ref: 8c42237bb200253931c49e2c530838f3a877dd65
parent: f60d1dcd3de46f72bafc5eeef481bd1a4e203301
author: Marco Paniconi <marpan@google.com>
date: Fri Aug 25 12:56:08 EDT 2017
Revert "quantize avx: copy 32x32 implementation" This reverts commit f60d1dcd3de46f72bafc5eeef481bd1a4e203301. Reason for revert: <INSERT REASONING HERE> Failures in AVX/VP9QuantizeTest in nightly tests. Original change's description: > quantize avx: copy 32x32 implementation > > Ensure avx and ssse3 stay in sync by testing them against each other. > > Change-Id: I699f3b48785c83260825402d7826231f475f697c TBR=slavarnway@google.com,johannkoenig@google.com,builds@webmproject.org Change-Id: Ibd38636212269328317dd0721be9d25452113d1c No-Presubmit: true No-Tree-Checks: true No-Try: true
--- a/test/vp9_quantize_test.cc
+++ b/test/vp9_quantize_test.cc
@@ -392,14 +392,16 @@
// TODO(johannkoenig): AVX optimizations do not yet pass the 32x32 test or
// highbitdepth configurations.
#if HAVE_AVX && !CONFIG_VP9_HIGHBITDEPTH
-INSTANTIATE_TEST_CASE_P(
- AVX, VP9QuantizeTest,
- ::testing::Values(make_tuple(&vpx_quantize_b_avx, &vpx_quantize_b_c,
- VPX_BITS_8, 16),
- // Even though SSSE3 and AVX do not match the reference
- // code, we can keep them in sync with each other.
- make_tuple(&vpx_quantize_b_32x32_avx,
- &vpx_quantize_b_32x32_ssse3, VPX_BITS_8, 32)));
+INSTANTIATE_TEST_CASE_P(AVX, VP9QuantizeTest,
+ ::testing::Values(make_tuple(&vpx_quantize_b_avx,
+ &vpx_quantize_b_c,
+ VPX_BITS_8, 16)));
+#if ARCH_X86_64
+INSTANTIATE_TEST_CASE_P(DISABLED_AVX, VP9QuantizeTest,
+ ::testing::Values(make_tuple(&vpx_quantize_b_32x32_avx,
+ &vpx_quantize_b_32x32_c,
+ VPX_BITS_8, 32)));
+#endif // ARCH_X86_64
#endif // HAVE_AVX && !CONFIG_VP9_HIGHBITDEPTH
// TODO(webm:1448): dqcoeff is not handled correctly in HBD builds.
--- a/vpx_dsp/vpx_dsp.mk
+++ b/vpx_dsp/vpx_dsp.mk
@@ -282,6 +282,9 @@
ifeq ($(CONFIG_VP9_HIGHBITDEPTH),yes)
DSP_SRCS-$(HAVE_SSE2) += x86/highbd_quantize_intrin_sse2.c
endif
+ifeq ($(ARCH_X86_64),yes)
+DSP_SRCS-$(HAVE_AVX) += x86/quantize_avx_x86_64.asm
+endif
# avg
DSP_SRCS-yes += avg.c
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -673,7 +673,7 @@
specialize qw/vpx_quantize_b neon sse2 ssse3 avx/;
add_proto qw/void vpx_quantize_b_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, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
- specialize qw/vpx_quantize_b_32x32 neon ssse3 avx/;
+ specialize qw/vpx_quantize_b_32x32 neon ssse3/, "$avx_x86_64";
if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {add_proto qw/void vpx_highbd_quantize_b/, "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, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
--- a/vpx_dsp/x86/quantize_avx.c
+++ b/vpx_dsp/x86/quantize_avx.c
@@ -46,7 +46,7 @@
// Setup global values.
zbin = _mm_load_si128((const __m128i *)zbin_ptr);
- // x86 has no "greater *or equal*" comparison. Subtract 1 from zbin so
+ // 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);
@@ -172,219 +172,6 @@
coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
coeff1 = _mm_mullo_epi16(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);
- 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);
- }
-}
-
-void vpx_quantize_b_32x32_avx(
- 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, uint16_t *eob_ptr,
- const int16_t *scan_ptr, const int16_t *iscan_ptr) {- const __m128i zero = _mm_setzero_si128();
- const __m128i one = _mm_set1_epi16(1);
- const __m256i big_zero = _mm256_setzero_si256();
- 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;
-
- (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);
- // Shift with rounding.
- zbin = _mm_add_epi16(zbin, one);
- zbin = _mm_srli_epi16(zbin, 1);
- // x86 has no "greater *or equal*" comparison. Subtract 1 from zbin so
- // it is a strict "greater" comparison.
- zbin = _mm_sub_epi16(zbin, one);
-
- round = _mm_load_si128((const __m128i *)round_ptr);
- round = _mm_add_epi16(round, one);
- round = _mm_srli_epi16(round, 1);
-
- 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);
- shift = _mm_slli_epi16(shift, 1);
-
- // Do DC and first 15 AC.
- coeff0 = load_tran_low(coeff_ptr);
- coeff1 = load_tran_low(coeff_ptr + 8);
-
- qcoeff0 = _mm_abs_epi16(coeff0);
- qcoeff1 = _mm_abs_epi16(coeff1);
-
- cmp_mask0 = _mm_cmpgt_epi16(qcoeff0, zbin);
- 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);
- if (_mm_test_all_zeros(all_zero, all_zero)) {- _mm256_store_si256((__m256i *)(qcoeff_ptr), big_zero);
- _mm256_store_si256((__m256i *)(dqcoeff_ptr), big_zero);
-#if CONFIG_VP9_HIGHBITDEPTH
- _mm256_store_si256((__m256i *)(qcoeff_ptr + 8), big_zero);
- _mm256_store_si256((__m256i *)(dqcoeff_ptr + 8), big_zero);
-#endif // CONFIG_VP9_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);
- 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);
-
- // Reinsert signs
- qcoeff0 = _mm_sign_epi16(qcoeff0, coeff0);
- qcoeff1 = _mm_sign_epi16(qcoeff1, coeff1);
-
- // Mask out zbin threshold coeffs
- qcoeff0 = _mm_and_si128(qcoeff0, cmp_mask0);
- qcoeff1 = _mm_and_si128(qcoeff1, cmp_mask1);
-
- store_tran_low(qcoeff0, qcoeff_ptr);
- store_tran_low(qcoeff1, qcoeff_ptr + 8);
-
- // Un-sign to bias rounding like C.
- // dequant is almost always negative, so this is probably the backwards way
- // to handle the sign. However, it matches the previous assembly.
- coeff0 = _mm_abs_epi16(qcoeff0);
- coeff1 = _mm_abs_epi16(qcoeff1);
-
- coeff0 = _mm_mullo_epi16(coeff0, dequant);
- dequant = _mm_unpackhi_epi64(dequant, dequant);
- coeff1 = _mm_mullo_epi16(coeff1, dequant);
-
- // "Divide" by 2.
- coeff0 = _mm_srli_epi16(coeff0, 1);
- coeff1 = _mm_srli_epi16(coeff1, 1);
-
- coeff0 = _mm_sign_epi16(coeff0, qcoeff0);
- coeff1 = _mm_sign_epi16(coeff1, qcoeff1);
-
- 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);
- }
-
- // AC only loop.
- for (index = 16; index < 32 * 32; index += 16) {- coeff0 = load_tran_low(coeff_ptr + index);
- coeff1 = load_tran_low(coeff_ptr + index + 8);
-
- qcoeff0 = _mm_abs_epi16(coeff0);
- qcoeff1 = _mm_abs_epi16(coeff1);
-
- cmp_mask0 = _mm_cmpgt_epi16(qcoeff0, zbin);
- cmp_mask1 = _mm_cmpgt_epi16(qcoeff1, zbin);
-
- all_zero = _mm_or_si128(cmp_mask0, cmp_mask1);
- if (_mm_test_all_zeros(all_zero, all_zero)) {- _mm256_store_si256((__m256i *)(qcoeff_ptr + index), big_zero);
- _mm256_store_si256((__m256i *)(dqcoeff_ptr + index), big_zero);
-#if CONFIG_VP9_HIGHBITDEPTH
- _mm256_store_si256((__m256i *)(qcoeff_ptr + index + 8), big_zero);
- _mm256_store_si256((__m256i *)(dqcoeff_ptr + index + 8), big_zero);
-#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);
-
- qcoeff0 = _mm_sign_epi16(qcoeff0, coeff0);
- qcoeff1 = _mm_sign_epi16(qcoeff1, coeff1);
-
- qcoeff0 = _mm_and_si128(qcoeff0, cmp_mask0);
- qcoeff1 = _mm_and_si128(qcoeff1, cmp_mask1);
-
- store_tran_low(qcoeff0, qcoeff_ptr + index);
- store_tran_low(qcoeff1, qcoeff_ptr + index + 8);
-
- coeff0 = _mm_abs_epi16(qcoeff0);
- coeff1 = _mm_abs_epi16(qcoeff1);
-
- coeff0 = _mm_mullo_epi16(coeff0, dequant);
- coeff1 = _mm_mullo_epi16(coeff1, dequant);
-
- coeff0 = _mm_srli_epi16(coeff0, 1);
- coeff1 = _mm_srli_epi16(coeff1, 1);
-
- coeff0 = _mm_sign_epi16(coeff0, qcoeff0);
- coeff1 = _mm_sign_epi16(coeff1, qcoeff1);
store_tran_low(coeff0, dqcoeff_ptr + index);
store_tran_low(coeff1, dqcoeff_ptr + index + 8);
--- /dev/null
+++ b/vpx_dsp/x86/quantize_avx_x86_64.asm
@@ -1,0 +1,497 @@
+;
+; Copyright (c) 2015 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 "third_party/x86inc/x86inc.asm"
+
+SECTION .text
+
+%macro QUANTIZE_FN 2
+cglobal quantize_%1, 0, %2, 15, coeff, ncoeff, skip, zbin, round, quant, \
+ shift, qcoeff, dqcoeff, dequant, \
+ eob, scan, iscan
+
+ vzeroupper
+
+%ifnidn %1, b_32x32
+
+ ; Special case for ncoeff == 16, as it is frequent and we can save on
+ ; not setting up a loop.
+ cmp ncoeffmp, 16
+ jne .generic
+
+ ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+ ;; Special case of ncoeff == 16
+ ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+.single:
+
+ movifnidn coeffq, coeffmp
+ movifnidn zbinq, zbinmp
+ mova m0, [zbinq] ; m0 = zbin
+
+ ; Get DC and first 15 AC coeffs - in this special case, that is all.
+%if CONFIG_VP9_HIGHBITDEPTH
+ ; coeff stored as 32bit numbers but we process them as 16 bit numbers
+ mova m9, [coeffq]
+ packssdw m9, [coeffq+16] ; m9 = c[i]
+ mova m10, [coeffq+32]
+ packssdw m10, [coeffq+48] ; m10 = c[i]
+%else
+ mova m9, [coeffq] ; m9 = c[i]
+ mova m10, [coeffq+16] ; m10 = c[i]
+%endif
+
+ mov r0, eobmp ; Output pointer
+ mov r1, qcoeffmp ; Output pointer
+ mov r2, dqcoeffmp ; Output pointer
+
+ pxor m5, m5 ; m5 = dedicated zero
+
+ pcmpeqw m4, m4 ; All word lanes -1
+ paddw m0, m4 ; m0 = zbin - 1
+
+ pabsw m6, m9 ; m6 = abs(m9)
+ pabsw m11, m10 ; m11 = abs(m10)
+ pcmpgtw m7, m6, m0 ; m7 = c[i] >= zbin
+ punpckhqdq m0, m0
+ pcmpgtw m12, m11, m0 ; m12 = c[i] >= zbin
+
+ ; Check if all coeffs are less than zbin. If yes, we just write zeros
+ ; to the outputs and we are done.
+ por m14, m7, m12
+ ptest m14, m14
+ jnz .single_nonzero
+
+%if CONFIG_VP9_HIGHBITDEPTH
+ mova [r1 ], ymm5
+ mova [r1+32], ymm5
+ mova [r2 ], ymm5
+ mova [r2+32], ymm5
+%else
+ mova [r1], ymm5
+ mova [r2], ymm5
+%endif
+ mov [r0], word 0
+
+ vzeroupper
+ RET
+
+.single_nonzero:
+
+ ; Actual quantization of size 16 block - setup pointers, rounders, etc.
+ movifnidn r4, roundmp
+ movifnidn r5, quantmp
+ mov r3, dequantmp
+ mov r6, shiftmp
+ mova m1, [r4] ; m1 = round
+ mova m2, [r5] ; m2 = quant
+ mova m3, [r3] ; m3 = dequant
+ mova m4, [r6] ; m4 = shift
+
+ mov r3, iscanmp
+
+ DEFINE_ARGS eob, qcoeff, dqcoeff, iscan
+
+ ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+ paddsw m6, m1 ; m6 += round
+ punpckhqdq m1, m1
+ paddsw m11, m1 ; m11 += round
+ pmulhw m8, m6, m2 ; m8 = m6*q>>16
+ punpckhqdq m2, m2
+ pmulhw m13, m11, m2 ; m13 = m11*q>>16
+ paddw m8, m6 ; m8 += m6
+ paddw m13, m11 ; m13 += m11
+ pmulhw m8, m4 ; m8 = m8*qsh>>16
+ punpckhqdq m4, m4
+ pmulhw m13, m4 ; m13 = m13*qsh>>16
+ psignw m8, m9 ; m8 = reinsert sign
+ psignw m13, m10 ; m13 = reinsert sign
+ pand m8, m7
+ pand m13, m12
+
+%if CONFIG_VP9_HIGHBITDEPTH
+ ; Store 16bit numbers as 32bit numbers in array pointed to by qcoeff
+ pcmpgtw m6, m5, m8
+ punpckhwd m6, m8, m6
+ pmovsxwd m11, m8
+ mova [qcoeffq ], m11
+ mova [qcoeffq+16], m6
+ pcmpgtw m6, m5, m13
+ punpckhwd m6, m13, m6
+ pmovsxwd m11, m13
+ mova [qcoeffq+32], m11
+ mova [qcoeffq+48], m6
+%else
+ mova [qcoeffq ], m8
+ mova [qcoeffq+16], m13
+%endif
+
+ pmullw m8, m3 ; dqc[i] = qc[i] * q
+ punpckhqdq m3, m3
+ pmullw m13, m3 ; dqc[i] = qc[i] * q
+
+%if CONFIG_VP9_HIGHBITDEPTH
+ ; Store 16bit numbers as 32bit numbers in array pointed to by qcoeff
+ pcmpgtw m6, m5, m8
+ punpckhwd m6, m8, m6
+ pmovsxwd m11, m8
+ mova [dqcoeffq ], m11
+ mova [dqcoeffq+16], m6
+ pcmpgtw m6, m5, m13
+ punpckhwd m6, m13, m6
+ pmovsxwd m11, m13
+ mova [dqcoeffq+32], m11
+ mova [dqcoeffq+48], m6
+%else
+ mova [dqcoeffq ], m8
+ mova [dqcoeffq+16], m13
+%endif
+
+ mova m6, [iscanq] ; m6 = scan[i]
+ mova m11, [iscanq+16] ; m11 = scan[i]
+
+ pcmpeqw m8, m8, m5 ; m8 = c[i] == 0
+ pcmpeqw m13, m13, m5 ; m13 = c[i] == 0
+ psubw m6, m6, m7 ; m6 = scan[i] + 1
+ psubw m11, m11, m12 ; m11 = scan[i] + 1
+ pandn m8, m8, m6 ; m8 = max(eob)
+ pandn m13, m13, m11 ; m13 = max(eob)
+ pmaxsw m8, m8, m13
+
+ ; Horizontally accumulate/max eobs and write into [eob] memory pointer
+ pshufd m7, m8, 0xe
+ pmaxsw m8, m7
+ pshuflw m7, m8, 0xe
+ pmaxsw m8, m7
+ pshuflw m7, m8, 0x1
+ pmaxsw m8, m7
+ movq rax, m8
+ mov [eobq], ax
+
+ vzeroupper
+ RET
+
+ ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+ ;; Generic case of ncoeff != 16
+ ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+.generic:
+
+%endif ; %ifnidn %1, b_32x32
+
+DEFINE_ARGS coeff, ncoeff, skip, zbin, round, quant, shift, \
+ qcoeff, dqcoeff, dequant, eob, scan, iscan
+
+ ; Actual quantization loop - setup pointers, rounders, etc.
+ movifnidn coeffq, coeffmp
+ movifnidn ncoeffq, ncoeffmp
+ mov r2, dequantmp
+ movifnidn zbinq, zbinmp
+ movifnidn roundq, roundmp
+ movifnidn quantq, quantmp
+ mova m0, [zbinq] ; m0 = zbin
+ mova m1, [roundq] ; m1 = round
+ mova m2, [quantq] ; m2 = quant
+ mova m3, [r2] ; m3 = dequant
+ pcmpeqw m4, m4 ; All lanes -1
+%ifidn %1, b_32x32
+ psubw m0, m4
+ psubw m1, m4
+ psrlw m0, 1 ; m0 = (m0 + 1) / 2
+ psrlw m1, 1 ; m1 = (m1 + 1) / 2
+%endif
+ paddw m0, m4 ; m0 = m0 + 1
+
+ mov r2, shiftmp
+ mov r3, qcoeffmp
+ mova m4, [r2] ; m4 = shift
+ mov r4, dqcoeffmp
+ mov r5, iscanmp
+%ifidn %1, b_32x32
+ psllw m4, 1
+%endif
+ pxor m5, m5 ; m5 = dedicated zero
+
+ DEFINE_ARGS coeff, ncoeff, d1, qcoeff, dqcoeff, iscan, d2, d3, d4, d5, eob
+
+%if CONFIG_VP9_HIGHBITDEPTH
+ lea coeffq, [ coeffq+ncoeffq*4]
+ lea qcoeffq, [ qcoeffq+ncoeffq*4]
+ lea dqcoeffq, [dqcoeffq+ncoeffq*4]
+%else
+ lea coeffq, [ coeffq+ncoeffq*2]
+ lea qcoeffq, [ qcoeffq+ncoeffq*2]
+ lea dqcoeffq, [dqcoeffq+ncoeffq*2]
+%endif
+ lea iscanq, [ iscanq+ncoeffq*2]
+ neg ncoeffq
+
+ ; get DC and first 15 AC coeffs
+%if CONFIG_VP9_HIGHBITDEPTH
+ ; coeff stored as 32bit numbers & require 16bit numbers
+ mova m9, [coeffq+ncoeffq*4+ 0]
+ packssdw m9, [coeffq+ncoeffq*4+16]
+ mova m10, [coeffq+ncoeffq*4+32]
+ packssdw m10, [coeffq+ncoeffq*4+48]
+%else
+ mova m9, [coeffq+ncoeffq*2+ 0] ; m9 = c[i]
+ mova m10, [coeffq+ncoeffq*2+16] ; m10 = c[i]
+%endif
+
+ pabsw m6, m9 ; m6 = abs(m9)
+ pabsw m11, m10 ; m11 = abs(m10)
+ pcmpgtw m7, m6, m0 ; m7 = c[i] >= zbin
+ punpckhqdq m0, m0
+ pcmpgtw m12, m11, m0 ; m12 = c[i] >= zbin
+
+ ; Check if all coeffs are less than zbin. If yes, skip forward quickly.
+ por m14, m7, m12
+ ptest m14, m14
+ jnz .first_nonzero
+
+%if CONFIG_VP9_HIGHBITDEPTH
+ mova [qcoeffq+ncoeffq*4 ], ymm5
+ mova [qcoeffq+ncoeffq*4+32], ymm5
+ mova [dqcoeffq+ncoeffq*4 ], ymm5
+ mova [dqcoeffq+ncoeffq*4+32], ymm5
+%else
+ mova [qcoeffq+ncoeffq*2], ymm5
+ mova [dqcoeffq+ncoeffq*2], ymm5
+%endif
+
+ add ncoeffq, mmsize
+
+ punpckhqdq m1, m1
+ punpckhqdq m2, m2
+ punpckhqdq m3, m3
+ punpckhqdq m4, m4
+ pxor m8, m8
+
+ jmp .ac_only_loop
+
+.first_nonzero:
+
+ paddsw m6, m1 ; m6 += round
+ punpckhqdq m1, m1
+ paddsw m11, m1 ; m11 += round
+ pmulhw m8, m6, m2 ; m8 = m6*q>>16
+ punpckhqdq m2, m2
+ pmulhw m13, m11, m2 ; m13 = m11*q>>16
+ paddw m8, m6 ; m8 += m6
+ paddw m13, m11 ; m13 += m11
+ pmulhw m8, m4 ; m8 = m8*qsh>>16
+ punpckhqdq m4, m4
+ pmulhw m13, m4 ; m13 = m13*qsh>>16
+ psignw m8, m9 ; m8 = reinsert sign
+ psignw m13, m10 ; m13 = reinsert sign
+ pand m8, m7
+ pand m13, m12
+
+%if CONFIG_VP9_HIGHBITDEPTH
+ ; store 16bit numbers as 32bit numbers in array pointed to by qcoeff
+ pcmpgtw m6, m5, m8
+ punpckhwd m6, m8, m6
+ pmovsxwd m11, m8
+ mova [qcoeffq+ncoeffq*4+ 0], m11
+ mova [qcoeffq+ncoeffq*4+16], m6
+ pcmpgtw m6, m5, m13
+ punpckhwd m6, m13, m6
+ pmovsxwd m11, m13
+ mova [qcoeffq+ncoeffq*4+32], m11
+ mova [qcoeffq+ncoeffq*4+48], m6
+%else
+ mova [qcoeffq+ncoeffq*2+ 0], m8
+ mova [qcoeffq+ncoeffq*2+16], m13
+%endif
+
+%ifidn %1, b_32x32
+ pabsw m8, m8
+ pabsw m13, m13
+%endif
+ pmullw m8, m3 ; dqc[i] = qc[i] * q
+ punpckhqdq m3, m3
+ pmullw m13, m3 ; dqc[i] = qc[i] * q
+%ifidn %1, b_32x32
+ psrlw m8, 1
+ psrlw m13, 1
+ psignw m8, m9
+ psignw m13, m10
+%endif
+
+%if CONFIG_VP9_HIGHBITDEPTH
+ ; store 16bit numbers as 32bit numbers in array pointed to by qcoeff
+ pcmpgtw m6, m5, m8
+ punpckhwd m6, m8, m6
+ pmovsxwd m11, m8
+ mova [dqcoeffq+ncoeffq*4+ 0], m11
+ mova [dqcoeffq+ncoeffq*4+16], m6
+ pcmpgtw m6, m5, m13
+ punpckhwd m6, m13, m6
+ pmovsxwd m11, m13
+ mova [dqcoeffq+ncoeffq*4+32], m11
+ mova [dqcoeffq+ncoeffq*4+48], m6
+%else
+ mova [dqcoeffq+ncoeffq*2+ 0], m8
+ mova [dqcoeffq+ncoeffq*2+16], m13
+%endif
+
+ pcmpeqw m8, m5 ; m8 = c[i] == 0
+ pcmpeqw m13, m5 ; m13 = c[i] == 0
+ mova m6, [iscanq+ncoeffq*2] ; m6 = scan[i]
+ mova m11, [iscanq+ncoeffq*2+16] ; m11 = scan[i]
+ psubw m6, m7 ; m6 = scan[i] + 1
+ psubw m11, m12 ; m11 = scan[i] + 1
+ pandn m8, m6 ; m8 = max(eob)
+ pandn m13, m11 ; m13 = max(eob)
+ pmaxsw m8, m13
+ add ncoeffq, mmsize
+
+.ac_only_loop:
+
+%if CONFIG_VP9_HIGHBITDEPTH
+ ; pack coeff from 32bit to 16bit array
+ mova m9, [coeffq+ncoeffq*4+ 0]
+ packssdw m9, [coeffq+ncoeffq*4+16]
+ mova m10, [coeffq+ncoeffq*4+32]
+ packssdw m10, [coeffq+ncoeffq*4+48]
+%else
+ mova m9, [coeffq+ncoeffq*2+ 0] ; m9 = c[i]
+ mova m10, [coeffq+ncoeffq*2+16] ; m10 = c[i]
+%endif
+
+ pabsw m6, m9 ; m6 = abs(m9)
+ pabsw m11, m10 ; m11 = abs(m10)
+ pcmpgtw m7, m6, m0 ; m7 = c[i] >= zbin
+ pcmpgtw m12, m11, m0 ; m12 = c[i] >= zbin
+
+ ; Check if all coeffs are less than zbin. If yes, skip this itertion.
+ ; And just write zeros as the result would be.
+ por m14, m7, m12
+ ptest m14, m14
+ jnz .rest_nonzero
+
+%if CONFIG_VP9_HIGHBITDEPTH
+ mova [qcoeffq+ncoeffq*4+ 0], ymm5
+ mova [qcoeffq+ncoeffq*4+32], ymm5
+ mova [dqcoeffq+ncoeffq*4+ 0], ymm5
+ mova [dqcoeffq+ncoeffq*4+32], ymm5
+%else
+ mova [qcoeffq+ncoeffq*2+ 0], ymm5
+ mova [dqcoeffq+ncoeffq*2+ 0], ymm5
+%endif
+ add ncoeffq, mmsize
+ jnz .ac_only_loop
+
+ ; Horizontally accumulate/max eobs and write into [eob] memory pointer
+ mov r2, eobmp
+ pshufd m7, m8, 0xe
+ pmaxsw m8, m7
+ pshuflw m7, m8, 0xe
+ pmaxsw m8, m7
+ pshuflw m7, m8, 0x1
+ pmaxsw m8, m7
+ movq rax, m8
+ mov [r2], ax
+ vzeroupper
+ RET
+
+.rest_nonzero:
+ paddsw m6, m1 ; m6 += round
+ paddsw m11, m1 ; m11 += round
+ pmulhw m14, m6, m2 ; m14 = m6*q>>16
+ pmulhw m13, m11, m2 ; m13 = m11*q>>16
+ paddw m14, m6 ; m14 += m6
+ paddw m13, m11 ; m13 += m11
+ pmulhw m14, m4 ; m14 = m14*qsh>>16
+ pmulhw m13, m4 ; m13 = m13*qsh>>16
+ psignw m14, m9 ; m14 = reinsert sign
+ psignw m13, m10 ; m13 = reinsert sign
+ pand m14, m7
+ pand m13, m12
+
+%if CONFIG_VP9_HIGHBITDEPTH
+ ; store 16bit numbers as 32bit numbers in array pointed to by qcoeff
+ pcmpgtw m6, m5, m14
+ punpckhwd m6, m14, m6
+ pmovsxwd m11, m14
+ mova [qcoeffq+ncoeffq*4+ 0], m11
+ mova [qcoeffq+ncoeffq*4+16], m6
+ pcmpgtw m6, m5, m13
+ punpckhwd m6, m13, m6
+ pmovsxwd m11, m13
+ mova [qcoeffq+ncoeffq*4+32], m11
+ mova [qcoeffq+ncoeffq*4+48], m6
+%else
+ mova [qcoeffq+ncoeffq*2+ 0], m14
+ mova [qcoeffq+ncoeffq*2+16], m13
+%endif
+
+%ifidn %1, b_32x32
+ pabsw m14, m14
+ pabsw m13, m13
+%endif
+ pmullw m14, m3 ; dqc[i] = qc[i] * q
+ pmullw m13, m3 ; dqc[i] = qc[i] * q
+%ifidn %1, b_32x32
+ psrlw m14, 1
+ psrlw m13, 1
+ psignw m14, m9
+ psignw m13, m10
+%endif
+
+%if CONFIG_VP9_HIGHBITDEPTH
+ ; store 16bit numbers as 32bit numbers in array pointed to by qcoeff
+ pcmpgtw m6, m5, m14
+ punpckhwd m6, m14, m6
+ pmovsxwd m11, m14
+ mova [dqcoeffq+ncoeffq*4+ 0], m11
+ mova [dqcoeffq+ncoeffq*4+16], m6
+ pcmpgtw m6, m5, m13
+ punpckhwd m6, m13, m6
+ pmovsxwd m11, m13
+ mova [dqcoeffq+ncoeffq*4+32], m11
+ mova [dqcoeffq+ncoeffq*4+48], m6
+%else
+ mova [dqcoeffq+ncoeffq*2+ 0], m14
+ mova [dqcoeffq+ncoeffq*2+16], m13
+%endif
+
+ pcmpeqw m14, m5 ; m14 = c[i] == 0
+ pcmpeqw m13, m5 ; m13 = c[i] == 0
+ mova m6, [iscanq+ncoeffq*2+ 0] ; m6 = scan[i]
+ mova m11, [iscanq+ncoeffq*2+16] ; m11 = scan[i]
+ psubw m6, m7 ; m6 = scan[i] + 1
+ psubw m11, m12 ; m11 = scan[i] + 1
+ pandn m14, m6 ; m14 = max(eob)
+ pandn m13, m11 ; m13 = max(eob)
+ pmaxsw m8, m14
+ pmaxsw m8, m13
+ add ncoeffq, mmsize
+ jnz .ac_only_loop
+
+ ; Horizontally accumulate/max eobs and write into [eob] memory pointer
+ mov r2, eobmp
+ pshufd m7, m8, 0xe
+ pmaxsw m8, m7
+ pshuflw m7, m8, 0xe
+ pmaxsw m8, m7
+ pshuflw m7, m8, 0x1
+ pmaxsw m8, m7
+ movq rax, m8
+ mov [r2], ax
+ vzeroupper
+ RET
+%endmacro
+
+INIT_XMM avx
+QUANTIZE_FN b_32x32, 7
+
+END