ref: ce77a7bcb05c1fec6a3b9af7e0ee0a11df2b1239
parent: 587f0cd39d9b949185f148a683db410d4e3f04b0
parent: 2d3cc8ea2b163fd7275ed0c44102659ff0acd68b
author: Jingning Han <jingning@google.com>
date: Wed Nov 19 06:17:36 EST 2014
Merge "Add sse2 version for vp9_quantize_fp"
--- a/vp9/common/vp9_rtcd_defs.pl
+++ b/vp9/common/vp9_rtcd_defs.pl
@@ -1154,7 +1154,7 @@
specialize qw/vp9_block_error avx2/, "$sse2_x86inc";
add_proto qw/void vp9_quantize_fp/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, int zbin_oq_value, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
- specialize qw/vp9_quantize_fp neon/, "$ssse3_x86_64";
+ specialize qw/vp9_quantize_fp neon sse2/, "$ssse3_x86_64";
add_proto qw/void vp9_quantize_fp_32x32/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, int zbin_oq_value, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
specialize qw/vp9_quantize_fp_32x32/, "$ssse3_x86_64";
--- a/vp9/encoder/x86/vp9_quantize_sse2.c
+++ b/vp9/encoder/x86/vp9_quantize_sse2.c
@@ -223,3 +223,185 @@
*eob_ptr = 0;
}
}
+
+void vp9_quantize_fp_sse2(const int16_t* coeff_ptr, intptr_t n_coeffs,
+ int skip_block, const int16_t* zbin_ptr,
+ const int16_t* round_ptr, const int16_t* quant_ptr,
+ const int16_t* quant_shift_ptr, int16_t* qcoeff_ptr,
+ int16_t* dqcoeff_ptr, const int16_t* dequant_ptr,
+ int zbin_oq_value, uint16_t* eob_ptr,
+ const int16_t* scan_ptr,
+ const int16_t* iscan_ptr) {+ __m128i zero;
+ (void)scan_ptr;
+ (void)zbin_ptr;
+ (void)quant_shift_ptr;
+ (void)zbin_oq_value;
+
+ coeff_ptr += n_coeffs;
+ iscan_ptr += n_coeffs;
+ qcoeff_ptr += n_coeffs;
+ dqcoeff_ptr += n_coeffs;
+ n_coeffs = -n_coeffs;
+ zero = _mm_setzero_si128();
+
+ if (!skip_block) {+ __m128i eob;
+ __m128i round, quant, dequant;
+ {+ __m128i coeff0, coeff1;
+
+ // Setup global values
+ {+ round = _mm_load_si128((const __m128i*)round_ptr);
+ quant = _mm_load_si128((const __m128i*)quant_ptr);
+ dequant = _mm_load_si128((const __m128i*)dequant_ptr);
+ }
+
+ {+ __m128i coeff0_sign, coeff1_sign;
+ __m128i qcoeff0, qcoeff1;
+ __m128i qtmp0, qtmp1;
+ // Do DC and first 15 AC
+ coeff0 = _mm_load_si128((const __m128i*)(coeff_ptr + n_coeffs));
+ coeff1 = _mm_load_si128((const __m128i*)(coeff_ptr + n_coeffs) + 1);
+
+ // Poor man's sign extract
+ 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 = _mm_adds_epi16(qcoeff0, round);
+ 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);
+
+ // Reinsert signs
+ qcoeff0 = _mm_xor_si128(qtmp0, coeff0_sign);
+ qcoeff1 = _mm_xor_si128(qtmp1, coeff1_sign);
+ qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+ qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+ _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs), qcoeff0);
+ _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs) + 1, qcoeff1);
+
+ coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
+ dequant = _mm_unpackhi_epi64(dequant, dequant);
+ coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
+
+ _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs), coeff0);
+ _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs) + 1, coeff1);
+ }
+
+ {+ // Scan for eob
+ __m128i zero_coeff0, zero_coeff1;
+ __m128i nzero_coeff0, nzero_coeff1;
+ __m128i iscan0, iscan1;
+ __m128i eob1;
+ zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
+ zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
+ nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero);
+ nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
+ iscan0 = _mm_load_si128((const __m128i*)(iscan_ptr + n_coeffs));
+ iscan1 = _mm_load_si128((const __m128i*)(iscan_ptr + n_coeffs) + 1);
+ // Add one to convert from indices to counts
+ iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0);
+ iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1);
+ eob = _mm_and_si128(iscan0, nzero_coeff0);
+ eob1 = _mm_and_si128(iscan1, nzero_coeff1);
+ eob = _mm_max_epi16(eob, eob1);
+ }
+ n_coeffs += 8 * 2;
+ }
+
+ // AC only loop
+ while (n_coeffs < 0) {+ __m128i coeff0, coeff1;
+ {+ __m128i coeff0_sign, coeff1_sign;
+ __m128i qcoeff0, qcoeff1;
+ __m128i qtmp0, qtmp1;
+
+ coeff0 = _mm_load_si128((const __m128i*)(coeff_ptr + n_coeffs));
+ coeff1 = _mm_load_si128((const __m128i*)(coeff_ptr + n_coeffs) + 1);
+
+ // Poor man's sign extract
+ 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 = _mm_adds_epi16(qcoeff0, round);
+ qcoeff1 = _mm_adds_epi16(qcoeff1, round);
+ qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
+ qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
+
+ // Reinsert signs
+ qcoeff0 = _mm_xor_si128(qtmp0, coeff0_sign);
+ qcoeff1 = _mm_xor_si128(qtmp1, coeff1_sign);
+ qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+ qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+ _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs), qcoeff0);
+ _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs) + 1, qcoeff1);
+
+ coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
+ coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
+
+ _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs), coeff0);
+ _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs) + 1, coeff1);
+ }
+
+ {+ // Scan for eob
+ __m128i zero_coeff0, zero_coeff1;
+ __m128i nzero_coeff0, nzero_coeff1;
+ __m128i iscan0, iscan1;
+ __m128i eob0, eob1;
+ zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
+ zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
+ nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero);
+ nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
+ iscan0 = _mm_load_si128((const __m128i*)(iscan_ptr + n_coeffs));
+ iscan1 = _mm_load_si128((const __m128i*)(iscan_ptr + n_coeffs) + 1);
+ // Add one to convert from indices to counts
+ iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0);
+ iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1);
+ eob0 = _mm_and_si128(iscan0, nzero_coeff0);
+ eob1 = _mm_and_si128(iscan1, nzero_coeff1);
+ eob0 = _mm_max_epi16(eob0, eob1);
+ eob = _mm_max_epi16(eob, eob0);
+ }
+ n_coeffs += 8 * 2;
+ }
+
+ // 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);
+ }
+ } else {+ do {+ _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs), zero);
+ _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs) + 1, zero);
+ _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs), zero);
+ _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs) + 1, zero);
+ n_coeffs += 8 * 2;
+ } while (n_coeffs < 0);
+ *eob_ptr = 0;
+ }
+}