ref: 7b13d99b981057199d9def0c7a61554609aa4ad9
parent: 13eed991f9450cdd8b03fc03c38ea0f9b2667bf8
author: Johann <johannkoenig@google.com>
date: Mon Aug 21 07:14:39 EDT 2017
quantize fp: ignore skip_block in x86 Change-Id: I1272917c49cf6e6710e52c36535b2fc8c8dced78
--- a/vp9/encoder/x86/vp9_quantize_sse2.c
+++ b/vp9/encoder/x86/vp9_quantize_sse2.c
@@ -8,6 +8,7 @@
* be found in the AUTHORS file in the root of the source tree.
*/
+#include <assert.h>
#include <emmintrin.h>
#include <xmmintrin.h>
@@ -25,8 +26,12 @@
__m128i zero;
__m128i thr;
int16_t nzflag;
+ __m128i eob;
+ __m128i round, quant, dequant;
(void)scan_ptr;
+ (void)skip_block;
+ assert(!skip_block);
coeff_ptr += n_coeffs;
iscan_ptr += n_coeffs;
@@ -35,40 +40,106 @@
n_coeffs = -n_coeffs;
zero = _mm_setzero_si128();
- if (!skip_block) {- __m128i eob;
- __m128i round, quant, dequant;
+ {+ __m128i coeff0, coeff1;
+
+ // Setup global values
{- __m128i coeff0, coeff1;
+ round = _mm_load_si128((const __m128i *)round_ptr);
+ quant = _mm_load_si128((const __m128i *)quant_ptr);
+ dequant = _mm_load_si128((const __m128i *)dequant_ptr);
+ }
- // 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 = load_tran_low(coeff_ptr + n_coeffs);
+ coeff1 = load_tran_low(coeff_ptr + n_coeffs + 8);
- {- __m128i coeff0_sign, coeff1_sign;
- __m128i qcoeff0, qcoeff1;
- __m128i qtmp0, qtmp1;
- // Do DC and first 15 AC
- coeff0 = load_tran_low(coeff_ptr + n_coeffs);
- coeff1 = load_tran_low(coeff_ptr + n_coeffs + 8);
+ // 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);
- // 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);
+
+ store_tran_low(qcoeff0, qcoeff_ptr + n_coeffs);
+ store_tran_low(qcoeff1, qcoeff_ptr + n_coeffs + 8);
+
+ coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
+ dequant = _mm_unpackhi_epi64(dequant, dequant);
+ coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
+
+ store_tran_low(coeff0, dqcoeff_ptr + n_coeffs);
+ store_tran_low(coeff1, dqcoeff_ptr + n_coeffs + 8);
+ }
+
+ {+ // 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;
+ }
+
+ thr = _mm_srai_epi16(dequant, 1);
+
+ // AC only loop
+ while (n_coeffs < 0) {+ __m128i coeff0, coeff1;
+ {+ __m128i coeff0_sign, coeff1_sign;
+ __m128i qcoeff0, qcoeff1;
+ __m128i qtmp0, qtmp1;
+
+ coeff0 = load_tran_low(coeff_ptr + n_coeffs);
+ coeff1 = load_tran_low(coeff_ptr + n_coeffs + 8);
+
+ // 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);
+
+ nzflag = _mm_movemask_epi8(_mm_cmpgt_epi16(qcoeff0, thr)) |
+ _mm_movemask_epi8(_mm_cmpgt_epi16(qcoeff1, thr));
+
+ if (nzflag) {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
@@ -81,131 +152,51 @@
store_tran_low(qcoeff1, qcoeff_ptr + n_coeffs + 8);
coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
- dequant = _mm_unpackhi_epi64(dequant, dequant);
coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
store_tran_low(coeff0, dqcoeff_ptr + n_coeffs);
store_tran_low(coeff1, dqcoeff_ptr + n_coeffs + 8);
- }
+ } else {+ store_zero_tran_low(qcoeff_ptr + n_coeffs);
+ store_zero_tran_low(qcoeff_ptr + n_coeffs + 8);
- {- // 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);
+ store_zero_tran_low(dqcoeff_ptr + n_coeffs);
+ store_zero_tran_low(dqcoeff_ptr + n_coeffs + 8);
}
- n_coeffs += 8 * 2;
}
- thr = _mm_srai_epi16(dequant, 1);
-
- // AC only loop
- while (n_coeffs < 0) {- __m128i coeff0, coeff1;
- {- __m128i coeff0_sign, coeff1_sign;
- __m128i qcoeff0, qcoeff1;
- __m128i qtmp0, qtmp1;
-
- coeff0 = load_tran_low(coeff_ptr + n_coeffs);
- coeff1 = load_tran_low(coeff_ptr + n_coeffs + 8);
-
- // 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);
-
- nzflag = _mm_movemask_epi8(_mm_cmpgt_epi16(qcoeff0, thr)) |
- _mm_movemask_epi8(_mm_cmpgt_epi16(qcoeff1, thr));
-
- if (nzflag) {- 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);
-
- store_tran_low(qcoeff0, qcoeff_ptr + n_coeffs);
- store_tran_low(qcoeff1, qcoeff_ptr + n_coeffs + 8);
-
- coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
- coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
-
- store_tran_low(coeff0, dqcoeff_ptr + n_coeffs);
- store_tran_low(coeff1, dqcoeff_ptr + n_coeffs + 8);
- } else {- store_zero_tran_low(qcoeff_ptr + n_coeffs);
- store_zero_tran_low(qcoeff_ptr + n_coeffs + 8);
-
- store_zero_tran_low(dqcoeff_ptr + n_coeffs);
- store_zero_tran_low(dqcoeff_ptr + n_coeffs + 8);
- }
- }
-
- if (nzflag) {- // 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;
+ if (nzflag) {+ // 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 {- store_zero_tran_low(qcoeff_ptr + n_coeffs);
- store_zero_tran_low(qcoeff_ptr + n_coeffs + 8);
-
- store_zero_tran_low(dqcoeff_ptr + n_coeffs);
- store_zero_tran_low(dqcoeff_ptr + n_coeffs + 8);
- n_coeffs += 8 * 2;
- } while (n_coeffs < 0);
- *eob_ptr = 0;
+ // 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);
}
}
--- a/vp9/encoder/x86/vp9_quantize_ssse3_x86_64.asm
+++ b/vp9/encoder/x86/vp9_quantize_ssse3_x86_64.asm
@@ -22,8 +22,6 @@
cglobal quantize_%1, 0, %2, 15, coeff, ncoeff, skip, round, quant, \
qcoeff, dqcoeff, dequant, \
eob, scan, iscan
- cmp dword skipm, 0
- jne .blank
; actual quantize loop - setup pointers, rounders, etc.
movifnidn coeffq, coeffmp
@@ -172,27 +170,6 @@
pmaxsw m8, m7
pextrw r6, m8, 0
mov [r2], r6
- RET
-
- ; skip-block, i.e. just write all zeroes
-.blank:
- mov r0, dqcoeffmp
- movifnidn ncoeffq, ncoeffmp
- mov r2, qcoeffmp
- mov r3, eobmp
-
- lea r0q, [r0q+ncoeffq*2]
- lea r2q, [r2q+ncoeffq*2]
- neg ncoeffq
- pxor m7, m7
-.blank_loop:
- STORE_ZERO_TRAN_LOW 7, r0q, ncoeffq
- STORE_ZERO_TRAN_LOW 7, r0q, ncoeffq + 8
- STORE_ZERO_TRAN_LOW 7, r2q, ncoeffq
- STORE_ZERO_TRAN_LOW 7, r2q, ncoeffq + 8
- add ncoeffq, mmsize
- jl .blank_loop
- mov word [r3q], 0
RET
%endmacro