SSE2 optimizations for _6/_16 lowbd lpf functions
Includes vertical and horizontal implementations
and to fix 5/13 TAPs/Parallel deblocking support.
Re-working internals of the filters for better
re-usage across different sizes.
Tests are enabled.
Performance changes, SSE2 over C:
Horizontal methods: up to 3-4x
Vertical methods: up to 1.5x-2x
Change-Id: I2e36035355d8c23c1d4b0d59d0e23f598e9d0e3f
diff --git a/aom_dsp/aom_dsp_rtcd_defs.pl b/aom_dsp/aom_dsp_rtcd_defs.pl
index e717f41..e1b9d52 100755
--- a/aom_dsp/aom_dsp_rtcd_defs.pl
+++ b/aom_dsp/aom_dsp_rtcd_defs.pl
@@ -444,6 +444,11 @@
$aom_lpf_vertical_16_dual_neon_asm=aom_lpf_vertical_16_dual_neon;
}
+add_proto qw/void aom_lpf_vertical_6/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh";
+if (aom_config("CONFIG_PARALLEL_DEBLOCKING") eq "yes") {
+ specialize qw/aom_lpf_vertical_6 sse2/;
+}
+
add_proto qw/void aom_lpf_vertical_8/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh";
if (aom_config("CONFIG_PARALLEL_DEBLOCKING") eq "yes") {
specialize qw/aom_lpf_vertical_8 sse2/;
@@ -485,6 +490,11 @@
$aom_lpf_horizontal_16_dual_neon_asm=aom_lpf_horizontal_16_dual_neon;
}
+add_proto qw/void aom_lpf_horizontal_6/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh";
+if (aom_config("CONFIG_PARALLEL_DEBLOCKING") eq "yes") {
+ specialize qw/aom_lpf_horizontal_6 sse2/;
+}
+
add_proto qw/void aom_lpf_horizontal_8/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh";
if (aom_config("CONFIG_PARALLEL_DEBLOCKING") eq "yes") {
specialize qw/aom_lpf_horizontal_8 sse2/;
diff --git a/aom_dsp/x86/loopfilter_sse2.c b/aom_dsp/x86/loopfilter_sse2.c
index e4da2d9..e604fcc 100644
--- a/aom_dsp/x86/loopfilter_sse2.c
+++ b/aom_dsp/x86/loopfilter_sse2.c
@@ -92,52 +92,55 @@
mask = _mm_and_si128(mask, _mm_srli_si128(mask, 8)); \
} while (0)
-#define FILTER4 \
- do { \
- const __m128i t3t4 = \
- _mm_set_epi8(3, 3, 3, 3, 3, 3, 3, 3, 4, 4, 4, 4, 4, 4, 4, 4); \
- const __m128i t80 = _mm_set1_epi8(0x80); \
- __m128i filter, filter2filter1, work; \
- \
- ps1ps0 = _mm_xor_si128(p1p0, t80); /* ^ 0x80 */ \
- qs1qs0 = _mm_xor_si128(q1q0, t80); \
- \
- /* int8_t filter = signed_char_clamp(ps1 - qs1) & hev; */ \
- work = _mm_subs_epi8(ps1ps0, qs1qs0); \
- filter = _mm_and_si128(_mm_srli_si128(work, 8), hev); \
- /* filter = signed_char_clamp(filter + 3 * (qs0 - ps0)) & mask; */ \
- filter = _mm_subs_epi8(filter, work); \
- filter = _mm_subs_epi8(filter, work); \
- filter = _mm_subs_epi8(filter, work); /* + 3 * (qs0 - ps0) */ \
- filter = _mm_and_si128(filter, mask); /* & mask */ \
- filter = _mm_unpacklo_epi64(filter, filter); \
- \
- /* filter1 = signed_char_clamp(filter + 4) >> 3; */ \
- /* filter2 = signed_char_clamp(filter + 3) >> 3; */ \
- filter2filter1 = _mm_adds_epi8(filter, t3t4); /* signed_char_clamp */ \
- filter = _mm_unpackhi_epi8(filter2filter1, filter2filter1); \
- filter2filter1 = _mm_unpacklo_epi8(filter2filter1, filter2filter1); \
- filter2filter1 = _mm_srai_epi16(filter2filter1, 11); /* >> 3 */ \
- filter = _mm_srai_epi16(filter, 11); /* >> 3 */ \
- filter2filter1 = _mm_packs_epi16(filter2filter1, filter); \
- \
- /* filter = ROUND_POWER_OF_TWO(filter1, 1) & ~hev; */ \
- filter = _mm_subs_epi8(filter2filter1, ff); /* + 1 */ \
- filter = _mm_unpacklo_epi8(filter, filter); \
- filter = _mm_srai_epi16(filter, 9); /* round */ \
- filter = _mm_packs_epi16(filter, filter); \
- filter = _mm_andnot_si128(hev, filter); \
- \
- hev = _mm_unpackhi_epi64(filter2filter1, filter); \
- filter2filter1 = _mm_unpacklo_epi64(filter2filter1, filter); \
- \
- /* signed_char_clamp(qs1 - filter), signed_char_clamp(qs0 - filter1) */ \
- qs1qs0 = _mm_subs_epi8(qs1qs0, filter2filter1); \
- /* signed_char_clamp(ps1 + filter), signed_char_clamp(ps0 + filter2) */ \
- ps1ps0 = _mm_adds_epi8(ps1ps0, hev); \
- qs1qs0 = _mm_xor_si128(qs1qs0, t80); /* ^ 0x80 */ \
- ps1ps0 = _mm_xor_si128(ps1ps0, t80); /* ^ 0x80 */ \
- } while (0)
+AOM_FORCE_INLINE void filter4_sse2(__m128i *p1p0, __m128i *q1q0, __m128i *hev,
+ __m128i *mask, __m128i *qs1qs0,
+ __m128i *ps1ps0) {
+ const __m128i t3t4 =
+ _mm_set_epi8(3, 3, 3, 3, 3, 3, 3, 3, 4, 4, 4, 4, 4, 4, 4, 4);
+ const __m128i t80 = _mm_set1_epi8(0x80);
+ __m128i filter, filter2filter1, work;
+ __m128i ps1ps0_work, qs1qs0_work;
+ const __m128i ff = _mm_cmpeq_epi8(t80, t80);
+
+ ps1ps0_work = _mm_xor_si128(*p1p0, t80); /* ^ 0x80 */
+ qs1qs0_work = _mm_xor_si128(*q1q0, t80);
+
+ /* int8_t filter = signed_char_clamp(ps1 - qs1) & hev; */
+ work = _mm_subs_epi8(ps1ps0_work, qs1qs0_work);
+ filter = _mm_and_si128(_mm_srli_si128(work, 8), *hev);
+ /* filter = signed_char_clamp(filter + 3 * (qs0 - ps0)) & mask; */
+ filter = _mm_subs_epi8(filter, work);
+ filter = _mm_subs_epi8(filter, work);
+ filter = _mm_subs_epi8(filter, work); /* + 3 * (qs0 - ps0) */
+ filter = _mm_and_si128(filter, *mask); /* & mask */
+ filter = _mm_unpacklo_epi64(filter, filter);
+
+ /* filter1 = signed_char_clamp(filter + 4) >> 3; */
+ /* filter2 = signed_char_clamp(filter + 3) >> 3; */
+ filter2filter1 = _mm_adds_epi8(filter, t3t4); /* signed_char_clamp */
+ filter = _mm_unpackhi_epi8(filter2filter1, filter2filter1);
+ filter2filter1 = _mm_unpacklo_epi8(filter2filter1, filter2filter1);
+ filter2filter1 = _mm_srai_epi16(filter2filter1, 11); /* >> 3 */
+ filter = _mm_srai_epi16(filter, 11); /* >> 3 */
+ filter2filter1 = _mm_packs_epi16(filter2filter1, filter);
+
+ /* filter = ROUND_POWER_OF_TWO(filter1, 1) & ~hev; */
+ filter = _mm_subs_epi8(filter2filter1, ff); /* + 1 */
+ filter = _mm_unpacklo_epi8(filter, filter);
+ filter = _mm_srai_epi16(filter, 9); /* round */
+ filter = _mm_packs_epi16(filter, filter);
+ filter = _mm_andnot_si128(*hev, filter);
+
+ *hev = _mm_unpackhi_epi64(filter2filter1, filter);
+ filter2filter1 = _mm_unpacklo_epi64(filter2filter1, filter);
+
+ /* signed_char_clamp(qs1 - filter), signed_char_clamp(qs0 - filter1) */
+ qs1qs0_work = _mm_subs_epi8(qs1qs0_work, filter2filter1);
+ /* signed_char_clamp(ps1 + filter), signed_char_clamp(ps0 + filter2) */
+ ps1ps0_work = _mm_adds_epi8(ps1ps0_work, *hev);
+ *qs1qs0 = _mm_xor_si128(qs1qs0_work, t80); /* ^ 0x80 */
+ *ps1ps0 = _mm_xor_si128(ps1ps0_work, t80); /* ^ 0x80 */
+}
void aom_lpf_horizontal_4_sse2(uint8_t *s, int p /* pitch */,
const uint8_t *_blimit, const uint8_t *_limit,
@@ -148,7 +151,6 @@
_mm_loadl_epi64((const __m128i *)_limit));
const __m128i thresh =
_mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)_thresh), zero);
- const __m128i ff = _mm_cmpeq_epi8(zero, zero);
#if !CONFIG_PARALLEL_DEBLOCKING
__m128i p3p2, p2p1, q3q2, q2q1;
#endif // !CONFIG_PARALLEL_DEBLOCKING
@@ -177,7 +179,7 @@
#else // CONFIG_PARALLEL_DEBLOCKING
FILTER_HEV_MASK4;
#endif // !CONFIG_PARALLEL_DEBLOCKING
- FILTER4;
+ filter4_sse2(&p1p0, &q1q0, &hev, &mask, &qs1qs0, &ps1ps0);
#if CONFIG_PARALLEL_DEBLOCKING
xx_storel_32(s - 1 * p, ps1ps0);
@@ -201,7 +203,7 @@
_mm_loadl_epi64((const __m128i *)_limit));
const __m128i thresh =
_mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)_thresh), zero);
- const __m128i ff = _mm_cmpeq_epi8(zero, zero);
+
__m128i x0, x1, x2, x3;
#if !CONFIG_PARALLEL_DEBLOCKING
__m128i p3p2, p2p1, q3q2, q2q1;
@@ -264,7 +266,7 @@
#else // CONFIG_PARALLEL_DEBLOCKING
FILTER_HEV_MASK4;
#endif // !CONFIG_PARALLEL_DEBLOCKING
- FILTER4;
+ filter4_sse2(&p1p0, &q1q0, &hev, &mask, &qs1qs0, &ps1ps0);
// Transpose 8x4 to 4x8
// qs1qs0: 20 21 22 23 24 25 26 27 30 31 32 33 34 34 36 37
@@ -314,7 +316,10 @@
const __m128i limit = _mm_load_si128((const __m128i *)_limit);
const __m128i thresh = _mm_load_si128((const __m128i *)_thresh);
__m128i mask, hev, flat, flat2;
- __m128i q7p7, q6p6, q5p5, q4p4, q3p3, q2p2, q1p1, q0p0, p0q0, p1q1;
+#if !CONFIG_DEBLOCK_13TAP
+ __m128i q7p7;
+#endif
+ __m128i q6p6, q5p5, q4p4, q3p3, q2p2, q1p1, q0p0, p0q0, p1q1;
__m128i abs_p1p0;
q4p4 = _mm_loadl_epi64((__m128i *)(s - 5 * p));
@@ -376,7 +381,10 @@
__m128i filt;
__m128i work_a;
__m128i filter1, filter2;
- __m128i flat2_q6p6, flat2_q5p5, flat2_q4p4, flat2_q3p3, flat2_q2p2;
+#if !CONFIG_DEBLOCK_13TAP
+ __m128i flat2_q6p6;
+#endif
+ __m128i flat2_q5p5, flat2_q4p4, flat2_q3p3, flat2_q2p2;
__m128i flat2_q1p1, flat2_q0p0, flat_q2p2, flat_q1p1, flat_q0p0;
filt = _mm_and_si128(_mm_subs_epi8(qs1ps1, qs1), hev);
@@ -425,30 +433,42 @@
q6p6 = _mm_castps_si128(
_mm_loadh_pi(_mm_castsi128_ps(q6p6), (__m64 *)(s + 6 * p)));
flat2 = _mm_max_epu8(abs_diff(q4p4, q0p0), abs_diff(q5p5, q0p0));
-
+#if !CONFIG_DEBLOCK_13TAP
q7p7 = _mm_loadl_epi64((__m128i *)(s - 8 * p));
q7p7 = _mm_castps_si128(
_mm_loadh_pi(_mm_castsi128_ps(q7p7), (__m64 *)(s + 7 * p)));
work = _mm_max_epu8(abs_diff(q6p6, q0p0), abs_diff(q7p7, q0p0));
+#else
+ work = abs_diff(q6p6, q0p0);
+#endif
flat2 = _mm_max_epu8(work, flat2);
flat2 = _mm_max_epu8(flat2, _mm_srli_si128(flat2, 8));
flat2 = _mm_subs_epu8(flat2, one);
flat2 = _mm_cmpeq_epi8(flat2, zero);
flat2 = _mm_and_si128(flat2, flat); // flat2 & flat & mask
}
-
// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
// flat and wide flat calculations
{
const __m128i eight = _mm_set1_epi16(8);
const __m128i four = _mm_set1_epi16(4);
- __m128i p7_16, p6_16, p5_16, p4_16, p3_16, p2_16, p1_16, p0_16;
- __m128i q7_16, q6_16, q5_16, q4_16, q3_16, q2_16, q1_16, q0_16;
+#if !CONFIG_DEBLOCK_13TAP
+ __m128i p7_16, q7_16;
+#endif
+ __m128i p6_16, p5_16, p4_16, p3_16, p2_16, p1_16, p0_16;
+ __m128i q6_16, q5_16, q4_16, q3_16, q2_16, q1_16, q0_16;
__m128i pixelFilter_p, pixelFilter_q;
__m128i pixetFilter_p2p1p0, pixetFilter_q2q1q0;
- __m128i sum_p7, sum_q7, sum_p3, sum_q3, res_p, res_q;
+#if !CONFIG_DEBLOCK_13TAP
+ __m128i sum_p7, sum_q7;
+#else
+ __m128i sum_p6, sum_q6;
+#endif
+ __m128i sum_p3, sum_q3, res_p, res_q;
+#if !CONFIG_DEBLOCK_13TAP
p7_16 = _mm_unpacklo_epi8(q7p7, zero);
+#endif
p6_16 = _mm_unpacklo_epi8(q6p6, zero);
p5_16 = _mm_unpacklo_epi8(q5p5, zero);
p4_16 = _mm_unpacklo_epi8(q4p4, zero);
@@ -463,12 +483,17 @@
q4_16 = _mm_unpackhi_epi8(q4p4, zero);
q5_16 = _mm_unpackhi_epi8(q5p5, zero);
q6_16 = _mm_unpackhi_epi8(q6p6, zero);
+#if !CONFIG_DEBLOCK_13TAP
q7_16 = _mm_unpackhi_epi8(q7p7, zero);
pixelFilter_p = _mm_add_epi16(_mm_add_epi16(p6_16, p5_16),
_mm_add_epi16(p4_16, p3_16));
pixelFilter_q = _mm_add_epi16(_mm_add_epi16(q6_16, q5_16),
_mm_add_epi16(q4_16, q3_16));
+#else
+ pixelFilter_p = _mm_add_epi16(p5_16, _mm_add_epi16(p4_16, p3_16));
+ pixelFilter_q = _mm_add_epi16(q5_16, _mm_add_epi16(q4_16, q3_16));
+#endif
pixetFilter_p2p1p0 = _mm_add_epi16(p0_16, _mm_add_epi16(p2_16, p1_16));
pixelFilter_p = _mm_add_epi16(pixelFilter_p, pixetFilter_p2p1p0);
@@ -480,10 +505,25 @@
pixetFilter_p2p1p0 = _mm_add_epi16(
four, _mm_add_epi16(pixetFilter_p2p1p0, pixetFilter_q2q1q0));
res_p = _mm_srli_epi16(
+#if !CONFIG_DEBLOCK_13TAP
_mm_add_epi16(pixelFilter_p, _mm_add_epi16(p7_16, p0_16)), 4);
+#else
+ _mm_add_epi16(pixelFilter_p,
+ _mm_add_epi16(_mm_add_epi16(p6_16, p0_16),
+ _mm_add_epi16(p1_16, q0_16))),
+ 4);
+#endif
res_q = _mm_srli_epi16(
+#if !CONFIG_DEBLOCK_13TAP
_mm_add_epi16(pixelFilter_p, _mm_add_epi16(q7_16, q0_16)), 4);
+#else
+ _mm_add_epi16(pixelFilter_p,
+ _mm_add_epi16(_mm_add_epi16(q6_16, q0_16),
+ _mm_add_epi16(p0_16, q1_16))),
+ 4);
+#endif
flat2_q0p0 = _mm_packus_epi16(res_p, res_q);
+
res_p = _mm_srli_epi16(
_mm_add_epi16(pixetFilter_p2p1p0, _mm_add_epi16(p3_16, p0_16)), 3);
res_q = _mm_srli_epi16(
@@ -491,17 +531,44 @@
flat_q0p0 = _mm_packus_epi16(res_p, res_q);
+#if !CONFIG_DEBLOCK_13TAP
sum_p7 = _mm_add_epi16(p7_16, p7_16);
sum_q7 = _mm_add_epi16(q7_16, q7_16);
+#else
+ sum_p6 = _mm_add_epi16(p6_16, p6_16);
+ sum_q6 = _mm_add_epi16(q6_16, q6_16);
+#endif
sum_p3 = _mm_add_epi16(p3_16, p3_16);
sum_q3 = _mm_add_epi16(q3_16, q3_16);
+#if !CONFIG_DEBLOCK_13TAP
pixelFilter_q = _mm_sub_epi16(pixelFilter_p, p6_16);
pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q6_16);
+#else
+ pixelFilter_q = _mm_sub_epi16(pixelFilter_p, p5_16);
+ pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q5_16);
+#endif
+
res_p = _mm_srli_epi16(
+#if !CONFIG_DEBLOCK_13TAP
_mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p1_16)), 4);
+#else
+ _mm_add_epi16(
+ pixelFilter_p,
+ _mm_add_epi16(sum_p6,
+ _mm_add_epi16(p1_16, _mm_add_epi16(p2_16, p0_16)))),
+ 4);
+#endif
res_q = _mm_srli_epi16(
+#if !CONFIG_DEBLOCK_13TAP
_mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q1_16)), 4);
+#else
+ _mm_add_epi16(
+ pixelFilter_q,
+ _mm_add_epi16(sum_q6,
+ _mm_add_epi16(q1_16, _mm_add_epi16(q0_16, q2_16)))),
+ 4);
+#endif
flat2_q1p1 = _mm_packus_epi16(res_p, res_q);
pixetFilter_q2q1q0 = _mm_sub_epi16(pixetFilter_p2p1p0, p2_16);
@@ -512,17 +579,44 @@
_mm_add_epi16(pixetFilter_q2q1q0, _mm_add_epi16(sum_q3, q1_16)), 3);
flat_q1p1 = _mm_packus_epi16(res_p, res_q);
+#if !CONFIG_DEBLOCK_13TAP
sum_p7 = _mm_add_epi16(sum_p7, p7_16);
sum_q7 = _mm_add_epi16(sum_q7, q7_16);
+#else
+ sum_p6 = _mm_add_epi16(sum_p6, p6_16);
+ sum_q6 = _mm_add_epi16(sum_q6, q6_16);
+#endif
sum_p3 = _mm_add_epi16(sum_p3, p3_16);
sum_q3 = _mm_add_epi16(sum_q3, q3_16);
+#if !CONFIG_DEBLOCK_13TAP
pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q5_16);
pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p5_16);
+#else
+ pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q4_16);
+ pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p4_16);
+#endif
+
res_p = _mm_srli_epi16(
+#if !CONFIG_DEBLOCK_13TAP
_mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p2_16)), 4);
+#else
+ _mm_add_epi16(
+ pixelFilter_p,
+ _mm_add_epi16(sum_p6,
+ _mm_add_epi16(p2_16, _mm_add_epi16(p3_16, p1_16)))),
+ 4);
+#endif
res_q = _mm_srli_epi16(
+#if !CONFIG_DEBLOCK_13TAP
_mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q2_16)), 4);
+#else
+ _mm_add_epi16(
+ pixelFilter_q,
+ _mm_add_epi16(sum_q6,
+ _mm_add_epi16(q2_16, _mm_add_epi16(q1_16, q3_16)))),
+ 4);
+#endif
flat2_q2p2 = _mm_packus_epi16(res_p, res_q);
pixetFilter_p2p1p0 = _mm_sub_epi16(pixetFilter_p2p1p0, q1_16);
@@ -534,36 +628,121 @@
_mm_add_epi16(pixetFilter_q2q1q0, _mm_add_epi16(sum_q3, q2_16)), 3);
flat_q2p2 = _mm_packus_epi16(res_p, res_q);
+#if !CONFIG_DEBLOCK_13TAP
sum_p7 = _mm_add_epi16(sum_p7, p7_16);
sum_q7 = _mm_add_epi16(sum_q7, q7_16);
+#else
+ sum_p6 = _mm_add_epi16(sum_p6, p6_16);
+ sum_q6 = _mm_add_epi16(sum_q6, q6_16);
+#endif
+
+#if !CONFIG_DEBLOCK_13TAP
pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q4_16);
pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p4_16);
- res_p = _mm_srli_epi16(
- _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p3_16)), 4);
- res_q = _mm_srli_epi16(
- _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q3_16)), 4);
- flat2_q3p3 = _mm_packus_epi16(res_p, res_q);
-
- sum_p7 = _mm_add_epi16(sum_p7, p7_16);
- sum_q7 = _mm_add_epi16(sum_q7, q7_16);
+#else
pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q3_16);
pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p3_16);
- res_p = _mm_srli_epi16(
- _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p4_16)), 4);
- res_q = _mm_srli_epi16(
- _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q4_16)), 4);
- flat2_q4p4 = _mm_packus_epi16(res_p, res_q);
+#endif
+ res_p = _mm_srli_epi16(
+#if !CONFIG_DEBLOCK_13TAP
+ _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p3_16)), 4);
+#else
+ _mm_add_epi16(
+ pixelFilter_p,
+ _mm_add_epi16(sum_p6,
+ _mm_add_epi16(p3_16, _mm_add_epi16(p4_16, p2_16)))),
+ 4);
+#endif
+ res_q = _mm_srli_epi16(
+#if !CONFIG_DEBLOCK_13TAP
+ _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q3_16)), 4);
+#else
+ _mm_add_epi16(
+ pixelFilter_q,
+ _mm_add_epi16(sum_q6,
+ _mm_add_epi16(q3_16, _mm_add_epi16(q2_16, q4_16)))),
+ 4);
+#endif
+ flat2_q3p3 = _mm_packus_epi16(res_p, res_q);
+
+#if !CONFIG_DEBLOCK_13TAP
sum_p7 = _mm_add_epi16(sum_p7, p7_16);
sum_q7 = _mm_add_epi16(sum_q7, q7_16);
+#else
+ sum_p6 = _mm_add_epi16(sum_p6, p6_16);
+ sum_q6 = _mm_add_epi16(sum_q6, q6_16);
+#endif
+
+#if !CONFIG_DEBLOCK_13TAP
+ pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q3_16);
+ pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p3_16);
+#else
pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q2_16);
pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p2_16);
+#endif
+
res_p = _mm_srli_epi16(
- _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p5_16)), 4);
+#if !CONFIG_DEBLOCK_13TAP
+ _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p4_16)), 4);
+#else
+ _mm_add_epi16(
+ pixelFilter_p,
+ _mm_add_epi16(sum_p6,
+ _mm_add_epi16(p4_16, _mm_add_epi16(p5_16, p3_16)))),
+ 4);
+#endif
res_q = _mm_srli_epi16(
+#if !CONFIG_DEBLOCK_13TAP
+ _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q4_16)), 4);
+#else
+ _mm_add_epi16(
+ pixelFilter_q,
+ _mm_add_epi16(sum_q6,
+ _mm_add_epi16(q4_16, _mm_add_epi16(q3_16, q5_16)))),
+ 4);
+#endif
+ flat2_q4p4 = _mm_packus_epi16(res_p, res_q);
+
+#if !CONFIG_DEBLOCK_13TAP
+ sum_p7 = _mm_add_epi16(sum_p7, p7_16);
+ sum_q7 = _mm_add_epi16(sum_q7, q7_16);
+#else
+ sum_p6 = _mm_add_epi16(sum_p6, p6_16);
+ sum_q6 = _mm_add_epi16(sum_q6, q6_16);
+#endif
+
+#if !CONFIG_DEBLOCK_13TAP
+ pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q2_16);
+ pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p2_16);
+#else
+ pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q1_16);
+ pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p1_16);
+#endif
+
+ res_p = _mm_srli_epi16(
+#if !CONFIG_DEBLOCK_13TAP
+ _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p5_16)), 4);
+#else
+ _mm_add_epi16(
+ pixelFilter_p,
+ _mm_add_epi16(sum_p6,
+ _mm_add_epi16(p5_16, _mm_add_epi16(p6_16, p4_16)))),
+ 4);
+#endif
+ res_q = _mm_srli_epi16(
+#if !CONFIG_DEBLOCK_13TAP
_mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q5_16)), 4);
+#else
+ _mm_add_epi16(
+ pixelFilter_q,
+ _mm_add_epi16(sum_q6,
+ _mm_add_epi16(q5_16, _mm_add_epi16(q6_16, q4_16)))),
+ 4);
+#endif
flat2_q5p5 = _mm_packus_epi16(res_p, res_q);
+#if !CONFIG_DEBLOCK_13TAP
sum_p7 = _mm_add_epi16(sum_p7, p7_16);
sum_q7 = _mm_add_epi16(sum_q7, q7_16);
pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q1_16);
@@ -573,6 +752,7 @@
res_q = _mm_srli_epi16(
_mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q6_16)), 4);
flat2_q6p6 = _mm_packus_epi16(res_p, res_q);
+#endif
}
// wide flat
// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
@@ -592,10 +772,12 @@
flat_q0p0 = _mm_and_si128(flat, flat_q0p0);
q0p0 = _mm_or_si128(qs0ps0, flat_q0p0);
+#if !CONFIG_DEBLOCK_13TAP
q6p6 = _mm_andnot_si128(flat2, q6p6);
flat2_q6p6 = _mm_and_si128(flat2, flat2_q6p6);
q6p6 = _mm_or_si128(q6p6, flat2_q6p6);
store_buffer_horz_8(q6p6, p, 6, s);
+#endif
q5p5 = _mm_andnot_si128(flat2, q5p5);
flat2_q5p5 = _mm_and_si128(flat2, flat2_q5p5);
@@ -1003,23 +1185,185 @@
}
}
+void aom_lpf_horizontal_6_sse2(unsigned char *s, int p,
+ const unsigned char *_blimit,
+ const unsigned char *_limit,
+ const unsigned char *_thresh) {
+ {
+ const __m128i zero = _mm_setzero_si128();
+ const __m128i blimit = _mm_load_si128((const __m128i *)_blimit);
+ const __m128i limit = _mm_load_si128((const __m128i *)_limit);
+ const __m128i thresh = _mm_load_si128((const __m128i *)_thresh);
+ __m128i mask, hev, flat;
+ int flatmmask, maskmmask;
+ __m128i p2, p1, p0, q0, q1, q2;
+ __m128i q2p2, q1p1, q0p0, p1q1, p0q0, flat_p1p0, flat_q0q1;
+ __m128i p1p0, q1q0, ps1ps0, qs1qs0;
+
+ q2p2 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 3 * p)),
+ _mm_loadl_epi64((__m128i *)(s + 2 * p)));
+ q1p1 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 2 * p)),
+ _mm_loadl_epi64((__m128i *)(s + 1 * p)));
+ q0p0 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 1 * p)),
+ _mm_loadl_epi64((__m128i *)(s - 0 * p)));
+ p1q1 = _mm_shuffle_epi32(q1p1, 78);
+ p0q0 = _mm_shuffle_epi32(q0p0, 78);
+
+ const __m128i one = _mm_set1_epi8(1);
+ const __m128i fe = _mm_set1_epi8(0xfe);
+ const __m128i ff = _mm_cmpeq_epi8(fe, fe);
+ {
+ // filter_mask and hev_mask
+ __m128i abs_p1q1, abs_p0q0, abs_q1q0, abs_p1p0, work;
+ abs_p1p0 = abs_diff(q1p1, q0p0);
+ abs_q1q0 = _mm_srli_si128(abs_p1p0, 8);
+
+ abs_p0q0 = abs_diff(q0p0, p0q0);
+ abs_p1q1 = abs_diff(q1p1, p1q1);
+
+ // considering sse doesn't have unsigned elements comparison the idea is
+ // to find at least one case when X > limit, it means the corresponding
+ // mask bit is set.
+ // to achieve that we find global max value of all inputs of abs(x-y) or
+ // (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 If it is > limit the mask is set
+ // otherwise - not
+
+ flat = _mm_max_epu8(abs_p1p0, abs_q1q0);
+ hev = _mm_subs_epu8(flat, thresh);
+ hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
+ // replicate for the further "merged variables" usage
+ hev = _mm_unpacklo_epi64(hev, hev);
+
+ abs_p0q0 = _mm_adds_epu8(abs_p0q0, abs_p0q0);
+ abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1);
+ mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit);
+ mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff);
+ // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1;
+ mask = _mm_max_epu8(abs_p1p0, mask);
+ // mask |= (abs(p1 - p0) > limit) * -1;
+ // mask |= (abs(q1 - q0) > limit) * -1;
+
+ work = abs_diff(q2p2, q1p1);
+ mask = _mm_max_epu8(work, mask);
+ mask = _mm_max_epu8(mask, _mm_srli_si128(mask, 8));
+ mask = _mm_subs_epu8(mask, limit);
+ mask = _mm_cmpeq_epi8(mask, zero);
+ // replicate for the further "merged variables" usage
+ mask = _mm_unpacklo_epi64(mask, mask);
+ maskmmask = _mm_movemask_epi8(mask); // if 0 we don't need to do 5 tap
+ // filter, otherwise don't need to
+ // do filter4_sse2
+
+ // flat_mask
+ flat = _mm_max_epu8(abs_diff(q2p2, q0p0), abs_p1p0);
+ flat = _mm_max_epu8(flat, _mm_srli_si128(flat, 8));
+ flat = _mm_subs_epu8(flat, one);
+ flat = _mm_cmpeq_epi8(flat, zero);
+ flat = _mm_and_si128(flat, mask);
+ // replicate for the further "merged variables" usage
+ flat = _mm_unpacklo_epi64(flat, flat);
+ flatmmask = _mm_movemask_epi8(flat); // if 0 we don't need to do 5 tap
+ // filter, otherwise don't need to
+ // do filter4_sse2
+ }
+
+ // 5 tap filter
+ if (flatmmask & maskmmask) {
+ const __m128i four = _mm_set1_epi16(4);
+ unsigned char *src = s;
+
+ __m128i workp_a, workp_b, workp_shft0, workp_shft1;
+ p2 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 3 * p)), zero);
+ p1 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 2 * p)), zero);
+ p0 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 1 * p)), zero);
+ q0 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 0 * p)), zero);
+ q1 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 1 * p)), zero);
+ q2 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 2 * p)), zero);
+
+ // op1
+ workp_a = _mm_add_epi16(_mm_add_epi16(p0, p0),
+ _mm_add_epi16(p1, p1)); // p0 *2 + p1 * 2
+ workp_a = _mm_add_epi16(_mm_add_epi16(workp_a, four),
+ p2); // p2 + p0 * 2 + p1 * 2 + 4
+
+ workp_b = _mm_add_epi16(_mm_add_epi16(p2, p2), q0);
+ workp_shft0 = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b),
+ 3); // p2 * 3 + p1 * 2 + p0 * 2 + q0 + 4
+
+ // op0
+ workp_b = _mm_add_epi16(_mm_add_epi16(q0, q0), q1); // q0 * 2 + q1
+ workp_a = _mm_add_epi16(
+ workp_a, workp_b); // p2 + p0 * 2 + p1 * 2 + q0 * 2 + q1 + 4
+ workp_shft1 = _mm_srli_epi16(workp_a, 3);
+
+ flat_p1p0 = _mm_packus_epi16(workp_shft1, workp_shft0);
+
+ // oq0
+ workp_a = _mm_sub_epi16(_mm_sub_epi16(workp_a, p2),
+ p1); // p0 * 2 + p1 + q0 * 2 + q1 + 4
+ workp_b = _mm_add_epi16(q1, q2);
+ workp_a = _mm_add_epi16(
+ workp_a, workp_b); // p0 * 2 + p1 + q0 * 2 + q1 * 2 + q2 + 4
+ workp_shft0 = _mm_srli_epi16(workp_a, 3);
+
+ // oq1
+ workp_a = _mm_sub_epi16(_mm_sub_epi16(workp_a, p1),
+ p0); // p0 + q0 * 2 + q1 * 2 + q2 + 4
+ workp_b = _mm_add_epi16(q2, q2);
+ workp_shft1 = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b),
+ 3); // p0 + q0 * 2 + q1 * 2 + q2 * 3 + 4
+
+ flat_q0q1 = _mm_packus_epi16(workp_shft0, workp_shft1);
+ } else {
+ flat_p1p0 = _mm_setzero_si128();
+ flat_q0q1 = _mm_setzero_si128();
+ }
+ if ((flatmmask & maskmmask) != 0xffff) {
+ // lp filter
+ p1p0 = _mm_unpacklo_epi64(q0p0, q1p1);
+ q1q0 = _mm_unpackhi_epi64(q0p0, q1p1);
+
+ filter4_sse2(&p1p0, &q1q0, &hev, &mask, &qs1qs0, &ps1ps0);
+ } else {
+ ps1ps0 = _mm_setzero_si128();
+ qs1qs0 = _mm_setzero_si128();
+ }
+ qs1qs0 = _mm_andnot_si128(flat, qs1qs0);
+ q1q0 = _mm_and_si128(flat, flat_q0q1);
+ q1q0 = _mm_or_si128(qs1qs0, q1q0);
+
+ ps1ps0 = _mm_andnot_si128(flat, ps1ps0);
+ p1p0 = _mm_and_si128(flat, flat_p1p0);
+ p1p0 = _mm_or_si128(ps1ps0, p1p0);
+
+#if CONFIG_PARALLEL_DEBLOCKING
+ xx_storel_32(s - 1 * p, p1p0);
+ xx_storel_32(s - 2 * p, _mm_srli_si128(p1p0, 8));
+ xx_storel_32(s + 0 * p, q1q0);
+ xx_storel_32(s + 1 * p, _mm_srli_si128(q1q0, 8));
+#else
+ xx_storel_64(s - 1 * p, p1p0);
+ xx_storel_64(s - 2 * p, _mm_srli_si128(p1p0, 8));
+ xx_storel_64(s + 0 * p, q1q0);
+ xx_storel_64(s + 1 * p, _mm_srli_si128(q1q0, 8));
+#endif
+ }
+}
+
void aom_lpf_horizontal_8_sse2(unsigned char *s, int p,
const unsigned char *_blimit,
const unsigned char *_limit,
const unsigned char *_thresh) {
- DECLARE_ALIGNED(16, unsigned char, flat_op2[16]);
- DECLARE_ALIGNED(16, unsigned char, flat_op1[16]);
- DECLARE_ALIGNED(16, unsigned char, flat_op0[16]);
- DECLARE_ALIGNED(16, unsigned char, flat_oq2[16]);
- DECLARE_ALIGNED(16, unsigned char, flat_oq1[16]);
- DECLARE_ALIGNED(16, unsigned char, flat_oq0[16]);
- const __m128i zero = _mm_set1_epi16(0);
+ const __m128i zero = _mm_setzero_si128();
const __m128i blimit = _mm_load_si128((const __m128i *)_blimit);
const __m128i limit = _mm_load_si128((const __m128i *)_limit);
const __m128i thresh = _mm_load_si128((const __m128i *)_thresh);
__m128i mask, hev, flat;
- __m128i p3, p2, p1, p0, q0, q1, q2, q3;
- __m128i q3p3, q2p2, q1p1, q0p0, p1q1, p0q0;
+ int flatmmask, maskmmask;
+ __m128i p2, p1, p0, q0, q1, q2, p3, q3, q3p3, flat_p1p0, flat_q0q1;
+ __m128i q2p2, q1p1, q0p0, p1q1, p0q0;
+ __m128i p1p0, q1q0, ps1ps0, qs1qs0;
+ __m128i work_a, op2, oq2;
q3p3 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 4 * p)),
_mm_loadl_epi64((__m128i *)(s + 3 * p)));
@@ -1034,10 +1378,19 @@
{
// filter_mask and hev_mask
+
+ // considering sse doesn't have unsigned elements comparison the idea is to
+ // find at least one case when X > limit, it means the corresponding mask
+ // bit is set.
+ // to achieve that we find global max value of all inputs of abs(x-y) or
+ // (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 If it is > limit the mask is set
+ // otherwise - not
+
const __m128i one = _mm_set1_epi8(1);
const __m128i fe = _mm_set1_epi8(0xfe);
const __m128i ff = _mm_cmpeq_epi8(fe, fe);
__m128i abs_p1q1, abs_p0q0, abs_q1q0, abs_p1p0, work;
+
abs_p1p0 = abs_diff(q1p1, q0p0);
abs_q1q0 = _mm_srli_si128(abs_p1p0, 8);
@@ -1046,6 +1399,8 @@
flat = _mm_max_epu8(abs_p1p0, abs_q1q0);
hev = _mm_subs_epu8(flat, thresh);
hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
+ // replicate for the further "merged variables" usage
+ hev = _mm_unpacklo_epi64(hev, hev);
abs_p0q0 = _mm_adds_epu8(abs_p0q0, abs_p0q0);
abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1);
@@ -1057,171 +1412,135 @@
// mask |= (abs(q1 - q0) > limit) * -1;
work = _mm_max_epu8(abs_diff(q2p2, q1p1), abs_diff(q3p3, q2p2));
+
mask = _mm_max_epu8(work, mask);
mask = _mm_max_epu8(mask, _mm_srli_si128(mask, 8));
mask = _mm_subs_epu8(mask, limit);
mask = _mm_cmpeq_epi8(mask, zero);
+ // replicate for the further "merged variables" usage
+ mask = _mm_unpacklo_epi64(mask, mask);
+ maskmmask = _mm_movemask_epi8(mask); // if 0 we don't need to do 5 tap
+ // filter, otherwise don't need to do
+ // filter4_sse2
// flat_mask4
flat = _mm_max_epu8(abs_diff(q2p2, q0p0), abs_diff(q3p3, q0p0));
flat = _mm_max_epu8(abs_p1p0, flat);
+
flat = _mm_max_epu8(flat, _mm_srli_si128(flat, 8));
flat = _mm_subs_epu8(flat, one);
flat = _mm_cmpeq_epi8(flat, zero);
flat = _mm_and_si128(flat, mask);
+ // replicate for the further "merged variables" usage
+ flat = _mm_unpacklo_epi64(flat, flat);
+ flatmmask = _mm_movemask_epi8(flat); // if 0 we don't need to do 5 tap
+ // filter, otherwise don't need to do
+ // filter4_sse2
}
- {
+ // filter8
+ if (flatmmask & maskmmask) {
const __m128i four = _mm_set1_epi16(4);
unsigned char *src = s;
- {
- __m128i workp_a, workp_b, workp_shft;
- p3 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 4 * p)), zero);
- p2 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 3 * p)), zero);
- p1 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 2 * p)), zero);
- p0 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 1 * p)), zero);
- q0 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 0 * p)), zero);
- q1 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 1 * p)), zero);
- q2 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 2 * p)), zero);
- q3 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 3 * p)), zero);
- workp_a = _mm_add_epi16(_mm_add_epi16(p3, p3), _mm_add_epi16(p2, p1));
- workp_a = _mm_add_epi16(_mm_add_epi16(workp_a, four), p0);
- workp_b = _mm_add_epi16(_mm_add_epi16(q0, p2), p3);
- workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
- _mm_storel_epi64((__m128i *)&flat_op2[0],
- _mm_packus_epi16(workp_shft, workp_shft));
+ __m128i workp_a, workp_b, workp_shft0, workp_shft1;
+ p2 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 3 * p)), zero);
+ p1 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 2 * p)), zero);
+ p0 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 1 * p)), zero);
+ q0 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 0 * p)), zero);
+ q1 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 1 * p)), zero);
+ q2 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 2 * p)), zero);
+ p3 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 4 * p)), zero);
+ q3 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 3 * p)), zero);
- workp_b = _mm_add_epi16(_mm_add_epi16(q0, q1), p1);
- workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
- _mm_storel_epi64((__m128i *)&flat_op1[0],
- _mm_packus_epi16(workp_shft, workp_shft));
+ // op2
+ workp_a = _mm_add_epi16(_mm_add_epi16(p3, p3), _mm_add_epi16(p2, p1));
+ workp_a = _mm_add_epi16(_mm_add_epi16(workp_a, four), p0);
+ workp_b = _mm_add_epi16(_mm_add_epi16(q0, p2), p3);
+ workp_shft0 = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+ op2 = _mm_packus_epi16(workp_shft0, workp_shft0);
- workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p3), q2);
- workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p1), p0);
- workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
- _mm_storel_epi64((__m128i *)&flat_op0[0],
- _mm_packus_epi16(workp_shft, workp_shft));
+ // op1
+ workp_b = _mm_add_epi16(_mm_add_epi16(q0, q1), p1);
+ workp_shft0 = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
- workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p3), q3);
- workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p0), q0);
- workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
- _mm_storel_epi64((__m128i *)&flat_oq0[0],
- _mm_packus_epi16(workp_shft, workp_shft));
+ // op0
+ workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p3), q2);
+ workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p1), p0);
+ workp_shft1 = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
- workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p2), q3);
- workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q0), q1);
- workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
- _mm_storel_epi64((__m128i *)&flat_oq1[0],
- _mm_packus_epi16(workp_shft, workp_shft));
+ flat_p1p0 = _mm_packus_epi16(workp_shft1, workp_shft0);
- workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p1), q3);
- workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q1), q2);
- workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
- _mm_storel_epi64((__m128i *)&flat_oq2[0],
- _mm_packus_epi16(workp_shft, workp_shft));
- }
+ // oq0
+ workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p3), q3);
+ workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p0), q0);
+ workp_shft0 = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+
+ // oq1
+ workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p2), q3);
+ workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q0), q1);
+ workp_shft1 = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+
+ flat_q0q1 = _mm_packus_epi16(workp_shft0, workp_shft1);
+
+ // oq2
+ workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p1), q3);
+ workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q1), q2);
+ workp_shft1 = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+ oq2 = _mm_packus_epi16(workp_shft1, workp_shft1);
+
+ } else {
+ flat_p1p0 = _mm_setzero_si128();
+ flat_q0q1 = _mm_setzero_si128();
+ op2 = _mm_setzero_si128();
+ oq2 = _mm_setzero_si128();
}
+
// lp filter
- {
- const __m128i t4 = _mm_set1_epi8(4);
- const __m128i t3 = _mm_set1_epi8(3);
- const __m128i t80 = _mm_set1_epi8(0x80);
- const __m128i t1 = _mm_set1_epi8(0x1);
- const __m128i ps1 =
- _mm_xor_si128(_mm_loadl_epi64((__m128i *)(s - 2 * p)), t80);
- const __m128i ps0 =
- _mm_xor_si128(_mm_loadl_epi64((__m128i *)(s - 1 * p)), t80);
- const __m128i qs0 =
- _mm_xor_si128(_mm_loadl_epi64((__m128i *)(s + 0 * p)), t80);
- const __m128i qs1 =
- _mm_xor_si128(_mm_loadl_epi64((__m128i *)(s + 1 * p)), t80);
- __m128i filt;
- __m128i work_a;
- __m128i filter1, filter2;
+ if ((flatmmask & maskmmask) != 0xffff) {
+ // lp filter - the same for 8 and 6 versions
+ p1p0 = _mm_unpacklo_epi64(q0p0, q1p1);
+ q1q0 = _mm_unpackhi_epi64(q0p0, q1p1);
- filt = _mm_and_si128(_mm_subs_epi8(ps1, qs1), hev);
- work_a = _mm_subs_epi8(qs0, ps0);
- filt = _mm_adds_epi8(filt, work_a);
- filt = _mm_adds_epi8(filt, work_a);
- filt = _mm_adds_epi8(filt, work_a);
- // (aom_filter + 3 * (qs0 - ps0)) & mask
- filt = _mm_and_si128(filt, mask);
+ filter4_sse2(&p1p0, &q1q0, &hev, &mask, &qs1qs0, &ps1ps0);
+ } else {
+ ps1ps0 = _mm_setzero_si128();
+ qs1qs0 = _mm_setzero_si128();
+ }
+ qs1qs0 = _mm_andnot_si128(flat, qs1qs0);
+ q1q0 = _mm_and_si128(flat, flat_q0q1);
+ q1q0 = _mm_or_si128(qs1qs0, q1q0);
- filter1 = _mm_adds_epi8(filt, t4);
- filter2 = _mm_adds_epi8(filt, t3);
+ ps1ps0 = _mm_andnot_si128(flat, ps1ps0);
+ p1p0 = _mm_and_si128(flat, flat_p1p0);
+ p1p0 = _mm_or_si128(ps1ps0, p1p0);
- // Filter1 >> 3
- filter1 = _mm_unpacklo_epi8(zero, filter1);
- filter1 = _mm_srai_epi16(filter1, 11);
- filter1 = _mm_packs_epi16(filter1, filter1);
+ work_a = _mm_loadu_si128((__m128i *)(s + 2 * p));
+ work_a = _mm_andnot_si128(flat, work_a);
+ q2 = _mm_and_si128(flat, oq2);
+ q2 = _mm_or_si128(work_a, q2);
- // Filter2 >> 3
- filter2 = _mm_unpacklo_epi8(zero, filter2);
- filter2 = _mm_srai_epi16(filter2, 11);
- filter2 = _mm_packs_epi16(filter2, zero);
-
- // filt >> 1
- filt = _mm_adds_epi8(filter1, t1);
- filt = _mm_unpacklo_epi8(zero, filt);
- filt = _mm_srai_epi16(filt, 9);
- filt = _mm_packs_epi16(filt, zero);
-
- filt = _mm_andnot_si128(hev, filt);
-
- work_a = _mm_xor_si128(_mm_subs_epi8(qs0, filter1), t80);
- q0 = _mm_loadl_epi64((__m128i *)flat_oq0);
- work_a = _mm_andnot_si128(flat, work_a);
- q0 = _mm_and_si128(flat, q0);
- q0 = _mm_or_si128(work_a, q0);
-
- work_a = _mm_xor_si128(_mm_subs_epi8(qs1, filt), t80);
- q1 = _mm_loadl_epi64((__m128i *)flat_oq1);
- work_a = _mm_andnot_si128(flat, work_a);
- q1 = _mm_and_si128(flat, q1);
- q1 = _mm_or_si128(work_a, q1);
-
- work_a = _mm_loadu_si128((__m128i *)(s + 2 * p));
- q2 = _mm_loadl_epi64((__m128i *)flat_oq2);
- work_a = _mm_andnot_si128(flat, work_a);
- q2 = _mm_and_si128(flat, q2);
- q2 = _mm_or_si128(work_a, q2);
-
- work_a = _mm_xor_si128(_mm_adds_epi8(ps0, filter2), t80);
- p0 = _mm_loadl_epi64((__m128i *)flat_op0);
- work_a = _mm_andnot_si128(flat, work_a);
- p0 = _mm_and_si128(flat, p0);
- p0 = _mm_or_si128(work_a, p0);
-
- work_a = _mm_xor_si128(_mm_adds_epi8(ps1, filt), t80);
- p1 = _mm_loadl_epi64((__m128i *)flat_op1);
- work_a = _mm_andnot_si128(flat, work_a);
- p1 = _mm_and_si128(flat, p1);
- p1 = _mm_or_si128(work_a, p1);
-
- work_a = _mm_loadu_si128((__m128i *)(s - 3 * p));
- p2 = _mm_loadl_epi64((__m128i *)flat_op2);
- work_a = _mm_andnot_si128(flat, work_a);
- p2 = _mm_and_si128(flat, p2);
- p2 = _mm_or_si128(work_a, p2);
+ work_a = _mm_loadu_si128((__m128i *)(s - 3 * p));
+ work_a = _mm_andnot_si128(flat, work_a);
+ p2 = _mm_and_si128(flat, op2);
+ p2 = _mm_or_si128(work_a, p2);
#if CONFIG_PARALLEL_DEBLOCKING
- xx_storel_32(s - 3 * p, p2);
- xx_storel_32(s - 2 * p, p1);
- xx_storel_32(s - 1 * p, p0);
- xx_storel_32(s + 0 * p, q0);
- xx_storel_32(s + 1 * p, q1);
- xx_storel_32(s + 2 * p, q2);
+ xx_storel_32(s - 1 * p, p1p0);
+ xx_storel_32(s - 2 * p, _mm_srli_si128(p1p0, 8));
+ xx_storel_32(s + 0 * p, q1q0);
+ xx_storel_32(s + 1 * p, _mm_srli_si128(q1q0, 8));
+ xx_storel_32(s - 3 * p, p2);
+ xx_storel_32(s + 2 * p, q2);
#else
- xx_storel_64(s - 3 * p, p2);
- xx_storel_64(s - 2 * p, p1);
- xx_storel_64(s - 1 * p, p0);
- xx_storel_64(s + 0 * p, q0);
- xx_storel_64(s + 1 * p, q1);
- xx_storel_64(s + 2 * p, q2);
+ xx_storel_64(s - 1 * p, p1p0);
+ xx_storel_64(s - 2 * p, _mm_srli_si128(p1p0, 8));
+ xx_storel_64(s + 0 * p, q1q0);
+ xx_storel_64(s + 1 * p, _mm_srli_si128(q1q0, 8));
+ xx_storel_64(s - 3 * p, p2);
+ xx_storel_64(s + 2 * p, q2);
#endif
- }
}
void aom_lpf_horizontal_16_dual_sse2(unsigned char *s, int p,
@@ -1735,9 +2054,69 @@
}
#endif // CONFIG_PARALLEL_DEBLOCKING
-static INLINE void transpose(unsigned char *src[], int in_p,
- unsigned char *dst[], int out_p,
- int num_8x8_to_transpose) {
+
+static INLINE void transpose6x6(unsigned char *src[], int in_p,
+ unsigned char *dst[], int out_p,
+ int num_6x6_to_transpose) {
+ int idx6x6 = 0;
+ __m128i x0, x1, x2, x3, x4, x5, x6;
+ DECLARE_ALIGNED(16, unsigned char, temp_dst[16]); //
+ do {
+ unsigned char *in = src[idx6x6];
+ unsigned char *out = dst[idx6x6];
+
+ x0 =
+ _mm_loadl_epi64((__m128i *)(in + 0 * in_p)); // 00 01 02 03 04 05 xx xx
+ x1 =
+ _mm_loadl_epi64((__m128i *)(in + 1 * in_p)); // 10 11 12 13 14 15 xx xx
+
+ x0 = _mm_unpacklo_epi8(
+ x0, x1); // 00 10 01 11 02 12 03 13 04 14 05 15 xx xx xx xx
+
+ x2 = _mm_loadl_epi64((__m128i *)(in + 2 * in_p)); // 20 21 22 23 24 25
+ x3 = _mm_loadl_epi64((__m128i *)(in + 3 * in_p)); // 30 31 32 33 34 35
+
+ x1 = _mm_unpacklo_epi8(
+ x2, x3); // 20 30 21 31 22 32 23 33 24 34 25 35 xx xx xx xx
+
+ x4 = _mm_loadl_epi64((__m128i *)(in + 4 * in_p)); // 40 41 42 43 44 45
+ x5 = _mm_loadl_epi64((__m128i *)(in + 5 * in_p)); // 50 51 52 53 54 55
+
+ x2 = _mm_unpacklo_epi8(x4, x5); // 40 50 41 51 42 52 43 53 44 54 45 55
+
+ x4 = _mm_unpacklo_epi16(
+ x0, x1); // 00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33
+ x5 = _mm_unpacklo_epi16(
+ x2, x0); // 40 50 xx xx 41 51 xx xx 42 52 xx xx 43 53 xx xx
+
+ x6 = _mm_unpacklo_epi32(
+ x4, x5); // 00 10 20 30 40 50 xx xx 01 11 21 31 41 51 xx xx
+ _mm_store_si128((__m128i *)(temp_dst), x6);
+ memcpy(out + 0 * out_p, temp_dst, 6);
+ memcpy(out + 1 * out_p, temp_dst + 8, 6);
+
+ x6 = _mm_unpackhi_epi32(
+ x4, x5); // 02 12 22 32 42 52 xx xx 03 13 23 33 43 53 xx xx
+ _mm_store_si128((__m128i *)(temp_dst), x6);
+ memcpy(out + 2 * out_p, temp_dst, 6);
+ memcpy(out + 3 * out_p, temp_dst + 8, 6);
+
+ x4 = _mm_unpackhi_epi16(
+ x0, x1); // 04 14 24 34 05 15 25 35 xx xx xx xx xx xx xx xx
+ x5 = _mm_unpackhi_epi16(
+ x2, x3); // 44 54 xx xx 45 55 xx xx xx xx xx xx xx xx xx xx
+ x6 = _mm_unpacklo_epi32(
+ x4, x5); // 04 14 24 34 44 54 xx xx 05 15 25 35 45 55 xx xx
+
+ _mm_store_si128((__m128i *)(temp_dst), x6);
+ memcpy(out + 4 * out_p, temp_dst, 6);
+ memcpy(out + 5 * out_p, temp_dst + 8, 6);
+ } while (++idx6x6 < num_6x6_to_transpose);
+}
+
+static INLINE void transpose8x8(unsigned char *src[], int in_p,
+ unsigned char *dst[], int out_p,
+ int num_8x8_to_transpose) {
int idx8x8 = 0;
__m128i x0, x1, x2, x3, x4, x5, x6, x7;
do {
@@ -1748,15 +2127,17 @@
_mm_loadl_epi64((__m128i *)(in + 0 * in_p)); // 00 01 02 03 04 05 06 07
x1 =
_mm_loadl_epi64((__m128i *)(in + 1 * in_p)); // 10 11 12 13 14 15 16 17
- // 00 10 01 11 02 12 03 13 04 14 05 15 06 16 07 17
- x0 = _mm_unpacklo_epi8(x0, x1);
+
+ x0 = _mm_unpacklo_epi8(
+ x0, x1); // 00 10 01 11 02 12 03 13 04 14 05 15 06 16 07 17
x2 =
_mm_loadl_epi64((__m128i *)(in + 2 * in_p)); // 20 21 22 23 24 25 26 27
x3 =
_mm_loadl_epi64((__m128i *)(in + 3 * in_p)); // 30 31 32 33 34 35 36 37
- // 20 30 21 31 22 32 23 33 24 34 25 35 26 36 27 37
- x1 = _mm_unpacklo_epi8(x2, x3);
+
+ x1 = _mm_unpacklo_epi8(
+ x2, x3); // 20 30 21 31 22 32 23 33 24 34 25 35 26 36 27 37
x4 =
_mm_loadl_epi64((__m128i *)(in + 4 * in_p)); // 40 41 42 43 44 45 46 47
@@ -1837,6 +2218,30 @@
#endif // !CONFIG_PARALLEL_DEBLOCKING
}
+void aom_lpf_vertical_6_sse2(unsigned char *s, int p,
+ const unsigned char *blimit,
+ const unsigned char *limit,
+ const unsigned char *thresh) {
+ DECLARE_ALIGNED(8, unsigned char, t_dst[38]);
+ unsigned char *src[1];
+ unsigned char *dst[1];
+
+ // Transpose 6x6
+ src[0] = s - 3;
+ dst[0] = t_dst;
+
+ transpose6x6(src, p, dst, 6, 1);
+
+ // Loop filtering
+ aom_lpf_horizontal_6_sse2(t_dst + 3 * 6, 6, blimit, limit, thresh);
+
+ src[0] = t_dst;
+ dst[0] = s - 3;
+
+ // Transpose back
+ transpose6x6(src, 6, dst, p, 1);
+}
+
void aom_lpf_vertical_8_sse2(unsigned char *s, int p,
const unsigned char *blimit,
const unsigned char *limit,
@@ -1849,7 +2254,7 @@
src[0] = s - 4;
dst[0] = t_dst;
- transpose(src, p, dst, 8, 1);
+ transpose8x8(src, p, dst, 8, 1);
// Loop filtering
aom_lpf_horizontal_8_sse2(t_dst + 4 * 8, 8, blimit, limit, thresh);
@@ -1858,7 +2263,7 @@
dst[0] = s - 4;
// Transpose back
- transpose(src, 8, dst, p, 1);
+ transpose8x8(src, 8, dst, p, 1);
}
void aom_lpf_vertical_8_dual_sse2(uint8_t *s, int p, const uint8_t *blimit0,
@@ -1882,7 +2287,7 @@
dst[1] = s - 4 + p * 8;
// Transpose back
- transpose(src, 16, dst, p, 2);
+ transpose8x8(src, 16, dst, p, 2);
}
void aom_lpf_vertical_16_sse2(unsigned char *s, int p,
@@ -1899,7 +2304,7 @@
dst[1] = t_dst + 8 * 8;
// Transpose 16x8
- transpose(src, p, dst, 8, 2);
+ transpose8x8(src, p, dst, 8, 2);
// Loop filtering
aom_lpf_horizontal_16_sse2(t_dst + 8 * 8, 8, blimit, limit, thresh);
@@ -1910,7 +2315,7 @@
dst[1] = s;
// Transpose back
- transpose(src, 8, dst, p, 2);
+ transpose8x8(src, 8, dst, p, 2);
}
void aom_lpf_vertical_16_dual_sse2(unsigned char *s, int p,
diff --git a/av1/common/av1_loopfilter.c b/av1/common/av1_loopfilter.c
index fafe2de..1732cb6 100644
--- a/av1/common/av1_loopfilter.c
+++ b/av1/common/av1_loopfilter.c
@@ -42,12 +42,6 @@
#endif
#if PARALLEL_DEBLOCKING_5_TAP_CHROMA
-extern void aom_lpf_vertical_6_c(uint8_t *s, int pitch, const uint8_t *blimit,
- const uint8_t *limit, const uint8_t *thresh);
-
-extern void aom_lpf_horizontal_6_c(uint8_t *s, int p, const uint8_t *blimit,
- const uint8_t *limit, const uint8_t *thresh);
-
extern void aom_highbd_lpf_horizontal_6_c(uint16_t *s, int p,
const uint8_t *blimit,
const uint8_t *limit,
@@ -2233,8 +2227,8 @@
params.mblim, params.lim,
params.hev_thr, cm->bit_depth);
else
- aom_lpf_vertical_6_c(p, dst_stride, params.mblim, params.lim,
- params.hev_thr);
+ aom_lpf_vertical_6(p, dst_stride, params.mblim, params.lim,
+ params.hev_thr);
break;
#endif
// apply 8-tap filtering
@@ -2261,13 +2255,8 @@
cm->bit_depth);
#endif
else
-#if CONFIG_DEBLOCK_13TAP
- aom_lpf_vertical_16_c(p, dst_stride, params.mblim, params.lim,
- params.hev_thr);
-#else
aom_lpf_vertical_16(p, dst_stride, params.mblim, params.lim,
params.hev_thr);
-#endif
break;
// no filtering
default: break;
@@ -2335,8 +2324,8 @@
params.mblim, params.lim,
params.hev_thr, cm->bit_depth);
else
- aom_lpf_horizontal_6_c(p, dst_stride, params.mblim, params.lim,
- params.hev_thr);
+ aom_lpf_horizontal_6(p, dst_stride, params.mblim, params.lim,
+ params.hev_thr);
break;
#endif
// apply 8-tap filtering
diff --git a/test/lpf_test.cc b/test/lpf_test.cc
index 7c54f93..12a25c6 100644
--- a/test/lpf_test.cc
+++ b/test/lpf_test.cc
@@ -37,20 +37,26 @@
const int kSpeedTestNum = 500000;
-typedef uint16_t Pixel;
-#define PIXEL_WIDTH 16
+#define LOOP_PARAM \
+ int p, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh
+#define DUAL_LOOP_PARAM \
+ int p, const uint8_t *blimit0, const uint8_t *limit0, \
+ const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, \
+ const uint8_t *thresh1
-typedef void (*loop_op_t)(Pixel *s, int p, const uint8_t *blimit,
- const uint8_t *limit, const uint8_t *thresh, int bd);
-typedef void (*dual_loop_op_t)(Pixel *s, int p, const uint8_t *blimit0,
- const uint8_t *limit0, const uint8_t *thresh0,
- const uint8_t *blimit1, const uint8_t *limit1,
- const uint8_t *thresh1, int bd);
+typedef void (*loop_op_t)(uint8_t *s, LOOP_PARAM);
+typedef void (*dual_loop_op_t)(uint8_t *s, DUAL_LOOP_PARAM);
+typedef void (*hbdloop_op_t)(uint16_t *s, LOOP_PARAM, int bd);
+typedef void (*hbddual_loop_op_t)(uint16_t *s, DUAL_LOOP_PARAM, int bd);
-typedef std::tr1::tuple<loop_op_t, loop_op_t, int> loop8_param_t;
-typedef std::tr1::tuple<dual_loop_op_t, dual_loop_op_t, int> dualloop8_param_t;
+typedef std::tr1::tuple<hbdloop_op_t, hbdloop_op_t, int> hbdloop_param_t;
+typedef std::tr1::tuple<hbddual_loop_op_t, hbddual_loop_op_t, int>
+ hbddual_loop_param_t;
+typedef std::tr1::tuple<loop_op_t, loop_op_t, int> loop_param_t;
+typedef std::tr1::tuple<dual_loop_op_t, dual_loop_op_t, int> dual_loop_param_t;
-void InitInput(Pixel *s, Pixel *ref_s, ACMRandom *rnd, const uint8_t limit,
+template <typename Pixel_t, int PIXEL_WIDTH_t>
+void InitInput(Pixel_t *s, Pixel_t *ref_s, ACMRandom *rnd, const uint8_t limit,
const int mask, const int32_t p, const int i) {
uint16_t tmp_s[kNumCoeffs];
@@ -117,13 +123,14 @@
return static_cast<uint8_t>(rnd->PseudoUniform(MAX_LOOP_FILTER + 1) >> 4);
}
-class Loop8Test6Param : public ::testing::TestWithParam<loop8_param_t> {
+template <typename func_type_t, typename params_t>
+class LoopTestParam : public ::testing::TestWithParam<params_t> {
public:
- virtual ~Loop8Test6Param() {}
+ virtual ~LoopTestParam() {}
virtual void SetUp() {
- loopfilter_op_ = GET_PARAM(0);
- ref_loopfilter_op_ = GET_PARAM(1);
- bit_depth_ = GET_PARAM(2);
+ loopfilter_op_ = std::tr1::get<0>(this->GetParam());
+ ref_loopfilter_op_ = std::tr1::get<1>(this->GetParam());
+ bit_depth_ = std::tr1::get<2>(this->GetParam());
mask_ = (1 << bit_depth_) - 1;
}
@@ -132,319 +139,318 @@
protected:
int bit_depth_;
int mask_;
- loop_op_t loopfilter_op_;
- loop_op_t ref_loopfilter_op_;
+ func_type_t loopfilter_op_;
+ func_type_t ref_loopfilter_op_;
};
-class Loop8Test9Param : public ::testing::TestWithParam<dualloop8_param_t> {
- public:
- virtual ~Loop8Test9Param() {}
- virtual void SetUp() {
- loopfilter_op_ = GET_PARAM(0);
- ref_loopfilter_op_ = GET_PARAM(1);
- bit_depth_ = GET_PARAM(2);
- mask_ = (1 << bit_depth_) - 1;
- }
-
- virtual void TearDown() { libaom_test::ClearSystemState(); }
-
- protected:
- int bit_depth_;
- int mask_;
- dual_loop_op_t loopfilter_op_;
- dual_loop_op_t ref_loopfilter_op_;
+void call_filter(uint16_t *s, LOOP_PARAM, int bd, hbdloop_op_t op) {
+ op(s, p, blimit, limit, thresh, bd);
+}
+void call_filter(uint8_t *s, LOOP_PARAM, int bd, loop_op_t op) {
+ (void)bd;
+ op(s, p, blimit, limit, thresh);
+}
+void call_dualfilter(uint16_t *s, DUAL_LOOP_PARAM, int bd,
+ hbddual_loop_op_t op) {
+ op(s, p, blimit0, limit0, thresh0, blimit1, limit1, thresh1, bd);
+}
+void call_dualfilter(uint8_t *s, DUAL_LOOP_PARAM, int bd, dual_loop_op_t op) {
+ (void)bd;
+ op(s, p, blimit0, limit0, thresh0, blimit1, limit1, thresh1);
};
-TEST_P(Loop8Test6Param, OperationCheck) {
- ACMRandom rnd(ACMRandom::DeterministicSeed());
- const int count_test_block = number_of_iterations;
- const int32_t p = kNumCoeffs / 32;
- DECLARE_ALIGNED(PIXEL_WIDTH, Pixel, s[kNumCoeffs]);
- DECLARE_ALIGNED(PIXEL_WIDTH, Pixel, ref_s[kNumCoeffs]);
- int err_count_total = 0;
- int first_failure = -1;
- for (int i = 0; i < count_test_block; ++i) {
- int err_count = 0;
- uint8_t tmp = GetOuterThresh(&rnd);
- DECLARE_ALIGNED(16, const uint8_t,
- blimit[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
- tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp };
- tmp = GetInnerThresh(&rnd);
- DECLARE_ALIGNED(16, const uint8_t,
- limit[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
- tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp };
- tmp = GetHevThresh(&rnd);
- DECLARE_ALIGNED(16, const uint8_t,
- thresh[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
- tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp };
- InitInput(s, ref_s, &rnd, *limit, mask_, p, i);
- ref_loopfilter_op_(ref_s + 8 + p * 8, p, blimit, limit, thresh, bit_depth_);
- ASM_REGISTER_STATE_CHECK(
- loopfilter_op_(s + 8 + p * 8, p, blimit, limit, thresh, bit_depth_));
+typedef LoopTestParam<hbdloop_op_t, hbdloop_param_t> Loop8Test6Param_hbd;
+typedef LoopTestParam<loop_op_t, loop_param_t> Loop8Test6Param_lbd;
+typedef LoopTestParam<hbddual_loop_op_t, hbddual_loop_param_t>
+ Loop8Test9Param_hbd;
+typedef LoopTestParam<dual_loop_op_t, dual_loop_param_t> Loop8Test9Param_lbd;
- for (int j = 0; j < kNumCoeffs; ++j) {
- err_count += ref_s[j] != s[j];
- }
- if (err_count && !err_count_total) {
- first_failure = i;
- }
- err_count_total += err_count;
- }
- EXPECT_EQ(0, err_count_total)
- << "Error: Loop8Test6Param, C output doesn't match SSE2 "
- "loopfilter output. "
+#define OPCHECK(a, b) \
+ ACMRandom rnd(ACMRandom::DeterministicSeed()); \
+ const int count_test_block = number_of_iterations; \
+ const int32_t p = kNumCoeffs / 32; \
+ DECLARE_ALIGNED(b, a, s[kNumCoeffs]); \
+ DECLARE_ALIGNED(b, a, ref_s[kNumCoeffs]); \
+ int err_count_total = 0; \
+ int first_failure = -1; \
+ for (int i = 0; i < count_test_block; ++i) { \
+ int err_count = 0; \
+ uint8_t tmp = GetOuterThresh(&rnd); \
+ DECLARE_ALIGNED(16, const uint8_t, \
+ blimit[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp, \
+ tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp }; \
+ tmp = GetInnerThresh(&rnd); \
+ DECLARE_ALIGNED(16, const uint8_t, \
+ limit[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp, \
+ tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp }; \
+ tmp = GetHevThresh(&rnd); \
+ DECLARE_ALIGNED(16, const uint8_t, \
+ thresh[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp, \
+ tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp }; \
+ InitInput<a, b>(s, ref_s, &rnd, *limit, mask_, p, i); \
+ call_filter(ref_s + 8 + p * 8, p, blimit, limit, thresh, bit_depth_, \
+ ref_loopfilter_op_); \
+ ASM_REGISTER_STATE_CHECK(call_filter(s + 8 + p * 8, p, blimit, limit, \
+ thresh, bit_depth_, loopfilter_op_)); \
+ for (int j = 0; j < kNumCoeffs; ++j) { \
+ err_count += ref_s[j] != s[j]; \
+ } \
+ if (err_count && !err_count_total) { \
+ first_failure = i; \
+ } \
+ err_count_total += err_count; \
+ } \
+ EXPECT_EQ(0, err_count_total) \
+ << "Error: Loop8Test6Param, C output doesn't match SIMD " \
+ "loopfilter output. " \
<< "First failed at test case " << first_failure;
-}
-TEST_P(Loop8Test6Param, ValueCheck) {
- ACMRandom rnd(ACMRandom::DeterministicSeed());
- const int count_test_block = number_of_iterations;
- DECLARE_ALIGNED(PIXEL_WIDTH, Pixel, s[kNumCoeffs]);
- DECLARE_ALIGNED(PIXEL_WIDTH, Pixel, ref_s[kNumCoeffs]);
- int err_count_total = 0;
- int first_failure = -1;
+TEST_P(Loop8Test6Param_hbd, OperationCheck) { OPCHECK(uint16_t, 16); }
+TEST_P(Loop8Test6Param_lbd, OperationCheck) { OPCHECK(uint8_t, 8); }
- // NOTE: The code in av1_loopfilter.c:update_sharpness computes mblim as a
- // function of sharpness_lvl and the loopfilter lvl as:
- // block_inside_limit = lvl >> ((sharpness_lvl > 0) + (sharpness_lvl > 4));
- // ...
- // memset(lfi->lfthr[lvl].mblim, (2 * (lvl + 2) + block_inside_limit),
- // SIMD_WIDTH);
- // This means that the largest value for mblim will occur when sharpness_lvl
- // is equal to 0, and lvl is equal to its greatest value (MAX_LOOP_FILTER).
- // In this case block_inside_limit will be equal to MAX_LOOP_FILTER and
- // therefore mblim will be equal to (2 * (lvl + 2) + block_inside_limit) =
- // 2 * (MAX_LOOP_FILTER + 2) + MAX_LOOP_FILTER = 3 * MAX_LOOP_FILTER + 4
-
- for (int i = 0; i < count_test_block; ++i) {
- int err_count = 0;
- uint8_t tmp = GetOuterThresh(&rnd);
- DECLARE_ALIGNED(16, const uint8_t,
- blimit[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
- tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp };
- tmp = GetInnerThresh(&rnd);
- DECLARE_ALIGNED(16, const uint8_t,
- limit[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
- tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp };
- tmp = GetHevThresh(&rnd);
- DECLARE_ALIGNED(16, const uint8_t,
- thresh[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
- tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp };
- int32_t p = kNumCoeffs / 32;
- for (int j = 0; j < kNumCoeffs; ++j) {
- s[j] = rnd.Rand16() & mask_;
- ref_s[j] = s[j];
- }
- ref_loopfilter_op_(ref_s + 8 + p * 8, p, blimit, limit, thresh, bit_depth_);
- ASM_REGISTER_STATE_CHECK(
- loopfilter_op_(s + 8 + p * 8, p, blimit, limit, thresh, bit_depth_));
-
- for (int j = 0; j < kNumCoeffs; ++j) {
- err_count += ref_s[j] != s[j];
- }
- if (err_count && !err_count_total) {
- first_failure = i;
- }
- err_count_total += err_count;
- }
- EXPECT_EQ(0, err_count_total)
- << "Error: Loop8Test6Param, C output doesn't match SSE2 "
- "loopfilter output. "
+#define VALCHECK(a, b) \
+ ACMRandom rnd(ACMRandom::DeterministicSeed()); \
+ const int count_test_block = number_of_iterations; \
+ DECLARE_ALIGNED(b, a, s[kNumCoeffs]); \
+ DECLARE_ALIGNED(b, a, ref_s[kNumCoeffs]); \
+ int err_count_total = 0; \
+ int first_failure = -1; \
+ for (int i = 0; i < count_test_block; ++i) { \
+ int err_count = 0; \
+ uint8_t tmp = GetOuterThresh(&rnd); \
+ DECLARE_ALIGNED(16, const uint8_t, \
+ blimit[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp, \
+ tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp }; \
+ tmp = GetInnerThresh(&rnd); \
+ DECLARE_ALIGNED(16, const uint8_t, \
+ limit[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp, \
+ tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp }; \
+ tmp = GetHevThresh(&rnd); \
+ DECLARE_ALIGNED(16, const uint8_t, \
+ thresh[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp, \
+ tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp }; \
+ int32_t p = kNumCoeffs / 32; \
+ for (int j = 0; j < kNumCoeffs; ++j) { \
+ s[j] = rnd.Rand16() & mask_; \
+ ref_s[j] = s[j]; \
+ } \
+ call_filter(ref_s + 8 + p * 8, p, blimit, limit, thresh, bit_depth_, \
+ ref_loopfilter_op_); \
+ ASM_REGISTER_STATE_CHECK(call_filter(s + 8 + p * 8, p, blimit, limit, \
+ thresh, bit_depth_, loopfilter_op_)); \
+ for (int j = 0; j < kNumCoeffs; ++j) { \
+ err_count += ref_s[j] != s[j]; \
+ } \
+ if (err_count && !err_count_total) { \
+ first_failure = i; \
+ } \
+ err_count_total += err_count; \
+ } \
+ EXPECT_EQ(0, err_count_total) \
+ << "Error: Loop8Test6Param, C output doesn't match SIMD " \
+ "loopfilter output. " \
<< "First failed at test case " << first_failure;
-}
-TEST_P(Loop8Test6Param, DISABLED_Speed) {
- ACMRandom rnd(ACMRandom::DeterministicSeed());
- const int count_test_block = kSpeedTestNum;
- const int32_t bd = bit_depth_;
- DECLARE_ALIGNED(16, uint16_t, s[kNumCoeffs]);
+TEST_P(Loop8Test6Param_hbd, ValueCheck) { VALCHECK(uint16_t, 16); }
+TEST_P(Loop8Test6Param_lbd, ValueCheck) { VALCHECK(uint8_t, 8); }
- uint8_t tmp = GetOuterThresh(&rnd);
- DECLARE_ALIGNED(16, const uint8_t,
- blimit[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
- tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp };
- tmp = GetInnerThresh(&rnd);
- DECLARE_ALIGNED(16, const uint8_t,
- limit[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
- tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp };
- tmp = GetHevThresh(&rnd);
- DECLARE_ALIGNED(16, const uint8_t,
- thresh[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
- tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp };
-
- int32_t p = kNumCoeffs / 32;
- for (int j = 0; j < kNumCoeffs; ++j) {
- s[j] = rnd.Rand16() & mask_;
+#define SPEEDCHECK(a, b) \
+ ACMRandom rnd(ACMRandom::DeterministicSeed()); \
+ const int count_test_block = kSpeedTestNum; \
+ const int32_t bd = bit_depth_; \
+ DECLARE_ALIGNED(b, a, s[kNumCoeffs]); \
+ uint8_t tmp = GetOuterThresh(&rnd); \
+ DECLARE_ALIGNED(16, const uint8_t, \
+ blimit[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp, \
+ tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp }; \
+ tmp = GetInnerThresh(&rnd); \
+ DECLARE_ALIGNED(16, const uint8_t, \
+ limit[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp, \
+ tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp }; \
+ tmp = GetHevThresh(&rnd); \
+ DECLARE_ALIGNED(16, const uint8_t, \
+ thresh[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp, \
+ tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp }; \
+ int32_t p = kNumCoeffs / 32; \
+ for (int j = 0; j < kNumCoeffs; ++j) { \
+ s[j] = rnd.Rand16() & mask_; \
+ } \
+ for (int i = 0; i < count_test_block; ++i) { \
+ call_filter(s + 8 + p * 8, p, blimit, limit, thresh, bd, loopfilter_op_); \
}
- for (int i = 0; i < count_test_block; ++i) {
- loopfilter_op_(s + 8 + p * 8, p, blimit, limit, thresh, bd);
- }
-}
+TEST_P(Loop8Test6Param_hbd, DISABLED_Speed) { SPEEDCHECK(uint16_t, 16); }
+TEST_P(Loop8Test6Param_lbd, DISABLED_Speed) { SPEEDCHECK(uint8_t, 8); }
-TEST_P(Loop8Test9Param, OperationCheck) {
- ACMRandom rnd(ACMRandom::DeterministicSeed());
- const int count_test_block = number_of_iterations;
- DECLARE_ALIGNED(PIXEL_WIDTH, Pixel, s[kNumCoeffs]);
- DECLARE_ALIGNED(PIXEL_WIDTH, Pixel, ref_s[kNumCoeffs]);
- int err_count_total = 0;
- int first_failure = -1;
- for (int i = 0; i < count_test_block; ++i) {
- int err_count = 0;
- uint8_t tmp = GetOuterThresh(&rnd);
- DECLARE_ALIGNED(16, const uint8_t,
- blimit0[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
- tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp };
- tmp = GetInnerThresh(&rnd);
- DECLARE_ALIGNED(16, const uint8_t,
- limit0[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
- tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp };
- tmp = GetHevThresh(&rnd);
- DECLARE_ALIGNED(16, const uint8_t,
- thresh0[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
- tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp };
- tmp = GetOuterThresh(&rnd);
- DECLARE_ALIGNED(16, const uint8_t,
- blimit1[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
- tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp };
- tmp = GetInnerThresh(&rnd);
- DECLARE_ALIGNED(16, const uint8_t,
- limit1[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
- tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp };
- tmp = GetHevThresh(&rnd);
- DECLARE_ALIGNED(16, const uint8_t,
- thresh1[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
- tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp };
- int32_t p = kNumCoeffs / 32;
- const uint8_t limit = *limit0 < *limit1 ? *limit0 : *limit1;
- InitInput(s, ref_s, &rnd, limit, mask_, p, i);
- ref_loopfilter_op_(ref_s + 8 + p * 8, p, blimit0, limit0, thresh0, blimit1,
- limit1, thresh1, bit_depth_);
- ASM_REGISTER_STATE_CHECK(loopfilter_op_(s + 8 + p * 8, p, blimit0, limit0,
- thresh0, blimit1, limit1, thresh1,
- bit_depth_));
-
- for (int j = 0; j < kNumCoeffs; ++j) {
- err_count += ref_s[j] != s[j];
- }
- if (err_count && !err_count_total) {
- first_failure = i;
- }
- err_count_total += err_count;
- }
- EXPECT_EQ(0, err_count_total)
- << "Error: Loop8Test9Param, C output doesn't match SSE2 "
- "loopfilter output. "
+#define OPCHECKd(a, b) \
+ ACMRandom rnd(ACMRandom::DeterministicSeed()); \
+ const int count_test_block = number_of_iterations; \
+ DECLARE_ALIGNED(b, a, s[kNumCoeffs]); \
+ DECLARE_ALIGNED(b, a, ref_s[kNumCoeffs]); \
+ int err_count_total = 0; \
+ int first_failure = -1; \
+ for (int i = 0; i < count_test_block; ++i) { \
+ int err_count = 0; \
+ uint8_t tmp = GetOuterThresh(&rnd); \
+ DECLARE_ALIGNED(16, const uint8_t, \
+ blimit0[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp, \
+ tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp }; \
+ tmp = GetInnerThresh(&rnd); \
+ DECLARE_ALIGNED(16, const uint8_t, \
+ limit0[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp, \
+ tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp }; \
+ tmp = GetHevThresh(&rnd); \
+ DECLARE_ALIGNED(16, const uint8_t, \
+ thresh0[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp, \
+ tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp }; \
+ tmp = GetOuterThresh(&rnd); \
+ DECLARE_ALIGNED(16, const uint8_t, \
+ blimit1[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp, \
+ tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp }; \
+ tmp = GetInnerThresh(&rnd); \
+ DECLARE_ALIGNED(16, const uint8_t, \
+ limit1[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp, \
+ tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp }; \
+ tmp = GetHevThresh(&rnd); \
+ DECLARE_ALIGNED(16, const uint8_t, \
+ thresh1[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp, \
+ tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp }; \
+ int32_t p = kNumCoeffs / 32; \
+ const uint8_t limit = *limit0 < *limit1 ? *limit0 : *limit1; \
+ InitInput<a, b>(s, ref_s, &rnd, limit, mask_, p, i); \
+ call_dualfilter(ref_s + 8 + p * 8, p, blimit0, limit0, thresh0, blimit1, \
+ limit1, thresh1, bit_depth_, ref_loopfilter_op_); \
+ ASM_REGISTER_STATE_CHECK( \
+ call_dualfilter(s + 8 + p * 8, p, blimit0, limit0, thresh0, blimit1, \
+ limit1, thresh1, bit_depth_, loopfilter_op_)); \
+ for (int j = 0; j < kNumCoeffs; ++j) { \
+ err_count += ref_s[j] != s[j]; \
+ } \
+ if (err_count && !err_count_total) { \
+ first_failure = i; \
+ } \
+ err_count_total += err_count; \
+ } \
+ EXPECT_EQ(0, err_count_total) \
+ << "Error: Loop8Test9Param, C output doesn't match SIMD " \
+ "loopfilter output. " \
<< "First failed at test case " << first_failure;
-}
-TEST_P(Loop8Test9Param, ValueCheck) {
- ACMRandom rnd(ACMRandom::DeterministicSeed());
- const int count_test_block = number_of_iterations;
- DECLARE_ALIGNED(PIXEL_WIDTH, Pixel, s[kNumCoeffs]);
- DECLARE_ALIGNED(PIXEL_WIDTH, Pixel, ref_s[kNumCoeffs]);
- int err_count_total = 0;
- int first_failure = -1;
- for (int i = 0; i < count_test_block; ++i) {
- int err_count = 0;
- uint8_t tmp = GetOuterThresh(&rnd);
- DECLARE_ALIGNED(16, const uint8_t,
- blimit0[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
- tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp };
- tmp = GetInnerThresh(&rnd);
- DECLARE_ALIGNED(16, const uint8_t,
- limit0[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
- tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp };
- tmp = GetHevThresh(&rnd);
- DECLARE_ALIGNED(16, const uint8_t,
- thresh0[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
- tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp };
- tmp = GetOuterThresh(&rnd);
- DECLARE_ALIGNED(16, const uint8_t,
- blimit1[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
- tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp };
- tmp = GetInnerThresh(&rnd);
- DECLARE_ALIGNED(16, const uint8_t,
- limit1[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
- tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp };
- tmp = GetHevThresh(&rnd);
- DECLARE_ALIGNED(16, const uint8_t,
- thresh1[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
- tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp };
- int32_t p = kNumCoeffs / 32; // TODO(pdlf) can we have non-square here?
- for (int j = 0; j < kNumCoeffs; ++j) {
- s[j] = rnd.Rand16() & mask_;
- ref_s[j] = s[j];
- }
- ref_loopfilter_op_(ref_s + 8 + p * 8, p, blimit0, limit0, thresh0, blimit1,
- limit1, thresh1, bit_depth_);
- ASM_REGISTER_STATE_CHECK(loopfilter_op_(s + 8 + p * 8, p, blimit0, limit0,
- thresh0, blimit1, limit1, thresh1,
- bit_depth_));
+TEST_P(Loop8Test9Param_hbd, OperationCheck) { OPCHECKd(uint16_t, 16); }
+TEST_P(Loop8Test9Param_lbd, OperationCheck) { OPCHECKd(uint8_t, 8); }
- for (int j = 0; j < kNumCoeffs; ++j) {
- err_count += ref_s[j] != s[j];
- }
- if (err_count && !err_count_total) {
- first_failure = i;
- }
- err_count_total += err_count;
- }
- EXPECT_EQ(0, err_count_total)
- << "Error: Loop8Test9Param, C output doesn't match SSE2"
- "loopfilter output. "
+#define VALCHECKd(a, b) \
+ ACMRandom rnd(ACMRandom::DeterministicSeed()); \
+ const int count_test_block = number_of_iterations; \
+ DECLARE_ALIGNED(b, a, s[kNumCoeffs]); \
+ DECLARE_ALIGNED(b, a, ref_s[kNumCoeffs]); \
+ int err_count_total = 0; \
+ int first_failure = -1; \
+ for (int i = 0; i < count_test_block; ++i) { \
+ int err_count = 0; \
+ uint8_t tmp = GetOuterThresh(&rnd); \
+ DECLARE_ALIGNED(16, const uint8_t, \
+ blimit0[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp, \
+ tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp }; \
+ tmp = GetInnerThresh(&rnd); \
+ DECLARE_ALIGNED(16, const uint8_t, \
+ limit0[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp, \
+ tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp }; \
+ tmp = GetHevThresh(&rnd); \
+ DECLARE_ALIGNED(16, const uint8_t, \
+ thresh0[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp, \
+ tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp }; \
+ tmp = GetOuterThresh(&rnd); \
+ DECLARE_ALIGNED(16, const uint8_t, \
+ blimit1[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp, \
+ tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp }; \
+ tmp = GetInnerThresh(&rnd); \
+ DECLARE_ALIGNED(16, const uint8_t, \
+ limit1[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp, \
+ tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp }; \
+ tmp = GetHevThresh(&rnd); \
+ DECLARE_ALIGNED(16, const uint8_t, \
+ thresh1[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp, \
+ tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp }; \
+ int32_t p = kNumCoeffs / 32; \
+ for (int j = 0; j < kNumCoeffs; ++j) { \
+ s[j] = rnd.Rand16() & mask_; \
+ ref_s[j] = s[j]; \
+ } \
+ call_dualfilter(ref_s + 8 + p * 8, p, blimit0, limit0, thresh0, blimit1, \
+ limit1, thresh1, bit_depth_, ref_loopfilter_op_); \
+ ASM_REGISTER_STATE_CHECK( \
+ call_dualfilter(s + 8 + p * 8, p, blimit0, limit0, thresh0, blimit1, \
+ limit1, thresh1, bit_depth_, loopfilter_op_)); \
+ for (int j = 0; j < kNumCoeffs; ++j) { \
+ err_count += ref_s[j] != s[j]; \
+ } \
+ if (err_count && !err_count_total) { \
+ first_failure = i; \
+ } \
+ err_count_total += err_count; \
+ } \
+ EXPECT_EQ(0, err_count_total) \
+ << "Error: Loop8Test9Param, C output doesn't match SIMD " \
+ "loopfilter output. " \
<< "First failed at test case " << first_failure;
-}
-TEST_P(Loop8Test9Param, DISABLED_Speed) {
- ACMRandom rnd(ACMRandom::DeterministicSeed());
- const int count_test_block = kSpeedTestNum;
- DECLARE_ALIGNED(16, uint16_t, s[kNumCoeffs]);
+TEST_P(Loop8Test9Param_hbd, ValueCheck) { VALCHECKd(uint16_t, 16); }
+TEST_P(Loop8Test9Param_lbd, ValueCheck) { VALCHECKd(uint8_t, 8); }
- uint8_t tmp = GetOuterThresh(&rnd);
- DECLARE_ALIGNED(16, const uint8_t,
- blimit0[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
- tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp };
- tmp = GetInnerThresh(&rnd);
- DECLARE_ALIGNED(16, const uint8_t,
- limit0[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
- tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp };
- tmp = GetHevThresh(&rnd);
- DECLARE_ALIGNED(16, const uint8_t,
- thresh0[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
- tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp };
- tmp = GetOuterThresh(&rnd);
- DECLARE_ALIGNED(16, const uint8_t,
- blimit1[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
- tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp };
- tmp = GetInnerThresh(&rnd);
- DECLARE_ALIGNED(16, const uint8_t,
- limit1[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
- tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp };
- tmp = GetHevThresh(&rnd);
- DECLARE_ALIGNED(16, const uint8_t,
- thresh1[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
- tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp };
- int32_t p = kNumCoeffs / 32; // TODO(pdlf) can we have non-square here?
- for (int j = 0; j < kNumCoeffs; ++j) {
- s[j] = rnd.Rand16() & mask_;
+#define SPEEDCHECKd(a, b) \
+ ACMRandom rnd(ACMRandom::DeterministicSeed()); \
+ const int count_test_block = kSpeedTestNum; \
+ DECLARE_ALIGNED(b, a, s[kNumCoeffs]); \
+ uint8_t tmp = GetOuterThresh(&rnd); \
+ DECLARE_ALIGNED(16, const uint8_t, \
+ blimit0[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp, \
+ tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp }; \
+ tmp = GetInnerThresh(&rnd); \
+ DECLARE_ALIGNED(16, const uint8_t, \
+ limit0[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp, \
+ tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp }; \
+ tmp = GetHevThresh(&rnd); \
+ DECLARE_ALIGNED(16, const uint8_t, \
+ thresh0[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp, \
+ tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp }; \
+ tmp = GetOuterThresh(&rnd); \
+ DECLARE_ALIGNED(16, const uint8_t, \
+ blimit1[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp, \
+ tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp }; \
+ tmp = GetInnerThresh(&rnd); \
+ DECLARE_ALIGNED(16, const uint8_t, \
+ limit1[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp, \
+ tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp }; \
+ tmp = GetHevThresh(&rnd); \
+ DECLARE_ALIGNED(16, const uint8_t, \
+ thresh1[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp, \
+ tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp }; \
+ int32_t p = kNumCoeffs / 32; \
+ for (int j = 0; j < kNumCoeffs; ++j) { \
+ s[j] = rnd.Rand16() & mask_; \
+ } \
+ for (int i = 0; i < count_test_block; ++i) { \
+ call_dualfilter(s + 8 + p * 8, p, blimit0, limit0, thresh0, blimit1, \
+ limit1, thresh1, bit_depth_, loopfilter_op_); \
}
- for (int i = 0; i < count_test_block; ++i) {
- const int32_t bd = bit_depth_;
- loopfilter_op_(s + 8 + p * 8, p, blimit0, limit0, thresh0, blimit1, limit1,
- thresh1, bd);
- }
-}
+TEST_P(Loop8Test9Param_hbd, DISABLED_Speed) { SPEEDCHECKd(uint16_t, 16); }
+TEST_P(Loop8Test9Param_lbd, DISABLED_Speed) { SPEEDCHECKd(uint8_t, 8); }
using std::tr1::make_tuple;
#if HAVE_SSE2
-const loop8_param_t kHbdLoop8Test6[] = {
+const hbdloop_param_t kHbdLoop8Test6[] = {
make_tuple(&aom_highbd_lpf_horizontal_4_sse2, &aom_highbd_lpf_horizontal_4_c,
8),
make_tuple(&aom_highbd_lpf_vertical_4_sse2, &aom_highbd_lpf_vertical_4_c, 8),
+#if PARALLEL_DEBLOCKING_5_TAP_CHROMA
+ make_tuple(&aom_highbd_lpf_horizontal_6_sse2, &aom_highbd_lpf_horizontal_6_c,
+ 8),
+ make_tuple(&aom_highbd_lpf_vertical_6_sse2, &aom_highbd_lpf_vertical_6_c, 8),
+#endif
make_tuple(&aom_highbd_lpf_horizontal_8_sse2, &aom_highbd_lpf_horizontal_8_c,
8),
#if !CONFIG_DEBLOCK_13TAP
@@ -508,13 +514,37 @@
make_tuple(&aom_highbd_lpf_vertical_8_sse2, &aom_highbd_lpf_vertical_8_c, 12)
};
-INSTANTIATE_TEST_CASE_P(SSE2, Loop8Test6Param,
+INSTANTIATE_TEST_CASE_P(SSE2, Loop8Test6Param_hbd,
::testing::ValuesIn(kHbdLoop8Test6));
+
+const loop_param_t kLoop8Test6[] = {
+ make_tuple(&aom_lpf_horizontal_4_sse2, &aom_lpf_horizontal_4_c, 8),
+ make_tuple(&aom_lpf_horizontal_8_sse2, &aom_lpf_horizontal_8_c, 8),
+#if CONFIG_DEBLOCK_13TAP
+ make_tuple(&aom_lpf_horizontal_6_sse2, &aom_lpf_horizontal_6_c, 8),
+ make_tuple(&aom_lpf_vertical_6_sse2, &aom_lpf_vertical_6_c, 8),
+#endif
+ make_tuple(&aom_lpf_horizontal_16_sse2, &aom_lpf_horizontal_16_c, 8),
+#if !CONFIG_DEBLOCK_13TAP // No SIMD implementation for deblock_13tap yet
+ make_tuple(&aom_lpf_horizontal_16_dual_sse2, &aom_lpf_horizontal_16_dual_c,
+ 8),
+#endif
+ make_tuple(&aom_lpf_vertical_4_sse2, &aom_lpf_vertical_4_c, 8),
+ make_tuple(&aom_lpf_vertical_8_sse2, &aom_lpf_vertical_8_c, 8),
+ make_tuple(&aom_lpf_vertical_16_sse2, &aom_lpf_vertical_16_c, 8),
+#if !CONFIG_DEBLOCK_13TAP
+ make_tuple(&aom_lpf_vertical_16_dual_sse2, &aom_lpf_vertical_16_dual_c, 8)
+#endif
+};
+
+INSTANTIATE_TEST_CASE_P(SSE2, Loop8Test6Param_lbd,
+ ::testing::ValuesIn(kLoop8Test6));
+
#endif // HAVE_SSE2
#if HAVE_AVX2
#if !CONFIG_DEBLOCK_13TAP // No SIMD implementation for deblock_13tap yet
-const loop8_param_t kHbdLoop8Test6Avx2[] = {
+const hbddual_loop_param_t kHbdLoop8Test9Avx2[] = {
make_tuple(&aom_highbd_lpf_horizontal_16_dual_avx2,
&aom_highbd_lpf_horizontal_16_dual_c, 8),
make_tuple(&aom_highbd_lpf_horizontal_16_dual_avx2,
@@ -529,14 +559,24 @@
&aom_highbd_lpf_vertical_16_dual_c, 12)
};
-INSTANTIATE_TEST_CASE_P(AVX2, Loop8Test6Param,
- ::testing::ValuesIn(kHbdLoop8Test6Avx2));
-
+INSTANTIATE_TEST_CASE_P(AVX2, Loop8Test9Param_hbd,
+ ::testing::ValuesIn(kHbdLoop8Test9Avx2));
#endif
#endif
+#if HAVE_AVX2 && (!CONFIG_PARALLEL_DEBLOCKING)
+INSTANTIATE_TEST_CASE_P(AVX2, Loop8Test6Param,
+ ::testing::Values(
+#if !CONFIG_DEBLOCK_13TAP // No SIMD implementation for deblock_13tap yet
+ make_tuple(&aom_lpf_horizontal_16_dual_avx2,
+ &aom_lpf_horizontal_16_dual_c, 8),
+#endif
+ make_tuple(&aom_lpf_horizontal_16_avx2,
+ &aom_lpf_horizontal_16_c, 8)));
+#endif
+
#if HAVE_SSE2
-const dualloop8_param_t kHbdLoop8Test9[] = {
+const hbddual_loop_param_t kHbdLoop8Test9[] = {
make_tuple(&aom_highbd_lpf_horizontal_4_dual_sse2,
&aom_highbd_lpf_horizontal_4_dual_c, 8),
make_tuple(&aom_highbd_lpf_horizontal_8_dual_sse2,
@@ -563,12 +603,24 @@
&aom_highbd_lpf_vertical_8_dual_c, 12)
};
-INSTANTIATE_TEST_CASE_P(SSE2, Loop8Test9Param,
+INSTANTIATE_TEST_CASE_P(SSE2, Loop8Test9Param_hbd,
::testing::ValuesIn(kHbdLoop8Test9));
+
+#if !CONFIG_PARALLEL_DEBLOCKING
+const hbddual_loop_param_t kLoop8Test9[] = {
+ make_tuple(&aom_lpf_horizontal_4_dual_sse2, &aom_lpf_horizontal_4_dual_c, 8),
+ make_tuple(&aom_lpf_horizontal_8_dual_sse2, &aom_lpf_horizontal_8_dual_c, 8),
+ make_tuple(&aom_lpf_vertical_4_dual_sse2, &aom_lpf_vertical_4_dual_c, 8),
+ make_tuple(&aom_lpf_vertical_8_dual_sse2, &aom_lpf_vertical_8_dual_c, 8)
+};
+
+INSTANTIATE_TEST_CASE_P(SSE2, Loop8Test9Param_lbd,
+ ::testing::ValuesIn(kLoop8Test9));
+#endif
#endif // HAVE_SSE2
#if HAVE_AVX2
-const dualloop8_param_t kHbdLoop8Test9Avx2[] = {
+const hbddual_loop_param_t kHbdLoop8Test9Avx2[] = {
make_tuple(&aom_highbd_lpf_horizontal_4_dual_avx2,
&aom_highbd_lpf_horizontal_4_dual_c, 8),
make_tuple(&aom_highbd_lpf_horizontal_4_dual_avx2,
@@ -595,8 +647,7 @@
&aom_highbd_lpf_vertical_8_dual_c, 12),
};
-INSTANTIATE_TEST_CASE_P(AVX2, Loop8Test9Param,
+INSTANTIATE_TEST_CASE_P(AVX2, Loop8Test9Param_hbd,
::testing::ValuesIn(kHbdLoop8Test9Avx2));
#endif
-
} // namespace
