2] = _mm_srai_epi16(t[2], 5); + t[3] = _mm_srai_epi16(t[3], 5); + t[4] = _mm_srai_epi16(t[4], 5); + t[5] = _mm_srai_epi16(t[5], 5); + t[6] = _mm_srai_epi16(t[6], 5); + t[7] = _mm_srai_epi16(t[7], 5); + + recon_and_store(dest + 0 * stride, t[0]); + recon_and_store(dest + 1 * stride, t[1]); + recon_and_store(dest + 2 * stride, t[2]); + recon_and_store(dest + 3 * stride, t[3]); + recon_and_store(dest + 4 * stride, t[4]); + recon_and_store(dest + 5 * stride, t[5]); + recon_and_store(dest + 6 * stride, t[6]); + recon_and_store(dest + 7 * stride, t[7]); +} + +static INLINE void write_buffer_8x16(uint8_t *const dest, __m128i *const in, + const int stride) { const __m128i final_rounding = _mm_set1_epi16(1 << 5); // Final rounding and shift in[0] = _mm_adds_epi16(in[0], final_rounding); diff --git a/vpx_dsp/x86/inv_txfm_ssse3.c b/vpx_dsp/x86/inv_txfm_ssse3.c index ae0b591..0e86e43 100644 --- a/vpx_dsp/x86/inv_txfm_ssse3.c +++ b/vpx_dsp/x86/inv_txfm_ssse3.c @@ -19,7 +19,6 @@ void vpx_idct8x8_12_add_ssse3(const tran_low_t *input, uint8_t *dest, int stride) { const __m128i zero = _mm_setzero_si128(); const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING); - const __m128i final_rounding = _mm_set1_epi16(1 << 4); const __m128i stg1_0 = pair_set_epi16(2 * cospi_28_64, 2 * cospi_28_64); const __m128i stg1_1 = pair_set_epi16(2 * cospi_4_64, 2 * cospi_4_64); const __m128i stg1_2 = pair_set_epi16(-2 * cospi_20_64, -2 * cospi_20_64); @@ -148,33 +147,7 @@ void vpx_idct8x8_12_add_ssse3(const tran_low_t *input, uint8_t *dest, in[6] = _mm_sub_epi16(stp1_1, stp1_6); in[7] = _mm_sub_epi16(stp1_0, stp2_7); - // Final rounding and shift - in[0] = _mm_adds_epi16(in[0], final_rounding); - in[1] = _mm_adds_epi16(in[1], final_rounding); - in[2] = _mm_adds_epi16(in[2], final_rounding); - in[3] = _mm_adds_epi16(in[3], final_rounding); - in[4] = _mm_adds_epi16(in[4], final_rounding); - in[5] = _mm_adds_epi16(in[5], final_rounding); - in[6] = _mm_adds_epi16(in[6], final_rounding); - in[7] = _mm_adds_epi16(in[7], final_rounding); - - in[0] = _mm_srai_epi16(in[0], 5); - in[1] = _mm_srai_epi16(in[1], 5); - in[2] = _mm_srai_epi16(in[2], 5); - in[3] = _mm_srai_epi16(in[3], 5); - in[4] = _mm_srai_epi16(in[4], 5); - in[5] = _mm_srai_epi16(in[5], 5); - in[6] = _mm_srai_epi16(in[6], 5); - in[7] = _mm_srai_epi16(in[7], 5); - - recon_and_store(dest + 0 * stride, in[0]); - recon_and_store(dest + 1 * stride, in[1]); - recon_and_store(dest + 2 * stride, in[2]); - recon_and_store(dest + 3 * stride, in[3]); - recon_and_store(dest + 4 * stride, in[4]); - recon_and_store(dest + 5 * stride, in[5]); - recon_and_store(dest + 6 * stride, in[6]); - recon_and_store(dest + 7 * stride, in[7]); + write_buffer_8x8(in, dest, stride); } // Only do addition and subtraction butterfly, size = 16, 32 -- 2.7.4