[Git][freetype/freetype][gsoc-anurag-2022-final] [dense] Re-enable SIMD to work with fixed-point

[Anurag Thakur (@AdbhutDev)]

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

Commits:

• 763f1108
  by Anurag Thakur at 2022-11-19T13:05:14+05:30

  [dense] Re-enable SIMD to work with fixed-point
  
  * src/dense/ftdense.c: Use integer SIMD functions for accumulation
  
  * src/dense/ftdense.h: Change types of FT26D6, FT20D12 to better fit
  their usage
  

2 changed files:

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

Changes:

src/dense/ftdense.c

---

...	...	@@ -24,7 +24,7 @@
24	24
25	25	#if FT_SSE4_1
26	26
27		- #include <tmmintrin.h>
	27	+ #include <immintrin.h>
28	28
29	29	#endif
30	30
...	...	@@ -384,26 +384,47 @@ dense_render_glyph( dense_worker* worker, const FT_Bitmap* target )
384	384	unsigned char* dest = target->buffer;
385	385	unsigned char* dest_end = target->buffer + worker->m_w * worker->m_h;
386	386
387		-//#if FT_SSE4_1
388		-
389		- // __m128 offset = _mm_setzero_ps();
390		- // __m128i mask = _mm_set1_epi32(0x0c080400);
391		- // __m128 sign_mask = _mm_set1_ps(-0.f);
392		- // for (int i = 0; i < worker->m_h*worker->m_w; i += 4) {
393		- // __m128 x = _mm_load_ps(&source[i]);
394		- // x = _mm_add_ps(x, _mm_castsi128_ps(_mm_slli_si128(_mm_castps_si128(x), 4)));
395		- // x = _mm_add_ps(x, _mm_shuffle_ps(_mm_setzero_ps(), x, 0x40));
396		- // x = _mm_add_ps(x, offset);
397		- // __m128 y = _mm_andnot_ps(sign_mask, x); // fabs(x)
398		- // y = _mm_min_ps(y, _mm_set1_ps(1.0f));
399		- // y = _mm_mul_ps(y, _mm_set1_ps(255.0f));
400		- // __m128i z = _mm_cvtps_epi32(y);
401		- // z = _mm_shuffle_epi8(z, mask);
402		- // _mm_store_ss((float *)&dest[i], (__m128)z);
403		- // offset = _mm_shuffle_ps(x, x, _MM_SHUFFLE(3, 3, 3, 3));
404		- // }
405		-
406		-//#else /* FT_SSE4_1 */
	387	+#if FT_SSE4_1
	388	+
	389	+__m128i offset = _mm_setzero_si128();
	390	+ __m128i mask = _mm_set1_epi32( 0x0c080400 );
	391	+
	392	+ for (int i = 0; i < worker->m_h*worker->m_w; i += 4)
	393	+ {
	394	+ // load 4 floats from source
	395	+
	396	+ __m128i x = _mm_load_si128( (__m128i*)&source[i] );
	397	+
	398	+ x = _mm_add_epi32( x, _mm_slli_si128( x, 4 ) );
	399	+
	400	+ x = _mm_add_epi32(
	401	+ x, _mm_castps_si128( _mm_shuffle_ps( _mm_setzero_ps(),
	402	+ _mm_castsi128_ps( x ), 0x40 ) ) );
	403	+
	404	+ // add the prefsum of previous 4 floats to all current floats
	405	+ x = _mm_add_epi32( x, offset );
	406	+
	407	+ // take absolute value
	408	+ __m128i y = _mm_abs_epi32( x ); // fabs(x)
	409	+
	410	+ // cap max value to 1
	411	+ y = _mm_min_epi32( y, _mm_set1_epi32( 4080 ) );
	412	+
	413	+ // reduce to 255
	414	+ y = _mm_srli_epi32( y, 4 );
	415	+
	416	+ // shuffle
	417	+ y = _mm_shuffle_epi8( y, mask );
	418	+
	419	+ _mm_store_ss( (float*)&dest[i], (__m128)y );
	420	+
	421	+ // store the current prefix sum in offset
	422	+ offset = _mm_castps_si128( _mm_shuffle_ps( _mm_castsi128_ps( x ),
	423	+ _mm_castsi128_ps( x ),
	424	+ _MM_SHUFFLE( 3, 3, 3, 3 ) ) );
	425	+ }
	426	+
	427	+#else /* FT_SSE4_1 */
407	428
408	429	FT20D12 value = 0;
409	430
...	...	@@ -422,7 +443,7 @@ dense_render_glyph( dense_worker* worker, const FT_Bitmap* target )
422	443	dest++;
423	444	}
424	445
425		-//#endif /* FT_SSE4_1 */
	446	+#endif /* FT_SSE4_1 */
426	447
427	448	free(worker->m_a);
428	449	return error;

src/dense/ftdense.h

---

...	...	@@ -20,8 +20,8 @@ extern "C"
20	20	#endif
21	21
22	22
23		- typedef signed long long FT26D6; /* 26.6 fixed-point representation */
24		- typedef signed long long FT20D12; /* 20.12 fixed-point representation */
	23	+ typedef signed long FT26D6; /* 26.6 fixed-point representation */
	24	+ typedef signed int FT20D12; /* 20.12 fixed-point representation */
25	25
26	26	typedef struct
27	27	{

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