ref: 41b0888a84d12926c66cb7f235ed0e850e0ca99e
parent: 27e1bacdb3ab3317c0a1c8f62fec3f5b0824cf99
author: Johann <[email protected]>
date: Mon Dec 12 11:47:05 EST 2016
postproc: neon down and across macroblock filter Implement vpx_post_proc_down_and_across_mb_row in NEON. Runs about 6-7x faster than C. BUG=webm:1320 Change-Id: Ic5c7d3552a88cfcf999ec5bf2bd46fee460642c2
--- a/test/pp_filter_test.cc
+++ b/test/pp_filter_test.cc
@@ -120,6 +120,94 @@
vpx_free(flimits);
};
+TEST_P(VpxPostProcDownAndAcrossMbRowTest, CheckCvsAssembly) {+ // Size of the underlying data block that will be filtered.
+ // Y blocks are always a multiple of 16 wide and exactly 16 high. U and V
+ // blocks are always a multiple of 8 wide and exactly 8 high.
+ const int block_width = 136;
+ const int block_height = 16;
+
+ // 5-tap filter needs 2 padding rows above and below the block in the input.
+ const int input_width = block_width;
+ const int input_height = block_height + 4;
+ const int input_stride = input_width;
+ const int input_size = input_stride * input_height;
+
+ // Filter extends output block by 8 samples at left and right edges.
+ const int output_width = block_width + 16;
+ const int output_height = block_height;
+ const int output_stride = output_width;
+ const int output_size = output_stride * output_height;
+
+ uint8_t *const src_image = new uint8_t[input_size];
+ ASSERT_TRUE(src_image != NULL);
+
+ // Though the left padding is only 8 bytes, the assembly code tries to
+ // read 16 bytes before the pointer.
+ uint8_t *const dst_image = new uint8_t[output_size + 8];
+ ASSERT_TRUE(dst_image != NULL);
+ uint8_t *const dst_image_ref = new uint8_t[output_size + 8];
+ ASSERT_TRUE(dst_image_ref != NULL);
+
+ // Pointers to top-left pixel of block in the input and output images.
+ uint8_t *const src_image_ptr = src_image + (input_stride << 1);
+
+ // The assembly works in increments of 16. The first read may be offset by
+ // this amount.
+ uint8_t *const dst_image_ptr = dst_image + 16;
+ uint8_t *const dst_image_ref_ptr = dst_image + 16;
+
+ // Filter values are set in blocks of 16 for Y and 8 for U/V. Each macroblock
+ // can have a different filter.
+ uint8_t *const flimits =
+ reinterpret_cast<uint8_t *>(vpx_memalign(16, block_width));
+
+ ACMRandom rnd;
+ rnd.Reset(ACMRandom::DeterministicSeed());
+ // Initialize pixels in the input:
+ // block pixels to random values.
+ // border pixels to value 10.
+ (void)memset(src_image, 10, input_size);
+ uint8_t *pixel_ptr = src_image_ptr;
+ for (int i = 0; i < block_height; ++i) {+ for (int j = 0; j < block_width; ++j) {+ pixel_ptr[j] = rnd.Rand8();
+ }
+ pixel_ptr += input_stride;
+ }
+
+ for (int blocks = 0; blocks < block_width; blocks += 8) {+ (void)memset(flimits, 0, sizeof(*flimits) * block_width);
+
+ for (int f = 0; f < 255; f++) {+ (void)memset(flimits + blocks, f, sizeof(*flimits) * 8);
+
+ (void)memset(dst_image, 0, output_size);
+ (void)memset(dst_image_ref, 0, output_size);
+
+ vpx_post_proc_down_and_across_mb_row_c(
+ src_image_ptr, dst_image_ref_ptr, input_stride, output_stride,
+ block_width, flimits, block_height);
+ ASM_REGISTER_STATE_CHECK(GetParam()(src_image_ptr, dst_image_ptr,
+ input_stride, output_stride,
+ block_width, flimits, 16));
+
+ for (int i = 0; i < block_height; ++i) {+ for (int j = 0; j < block_width; ++j) {+ ASSERT_EQ(dst_image_ref_ptr[j + i * output_stride],
+ dst_image_ptr[j + i * output_stride])
+ << "at (" << i << ", " << j << ")";+ }
+ }
+ }
+ }
+
+ delete[] src_image;
+ delete[] dst_image;
+ delete[] dst_image_ref;
+ vpx_free(flimits);
+}
+
class VpxMbPostProcAcrossIpTest
: public ::testing::TestWithParam<VpxMbPostProcAcrossIpFunc> {public:
@@ -497,8 +585,14 @@
INSTANTIATE_TEST_CASE_P(SSE2, VpxMbPostProcDownTest,
::testing::Values(vpx_mbpost_proc_down_sse2));
-#endif
+#endif // HAVE_SSE2
+#if HAVE_NEON
+INSTANTIATE_TEST_CASE_P(
+ NEON, VpxPostProcDownAndAcrossMbRowTest,
+ ::testing::Values(vpx_post_proc_down_and_across_mb_row_neon));
+#endif // HAVE_NEON
+
#if HAVE_MSA
INSTANTIATE_TEST_CASE_P(
MSA, VpxPostProcDownAndAcrossMbRowTest,
@@ -509,6 +603,6 @@
INSTANTIATE_TEST_CASE_P(MSA, VpxMbPostProcDownTest,
::testing::Values(vpx_mbpost_proc_down_msa));
-#endif
+#endif // HAVE_MSA
} // namespace
--- /dev/null
+++ b/vpx_dsp/arm/deblock_neon.c
@@ -1,0 +1,259 @@
+/*
+ * Copyright (c) 2016 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 <arm_neon.h>
+#include <assert.h>
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx/vpx_integer.h"
+#include "vpx_dsp/arm/transpose_neon.h"
+
+static uint8x8_t average_k_out(const uint8x8_t a2, const uint8x8_t a1,
+ const uint8x8_t v0, const uint8x8_t b1,
+ const uint8x8_t b2) {+ const uint8x8_t k1 = vrhadd_u8(a2, a1);
+ const uint8x8_t k2 = vrhadd_u8(b2, b1);
+ const uint8x8_t k3 = vrhadd_u8(k1, k2);
+ return vrhadd_u8(k3, v0);
+}
+
+static uint8x8_t generate_mask(const uint8x8_t a2, const uint8x8_t a1,
+ const uint8x8_t v0, const uint8x8_t b1,
+ const uint8x8_t b2, const uint8x8_t filter) {+ const uint8x8_t a2_v0 = vabd_u8(a2, v0);
+ const uint8x8_t a1_v0 = vabd_u8(a1, v0);
+ const uint8x8_t b1_v0 = vabd_u8(b1, v0);
+ const uint8x8_t b2_v0 = vabd_u8(b2, v0);
+
+ uint8x8_t max = vmax_u8(a2_v0, a1_v0);
+ max = vmax_u8(b1_v0, max);
+ max = vmax_u8(b2_v0, max);
+ return vclt_u8(max, filter);
+}
+
+static uint8x8_t generate_output(const uint8x8_t a2, const uint8x8_t a1,
+ const uint8x8_t v0, const uint8x8_t b1,
+ const uint8x8_t b2, const uint8x8_t filter) {+ const uint8x8_t k_out = average_k_out(a2, a1, v0, b1, b2);
+ const uint8x8_t mask = generate_mask(a2, a1, v0, b1, b2, filter);
+
+ return vbsl_u8(mask, k_out, v0);
+}
+
+// Same functions but for uint8x16_t.
+static uint8x16_t average_k_outq(const uint8x16_t a2, const uint8x16_t a1,
+ const uint8x16_t v0, const uint8x16_t b1,
+ const uint8x16_t b2) {+ const uint8x16_t k1 = vrhaddq_u8(a2, a1);
+ const uint8x16_t k2 = vrhaddq_u8(b2, b1);
+ const uint8x16_t k3 = vrhaddq_u8(k1, k2);
+ return vrhaddq_u8(k3, v0);
+}
+
+static uint8x16_t generate_maskq(const uint8x16_t a2, const uint8x16_t a1,
+ const uint8x16_t v0, const uint8x16_t b1,
+ const uint8x16_t b2, const uint8x16_t filter) {+ const uint8x16_t a2_v0 = vabdq_u8(a2, v0);
+ const uint8x16_t a1_v0 = vabdq_u8(a1, v0);
+ const uint8x16_t b1_v0 = vabdq_u8(b1, v0);
+ const uint8x16_t b2_v0 = vabdq_u8(b2, v0);
+
+ uint8x16_t max = vmaxq_u8(a2_v0, a1_v0);
+ max = vmaxq_u8(b1_v0, max);
+ max = vmaxq_u8(b2_v0, max);
+ return vcltq_u8(max, filter);
+}
+
+static uint8x16_t generate_outputq(const uint8x16_t a2, const uint8x16_t a1,
+ const uint8x16_t v0, const uint8x16_t b1,
+ const uint8x16_t b2,
+ const uint8x16_t filter) {+ const uint8x16_t k_out = average_k_outq(a2, a1, v0, b1, b2);
+ const uint8x16_t mask = generate_maskq(a2, a1, v0, b1, b2, filter);
+
+ return vbslq_u8(mask, k_out, v0);
+}
+
+void vpx_post_proc_down_and_across_mb_row_neon(uint8_t *src_ptr,
+ uint8_t *dst_ptr, int src_stride,
+ int dst_stride, int cols,
+ uint8_t *f, int size) {+ uint8_t *src, *dst;
+ int row;
+ int col;
+
+ // Process a stripe of macroblocks. The stripe will be a multiple of 16 (for
+ // Y) or 8 (for U/V) wide (cols) and the height (size) will be 16 (for Y) or 8
+ // (for U/V).
+ assert((size == 8 || size == 16) && cols % 8 == 0);
+
+ // While columns of length 16 can be processed, load them.
+ for (col = 0; col < cols - 8; col += 16) {+ uint8x16_t a0, a1, a2, a3, a4, a5, a6, a7;
+ src = src_ptr - 2 * src_stride;
+ dst = dst_ptr;
+
+ a0 = vld1q_u8(src);
+ src += src_stride;
+ a1 = vld1q_u8(src);
+ src += src_stride;
+ a2 = vld1q_u8(src);
+ src += src_stride;
+ a3 = vld1q_u8(src);
+ src += src_stride;
+
+ for (row = 0; row < size; row += 4) {+ uint8x16_t v_out_0, v_out_1, v_out_2, v_out_3;
+ const uint8x16_t filterq = vld1q_u8(f + col);
+
+ a4 = vld1q_u8(src);
+ src += src_stride;
+ a5 = vld1q_u8(src);
+ src += src_stride;
+ a6 = vld1q_u8(src);
+ src += src_stride;
+ a7 = vld1q_u8(src);
+ src += src_stride;
+
+ v_out_0 = generate_outputq(a0, a1, a2, a3, a4, filterq);
+ v_out_1 = generate_outputq(a1, a2, a3, a4, a5, filterq);
+ v_out_2 = generate_outputq(a2, a3, a4, a5, a6, filterq);
+ v_out_3 = generate_outputq(a3, a4, a5, a6, a7, filterq);
+
+ vst1q_u8(dst, v_out_0);
+ dst += dst_stride;
+ vst1q_u8(dst, v_out_1);
+ dst += dst_stride;
+ vst1q_u8(dst, v_out_2);
+ dst += dst_stride;
+ vst1q_u8(dst, v_out_3);
+ dst += dst_stride;
+
+ // Rotate over to the next slot.
+ a0 = a4;
+ a1 = a5;
+ a2 = a6;
+ a3 = a7;
+ }
+
+ src_ptr += 16;
+ dst_ptr += 16;
+ }
+
+ // Clean up any left over column of length 8.
+ if (col != cols) {+ uint8x8_t a0, a1, a2, a3, a4, a5, a6, a7;
+ src = src_ptr - 2 * src_stride;
+ dst = dst_ptr;
+
+ a0 = vld1_u8(src);
+ src += src_stride;
+ a1 = vld1_u8(src);
+ src += src_stride;
+ a2 = vld1_u8(src);
+ src += src_stride;
+ a3 = vld1_u8(src);
+ src += src_stride;
+
+ for (row = 0; row < size; row += 4) {+ uint8x8_t v_out_0, v_out_1, v_out_2, v_out_3;
+ const uint8x8_t filter = vld1_u8(f + col);
+
+ a4 = vld1_u8(src);
+ src += src_stride;
+ a5 = vld1_u8(src);
+ src += src_stride;
+ a6 = vld1_u8(src);
+ src += src_stride;
+ a7 = vld1_u8(src);
+ src += src_stride;
+
+ v_out_0 = generate_output(a0, a1, a2, a3, a4, filter);
+ v_out_1 = generate_output(a1, a2, a3, a4, a5, filter);
+ v_out_2 = generate_output(a2, a3, a4, a5, a6, filter);
+ v_out_3 = generate_output(a3, a4, a5, a6, a7, filter);
+
+ vst1_u8(dst, v_out_0);
+ dst += dst_stride;
+ vst1_u8(dst, v_out_1);
+ dst += dst_stride;
+ vst1_u8(dst, v_out_2);
+ dst += dst_stride;
+ vst1_u8(dst, v_out_3);
+ dst += dst_stride;
+
+ // Rotate over to the next slot.
+ a0 = a4;
+ a1 = a5;
+ a2 = a6;
+ a3 = a7;
+ }
+
+ // Not strictly necessary but makes resetting dst_ptr easier.
+ dst_ptr += 8;
+ }
+
+ dst_ptr -= cols;
+
+ for (row = 0; row < size; row += 8) {+ uint8x8_t a0, a1, a2, a3;
+ uint8x8_t b0, b1, b2, b3, b4, b5, b6, b7;
+
+ src = dst_ptr;
+ dst = dst_ptr;
+
+ // Load 8 values, transpose 4 of them, and discard 2 because they will be
+ // reloaded later.
+ load_and_transpose_u8_4x8(src, dst_stride, &a0, &a1, &a2, &a3);
+ a3 = a1;
+ a2 = a1 = a0; // Extend left border.
+
+ src += 2;
+
+ for (col = 0; col < cols; col += 8) {+ uint8x8_t v_out_0, v_out_1, v_out_2, v_out_3, v_out_4, v_out_5, v_out_6,
+ v_out_7;
+ // Although the filter is meant to be applied vertically and is instead
+ // being applied horizontally here it's OK because it's set in blocks of 8
+ // (or 16).
+ const uint8x8_t filter = vld1_u8(f + col);
+
+ load_and_transpose_u8_8x8(src, dst_stride, &b0, &b1, &b2, &b3, &b4, &b5,
+ &b6, &b7);
+
+ if (col + 8 == cols) {+ // Last row. Extend border (b5).
+ b6 = b7 = b5;
+ }
+
+ v_out_0 = generate_output(a0, a1, a2, a3, b0, filter);
+ v_out_1 = generate_output(a1, a2, a3, b0, b1, filter);
+ v_out_2 = generate_output(a2, a3, b0, b1, b2, filter);
+ v_out_3 = generate_output(a3, b0, b1, b2, b3, filter);
+ v_out_4 = generate_output(b0, b1, b2, b3, b4, filter);
+ v_out_5 = generate_output(b1, b2, b3, b4, b5, filter);
+ v_out_6 = generate_output(b2, b3, b4, b5, b6, filter);
+ v_out_7 = generate_output(b3, b4, b5, b6, b7, filter);
+
+ transpose_and_store_u8_8x8(dst, dst_stride, v_out_0, v_out_1, v_out_2,
+ v_out_3, v_out_4, v_out_5, v_out_6, v_out_7);
+
+ a0 = b4;
+ a1 = b5;
+ a2 = b6;
+ a3 = b7;
+
+ src += 8;
+ dst += 8;
+ }
+
+ dst_ptr += 8 * dst_stride;
+ }
+}
--- a/vpx_dsp/arm/transpose_neon.h
+++ b/vpx_dsp/arm/transpose_neon.h
@@ -179,6 +179,62 @@
*a1 = d0.val[1];
}
+static INLINE void transpose_u8_4x8(uint8x8_t *a0, uint8x8_t *a1, uint8x8_t *a2,
+ uint8x8_t *a3, const uint8x8_t a4,
+ const uint8x8_t a5, const uint8x8_t a6,
+ const uint8x8_t a7) {+ // Swap 32 bit elements. Goes from:
+ // a0: 00 01 02 03 XX XX XX XX
+ // a1: 10 11 12 13 XX XX XX XX
+ // a2: 20 21 22 23 XX XX XX XX
+ // a3; 30 31 32 33 XX XX XX XX
+ // a4: 40 41 42 43 XX XX XX XX
+ // a5: 50 51 52 53 XX XX XX XX
+ // a6: 60 61 62 63 XX XX XX XX
+ // a7: 70 71 72 73 XX XX XX XX
+ // to:
+ // b0.val[0]: 00 01 02 03 40 41 42 43
+ // b1.val[0]: 10 11 12 13 50 51 52 53
+ // b2.val[0]: 20 21 22 23 60 61 62 63
+ // b3.val[0]: 30 31 32 33 70 71 72 73
+
+ const uint32x2x2_t b0 =
+ vtrn_u32(vreinterpret_u32_u8(*a0), vreinterpret_u32_u8(a4));
+ const uint32x2x2_t b1 =
+ vtrn_u32(vreinterpret_u32_u8(*a1), vreinterpret_u32_u8(a5));
+ const uint32x2x2_t b2 =
+ vtrn_u32(vreinterpret_u32_u8(*a2), vreinterpret_u32_u8(a6));
+ const uint32x2x2_t b3 =
+ vtrn_u32(vreinterpret_u32_u8(*a3), vreinterpret_u32_u8(a7));
+
+ // Swap 16 bit elements resulting in:
+ // c0.val[0]: 00 01 20 21 40 41 60 61
+ // c0.val[1]: 02 03 22 23 42 43 62 63
+ // c1.val[0]: 10 11 30 31 50 51 70 71
+ // c1.val[1]: 12 13 32 33 52 53 72 73
+
+ const uint16x4x2_t c0 = vtrn_u16(vreinterpret_u16_u32(b0.val[0]),
+ vreinterpret_u16_u32(b2.val[0]));
+ const uint16x4x2_t c1 = vtrn_u16(vreinterpret_u16_u32(b1.val[0]),
+ vreinterpret_u16_u32(b3.val[0]));
+
+ // Swap 8 bit elements resulting in:
+ // d0.val[0]: 00 10 20 30 40 50 60 70
+ // d0.val[1]: 01 11 21 31 41 51 61 71
+ // d1.val[0]: 02 12 22 32 42 52 62 72
+ // d1.val[1]: 03 13 23 33 43 53 63 73
+
+ const uint8x8x2_t d0 =
+ vtrn_u8(vreinterpret_u8_u16(c0.val[0]), vreinterpret_u8_u16(c1.val[0]));
+ const uint8x8x2_t d1 =
+ vtrn_u8(vreinterpret_u8_u16(c0.val[1]), vreinterpret_u8_u16(c1.val[1]));
+
+ *a0 = d0.val[0];
+ *a1 = d0.val[1];
+ *a2 = d1.val[0];
+ *a3 = d1.val[1];
+}
+
static INLINE void transpose_s32_4x4(int32x4_t *a0, int32x4_t *a1,
int32x4_t *a2, int32x4_t *a3) {// Swap 32 bit elements. Goes from:
@@ -936,11 +992,85 @@
*o15 = e7.val[1];
}
-static INLINE void load_and_transpose_s16_8x8(const int16_t *a, int a_stride,
- int16x8_t *a0, int16x8_t *a1,
- int16x8_t *a2, int16x8_t *a3,
- int16x8_t *a4, int16x8_t *a5,
- int16x8_t *a6, int16x8_t *a7) {+static INLINE void load_and_transpose_u8_4x8(const uint8_t *a,
+ const int a_stride, uint8x8_t *a0,
+ uint8x8_t *a1, uint8x8_t *a2,
+ uint8x8_t *a3) {+ uint8x8_t a4, a5, a6, a7;
+ *a0 = vld1_u8(a);
+ a += a_stride;
+ *a1 = vld1_u8(a);
+ a += a_stride;
+ *a2 = vld1_u8(a);
+ a += a_stride;
+ *a3 = vld1_u8(a);
+ a += a_stride;
+ a4 = vld1_u8(a);
+ a += a_stride;
+ a5 = vld1_u8(a);
+ a += a_stride;
+ a6 = vld1_u8(a);
+ a += a_stride;
+ a7 = vld1_u8(a);
+
+ transpose_u8_4x8(a0, a1, a2, a3, a4, a5, a6, a7);
+}
+
+static INLINE void load_and_transpose_u8_8x8(const uint8_t *a,
+ const int a_stride, uint8x8_t *a0,
+ uint8x8_t *a1, uint8x8_t *a2,
+ uint8x8_t *a3, uint8x8_t *a4,
+ uint8x8_t *a5, uint8x8_t *a6,
+ uint8x8_t *a7) {+ *a0 = vld1_u8(a);
+ a += a_stride;
+ *a1 = vld1_u8(a);
+ a += a_stride;
+ *a2 = vld1_u8(a);
+ a += a_stride;
+ *a3 = vld1_u8(a);
+ a += a_stride;
+ *a4 = vld1_u8(a);
+ a += a_stride;
+ *a5 = vld1_u8(a);
+ a += a_stride;
+ *a6 = vld1_u8(a);
+ a += a_stride;
+ *a7 = vld1_u8(a);
+
+ transpose_u8_8x8(a0, a1, a2, a3, a4, a5, a6, a7);
+}
+
+static INLINE void transpose_and_store_u8_8x8(uint8_t *a, const int a_stride,
+ uint8x8_t a0, uint8x8_t a1,
+ uint8x8_t a2, uint8x8_t a3,
+ uint8x8_t a4, uint8x8_t a5,
+ uint8x8_t a6, uint8x8_t a7) {+ transpose_u8_8x8(&a0, &a1, &a2, &a3, &a4, &a5, &a6, &a7);
+
+ vst1_u8(a, a0);
+ a += a_stride;
+ vst1_u8(a, a1);
+ a += a_stride;
+ vst1_u8(a, a2);
+ a += a_stride;
+ vst1_u8(a, a3);
+ a += a_stride;
+ vst1_u8(a, a4);
+ a += a_stride;
+ vst1_u8(a, a5);
+ a += a_stride;
+ vst1_u8(a, a6);
+ a += a_stride;
+ vst1_u8(a, a7);
+}
+
+static INLINE void load_and_transpose_s16_8x8(const int16_t *a,
+ const int a_stride, int16x8_t *a0,
+ int16x8_t *a1, int16x8_t *a2,
+ int16x8_t *a3, int16x8_t *a4,
+ int16x8_t *a5, int16x8_t *a6,
+ int16x8_t *a7) {*a0 = vld1q_s16(a);
a += a_stride;
*a1 = vld1q_s16(a);
--- a/vpx_dsp/vpx_dsp.mk
+++ b/vpx_dsp/vpx_dsp.mk
@@ -57,6 +57,7 @@
DSP_SRCS-yes += postproc.h
DSP_SRCS-$(HAVE_MSA) += mips/add_noise_msa.c
DSP_SRCS-$(HAVE_MSA) += mips/deblock_msa.c
+DSP_SRCS-$(HAVE_NEON) += arm/deblock_neon.c
DSP_SRCS-$(HAVE_SSE2) += x86/add_noise_sse2.asm
DSP_SRCS-$(HAVE_SSE2) += x86/deblock_sse2.asm
endif # CONFIG_POSTPROC
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -1756,7 +1756,7 @@
specialize qw/vpx_mbpost_proc_across_ip sse2 msa/;
add_proto qw/void vpx_post_proc_down_and_across_mb_row/, "unsigned char *src, unsigned char *dst, int src_pitch, int dst_pitch, int cols, unsigned char *flimits, int size";
- specialize qw/vpx_post_proc_down_and_across_mb_row sse2 msa/;
+ specialize qw/vpx_post_proc_down_and_across_mb_row sse2 neon msa/;
}