post-proc: deblock filter optimization

1. Algorithm modification: Instead of having same filter threshold for a whole frame, now we allow the thresholds to be adjusted for each macroblock. In current implementation, to avoid excessive blur on background as reported in issue480(http://code.google.com/p/webm/issues/detail?id=480), we reduce the thresholds for skipped macroblocks. 2. SSE2 optimization: As started in issue479(http://code.google.com/p/webm/issues/detail?id=479), the filter calculation was adjusted for better performance. The c code was also modified accordingly. This made the deblock filter 2x faster, and the decoder was 1.2x faster overall. Next, the demacroblock filter will be modified similarly.

Yunqing Wang committed 14 years ago

Commit 4c53bac1 parent 9704cde8 files · +293 −564Browse files

Unified · Split

8 files

test/pp_filter_test.cc +11 −14
vp8/common/postproc.c +123 −86
vp8/common/postproc.h +2 −1
vp8/common/ppc/systemdependent.c +7 −7
vp8/common/rtcd_defs.sh +2 −3
vp8/common/x86/postproc_mmx.asm +0 −265
vp8/common/x86/postproc_sse2.asm +147 −187
vp8/encoder/onyx_if.c +1 −1

test/pp_filter_test.cc+11−14

@@ -19,9 +19,9 @@ typedef void (post_proc_func_t)(unsigned char src_ptr, · 18 unmodified lines
19	19	unsigned char *dst_ptr,
20	20	int src_pixels_per_line,
21	21	int dst_pixels_per_line,
22		− int rows,
23	22	int cols,
24		− int flimit);
	23	+ unsigned char *flimit,
	24	+ int size);
25	25
26	26	namespace {
27	27
@@ -29,7 +29,7 @@ class Vp8PostProcessingFilterTest · 1 unmodified line
29	29	: public ::testing::TestWithParam<post_proc_func_t> {};
30	30
31	31	// Test routine for the VP8 post-processing function
32		−// vp8_post_proc_down_and_across_c.
	32	+// vp8_post_proc_down_and_across_mb_row_c.
33	33
34	34	TEST_P(Vp8PostProcessingFilterTest, FilterOutputCheck) {
35	35	// Size of the underlying data block that will be filtered.
@@ -56,6 +56,8 @@ TEST_P(Vp8PostProcessingFilterTest, FilterOutputCheck) { · 20 unmodified lines
56	56	// Pointers to top-left pixel of block in the input and output images.
57	57	uint8_t *const src_image_ptr = src_image + (input_stride << 1);
58	58	uint8_t *const dst_image_ptr = dst_image + 8;
	59	+ uint8_t const flimits = reinterpret_cast<uint8_t >(vpx_memalign(16, block_width));
	60	+ (void)vpx_memset(flimits, 255, block_width);
59	61
60	62	// Initialize pixels in the input:
61	63	// block pixels to value 1,
@@ -73,14 +75,13 @@ TEST_P(Vp8PostProcessingFilterTest, FilterOutputCheck) { · 11 unmodified lines
73	75	(void)vpx_memset(dst_image, 99, output_size);
74	76
75	77	GetParam()(src_image_ptr, dst_image_ptr, input_stride,
76		− output_stride, block_height, block_width,
77		− 255);
	78	+ output_stride, block_width, flimits, 16);
78	79
79	80	static const uint8_t expected_data[block_height] = {
80		− 3, 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 3
	81	+ 4, 3, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 3, 4
81	82	};
82	83
83		− pixel_ptr = dst_image;
	84	+ pixel_ptr = dst_image_ptr;
84	85	for (int i = 0; i < block_height; ++i) {
85	86	for (int j = 0; j < block_width; ++j) {
86	87	EXPECT_EQ(expected_data[i], pixel_ptr[j])
@@ -91,19 +92,15 @@ TEST_P(Vp8PostProcessingFilterTest, FilterOutputCheck) { · 4 unmodified lines
91	92
92	93	vpx_free(src_image);
93	94	vpx_free(dst_image);
	95	+ vpx_free(flimits);
94	96	};
95	97
96	98	INSTANTIATE_TEST_CASE_P(C, Vp8PostProcessingFilterTest,
97		− ::testing::Values(vp8_post_proc_down_and_across_c));
98		−
99		−#if HAVE_MMX
100		−INSTANTIATE_TEST_CASE_P(MMX, Vp8PostProcessingFilterTest,
101		− ::testing::Values(vp8_post_proc_down_and_across_mmx));
102		−#endif
	99	+ ::testing::Values(vp8_post_proc_down_and_across_mb_row_c));
103	100
104	101	#if HAVE_SSE2
105	102	INSTANTIATE_TEST_CASE_P(SSE2, Vp8PostProcessingFilterTest,
106		− ::testing::Values(vp8_post_proc_down_and_across_xmm));
	103	+ ::testing::Values(vp8_post_proc_down_and_across_mb_row_sse2));
107	104	#endif
108	105
109	106	} // namespace

vp8/common/postproc.c+123−86

@@ -127,25 +127,24 @@ extern void vp8_blit_text(const char msg, unsigned char address, const int pit · 126 unmodified lines
127	127	extern void vp8_blit_line(int x0, int x1, int y0, int y1, unsigned char *image, const int pitch);
128	128	/***********************************************************************************************************
129	129	*/
130		−void vp8_post_proc_down_and_across_c
	130	+void vp8_post_proc_down_and_across_mb_row_c
131	131	(
132	132	unsigned char *src_ptr,
133	133	unsigned char *dst_ptr,
134	134	int src_pixels_per_line,
135	135	int dst_pixels_per_line,
136		− int rows,
137	136	int cols,
138		− int flimit
	137	+ unsigned char *f,
	138	+ int size
139	139	)
140	140	{
141	141	unsigned char p_src, p_dst;
142	142	int row;
143	143	int col;
144		− int i;
145		− int v;
146		− unsigned char d[8];
	144	+ unsigned char v;
	145	+ unsigned char d[4];
147	146
148		− for (row = 0; row < rows; row++)
	147	+ for (row = 0; row < size; row++)
149	148	{
150	149	/* post_proc_down for one row */
151	150	p_src = src_ptr;
@@ -153,20 +152,23 @@ void vp8_post_proc_down_and_across_c · 1 unmodified line
153	152
154	153	for (col = 0; col < cols; col++)
155	154	{
	155	+ unsigned char p_above2 = p_src[col - 2 * src_pixels_per_line];
	156	+ unsigned char p_above1 = p_src[col - src_pixels_per_line];
	157	+ unsigned char p_below1 = p_src[col + src_pixels_per_line];
	158	+ unsigned char p_below2 = p_src[col + 2 * src_pixels_per_line];
156	159
157		− int kernel = 4;
158		− int v = p_src[col];
	160	+ v = p_src[col];
159	161
160		− for (i = -2; i <= 2; i++)
	162	+ if ((abs(v - p_above2) < f[col]) && (abs(v - p_above1) < f[col])
	163	+ && (abs(v - p_below1) < f[col]) && (abs(v - p_below2) < f[col]))
161	164	{
162		− if (abs(v - p_src[col+i*src_pixels_per_line]) > flimit)
163		− goto down_skip_convolve;
164		−
165		− kernel += kernel5[2+i] * p_src[col+i*src_pixels_per_line];
	165	+ unsigned char k1, k2, k3;
	166	+ k1 = (p_above2 + p_above1 + 1) >> 1;
	167	+ k2 = (p_below2 + p_below1 + 1) >> 1;
	168	+ k3 = (k1 + k2 + 1) >> 1;
	169	+ v = (k3 + v + 1) >> 1;
166	170	}
167	171
168		− v = (kernel >> 3);
169		− down_skip_convolve:
170	172	p_dst[col] = v;
171	173	}
172	174
@@ -174,40 +176,34 @@ void vp8_post_proc_down_and_across_c · 1 unmodified line
174	176	p_src = dst_ptr;
175	177	p_dst = dst_ptr;
176	178
177		− for (i = -8; i<0; i++)
178		− p_src[i]=p_src[0];
179		−
180		− for (i = cols; i<cols+8; i++)
181		− p_src[i]=p_src[cols-1];
182		−
183		− for (i = 0; i < 8; i++)
184		− d[i] = p_src[i];
	179	+ p_src[-2] = p_src[-1] = p_src[0];
	180	+ p_src[cols] = p_src[cols + 1] = p_src[cols - 1];
185	181
186	182	for (col = 0; col < cols; col++)
187	183	{
188		− int kernel = 4;
189	184	v = p_src[col];
190	185
191		− d[col&7] = v;
192		−
193		− for (i = -2; i <= 2; i++)
	186	+ if ((abs(v - p_src[col - 2]) < f[col])
	187	+ && (abs(v - p_src[col - 1]) < f[col])
	188	+ && (abs(v - p_src[col + 1]) < f[col])
	189	+ && (abs(v - p_src[col + 2]) < f[col]))
194	190	{
195		− if (abs(v - p_src[col+i]) > flimit)
196		− goto across_skip_convolve;
197		−
198		− kernel += kernel5[2+i] * p_src[col+i];
	191	+ unsigned char k1, k2, k3;
	192	+ k1 = (p_src[col - 2] + p_src[col - 1] + 1) >> 1;
	193	+ k2 = (p_src[col + 2] + p_src[col + 1] + 1) >> 1;
	194	+ k3 = (k1 + k2 + 1) >> 1;
	195	+ v = (k3 + v + 1) >> 1;
199	196	}
200	197
201		− d[col&7] = (kernel >> 3);
202		− across_skip_convolve:
	198	+ d[col & 3] = v;
203	199
204	200	if (col >= 2)
205		− p_dst[col-2] = d[(col-2)&7];
	201	+ p_dst[col - 2] = d[(col - 2) & 3];
206	202	}
207	203
208	204	/* handle the last two pixels */
209		− p_dst[col-2] = d[(col-2)&7];
210		− p_dst[col-1] = d[(col-1)&7];
	205	+ p_dst[col - 2] = d[(col - 2) & 3];
	206	+ p_dst[col - 1] = d[(col - 1) & 3];
211	207
212	208	/* next row */
213	209	src_ptr += src_pixels_per_line;
@@ -318,28 +314,17 @@ void vp8_mbpost_proc_down_c(unsigned char *dst, int pitch, int rows, int cols, i · 104 unmodified lines
318	314	}
319	315	}
320	316
321		−
322		−static void vp8_deblock_and_de_macro_block(YV12_BUFFER_CONFIG *source,
323		− YV12_BUFFER_CONFIG *post,
324		− int q,
325		− int low_var_thresh,
326		− int flag)
	317	+static void vp8_de_mblock(YV12_BUFFER_CONFIG *post,
	318	+ int q)
327	319	{
328		− double level = 6.0e-05 * q * q * q - .0067 * q * q + .306 * q + .0065;
329		− int ppl = (int)(level + .5);
330		− (void) low_var_thresh;
331		− (void) flag;
332		−
333		− vp8_post_proc_down_and_across(source->y_buffer, post->y_buffer, source->y_stride, post->y_stride, source->y_height, source->y_width, ppl);
334		− vp8_mbpost_proc_across_ip(post->y_buffer, post->y_stride, post->y_height, post->y_width, q2mbl(q));
335		− vp8_mbpost_proc_down(post->y_buffer, post->y_stride, post->y_height, post->y_width, q2mbl(q));
336		−
337		− vp8_post_proc_down_and_across(source->u_buffer, post->u_buffer, source->uv_stride, post->uv_stride, source->uv_height, source->uv_width, ppl);
338		− vp8_post_proc_down_and_across(source->v_buffer, post->v_buffer, source->uv_stride, post->uv_stride, source->uv_height, source->uv_width, ppl);
339		−
	320	+ vp8_mbpost_proc_across_ip(post->y_buffer, post->y_stride, post->y_height,
	321	+ post->y_width, q2mbl(q));
	322	+ vp8_mbpost_proc_down(post->y_buffer, post->y_stride, post->y_height,
	323	+ post->y_width, q2mbl(q));
340	324	}
341	325
342		−void vp8_deblock(YV12_BUFFER_CONFIG *source,
	326	+void vp8_deblock(VP8_COMMON *cm,
	327	+ YV12_BUFFER_CONFIG *source,
343	328	YV12_BUFFER_CONFIG *post,
344	329	int q,
345	330	int low_var_thresh,
@@ -347,12 +332,58 @@ void vp8_deblock(YV12_BUFFER_CONFIG *source, · 1 unmodified line
347	332	{
348	333	double level = 6.0e-05 * q * q * q - .0067 * q * q + .306 * q + .0065;
349	334	int ppl = (int)(level + .5);
	335	+
	336	+ const MODE_INFO *mode_info_context = cm->mi;
	337	+ int mbr, mbc;
	338	+
	339	+ /* The pixel thresholds are adjusted according to if or not the macroblock
	340	+ * is a skipped block. */
	341	+ unsigned char ylimits = (unsigned char )vpx_memalign(16, 16 * cm->mb_cols);
	342	+ unsigned char uvlimits = (unsigned char )vpx_memalign(16, 8 * cm->mb_cols);
350	343	(void) low_var_thresh;
351	344	(void) flag;
352	345
353		− vp8_post_proc_down_and_across(source->y_buffer, post->y_buffer, source->y_stride, post->y_stride, source->y_height, source->y_width, ppl);
354		− vp8_post_proc_down_and_across(source->u_buffer, post->u_buffer, source->uv_stride, post->uv_stride, source->uv_height, source->uv_width, ppl);
355		− vp8_post_proc_down_and_across(source->v_buffer, post->v_buffer, source->uv_stride, post->uv_stride, source->uv_height, source->uv_width, ppl);
	346	+ if (ppl > 0)
	347	+ {
	348	+ for (mbr = 0; mbr < cm->mb_rows; mbr++)
	349	+ {
	350	+ unsigned char *ylptr = ylimits;
	351	+ unsigned char *uvlptr = uvlimits;
	352	+ for (mbc = 0; mbc < cm->mb_cols; mbc++)
	353	+ {
	354	+ unsigned char mb_ppl;
	355	+
	356	+ if (mode_info_context->mbmi.mb_skip_coeff)
	357	+ mb_ppl = (unsigned char)ppl >> 1;
	358	+ else
	359	+ mb_ppl = (unsigned char)ppl;
	360	+
	361	+ vpx_memset(ylptr, mb_ppl, 16);
	362	+ vpx_memset(uvlptr, mb_ppl, 8);
	363	+
	364	+ ylptr += 16;
	365	+ uvlptr += 8;
	366	+ mode_info_context++;
	367	+ }
	368	+ mode_info_context++;
	369	+
	370	+ vp8_post_proc_down_and_across_mb_row(
	371	+ source->y_buffer + 16 * mbr * source->y_stride,
	372	+ post->y_buffer + 16 * mbr * post->y_stride, source->y_stride,
	373	+ post->y_stride, source->y_width, ylimits, 16);
	374	+
	375	+ vp8_post_proc_down_and_across_mb_row(
	376	+ source->u_buffer + 8 * mbr * source->uv_stride,
	377	+ post->u_buffer + 8 * mbr * post->uv_stride, source->uv_stride,
	378	+ post->uv_stride, source->uv_width, uvlimits, 8);
	379	+ vp8_post_proc_down_and_across_mb_row(
	380	+ source->v_buffer + 8 * mbr * source->uv_stride,
	381	+ post->v_buffer + 8 * mbr * post->uv_stride, source->uv_stride,
	382	+ post->uv_stride, source->uv_width, uvlimits, 8);
	383	+ }
	384	+ }
	385	+ vpx_free(ylimits);
	386	+ vpx_free(uvlimits);
356	387	}
357	388
358	389	#if !(CONFIG_TEMPORAL_DENOISING)
@@ -364,33 +395,35 @@ void vp8_de_noise(YV12_BUFFER_CONFIG *source, · 5 unmodified lines
364	395	{
365	396	double level = 6.0e-05 * q * q * q - .0067 * q * q + .306 * q + .0065;
366	397	int ppl = (int)(level + .5);
	398	+ int mb_rows = source->y_width >> 4;
	399	+ int mb_cols = source->y_height >> 4;
	400	+ unsigned char limits = (unsigned char )vpx_memalign(16, 16 * mb_cols);
	401	+ int mbr, mbc;
367	402	(void) post;
368	403	(void) low_var_thresh;
369	404	(void) flag;
370	405
371		− vp8_post_proc_down_and_across(
372		− source->y_buffer + 2 * source->y_stride + 2,
373		− source->y_buffer + 2 * source->y_stride + 2,
374		− source->y_stride,
375		− source->y_stride,
376		− source->y_height - 4,
377		− source->y_width - 4,
378		− ppl);
379		− vp8_post_proc_down_and_across(
380		− source->u_buffer + 2 * source->uv_stride + 2,
381		− source->u_buffer + 2 * source->uv_stride + 2,
382		− source->uv_stride,
383		− source->uv_stride,
384		− source->uv_height - 4,
385		− source->uv_width - 4, ppl);
386		− vp8_post_proc_down_and_across(
387		− source->v_buffer + 2 * source->uv_stride + 2,
388		− source->v_buffer + 2 * source->uv_stride + 2,
389		− source->uv_stride,
390		− source->uv_stride,
391		− source->uv_height - 4,
392		− source->uv_width - 4, ppl);
	406	+ /* TODO: The original code don't filter the 2 outer rows and columns. */
	407	+ vpx_memset(limits, (unsigned char)ppl, 16 * mb_cols);
393	408
	409	+ for (mbr = 0; mbr < mb_rows; mbr++)
	410	+ {
	411	+ vp8_post_proc_down_and_across_mb_row(
	412	+ source->y_buffer + 16 * mbr * source->y_stride,
	413	+ source->y_buffer + 16 * mbr * source->y_stride,
	414	+ source->y_stride, source->y_stride, source->y_width, limits, 16);
	415	+
	416	+ vp8_post_proc_down_and_across_mb_row(
	417	+ source->u_buffer + 8 * mbr * source->uv_stride,
	418	+ source->u_buffer + 8 * mbr * source->uv_stride,
	419	+ source->uv_stride, source->uv_stride, source->uv_width, limits, 8);
	420	+ vp8_post_proc_down_and_across_mb_row(
	421	+ source->v_buffer + 8 * mbr * source->uv_stride,
	422	+ source->v_buffer + 8 * mbr * source->uv_stride,
	423	+ source->uv_stride, source->uv_stride, source->uv_width, limits, 8);
	424	+ }
	425	+
	426	+ vpx_free(limits);
394	427	}
395	428	#endif
396	429
@@ -752,12 +785,14 @@ int vp8_post_proc_frame(VP8_COMMON oci, YV12_BUFFER_CONFIG dest, vp8_ppflags_t · 355 unmodified lines
752	785	vp8_yv12_copy_frame(&oci->post_proc_buffer, &oci->post_proc_buffer_int);
753	786	if (flags & VP8D_DEMACROBLOCK)
754	787	{
755		− vp8_deblock_and_de_macro_block(&oci->post_proc_buffer_int, &oci->post_proc_buffer,
	788	+ vp8_deblock(oci, &oci->post_proc_buffer_int, &oci->post_proc_buffer,
756	789	q + (deblock_level - 5) * 10, 1, 0);
	790	+ vp8_de_mblock(&oci->post_proc_buffer,
	791	+ q + (deblock_level - 5) * 10);
757	792	}
758	793	else if (flags & VP8D_DEBLOCK)
759	794	{
760		− vp8_deblock(&oci->post_proc_buffer_int, &oci->post_proc_buffer,
	795	+ vp8_deblock(oci, &oci->post_proc_buffer_int, &oci->post_proc_buffer,
761	796	q, 1, 0);
762	797	}
763	798	}
@@ -766,13 +801,15 @@ int vp8_post_proc_frame(VP8_COMMON oci, YV12_BUFFER_CONFIG dest, vp8_ppflags_t · 2 unmodified lines
766	801	}
767	802	else if (flags & VP8D_DEMACROBLOCK)
768	803	{
769		− vp8_deblock_and_de_macro_block(oci->frame_to_show, &oci->post_proc_buffer,
770		− q + (deblock_level - 5) * 10, 1, 0);
	804	+ vp8_deblock(oci, oci->frame_to_show, &oci->post_proc_buffer,
	805	+ q + (deblock_level - 5) * 10, 1, 0);
	806	+ vp8_de_mblock(&oci->post_proc_buffer, q + (deblock_level - 5) * 10);
	807	+
771	808	oci->postproc_state.last_base_qindex = oci->base_qindex;
772	809	}
773	810	else if (flags & VP8D_DEBLOCK)
774	811	{
775		− vp8_deblock(oci->frame_to_show, &oci->post_proc_buffer,
	812	+ vp8_deblock(oci, oci->frame_to_show, &oci->post_proc_buffer,
776	813	q, 1, 0);
777	814	oci->postproc_state.last_base_qindex = oci->base_qindex;
778	815	}

vp8/common/postproc.h+2−1

@@ -36,7 +36,8 @@ void vp8_de_noise(YV12_BUFFER_CONFIG *source, · 35 unmodified lines
36	36	int low_var_thresh,
37	37	int flag);
38	38
39		−void vp8_deblock(YV12_BUFFER_CONFIG *source,
	39	+void vp8_deblock(struct VP8Common *oci,
	40	+ YV12_BUFFER_CONFIG *source,
40	41	YV12_BUFFER_CONFIG *post,
41	42	int q,
42	43	int low_var_thresh,

vp8/common/ppc/systemdependent.c+7−7

@@ -19,14 +19,14 @@ void (vp8_short_idct4x4)(short input, short *output, int pitch); · 18 unmodified lines
19	19	void (vp8_short_idct4x4_1)(short input, short *output, int pitch);
20	20	void (vp8_dc_only_idct)(short input_dc, short output, int pitch);
21	21
22		−extern void (*vp8_post_proc_down_and_across)(
	22	+extern void (*vp8_post_proc_down_and_across_mb_row)(
23	23	unsigned char *src_ptr,
24	24	unsigned char *dst_ptr,
25	25	int src_pixels_per_line,
26	26	int dst_pixels_per_line,
27		− int rows,
28	27	int cols,
29		− int flimit
	28	+ unsigned char *f,
	29	+ int size
30	30	);
31	31
32	32	extern void (vp8_mbpost_proc_down)(unsigned char dst, int pitch, int rows, int cols, int flimit);
@@ -34,15 +34,15 @@ extern void vp8_mbpost_proc_down_c(unsigned char *dst, int pitch, int rows, int · 1 unmodified line
34	34	extern void (vp8_mbpost_proc_across_ip)(unsigned char src, int pitch, int rows, int cols, int flimit);
35	35	extern void vp8_mbpost_proc_across_ip_c(unsigned char *src, int pitch, int rows, int cols, int flimit);
36	36
37		−extern void vp8_post_proc_down_and_across_c
	37	+extern void vp8_post_proc_down_and_across_mb_row_c
38	38	(
39	39	unsigned char *src_ptr,
40	40	unsigned char *dst_ptr,
41	41	int src_pixels_per_line,
42	42	int dst_pixels_per_line,
43		− int rows,
44	43	int cols,
45		− int flimit
	44	+ unsigned char *f,
	45	+ int size
46	46	);
47	47	void vp8_plane_add_noise_c(unsigned char *Start, unsigned int Width, unsigned int Height, int Pitch, int q, int a);
48	48
@@ -158,7 +158,7 @@ void vp8_machine_specific_config(void) · 109 unmodified lines
158	158	vp8_lf_mbhsimple = loop_filter_mbhs_ppc;
159	159	vp8_lf_bhsimple = loop_filter_bhs_ppc;
160	160
161		− vp8_post_proc_down_and_across = vp8_post_proc_down_and_across_c;
	161	+ vp8_post_proc_down_and_across_mb_row = vp8_post_proc_down_and_across_mb_row_c;
162	162	vp8_mbpost_proc_down = vp8_mbpost_proc_down_c;
163	163	vp8_mbpost_proc_across_ip = vp8_mbpost_proc_across_ip_c;
164	164	vp8_plane_add_noise = vp8_plane_add_noise_c;

vp8/common/rtcd_defs.sh+2−3

@@ -162,9 +162,8 @@ if [ "$CONFIG_POSTPROC" = "yes" ]; then · 161 unmodified lines
162	162	specialize vp8_mbpost_proc_across_ip sse2
163	163	vp8_mbpost_proc_across_ip_sse2=vp8_mbpost_proc_across_ip_xmm
164	164
165		− prototype void vp8_post_proc_down_and_across "unsigned char src, unsigned char dst, int src_pitch, int dst_pitch, int rows, int cols, int flimit"
166		− specialize vp8_post_proc_down_and_across mmx sse2
167		− vp8_post_proc_down_and_across_sse2=vp8_post_proc_down_and_across_xmm
	165	+ prototype void vp8_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"
	166	+ specialize vp8_post_proc_down_and_across_mb_row sse2
168	167
169	168	prototype void vp8_plane_add_noise "unsigned char s, char noise, char blackclamp[16], char whiteclamp[16], char bothclamp[16], unsigned int w, unsigned int h, int pitch"
170	169	specialize vp8_plane_add_noise mmx sse2

vp8/common/x86/postproc_mmx.asm+0−265

@@ -14,271 +14,6 @@ · 13 unmodified lines
14	14	%define VP8_FILTER_WEIGHT 128
15	15	%define VP8_FILTER_SHIFT 7
16	16
17		−;void vp8_post_proc_down_and_across_mmx
18		−;(
19		−; unsigned char *src_ptr,
20		−; unsigned char *dst_ptr,
21		−; int src_pixels_per_line,
22		−; int dst_pixels_per_line,
23		−; int rows,
24		−; int cols,
25		−; int flimit
26		−;)
27		−global sym(vp8_post_proc_down_and_across_mmx) PRIVATE
28		−sym(vp8_post_proc_down_and_across_mmx):
29		− push rbp
30		− mov rbp, rsp
31		− SHADOW_ARGS_TO_STACK 7
32		− GET_GOT rbx
33		− push rsi
34		− push rdi
35		− ; end prolog
36		−
37		−%if ABI_IS_32BIT=1 && CONFIG_PIC=1
38		− ; move the global rd onto the stack, since we don't have enough registers
39		− ; to do PIC addressing
40		− movq mm0, [GLOBAL(rd)]
41		− sub rsp, 8
42		− movq [rsp], mm0
43		−%define RD [rsp]
44		−%else
45		−%define RD [GLOBAL(rd)]
46		−%endif
47		−
48		− push rbx
49		− lea rbx, [GLOBAL(Blur)]
50		− movd mm2, dword ptr arg(6) ;flimit
51		− punpcklwd mm2, mm2
52		− punpckldq mm2, mm2
53		−
54		− mov rsi, arg(0) ;src_ptr
55		− mov rdi, arg(1) ;dst_ptr
56		−
57		− movsxd rcx, DWORD PTR arg(4) ;rows
58		− movsxd rax, DWORD PTR arg(2) ;src_pixels_per_line ; destination pitch?
59		− pxor mm0, mm0 ; mm0 = 00000000
60		−
61		−.nextrow:
62		−
63		− xor rdx, rdx ; clear out rdx for use as loop counter
64		−.nextcol:
65		−
66		− pxor mm7, mm7 ; mm7 = 00000000
67		− movq mm6, [rbx + 32 ] ; mm6 = kernel 2 taps
68		− movq mm3, [rsi] ; mm4 = r0 p0..p7
69		− punpcklbw mm3, mm0 ; mm3 = p0..p3
70		− movq mm1, mm3 ; mm1 = p0..p3
71		− pmullw mm3, mm6 ; mm3 *= kernel 2 modifiers
72		−
73		− movq mm6, [rbx + 48] ; mm6 = kernel 3 taps
74		− movq mm5, [rsi + rax] ; mm4 = r1 p0..p7
75		− punpcklbw mm5, mm0 ; mm5 = r1 p0..p3
76		− pmullw mm6, mm5 ; mm6 = p0..p3 kernel 3 modifiers
77		− paddusw mm3, mm6 ; mm3 += mm6
78		−
79		− ; thresholding
80		− movq mm7, mm1 ; mm7 = r0 p0..p3
81		− psubusw mm7, mm5 ; mm7 = r0 p0..p3 - r1 p0..p3
82		− psubusw mm5, mm1 ; mm5 = r1 p0..p3 - r0 p0..p3
83		− paddusw mm7, mm5 ; mm7 = abs(r0 p0..p3 - r1 p0..p3)
84		− pcmpgtw mm7, mm2
85		−
86		− movq mm6, [rbx + 64 ] ; mm6 = kernel 4 modifiers
87		− movq mm5, [rsi + 2*rax] ; mm4 = r2 p0..p7
88		− punpcklbw mm5, mm0 ; mm5 = r2 p0..p3
89		− pmullw mm6, mm5 ; mm5 *= kernel 4 modifiers
90		− paddusw mm3, mm6 ; mm3 += mm5
91		−
92		− ; thresholding
93		− movq mm6, mm1 ; mm6 = r0 p0..p3
94		− psubusw mm6, mm5 ; mm6 = r0 p0..p3 - r2 p0..p3
95		− psubusw mm5, mm1 ; mm5 = r2 p0..p3 - r2 p0..p3
96		− paddusw mm6, mm5 ; mm6 = abs(r0 p0..p3 - r2 p0..p3)
97		− pcmpgtw mm6, mm2
98		− por mm7, mm6 ; accumulate thresholds
99		−
100		−
101		− neg rax
102		− movq mm6, [rbx ] ; kernel 0 taps
103		− movq mm5, [rsi+2*rax] ; mm4 = r-2 p0..p7
104		− punpcklbw mm5, mm0 ; mm5 = r-2 p0..p3
105		− pmullw mm6, mm5 ; mm5 *= kernel 0 modifiers
106		− paddusw mm3, mm6 ; mm3 += mm5
107		−
108		− ; thresholding
109		− movq mm6, mm1 ; mm6 = r0 p0..p3
110		− psubusw mm6, mm5 ; mm6 = p0..p3 - r-2 p0..p3
111		− psubusw mm5, mm1 ; mm5 = r-2 p0..p3 - p0..p3
112		− paddusw mm6, mm5 ; mm6 = abs(r0 p0..p3 - r-2 p0..p3)
113		− pcmpgtw mm6, mm2
114		− por mm7, mm6 ; accumulate thresholds
115		−
116		− movq mm6, [rbx + 16] ; kernel 1 taps
117		− movq mm4, [rsi+rax] ; mm4 = r-1 p0..p7
118		− punpcklbw mm4, mm0 ; mm4 = r-1 p0..p3
119		− pmullw mm6, mm4 ; mm4 *= kernel 1 modifiers.
120		− paddusw mm3, mm6 ; mm3 += mm5
121		−
122		− ; thresholding
123		− movq mm6, mm1 ; mm6 = r0 p0..p3
124		− psubusw mm6, mm4 ; mm6 = p0..p3 - r-2 p0..p3
125		− psubusw mm4, mm1 ; mm5 = r-1 p0..p3 - p0..p3
126		− paddusw mm6, mm4 ; mm6 = abs(r0 p0..p3 - r-1 p0..p3)
127		− pcmpgtw mm6, mm2
128		− por mm7, mm6 ; accumulate thresholds
129		−
130		−
131		− paddusw mm3, RD ; mm3 += round value
132		− psraw mm3, VP8_FILTER_SHIFT ; mm3 /= 128
133		−
134		− pand mm1, mm7 ; mm1 select vals > thresh from source
135		− pandn mm7, mm3 ; mm7 select vals < thresh from blurred result
136		− paddusw mm1, mm7 ; combination
137		−
138		− packuswb mm1, mm0 ; pack to bytes
139		−
140		− movd [rdi], mm1 ;
141		− neg rax ; pitch is positive
142		−
143		−
144		− add rsi, 4
145		− add rdi, 4
146		− add rdx, 4
147		−
148		− cmp edx, dword ptr arg(5) ;cols
149		− jl .nextcol
150		− ; done with the all cols, start the across filtering in place
151		− sub rsi, rdx
152		− sub rdi, rdx
153		−
154		− ; dup the first byte into the left border 8 times
155		− movq mm1, [rdi]
156		− punpcklbw mm1, mm1
157		− punpcklwd mm1, mm1
158		− punpckldq mm1, mm1
159		−
160		− mov rdx, -8
161		− movq [rdi+rdx], mm1
162		−
163		− ; dup the last byte into the right border
164		− movsxd rdx, dword arg(5)
165		− movq mm1, [rdi + rdx + -1]
166		− punpcklbw mm1, mm1
167		− punpcklwd mm1, mm1
168		− punpckldq mm1, mm1
169		− movq [rdi+rdx], mm1
170		−
171		−
172		− push rax
173		− xor rdx, rdx
174		− mov rax, [rdi-4];
175		−
176		−.acrossnextcol:
177		− pxor mm7, mm7 ; mm7 = 00000000
178		− movq mm6, [rbx + 32 ] ;
179		− movq mm4, [rdi+rdx] ; mm4 = p0..p7
180		− movq mm3, mm4 ; mm3 = p0..p7
181		− punpcklbw mm3, mm0 ; mm3 = p0..p3
182		− movq mm1, mm3 ; mm1 = p0..p3
183		− pmullw mm3, mm6 ; mm3 *= kernel 2 modifiers
184		−
185		− movq mm6, [rbx + 48]
186		− psrlq mm4, 8 ; mm4 = p1..p7
187		− movq mm5, mm4 ; mm5 = p1..p7
188		− punpcklbw mm5, mm0 ; mm5 = p1..p4
189		− pmullw mm6, mm5 ; mm6 = p1..p4 kernel 3 modifiers
190		− paddusw mm3, mm6 ; mm3 += mm6
191		−
192		− ; thresholding
193		− movq mm7, mm1 ; mm7 = p0..p3
194		− psubusw mm7, mm5 ; mm7 = p0..p3 - p1..p4
195		− psubusw mm5, mm1 ; mm5 = p1..p4 - p0..p3
196		− paddusw mm7, mm5 ; mm7 = abs(p0..p3 - p1..p4)
197		− pcmpgtw mm7, mm2
198		−
199		− movq mm6, [rbx + 64 ]
200		− psrlq mm4, 8 ; mm4 = p2..p7
201		− movq mm5, mm4 ; mm5 = p2..p7
202		− punpcklbw mm5, mm0 ; mm5 = p2..p5
203		− pmullw mm6, mm5 ; mm5 *= kernel 4 modifiers
204		− paddusw mm3, mm6 ; mm3 += mm5
205		−
206		− ; thresholding
207		− movq mm6, mm1 ; mm6 = p0..p3
208		− psubusw mm6, mm5 ; mm6 = p0..p3 - p1..p4
209		− psubusw mm5, mm1 ; mm5 = p1..p4 - p0..p3
210		− paddusw mm6, mm5 ; mm6 = abs(p0..p3 - p1..p4)
211		− pcmpgtw mm6, mm2
212		− por mm7, mm6 ; accumulate thresholds
213		−
214		−
215		− movq mm6, [rbx ]
216		− movq mm4, [rdi+rdx-2] ; mm4 = p-2..p5
217		− movq mm5, mm4 ; mm5 = p-2..p5
218		− punpcklbw mm5, mm0 ; mm5 = p-2..p1
219		− pmullw mm6, mm5 ; mm5 *= kernel 0 modifiers
220		− paddusw mm3, mm6 ; mm3 += mm5
221		−
222		− ; thresholding
223		− movq mm6, mm1 ; mm6 = p0..p3
224		− psubusw mm6, mm5 ; mm6 = p0..p3 - p1..p4
225		− psubusw mm5, mm1 ; mm5 = p1..p4 - p0..p3
226		− paddusw mm6, mm5 ; mm6 = abs(p0..p3 - p1..p4)
227		− pcmpgtw mm6, mm2
228		− por mm7, mm6 ; accumulate thresholds
229		−
230		− movq mm6, [rbx + 16]
231		− psrlq mm4, 8 ; mm4 = p-1..p5
232		− punpcklbw mm4, mm0 ; mm4 = p-1..p2
233		− pmullw mm6, mm4 ; mm4 *= kernel 1 modifiers.
234		− paddusw mm3, mm6 ; mm3 += mm5
235		−
236		− ; thresholding
237		− movq mm6, mm1 ; mm6 = p0..p3
238		− psubusw mm6, mm4 ; mm6 = p0..p3 - p1..p4
239		− psubusw mm4, mm1 ; mm5 = p1..p4 - p0..p3
240		− paddusw mm6, mm4 ; mm6 = abs(p0..p3 - p1..p4)
241		− pcmpgtw mm6, mm2
242		− por mm7, mm6 ; accumulate thresholds
243		−
244		− paddusw mm3, RD ; mm3 += round value
245		− psraw mm3, VP8_FILTER_SHIFT ; mm3 /= 128
246		−
247		− pand mm1, mm7 ; mm1 select vals > thresh from source
248		− pandn mm7, mm3 ; mm7 select vals < thresh from blurred result
249		− paddusw mm1, mm7 ; combination
250		−
251		− packuswb mm1, mm0 ; pack to bytes
252		− mov DWORD PTR [rdi+rdx-4], eax ; store previous four bytes
253		− movd eax, mm1
254		−
255		− add rdx, 4
256		− cmp edx, dword ptr arg(5) ;cols
257		− jl .acrossnextcol;
258		−
259		− mov DWORD PTR [rdi+rdx-4], eax
260		− pop rax
261		−
262		− ; done with this rwo
263		− add rsi,rax ; next line
264		− movsxd rax, dword ptr arg(3) ;dst_pixels_per_line ; destination pitch?
265		− add rdi,rax ; next destination
266		− movsxd rax, dword ptr arg(2) ;src_pixels_per_line ; destination pitch?
267		−
268		− dec rcx ; decrement count
269		− jnz .nextrow ; next row
270		− pop rbx
271		−
272		− ; begin epilog
273		− pop rdi
274		− pop rsi
275		− RESTORE_GOT
276		− UNSHADOW_ARGS
277		− pop rbp
278		− ret
279		−%undef RD
280		−
281		−
282	17	;void vp8_mbpost_proc_down_mmx(unsigned char *dst,
283	18	; int pitch, int rows, int cols,int flimit)
284	19	extern sym(vp8_rv)

vp8/common/x86/postproc_sse2.asm+147−187

@@ -11,146 +11,158 @@ · 10 unmodified lines
11	11
12	12	%include "vpx_ports/x86_abi_support.asm"
13	13
14		−;void vp8_post_proc_down_and_across_xmm
	14	+;macro in deblock functions
	15	+%macro FIRST_2_ROWS 0
	16	+ movdqa xmm4, xmm0
	17	+ movdqa xmm6, xmm0
	18	+ movdqa xmm5, xmm1
	19	+ pavgb xmm5, xmm3
	20	+
	21	+ ;calculate absolute value
	22	+ psubusb xmm4, xmm1
	23	+ psubusb xmm1, xmm0
	24	+ psubusb xmm6, xmm3
	25	+ psubusb xmm3, xmm0
	26	+ paddusb xmm4, xmm1
	27	+ paddusb xmm6, xmm3
	28	+
	29	+ ;get threshold
	30	+ movdqa xmm2, flimit
	31	+ pxor xmm1, xmm1
	32	+ movdqa xmm7, xmm2
	33	+
	34	+ ;get mask
	35	+ psubusb xmm2, xmm4
	36	+ psubusb xmm7, xmm6
	37	+ pcmpeqb xmm2, xmm1
	38	+ pcmpeqb xmm7, xmm1
	39	+ por xmm7, xmm2
	40	+%endmacro
	41	+
	42	+%macro SECOND_2_ROWS 0
	43	+ movdqa xmm6, xmm0
	44	+ movdqa xmm4, xmm0
	45	+ movdqa xmm2, xmm1
	46	+ pavgb xmm1, xmm3
	47	+
	48	+ ;calculate absolute value
	49	+ psubusb xmm6, xmm2
	50	+ psubusb xmm2, xmm0
	51	+ psubusb xmm4, xmm3
	52	+ psubusb xmm3, xmm0
	53	+ paddusb xmm6, xmm2
	54	+ paddusb xmm4, xmm3
	55	+
	56	+ pavgb xmm5, xmm1
	57	+
	58	+ ;get threshold
	59	+ movdqa xmm2, flimit
	60	+ pxor xmm1, xmm1
	61	+ movdqa xmm3, xmm2
	62	+
	63	+ ;get mask
	64	+ psubusb xmm2, xmm6
	65	+ psubusb xmm3, xmm4
	66	+ pcmpeqb xmm2, xmm1
	67	+ pcmpeqb xmm3, xmm1
	68	+
	69	+ por xmm7, xmm2
	70	+ por xmm7, xmm3
	71	+
	72	+ pavgb xmm5, xmm0
	73	+
	74	+ ;decide if or not to use filtered value
	75	+ pand xmm0, xmm7
	76	+ pandn xmm7, xmm5
	77	+ paddusb xmm0, xmm7
	78	+%endmacro
	79	+
	80	+%macro UPDATE_FLIMIT 0
	81	+ movdqa xmm2, XMMWORD PTR [rbx]
	82	+ movdqa [rsp], xmm2
	83	+ add rbx, 16
	84	+%endmacro
	85	+
	86	+;void vp8_post_proc_down_and_across_mb_row_sse2
15	87	;(
16	88	; unsigned char *src_ptr,
17	89	; unsigned char *dst_ptr,
18	90	; int src_pixels_per_line,
19	91	; int dst_pixels_per_line,
20		−; int rows,
21	92	; int cols,
22		−; int flimit
	93	+; int *flimits,
	94	+; int size
23	95	;)
24		−global sym(vp8_post_proc_down_and_across_xmm) PRIVATE
25		−sym(vp8_post_proc_down_and_across_xmm):
	96	+global sym(vp8_post_proc_down_and_across_mb_row_sse2) PRIVATE
	97	+sym(vp8_post_proc_down_and_across_mb_row_sse2):
26	98	push rbp
27	99	mov rbp, rsp
28	100	SHADOW_ARGS_TO_STACK 7
29	101	SAVE_XMM 7
30		− GET_GOT rbx
	102	+ push rbx
31	103	push rsi
32	104	push rdi
33	105	; end prolog
34		−
35		−%if ABI_IS_32BIT=1 && CONFIG_PIC=1
36	106	ALIGN_STACK 16, rax
37		− ; move the global rd onto the stack, since we don't have enough registers
38		− ; to do PIC addressing
39		− movdqa xmm0, [GLOBAL(rd42)]
40	107	sub rsp, 16
41		− movdqa [rsp], xmm0
42		−%define RD42 [rsp]
43		−%else
44		−%define RD42 [GLOBAL(rd42)]
45		−%endif
46		−
47	108
48		− movd xmm2, dword ptr arg(6) ;flimit
49		− punpcklwd xmm2, xmm2
50		− punpckldq xmm2, xmm2
51		− punpcklqdq xmm2, xmm2
	109	+ ; put flimit on stack
	110	+ mov rbx, arg(5) ;flimits ptr
	111	+ UPDATE_FLIMIT
52	112
53		− mov rsi, arg(0) ;src_ptr
54		− mov rdi, arg(1) ;dst_ptr
	113	+%define flimit [rsp]
55	114
56		− movsxd rcx, DWORD PTR arg(4) ;rows
57		− movsxd rax, DWORD PTR arg(2) ;src_pixels_per_line ; destination pitch?
58		− pxor xmm0, xmm0 ; mm0 = 00000000
	115	+ mov rsi, arg(0) ;src_ptr
	116	+ mov rdi, arg(1) ;dst_ptr
59	117
	118	+ movsxd rax, DWORD PTR arg(2) ;src_pixels_per_line
	119	+ movsxd rcx, DWORD PTR arg(6) ;rows in a macroblock
60	120	.nextrow:
61		−
62		− xor rdx, rdx ; clear out rdx for use as loop counter
	121	+ xor rdx, rdx ;col
63	122	.nextcol:
64		− movq xmm3, QWORD PTR [rsi] ; mm4 = r0 p0..p7
65		− punpcklbw xmm3, xmm0 ; mm3 = p0..p3
66		− movdqa xmm1, xmm3 ; mm1 = p0..p3
67		− psllw xmm3, 2 ;
68		−
69		− movq xmm5, QWORD PTR [rsi + rax] ; mm4 = r1 p0..p7
70		− punpcklbw xmm5, xmm0 ; mm5 = r1 p0..p3
71		− paddusw xmm3, xmm5 ; mm3 += mm6
72		−
73		− ; thresholding
74		− movdqa xmm7, xmm1 ; mm7 = r0 p0..p3
75		− psubusw xmm7, xmm5 ; mm7 = r0 p0..p3 - r1 p0..p3
76		− psubusw xmm5, xmm1 ; mm5 = r1 p0..p3 - r0 p0..p3
77		− paddusw xmm7, xmm5 ; mm7 = abs(r0 p0..p3 - r1 p0..p3)
78		− pcmpgtw xmm7, xmm2
79		−
80		− movq xmm5, QWORD PTR [rsi + 2*rax] ; mm4 = r2 p0..p7
81		− punpcklbw xmm5, xmm0 ; mm5 = r2 p0..p3
82		− paddusw xmm3, xmm5 ; mm3 += mm5
83		−
84		− ; thresholding
85		− movdqa xmm6, xmm1 ; mm6 = r0 p0..p3
86		− psubusw xmm6, xmm5 ; mm6 = r0 p0..p3 - r2 p0..p3
87		− psubusw xmm5, xmm1 ; mm5 = r2 p0..p3 - r2 p0..p3
88		− paddusw xmm6, xmm5 ; mm6 = abs(r0 p0..p3 - r2 p0..p3)
89		− pcmpgtw xmm6, xmm2
90		− por xmm7, xmm6 ; accumulate thresholds
	123	+ ;load current and next 2 rows
	124	+ movdqu xmm0, XMMWORD PTR [rsi]
	125	+ movdqu xmm1, XMMWORD PTR [rsi + rax]
	126	+ movdqu xmm3, XMMWORD PTR [rsi + 2*rax]
91	127
	128	+ FIRST_2_ROWS
92	129
	130	+ ;load above 2 rows
93	131	neg rax
94		− movq xmm5, QWORD PTR [rsi+2*rax] ; mm4 = r-2 p0..p7
95		− punpcklbw xmm5, xmm0 ; mm5 = r-2 p0..p3
96		− paddusw xmm3, xmm5 ; mm3 += mm5
97		−
98		− ; thresholding
99		− movdqa xmm6, xmm1 ; mm6 = r0 p0..p3
100		− psubusw xmm6, xmm5 ; mm6 = p0..p3 - r-2 p0..p3
101		− psubusw xmm5, xmm1 ; mm5 = r-2 p0..p3 - p0..p3
102		− paddusw xmm6, xmm5 ; mm6 = abs(r0 p0..p3 - r-2 p0..p3)
103		− pcmpgtw xmm6, xmm2
104		− por xmm7, xmm6 ; accumulate thresholds
105		−
106		− movq xmm4, QWORD PTR [rsi+rax] ; mm4 = r-1 p0..p7
107		− punpcklbw xmm4, xmm0 ; mm4 = r-1 p0..p3
108		− paddusw xmm3, xmm4 ; mm3 += mm5
109		−
110		− ; thresholding
111		− movdqa xmm6, xmm1 ; mm6 = r0 p0..p3
112		− psubusw xmm6, xmm4 ; mm6 = p0..p3 - r-2 p0..p3
113		− psubusw xmm4, xmm1 ; mm5 = r-1 p0..p3 - p0..p3
114		− paddusw xmm6, xmm4 ; mm6 = abs(r0 p0..p3 - r-1 p0..p3)
115		− pcmpgtw xmm6, xmm2
116		− por xmm7, xmm6 ; accumulate thresholds
117		−
118		−
119		− paddusw xmm3, RD42 ; mm3 += round value
120		− psraw xmm3, 3 ; mm3 /= 8
	132	+ movdqu xmm1, XMMWORD PTR [rsi + 2*rax]
	133	+ movdqu xmm3, XMMWORD PTR [rsi + rax]
121	134
122		− pand xmm1, xmm7 ; mm1 select vals > thresh from source
123		− pandn xmm7, xmm3 ; mm7 select vals < thresh from blurred result
124		− paddusw xmm1, xmm7 ; combination
	135	+ SECOND_2_ROWS
125	136
126		− packuswb xmm1, xmm0 ; pack to bytes
127		− movq QWORD PTR [rdi], xmm1 ;
	137	+ movdqu XMMWORD PTR [rdi], xmm0
128	138
129		− neg rax ; pitch is positive
130		− add rsi, 8
131		− add rdi, 8
	139	+ neg rax ; positive stride
	140	+ add rsi, 16
	141	+ add rdi, 16
132	142
133		− add rdx, 8
134		− cmp edx, dword arg(5) ;cols
	143	+ UPDATE_FLIMIT
135	144
	145	+ add rdx, 16
	146	+ cmp edx, dword arg(4) ;cols
136	147	jl .nextcol
137	148
138	149	; done with the all cols, start the across filtering in place
139	150	sub rsi, rdx
140	151	sub rdi, rdx
141	152
	153	+ mov rbx, arg(5) ; flimits
	154	+ UPDATE_FLIMIT
142	155
143	156	; dup the first byte into the left border 8 times
144	157	movq mm1, [rdi]
145	158	punpcklbw mm1, mm1
146	159	punpcklwd mm1, mm1
147	160	punpckldq mm1, mm1
148		−
149	161	mov rdx, -8
150	162	movq [rdi+rdx], mm1
151	163
152	164	; dup the last byte into the right border
153		− movsxd rdx, dword arg(5)
	165	+ movsxd rdx, dword arg(4)
154	166	movq mm1, [rdi + rdx + -1]
155	167	punpcklbw mm1, mm1
156	168	punpcklwd mm1, mm1
@@ -158,113 +170,63 @@ sym(vp8_post_proc_down_and_across_xmm): · 1 unmodified line
158	170	movq [rdi+rdx], mm1
159	171
160	172	xor rdx, rdx
161		− movq mm0, QWORD PTR [rdi-8];
	173	+ movq mm0, QWORD PTR [rdi-16];
	174	+ movq mm1, QWORD PTR [rdi-8];
162	175
163	176	.acrossnextcol:
164		− movq xmm7, QWORD PTR [rdi +rdx -2]
165		− movd xmm4, DWORD PTR [rdi +rdx +6]
166		−
167		− pslldq xmm4, 8
168		− por xmm4, xmm7
169		−
170		− movdqa xmm3, xmm4
171		− psrldq xmm3, 2
172		− punpcklbw xmm3, xmm0 ; mm3 = p0..p3
173		− movdqa xmm1, xmm3 ; mm1 = p0..p3
174		− psllw xmm3, 2
175		−
176		−
177		− movdqa xmm5, xmm4
178		− psrldq xmm5, 3
179		− punpcklbw xmm5, xmm0 ; mm5 = p1..p4
180		− paddusw xmm3, xmm5 ; mm3 += mm6
181		−
182		− ; thresholding
183		− movdqa xmm7, xmm1 ; mm7 = p0..p3
184		− psubusw xmm7, xmm5 ; mm7 = p0..p3 - p1..p4
185		− psubusw xmm5, xmm1 ; mm5 = p1..p4 - p0..p3
186		− paddusw xmm7, xmm5 ; mm7 = abs(p0..p3 - p1..p4)
187		− pcmpgtw xmm7, xmm2
188		−
189		− movdqa xmm5, xmm4
190		− psrldq xmm5, 4
191		− punpcklbw xmm5, xmm0 ; mm5 = p2..p5
192		− paddusw xmm3, xmm5 ; mm3 += mm5
193		−
194		− ; thresholding
195		− movdqa xmm6, xmm1 ; mm6 = p0..p3
196		− psubusw xmm6, xmm5 ; mm6 = p0..p3 - p1..p4
197		− psubusw xmm5, xmm1 ; mm5 = p1..p4 - p0..p3
198		− paddusw xmm6, xmm5 ; mm6 = abs(p0..p3 - p1..p4)
199		− pcmpgtw xmm6, xmm2
200		− por xmm7, xmm6 ; accumulate thresholds
201		−
202		−
203		− movdqa xmm5, xmm4 ; mm5 = p-2..p5
204		− punpcklbw xmm5, xmm0 ; mm5 = p-2..p1
205		− paddusw xmm3, xmm5 ; mm3 += mm5
206		−
207		− ; thresholding
208		− movdqa xmm6, xmm1 ; mm6 = p0..p3
209		− psubusw xmm6, xmm5 ; mm6 = p0..p3 - p1..p4
210		− psubusw xmm5, xmm1 ; mm5 = p1..p4 - p0..p3
211		− paddusw xmm6, xmm5 ; mm6 = abs(p0..p3 - p1..p4)
212		− pcmpgtw xmm6, xmm2
213		− por xmm7, xmm6 ; accumulate thresholds
214		−
215		− psrldq xmm4, 1 ; mm4 = p-1..p5
216		− punpcklbw xmm4, xmm0 ; mm4 = p-1..p2
217		− paddusw xmm3, xmm4 ; mm3 += mm5
218		−
219		− ; thresholding
220		− movdqa xmm6, xmm1 ; mm6 = p0..p3
221		− psubusw xmm6, xmm4 ; mm6 = p0..p3 - p1..p4
222		− psubusw xmm4, xmm1 ; mm5 = p1..p4 - p0..p3
223		− paddusw xmm6, xmm4 ; mm6 = abs(p0..p3 - p1..p4)
224		− pcmpgtw xmm6, xmm2
225		− por xmm7, xmm6 ; accumulate thresholds
226		−
227		− paddusw xmm3, RD42 ; mm3 += round value
228		− psraw xmm3, 3 ; mm3 /= 8
229		−
230		− pand xmm1, xmm7 ; mm1 select vals > thresh from source
231		− pandn xmm7, xmm3 ; mm7 select vals < thresh from blurred result
232		− paddusw xmm1, xmm7 ; combination
233		−
234		− packuswb xmm1, xmm0 ; pack to bytes
235		− movq QWORD PTR [rdi+rdx-8], mm0 ; store previous four bytes
236		− movdq2q mm0, xmm1
237		−
238		− add rdx, 8
239		− cmp edx, dword arg(5) ;cols
	177	+ movdqu xmm0, XMMWORD PTR [rdi + rdx]
	178	+ movdqu xmm1, XMMWORD PTR [rdi + rdx -2]
	179	+ movdqu xmm3, XMMWORD PTR [rdi + rdx -1]
	180	+
	181	+ FIRST_2_ROWS
	182	+
	183	+ movdqu xmm1, XMMWORD PTR [rdi + rdx +1]
	184	+ movdqu xmm3, XMMWORD PTR [rdi + rdx +2]
	185	+
	186	+ SECOND_2_ROWS
	187	+
	188	+ movq QWORD PTR [rdi+rdx-16], mm0 ; store previous 8 bytes
	189	+ movq QWORD PTR [rdi+rdx-8], mm1 ; store previous 8 bytes
	190	+ movdq2q mm0, xmm0
	191	+ psrldq xmm0, 8
	192	+ movdq2q mm1, xmm0
	193	+
	194	+ UPDATE_FLIMIT
	195	+
	196	+ add rdx, 16
	197	+ cmp edx, dword arg(4) ;cols
240	198	jl .acrossnextcol;
241	199
242		− ; last 8 pixels
243		− movq QWORD PTR [rdi+rdx-8], mm0
	200	+ ; last 16 pixels
	201	+ movq QWORD PTR [rdi+rdx-16], mm0
244	202
	203	+ cmp edx, dword arg(4)
	204	+ jne .throw_last_8
	205	+ movq QWORD PTR [rdi+rdx-8], mm1
	206	+.throw_last_8:
245	207	; done with this rwo
246		− add rsi,rax ; next line
247		− mov eax, dword arg(3) ;dst_pixels_per_line ; destination pitch?
248		− add rdi,rax ; next destination
249		− mov eax, dword arg(2) ;src_pixels_per_line ; destination pitch?
	208	+ add rsi,rax ;next src line
	209	+ mov eax, dword arg(3) ;dst_pixels_per_line
	210	+ add rdi,rax ;next destination
	211	+ mov eax, dword arg(2) ;src_pixels_per_line
250	212
251		− dec rcx ; decrement count
252		− jnz .nextrow ; next row
	213	+ mov rbx, arg(5) ;flimits
	214	+ UPDATE_FLIMIT
253	215
254		−%if ABI_IS_32BIT=1 && CONFIG_PIC=1
255		− add rsp,16
	216	+ dec rcx ;decrement count
	217	+ jnz .nextrow ;next row
	218	+
	219	+ add rsp, 16
256	220	pop rsp
257		−%endif
258	221	; begin epilog
259	222	pop rdi
260	223	pop rsi
261		− RESTORE_GOT
	224	+ pop rbx
262	225	RESTORE_XMM
263	226	UNSHADOW_ARGS
264	227	pop rbp
265	228	ret
266		−%undef RD42
267		−
	229	+%undef flimit
268	230
269	231	;void vp8_mbpost_proc_down_xmm(unsigned char *dst,
270	232	; int pitch, int rows, int cols,int flimit)
@@ -753,7 +715,5 @@ sym(vp8_plane_add_noise_wmt): · 482 unmodified lines
753	715
754	716	SECTION_RODATA
755	717	align 16
756		−rd42:
757		− times 8 dw 0x04
758	718	four8s:
759	719	times 4 dd 8

vp8/encoder/onyx_if.c+1−1

@@ -5301,7 +5301,7 @@ int vp8_get_compressed_data(VP8_COMP cpi, unsigned int frame_flags, unsigned l · 5300 unmodified lines
5301	5301	double frame_psnr2, frame_ssim2 = 0;
5302	5302	double weight = 0;
5303	5303
5304		− vp8_deblock(cm->frame_to_show, &cm->post_proc_buffer, cm->filter_level * 10 / 6, 1, 0);
	5304	+ vp8_deblock(cm, cm->frame_to_show, &cm->post_proc_buffer, cm->filter_level * 10 / 6, 1, 0);
5305	5305	vp8_clear_system_state();
5306	5306
5307	5307	ye = calc_plane_error(orig->y_buffer, orig->y_stride,

post-proc: deblock filter optimization

Keyboard shortcuts

General

Issues and pull requests

Issues, pull requests and Inbox

Repository code