#include "vpx/vpx_integer.h"
#include "vpx_ports/msvc.h"
#include "vpx_ports/vpx_timer.h"
-#include "vpx_dsp/quantize.h"
using libvpx_test::ACMRandom;
using libvpx_test::Buffer;
#if CONFIG_VP9_HIGHBITDEPTH
INSTANTIATE_TEST_SUITE_P(
SSE2, VP9QuantizeTest,
- ::testing::Values(make_tuple(&vpx_quantize_b_sse2, &vpx_quantize_b_c,
- VPX_BITS_8, 16, false),
- make_tuple(&vpx_quantize_b_sse2, &vpx_highbd_quantize_b_c,
- VPX_BITS_10, 16, false),
- make_tuple(&vpx_quantize_b_sse2, &vpx_highbd_quantize_b_c,
- VPX_BITS_12, 16, false)));
+ ::testing::Values(
+ make_tuple(&vpx_quantize_b_sse2, &vpx_quantize_b_c, VPX_BITS_8, 16,
+ false),
+ make_tuple(&vpx_highbd_quantize_b_sse2, &vpx_highbd_quantize_b_c,
+ VPX_BITS_8, 16, false),
+ make_tuple(&vpx_highbd_quantize_b_sse2, &vpx_highbd_quantize_b_c,
+ VPX_BITS_10, 16, false),
+ make_tuple(&vpx_highbd_quantize_b_sse2, &vpx_highbd_quantize_b_c,
+ VPX_BITS_12, 16, false),
+ make_tuple(&vpx_highbd_quantize_b_32x32_sse2,
+ &vpx_highbd_quantize_b_32x32_c, VPX_BITS_8, 32, false),
+ make_tuple(&vpx_highbd_quantize_b_32x32_sse2,
+ &vpx_highbd_quantize_b_32x32_c, VPX_BITS_10, 32, false),
+ make_tuple(&vpx_highbd_quantize_b_32x32_sse2,
+ &vpx_highbd_quantize_b_32x32_c, VPX_BITS_12, 32, false)));
+
#else
INSTANTIATE_TEST_SUITE_P(
SSE2, VP9QuantizeTest,
#endif // HAVE_SSSE3
#if HAVE_AVX
-#if CONFIG_VP9_HIGHBITDEPTH
-INSTANTIATE_TEST_SUITE_P(
- AVX, VP9QuantizeTest,
- ::testing::Values(
- make_tuple(&vpx_quantize_b_avx, &vpx_quantize_b_c, VPX_BITS_8, 16,
- false),
- make_tuple(&vpx_quantize_b_avx, &vpx_highbd_quantize_b_c, VPX_BITS_10,
- 16, false),
- make_tuple(&vpx_quantize_b_avx, &vpx_highbd_quantize_b_c, VPX_BITS_12,
- 16, false),
- make_tuple(&vpx_quantize_b_32x32_avx, &vpx_highbd_quantize_b_32x32_c,
- VPX_BITS_8, 32, false),
- make_tuple(&vpx_quantize_b_32x32_avx, &vpx_highbd_quantize_b_32x32_c,
- VPX_BITS_10, 32, false),
- make_tuple(&vpx_quantize_b_32x32_avx, &vpx_highbd_quantize_b_32x32_c,
- VPX_BITS_12, 32, false)));
-
-#else
INSTANTIATE_TEST_SUITE_P(AVX, VP9QuantizeTest,
::testing::Values(make_tuple(&vpx_quantize_b_avx,
&vpx_quantize_b_c,
make_tuple(&vpx_quantize_b_32x32_avx,
&vpx_quantize_b_32x32_c,
VPX_BITS_8, 32, false)));
-#endif // CONFIG_VP9_HIGHBITDEPTH
#endif // HAVE_AVX
#if VPX_ARCH_X86_64 && HAVE_AVX2
switch (tx_size) {
case TX_32X32:
highbd_fdct32x32(x->use_lp32x32fdct, src_diff, coeff, diff_stride);
- vpx_quantize_b_32x32(coeff, 1024, p->zbin, p->round, p->quant,
- p->quant_shift, qcoeff, dqcoeff, pd->dequant, eob,
- scan_order->scan, scan_order->iscan);
+ vpx_highbd_quantize_b_32x32(
+ coeff, 1024, p->zbin, p->round, p->quant, p->quant_shift, qcoeff,
+ dqcoeff, pd->dequant, eob, scan_order->scan, scan_order->iscan);
break;
case TX_16X16:
vpx_highbd_fdct16x16(src_diff, coeff, diff_stride);
- vpx_quantize_b(coeff, 256, p->zbin, p->round, p->quant, p->quant_shift,
- qcoeff, dqcoeff, pd->dequant, eob, scan_order->scan,
- scan_order->iscan);
+ vpx_highbd_quantize_b(coeff, 256, p->zbin, p->round, p->quant,
+ p->quant_shift, qcoeff, dqcoeff, pd->dequant, eob,
+ scan_order->scan, scan_order->iscan);
break;
case TX_8X8:
vpx_highbd_fdct8x8(src_diff, coeff, diff_stride);
- vpx_quantize_b(coeff, 64, p->zbin, p->round, p->quant, p->quant_shift,
- qcoeff, dqcoeff, pd->dequant, eob, scan_order->scan,
- scan_order->iscan);
+ vpx_highbd_quantize_b(coeff, 64, p->zbin, p->round, p->quant,
+ p->quant_shift, qcoeff, dqcoeff, pd->dequant, eob,
+ scan_order->scan, scan_order->iscan);
break;
default:
assert(tx_size == TX_4X4);
x->fwd_txfm4x4(src_diff, coeff, diff_stride);
- vpx_quantize_b(coeff, 16, p->zbin, p->round, p->quant, p->quant_shift,
- qcoeff, dqcoeff, pd->dequant, eob, scan_order->scan,
- scan_order->iscan);
+ vpx_highbd_quantize_b(coeff, 16, p->zbin, p->round, p->quant,
+ p->quant_shift, qcoeff, dqcoeff, pd->dequant, eob,
+ scan_order->scan, scan_order->iscan);
break;
}
return;
vpx_highbd_subtract_block(32, 32, src_diff, diff_stride, src,
src_stride, dst, dst_stride, xd->bd);
highbd_fdct32x32(x->use_lp32x32fdct, src_diff, coeff, diff_stride);
- vpx_quantize_b_32x32(coeff, 1024, p->zbin, p->round, p->quant,
- p->quant_shift, qcoeff, dqcoeff, pd->dequant,
- eob, scan_order->scan, scan_order->iscan);
+ vpx_highbd_quantize_b_32x32(
+ coeff, 1024, p->zbin, p->round, p->quant, p->quant_shift, qcoeff,
+ dqcoeff, pd->dequant, eob, scan_order->scan, scan_order->iscan);
}
if (args->enable_coeff_opt && !x->skip_recode) {
*a = *l = vp9_optimize_b(x, plane, block, tx_size, entropy_ctx) > 0;
vpx_highbd_fdct16x16(src_diff, coeff, diff_stride);
else
vp9_highbd_fht16x16(src_diff, coeff, diff_stride, tx_type);
- vpx_quantize_b(coeff, 256, p->zbin, p->round, p->quant,
- p->quant_shift, qcoeff, dqcoeff, pd->dequant, eob,
- scan_order->scan, scan_order->iscan);
+ vpx_highbd_quantize_b(coeff, 256, p->zbin, p->round, p->quant,
+ p->quant_shift, qcoeff, dqcoeff, pd->dequant,
+ eob, scan_order->scan, scan_order->iscan);
}
if (args->enable_coeff_opt && !x->skip_recode) {
*a = *l = vp9_optimize_b(x, plane, block, tx_size, entropy_ctx) > 0;
vpx_highbd_fdct8x8(src_diff, coeff, diff_stride);
else
vp9_highbd_fht8x8(src_diff, coeff, diff_stride, tx_type);
- vpx_quantize_b(coeff, 64, p->zbin, p->round, p->quant, p->quant_shift,
- qcoeff, dqcoeff, pd->dequant, eob, scan_order->scan,
- scan_order->iscan);
+ vpx_highbd_quantize_b(coeff, 64, p->zbin, p->round, p->quant,
+ p->quant_shift, qcoeff, dqcoeff, pd->dequant,
+ eob, scan_order->scan, scan_order->iscan);
}
if (args->enable_coeff_opt && !x->skip_recode) {
*a = *l = vp9_optimize_b(x, plane, block, tx_size, entropy_ctx) > 0;
vp9_highbd_fht4x4(src_diff, coeff, diff_stride, tx_type);
else
x->fwd_txfm4x4(src_diff, coeff, diff_stride);
- vpx_quantize_b(coeff, 16, p->zbin, p->round, p->quant, p->quant_shift,
- qcoeff, dqcoeff, pd->dequant, eob, scan_order->scan,
- scan_order->iscan);
+ vpx_highbd_quantize_b(coeff, 16, p->zbin, p->round, p->quant,
+ p->quant_shift, qcoeff, dqcoeff, pd->dequant,
+ eob, scan_order->scan, scan_order->iscan);
}
if (args->enable_coeff_opt && !x->skip_recode) {
*a = *l = vp9_optimize_b(x, plane, block, tx_size, entropy_ctx) > 0;
return;
}
+#if CONFIG_VP9_HIGHBITDEPTH
+ if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+ vpx_highbd_quantize_b(BLOCK_OFFSET(p->coeff, block), n_coeffs, p->zbin,
+ p->round, p->quant, p->quant_shift, qcoeff, dqcoeff,
+ pd->dequant, &p->eobs[block], scan, iscan);
+ return;
+ }
+#endif
vpx_quantize_b(BLOCK_OFFSET(p->coeff, block), n_coeffs, p->zbin, p->round,
p->quant, p->quant_shift, qcoeff, dqcoeff, pd->dequant,
&p->eobs[block], scan, iscan);
tran_low_t *qcoeff_ptr,
tran_low_t *dqcoeff_ptr,
const int16_t dequant, uint16_t *eob_ptr);
-
-// Only used for reference. The optimized versions can handle HBD.
-void vpx_highbd_quantize_b_c(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);
-
-void vpx_highbd_quantize_b_32x32_c(
- 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);
#endif
#ifdef __cplusplus
DSP_SRCS-$(HAVE_AVX) += x86/quantize_avx.c
DSP_SRCS-$(HAVE_NEON) += arm/quantize_neon.c
DSP_SRCS-$(HAVE_VSX) += ppc/quantize_vsx.c
+ifeq ($(CONFIG_VP9_HIGHBITDEPTH),yes)
+DSP_SRCS-$(HAVE_SSE2) += x86/highbd_quantize_intrin_sse2.c
+endif
# avg
DSP_SRCS-yes += avg.c
add_proto qw/void vpx_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/vpx_quantize_b_32x32 neon ssse3 avx vsx/;
+
+ 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, 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_highbd_quantize_b sse2/;
+
+ add_proto qw/void vpx_highbd_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/vpx_highbd_quantize_b_32x32 sse2/;
+ } # CONFIG_VP9_HIGHBITDEPTH
} # CONFIG_VP9_ENCODER
if (vpx_config("CONFIG_ENCODERS") eq "yes") {
--- /dev/null
+/*
+ * 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 <assert.h>
+#include <emmintrin.h>
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_dsp/vpx_dsp_common.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vpx_ports/mem.h"
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void vpx_highbd_quantize_b_sse2(const tran_low_t *coeff_ptr, intptr_t count,
+ 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) {
+ int i, j, non_zero_regs = (int)count / 4, eob_i = -1;
+ __m128i zbins[2];
+ __m128i nzbins[2];
+
+ zbins[0] = _mm_set_epi32((int)zbin_ptr[1], (int)zbin_ptr[1], (int)zbin_ptr[1],
+ (int)zbin_ptr[0]);
+ zbins[1] = _mm_set1_epi32((int)zbin_ptr[1]);
+
+ nzbins[0] = _mm_setzero_si128();
+ nzbins[1] = _mm_setzero_si128();
+ nzbins[0] = _mm_sub_epi32(nzbins[0], zbins[0]);
+ nzbins[1] = _mm_sub_epi32(nzbins[1], zbins[1]);
+
+ (void)scan;
+
+ memset(qcoeff_ptr, 0, count * sizeof(*qcoeff_ptr));
+ memset(dqcoeff_ptr, 0, count * sizeof(*dqcoeff_ptr));
+
+ // Pre-scan pass
+ for (i = ((int)count / 4) - 1; i >= 0; i--) {
+ __m128i coeffs, cmp1, cmp2;
+ int test;
+ coeffs = _mm_load_si128((const __m128i *)(coeff_ptr + i * 4));
+ cmp1 = _mm_cmplt_epi32(coeffs, zbins[i != 0]);
+ cmp2 = _mm_cmpgt_epi32(coeffs, nzbins[i != 0]);
+ cmp1 = _mm_and_si128(cmp1, cmp2);
+ test = _mm_movemask_epi8(cmp1);
+ if (test == 0xffff)
+ non_zero_regs--;
+ else
+ break;
+ }
+
+ // Quantization pass:
+ for (i = 0; i < non_zero_regs; i++) {
+ __m128i coeffs, coeffs_sign, tmp1, tmp2;
+ int test;
+ int abs_coeff[4];
+ int coeff_sign[4];
+
+ coeffs = _mm_load_si128((const __m128i *)(coeff_ptr + i * 4));
+ coeffs_sign = _mm_srai_epi32(coeffs, 31);
+ coeffs = _mm_sub_epi32(_mm_xor_si128(coeffs, coeffs_sign), coeffs_sign);
+ tmp1 = _mm_cmpgt_epi32(coeffs, zbins[i != 0]);
+ tmp2 = _mm_cmpeq_epi32(coeffs, zbins[i != 0]);
+ tmp1 = _mm_or_si128(tmp1, tmp2);
+ test = _mm_movemask_epi8(tmp1);
+ _mm_storeu_si128((__m128i *)abs_coeff, coeffs);
+ _mm_storeu_si128((__m128i *)coeff_sign, coeffs_sign);
+
+ for (j = 0; j < 4; j++) {
+ if (test & (1 << (4 * j))) {
+ int k = 4 * i + j;
+ const int64_t tmp3 = abs_coeff[j] + round_ptr[k != 0];
+ const int64_t tmp4 = ((tmp3 * quant_ptr[k != 0]) >> 16) + tmp3;
+ const uint32_t abs_qcoeff =
+ (uint32_t)((tmp4 * quant_shift_ptr[k != 0]) >> 16);
+ qcoeff_ptr[k] = (int)(abs_qcoeff ^ coeff_sign[j]) - coeff_sign[j];
+ dqcoeff_ptr[k] = qcoeff_ptr[k] * dequant_ptr[k != 0];
+ if (abs_qcoeff) eob_i = iscan[k] > eob_i ? iscan[k] : eob_i;
+ }
+ }
+ }
+ *eob_ptr = eob_i + 1;
+}
+
+void vpx_highbd_quantize_b_32x32_sse2(
+ 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) {
+ __m128i zbins[2];
+ __m128i nzbins[2];
+ int idx = 0;
+ int idx_arr[1024];
+ int i, eob = -1;
+ const int zbin0_tmp = ROUND_POWER_OF_TWO(zbin_ptr[0], 1);
+ const int zbin1_tmp = ROUND_POWER_OF_TWO(zbin_ptr[1], 1);
+ (void)scan;
+
+ zbins[0] = _mm_set_epi32(zbin1_tmp, zbin1_tmp, zbin1_tmp, zbin0_tmp);
+ zbins[1] = _mm_set1_epi32(zbin1_tmp);
+
+ nzbins[0] = _mm_setzero_si128();
+ nzbins[1] = _mm_setzero_si128();
+ nzbins[0] = _mm_sub_epi32(nzbins[0], zbins[0]);
+ nzbins[1] = _mm_sub_epi32(nzbins[1], zbins[1]);
+
+ memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+ memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+
+ // Pre-scan pass
+ for (i = 0; i < n_coeffs / 4; i++) {
+ __m128i coeffs, cmp1, cmp2;
+ int test;
+ coeffs = _mm_load_si128((const __m128i *)(coeff_ptr + i * 4));
+ cmp1 = _mm_cmplt_epi32(coeffs, zbins[i != 0]);
+ cmp2 = _mm_cmpgt_epi32(coeffs, nzbins[i != 0]);
+ cmp1 = _mm_and_si128(cmp1, cmp2);
+ test = _mm_movemask_epi8(cmp1);
+ if (!(test & 0xf)) idx_arr[idx++] = i * 4;
+ if (!(test & 0xf0)) idx_arr[idx++] = i * 4 + 1;
+ if (!(test & 0xf00)) idx_arr[idx++] = i * 4 + 2;
+ if (!(test & 0xf000)) idx_arr[idx++] = i * 4 + 3;
+ }
+
+ // Quantization pass: only process the coefficients selected in
+ // pre-scan pass. Note: idx can be zero.
+ for (i = 0; i < idx; i++) {
+ const int rc = idx_arr[i];
+ const int coeff = coeff_ptr[rc];
+ const int coeff_sign = (coeff >> 31);
+ const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+ const int64_t tmp1 = abs_coeff + ROUND_POWER_OF_TWO(round_ptr[rc != 0], 1);
+ const int64_t tmp2 = ((tmp1 * quant_ptr[rc != 0]) >> 16) + tmp1;
+ const uint32_t abs_qcoeff =
+ (uint32_t)((tmp2 * quant_shift_ptr[rc != 0]) >> 15);
+ qcoeff_ptr[rc] = (int)(abs_qcoeff ^ coeff_sign) - coeff_sign;
+ dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0] / 2;
+ if (abs_qcoeff) eob = iscan[idx_arr[i]] > eob ? iscan[idx_arr[i]] : eob;
+ }
+ *eob_ptr = eob + 1;
+}
+#endif
projects / platform / upstream / libvpx.git / commitdiff