From e83d99d7b8e009fd847ff91319f57abd5f7f96eb Mon Sep 17 00:00:00 2001 From: Johann Date: Thu, 10 Aug 2017 15:02:22 -0700 Subject: [PATCH] quantize fp: neon implementation About 4x faster when values are below the dequant threshold and 10x faster if everything needs to be calculated. Both numbers would improve if the division for dqcoeff could be simplified. BUG=webm:1426 Change-Id: I8da67c1f3fcb4abed8751990c1afe00bc841f4b2 --- test/vp9_quantize_test.cc | 4 +- vp9/common/vp9_rtcd_defs.pl | 4 +- vp9/encoder/arm/neon/vp9_quantize_neon.c | 124 +++++++++++++++++++++++ 3 files changed, 129 insertions(+), 3 deletions(-) diff --git a/test/vp9_quantize_test.cc b/test/vp9_quantize_test.cc index 9e1b0e2f8..a0a791a99 100644 --- a/test/vp9_quantize_test.cc +++ b/test/vp9_quantize_test.cc @@ -409,7 +409,9 @@ INSTANTIATE_TEST_CASE_P( make_tuple(&vpx_quantize_b_32x32_neon, &vpx_quantize_b_32x32_c, VPX_BITS_8, 32), make_tuple(&QuantFPWrapper, - &QuantFPWrapper, VPX_BITS_8, 16))); + &QuantFPWrapper, VPX_BITS_8, 16), + make_tuple(&QuantFPWrapper, + &QuantFPWrapper, VPX_BITS_8, 32))); #endif // HAVE_NEON && !CONFIG_VP9_HIGHBITDEPTH // Only useful to compare "Speed" test results. diff --git a/vp9/common/vp9_rtcd_defs.pl b/vp9/common/vp9_rtcd_defs.pl index 15e19035b..9fc5ef7f3 100644 --- a/vp9/common/vp9_rtcd_defs.pl +++ b/vp9/common/vp9_rtcd_defs.pl @@ -138,7 +138,7 @@ if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") { specialize qw/vp9_quantize_fp neon sse2/, "$ssse3_x86_64"; add_proto qw/void vp9_quantize_fp_32x32/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *round_ptr, const int16_t *quant_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan"; - specialize qw/vp9_quantize_fp_32x32/, "$ssse3_x86_64"; + specialize qw/vp9_quantize_fp_32x32 neon/, "$ssse3_x86_64"; add_proto qw/void vp9_fdct8x8_quant/, "const int16_t *input, int stride, tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *round_ptr, const int16_t *quant_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan"; specialize qw/vp9_fdct8x8_quant neon ssse3/; @@ -153,7 +153,7 @@ if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") { specialize qw/vp9_quantize_fp neon sse2/, "$ssse3_x86_64"; add_proto qw/void vp9_quantize_fp_32x32/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *round_ptr, const int16_t *quant_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan"; - specialize qw/vp9_quantize_fp_32x32/, "$ssse3_x86_64"; + specialize qw/vp9_quantize_fp_32x32 neon/, "$ssse3_x86_64"; add_proto qw/void vp9_fdct8x8_quant/, "const int16_t *input, int stride, tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *round_ptr, const int16_t *quant_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan"; specialize qw/vp9_fdct8x8_quant sse2 ssse3 neon/; diff --git a/vp9/encoder/arm/neon/vp9_quantize_neon.c b/vp9/encoder/arm/neon/vp9_quantize_neon.c index c73b5123d..97a09bdff 100644 --- a/vp9/encoder/arm/neon/vp9_quantize_neon.c +++ b/vp9/encoder/arm/neon/vp9_quantize_neon.c @@ -12,6 +12,7 @@ #include #include +#include "./vpx_config.h" #include "vpx_mem/vpx_mem.h" #include "vp9/common/vp9_quant_common.h" @@ -111,3 +112,126 @@ void vp9_quantize_fp_neon(const tran_low_t *coeff_ptr, intptr_t count, *eob_ptr = (uint16_t)vget_lane_s16(v_eobmax_final, 0); } } + +static INLINE int32x4_t extract_sign_bit(int32x4_t a) { + return vreinterpretq_s32_u32(vshrq_n_u32(vreinterpretq_u32_s32(a), 31)); +} + +void vp9_quantize_fp_32x32_neon(const tran_low_t *coeff_ptr, intptr_t count, + int skip_block, const int16_t *round_ptr, + const int16_t *quant_ptr, + tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, + const int16_t *dequant_ptr, uint16_t *eob_ptr, + const int16_t *scan, const int16_t *iscan_ptr) { + const int16x8_t one = vdupq_n_s16(1); + const int16x8_t neg_one = vdupq_n_s16(-1); + + // ROUND_POWER_OF_TWO(round_ptr[], 1) + const int16x8_t round = vrshrq_n_s16(vld1q_s16(round_ptr), 1); + const int16x8_t quant = vld1q_s16(quant_ptr); + const int16x4_t dequant = vld1_s16(dequant_ptr); + // dequant >> 2 is used similar to zbin as a threshold. + const int16x8_t dequant_thresh = vshrq_n_s16(vld1q_s16(dequant_ptr), 2); + + // Process dc and the first seven ac coeffs. + const uint16x8_t iscan = + vreinterpretq_u16_s16(vaddq_s16(vld1q_s16(iscan_ptr), one)); + const int16x8_t coeff = load_tran_low_to_s16q(coeff_ptr); + const int16x8_t coeff_sign = vshrq_n_s16(coeff, 15); + const int16x8_t coeff_abs = vabsq_s16(coeff); + const int16x8_t dequant_mask = + vreinterpretq_s16_u16(vcgeq_s16(coeff_abs, dequant_thresh)); + + int16x8_t qcoeff = vaddq_s16(coeff_abs, round); + int32x4_t dqcoeff_0, dqcoeff_1; + int16x8_t dqcoeff; + uint16x8_t eob_max; + (void)scan; + (void)count; + (void)skip_block; + assert(!skip_block); + + // coeff * quant_ptr[]) >> 15 + qcoeff = vqdmulhq_s16(qcoeff, quant); + + // Restore sign. + qcoeff = veorq_s16(qcoeff, coeff_sign); + qcoeff = vsubq_s16(qcoeff, coeff_sign); + qcoeff = vandq_s16(qcoeff, dequant_mask); + + // qcoeff * dequant[] / 2 + dqcoeff_0 = vmull_s16(vget_low_s16(qcoeff), dequant); + dqcoeff_1 = vmull_n_s16(vget_high_s16(qcoeff), dequant_ptr[1]); + + // Add 1 if negative to round towards zero because the C uses division. + dqcoeff_0 = vaddq_s32(dqcoeff_0, extract_sign_bit(dqcoeff_0)); + dqcoeff_1 = vaddq_s32(dqcoeff_1, extract_sign_bit(dqcoeff_1)); + + dqcoeff = vcombine_s16(vshrn_n_s32(dqcoeff_0, 1), vshrn_n_s32(dqcoeff_1, 1)); + + eob_max = vandq_u16(vtstq_s16(qcoeff, neg_one), iscan); + + store_s16q_to_tran_low(qcoeff_ptr, qcoeff); + store_s16q_to_tran_low(dqcoeff_ptr, dqcoeff); + + iscan_ptr += 8; + coeff_ptr += 8; + qcoeff_ptr += 8; + dqcoeff_ptr += 8; + + { + int i; + const int16x8_t round = vrshrq_n_s16(vmovq_n_s16(round_ptr[1]), 1); + const int16x8_t quant = vmovq_n_s16(quant_ptr[1]); + const int16x8_t dequant_thresh = + vshrq_n_s16(vmovq_n_s16(dequant_ptr[1]), 2); + + // Process the rest of the ac coeffs. + for (i = 8; i < 32 * 32; i += 8) { + const uint16x8_t iscan = + vreinterpretq_u16_s16(vaddq_s16(vld1q_s16(iscan_ptr), one)); + const int16x8_t coeff = load_tran_low_to_s16q(coeff_ptr); + const int16x8_t coeff_sign = vshrq_n_s16(coeff, 15); + const int16x8_t coeff_abs = vabsq_s16(coeff); + const int16x8_t dequant_mask = + vreinterpretq_s16_u16(vcgeq_s16(coeff_abs, dequant_thresh)); + + int16x8_t qcoeff = vaddq_s16(coeff_abs, round); + int32x4_t dqcoeff_0, dqcoeff_1; + int16x8_t dqcoeff; + + qcoeff = vqdmulhq_s16(qcoeff, quant); + qcoeff = veorq_s16(qcoeff, coeff_sign); + qcoeff = vsubq_s16(qcoeff, coeff_sign); + qcoeff = vandq_s16(qcoeff, dequant_mask); + + dqcoeff_0 = vmull_n_s16(vget_low_s16(qcoeff), dequant_ptr[1]); + dqcoeff_1 = vmull_n_s16(vget_high_s16(qcoeff), dequant_ptr[1]); + + dqcoeff_0 = vaddq_s32(dqcoeff_0, extract_sign_bit(dqcoeff_0)); + dqcoeff_1 = vaddq_s32(dqcoeff_1, extract_sign_bit(dqcoeff_1)); + + dqcoeff = + vcombine_s16(vshrn_n_s32(dqcoeff_0, 1), vshrn_n_s32(dqcoeff_1, 1)); + + eob_max = + vmaxq_u16(eob_max, vandq_u16(vtstq_s16(qcoeff, neg_one), iscan)); + + store_s16q_to_tran_low(qcoeff_ptr, qcoeff); + store_s16q_to_tran_low(dqcoeff_ptr, dqcoeff); + + iscan_ptr += 8; + coeff_ptr += 8; + qcoeff_ptr += 8; + dqcoeff_ptr += 8; + } + + { + const uint16x4_t eob_max_0 = + vmax_u16(vget_low_u16(eob_max), vget_high_u16(eob_max)); + const uint16x4_t eob_max_1 = vpmax_u16(eob_max_0, eob_max_0); + const uint16x4_t eob_max_2 = vpmax_u16(eob_max_1, eob_max_1); + vst1_lane_u16(eob_ptr, eob_max_2, 0); + } + } +} -- 2.40.0