[Git][freetype/freetype][gsoc-anurag-2022] 3 commits: lkw

[Anurag Thakur (@AdbhutDev)]

Anurag Thakur pushed to branch gsoc-anurag-2022 at FreeType / FreeType

Commits:

• c03591c4
  by Anurag Thakur at 2022-10-06T01:55:06+05:30

  lkw
  

• 9538d8e7
  by Anurag Thakur at 2022-10-06T03:06:00+05:30

  Fixed first pass
  

• d2574056
  by Anurag Thakur at 2022-10-09T18:20:49+05:30

  Fixed point working now
  

2 changed files:

• src/dense/ftdense.c
• src/dense/ftdense.h

Changes:

src/dense/ftdense.c

---

...	...	@@ -58,6 +58,19 @@ dense_line_to( const FT_Vector* to, dense_worker* worker )
58	58	dense_move_to( to, worker );
59	59	return 0;
60	60	}
	61	+
	62	+FT26D6 max(FT26D6 x, FT26D6 y){
	63	+ if(x > y)
	64	+ return x;
	65	+ return y;
	66	+}
	67	+
	68	+FT26D6 min(FT26D6 x, FT26D6 y){
	69	+ if(x < y)
	70	+ return x;
	71	+ return y;
	72	+}
	73	+
61	74	void
62	75	swap( long int* a, long int* b )
63	76	{
...	...	@@ -72,33 +85,39 @@ dense_render_line( dense_worker* worker, TPos tox, TPos toy )
72	85	{
73	86	// printf("line from: %d, %d to %d, %d\n", worker->prev_x, worker->prev_y,
74	87	// to_x, to_y);
75		- float from_x = worker->prev_x;
76		- float from_y = worker->prev_y;
77		- if ( from_y == toy )
78		- return;
79	88
80		- // aP0.m_x -= worker->m_origin_x;
81		- // aP0.m_y -= worker->m_origin_y;
82		- // aP1.m_x -= worker->m_origin_x;
83		- // aP1.m_y -= worker->m_origin_y;
84	89
85		- // from_x = TRUNC( (int)from_x );
86		- // from_y = TRUNC( (int)from_y );
87		- // to_x = TRUNC( (int)to_x );
88		- // to_y = TRUNC( (int)to_y );
	90	+ FT26D6 fx = worker->prev_x>>2;
	91	+ FT26D6 fy = worker->prev_y>>2;
	92	+
	93	+
	94	+ // float from_x = fx;
	95	+ // float from_y = fy;
	96	+
	97	+ FT26D6 from_x = fx;
	98	+ FT26D6 from_y = fy;
	99	+
	100	+
	101	+ // from_x /= 64.0;
	102	+ // from_y /= 64.0;
	103	+
	104	+ FT26D6 tx = tox>>2;
	105	+ FT26D6 ty = toy>>2;
	106	+
	107	+ if ( fy == ty )
	108	+ return;
	109	+
	110	+ // float to_x = tx / 64.0;
	111	+ // float to_y = ty / 64.0;
	112	+ FT26D6 to_x = tx;
	113	+ FT26D6 to_y = ty;
89	114
90		- from_x /= 256.0;
91		- from_y /= 256.0;
92		- float to_x = tox / 256.0;
93		- float to_y = toy / 256.0;
94	115
95	116
96	117	//printf("line from: %f, %f to %f, %f\n", from_x, from_y, to_x, to_y);
97	118
98		- float dir;
99		- if ( from_y < to_y )
100		- dir = 1;
101		- else
	119	+ int dir = 1;
	120	+ if ( from_y >= to_y )
102	121	{
103	122	dir = -1;
104	123	swap( &from_x, &to_x );
...	...	@@ -106,35 +125,41 @@ dense_render_line( dense_worker* worker, TPos tox, TPos toy )
106	125	}
107	126
108	127	// Clip to the height.
109		- if ( from_y >= worker->m_h || to_y <= 0 )
	128	+ if ( from_y >= worker->m_h<<6 || to_y <= 0 )
110	129	return;
111	130
112		- float dxdy = ( to_x - from_x ) / (float)( to_y - from_y );
	131	+ //float dxdy = ( to_x - from_x ) / (float)( to_y - from_y );
	132	+ FT26D6 deltax,deltay;
	133	+ deltax = to_x - from_x;
	134	+ deltay = to_y - from_y;
	135	+
113	136	if ( from_y < 0 )
114	137	{
115		- from_x -= from_y * dxdy;
	138	+ from_x -= from_y * deltax/deltay;
116	139	from_y = 0;
117	140	}
118		- if ( to_y > worker->m_h )
	141	+
	142	+ // TODO: test, imp line
	143	+ if ( to_y > worker->m_h<<6 )
119	144	{
120		- to_x -= ( to_y - worker->m_h ) * dxdy;
121		- to_y = (float)worker->m_h;
	145	+ to_x -= (( to_y - worker->m_h<<6 ) * deltax/deltay);
	146	+ to_y = worker->m_h<<6;
122	147	}
123	148
124		- float x = from_x;
125		- int y0 = (int)from_y;
126		- int y_limit = (int)ceil( to_y );
127		- float* m_a = worker->m_a;
	149	+ int x = from_x;
	150	+ int y0 = from_y>>6;
	151	+ int y_limit = (to_y + 0x3f)>>6;
	152	+ FT20D12* m_a = worker->m_a;
128	153
129	154	for ( int y = y0; y < y_limit; y++ )
130	155	{
131	156	// printf("y is %d\n", y);
132	157	int linestart = y * worker->m_w;
133		- float dy = fmin( y + 1.0f, to_y ) - fmax( (float)y, from_y );
134		- float xnext = x + dxdy * dy;
135		- float d = dy * dir;
	158	+ FT26D6 dy = min( (y + 1)<<6, to_y ) - max( y<<6, from_y );
	159	+ FT26D6 xnext = x + dy * deltax/deltay;
	160	+ FT26D6 d = dy * dir;
136	161
137		- float x0, x1;
	162	+ FT26D6 x0, x1;
138	163	if ( x < xnext )
139	164	{
140	165	x0 = x;
...	...	@@ -151,35 +176,59 @@ dense_render_line( dense_worker* worker, TPos tox, TPos toy )
151	176	floating-point inaccuracy That would cause an out-of-bounds array access at
152	177	index -1.
153	178	*/
154		- float x0floor = x0 <= 0.0f ? 0.0f : (float)floor( x0 );
	179	+ // float x0floor = x0 <= 0.0f ? 0.0f : (float)floor( x0 );
	180	+
	181	+ int x0i = x0>>6;
	182	+ FT26D6 x0floor = x0i<<6;
	183	+
	184	+
	185	+ // float x1ceil = (float)ceil( x1 );
	186	+ int x1i = (x1+0x3f)>>6;
	187	+ FT26D6 x1ceil = x1i <<6;
155	188
156		- int x0i = (int)x0floor;
157		- float x1ceil = (float)ceil( x1 );
158		- int x1i = (int)x1ceil;
159	189	if ( x1i <= x0i + 1 )
160	190	{
161		- float xmf = 0.5f * ( x + xnext ) - x0floor;
162		- m_a[linestart + x0i] += d - d * xmf;
	191	+ FT26D6 xmf = ( ( x + xnext )>>1) - x0floor;
	192	+ m_a[linestart + x0i] += d * ((1<<6) - xmf);
163	193	m_a[linestart + ( x0i + 1 )] += d * xmf;
164	194	}
165	195	else
166	196	{
167		- float s = 1.0f / ( x1 - x0 );
168		- float x0f = x0 - x0floor;
169		- float a0 = 0.5f * s * ( 1.0f - x0f ) * ( 1.0f - x0f );
170		- float x1f = x1 - x1ceil + 1.0f;
171		- float am = 0.5f * s * x1f * x1f;
	197	+ // float s = 1.0f / ( x1 - x0 );
	198	+ // float x0f = x0 - x0floor;
	199	+ // float a0 = 0.5f * s * ( 1.0f - x0f ) * ( 1.0f - x0f );
	200	+ // float x1f = x1 - x1ceil + 1.0f;
	201	+ // float am = 0.5f * s * x1f * x1f;
	202	+
	203	+ FT26D6 oneOverS = x1 - x0;
	204	+ FT26D6 x0f = x0 - x0floor;
	205	+
	206	+
	207	+ FT26D6 oneMinusX0f = (1<<6) - x0f;
	208	+ FT26D6 a0 = ((oneMinusX0f * oneMinusX0f) >> 1) / oneOverS;
	209	+ FT26D6 x1f = x1 - x1ceil + (1<<6);
	210	+ FT26D6 am = ((x1f * x1f) >> 1) / oneOverS;
	211	+
	212	+ // printf("x0 is %lld, x1 is %lld\n",x0,x1 );
	213	+
	214	+
	215	+
172	216	m_a[linestart + x0i] += d * a0;
	217	+
173	218	if ( x1i == x0i + 2 )
174		- m_a[linestart + ( x0i + 1 )] += d * ( 1.0f - a0 - am );
	219	+ m_a[linestart + ( x0i + 1 )] += d * ( (1<<6) - a0 - am );
175	220	else
176	221	{
177		- float a1 = s * ( 1.5f - x0f );
	222	+ FT26D6 a1 = (((1<<6) + (1<<5) - x0f) << 6) / oneOverS;
178	223	m_a[linestart + ( x0i + 1 )] += d * ( a1 - a0 );
179		- for ( int xi = x0i + 2; xi < x1i - 1; xi++ )
180		- m_a[linestart + xi] += d * s;
181		- float a2 = a1 + ( x1i - x0i - 3 ) * s;
182		- m_a[linestart + ( x1i - 1 )] += d * ( 1.0f - a2 - am );
	224	+
	225	+ FT26D6 dTimesS = (d << 12) / oneOverS;
	226	+
	227	+ for ( FT26D6 xi = x0i + 2; xi < x1i - 1; xi++ )
	228	+ m_a[linestart + xi] += dTimesS;
	229	+
	230	+ FT26D6 a2 = a1 + (( x1i - x0i - 3 )<<12)/oneOverS;
	231	+ m_a[linestart + ( x1i - 1 )] += d * ( (1<<6) - a2 - am );
183	232	}
184	233	m_a[linestart + x1i] += d * am;
185	234	}
...	...	@@ -369,26 +418,45 @@ dense_render_glyph( dense_worker* worker, const FT_Bitmap* target )
369	418	FT_Error error = FT_Outline_Decompose( &( worker->outline ),
370	419	&dense_decompose_funcs, worker );
371	420	// Render into bitmap
372		- const float* source = worker->m_a;
	421	+ const FT20D12* source = worker->m_a;
373	422
374	423	unsigned char* dest = target->buffer;
375	424	unsigned char* dest_end = target->buffer + worker->m_w * worker->m_h;
376	425
377		- __m128 offset = _mm_setzero_ps();
378		- __m128i mask = _mm_set1_epi32(0x0c080400);
379		- __m128 sign_mask = _mm_set1_ps(-0.f);
380		- for (int i = 0; i < worker->m_h*worker->m_w; i += 4) {
381		- __m128 x = _mm_load_ps(&source[i]);
382		- x = _mm_add_ps(x, _mm_castsi128_ps(_mm_slli_si128(_mm_castps_si128(x), 4)));
383		- x = _mm_add_ps(x, _mm_shuffle_ps(_mm_setzero_ps(), x, 0x40));
384		- x = _mm_add_ps(x, offset);
385		- __m128 y = _mm_andnot_ps(sign_mask, x); // fabs(x)
386		- y = _mm_min_ps(y, _mm_set1_ps(1.0f));
387		- y = _mm_mul_ps(y, _mm_set1_ps(255.0f));
388		- __m128i z = _mm_cvtps_epi32(y);
389		- z = _mm_shuffle_epi8(z, mask);
390		- _mm_store_ss((float *)&dest[i], (__m128)z);
391		- offset = _mm_shuffle_ps(x, x, _MM_SHUFFLE(3, 3, 3, 3));
	426	+ // __m128 offset = _mm_setzero_ps();
	427	+ // __m128i mask = _mm_set1_epi32(0x0c080400);
	428	+ // __m128 sign_mask = _mm_set1_ps(-0.f);
	429	+ // for (int i = 0; i < worker->m_h*worker->m_w; i += 4) {
	430	+ // __m128 x = _mm_load_ps(&source[i]);
	431	+ // x = _mm_add_ps(x, _mm_castsi128_ps(_mm_slli_si128(_mm_castps_si128(x), 4)));
	432	+ // x = _mm_add_ps(x, _mm_shuffle_ps(_mm_setzero_ps(), x, 0x40));
	433	+ // x = _mm_add_ps(x, offset);
	434	+ // __m128 y = _mm_andnot_ps(sign_mask, x); // fabs(x)
	435	+ // y = _mm_min_ps(y, _mm_set1_ps(1.0f));
	436	+ // y = _mm_mul_ps(y, _mm_set1_ps(255.0f));
	437	+ // __m128i z = _mm_cvtps_epi32(y);
	438	+ // z = _mm_shuffle_epi8(z, mask);
	439	+ // _mm_store_ss((float *)&dest[i], (__m128)z);
	440	+ // offset = _mm_shuffle_ps(x, x, _MM_SHUFFLE(3, 3, 3, 3));
	441	+ // }
	442	+
	443	+ FT20D12 valnew = 0;
	444	+ //float value = 0.0f;
	445	+ while ( dest < dest_end )
	446	+ {
	447	+ valnew += *source++;
	448	+
	449	+ // printf("%d\n", *source);
	450	+
	451	+ if(valnew > 0){
	452	+ int nnew = valnew >>4;
	453	+
	454	+ if(nnew>255)nnew=255;
	455	+ *dest = (unsigned char)nnew;
	456	+ }else{
	457	+ *dest = 0;
	458	+ }
	459	+ dest++;
392	460	}
393	461
394	462	// float value = 0.0f;
...	...	@@ -445,10 +513,10 @@ dense_raster_render( FT_Raster raster, const FT_Raster_Params* params )
445	513
446	514	int size = worker->m_w * worker->m_h + 4;
447	515
448		- worker->m_a = malloc( sizeof( float ) * size );
	516	+ worker->m_a = malloc( sizeof( FT20D12 ) * size );
449	517	worker->m_a_size = size;
450	518
451		- memset( worker->m_a, 0, ( sizeof( float ) * size ) );
	519	+ memset( worker->m_a, 0, ( sizeof( FT20D12 ) * size ) );
452	520	/* exit if nothing to do */
453	521	if ( worker->m_w <= worker->m_origin_x || worker->m_h <= worker->m_origin_y )
454	522	{

src/dense/ftdense.h

---

...	...	@@ -40,12 +40,14 @@ extern "C"
40	40	{
41	41	#endif
42	42
43		- typedef long TPos;
	43	+ typedef signed long TPos;
	44	+ typedef signed long long FT26D6;
	45	+ typedef signed long long FT20D12;
44	46
45	47	typedef struct
46	48	{
47	49	/** The array used to store signed area differences. */
48		- float* m_a;
	50	+ FT20D12* m_a;
49	51	/** The number of elements in m_a. */
50	52	int m_a_size;
51	53	/** The width of the current raster in pixels. */

—
View it on GitLab.
You're receiving this email because of your account on gitlab.freedesktop.org. Manage all notifications · Help