quantize_b and *_32x32: implementation in avx intrinsics
BUG=1075935
Change-Id: Ie8a093ff2a26f640ca5080ad70052932b3427aaa
diff --git a/aom_dsp/aom_dsp.cmake b/aom_dsp/aom_dsp.cmake
index 6dbb2f5..0fac829 100644
--- a/aom_dsp/aom_dsp.cmake
+++ b/aom_dsp/aom_dsp.cmake
@@ -250,8 +250,8 @@
"${AOM_ROOT}/aom_dsp/x86/blk_sse_sum_avx2.c"
"${AOM_ROOT}/aom_dsp/x86/sum_squares_avx2.c")
- list(APPEND AOM_DSP_ENCODER_AVX_ASM_X86_64
- "${AOM_ROOT}/aom_dsp/x86/quantize_avx_x86_64.asm")
+ list(APPEND AOM_DSP_ENCODER_INTRIN_AVX
+ "${AOM_ROOT}/aom_dsp/x86/aom_quantize_avx.c")
list(APPEND AOM_DSP_ENCODER_INTRIN_SSSE3
"${AOM_ROOT}/aom_dsp/x86/masked_sad_intrin_ssse3.h"
@@ -375,9 +375,10 @@
endif()
endif()
- if(HAVE_AVX AND "${AOM_TARGET_CPU}" STREQUAL "x86_64")
+ if(HAVE_AVX)
if(CONFIG_AV1_ENCODER)
- add_asm_library("aom_dsp_encoder_avx" "AOM_DSP_ENCODER_AVX_ASM_X86_64")
+ add_intrinsics_object_library("-mavx" "avx" "aom_dsp_encoder"
+ "AOM_DSP_ENCODER_INTRIN_AVX")
endif()
endif()
diff --git a/aom_dsp/aom_dsp_rtcd_defs.pl b/aom_dsp/aom_dsp_rtcd_defs.pl
index e0ce1a1..d9eb518 100755
--- a/aom_dsp/aom_dsp_rtcd_defs.pl
+++ b/aom_dsp/aom_dsp_rtcd_defs.pl
@@ -529,13 +529,13 @@
#
if (aom_config("CONFIG_AV1_ENCODER") eq "yes") {
add_proto qw/void aom_quantize_b/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, 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/aom_quantize_b sse2 neon/, "$ssse3_x86_64", "$avx_x86_64";
+ specialize qw/aom_quantize_b sse2 neon avx/, "$ssse3_x86_64";
add_proto qw/void aom_quantize_b_adaptive/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
specialize qw/aom_quantize_b_adaptive sse2 avx2/;
add_proto qw/void aom_quantize_b_32x32/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
- specialize qw/aom_quantize_b_32x32 neon/, "$ssse3_x86_64", "$avx_x86_64";
+ specialize qw/aom_quantize_b_32x32 neon avx/, "$ssse3_x86_64";
add_proto qw/void aom_quantize_b_32x32_adaptive/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, 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/aom_quantize_b_32x32_adaptive sse2/;
diff --git a/aom_dsp/x86/aom_quantize_avx.c b/aom_dsp/x86/aom_quantize_avx.c
new file mode 100644
index 0000000..b2d6d4b
--- /dev/null
+++ b/aom_dsp/x86/aom_quantize_avx.c
@@ -0,0 +1,282 @@
+/*
+ * Copyright (c) 2020, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <immintrin.h>
+
+#include "config/aom_dsp_rtcd.h"
+#include "aom/aom_integer.h"
+#include "aom_dsp/x86/bitdepth_conversion_sse2.h"
+#include "aom_dsp/x86/quantize_x86.h"
+
+static INLINE void calculate_dqcoeff_and_store(__m128i qcoeff, __m128i dequant,
+ tran_low_t *dqcoeff) {
+ const __m128i low = _mm_mullo_epi16(qcoeff, dequant);
+ const __m128i high = _mm_mulhi_epi16(qcoeff, dequant);
+
+ const __m128i dqcoeff32_0 = _mm_unpacklo_epi16(low, high);
+ const __m128i dqcoeff32_1 = _mm_unpackhi_epi16(low, high);
+
+ _mm_store_si128((__m128i *)(dqcoeff), dqcoeff32_0);
+ _mm_store_si128((__m128i *)(dqcoeff + 4), dqcoeff32_1);
+}
+
+void aom_quantize_b_avx(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+ 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) {
+ const __m128i zero = _mm_setzero_si128();
+ 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 eob = zero, eob0;
+
+ (void)scan;
+
+ *eob_ptr = 0;
+
+ 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);
+ 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);
+ _mm256_store_si256((__m256i *)(qcoeff_ptr + 8), big_zero);
+ _mm256_store_si256((__m256i *)(dqcoeff_ptr + 8), big_zero);
+
+ if (n_coeffs == 16) return;
+
+ 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 {
+ calculate_qcoeff(&qcoeff0, round, quant, shift);
+ round = _mm_unpackhi_epi64(round, round);
+ quant = _mm_unpackhi_epi64(quant, quant);
+ shift = _mm_unpackhi_epi64(shift, shift);
+ calculate_qcoeff(&qcoeff1, round, quant, 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);
+
+ calculate_dqcoeff_and_store(qcoeff0, dequant, dqcoeff_ptr);
+ dequant = _mm_unpackhi_epi64(dequant, dequant);
+ calculate_dqcoeff_and_store(qcoeff1, dequant, dqcoeff_ptr + 8);
+
+ eob =
+ scan_for_eob(&qcoeff0, &qcoeff1, cmp_mask0, cmp_mask1, iscan, 0, zero);
+ }
+
+ // AC only loop.
+ for (index = 16; index < n_coeffs; 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);
+ _mm256_store_si256((__m256i *)(qcoeff_ptr + index + 8), big_zero);
+ _mm256_store_si256((__m256i *)(dqcoeff_ptr + index + 8), big_zero);
+ continue;
+ }
+
+ calculate_qcoeff(&qcoeff0, round, quant, shift);
+ calculate_qcoeff(&qcoeff1, round, quant, 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);
+
+ calculate_dqcoeff_and_store(qcoeff0, dequant, dqcoeff_ptr + index);
+ calculate_dqcoeff_and_store(qcoeff1, dequant, dqcoeff_ptr + index + 8);
+
+ eob0 = scan_for_eob(&qcoeff0, &qcoeff1, cmp_mask0, cmp_mask1, iscan, index,
+ zero);
+ eob = _mm_max_epi16(eob, eob0);
+ }
+
+ *eob_ptr = accumulate_eob(eob);
+}
+
+void aom_quantize_b_32x32_avx(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+ 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) {
+ const __m128i zero = _mm_setzero_si128();
+ const __m128i one = _mm_set1_epi16(1);
+ const __m256i big_zero = _mm256_setzero_si256();
+ int index;
+ const int log_scale = 1;
+
+ __m128i zbin, round, quant, dequant, shift;
+ __m128i coeff0, coeff1;
+ __m128i qcoeff0, qcoeff1;
+ __m128i cmp_mask0, cmp_mask1;
+ __m128i all_zero;
+ __m128i eob = zero, eob0;
+
+ (void)scan;
+
+ // 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);
+
+ // 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);
+ _mm256_store_si256((__m256i *)(qcoeff_ptr + 8), big_zero);
+ _mm256_store_si256((__m256i *)(dqcoeff_ptr + 8), big_zero);
+
+ 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 {
+ calculate_qcoeff_log_scale(&qcoeff0, round, quant, &shift, &log_scale);
+ round = _mm_unpackhi_epi64(round, round);
+ quant = _mm_unpackhi_epi64(quant, quant);
+ shift = _mm_unpackhi_epi64(shift, shift);
+ calculate_qcoeff_log_scale(&qcoeff1, round, quant, &shift, &log_scale);
+
+ // 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);
+
+ calculate_dqcoeff_and_store_log_scale(qcoeff0, dequant, zero, dqcoeff_ptr,
+ &log_scale);
+ dequant = _mm_unpackhi_epi64(dequant, dequant);
+ calculate_dqcoeff_and_store_log_scale(qcoeff1, dequant, zero,
+ dqcoeff_ptr + 8, &log_scale);
+
+ eob =
+ scan_for_eob(&qcoeff0, &qcoeff1, cmp_mask0, cmp_mask1, iscan, 0, zero);
+ }
+
+ // AC only loop.
+ for (index = 16; index < n_coeffs; 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);
+ _mm256_store_si256((__m256i *)(qcoeff_ptr + index + 8), big_zero);
+ _mm256_store_si256((__m256i *)(dqcoeff_ptr + index + 8), big_zero);
+ continue;
+ }
+
+ calculate_qcoeff_log_scale(&qcoeff0, round, quant, &shift, &log_scale);
+ calculate_qcoeff_log_scale(&qcoeff1, round, quant, &shift, &log_scale);
+
+ 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);
+
+ calculate_dqcoeff_and_store_log_scale(qcoeff0, dequant, zero,
+ dqcoeff_ptr + index, &log_scale);
+ calculate_dqcoeff_and_store_log_scale(qcoeff1, dequant, zero,
+ dqcoeff_ptr + index + 8, &log_scale);
+
+ eob0 = scan_for_eob(&qcoeff0, &qcoeff1, cmp_mask0, cmp_mask1, iscan, index,
+ zero);
+ eob = _mm_max_epi16(eob, eob0);
+ }
+
+ *eob_ptr = accumulate_eob(eob);
+}
diff --git a/aom_dsp/x86/quantize_avx_x86_64.asm b/aom_dsp/x86/quantize_avx_x86_64.asm
deleted file mode 100644
index d6e15c4..0000000
--- a/aom_dsp/x86/quantize_avx_x86_64.asm
+++ /dev/null
@@ -1,464 +0,0 @@
-;
-; Copyright (c) 2016, Alliance for Open Media. All rights reserved
-;
-; This source code is subject to the terms of the BSD 2 Clause License and
-; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
-; was not distributed with this source code in the LICENSE file, you can
-; obtain it at www.aomedia.org/license/software. If the Alliance for Open
-; Media Patent License 1.0 was not distributed with this source code in the
-; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
-;
-
-;
-
-%include "third_party/x86inc/x86inc.asm"
-
-SECTION .text
-
-%macro QUANTIZE_FN 2
-cglobal quantize_%1, 0, %2, 15, coeff, ncoeff, 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.
- ; 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]
-
- 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
-
- mova [r1 ], ymm5
- mova [r1+32], ymm5
- mova [r2 ], ymm5
- mova [r2+32], ymm5
- mov [r0], word 0
-
- vzeroupper
- RET
-
-.single_nonzero:
-
- ; Actual quantization of size 16 block - setup pointers, rounders, etc.
- movifnidn r3, roundmp
- movifnidn r4, quantmp
- mov r6, dequantmp
- mov r5, shiftmp
- mova m1, [r3] ; m1 = round
- mova m2, [r4] ; m2 = quant
- mova m3, [r6] ; m3 = dequant
- mova m4, [r5] ; 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
-
- ; 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
-
- pmullw m8, m3 ; dqc[i] = qc[i] * q
- punpckhqdq m3, m3
- pmullw m13, m3 ; dqc[i] = qc[i] * q
-
- ; Store 16bit numbers as 32bit numbers in array pointed to by dqcoeff
- 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
-
- 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, zbin, round, quant, shift, \
- qcoeff, dqcoeff, dequant, eob, scan, iscan
-
- ; Actual quantization loop - setup pointers, rounders, etc.
- movifnidn coeffq, coeffmp
- movifnidn ncoeffq, ncoeffmp
- movifnidn zbinq, zbinmp
- movifnidn roundq, roundmp
- movifnidn quantq, quantmp
- movifnidn dequantq, dequantmp
- mova m0, [zbinq] ; m0 = zbin
- mova m1, [roundq] ; m1 = round
- mova m2, [quantq] ; m2 = quant
- mova m3, [dequantq] ; 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
- pxor m5, m5 ; m5 = dedicated zero
-
- DEFINE_ARGS coeff, ncoeff, d1, qcoeff, dqcoeff, iscan, d2, d3, d4, eob
-
-
- lea coeffq, [ coeffq+ncoeffq*4]
- lea qcoeffq, [ qcoeffq+ncoeffq*4]
- lea dqcoeffq, [dqcoeffq+ncoeffq*4]
-
- lea iscanq, [ iscanq+ncoeffq*2]
- neg ncoeffq
-
- ; get DC and first 15 AC coeffs
- ; 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]
-
- 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
-
- mova [qcoeffq+ncoeffq*4 ], ymm5
- mova [qcoeffq+ncoeffq*4+32], ymm5
- mova [dqcoeffq+ncoeffq*4 ], ymm5
- mova [dqcoeffq+ncoeffq*4+32], ymm5
- 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
- %ifidn %1, b_32x32
- pmullw m5, m8, m4 ; store the lower 16 bits of m8*qsh
- %endif
- pmulhw m8, m4 ; m8 = m8*qsh>>16
- %ifidn %1, b_32x32
- psllw m8, 1
- psrlw m5, 15
- por m8, m5
- %endif
- punpckhqdq m4, m4
- %ifidn %1, b_32x32
- pmullw m5, m13, m4 ; store the lower 16 bits of m13*qsh
- %endif
- pmulhw m13, m4 ; m13 = m13*qsh>>16
- %ifidn %1, b_32x32
- psllw m13, 1
- psrlw m5, 15
- por m13, m5
- pxor m5, m5 ; reset m5 to zero register
- %endif
- psignw m8, m9 ; m8 = reinsert sign
- psignw m13, m10 ; m13 = reinsert sign
- pand m8, m7
- pand m13, m12
-
- ; 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
-
-%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
-
- ; store 16bit numbers as 32bit numbers in array pointed to by dqcoeff
- 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
-
- 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:
-
- ; 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]
-
- 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
-
- mova [qcoeffq+ncoeffq*4+ 0], ymm5
- mova [qcoeffq+ncoeffq*4+32], ymm5
- mova [dqcoeffq+ncoeffq*4+ 0], ymm5
- mova [dqcoeffq+ncoeffq*4+32], ymm5
-
- 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
- %ifidn %1, b_32x32
- pmullw m5, m14, m4 ; store the lower 16 bits of m14*qsh
- %endif
- pmulhw m14, m4 ; m14 = m14*qsh>>16
- %ifidn %1, b_32x32
- psllw m14, 1
- psrlw m5, 15
- por m14, m5
- pmullw m5, m13, m4 ; store the lower 16 bits of m13*qsh
- %endif
- pmulhw m13, m4 ; m13 = m13*qsh>>16
- %ifidn %1, b_32x32
- psllw m13, 1
- psrlw m5, 15
- por m13, m5
- pxor m5, m5 ; reset m5 to zero register
- %endif
- psignw m14, m9 ; m14 = reinsert sign
- psignw m13, m10 ; m13 = reinsert sign
- pand m14, m7
- pand m13, m12
-
- ; 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
-
-%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
-
- ; store 16bit numbers as 32bit numbers in array pointed to by dqcoeff
- 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
-
- 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, 9
-QUANTIZE_FN b_32x32, 9
diff --git a/test/quantize_func_test.cc b/test/quantize_func_test.cc
index d69820e..a347c38 100644
--- a/test/quantize_func_test.cc
+++ b/test/quantize_func_test.cc
@@ -544,7 +544,7 @@
#endif // HAVE_SSSE3 && ARCH_X86_64
-#if HAVE_AVX && ARCH_X86_64
+#if HAVE_AVX
INSTANTIATE_TEST_SUITE_P(
AVX, QuantizeTest,
::testing::Values(
@@ -553,5 +553,5 @@
make_tuple(&aom_quantize_b_32x32_c, &aom_quantize_b_32x32_avx,
static_cast<TX_SIZE>(TX_32X32), TYPE_B, AOM_BITS_8)));
-#endif // HAVE_AVX && ARCH_X86_64
+#endif // HAVE_AVX
} // namespace
