quantize: copy ssse3 optimizations to intrinsics

Fairly minor differences from sse2. pabsw and psignw are the big gains.
Also re-uses some values in eob calculation to avoid an extra pcmp.

Fixes test failures in HBD and OS X builds.

Allows using it in 32bit builds, where it is about 40% faster than sse2.

Substantially faster than the assembly for skip_block. 10-20% faster the
rest of the time.

Change-Id: If783bb3567e561e47667e10133b9c84414a334e2

diff --git a/test/vp9_quantize_test.cc b/test/vp9_quantize_test.cc
index 7ec8f585b03aaa636030b2fcd29e48fd296952ff..c4aae2d9d7c73f9e9cd27bdf865c76739d6e80a2 100644 (file)
--- a/test/vp9_quantize_test.cc
+++ b/test/vp9_quantize_test.cc
@@ -333,15 +333,20 @@ INSTANTIATE_TEST_CASE_P(SSE2, VP9QuantizeTest,
 #endif  // CONFIG_VP9_HIGHBITDEPTH
 #endif  // HAVE_SSE2
 
-// TODO(johannkoenig): SSSE3 optimizations do not yet pass these tests.
-#if HAVE_SSSE3 && ARCH_X86_64
+#if HAVE_SSSE3
+INSTANTIATE_TEST_CASE_P(SSSE3, VP9QuantizeTest,
+                        ::testing::Values(make_tuple(&vpx_quantize_b_ssse3,
+                                                     &vpx_quantize_b_c,
+                                                     VPX_BITS_8, 16)));
+
+#if ARCH_X86_64
+// TODO(johannkoenig): SSSE3 optimizations do not yet pass this test.
 INSTANTIATE_TEST_CASE_P(
     DISABLED_SSSE3, VP9QuantizeTest,
-    ::testing::Values(make_tuple(&vpx_quantize_b_ssse3, &vpx_quantize_b_c,
-                                 VPX_BITS_8, 16),
-                      make_tuple(&vpx_quantize_b_32x32_ssse3,
+    ::testing::Values(make_tuple(&vpx_quantize_b_32x32_ssse3,
                                  &vpx_quantize_b_32x32_c, VPX_BITS_8, 32)));
-#endif  // HAVE_SSSE3 && ARCH_X86_64
+#endif  // ARCH_X86_64
+#endif  // HAVE_SSSE3
 
 // TODO(johannkoenig): AVX optimizations do not yet pass the 32x32 test or
 // highbitdepth configurations.

diff --git a/vpx_dsp/vpx_dsp.mk b/vpx_dsp/vpx_dsp.mk
index ae98eb23dde996478f2eb4761d2f5d5b1feb12d9..98aa2173bd3ac2763c0d3fb02619881471a03382 100644 (file)
--- a/vpx_dsp/vpx_dsp.mk
+++ b/vpx_dsp/vpx_dsp.mk
@@ -274,6 +274,7 @@ DSP_SRCS-yes            += quantize.c
 DSP_SRCS-yes            += quantize.h
 
 DSP_SRCS-$(HAVE_SSE2)   += x86/quantize_sse2.c
+DSP_SRCS-$(HAVE_SSSE3)  += x86/quantize_ssse3.c
 DSP_SRCS-$(HAVE_NEON)   += arm/quantize_neon.c
 ifeq ($(CONFIG_VP9_HIGHBITDEPTH),yes)
 DSP_SRCS-$(HAVE_SSE2)   += x86/highbd_quantize_intrin_sse2.c

diff --git a/vpx_dsp/vpx_dsp_rtcd_defs.pl b/vpx_dsp/vpx_dsp_rtcd_defs.pl
index 99ef262b12ccc2a82a69e08ad66f2c9e769495e8..a78b1cff7d6cfc2d8cda86f7e2850ed57da26e72 100644 (file)
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -671,7 +671,7 @@ if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
 #
 if (vpx_config("CONFIG_VP9_ENCODER") eq "yes") {
   add_proto qw/void vpx_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";
-  specialize qw/vpx_quantize_b neon sse2/, "$ssse3_x86_64", "$avx_x86_64";
+  specialize qw/vpx_quantize_b neon sse2 ssse3/, "$avx_x86_64";
 
   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/, "$ssse3_x86_64", "$avx_x86_64";

diff --git a/vpx_dsp/x86/quantize_ssse3.c b/vpx_dsp/x86/quantize_ssse3.c
new file mode 100644 (file)
index 0000000..014acb4
--- /dev/null
+++ b/vpx_dsp/x86/quantize_ssse3.c
@@ -0,0 +1,198 @@
+/*
+ *  Copyright (c) 2017 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 <tmmintrin.h>
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx/vpx_integer.h"
+#include "vpx_dsp/x86/bitdepth_conversion_sse2.h"
+
+void vpx_quantize_b_ssse3(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();
+  __m128i coeff0, coeff1;
+  __m128i eob;
+  __m128i zbin;
+  __m128i round, quant, dequant, shift;
+  intptr_t index = 0;
+  (void)scan_ptr;
+
+  if (skip_block) {
+    do {
+      store_tran_low(zero, dqcoeff_ptr + index);
+      store_tran_low(zero, dqcoeff_ptr + index + 8);
+      store_tran_low(zero, qcoeff_ptr + index);
+      store_tran_low(zero, qcoeff_ptr + index + 8);
+      index += 16;
+    } while (index < n_coeffs);
+    *eob_ptr = 0;
+    return;
+  }
+
+  // Setup global values
+  {
+    const __m128i one = _mm_set1_epi16(1);
+    zbin = _mm_load_si128((const __m128i *)zbin_ptr);
+    // 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);
+    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);
+  }
+
+  {
+    __m128i qcoeff0, qcoeff1;
+    __m128i qtmp0, qtmp1;
+    __m128i cmp_mask0, cmp_mask1;
+    __m128i zero_coeff0, zero_coeff1;
+    __m128i iscan0, iscan1;
+    __m128i eob1;
+
+    // Do DC and first 15 AC
+    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);
+    // Overwrite DC component.
+    zbin = _mm_unpackhi_epi64(zbin, zbin);
+    cmp_mask1 = _mm_cmpgt_epi16(qcoeff1, zbin);
+
+    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 + index);
+    store_tran_low(qcoeff1, qcoeff_ptr + index + 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 + index);
+    store_tran_low(coeff1, dqcoeff_ptr + index + 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 + index));
+    iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + index + 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);
+  }
+  index += 16;
+
+  // AC only loop
+  while (index < n_coeffs) {
+    __m128i qcoeff0, qcoeff1;
+    __m128i qtmp0, qtmp1;
+    __m128i cmp_mask0, cmp_mask1;
+    __m128i zero_coeff0, zero_coeff1;
+    __m128i iscan0, iscan1;
+    __m128i eob0, eob1;
+
+    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);
+
+    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);
+
+    // 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 + index);
+    store_tran_low(qcoeff1, qcoeff_ptr + index + 8);
+
+    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);
+
+    // 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 + index));
+    iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + index + 8));
+    // Add one to convert from indices to counts
+    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);
+
+    index += 16;
+  }
+
+  // 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);
+  }
+}

diff --git a/vpx_dsp/x86/quantize_ssse3_x86_64.asm b/vpx_dsp/x86/quantize_ssse3_x86_64.asm
index ec2cafb94cd6ffd2e3e9cfe2b50590a18d8cc0b2..19798422b4aee491c058101c6c4bb3a302deeee3 100644 (file)
--- a/vpx_dsp/x86/quantize_ssse3_x86_64.asm
+++ b/vpx_dsp/x86/quantize_ssse3_x86_64.asm
@@ -15,7 +15,6 @@ pw_1: times 8 dw 1
 
 SECTION .text
 
-; TODO(yunqingwang)fix quantize_b code for skip=1 case.
 %macro QUANTIZE_FN 2
 cglobal quantize_%1, 0, %2, 15, coeff, ncoeff, skip, zbin, round, quant, \
                                 shift, qcoeff, dqcoeff, dequant, \
@@ -304,6 +303,9 @@ cglobal quantize_%1, 0, %2, 15, coeff, ncoeff, skip, zbin, round, quant, \
 
   ; skip-block, i.e. just write all zeroes
 .blank:
+DEFINE_ARGS coeff, ncoeff, skip, zbin, round, quant, shift, \
+            qcoeff, dqcoeff, dequant, eob, scan, iscan
+
   mov                             r0, dqcoeffmp
   movifnidn                  ncoeffq, ncoeffmp
   mov                             r2, qcoeffmp
@@ -341,5 +343,4 @@ cglobal quantize_%1, 0, %2, 15, coeff, ncoeff, skip, zbin, round, quant, \
 %endmacro
 
 INIT_XMM ssse3
-QUANTIZE_FN b, 7
 QUANTIZE_FN b_32x32, 7