[Git][freetype/freetype][gsoc-anurag-2023] Optimize SIMD even more

[Anurag Thakur (@AdbhutDev)]

Anurag Thakur pushed to branch gsoc-anurag-2023 at FreeType / FreeType

Commits:

• 55c25257
  by Anurag Thakur at 2023-10-06T03:10:27+05:30

  Optimize SIMD even more
  

2 changed files:

• src/base/ftobjs.c
• src/dense/ftdense.c

Changes:

src/base/ftobjs.c

---

...	...	@@ -3154,7 +3154,7 @@ int conic_to2(FT_GlyphSlot* slot, FT_Vector *control, FT_Vector *from, FT_Vector
3154	3154	face->garray = (FT_GlyphSlot*)malloc(
3155	3155	face->driver->clazz->slot_object_size * face->num_glyphs );
3156	3156	//error = FT_Set_Char_Size( face, 0, 160 * 64, 300, 300 );
3157		- error = FT_Set_Pixel_Sizes( face, 0, 500);
	3157	+ error = FT_Set_Pixel_Sizes( face, 0, 100);
3158	3158	// int glyph_index = FT_Get_Char_Index( face, 'A' );
3159	3159	// error = FT_Load_Glyph( face, glyph_index, FT_LOAD_NO_HINTING );
3160	3160

src/dense/ftdense.c

---

...	...	@@ -443,8 +443,7 @@ dense_render_glyph( dense_worker* worker, const FT_Bitmap* target, FT_PreLine pl
443	443	}
444	444	// point.x = 100;
445	445	// point.y = 100;
446		- // dense_move_to(&point, worker);
447		-
	446	+ // dense_move_to(&point, worker);
448	447	// Render into bitmap
449	448	const FT20D12* source = worker->m_a;
450	449	unsigned char* dest = target->buffer;
...	...	@@ -460,6 +459,7 @@ __m128i offset = _mm_setzero_si128();
460	459	{
461	460	// load 4 floats from source
462	461
	462	+ //printf("%d\n", source[i]);
463	463	__m128i x = _mm_load_si128( (__m128i*)&source[i] );
464	464
465	465	x = _mm_add_epi32( x, _mm_slli_si128( x, 4 ) );
...	...	@@ -481,27 +481,40 @@ __m128i offset = _mm_setzero_si128();
481	481	// __m128i z = _mm_cvttps_epi32(y);
482	482	// z = _mm_packus_epi16(_mm_packs_epi32(z, nzero), nzero);
483	483
484		- // *((int*)dest+i) = _mm_extract_epi16(z, 0);
	484	+
	485	+
	486	+ // int yu = ;
	487	+ // *((int*)dest+i) = yu;
485	488
486	489
487	490	// take absolute value
488		- __m128i y = _mm_abs_epi32( x ); // fabs(x)
	491	+ //__m128i y = _mm_abs_epi32( x ); // fabs(x)
	492	+
489	493
490	494	// cap max value to 1
491		- y = _mm_min_epi32( y, _mm_set1_epi32( 4094 ) );
	495	+ //y = _mm_min_epi32( _mm_srli_epi32( y, 4 ), _mm_set1_epi32( 255 ) );
	496	+ __m128i y = _mm_abs_epi32(_mm_srai_epi32( x , 4 ));
492	497
493	498	// reduce to 255
494		- y = _mm_srli_epi32( y, 4 );
	499	+ // y =
	500	+
	501	+ // // shuffle
	502	+ //y = _mm_shuffle_epi8( y, mask );
	503	+
	504	+ y = _mm_packus_epi16(_mm_packs_epi32(y, nzero), nzero);
	505	+ //__m128i z = _mm_packus_epi16(_mm_packs_epi32(z, nzero), nzero);
495	506
496		- // shuffle
497		- y = _mm_shuffle_epi8( y, mask );
	507	+ // int* ptr = (int*)&dest[i];
	508	+ *(int*)&dest[i] = *(int*)&y;
	509	+ //*(int*)&dest[i] = _mm_extract_epi32(y, 0);
498	510
499		- _mm_store_ss( (float*)&dest[i], _mm_castsi128_ps(y) );
	511	+ //_mm_store_ss( (float*)&dest[i], _mm_castsi128_ps(y) );
500	512
501	513	// store the current prefix sum in offset
502		- offset = _mm_castps_si128( _mm_shuffle_ps( _mm_castsi128_ps( x ),
503		- _mm_castsi128_ps( x ),
504		- _MM_SHUFFLE( 3, 3, 3, 3 ) ) );
	514	+ // offset = _mm_castps_si128( _mm_shuffle_ps( _mm_castsi128_ps( x ),
	515	+ // _mm_castsi128_ps( x ),
	516	+ // _MM_SHUFFLE( 3, 3, 3, 3 ) ) );
	517	+ offset = _mm_shuffle_epi32(x,_MM_SHUFFLE( 3, 3, 3, 3 ) );
505	518	//offset = _mm_set1_epi32(_mm_extract_epi32(x, 3));
506	519	}
507	520
...	...	@@ -566,7 +579,7 @@ dense_raster_render( FT_Raster raster, const FT_Raster_Params* params )
566	579	worker->m_w = target_map->pitch;
567	580	worker->m_h = target_map->rows;
568	581
569		- int size = worker->m_w * worker->m_h + 4;
	582	+ int size = (worker->m_w * worker->m_h + 3) & ~3;
570	583
571	584	worker->m_a = malloc( sizeof( FT20D12 ) * size );
572	585	worker->m_a_size = size;

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