summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
Diffstat
-rw-r--r--test/convolve_test.cc52
-rw-r--r--test/vpx_scale_test.h28
-rw-r--r--vp8/common/mips/mmi/idct_blk_mmi.c71
-rw-r--r--vp8/common/rtcd_defs.pl4
-rw-r--r--vp8/vp8_common.mk1
-rw-r--r--vp9/encoder/vp9_frame_scale.c2
-rw-r--r--vp9/encoder/x86/vp9_frame_scale_ssse3.c226
-rw-r--r--vpx_dsp/vpx_dsp.mk1
-rw-r--r--vpx_dsp/x86/convolve_avx2.h99
-rw-r--r--vpx_dsp/x86/convolve_ssse3.h61
-rw-r--r--vpx_dsp/x86/mem_sse2.h8
-rw-r--r--vpx_dsp/x86/vpx_subpixel_8t_intrin_avx2.c557
12 files changed, 678 insertions, 432 deletions
diff --git a/test/convolve_test.cc b/test/convolve_test.cc
index 9f6f795c9..08ef57224 100644
--- a/test/convolve_test.cc
+++ b/test/convolve_test.cc
@@ -925,33 +925,51 @@ TEST_P(ConvolveTest, FilterExtremes) {
/* This test exercises that enough rows and columns are filtered with every
possible initial fractional positions and scaling steps. */
+#if !CONFIG_VP9_HIGHBITDEPTH
+static const ConvolveFunc scaled_2d_c_funcs[2] = { vpx_scaled_2d_c,
+ vpx_scaled_avg_2d_c };
+
TEST_P(ConvolveTest, CheckScalingFiltering) {
uint8_t *const in = input();
uint8_t *const out = output();
- const InterpKernel *const eighttap = vp9_filter_kernels[EIGHTTAP];
+ uint8_t ref[kOutputStride * kMaxDimension];
- SetConstantInput(127);
+ ::libvpx_test::ACMRandom prng;
+ for (int y = 0; y < Height(); ++y) {
+ for (int x = 0; x < Width(); ++x) {
+ const uint16_t r = prng.Rand8Extremes();
+ assign_val(in, y * kInputStride + x, r);
+ }
+ }
- for (int frac = 0; frac < 16; ++frac) {
- for (int step = 1; step <= 32; ++step) {
- /* Test the horizontal and vertical filters in combination. */
- ASM_REGISTER_STATE_CHECK(
- UUT_->shv8_[0](in, kInputStride, out, kOutputStride, eighttap, frac,
- step, frac, step, Width(), Height()));
-
- CheckGuardBlocks();
-
- for (int y = 0; y < Height(); ++y) {
- for (int x = 0; x < Width(); ++x) {
- ASSERT_EQ(lookup(in, y * kInputStride + x),
- lookup(out, y * kOutputStride + x))
- << "x == " << x << ", y == " << y << ", frac == " << frac
- << ", step == " << step;
+ for (int i = 0; i < 2; ++i) {
+ for (INTERP_FILTER filter_type = 0; filter_type < 4; ++filter_type) {
+ const InterpKernel *const eighttap = vp9_filter_kernels[filter_type];
+ for (int frac = 0; frac < 16; ++frac) {
+ for (int step = 1; step <= 32; ++step) {
+ /* Test the horizontal and vertical filters in combination. */
+ scaled_2d_c_funcs[i](in, kInputStride, ref, kOutputStride, eighttap,
+ frac, step, frac, step, Width(), Height());
+ ASM_REGISTER_STATE_CHECK(
+ UUT_->shv8_[i](in, kInputStride, out, kOutputStride, eighttap,
+ frac, step, frac, step, Width(), Height()));
+
+ CheckGuardBlocks();
+
+ for (int y = 0; y < Height(); ++y) {
+ for (int x = 0; x < Width(); ++x) {
+ ASSERT_EQ(lookup(ref, y * kOutputStride + x),
+ lookup(out, y * kOutputStride + x))
+ << "x == " << x << ", y == " << y << ", frac == " << frac
+ << ", step == " << step;
+ }
+ }
}
}
}
}
}
+#endif
using std::tr1::make_tuple;
diff --git a/test/vpx_scale_test.h b/test/vpx_scale_test.h
index 18909d1b5..dcbd02b91 100644
--- a/test/vpx_scale_test.h
+++ b/test/vpx_scale_test.h
@@ -15,11 +15,14 @@
#include "./vpx_config.h"
#include "./vpx_scale_rtcd.h"
+#include "test/acm_random.h"
#include "test/clear_system_state.h"
#include "test/register_state_check.h"
#include "vpx_mem/vpx_mem.h"
#include "vpx_scale/yv12config.h"
+using libvpx_test::ACMRandom;
+
namespace libvpx_test {
class VpxScaleBase {
@@ -65,12 +68,12 @@ class VpxScaleBase {
ResetScaleImage(&img_, src_width, src_height);
ResetScaleImage(&ref_img_, dst_width, dst_height);
ResetScaleImage(&dst_img_, dst_width, dst_height);
- FillPlane(img_.y_buffer, img_.y_crop_width, img_.y_crop_height,
- img_.y_stride);
- FillPlane(img_.u_buffer, img_.uv_crop_width, img_.uv_crop_height,
- img_.uv_stride);
- FillPlane(img_.v_buffer, img_.uv_crop_width, img_.uv_crop_height,
- img_.uv_stride);
+ FillPlaneExtreme(img_.y_buffer, img_.y_crop_width, img_.y_crop_height,
+ img_.y_stride);
+ FillPlaneExtreme(img_.u_buffer, img_.uv_crop_width, img_.uv_crop_height,
+ img_.uv_stride);
+ FillPlaneExtreme(img_.v_buffer, img_.uv_crop_width, img_.uv_crop_height,
+ img_.uv_stride);
}
void DeallocImages() {
@@ -89,7 +92,8 @@ class VpxScaleBase {
static const int kBufFiller = 123;
static const int kBufMax = kBufFiller - 1;
- static void FillPlane(uint8_t *buf, int width, int height, int stride) {
+ static void FillPlane(uint8_t *const buf, const int width, const int height,
+ const int stride) {
for (int y = 0; y < height; ++y) {
for (int x = 0; x < width; ++x) {
buf[x + (y * stride)] = (x + (width * y)) % kBufMax;
@@ -97,6 +101,16 @@ class VpxScaleBase {
}
}
+ static void FillPlaneExtreme(uint8_t *const buf, const int width,
+ const int height, const int stride) {
+ ACMRandom rnd;
+ for (int y = 0; y < height; ++y) {
+ for (int x = 0; x < width; ++x) {
+ buf[x + (y * stride)] = rnd.Rand8() % 2 ? 255 : 0;
+ }
+ }
+ }
+
static void ExtendPlane(uint8_t *buf, int crop_width, int crop_height,
int width, int height, int stride, int padding) {
// Copy the outermost visible pixel to a distance of at least 'padding.'
diff --git a/vp8/common/mips/mmi/idct_blk_mmi.c b/vp8/common/mips/mmi/idct_blk_mmi.c
new file mode 100644
index 000000000..f6020ab46
--- /dev/null
+++ b/vp8/common/mips/mmi/idct_blk_mmi.c
@@ -0,0 +1,71 @@
+/*
+ * 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 "./vp8_rtcd.h"
+#include "vpx_mem/vpx_mem.h"
+
+void vp8_dequant_idct_add_y_block_mmi(int16_t *q, int16_t *dq, uint8_t *dst,
+ int stride, int8_t *eobs) {
+ int i, j;
+
+ for (i = 0; i < 4; i++) {
+ for (j = 0; j < 4; j++) {
+ if (*eobs++ > 1) {
+ vp8_dequant_idct_add_mmi(q, dq, dst, stride);
+ } else {
+ vp8_dc_only_idct_add_mmi(q[0] * dq[0], dst, stride, dst, stride);
+ memset(q, 0, 2 * sizeof(q[0]));
+ }
+
+ q += 16;
+ dst += 4;
+ }
+
+ dst += 4 * stride - 16;
+ }
+}
+
+void vp8_dequant_idct_add_uv_block_mmi(int16_t *q, int16_t *dq, uint8_t *dstu,
+ uint8_t *dstv, int stride,
+ int8_t *eobs) {
+ int i, j;
+
+ for (i = 0; i < 2; i++) {
+ for (j = 0; j < 2; j++) {
+ if (*eobs++ > 1) {
+ vp8_dequant_idct_add_mmi(q, dq, dstu, stride);
+ } else {
+ vp8_dc_only_idct_add_mmi(q[0] * dq[0], dstu, stride, dstu, stride);
+ memset(q, 0, 2 * sizeof(q[0]));
+ }
+
+ q += 16;
+ dstu += 4;
+ }
+
+ dstu += 4 * stride - 8;
+ }
+
+ for (i = 0; i < 2; i++) {
+ for (j = 0; j < 2; j++) {
+ if (*eobs++ > 1) {
+ vp8_dequant_idct_add_mmi(q, dq, dstv, stride);
+ } else {
+ vp8_dc_only_idct_add_mmi(q[0] * dq[0], dstv, stride, dstv, stride);
+ memset(q, 0, 2 * sizeof(q[0]));
+ }
+
+ q += 16;
+ dstv += 4;
+ }
+
+ dstv += 4 * stride - 8;
+ }
+}
diff --git a/vp8/common/rtcd_defs.pl b/vp8/common/rtcd_defs.pl
index be2ac0054..ece2785eb 100644
--- a/vp8/common/rtcd_defs.pl
+++ b/vp8/common/rtcd_defs.pl
@@ -28,10 +28,10 @@ add_proto qw/void vp8_dequant_idct_add/, "short *input, short *dq, unsigned char
specialize qw/vp8_dequant_idct_add mmx neon dspr2 msa mmi/;
add_proto qw/void vp8_dequant_idct_add_y_block/, "short *q, short *dq, unsigned char *dst, int stride, char *eobs";
-specialize qw/vp8_dequant_idct_add_y_block sse2 neon dspr2 msa/;
+specialize qw/vp8_dequant_idct_add_y_block sse2 neon dspr2 msa mmi/;
add_proto qw/void vp8_dequant_idct_add_uv_block/, "short *q, short *dq, unsigned char *dst_u, unsigned char *dst_v, int stride, char *eobs";
-specialize qw/vp8_dequant_idct_add_uv_block sse2 neon dspr2 msa/;
+specialize qw/vp8_dequant_idct_add_uv_block sse2 neon dspr2 msa mmi/;
#
# Loopfilter
diff --git a/vp8/vp8_common.mk b/vp8/vp8_common.mk
index 5813c81c4..246fe6a67 100644
--- a/vp8/vp8_common.mk
+++ b/vp8/vp8_common.mk
@@ -122,6 +122,7 @@ VP8_COMMON_SRCS-$(HAVE_MMI) += common/mips/mmi/loopfilter_filters_mmi.c
VP8_COMMON_SRCS-$(HAVE_MMI) += common/mips/mmi/idctllm_mmi.c
VP8_COMMON_SRCS-$(HAVE_MMI) += common/mips/mmi/dequantize_mmi.c
VP8_COMMON_SRCS-$(HAVE_MMI) += common/mips/mmi/copymem_mmi.c
+VP8_COMMON_SRCS-$(HAVE_MMI) += common/mips/mmi/idct_blk_mmi.c
ifeq ($(CONFIG_POSTPROC),yes)
VP8_COMMON_SRCS-$(HAVE_MSA) += common/mips/msa/mfqe_msa.c
diff --git a/vp9/encoder/vp9_frame_scale.c b/vp9/encoder/vp9_frame_scale.c
index 832df18c8..a410d0407 100644
--- a/vp9/encoder/vp9_frame_scale.c
+++ b/vp9/encoder/vp9_frame_scale.c
@@ -28,7 +28,7 @@ void vp9_scale_and_extend_frame_c(const YV12_BUFFER_CONFIG *src,
const InterpKernel *const kernel = vp9_filter_kernels[filter_type];
int x, y, i;
-#if HAVE_NEON
+#if HAVE_SSSE3 || HAVE_NEON
// TODO(linfengz): The 4:3 specialized C code is disabled by default since
// it's much slower than the general version which calls vpx_scaled_2d() even
// if vpx_scaled_2d() is not optimized. It will only be enabled as a reference
diff --git a/vp9/encoder/x86/vp9_frame_scale_ssse3.c b/vp9/encoder/x86/vp9_frame_scale_ssse3.c
index 81e5b4229..7685e7bc3 100644
--- a/vp9/encoder/x86/vp9_frame_scale_ssse3.c
+++ b/vp9/encoder/x86/vp9_frame_scale_ssse3.c
@@ -438,6 +438,202 @@ static void scale_plane_4_to_1_general(const uint8_t *src, const int src_stride,
} while (x);
}
+typedef void (*shuffle_filter_funcs)(const int16_t *const filter,
+ __m128i *const f);
+
+typedef __m128i (*convolve8_funcs)(const __m128i *const s,
+ const __m128i *const f);
+
+static void scale_plane_4_to_3_general(const uint8_t *src, const int src_stride,
+ uint8_t *dst, const int dst_stride,
+ const int w, const int h,
+ const InterpKernel *const coef,
+ const int phase_scaler,
+ uint8_t *const temp_buffer) {
+ static const int step_q4 = 16 * 4 / 3;
+ const int width_hor = (w + 5) - ((w + 5) % 6);
+ const int stride_hor = 2 * width_hor + 4; // store 4 extra pixels
+ const int width_ver = (w + 7) & ~7;
+ // We need (SUBPEL_TAPS - 1) extra rows: (SUBPEL_TAPS / 2 - 1) extra rows
+ // above and (SUBPEL_TAPS / 2) extra rows below.
+ const int height_hor = (4 * h / 3 + SUBPEL_TAPS - 1 + 7) & ~7;
+ const int height_ver = (h + 5) - ((h + 5) % 6);
+ int x, y = height_hor;
+ uint8_t *t = temp_buffer;
+ __m128i s[12], d[6], dd[4];
+ __m128i f0[4], f1[5], f2[5];
+ // The offset of the first row is always less than 1 pixel.
+ const int offset1_q4 = phase_scaler + 1 * step_q4;
+ const int offset2_q4 = phase_scaler + 2 * step_q4;
+ // offset_idxx indicates the pixel offset is even (0) or odd (1).
+ // It's used to choose the src offset and filter coefficient offset.
+ const int offset_idx1 = (offset1_q4 >> 4) & 1;
+ const int offset_idx2 = (offset2_q4 >> 4) & 1;
+ static const shuffle_filter_funcs shuffle_filter_funcs[2] = {
+ shuffle_filter_ssse3, shuffle_filter_odd_ssse3
+ };
+ static const convolve8_funcs convolve8_funcs[2] = {
+ convolve8_8_even_offset_ssse3, convolve8_8_odd_offset_ssse3
+ };
+
+ assert(w && h);
+
+ shuffle_filter_ssse3(coef[(phase_scaler + 0 * step_q4) & SUBPEL_MASK], f0);
+ shuffle_filter_funcs[offset_idx1](coef[offset1_q4 & SUBPEL_MASK], f1);
+ shuffle_filter_funcs[offset_idx2](coef[offset2_q4 & SUBPEL_MASK], f2);
+
+ // Sub 64 to avoid overflow.
+ // Coef 128 would be treated as -128 in PMADDUBSW. Sub 64 here.
+ // Coef 128 is in either fx[1] or fx[2] depending on the phase idx.
+ // When filter phase idx is 1, the two biggest coefficients are shuffled
+ // together, and the sum of them are always no less than 128. Sub 64 here.
+ // After the subtraction, when the sum of all positive coefficients are no
+ // larger than 128, and the sum of all negative coefficients are no
+ // less than -128, there will be no overflow in the convolve8 functions.
+ f0[1] = _mm_sub_epi8(f0[1], _mm_set1_epi8(64));
+ f1[1 + offset_idx1] = _mm_sub_epi8(f1[1 + offset_idx1], _mm_set1_epi8(64));
+ f2[1 + offset_idx2] = _mm_sub_epi8(f2[1 + offset_idx2], _mm_set1_epi8(64));
+
+ src -= (SUBPEL_TAPS / 2 - 1) * src_stride + SUBPEL_TAPS / 2 - 1;
+
+ // horizontal 6x8
+ do {
+ load_8bit_8x8(src, src_stride, s);
+ // 00 01 10 11 20 21 30 31 40 41 50 51 60 61 70 71
+ // 02 03 12 13 22 23 32 33 42 43 52 53 62 63 72 73
+ // 04 05 14 15 24 25 34 35 44 45 54 55 64 65 74 75
+ // 06 07 16 17 26 27 36 37 46 47 56 57 66 67 76 77
+ transpose_16bit_4x8(s, s);
+ x = width_hor;
+
+ do {
+ src += 8;
+ load_8bit_8x8(src, src_stride, &s[4]);
+ // 08 09 18 19 28 29 38 39 48 49 58 59 68 69 78 79
+ // 0A 0B 1A 1B 2A 2B 3A 3B 4A 4B 5A 5B 6A 6B 7A 7B
+ // OC 0D 1C 1D 2C 2D 3C 3D 4C 4D 5C 5D 6C 6D 7C 7D
+ // 0E 0F 1E 1F 2E 2F 3E 3F 4E 4F 5E 5F 6E 6F 7E 7F
+ transpose_16bit_4x8(&s[4], &s[4]);
+
+ // 00 10 20 30 40 50 60 70
+ // 01 11 21 31 41 51 61 71
+ // 02 12 22 32 42 52 62 72
+ // 03 13 23 33 43 53 63 73
+ // 04 14 24 34 44 54 64 74
+ // 05 15 25 35 45 55 65 75
+ d[0] = convolve8_8_even_offset_ssse3(&s[0], f0);
+ d[1] = convolve8_funcs[offset_idx1](&s[offset1_q4 >> 5], f1);
+ d[2] = convolve8_funcs[offset_idx2](&s[offset2_q4 >> 5], f2);
+ d[3] = convolve8_8_even_offset_ssse3(&s[2], f0);
+ d[4] = convolve8_funcs[offset_idx1](&s[2 + (offset1_q4 >> 5)], f1);
+ d[5] = convolve8_funcs[offset_idx2](&s[2 + (offset2_q4 >> 5)], f2);
+
+ // 00 10 20 30 40 50 60 70 02 12 22 32 42 52 62 72
+ // 01 11 21 31 41 51 61 71 03 13 23 33 43 53 63 73
+ // 04 14 24 34 44 54 64 74 xx xx xx xx xx xx xx xx
+ // 05 15 25 35 45 55 65 75 xx xx xx xx xx xx xx xx
+ dd[0] = _mm_packus_epi16(d[0], d[2]);
+ dd[1] = _mm_packus_epi16(d[1], d[3]);
+ dd[2] = _mm_packus_epi16(d[4], d[4]);
+ dd[3] = _mm_packus_epi16(d[5], d[5]);
+
+ // 00 10 01 11 20 30 21 31 40 50 41 51 60 70 61 71
+ // 02 12 03 13 22 32 23 33 42 52 43 53 62 72 63 73
+ // 04 14 05 15 24 34 25 35 44 54 45 55 64 74 65 75
+ d[0] = _mm_unpacklo_epi16(dd[0], dd[1]);
+ d[1] = _mm_unpackhi_epi16(dd[0], dd[1]);
+ d[2] = _mm_unpacklo_epi16(dd[2], dd[3]);
+
+ // 00 10 01 11 02 12 03 13 20 30 21 31 22 32 23 33
+ // 40 50 41 51 42 52 43 53 60 70 61 71 62 72 63 73
+ // 04 14 05 15 xx xx xx xx 24 34 25 35 xx xx xx xx
+ // 44 54 45 55 xx xx xx xx 64 74 65 75 xx xx xx xx
+ dd[0] = _mm_unpacklo_epi32(d[0], d[1]);
+ dd[1] = _mm_unpackhi_epi32(d[0], d[1]);
+ dd[2] = _mm_unpacklo_epi32(d[2], d[2]);
+ dd[3] = _mm_unpackhi_epi32(d[2], d[2]);
+
+ // 00 10 01 11 02 12 03 13 04 14 05 15 xx xx xx xx
+ // 20 30 21 31 22 32 23 33 24 34 25 35 xx xx xx xx
+ // 40 50 41 51 42 52 43 53 44 54 45 55 xx xx xx xx
+ // 60 70 61 71 62 72 63 73 64 74 65 75 xx xx xx xx
+ d[0] = _mm_unpacklo_epi64(dd[0], dd[2]);
+ d[1] = _mm_unpackhi_epi64(dd[0], dd[2]);
+ d[2] = _mm_unpacklo_epi64(dd[1], dd[3]);
+ d[3] = _mm_unpackhi_epi64(dd[1], dd[3]);
+
+ // store 4 extra pixels
+ storeu_8bit_16x4(d, t, stride_hor);
+
+ s[0] = s[4];
+ s[1] = s[5];
+ s[2] = s[6];
+ s[3] = s[7];
+
+ t += 12;
+ x -= 6;
+ } while (x);
+ src += 8 * src_stride - 4 * width_hor / 3;
+ t += 3 * stride_hor + 4;
+ y -= 8;
+ } while (y);
+
+ // vertical 8x6
+ x = width_ver;
+ t = temp_buffer;
+ do {
+ // 00 10 01 11 02 12 03 13 04 14 05 15 06 16 07 17
+ // 20 30 21 31 22 32 23 33 24 34 25 35 26 36 27 37
+ // 40 50 41 51 42 52 43 53 44 54 45 55 46 56 47 57
+ // 60 70 61 71 62 72 63 73 64 74 65 75 66 76 67 77
+ loadu_8bit_16x4(t, stride_hor, s);
+ y = height_ver;
+
+ do {
+ // 80 90 81 91 82 92 83 93 84 94 85 95 86 96 87 97
+ // A0 B0 A1 B1 A2 B2 A3 B3 A4 B4 A5 B5 A6 B6 A7 B7
+ // C0 D0 C1 D1 C2 D2 C3 D3 C4 D4 C5 D5 C6 D6 C7 D7
+ // E0 F0 E1 F1 E2 F2 E3 F3 E4 F4 E5 F5 E6 F6 E7 F7
+ t += 4 * stride_hor;
+ loadu_8bit_16x4(t, stride_hor, &s[4]);
+
+ d[0] = convolve8_8_even_offset_ssse3(&s[0], f0);
+ d[1] = convolve8_funcs[offset_idx1](&s[offset1_q4 >> 5], f1);
+ d[2] = convolve8_funcs[offset_idx2](&s[offset2_q4 >> 5], f2);
+ d[3] = convolve8_8_even_offset_ssse3(&s[2], f0);
+ d[4] = convolve8_funcs[offset_idx1](&s[2 + (offset1_q4 >> 5)], f1);
+ d[5] = convolve8_funcs[offset_idx2](&s[2 + (offset2_q4 >> 5)], f2);
+
+ // 00 01 02 03 04 05 06 07 10 11 12 13 14 15 16 17
+ // 20 21 22 23 24 25 26 27 30 31 32 33 34 35 36 37
+ // 40 41 42 43 44 45 46 47 50 51 52 53 54 55 56 57
+ d[0] = _mm_packus_epi16(d[0], d[1]);
+ d[2] = _mm_packus_epi16(d[2], d[3]);
+ d[4] = _mm_packus_epi16(d[4], d[5]);
+
+ _mm_storel_epi64((__m128i *)(dst + 0 * dst_stride), d[0]);
+ _mm_storeh_epi64((__m128i *)(dst + 1 * dst_stride), d[0]);
+ _mm_storel_epi64((__m128i *)(dst + 2 * dst_stride), d[2]);
+ _mm_storeh_epi64((__m128i *)(dst + 3 * dst_stride), d[2]);
+ _mm_storel_epi64((__m128i *)(dst + 4 * dst_stride), d[4]);
+ _mm_storeh_epi64((__m128i *)(dst + 5 * dst_stride), d[4]);
+
+ s[0] = s[4];
+ s[1] = s[5];
+ s[2] = s[6];
+ s[3] = s[7];
+
+ dst += 6 * dst_stride;
+ y -= 6;
+ } while (y);
+ t -= stride_hor * 2 * height_ver / 3;
+ t += 16;
+ dst -= height_ver * dst_stride;
+ dst += 8;
+ x -= 8;
+ } while (x);
+}
+
static INLINE __m128i scale_1_to_2_phase_0_kernel(const __m128i *const s,
const __m128i *const f) {
__m128i ss[4], temp;
@@ -652,6 +848,36 @@ void vp9_scale_and_extend_frame_ssse3(const YV12_BUFFER_CONFIG *src,
scaled = 0;
}
}
+ } else if (4 * dst_w == 3 * src_w && 4 * dst_h == 3 * src_h) {
+ // 4 to 3
+ const int buffer_stride_hor = (dst_w + 5) - ((dst_w + 5) % 6) + 2;
+ const int buffer_stride_ver = (dst_w + 7) & ~7;
+ const int buffer_height = (4 * dst_h / 3 + SUBPEL_TAPS - 1 + 7) & ~7;
+ // When the vertical filter reads more pixels than the horizontal filter
+ // generated in each row, we need extra padding to avoid heap read overflow.
+ // For example, the horizontal filter generates 18 pixels but the vertical
+ // filter reads 24 pixels in a row. The difference is multiplied by 2 since
+ // two rows are interlaced together in the optimization.
+ const int extra_padding = (buffer_stride_ver > buffer_stride_hor)
+ ? 2 * (buffer_stride_ver - buffer_stride_hor)
+ : 0;
+ const int buffer_size = buffer_stride_hor * buffer_height + extra_padding;
+ uint8_t *const temp_buffer = (uint8_t *)malloc(buffer_size);
+ if (temp_buffer) {
+ scaled = 1;
+ scale_plane_4_to_3_general(
+ src->y_buffer, src->y_stride, dst->y_buffer, dst->y_stride, dst_w,
+ dst_h, vp9_filter_kernels[filter_type], phase_scaler, temp_buffer);
+ scale_plane_4_to_3_general(src->u_buffer, src->uv_stride, dst->u_buffer,
+ dst->uv_stride, dst_uv_w, dst_uv_h,
+ vp9_filter_kernels[filter_type], phase_scaler,
+ temp_buffer);
+ scale_plane_4_to_3_general(src->v_buffer, src->uv_stride, dst->v_buffer,
+ dst->uv_stride, dst_uv_w, dst_uv_h,
+ vp9_filter_kernels[filter_type], phase_scaler,
+ temp_buffer);
+ free(temp_buffer);
+ }
} else if (dst_w == src_w * 2 && dst_h == src_h * 2 && phase_scaler == 0) {
// 1 to 2
uint8_t *const temp_buffer = (uint8_t *)malloc(8 * ((src_w + 7) & ~7));
diff --git a/vpx_dsp/vpx_dsp.mk b/vpx_dsp/vpx_dsp.mk
index 2d071b96a..fa5feca16 100644
--- a/vpx_dsp/vpx_dsp.mk
+++ b/vpx_dsp/vpx_dsp.mk
@@ -89,6 +89,7 @@ DSP_SRCS-yes += vpx_filter.h
DSP_SRCS-$(ARCH_X86)$(ARCH_X86_64) += x86/convolve.h
DSP_SRCS-$(ARCH_X86)$(ARCH_X86_64) += x86/vpx_asm_stubs.c
DSP_SRCS-$(HAVE_SSSE3) += x86/convolve_ssse3.h
+DSP_SRCS-$(HAVE_AVX2) += x86/convolve_avx2.h
DSP_SRCS-$(HAVE_SSE2) += x86/vpx_subpixel_8t_sse2.asm
DSP_SRCS-$(HAVE_SSE2) += x86/vpx_subpixel_bilinear_sse2.asm
DSP_SRCS-$(HAVE_SSSE3) += x86/vpx_subpixel_8t_ssse3.asm
diff --git a/vpx_dsp/x86/convolve_avx2.h b/vpx_dsp/x86/convolve_avx2.h
new file mode 100644
index 000000000..c2e83b53f
--- /dev/null
+++ b/vpx_dsp/x86/convolve_avx2.h
@@ -0,0 +1,99 @@
+/*
+ * 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.
+ */
+
+#ifndef VPX_DSP_X86_CONVOLVE_AVX2_H_
+#define VPX_DSP_X86_CONVOLVE_AVX2_H_
+
+#include <immintrin.h> // AVX2
+
+#include "./vpx_config.h"
+
+#if defined(__clang__)
+#if (__clang_major__ > 0 && __clang_major__ < 3) || \
+ (__clang_major__ == 3 && __clang_minor__ <= 3) || \
+ (defined(__APPLE__) && defined(__apple_build_version__) && \
+ ((__clang_major__ == 4 && __clang_minor__ <= 2) || \
+ (__clang_major__ == 5 && __clang_minor__ == 0)))
+#define MM256_BROADCASTSI128_SI256(x) \
+ _mm_broadcastsi128_si256((__m128i const *)&(x))
+#else // clang > 3.3, and not 5.0 on macosx.
+#define MM256_BROADCASTSI128_SI256(x) _mm256_broadcastsi128_si256(x)
+#endif // clang <= 3.3
+#elif defined(__GNUC__)
+#if __GNUC__ < 4 || (__GNUC__ == 4 && __GNUC_MINOR__ <= 6)
+#define MM256_BROADCASTSI128_SI256(x) \
+ _mm_broadcastsi128_si256((__m128i const *)&(x))
+#elif __GNUC__ == 4 && __GNUC_MINOR__ == 7
+#define MM256_BROADCASTSI128_SI256(x) _mm_broadcastsi128_si256(x)
+#else // gcc > 4.7
+#define MM256_BROADCASTSI128_SI256(x) _mm256_broadcastsi128_si256(x)
+#endif // gcc <= 4.6
+#else // !(gcc || clang)
+#define MM256_BROADCASTSI128_SI256(x) _mm256_broadcastsi128_si256(x)
+#endif // __clang__
+
+static INLINE void shuffle_filter_avx2(const int16_t *const filter,
+ __m256i *const f) {
+ const __m256i f_values =
+ MM256_BROADCASTSI128_SI256(_mm_load_si128((const __m128i *)filter));
+ // pack and duplicate the filter values
+ f[0] = _mm256_shuffle_epi8(f_values, _mm256_set1_epi16(0x0200u));
+ f[1] = _mm256_shuffle_epi8(f_values, _mm256_set1_epi16(0x0604u));
+ f[2] = _mm256_shuffle_epi8(f_values, _mm256_set1_epi16(0x0a08u));
+ f[3] = _mm256_shuffle_epi8(f_values, _mm256_set1_epi16(0x0e0cu));
+}
+
+static INLINE __m256i convolve8_16_avx2(const __m256i *const s,
+ const __m256i *const f) {
+ // multiply 2 adjacent elements with the filter and add the result
+ const __m256i k_64 = _mm256_set1_epi16(1 << 6);
+ const __m256i x0 = _mm256_maddubs_epi16(s[0], f[0]);
+ const __m256i x1 = _mm256_maddubs_epi16(s[1], f[1]);
+ const __m256i x2 = _mm256_maddubs_epi16(s[2], f[2]);
+ const __m256i x3 = _mm256_maddubs_epi16(s[3], f[3]);
+ // add and saturate the results together
+ const __m256i min_x2x1 = _mm256_min_epi16(x2, x1);
+ const __m256i max_x2x1 = _mm256_max_epi16(x2, x1);
+ __m256i temp = _mm256_adds_epi16(x0, x3);
+ temp = _mm256_adds_epi16(temp, min_x2x1);
+ temp = _mm256_adds_epi16(temp, max_x2x1);
+ // round and shift by 7 bit each 16 bit
+ temp = _mm256_adds_epi16(temp, k_64);
+ temp = _mm256_srai_epi16(temp, 7);
+ return temp;
+}
+
+static INLINE __m128i convolve8_8_avx2(const __m256i *const s,
+ const __m256i *const f) {
+ // multiply 2 adjacent elements with the filter and add the result
+ const __m128i k_64 = _mm_set1_epi16(1 << 6);
+ const __m128i x0 = _mm_maddubs_epi16(_mm256_castsi256_si128(s[0]),
+ _mm256_castsi256_si128(f[0]));
+ const __m128i x1 = _mm_maddubs_epi16(_mm256_castsi256_si128(s[1]),
+ _mm256_castsi256_si128(f[1]));
+ const __m128i x2 = _mm_maddubs_epi16(_mm256_castsi256_si128(s[2]),
+ _mm256_castsi256_si128(f[2]));
+ const __m128i x3 = _mm_maddubs_epi16(_mm256_castsi256_si128(s[3]),
+ _mm256_castsi256_si128(f[3]));
+ // add and saturate the results together
+ const __m128i min_x2x1 = _mm_min_epi16(x2, x1);
+ const __m128i max_x2x1 = _mm_max_epi16(x2, x1);
+ __m128i temp = _mm_adds_epi16(x0, x3);
+ temp = _mm_adds_epi16(temp, min_x2x1);
+ temp = _mm_adds_epi16(temp, max_x2x1);
+ // round and shift by 7 bit each 16 bit
+ temp = _mm_adds_epi16(temp, k_64);
+ temp = _mm_srai_epi16(temp, 7);
+ return temp;
+}
+
+#undef MM256_BROADCASTSI128_SI256
+
+#endif // VPX_DSP_X86_CONVOLVE_AVX2_H_
diff --git a/vpx_dsp/x86/convolve_ssse3.h b/vpx_dsp/x86/convolve_ssse3.h
index b71da0e4e..8da28f0b2 100644
--- a/vpx_dsp/x86/convolve_ssse3.h
+++ b/vpx_dsp/x86/convolve_ssse3.h
@@ -11,6 +11,7 @@
#ifndef VPX_DSP_X86_CONVOLVE_SSSE3_H_
#define VPX_DSP_X86_CONVOLVE_SSSE3_H_
+#include <assert.h>
#include <tmmintrin.h> // SSSE3
#include "./vpx_config.h"
@@ -25,6 +26,20 @@ static INLINE void shuffle_filter_ssse3(const int16_t *const filter,
f[3] = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0e0cu));
}
+static INLINE void shuffle_filter_odd_ssse3(const int16_t *const filter,
+ __m128i *const f) {
+ const __m128i f_values = _mm_load_si128((const __m128i *)filter);
+ // pack and duplicate the filter values
+ // It utilizes the fact that the high byte of filter[3] is always 0 to clean
+ // half of f[0] and f[4].
+ assert(filter[3] >= 0 && filter[3] < 256);
+ f[0] = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0007u));
+ f[1] = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0402u));
+ f[2] = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0806u));
+ f[3] = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0c0au));
+ f[4] = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x070eu));
+}
+
static INLINE __m128i convolve8_8_ssse3(const __m128i *const s,
const __m128i *const f) {
// multiply 2 adjacent elements with the filter and add the result
@@ -45,4 +60,50 @@ static INLINE __m128i convolve8_8_ssse3(const __m128i *const s,
return temp;
}
+static INLINE __m128i convolve8_8_even_offset_ssse3(const __m128i *const s,
+ const __m128i *const f) {
+ // multiply 2 adjacent elements with the filter and add the result
+ const __m128i k_64 = _mm_set1_epi16(1 << 6);
+ const __m128i x0 = _mm_maddubs_epi16(s[0], f[0]);
+ const __m128i x1 = _mm_maddubs_epi16(s[1], f[1]);
+ const __m128i x2 = _mm_maddubs_epi16(s[2], f[2]);
+ const __m128i x3 = _mm_maddubs_epi16(s[3], f[3]);
+ // compensate the subtracted 64 in f[1]. x4 is always non negative.
+ const __m128i x4 = _mm_maddubs_epi16(s[1], _mm_set1_epi8(64));
+ // add and saturate the results together
+ __m128i temp = _mm_adds_epi16(x0, x3);
+ temp = _mm_adds_epi16(temp, x1);
+ temp = _mm_adds_epi16(temp, x2);
+ temp = _mm_adds_epi16(temp, x4);
+ // round and shift by 7 bit each 16 bit
+ temp = _mm_adds_epi16(temp, k_64);
+ temp = _mm_srai_epi16(temp, 7);
+ return temp;
+}
+
+static INLINE __m128i convolve8_8_odd_offset_ssse3(const __m128i *const s,
+ const __m128i *const f) {
+ // multiply 2 adjacent elements with the filter and add the result
+ const __m128i k_64 = _mm_set1_epi16(1 << 6);
+ const __m128i x0 = _mm_maddubs_epi16(s[0], f[0]);
+ const __m128i x1 = _mm_maddubs_epi16(s[1], f[1]);
+ const __m128i x2 = _mm_maddubs_epi16(s[2], f[2]);
+ const __m128i x3 = _mm_maddubs_epi16(s[3], f[3]);
+ const __m128i x4 = _mm_maddubs_epi16(s[4], f[4]);
+ // compensate the subtracted 64 in f[2]. x5 is always non negative.
+ const __m128i x5 = _mm_maddubs_epi16(s[2], _mm_set1_epi8(64));
+ __m128i temp;
+
+ // add and saturate the results together
+ temp = _mm_adds_epi16(x0, x1);
+ temp = _mm_adds_epi16(temp, x2);
+ temp = _mm_adds_epi16(temp, x3);
+ temp = _mm_adds_epi16(temp, x4);
+ temp = _mm_adds_epi16(temp, x5);
+ // round and shift by 7 bit each 16 bit
+ temp = _mm_adds_epi16(temp, k_64);
+ temp = _mm_srai_epi16(temp, 7);
+ return temp;
+}
+
#endif // VPX_DSP_X86_CONVOLVE_SSSE3_H_
diff --git a/vpx_dsp/x86/mem_sse2.h b/vpx_dsp/x86/mem_sse2.h
index f9f0a48a0..2ce738fb7 100644
--- a/vpx_dsp/x86/mem_sse2.h
+++ b/vpx_dsp/x86/mem_sse2.h
@@ -113,4 +113,12 @@ static INLINE void store_8bit_8x8(const __m128i *const s, uint8_t *const d,
_mm_storel_epi64((__m128i *)(d + 7 * stride), s[7]);
}
+static INLINE void storeu_8bit_16x4(const __m128i *const s, uint8_t *const d,
+ const ptrdiff_t stride) {
+ _mm_storeu_si128((__m128i *)(d + 0 * stride), s[0]);
+ _mm_storeu_si128((__m128i *)(d + 1 * stride), s[1]);
+ _mm_storeu_si128((__m128i *)(d + 2 * stride), s[2]);
+ _mm_storeu_si128((__m128i *)(d + 3 * stride), s[3]);
+}
+
#endif // VPX_DSP_X86_MEM_SSE2_H_
diff --git a/vpx_dsp/x86/vpx_subpixel_8t_intrin_avx2.c b/vpx_dsp/x86/vpx_subpixel_8t_intrin_avx2.c
index 4e851b58e..d0919695c 100644
--- a/vpx_dsp/x86/vpx_subpixel_8t_intrin_avx2.c
+++ b/vpx_dsp/x86/vpx_subpixel_8t_intrin_avx2.c
@@ -12,9 +12,10 @@
#include "./vpx_dsp_rtcd.h"
#include "vpx_dsp/x86/convolve.h"
+#include "vpx_dsp/x86/convolve_avx2.h"
#include "vpx_ports/mem.h"
-// filters for 16_h8 and 16_v8
+// filters for 16_h8
DECLARE_ALIGNED(32, static const uint8_t, filt1_global_avx2[32]) = {
0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8,
0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8
@@ -35,160 +36,68 @@ DECLARE_ALIGNED(32, static const uint8_t, filt4_global_avx2[32]) = {
6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14
};
-#if defined(__clang__)
-#if (__clang_major__ > 0 && __clang_major__ < 3) || \
- (__clang_major__ == 3 && __clang_minor__ <= 3) || \
- (defined(__APPLE__) && defined(__apple_build_version__) && \
- ((__clang_major__ == 4 && __clang_minor__ <= 2) || \
- (__clang_major__ == 5 && __clang_minor__ == 0)))
-#define MM256_BROADCASTSI128_SI256(x) \
- _mm_broadcastsi128_si256((__m128i const *)&(x))
-#else // clang > 3.3, and not 5.0 on macosx.
-#define MM256_BROADCASTSI128_SI256(x) _mm256_broadcastsi128_si256(x)
-#endif // clang <= 3.3
-#elif defined(__GNUC__)
-#if __GNUC__ < 4 || (__GNUC__ == 4 && __GNUC_MINOR__ <= 6)
-#define MM256_BROADCASTSI128_SI256(x) \
- _mm_broadcastsi128_si256((__m128i const *)&(x))
-#elif __GNUC__ == 4 && __GNUC_MINOR__ == 7
-#define MM256_BROADCASTSI128_SI256(x) _mm_broadcastsi128_si256(x)
-#else // gcc > 4.7
-#define MM256_BROADCASTSI128_SI256(x) _mm256_broadcastsi128_si256(x)
-#endif // gcc <= 4.6
-#else // !(gcc || clang)
-#define MM256_BROADCASTSI128_SI256(x) _mm256_broadcastsi128_si256(x)
-#endif // __clang__
-
-static INLINE void vpx_filter_block1d16_h8_X_avx2(
+static INLINE void vpx_filter_block1d16_h8_x_avx2(
const uint8_t *src_ptr, ptrdiff_t src_pixels_per_line, uint8_t *output_ptr,
ptrdiff_t output_pitch, uint32_t output_height, const int16_t *filter,
const int avg) {
- __m128i filtersReg, outReg1, outReg2;
- __m256i addFilterReg64, filt1Reg, filt2Reg, filt3Reg, filt4Reg;
- __m256i firstFilters, secondFilters, thirdFilters, forthFilters;
- __m256i srcRegFilt32b1_1, srcRegFilt32b2_1, srcRegFilt32b2, srcRegFilt32b3;
- __m256i srcReg32b1, srcReg32b2, filtersReg32;
+ __m128i outReg1, outReg2;
+ __m256i outReg32b1, outReg32b2;
unsigned int i;
ptrdiff_t src_stride, dst_stride;
+ __m256i f[4], filt[4], s[4];
- // create a register with 0,64,0,64,0,64,0,64,0,64,0,64,0,64,0,64
- addFilterReg64 = _mm256_set1_epi32((int)0x0400040u);
- filtersReg = _mm_loadu_si128((const __m128i *)filter);
- // converting the 16 bit (short) to 8 bit (byte) and have the same data
- // in both lanes of 128 bit register.
- filtersReg = _mm_packs_epi16(filtersReg, filtersReg);
- // have the same data in both lanes of a 256 bit register
- filtersReg32 = MM256_BROADCASTSI128_SI256(filtersReg);
-
- // duplicate only the first 16 bits (first and second byte)
- // across 256 bit register
- firstFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x100u));
- // duplicate only the second 16 bits (third and forth byte)
- // across 256 bit register
- secondFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x302u));
- // duplicate only the third 16 bits (fifth and sixth byte)
- // across 256 bit register
- thirdFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x504u));
- // duplicate only the forth 16 bits (seventh and eighth byte)
- // across 256 bit register
- forthFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x706u));
-
- filt1Reg = _mm256_load_si256((__m256i const *)filt1_global_avx2);
- filt2Reg = _mm256_load_si256((__m256i const *)filt2_global_avx2);
- filt3Reg = _mm256_load_si256((__m256i const *)filt3_global_avx2);
- filt4Reg = _mm256_load_si256((__m256i const *)filt4_global_avx2);
+ shuffle_filter_avx2(filter, f);
+ filt[0] = _mm256_load_si256((__m256i const *)filt1_global_avx2);
+ filt[1] = _mm256_load_si256((__m256i const *)filt2_global_avx2);
+ filt[2] = _mm256_load_si256((__m256i const *)filt3_global_avx2);
+ filt[3] = _mm256_load_si256((__m256i const *)filt4_global_avx2);
// multiple the size of the source and destination stride by two
src_stride = src_pixels_per_line << 1;
dst_stride = output_pitch << 1;
for (i = output_height; i > 1; i -= 2) {
+ __m256i srcReg;
+
// load the 2 strides of source
- srcReg32b1 =
+ srcReg =
_mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(src_ptr - 3)));
- srcReg32b1 = _mm256_inserti128_si256(
- srcReg32b1,
+ srcReg = _mm256_inserti128_si256(
+ srcReg,
_mm_loadu_si128((const __m128i *)(src_ptr + src_pixels_per_line - 3)),
1);
// filter the source buffer
- srcRegFilt32b1_1 = _mm256_shuffle_epi8(srcReg32b1, filt1Reg);
- srcRegFilt32b2 = _mm256_shuffle_epi8(srcReg32b1, filt4Reg);
-
- // multiply 2 adjacent elements with the filter and add the result
- srcRegFilt32b1_1 = _mm256_maddubs_epi16(srcRegFilt32b1_1, firstFilters);
- srcRegFilt32b2 = _mm256_maddubs_epi16(srcRegFilt32b2, forthFilters);
-
- // add and saturate the results together
- srcRegFilt32b1_1 = _mm256_adds_epi16(srcRegFilt32b1_1, srcRegFilt32b2);
-
- // filter the source buffer
- srcRegFilt32b3 = _mm256_shuffle_epi8(srcReg32b1, filt2Reg);
- srcRegFilt32b2 = _mm256_shuffle_epi8(srcReg32b1, filt3Reg);
-
- // multiply 2 adjacent elements with the filter and add the result
- srcRegFilt32b3 = _mm256_maddubs_epi16(srcRegFilt32b3, secondFilters);
- srcRegFilt32b2 = _mm256_maddubs_epi16(srcRegFilt32b2, thirdFilters);
-
- // add and saturate the results together
- srcRegFilt32b1_1 = _mm256_adds_epi16(
- srcRegFilt32b1_1, _mm256_min_epi16(srcRegFilt32b3, srcRegFilt32b2));
+ s[0] = _mm256_shuffle_epi8(srcReg, filt[0]);
+ s[1] = _mm256_shuffle_epi8(srcReg, filt[1]);
+ s[2] = _mm256_shuffle_epi8(srcReg, filt[2]);
+ s[3] = _mm256_shuffle_epi8(srcReg, filt[3]);
+ outReg32b1 = convolve8_16_avx2(s, f);
// reading 2 strides of the next 16 bytes
// (part of it was being read by earlier read)
- srcReg32b2 =
+ srcReg =
_mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(src_ptr + 5)));
- srcReg32b2 = _mm256_inserti128_si256(
- srcReg32b2,
+ srcReg = _mm256_inserti128_si256(
+ srcReg,
_mm_loadu_si128((const __m128i *)(src_ptr + src_pixels_per_line + 5)),
1);
- // add and saturate the results together
- srcRegFilt32b1_1 = _mm256_adds_epi16(
- srcRegFilt32b1_1, _mm256_max_epi16(srcRegFilt32b3, srcRegFilt32b2));
-
// filter the source buffer
- srcRegFilt32b2_1 = _mm256_shuffle_epi8(srcReg32b2, filt1Reg);
- srcRegFilt32b2 = _mm256_shuffle_epi8(srcReg32b2, filt4Reg);
-
- // multiply 2 adjacent elements with the filter and add the result
- srcRegFilt32b2_1 = _mm256_maddubs_epi16(srcRegFilt32b2_1, firstFilters);
- srcRegFilt32b2 = _mm256_maddubs_epi16(srcRegFilt32b2, forthFilters);
+ s[0] = _mm256_shuffle_epi8(srcReg, filt[0]);
+ s[1] = _mm256_shuffle_epi8(srcReg, filt[1]);
+ s[2] = _mm256_shuffle_epi8(srcReg, filt[2]);
+ s[3] = _mm256_shuffle_epi8(srcReg, filt[3]);
+ outReg32b2 = convolve8_16_avx2(s, f);
- // add and saturate the results together
- srcRegFilt32b2_1 = _mm256_adds_epi16(srcRegFilt32b2_1, srcRegFilt32b2);
-
- // filter the source buffer
- srcRegFilt32b3 = _mm256_shuffle_epi8(srcReg32b2, filt2Reg);
- srcRegFilt32b2 = _mm256_shuffle_epi8(srcReg32b2, filt3Reg);
-
- // multiply 2 adjacent elements with the filter and add the result
- srcRegFilt32b3 = _mm256_maddubs_epi16(srcRegFilt32b3, secondFilters);
- srcRegFilt32b2 = _mm256_maddubs_epi16(srcRegFilt32b2, thirdFilters);
-
- // add and saturate the results together
- srcRegFilt32b2_1 = _mm256_adds_epi16(
- srcRegFilt32b2_1, _mm256_min_epi16(srcRegFilt32b3, srcRegFilt32b2));
- srcRegFilt32b2_1 = _mm256_adds_epi16(
- srcRegFilt32b2_1, _mm256_max_epi16(srcRegFilt32b3, srcRegFilt32b2));
-
- srcRegFilt32b1_1 = _mm256_adds_epi16(srcRegFilt32b1_1, addFilterReg64);
-
- srcRegFilt32b2_1 = _mm256_adds_epi16(srcRegFilt32b2_1, addFilterReg64);
-
- // shift by 7 bit each 16 bit
- srcRegFilt32b1_1 = _mm256_srai_epi16(srcRegFilt32b1_1, 7);
- srcRegFilt32b2_1 = _mm256_srai_epi16(srcRegFilt32b2_1, 7);
-
- // shrink to 8 bit each 16 bits, the first lane contain the first
- // convolve result and the second lane contain the second convolve
- // result
- srcRegFilt32b1_1 = _mm256_packus_epi16(srcRegFilt32b1_1, srcRegFilt32b2_1);
+ // shrink to 8 bit each 16 bits, the low and high 64-bits of each lane
+ // contain the first and second convolve result respectively
+ outReg32b1 = _mm256_packus_epi16(outReg32b1, outReg32b2);
src_ptr += src_stride;
// average if necessary
- outReg1 = _mm256_castsi256_si128(srcRegFilt32b1_1);
- outReg2 = _mm256_extractf128_si256(srcRegFilt32b1_1, 1);
+ outReg1 = _mm256_castsi256_si128(outReg32b1);
+ outReg2 = _mm256_extractf128_si256(outReg32b1, 1);
if (avg) {
outReg1 = _mm_avg_epu8(outReg1, _mm_load_si128((__m128i *)output_ptr));
outReg2 = _mm_avg_epu8(
@@ -207,89 +116,40 @@ static INLINE void vpx_filter_block1d16_h8_X_avx2(
// if the number of strides is odd.
// process only 16 bytes
if (i > 0) {
- __m128i srcReg1, srcReg2, srcRegFilt1_1, srcRegFilt2_1;
- __m128i srcRegFilt2, srcRegFilt3;
-
- srcReg1 = _mm_loadu_si128((const __m128i *)(src_ptr - 3));
-
- // filter the source buffer
- srcRegFilt1_1 = _mm_shuffle_epi8(srcReg1, _mm256_castsi256_si128(filt1Reg));
- srcRegFilt2 = _mm_shuffle_epi8(srcReg1, _mm256_castsi256_si128(filt4Reg));
-
- // multiply 2 adjacent elements with the filter and add the result
- srcRegFilt1_1 =
- _mm_maddubs_epi16(srcRegFilt1_1, _mm256_castsi256_si128(firstFilters));
- srcRegFilt2 =
- _mm_maddubs_epi16(srcRegFilt2, _mm256_castsi256_si128(forthFilters));
+ __m128i srcReg;
- // add and saturate the results together
- srcRegFilt1_1 = _mm_adds_epi16(srcRegFilt1_1, srcRegFilt2);
+ // load the first 16 bytes of the last row
+ srcReg = _mm_loadu_si128((const __m128i *)(src_ptr - 3));
// filter the source buffer
- srcRegFilt3 = _mm_shuffle_epi8(srcReg1, _mm256_castsi256_si128(filt2Reg));
- srcRegFilt2 = _mm_shuffle_epi8(srcReg1, _mm256_castsi256_si128(filt3Reg));
-
- // multiply 2 adjacent elements with the filter and add the result
- srcRegFilt3 =
- _mm_maddubs_epi16(srcRegFilt3, _mm256_castsi256_si128(secondFilters));
- srcRegFilt2 =
- _mm_maddubs_epi16(srcRegFilt2, _mm256_castsi256_si128(thirdFilters));
-
- // add and saturate the results together
- srcRegFilt1_1 =
- _mm_adds_epi16(srcRegFilt1_1, _mm_min_epi16(srcRegFilt3, srcRegFilt2));
+ s[0] = _mm256_castsi128_si256(
+ _mm_shuffle_epi8(srcReg, _mm256_castsi256_si128(filt[0])));
+ s[1] = _mm256_castsi128_si256(
+ _mm_shuffle_epi8(srcReg, _mm256_castsi256_si128(filt[1])));
+ s[2] = _mm256_castsi128_si256(
+ _mm_shuffle_epi8(srcReg, _mm256_castsi256_si128(filt[2])));
+ s[3] = _mm256_castsi128_si256(
+ _mm_shuffle_epi8(srcReg, _mm256_castsi256_si128(filt[3])));
+ outReg1 = convolve8_8_avx2(s, f);
// reading the next 16 bytes
// (part of it was being read by earlier read)
- srcReg2 = _mm_loadu_si128((const __m128i *)(src_ptr + 5));
-
- // add and saturate the results together
- srcRegFilt1_1 =
- _mm_adds_epi16(srcRegFilt1_1, _mm_max_epi16(srcRegFilt3, srcRegFilt2));
-
- // filter the source buffer
- srcRegFilt2_1 = _mm_shuffle_epi8(srcReg2, _mm256_castsi256_si128(filt1Reg));
- srcRegFilt2 = _mm_shuffle_epi8(srcReg2, _mm256_castsi256_si128(filt4Reg));
-
- // multiply 2 adjacent elements with the filter and add the result
- srcRegFilt2_1 =
- _mm_maddubs_epi16(srcRegFilt2_1, _mm256_castsi256_si128(firstFilters));
- srcRegFilt2 =
- _mm_maddubs_epi16(srcRegFilt2, _mm256_castsi256_si128(forthFilters));
-
- // add and saturate the results together
- srcRegFilt2_1 = _mm_adds_epi16(srcRegFilt2_1, srcRegFilt2);
+ srcReg = _mm_loadu_si128((const __m128i *)(src_ptr + 5));
// filter the source buffer
- srcRegFilt3 = _mm_shuffle_epi8(srcReg2, _mm256_castsi256_si128(filt2Reg));
- srcRegFilt2 = _mm_shuffle_epi8(srcReg2, _mm256_castsi256_si128(filt3Reg));
-
- // multiply 2 adjacent elements with the filter and add the result
- srcRegFilt3 =
- _mm_maddubs_epi16(srcRegFilt3, _mm256_castsi256_si128(secondFilters));
- srcRegFilt2 =
- _mm_maddubs_epi16(srcRegFilt2, _mm256_castsi256_si128(thirdFilters));
-
- // add and saturate the results together
- srcRegFilt2_1 =
- _mm_adds_epi16(srcRegFilt2_1, _mm_min_epi16(srcRegFilt3, srcRegFilt2));
- srcRegFilt2_1 =
- _mm_adds_epi16(srcRegFilt2_1, _mm_max_epi16(srcRegFilt3, srcRegFilt2));
-
- srcRegFilt1_1 =
- _mm_adds_epi16(srcRegFilt1_1, _mm256_castsi256_si128(addFilterReg64));
-
- srcRegFilt2_1 =
- _mm_adds_epi16(srcRegFilt2_1, _mm256_castsi256_si128(addFilterReg64));
-
- // shift by 7 bit each 16 bit
- srcRegFilt1_1 = _mm_srai_epi16(srcRegFilt1_1, 7);
- srcRegFilt2_1 = _mm_srai_epi16(srcRegFilt2_1, 7);
-
- // shrink to 8 bit each 16 bits, the first lane contain the first
- // convolve result and the second lane contain the second convolve
- // result
- outReg1 = _mm_packus_epi16(srcRegFilt1_1, srcRegFilt2_1);
+ s[0] = _mm256_castsi128_si256(
+ _mm_shuffle_epi8(srcReg, _mm256_castsi256_si128(filt[0])));
+ s[1] = _mm256_castsi128_si256(
+ _mm_shuffle_epi8(srcReg, _mm256_castsi256_si128(filt[1])));
+ s[2] = _mm256_castsi128_si256(
+ _mm_shuffle_epi8(srcReg, _mm256_castsi256_si128(filt[2])));
+ s[3] = _mm256_castsi128_si256(
+ _mm_shuffle_epi8(srcReg, _mm256_castsi256_si128(filt[3])));
+ outReg2 = convolve8_8_avx2(s, f);
+
+ // shrink to 8 bit each 16 bits, the low and high 64-bits of each lane
+ // contain the first and second convolve result respectively
+ outReg1 = _mm_packus_epi16(outReg1, outReg2);
// average if necessary
if (avg) {
@@ -304,169 +164,99 @@ static INLINE void vpx_filter_block1d16_h8_X_avx2(
static void vpx_filter_block1d16_h8_avx2(
const uint8_t *src_ptr, ptrdiff_t src_stride, uint8_t *output_ptr,
ptrdiff_t dst_stride, uint32_t output_height, const int16_t *filter) {
- vpx_filter_block1d16_h8_X_avx2(src_ptr, src_stride, output_ptr, dst_stride,
+ vpx_filter_block1d16_h8_x_avx2(src_ptr, src_stride, output_ptr, dst_stride,
output_height, filter, 0);
}
static void vpx_filter_block1d16_h8_avg_avx2(
const uint8_t *src_ptr, ptrdiff_t src_stride, uint8_t *output_ptr,
ptrdiff_t dst_stride, uint32_t output_height, const int16_t *filter) {
- vpx_filter_block1d16_h8_X_avx2(src_ptr, src_stride, output_ptr, dst_stride,
+ vpx_filter_block1d16_h8_x_avx2(src_ptr, src_stride, output_ptr, dst_stride,
output_height, filter, 1);
}
-static INLINE void vpx_filter_block1d16_v8_X_avx2(
+static INLINE void vpx_filter_block1d16_v8_x_avx2(
const uint8_t *src_ptr, ptrdiff_t src_pitch, uint8_t *output_ptr,
ptrdiff_t out_pitch, uint32_t output_height, const int16_t *filter,
const int avg) {
- __m128i filtersReg, outReg1, outReg2;
- __m256i addFilterReg64;
- __m256i srcReg32b1, srcReg32b2, srcReg32b3, srcReg32b4, srcReg32b5;
- __m256i srcReg32b6, srcReg32b7, srcReg32b8, srcReg32b9, srcReg32b10;
- __m256i srcReg32b11, srcReg32b12, filtersReg32;
- __m256i firstFilters, secondFilters, thirdFilters, forthFilters;
+ __m128i outReg1, outReg2;
+ __m256i srcRegHead1;
unsigned int i;
ptrdiff_t src_stride, dst_stride;
+ __m256i f[4], s1[4], s2[4];
- // create a register with 0,64,0,64,0,64,0,64,0,64,0,64,0,64,0,64
- addFilterReg64 = _mm256_set1_epi32((int)0x0400040u);
- filtersReg = _mm_loadu_si128((const __m128i *)filter);
- // converting the 16 bit (short) to 8 bit (byte) and have the
- // same data in both lanes of 128 bit register.
- filtersReg = _mm_packs_epi16(filtersReg, filtersReg);
- // have the same data in both lanes of a 256 bit register
- filtersReg32 = MM256_BROADCASTSI128_SI256(filtersReg);
-
- // duplicate only the first 16 bits (first and second byte)
- // across 256 bit register
- firstFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x100u));
- // duplicate only the second 16 bits (third and forth byte)
- // across 256 bit register
- secondFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x302u));
- // duplicate only the third 16 bits (fifth and sixth byte)
- // across 256 bit register
- thirdFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x504u));
- // duplicate only the forth 16 bits (seventh and eighth byte)
- // across 256 bit register
- forthFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x706u));
+ shuffle_filter_avx2(filter, f);
// multiple the size of the source and destination stride by two
src_stride = src_pitch << 1;
dst_stride = out_pitch << 1;
- // load 16 bytes 7 times in stride of src_pitch
- srcReg32b1 =
- _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(src_ptr)));
- srcReg32b2 = _mm256_castsi128_si256(
- _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch)));
- srcReg32b3 = _mm256_castsi128_si256(
- _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 2)));
- srcReg32b4 = _mm256_castsi128_si256(
- _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 3)));
- srcReg32b5 = _mm256_castsi128_si256(
- _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 4)));
- srcReg32b6 = _mm256_castsi128_si256(
- _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 5)));
- srcReg32b7 = _mm256_castsi128_si256(
- _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 6)));
-
- // have each consecutive loads on the same 256 register
- srcReg32b1 = _mm256_inserti128_si256(srcReg32b1,
- _mm256_castsi256_si128(srcReg32b2), 1);
- srcReg32b2 = _mm256_inserti128_si256(srcReg32b2,
- _mm256_castsi256_si128(srcReg32b3), 1);
- srcReg32b3 = _mm256_inserti128_si256(srcReg32b3,
- _mm256_castsi256_si128(srcReg32b4), 1);
- srcReg32b4 = _mm256_inserti128_si256(srcReg32b4,
- _mm256_castsi256_si128(srcReg32b5), 1);
- srcReg32b5 = _mm256_inserti128_si256(srcReg32b5,
- _mm256_castsi256_si128(srcReg32b6), 1);
- srcReg32b6 = _mm256_inserti128_si256(srcReg32b6,
- _mm256_castsi256_si128(srcReg32b7), 1);
-
- // merge every two consecutive registers except the last one
- srcReg32b10 = _mm256_unpacklo_epi8(srcReg32b1, srcReg32b2);
- srcReg32b1 = _mm256_unpackhi_epi8(srcReg32b1, srcReg32b2);
-
- // save
- srcReg32b11 = _mm256_unpacklo_epi8(srcReg32b3, srcReg32b4);
-
- // save
- srcReg32b3 = _mm256_unpackhi_epi8(srcReg32b3, srcReg32b4);
-
- // save
- srcReg32b2 = _mm256_unpacklo_epi8(srcReg32b5, srcReg32b6);
-
- // save
- srcReg32b5 = _mm256_unpackhi_epi8(srcReg32b5, srcReg32b6);
+ {
+ __m128i s[6];
+ __m256i s32b[6];
+
+ // load 16 bytes 7 times in stride of src_pitch
+ s[0] = _mm_loadu_si128((const __m128i *)(src_ptr + 0 * src_pitch));
+ s[1] = _mm_loadu_si128((const __m128i *)(src_ptr + 1 * src_pitch));
+ s[2] = _mm_loadu_si128((const __m128i *)(src_ptr + 2 * src_pitch));
+ s[3] = _mm_loadu_si128((const __m128i *)(src_ptr + 3 * src_pitch));
+ s[4] = _mm_loadu_si128((const __m128i *)(src_ptr + 4 * src_pitch));
+ s[5] = _mm_loadu_si128((const __m128i *)(src_ptr + 5 * src_pitch));
+ srcRegHead1 = _mm256_castsi128_si256(
+ _mm_loadu_si128((const __m128i *)(src_ptr + 6 * src_pitch)));
+
+ // have each consecutive loads on the same 256 register
+ s32b[0] = _mm256_inserti128_si256(_mm256_castsi128_si256(s[0]), s[1], 1);
+ s32b[1] = _mm256_inserti128_si256(_mm256_castsi128_si256(s[1]), s[2], 1);
+ s32b[2] = _mm256_inserti128_si256(_mm256_castsi128_si256(s[2]), s[3], 1);
+ s32b[3] = _mm256_inserti128_si256(_mm256_castsi128_si256(s[3]), s[4], 1);
+ s32b[4] = _mm256_inserti128_si256(_mm256_castsi128_si256(s[4]), s[5], 1);
+ s32b[5] = _mm256_inserti128_si256(_mm256_castsi128_si256(s[5]),
+ _mm256_castsi256_si128(srcRegHead1), 1);
+
+ // merge every two consecutive registers except the last one
+ // the first lanes contain values for filtering odd rows (1,3,5...) and
+ // the second lanes contain values for filtering even rows (2,4,6...)
+ s1[0] = _mm256_unpacklo_epi8(s32b[0], s32b[1]);
+ s2[0] = _mm256_unpackhi_epi8(s32b[0], s32b[1]);
+ s1[1] = _mm256_unpacklo_epi8(s32b[2], s32b[3]);
+ s2[1] = _mm256_unpackhi_epi8(s32b[2], s32b[3]);
+ s1[2] = _mm256_unpacklo_epi8(s32b[4], s32b[5]);
+ s2[2] = _mm256_unpackhi_epi8(s32b[4], s32b[5]);
+ }
for (i = output_height; i > 1; i -= 2) {
- // load the last 2 loads of 16 bytes and have every two
+ __m256i srcRegHead2, srcRegHead3;
+
+ // load the next 2 loads of 16 bytes and have every two
// consecutive loads in the same 256 bit register
- srcReg32b8 = _mm256_castsi128_si256(
- _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 7)));
- srcReg32b7 = _mm256_inserti128_si256(srcReg32b7,
- _mm256_castsi256_si128(srcReg32b8), 1);
- srcReg32b9 = _mm256_castsi128_si256(
- _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 8)));
- srcReg32b8 = _mm256_inserti128_si256(srcReg32b8,
- _mm256_castsi256_si128(srcReg32b9), 1);
-
- // merge every two consecutive registers
- // save
- srcReg32b4 = _mm256_unpacklo_epi8(srcReg32b7, srcReg32b8);
- srcReg32b7 = _mm256_unpackhi_epi8(srcReg32b7, srcReg32b8);
-
- // multiply 2 adjacent elements with the filter and add the result
- srcReg32b10 = _mm256_maddubs_epi16(srcReg32b10, firstFilters);
- srcReg32b6 = _mm256_maddubs_epi16(srcReg32b4, forthFilters);
-
- // add and saturate the results together
- srcReg32b10 = _mm256_adds_epi16(srcReg32b10, srcReg32b6);
-
- // multiply 2 adjacent elements with the filter and add the result
- srcReg32b8 = _mm256_maddubs_epi16(srcReg32b11, secondFilters);
- srcReg32b12 = _mm256_maddubs_epi16(srcReg32b2, thirdFilters);
-
- // add and saturate the results together
- srcReg32b10 = _mm256_adds_epi16(srcReg32b10,
- _mm256_min_epi16(srcReg32b8, srcReg32b12));
- srcReg32b10 = _mm256_adds_epi16(srcReg32b10,
- _mm256_max_epi16(srcReg32b8, srcReg32b12));
-
- // multiply 2 adjacent elements with the filter and add the result
- srcReg32b1 = _mm256_maddubs_epi16(srcReg32b1, firstFilters);
- srcReg32b6 = _mm256_maddubs_epi16(srcReg32b7, forthFilters);
-
- srcReg32b1 = _mm256_adds_epi16(srcReg32b1, srcReg32b6);
-
- // multiply 2 adjacent elements with the filter and add the result
- srcReg32b8 = _mm256_maddubs_epi16(srcReg32b3, secondFilters);
- srcReg32b12 = _mm256_maddubs_epi16(srcReg32b5, thirdFilters);
-
- // add and saturate the results together
- srcReg32b1 = _mm256_adds_epi16(srcReg32b1,
- _mm256_min_epi16(srcReg32b8, srcReg32b12));
- srcReg32b1 = _mm256_adds_epi16(srcReg32b1,
- _mm256_max_epi16(srcReg32b8, srcReg32b12));
-
- srcReg32b10 = _mm256_adds_epi16(srcReg32b10, addFilterReg64);
- srcReg32b1 = _mm256_adds_epi16(srcReg32b1, addFilterReg64);
-
- // shift by 7 bit each 16 bit
- srcReg32b10 = _mm256_srai_epi16(srcReg32b10, 7);
- srcReg32b1 = _mm256_srai_epi16(srcReg32b1, 7);
-
- // shrink to 8 bit each 16 bits, the first lane contain the first
- // convolve result and the second lane contain the second convolve
- // result
- srcReg32b1 = _mm256_packus_epi16(srcReg32b10, srcReg32b1);
+ srcRegHead2 = _mm256_castsi128_si256(
+ _mm_loadu_si128((const __m128i *)(src_ptr + 7 * src_pitch)));
+ srcRegHead1 = _mm256_inserti128_si256(
+ srcRegHead1, _mm256_castsi256_si128(srcRegHead2), 1);
+ srcRegHead3 = _mm256_castsi128_si256(
+ _mm_loadu_si128((const __m128i *)(src_ptr + 8 * src_pitch)));
+ srcRegHead2 = _mm256_inserti128_si256(
+ srcRegHead2, _mm256_castsi256_si128(srcRegHead3), 1);
+
+ // merge the two new consecutive registers
+ // the first lane contain values for filtering odd rows (1,3,5...) and
+ // the second lane contain values for filtering even rows (2,4,6...)
+ s1[3] = _mm256_unpacklo_epi8(srcRegHead1, srcRegHead2);
+ s2[3] = _mm256_unpackhi_epi8(srcRegHead1, srcRegHead2);
+
+ s1[0] = convolve8_16_avx2(s1, f);
+ s2[0] = convolve8_16_avx2(s2, f);
+
+ // shrink to 8 bit each 16 bits, the low and high 64-bits of each lane
+ // contain the first and second convolve result respectively
+ s1[0] = _mm256_packus_epi16(s1[0], s2[0]);
src_ptr += src_stride;
// average if necessary
- outReg1 = _mm256_castsi256_si128(srcReg32b1);
- outReg2 = _mm256_extractf128_si256(srcReg32b1, 1);
+ outReg1 = _mm256_castsi256_si128(s1[0]);
+ outReg2 = _mm256_extractf128_si256(s1[0], 1);
if (avg) {
outReg1 = _mm_avg_epu8(outReg1, _mm_load_si128((__m128i *)output_ptr));
outReg2 = _mm_avg_epu8(
@@ -481,78 +271,35 @@ static INLINE void vpx_filter_block1d16_v8_X_avx2(
output_ptr += dst_stride;
- // save part of the registers for next strides
- srcReg32b10 = srcReg32b11;
- srcReg32b1 = srcReg32b3;
- srcReg32b11 = srcReg32b2;
- srcReg32b3 = srcReg32b5;
- srcReg32b2 = srcReg32b4;
- srcReg32b5 = srcReg32b7;
- srcReg32b7 = srcReg32b9;
+ // shift down by two rows
+ s1[0] = s1[1];
+ s2[0] = s2[1];
+ s1[1] = s1[2];
+ s2[1] = s2[2];
+ s1[2] = s1[3];
+ s2[2] = s2[3];
+ srcRegHead1 = srcRegHead3;
}
+
+ // if the number of strides is odd.
+ // process only 16 bytes
if (i > 0) {
- __m128i srcRegFilt1, srcRegFilt3, srcRegFilt4, srcRegFilt5;
- __m128i srcRegFilt6, srcRegFilt7, srcRegFilt8;
// load the last 16 bytes
- srcRegFilt8 = _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 7));
+ const __m128i srcRegHead2 =
+ _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 7));
// merge the last 2 results together
- srcRegFilt4 =
- _mm_unpacklo_epi8(_mm256_castsi256_si128(srcReg32b7), srcRegFilt8);
- srcRegFilt7 =
- _mm_unpackhi_epi8(_mm256_castsi256_si128(srcReg32b7), srcRegFilt8);
-
- // multiply 2 adjacent elements with the filter and add the result
- srcRegFilt1 = _mm_maddubs_epi16(_mm256_castsi256_si128(srcReg32b10),
- _mm256_castsi256_si128(firstFilters));
- srcRegFilt4 =
- _mm_maddubs_epi16(srcRegFilt4, _mm256_castsi256_si128(forthFilters));
- srcRegFilt3 = _mm_maddubs_epi16(_mm256_castsi256_si128(srcReg32b1),
- _mm256_castsi256_si128(firstFilters));
- srcRegFilt7 =
- _mm_maddubs_epi16(srcRegFilt7, _mm256_castsi256_si128(forthFilters));
-
- // add and saturate the results together
- srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, srcRegFilt4);
- srcRegFilt3 = _mm_adds_epi16(srcRegFilt3, srcRegFilt7);
-
- // multiply 2 adjacent elements with the filter and add the result
- srcRegFilt4 = _mm_maddubs_epi16(_mm256_castsi256_si128(srcReg32b11),
- _mm256_castsi256_si128(secondFilters));
- srcRegFilt5 = _mm_maddubs_epi16(_mm256_castsi256_si128(srcReg32b3),
- _mm256_castsi256_si128(secondFilters));
-
- // multiply 2 adjacent elements with the filter and add the result
- srcRegFilt6 = _mm_maddubs_epi16(_mm256_castsi256_si128(srcReg32b2),
- _mm256_castsi256_si128(thirdFilters));
- srcRegFilt7 = _mm_maddubs_epi16(_mm256_castsi256_si128(srcReg32b5),
- _mm256_castsi256_si128(thirdFilters));
-
- // add and saturate the results together
- srcRegFilt1 =
- _mm_adds_epi16(srcRegFilt1, _mm_min_epi16(srcRegFilt4, srcRegFilt6));
- srcRegFilt3 =
- _mm_adds_epi16(srcRegFilt3, _mm_min_epi16(srcRegFilt5, srcRegFilt7));
-
- // add and saturate the results together
- srcRegFilt1 =
- _mm_adds_epi16(srcRegFilt1, _mm_max_epi16(srcRegFilt4, srcRegFilt6));
- srcRegFilt3 =
- _mm_adds_epi16(srcRegFilt3, _mm_max_epi16(srcRegFilt5, srcRegFilt7));
-
- srcRegFilt1 =
- _mm_adds_epi16(srcRegFilt1, _mm256_castsi256_si128(addFilterReg64));
- srcRegFilt3 =
- _mm_adds_epi16(srcRegFilt3, _mm256_castsi256_si128(addFilterReg64));
-
- // shift by 7 bit each 16 bit
- srcRegFilt1 = _mm_srai_epi16(srcRegFilt1, 7);
- srcRegFilt3 = _mm_srai_epi16(srcRegFilt3, 7);
-
- // shrink to 8 bit each 16 bits, the first lane contain the first
- // convolve result and the second lane contain the second convolve
- // result
- outReg1 = _mm_packus_epi16(srcRegFilt1, srcRegFilt3);
+ s1[0] = _mm256_castsi128_si256(
+ _mm_unpacklo_epi8(_mm256_castsi256_si128(srcRegHead1), srcRegHead2));
+ s2[0] = _mm256_castsi128_si256(
+ _mm_unpackhi_epi8(_mm256_castsi256_si128(srcRegHead1), srcRegHead2));
+
+ outReg1 = convolve8_8_avx2(s1, f);
+ outReg2 = convolve8_8_avx2(s2, f);
+
+ // shrink to 8 bit each 16 bits, the low and high 64-bits of each lane
+ // contain the first and second convolve result respectively
+ outReg1 = _mm_packus_epi16(outReg1, outReg2);
// average if necessary
if (avg) {
@@ -568,14 +315,14 @@ static void vpx_filter_block1d16_v8_avx2(const uint8_t *src_ptr,
ptrdiff_t src_stride, uint8_t *dst_ptr,
ptrdiff_t dst_stride, uint32_t height,
const int16_t *filter) {
- vpx_filter_block1d16_v8_X_avx2(src_ptr, src_stride, dst_ptr, dst_stride,
+ vpx_filter_block1d16_v8_x_avx2(src_ptr, src_stride, dst_ptr, dst_stride,
height, filter, 0);
}
static void vpx_filter_block1d16_v8_avg_avx2(
const uint8_t *src_ptr, ptrdiff_t src_stride, uint8_t *dst_ptr,
ptrdiff_t dst_stride, uint32_t height, const int16_t *filter) {
- vpx_filter_block1d16_v8_X_avx2(src_ptr, src_stride, dst_ptr, dst_stride,
+ vpx_filter_block1d16_v8_x_avx2(src_ptr, src_stride, dst_ptr, dst_stride,
height, filter, 1);
}
generated by cgit v1.2.3 (git 2.25.1) at 2024-10-09 23:22:54 +0000