From bd990cad72ecdc864ebfd4470dbbd747974f3aa9 Mon Sep 17 00:00:00 2001 From: Johann Date: Wed, 23 Aug 2017 15:27:25 -0700 Subject: [PATCH] quantize x86: dedup some parts Change-Id: I9f95f47bc7ecbb7980f21cbc3a91f699624141af --- vpx_dsp/vpx_dsp.mk | 1 + vpx_dsp/x86/quantize_avx.c | 168 +++++++---------------------------- vpx_dsp/x86/quantize_sse2.c | 109 ++++++----------------- vpx_dsp/x86/quantize_ssse3.c | 167 ++++++++-------------------------- vpx_dsp/x86/quantize_x86.h | 78 ++++++++++++++++ 5 files changed, 175 insertions(+), 348 deletions(-) create mode 100644 vpx_dsp/x86/quantize_x86.h diff --git a/vpx_dsp/vpx_dsp.mk b/vpx_dsp/vpx_dsp.mk index d18dd3107..3b1a873cd 100644 --- a/vpx_dsp/vpx_dsp.mk +++ b/vpx_dsp/vpx_dsp.mk @@ -279,6 +279,7 @@ ifeq ($(CONFIG_VP9_ENCODER),yes) DSP_SRCS-yes += quantize.c DSP_SRCS-yes += quantize.h +DSP_SRCS-$(HAVE_SSE2) += x86/quantize_x86.h DSP_SRCS-$(HAVE_SSE2) += x86/quantize_sse2.c DSP_SRCS-$(HAVE_SSSE3) += x86/quantize_ssse3.c DSP_SRCS-$(HAVE_AVX) += x86/quantize_avx.c diff --git a/vpx_dsp/x86/quantize_avx.c b/vpx_dsp/x86/quantize_avx.c index 3f28af2ca..6f4489004 100644 --- a/vpx_dsp/x86/quantize_avx.c +++ b/vpx_dsp/x86/quantize_avx.c @@ -17,6 +17,7 @@ #include "./vpx_dsp_rtcd.h" #include "vpx/vpx_integer.h" #include "vpx_dsp/x86/bitdepth_conversion_sse2.h" +#include "vpx_dsp/x86/quantize_x86.h" void vpx_quantize_b_avx(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, @@ -34,9 +35,7 @@ void vpx_quantize_b_avx(const tran_low_t *coeff_ptr, intptr_t n_coeffs, __m128i qcoeff0, qcoeff1; __m128i cmp_mask0, cmp_mask1; __m128i all_zero; - __m128i qtmp0, qtmp1; - __m128i zero_coeff0, zero_coeff1, iscan0, iscan1; - __m128i eob = zero, eob0, eob1; + __m128i eob = zero, eob0; (void)scan_ptr; (void)skip_block; @@ -44,15 +43,8 @@ void vpx_quantize_b_avx(const tran_low_t *coeff_ptr, intptr_t n_coeffs, *eob_ptr = 0; - // Setup global values. - zbin = _mm_load_si128((const __m128i *)zbin_ptr); - // x86 has no "greater *or equal*" comparison. Subtract 1 from zbin so - // it is a strict "greater" comparison. - zbin = _mm_sub_epi16(zbin, _mm_set1_epi16(1)); - round = _mm_load_si128((const __m128i *)round_ptr); - quant = _mm_load_si128((const __m128i *)quant_ptr); - dequant = _mm_load_si128((const __m128i *)dequant_ptr); - shift = _mm_load_si128((const __m128i *)quant_shift_ptr); + load_b_values(zbin_ptr, &zbin, round_ptr, &round, quant_ptr, &quant, + dequant_ptr, &dequant, quant_shift_ptr, &shift); // Do DC and first 15 AC. coeff0 = load_tran_low(coeff_ptr); @@ -81,20 +73,11 @@ void vpx_quantize_b_avx(const tran_low_t *coeff_ptr, intptr_t n_coeffs, shift = _mm_unpackhi_epi64(shift, shift); dequant = _mm_unpackhi_epi64(dequant, dequant); } else { - qcoeff0 = _mm_adds_epi16(qcoeff0, round); + calculate_qcoeff(&qcoeff0, round, quant, shift); round = _mm_unpackhi_epi64(round, round); - qcoeff1 = _mm_adds_epi16(qcoeff1, round); - - qtmp0 = _mm_mulhi_epi16(qcoeff0, quant); quant = _mm_unpackhi_epi64(quant, quant); - qtmp1 = _mm_mulhi_epi16(qcoeff1, quant); - - qtmp0 = _mm_add_epi16(qtmp0, qcoeff0); - qtmp1 = _mm_add_epi16(qtmp1, qcoeff1); - - qcoeff0 = _mm_mulhi_epi16(qtmp0, shift); shift = _mm_unpackhi_epi64(shift, shift); - qcoeff1 = _mm_mulhi_epi16(qtmp1, shift); + calculate_qcoeff(&qcoeff1, round, quant, shift); // Reinsert signs qcoeff0 = _mm_sign_epi16(qcoeff0, coeff0); @@ -107,24 +90,15 @@ void vpx_quantize_b_avx(const tran_low_t *coeff_ptr, intptr_t n_coeffs, store_tran_low(qcoeff0, qcoeff_ptr); store_tran_low(qcoeff1, qcoeff_ptr + 8); - coeff0 = _mm_mullo_epi16(qcoeff0, dequant); + coeff0 = calculate_dqcoeff(qcoeff0, dequant); dequant = _mm_unpackhi_epi64(dequant, dequant); - coeff1 = _mm_mullo_epi16(qcoeff1, dequant); + coeff1 = calculate_dqcoeff(qcoeff1, dequant); store_tran_low(coeff0, dqcoeff_ptr); store_tran_low(coeff1, dqcoeff_ptr + 8); - // Scan for eob. - zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero); - zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero); - iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr)); - iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + 8)); - // Add one to convert from indices to counts - iscan0 = _mm_sub_epi16(iscan0, cmp_mask0); - iscan1 = _mm_sub_epi16(iscan1, cmp_mask1); - eob = _mm_andnot_si128(zero_coeff0, iscan0); - eob1 = _mm_andnot_si128(zero_coeff1, iscan1); - eob = _mm_max_epi16(eob, eob1); + eob = scan_for_eob(&coeff0, &coeff1, cmp_mask0, cmp_mask1, iscan_ptr, 0, + zero); } // AC only loop. @@ -149,17 +123,8 @@ void vpx_quantize_b_avx(const tran_low_t *coeff_ptr, intptr_t n_coeffs, continue; } - qcoeff0 = _mm_adds_epi16(qcoeff0, round); - qcoeff1 = _mm_adds_epi16(qcoeff1, round); - - qtmp0 = _mm_mulhi_epi16(qcoeff0, quant); - qtmp1 = _mm_mulhi_epi16(qcoeff1, quant); - - qtmp0 = _mm_add_epi16(qtmp0, qcoeff0); - qtmp1 = _mm_add_epi16(qtmp1, qcoeff1); - - qcoeff0 = _mm_mulhi_epi16(qtmp0, shift); - qcoeff1 = _mm_mulhi_epi16(qtmp1, shift); + calculate_qcoeff(&qcoeff0, round, quant, shift); + calculate_qcoeff(&qcoeff1, round, quant, shift); qcoeff0 = _mm_sign_epi16(qcoeff0, coeff0); qcoeff1 = _mm_sign_epi16(qcoeff1, coeff1); @@ -170,35 +135,18 @@ void vpx_quantize_b_avx(const tran_low_t *coeff_ptr, intptr_t n_coeffs, store_tran_low(qcoeff0, qcoeff_ptr + index); store_tran_low(qcoeff1, qcoeff_ptr + index + 8); - coeff0 = _mm_mullo_epi16(qcoeff0, dequant); - coeff1 = _mm_mullo_epi16(qcoeff1, dequant); + coeff0 = calculate_dqcoeff(qcoeff0, dequant); + coeff1 = calculate_dqcoeff(qcoeff1, dequant); store_tran_low(coeff0, dqcoeff_ptr + index); store_tran_low(coeff1, dqcoeff_ptr + index + 8); - zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero); - zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero); - iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr + index)); - iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + index + 8)); - iscan0 = _mm_sub_epi16(iscan0, cmp_mask0); - iscan1 = _mm_sub_epi16(iscan1, cmp_mask1); - eob0 = _mm_andnot_si128(zero_coeff0, iscan0); - eob1 = _mm_andnot_si128(zero_coeff1, iscan1); - eob0 = _mm_max_epi16(eob0, eob1); + eob0 = scan_for_eob(&coeff0, &coeff1, cmp_mask0, cmp_mask1, iscan_ptr, + index, zero); eob = _mm_max_epi16(eob, eob0); } - // Accumulate eob. - { - __m128i eob_shuffled; - eob_shuffled = _mm_shuffle_epi32(eob, 0xe); - eob = _mm_max_epi16(eob, eob_shuffled); - eob_shuffled = _mm_shufflelo_epi16(eob, 0xe); - eob = _mm_max_epi16(eob, eob_shuffled); - eob_shuffled = _mm_shufflelo_epi16(eob, 0x1); - eob = _mm_max_epi16(eob, eob_shuffled); - *eob_ptr = _mm_extract_epi16(eob, 1); - } + *eob_ptr = accumulate_eob(eob); } void vpx_quantize_b_32x32_avx( @@ -217,17 +165,13 @@ void vpx_quantize_b_32x32_avx( __m128i qcoeff0, qcoeff1; __m128i cmp_mask0, cmp_mask1; __m128i all_zero; - __m128i qtmp0, qtmp1; - __m128i zero_coeff0, zero_coeff1, iscan0, iscan1; - __m128i eob = zero, eob0, eob1; + __m128i eob = zero, eob0; (void)scan_ptr; (void)n_coeffs; (void)skip_block; assert(!skip_block); - *eob_ptr = 0; - // Setup global values. // The 32x32 halves zbin and round. zbin = _mm_load_si128((const __m128i *)zbin_ptr); @@ -255,7 +199,7 @@ void vpx_quantize_b_32x32_avx( qcoeff1 = _mm_abs_epi16(coeff1); cmp_mask0 = _mm_cmpgt_epi16(qcoeff0, zbin); - zbin = _mm_unpackhi_epi64(zbin, zbin); // Switch DC to AC + zbin = _mm_unpackhi_epi64(zbin, zbin); // Switch DC to AC. cmp_mask1 = _mm_cmpgt_epi16(qcoeff1, zbin); all_zero = _mm_or_si128(cmp_mask0, cmp_mask1); @@ -272,26 +216,17 @@ void vpx_quantize_b_32x32_avx( shift = _mm_unpackhi_epi64(shift, shift); dequant = _mm_unpackhi_epi64(dequant, dequant); } else { - qcoeff0 = _mm_adds_epi16(qcoeff0, round); + calculate_qcoeff(&qcoeff0, round, quant, shift); round = _mm_unpackhi_epi64(round, round); - qcoeff1 = _mm_adds_epi16(qcoeff1, round); - - qtmp0 = _mm_mulhi_epi16(qcoeff0, quant); quant = _mm_unpackhi_epi64(quant, quant); - qtmp1 = _mm_mulhi_epi16(qcoeff1, quant); - - qtmp0 = _mm_add_epi16(qtmp0, qcoeff0); - qtmp1 = _mm_add_epi16(qtmp1, qcoeff1); - - qcoeff0 = _mm_mulhi_epi16(qtmp0, shift); shift = _mm_unpackhi_epi64(shift, shift); - qcoeff1 = _mm_mulhi_epi16(qtmp1, shift); + calculate_qcoeff(&qcoeff1, round, quant, shift); - // Reinsert signs + // Reinsert signs. qcoeff0 = _mm_sign_epi16(qcoeff0, coeff0); qcoeff1 = _mm_sign_epi16(qcoeff1, coeff1); - // Mask out zbin threshold coeffs + // Mask out zbin threshold coeffs. qcoeff0 = _mm_and_si128(qcoeff0, cmp_mask0); qcoeff1 = _mm_and_si128(qcoeff1, cmp_mask1); @@ -304,9 +239,9 @@ void vpx_quantize_b_32x32_avx( coeff0 = _mm_abs_epi16(qcoeff0); coeff1 = _mm_abs_epi16(qcoeff1); - coeff0 = _mm_mullo_epi16(coeff0, dequant); + coeff0 = calculate_dqcoeff(coeff0, dequant); dequant = _mm_unpackhi_epi64(dequant, dequant); - coeff1 = _mm_mullo_epi16(coeff1, dequant); + coeff1 = calculate_dqcoeff(coeff1, dequant); // "Divide" by 2. coeff0 = _mm_srli_epi16(coeff0, 1); @@ -318,17 +253,8 @@ void vpx_quantize_b_32x32_avx( store_tran_low(coeff0, dqcoeff_ptr); store_tran_low(coeff1, dqcoeff_ptr + 8); - // Scan for eob. - zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero); - zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero); - iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr)); - iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + 8)); - // Add one to convert from indices to counts - iscan0 = _mm_sub_epi16(iscan0, cmp_mask0); - iscan1 = _mm_sub_epi16(iscan1, cmp_mask1); - eob = _mm_andnot_si128(zero_coeff0, iscan0); - eob1 = _mm_andnot_si128(zero_coeff1, iscan1); - eob = _mm_max_epi16(eob, eob1); + eob = scan_for_eob(&coeff0, &coeff1, cmp_mask0, cmp_mask1, iscan_ptr, 0, + zero); } // AC only loop. @@ -353,17 +279,8 @@ void vpx_quantize_b_32x32_avx( continue; } - qcoeff0 = _mm_adds_epi16(qcoeff0, round); - qcoeff1 = _mm_adds_epi16(qcoeff1, round); - - qtmp0 = _mm_mulhi_epi16(qcoeff0, quant); - qtmp1 = _mm_mulhi_epi16(qcoeff1, quant); - - qtmp0 = _mm_add_epi16(qtmp0, qcoeff0); - qtmp1 = _mm_add_epi16(qtmp1, qcoeff1); - - qcoeff0 = _mm_mulhi_epi16(qtmp0, shift); - qcoeff1 = _mm_mulhi_epi16(qtmp1, shift); + calculate_qcoeff(&qcoeff0, round, quant, shift); + calculate_qcoeff(&qcoeff1, round, quant, shift); qcoeff0 = _mm_sign_epi16(qcoeff0, coeff0); qcoeff1 = _mm_sign_epi16(qcoeff1, coeff1); @@ -377,8 +294,8 @@ void vpx_quantize_b_32x32_avx( coeff0 = _mm_abs_epi16(qcoeff0); coeff1 = _mm_abs_epi16(qcoeff1); - coeff0 = _mm_mullo_epi16(coeff0, dequant); - coeff1 = _mm_mullo_epi16(coeff1, dequant); + coeff0 = calculate_dqcoeff(coeff0, dequant); + coeff1 = calculate_dqcoeff(coeff1, dequant); coeff0 = _mm_srli_epi16(coeff0, 1); coeff1 = _mm_srli_epi16(coeff1, 1); @@ -389,27 +306,10 @@ void vpx_quantize_b_32x32_avx( store_tran_low(coeff0, dqcoeff_ptr + index); store_tran_low(coeff1, dqcoeff_ptr + index + 8); - zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero); - zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero); - iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr + index)); - iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + index + 8)); - iscan0 = _mm_sub_epi16(iscan0, cmp_mask0); - iscan1 = _mm_sub_epi16(iscan1, cmp_mask1); - eob0 = _mm_andnot_si128(zero_coeff0, iscan0); - eob1 = _mm_andnot_si128(zero_coeff1, iscan1); - eob0 = _mm_max_epi16(eob0, eob1); + eob0 = scan_for_eob(&coeff0, &coeff1, cmp_mask0, cmp_mask1, iscan_ptr, + index, zero); eob = _mm_max_epi16(eob, eob0); } - // Accumulate eob. - { - __m128i eob_shuffled; - eob_shuffled = _mm_shuffle_epi32(eob, 0xe); - eob = _mm_max_epi16(eob, eob_shuffled); - eob_shuffled = _mm_shufflelo_epi16(eob, 0xe); - eob = _mm_max_epi16(eob, eob_shuffled); - eob_shuffled = _mm_shufflelo_epi16(eob, 0x1); - eob = _mm_max_epi16(eob, eob_shuffled); - *eob_ptr = _mm_extract_epi16(eob, 1); - } + *eob_ptr = accumulate_eob(eob); } diff --git a/vpx_dsp/x86/quantize_sse2.c b/vpx_dsp/x86/quantize_sse2.c index 2635b27ef..c020b398c 100644 --- a/vpx_dsp/x86/quantize_sse2.c +++ b/vpx_dsp/x86/quantize_sse2.c @@ -15,6 +15,7 @@ #include "./vpx_dsp_rtcd.h" #include "vpx/vpx_integer.h" #include "vpx_dsp/x86/bitdepth_conversion_sse2.h" +#include "vpx_dsp/x86/quantize_x86.h" void vpx_quantize_b_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, @@ -30,21 +31,15 @@ void vpx_quantize_b_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, __m128i coeff0, coeff1, coeff0_sign, coeff1_sign; __m128i qcoeff0, qcoeff1; __m128i cmp_mask0, cmp_mask1; - __m128i qtmp0, qtmp1; - __m128i zero_coeff0, zero_coeff1, iscan0, iscan1; - __m128i eob, eob0, eob1; + __m128i eob, eob0; (void)scan_ptr; (void)skip_block; assert(!skip_block); // Setup global values. - zbin = _mm_load_si128((const __m128i *)zbin_ptr); - round = _mm_load_si128((const __m128i *)round_ptr); - quant = _mm_load_si128((const __m128i *)quant_ptr); - zbin = _mm_sub_epi16(zbin, _mm_set1_epi16(1)); - dequant = _mm_load_si128((const __m128i *)dequant_ptr); - shift = _mm_load_si128((const __m128i *)quant_shift_ptr); + load_b_values(zbin_ptr, &zbin, round_ptr, &round, quant_ptr, &quant, + dequant_ptr, &dequant, quant_shift_ptr, &shift); // Do DC and first 15 AC. coeff0 = load_tran_low(coeff_ptr); @@ -53,35 +48,24 @@ void vpx_quantize_b_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, // Poor man's abs(). coeff0_sign = _mm_srai_epi16(coeff0, 15); coeff1_sign = _mm_srai_epi16(coeff1, 15); - qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign); - qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign); - qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign); - qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign); + qcoeff0 = invert_sign_sse2(coeff0, coeff0_sign); + qcoeff1 = invert_sign_sse2(coeff1, coeff1_sign); cmp_mask0 = _mm_cmpgt_epi16(qcoeff0, zbin); zbin = _mm_unpackhi_epi64(zbin, zbin); // Switch DC to AC cmp_mask1 = _mm_cmpgt_epi16(qcoeff1, zbin); - qcoeff0 = _mm_adds_epi16(qcoeff0, round); + calculate_qcoeff(&qcoeff0, round, quant, shift); + round = _mm_unpackhi_epi64(round, round); - qcoeff1 = _mm_adds_epi16(qcoeff1, round); - - qtmp0 = _mm_mulhi_epi16(qcoeff0, quant); quant = _mm_unpackhi_epi64(quant, quant); - qtmp1 = _mm_mulhi_epi16(qcoeff1, quant); - - qtmp0 = _mm_add_epi16(qtmp0, qcoeff0); - qtmp1 = _mm_add_epi16(qtmp1, qcoeff1); - - qcoeff0 = _mm_mulhi_epi16(qtmp0, shift); shift = _mm_unpackhi_epi64(shift, shift); - qcoeff1 = _mm_mulhi_epi16(qtmp1, shift); + + calculate_qcoeff(&qcoeff1, round, quant, shift); // Reinsert signs - qcoeff0 = _mm_xor_si128(qcoeff0, coeff0_sign); - qcoeff1 = _mm_xor_si128(qcoeff1, coeff1_sign); - qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign); - qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign); + qcoeff0 = invert_sign_sse2(qcoeff0, coeff0_sign); + qcoeff1 = invert_sign_sse2(qcoeff1, coeff1_sign); // Mask out zbin threshold coeffs qcoeff0 = _mm_and_si128(qcoeff0, cmp_mask0); @@ -90,24 +74,15 @@ void vpx_quantize_b_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, store_tran_low(qcoeff0, qcoeff_ptr); store_tran_low(qcoeff1, qcoeff_ptr + 8); - coeff0 = _mm_mullo_epi16(qcoeff0, dequant); + coeff0 = calculate_dqcoeff(qcoeff0, dequant); dequant = _mm_unpackhi_epi64(dequant, dequant); - coeff1 = _mm_mullo_epi16(qcoeff1, dequant); + coeff1 = calculate_dqcoeff(qcoeff1, dequant); store_tran_low(coeff0, dqcoeff_ptr); store_tran_low(coeff1, dqcoeff_ptr + 8); - // Scan for eob. - zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero); - zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero); - iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr)); - iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + 8)); - // Add one to convert from indices to counts - iscan0 = _mm_sub_epi16(iscan0, cmp_mask0); - iscan1 = _mm_sub_epi16(iscan1, cmp_mask1); - eob = _mm_andnot_si128(zero_coeff0, iscan0); - eob1 = _mm_andnot_si128(zero_coeff1, iscan1); - eob = _mm_max_epi16(eob, eob1); + eob = + scan_for_eob(&coeff0, &coeff1, cmp_mask0, cmp_mask1, iscan_ptr, 0, zero); // AC only loop. while (index < n_coeffs) { @@ -116,30 +91,17 @@ void vpx_quantize_b_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, coeff0_sign = _mm_srai_epi16(coeff0, 15); coeff1_sign = _mm_srai_epi16(coeff1, 15); - qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign); - qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign); - qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign); - qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign); + qcoeff0 = invert_sign_sse2(coeff0, coeff0_sign); + qcoeff1 = invert_sign_sse2(coeff1, coeff1_sign); cmp_mask0 = _mm_cmpgt_epi16(qcoeff0, zbin); cmp_mask1 = _mm_cmpgt_epi16(qcoeff1, zbin); - qcoeff0 = _mm_adds_epi16(qcoeff0, round); - qcoeff1 = _mm_adds_epi16(qcoeff1, round); + calculate_qcoeff(&qcoeff0, round, quant, shift); + calculate_qcoeff(&qcoeff1, round, quant, shift); - qtmp0 = _mm_mulhi_epi16(qcoeff0, quant); - qtmp1 = _mm_mulhi_epi16(qcoeff1, quant); - - qtmp0 = _mm_add_epi16(qtmp0, qcoeff0); - qtmp1 = _mm_add_epi16(qtmp1, qcoeff1); - - qcoeff0 = _mm_mulhi_epi16(qtmp0, shift); - qcoeff1 = _mm_mulhi_epi16(qtmp1, shift); - - qcoeff0 = _mm_xor_si128(qcoeff0, coeff0_sign); - qcoeff1 = _mm_xor_si128(qcoeff1, coeff1_sign); - qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign); - qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign); + qcoeff0 = invert_sign_sse2(qcoeff0, coeff0_sign); + qcoeff1 = invert_sign_sse2(qcoeff1, coeff1_sign); qcoeff0 = _mm_and_si128(qcoeff0, cmp_mask0); qcoeff1 = _mm_and_si128(qcoeff1, cmp_mask1); @@ -147,35 +109,18 @@ void vpx_quantize_b_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, store_tran_low(qcoeff0, qcoeff_ptr + index); store_tran_low(qcoeff1, qcoeff_ptr + index + 8); - coeff0 = _mm_mullo_epi16(qcoeff0, dequant); - coeff1 = _mm_mullo_epi16(qcoeff1, dequant); + coeff0 = calculate_dqcoeff(qcoeff0, dequant); + coeff1 = calculate_dqcoeff(qcoeff1, dequant); store_tran_low(coeff0, dqcoeff_ptr + index); store_tran_low(coeff1, dqcoeff_ptr + index + 8); - zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero); - zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero); - iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr + index)); - iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + index + 8)); - iscan0 = _mm_sub_epi16(iscan0, cmp_mask0); - iscan1 = _mm_sub_epi16(iscan1, cmp_mask1); - eob0 = _mm_andnot_si128(zero_coeff0, iscan0); - eob1 = _mm_andnot_si128(zero_coeff1, iscan1); - eob0 = _mm_max_epi16(eob0, eob1); + eob0 = scan_for_eob(&coeff0, &coeff1, cmp_mask0, cmp_mask1, iscan_ptr, + index, zero); eob = _mm_max_epi16(eob, eob0); index += 16; } - // Accumulate eob. - { - __m128i eob_shuffled; - eob_shuffled = _mm_shuffle_epi32(eob, 0xe); - eob = _mm_max_epi16(eob, eob_shuffled); - eob_shuffled = _mm_shufflelo_epi16(eob, 0xe); - eob = _mm_max_epi16(eob, eob_shuffled); - eob_shuffled = _mm_shufflelo_epi16(eob, 0x1); - eob = _mm_max_epi16(eob, eob_shuffled); - *eob_ptr = _mm_extract_epi16(eob, 1); - } + *eob_ptr = accumulate_eob(eob); } diff --git a/vpx_dsp/x86/quantize_ssse3.c b/vpx_dsp/x86/quantize_ssse3.c index 2a0eacb51..3f528e1a9 100644 --- a/vpx_dsp/x86/quantize_ssse3.c +++ b/vpx_dsp/x86/quantize_ssse3.c @@ -14,6 +14,7 @@ #include "./vpx_dsp_rtcd.h" #include "vpx/vpx_integer.h" #include "vpx_dsp/x86/bitdepth_conversion_sse2.h" +#include "vpx_dsp/x86/quantize_x86.h" void vpx_quantize_b_ssse3(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, @@ -23,29 +24,20 @@ void vpx_quantize_b_ssse3(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan_ptr, const int16_t *iscan_ptr) { const __m128i zero = _mm_setzero_si128(); - intptr_t index = 16; + int index = 16; __m128i zbin, round, quant, dequant, shift; __m128i coeff0, coeff1; __m128i qcoeff0, qcoeff1; __m128i cmp_mask0, cmp_mask1; - __m128i qtmp0, qtmp1; - __m128i zero_coeff0, zero_coeff1, iscan0, iscan1; - __m128i eob, eob0, eob1; + __m128i eob, eob0; (void)scan_ptr; (void)skip_block; assert(!skip_block); - // Setup global values. - zbin = _mm_load_si128((const __m128i *)zbin_ptr); - // x86 has no "greater *or equal*" comparison. Subtract 1 from zbin so - // it is a strict "greater" comparison. - zbin = _mm_sub_epi16(zbin, _mm_set1_epi16(1)); - round = _mm_load_si128((const __m128i *)round_ptr); - quant = _mm_load_si128((const __m128i *)quant_ptr); - dequant = _mm_load_si128((const __m128i *)dequant_ptr); - shift = _mm_load_si128((const __m128i *)quant_shift_ptr); + load_b_values(zbin_ptr, &zbin, round_ptr, &round, quant_ptr, &quant, + dequant_ptr, &dequant, quant_shift_ptr, &shift); // Do DC and first 15 AC. coeff0 = load_tran_low(coeff_ptr); @@ -58,20 +50,11 @@ void vpx_quantize_b_ssse3(const tran_low_t *coeff_ptr, intptr_t n_coeffs, zbin = _mm_unpackhi_epi64(zbin, zbin); // Switch DC to AC cmp_mask1 = _mm_cmpgt_epi16(qcoeff1, zbin); - qcoeff0 = _mm_adds_epi16(qcoeff0, round); + calculate_qcoeff(&qcoeff0, round, quant, shift); round = _mm_unpackhi_epi64(round, round); - qcoeff1 = _mm_adds_epi16(qcoeff1, round); - - qtmp0 = _mm_mulhi_epi16(qcoeff0, quant); quant = _mm_unpackhi_epi64(quant, quant); - qtmp1 = _mm_mulhi_epi16(qcoeff1, quant); - - qtmp0 = _mm_add_epi16(qtmp0, qcoeff0); - qtmp1 = _mm_add_epi16(qtmp1, qcoeff1); - - qcoeff0 = _mm_mulhi_epi16(qtmp0, shift); shift = _mm_unpackhi_epi64(shift, shift); - qcoeff1 = _mm_mulhi_epi16(qtmp1, shift); + calculate_qcoeff(&qcoeff1, round, quant, shift); // Reinsert signs qcoeff0 = _mm_sign_epi16(qcoeff0, coeff0); @@ -84,24 +67,15 @@ void vpx_quantize_b_ssse3(const tran_low_t *coeff_ptr, intptr_t n_coeffs, store_tran_low(qcoeff0, qcoeff_ptr); store_tran_low(qcoeff1, qcoeff_ptr + 8); - coeff0 = _mm_mullo_epi16(qcoeff0, dequant); + coeff0 = calculate_dqcoeff(qcoeff0, dequant); dequant = _mm_unpackhi_epi64(dequant, dequant); - coeff1 = _mm_mullo_epi16(qcoeff1, dequant); + coeff1 = calculate_dqcoeff(qcoeff1, dequant); store_tran_low(coeff0, dqcoeff_ptr); store_tran_low(coeff1, dqcoeff_ptr + 8); - // Scan for eob. - zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero); - zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero); - iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr)); - iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + 8)); - // Add one to convert from indices to counts - iscan0 = _mm_sub_epi16(iscan0, cmp_mask0); - iscan1 = _mm_sub_epi16(iscan1, cmp_mask1); - eob = _mm_andnot_si128(zero_coeff0, iscan0); - eob1 = _mm_andnot_si128(zero_coeff1, iscan1); - eob = _mm_max_epi16(eob, eob1); + eob = + scan_for_eob(&coeff0, &coeff1, cmp_mask0, cmp_mask1, iscan_ptr, 0, zero); // AC only loop. while (index < n_coeffs) { @@ -114,17 +88,8 @@ void vpx_quantize_b_ssse3(const tran_low_t *coeff_ptr, intptr_t n_coeffs, cmp_mask0 = _mm_cmpgt_epi16(qcoeff0, zbin); cmp_mask1 = _mm_cmpgt_epi16(qcoeff1, zbin); - qcoeff0 = _mm_adds_epi16(qcoeff0, round); - qcoeff1 = _mm_adds_epi16(qcoeff1, round); - - qtmp0 = _mm_mulhi_epi16(qcoeff0, quant); - qtmp1 = _mm_mulhi_epi16(qcoeff1, quant); - - qtmp0 = _mm_add_epi16(qtmp0, qcoeff0); - qtmp1 = _mm_add_epi16(qtmp1, qcoeff1); - - qcoeff0 = _mm_mulhi_epi16(qtmp0, shift); - qcoeff1 = _mm_mulhi_epi16(qtmp1, shift); + calculate_qcoeff(&qcoeff0, round, quant, shift); + calculate_qcoeff(&qcoeff1, round, quant, shift); qcoeff0 = _mm_sign_epi16(qcoeff0, coeff0); qcoeff1 = _mm_sign_epi16(qcoeff1, coeff1); @@ -135,37 +100,20 @@ void vpx_quantize_b_ssse3(const tran_low_t *coeff_ptr, intptr_t n_coeffs, store_tran_low(qcoeff0, qcoeff_ptr + index); store_tran_low(qcoeff1, qcoeff_ptr + index + 8); - coeff0 = _mm_mullo_epi16(qcoeff0, dequant); - coeff1 = _mm_mullo_epi16(qcoeff1, dequant); + coeff0 = calculate_dqcoeff(qcoeff0, dequant); + coeff1 = calculate_dqcoeff(qcoeff1, dequant); store_tran_low(coeff0, dqcoeff_ptr + index); store_tran_low(coeff1, dqcoeff_ptr + index + 8); - zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero); - zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero); - iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr + index)); - iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + index + 8)); - iscan0 = _mm_sub_epi16(iscan0, cmp_mask0); - iscan1 = _mm_sub_epi16(iscan1, cmp_mask1); - eob0 = _mm_andnot_si128(zero_coeff0, iscan0); - eob1 = _mm_andnot_si128(zero_coeff1, iscan1); - eob0 = _mm_max_epi16(eob0, eob1); + eob0 = scan_for_eob(&coeff0, &coeff1, cmp_mask0, cmp_mask1, iscan_ptr, + index, zero); eob = _mm_max_epi16(eob, eob0); index += 16; } - // Accumulate eob. - { - __m128i eob_shuffled; - eob_shuffled = _mm_shuffle_epi32(eob, 0xe); - eob = _mm_max_epi16(eob, eob_shuffled); - eob_shuffled = _mm_shufflelo_epi16(eob, 0xe); - eob = _mm_max_epi16(eob, eob_shuffled); - eob_shuffled = _mm_shufflelo_epi16(eob, 0x1); - eob = _mm_max_epi16(eob, eob_shuffled); - *eob_ptr = _mm_extract_epi16(eob, 1); - } + *eob_ptr = accumulate_eob(eob); } void vpx_quantize_b_32x32_ssse3( @@ -176,16 +124,14 @@ void vpx_quantize_b_32x32_ssse3( const int16_t *scan_ptr, const int16_t *iscan_ptr) { const __m128i zero = _mm_setzero_si128(); const __m128i one = _mm_set1_epi16(1); - intptr_t index = 16; + int index; __m128i zbin, round, quant, dequant, shift; __m128i coeff0, coeff1; __m128i qcoeff0, qcoeff1; __m128i cmp_mask0, cmp_mask1; __m128i all_zero; - __m128i qtmp0, qtmp1; - __m128i zero_coeff0, zero_coeff1, iscan0, iscan1; - __m128i eob = zero, eob0, eob1; + __m128i eob = zero, eob0; (void)scan_ptr; (void)n_coeffs; @@ -236,27 +182,18 @@ void vpx_quantize_b_32x32_ssse3( _mm_store_si128((__m128i *)(qcoeff_ptr + 12), zero); _mm_store_si128((__m128i *)(dqcoeff_ptr + 4), zero); _mm_store_si128((__m128i *)(dqcoeff_ptr + 12), zero); -#endif +#endif // CONFIG_HIGHBITDEPTH round = _mm_unpackhi_epi64(round, round); quant = _mm_unpackhi_epi64(quant, quant); shift = _mm_unpackhi_epi64(shift, shift); dequant = _mm_unpackhi_epi64(dequant, dequant); } else { - qcoeff0 = _mm_adds_epi16(qcoeff0, round); + calculate_qcoeff(&qcoeff0, round, quant, shift); round = _mm_unpackhi_epi64(round, round); - qcoeff1 = _mm_adds_epi16(qcoeff1, round); - - qtmp0 = _mm_mulhi_epi16(qcoeff0, quant); quant = _mm_unpackhi_epi64(quant, quant); - qtmp1 = _mm_mulhi_epi16(qcoeff1, quant); - - qtmp0 = _mm_add_epi16(qtmp0, qcoeff0); - qtmp1 = _mm_add_epi16(qtmp1, qcoeff1); - - qcoeff0 = _mm_mulhi_epi16(qtmp0, shift); shift = _mm_unpackhi_epi64(shift, shift); - qcoeff1 = _mm_mulhi_epi16(qtmp1, shift); + calculate_qcoeff(&qcoeff1, round, quant, shift); // Reinsert signs. qcoeff0 = _mm_sign_epi16(qcoeff0, coeff0); @@ -275,9 +212,9 @@ void vpx_quantize_b_32x32_ssse3( coeff0 = _mm_abs_epi16(qcoeff0); coeff1 = _mm_abs_epi16(qcoeff1); - coeff0 = _mm_mullo_epi16(coeff0, dequant); + coeff0 = calculate_dqcoeff(coeff0, dequant); dequant = _mm_unpackhi_epi64(dequant, dequant); - coeff1 = _mm_mullo_epi16(coeff1, dequant); + coeff1 = calculate_dqcoeff(coeff1, dequant); // "Divide" by 2. coeff0 = _mm_srli_epi16(coeff0, 1); @@ -289,17 +226,8 @@ void vpx_quantize_b_32x32_ssse3( store_tran_low(coeff0, dqcoeff_ptr); store_tran_low(coeff1, dqcoeff_ptr + 8); - // Scan for eob. - zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero); - zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero); - iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr)); - iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + 8)); - // Add one to convert from indices to counts. - iscan0 = _mm_sub_epi16(iscan0, cmp_mask0); - iscan1 = _mm_sub_epi16(iscan1, cmp_mask1); - eob = _mm_andnot_si128(zero_coeff0, iscan0); - eob1 = _mm_andnot_si128(zero_coeff1, iscan1); - eob = _mm_max_epi16(eob, eob1); + eob = scan_for_eob(&coeff0, &coeff1, cmp_mask0, cmp_mask1, iscan_ptr, 0, + zero); } // AC only loop. @@ -324,21 +252,12 @@ void vpx_quantize_b_32x32_ssse3( _mm_store_si128((__m128i *)(qcoeff_ptr + index + 12), zero); _mm_store_si128((__m128i *)(dqcoeff_ptr + index + 4), zero); _mm_store_si128((__m128i *)(dqcoeff_ptr + index + 12), zero); -#endif +#endif // CONFIG_VP9_HIGHBITDEPTH continue; } - qcoeff0 = _mm_adds_epi16(qcoeff0, round); - qcoeff1 = _mm_adds_epi16(qcoeff1, round); - - qtmp0 = _mm_mulhi_epi16(qcoeff0, quant); - qtmp1 = _mm_mulhi_epi16(qcoeff1, quant); - - qtmp0 = _mm_add_epi16(qtmp0, qcoeff0); - qtmp1 = _mm_add_epi16(qtmp1, qcoeff1); - - qcoeff0 = _mm_mulhi_epi16(qtmp0, shift); - qcoeff1 = _mm_mulhi_epi16(qtmp1, shift); + calculate_qcoeff(&qcoeff0, round, quant, shift); + calculate_qcoeff(&qcoeff1, round, quant, shift); qcoeff0 = _mm_sign_epi16(qcoeff0, coeff0); qcoeff1 = _mm_sign_epi16(qcoeff1, coeff1); @@ -352,8 +271,8 @@ void vpx_quantize_b_32x32_ssse3( coeff0 = _mm_abs_epi16(qcoeff0); coeff1 = _mm_abs_epi16(qcoeff1); - coeff0 = _mm_mullo_epi16(coeff0, dequant); - coeff1 = _mm_mullo_epi16(coeff1, dequant); + coeff0 = calculate_dqcoeff(coeff0, dequant); + coeff1 = calculate_dqcoeff(coeff1, dequant); coeff0 = _mm_srli_epi16(coeff0, 1); coeff1 = _mm_srli_epi16(coeff1, 1); @@ -364,26 +283,10 @@ void vpx_quantize_b_32x32_ssse3( store_tran_low(coeff0, dqcoeff_ptr + index); store_tran_low(coeff1, dqcoeff_ptr + index + 8); - zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero); - zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero); - iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr + index)); - iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + index + 8)); - iscan0 = _mm_sub_epi16(iscan0, cmp_mask0); - iscan1 = _mm_sub_epi16(iscan1, cmp_mask1); - eob0 = _mm_andnot_si128(zero_coeff0, iscan0); - eob1 = _mm_andnot_si128(zero_coeff1, iscan1); - eob0 = _mm_max_epi16(eob0, eob1); + eob0 = scan_for_eob(&coeff0, &coeff1, cmp_mask0, cmp_mask1, iscan_ptr, + index, zero); eob = _mm_max_epi16(eob, eob0); } - { - __m128i eob_shuffled; - eob_shuffled = _mm_shuffle_epi32(eob, 0xe); - eob = _mm_max_epi16(eob, eob_shuffled); - eob_shuffled = _mm_shufflelo_epi16(eob, 0xe); - eob = _mm_max_epi16(eob, eob_shuffled); - eob_shuffled = _mm_shufflelo_epi16(eob, 0x1); - eob = _mm_max_epi16(eob, eob_shuffled); - *eob_ptr = _mm_extract_epi16(eob, 1); - } + *eob_ptr = accumulate_eob(eob); } diff --git a/vpx_dsp/x86/quantize_x86.h b/vpx_dsp/x86/quantize_x86.h new file mode 100644 index 000000000..34928fbb5 --- /dev/null +++ b/vpx_dsp/x86/quantize_x86.h @@ -0,0 +1,78 @@ +/* + * Copyright (c) 2017 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 + +#include "./vpx_config.h" +#include "vpx/vpx_integer.h" +#include "vpx_dsp/x86/bitdepth_conversion_sse2.h" + +static INLINE void load_b_values(const int16_t *zbin_ptr, __m128i *zbin, + const int16_t *round_ptr, __m128i *round, + const int16_t *quant_ptr, __m128i *quant, + const int16_t *dequant_ptr, __m128i *dequant, + const int16_t *shift_ptr, __m128i *shift) { + *zbin = _mm_load_si128((const __m128i *)zbin_ptr); + *round = _mm_load_si128((const __m128i *)round_ptr); + *quant = _mm_load_si128((const __m128i *)quant_ptr); + *zbin = _mm_sub_epi16(*zbin, _mm_set1_epi16(1)); + *dequant = _mm_load_si128((const __m128i *)dequant_ptr); + *shift = _mm_load_si128((const __m128i *)shift_ptr); +} + +// With ssse3 and later abs() and sign() are preferred. +static INLINE __m128i invert_sign_sse2(__m128i a, __m128i sign) { + a = _mm_xor_si128(a, sign); + return _mm_sub_epi16(a, sign); +} + +static INLINE void calculate_qcoeff(__m128i *coeff, const __m128i round, + const __m128i quant, const __m128i shift) { + __m128i tmp, qcoeff; + qcoeff = _mm_adds_epi16(*coeff, round); + tmp = _mm_mulhi_epi16(qcoeff, quant); + qcoeff = _mm_add_epi16(tmp, qcoeff); + *coeff = _mm_mulhi_epi16(qcoeff, shift); +} + +static INLINE __m128i calculate_dqcoeff(__m128i qcoeff, __m128i dequant) { + return _mm_mullo_epi16(qcoeff, dequant); +} + +// Scan 16 values for eob reference in scan_ptr. Use masks (-1) from comparing +// to zbin to add 1 to the index in 'scan'. +static INLINE __m128i scan_for_eob(__m128i *coeff0, __m128i *coeff1, + const __m128i zbin_mask0, + const __m128i zbin_mask1, + const int16_t *scan_ptr, const int index, + const __m128i zero) { + const __m128i zero_coeff0 = _mm_cmpeq_epi16(*coeff0, zero); + const __m128i zero_coeff1 = _mm_cmpeq_epi16(*coeff1, zero); + __m128i scan0 = _mm_load_si128((const __m128i *)(scan_ptr + index)); + __m128i scan1 = _mm_load_si128((const __m128i *)(scan_ptr + index + 8)); + __m128i eob0, eob1; + // Add one to convert from indices to counts + scan0 = _mm_sub_epi16(scan0, zbin_mask0); + scan1 = _mm_sub_epi16(scan1, zbin_mask1); + eob0 = _mm_andnot_si128(zero_coeff0, scan0); + eob1 = _mm_andnot_si128(zero_coeff1, scan1); + return _mm_max_epi16(eob0, eob1); +} + +static INLINE int16_t accumulate_eob(__m128i eob) { + __m128i eob_shuffled; + eob_shuffled = _mm_shuffle_epi32(eob, 0xe); + eob = _mm_max_epi16(eob, eob_shuffled); + eob_shuffled = _mm_shufflelo_epi16(eob, 0xe); + eob = _mm_max_epi16(eob, eob_shuffled); + eob_shuffled = _mm_shufflelo_epi16(eob, 0x1); + eob = _mm_max_epi16(eob, eob_shuffled); + return _mm_extract_epi16(eob, 1); +}