... |
... |
@@ -12,6 +12,22 @@ |
12
|
12
|
#include "ftdense.h"
|
13
|
13
|
#include "ftdenseerrs.h"
|
14
|
14
|
|
|
15
|
+#if defined( __SSE4_1__ ) || \
|
|
16
|
+ defined( __x86_64__ ) || \
|
|
17
|
+ defined( _M_AMD64 ) || \
|
|
18
|
+ ( defined( _M_IX86_FP ) && _M_IX86_FP >= 2 )
|
|
19
|
+# define FT_SSE4_1 1
|
|
20
|
+#else
|
|
21
|
+# define FT_SSE4_1 0
|
|
22
|
+#endif
|
|
23
|
+
|
|
24
|
+
|
|
25
|
+#if FT_SSE4_1
|
|
26
|
+
|
|
27
|
+ #include <tmmintrin.h>
|
|
28
|
+
|
|
29
|
+#endif
|
|
30
|
+
|
15
|
31
|
#define PIXEL_BITS 8
|
16
|
32
|
|
17
|
33
|
#define ONE_PIXEL ( 1 << PIXEL_BITS )
|
... |
... |
@@ -349,9 +365,30 @@ dense_render_glyph( dense_worker* worker, const FT_Bitmap* target ) |
349
|
365
|
&dense_decompose_funcs, worker );
|
350
|
366
|
// Render into bitmap
|
351
|
367
|
const float* source = worker->m_a;
|
352
|
|
-
|
353
|
368
|
unsigned char* dest = target->buffer;
|
354
|
369
|
unsigned char* dest_end = target->buffer + worker->m_w * worker->m_h;
|
|
370
|
+
|
|
371
|
+#if FT_SSE4_1
|
|
372
|
+
|
|
373
|
+ __m128 offset = _mm_setzero_ps();
|
|
374
|
+ __m128i mask = _mm_set1_epi32(0x0c080400);
|
|
375
|
+ __m128 sign_mask = _mm_set1_ps(-0.f);
|
|
376
|
+ for (int i = 0; i < worker->m_h*worker->m_w; i += 4) {
|
|
377
|
+ __m128 x = _mm_load_ps(&source[i]);
|
|
378
|
+ x = _mm_add_ps(x, _mm_castsi128_ps(_mm_slli_si128(_mm_castps_si128(x), 4)));
|
|
379
|
+ x = _mm_add_ps(x, _mm_shuffle_ps(_mm_setzero_ps(), x, 0x40));
|
|
380
|
+ x = _mm_add_ps(x, offset);
|
|
381
|
+ __m128 y = _mm_andnot_ps(sign_mask, x); // fabs(x)
|
|
382
|
+ y = _mm_min_ps(y, _mm_set1_ps(1.0f));
|
|
383
|
+ y = _mm_mul_ps(y, _mm_set1_ps(255.0f));
|
|
384
|
+ __m128i z = _mm_cvtps_epi32(y);
|
|
385
|
+ z = _mm_shuffle_epi8(z, mask);
|
|
386
|
+ _mm_store_ss((float *)&dest[i], (__m128)z);
|
|
387
|
+ offset = _mm_shuffle_ps(x, x, _MM_SHUFFLE(3, 3, 3, 3));
|
|
388
|
+ }
|
|
389
|
+
|
|
390
|
+#else /* FT_SSE4_1 */
|
|
391
|
+
|
355
|
392
|
float value = 0.0f;
|
356
|
393
|
while ( dest < dest_end )
|
357
|
394
|
{
|
... |
... |
@@ -368,6 +405,8 @@ dense_render_glyph( dense_worker* worker, const FT_Bitmap* target ) |
368
|
405
|
dest++;
|
369
|
406
|
}
|
370
|
407
|
|
|
408
|
+#endif /* FT_SSE4_1 */
|
|
409
|
+
|
371
|
410
|
free(worker->m_a);
|
372
|
411
|
return error;
|
373
|
412
|
}
|