... |
... |
@@ -11,7 +11,7 @@ |
11
|
11
|
#include "ftdense.h"
|
12
|
12
|
|
13
|
13
|
#include <math.h>
|
14
|
|
-#include <tmmintrin.h>
|
|
14
|
+#include <immintrin.h>
|
15
|
15
|
#include "ftdenseerrs.h"
|
16
|
16
|
|
17
|
17
|
#define PIXEL_BITS 8
|
... |
... |
@@ -419,41 +419,66 @@ dense_render_glyph( dense_worker* worker, const FT_Bitmap* target ) |
419
|
419
|
unsigned char* dest = target->buffer;
|
420
|
420
|
unsigned char* dest_end = target->buffer + worker->m_w * worker->m_h;
|
421
|
421
|
|
422
|
|
- // __m128 offset = _mm_setzero_ps();
|
423
|
|
- // __m128i mask = _mm_set1_epi32(0x0c080400);
|
424
|
|
- // __m128 sign_mask = _mm_set1_ps(-0.f);
|
425
|
|
- // for (int i = 0; i < worker->m_h*worker->m_w; i += 4) {
|
426
|
|
- // __m128 x = _mm_load_ps(&source[i]);
|
427
|
|
- // x = _mm_add_ps(x, _mm_castsi128_ps(_mm_slli_si128(_mm_castps_si128(x), 4)));
|
428
|
|
- // x = _mm_add_ps(x, _mm_shuffle_ps(_mm_setzero_ps(), x, 0x40));
|
429
|
|
- // x = _mm_add_ps(x, offset);
|
430
|
|
- // __m128 y = _mm_andnot_ps(sign_mask, x); // fabs(x)
|
431
|
|
- // y = _mm_min_ps(y, _mm_set1_ps(1.0f));
|
432
|
|
- // y = _mm_mul_ps(y, _mm_set1_ps(255.0f));
|
433
|
|
- // __m128i z = _mm_cvtps_epi32(y);
|
434
|
|
- // z = _mm_shuffle_epi8(z, mask);
|
435
|
|
- // _mm_store_ss((float *)&dest[i], (__m128)z);
|
436
|
|
- // offset = _mm_shuffle_ps(x, x, _MM_SHUFFLE(3, 3, 3, 3));
|
437
|
|
- // }
|
438
|
|
-
|
439
|
|
- FT20D12 valnew = 0;
|
440
|
|
- //float value = 0.0f;
|
441
|
|
- while ( dest < dest_end )
|
|
422
|
+ __m128i offset = _mm_setzero_si128();
|
|
423
|
+ __m128i mask = _mm_set1_epi32( 0x0c080400 );
|
|
424
|
+
|
|
425
|
+ for (int i = 0; i < worker->m_h*worker->m_w; i += 4)
|
442
|
426
|
{
|
443
|
|
- valnew += *source++;
|
|
427
|
+ // load 4 floats from source
|
444
|
428
|
|
445
|
|
- // printf("%d\n", *source);
|
|
429
|
+ __m128i x = _mm_load_si128( (__m128i*)&source[i] );
|
446
|
430
|
|
447
|
|
- if(valnew > 0){
|
448
|
|
- int nnew = valnew >>4;
|
|
431
|
+ // bkc
|
|
432
|
+ x = _mm_add_epi32( x, _mm_slli_si128( x, 4 ) );
|
449
|
433
|
|
450
|
|
- if(nnew>255)nnew=255;
|
451
|
|
- *dest = (unsigned char)nnew;
|
452
|
|
- }else{
|
453
|
|
- *dest = 0;
|
454
|
|
- }
|
455
|
|
- dest++;
|
|
434
|
+ // more bkc
|
|
435
|
+ x = _mm_add_epi32(
|
|
436
|
+ x, _mm_castps_si128( _mm_shuffle_ps( _mm_setzero_ps(),
|
|
437
|
+ _mm_castsi128_ps( x ), 0x40 ) ) );
|
|
438
|
+
|
|
439
|
+ // add the prefsum of previous 4 floats to all current floats
|
|
440
|
+ x = _mm_add_epi32( x, offset );
|
|
441
|
+
|
|
442
|
+ // take absolute value
|
|
443
|
+ __m128i y = _mm_abs_epi32( x ); // fabs(x)
|
|
444
|
+
|
|
445
|
+ // cap max value to 1
|
|
446
|
+ y = _mm_min_epi32( y, _mm_set1_epi32( 4080 ) );
|
|
447
|
+
|
|
448
|
+ // reduce to 255
|
|
449
|
+ y = _mm_srli_epi32( y, 4 );
|
|
450
|
+
|
|
451
|
+ // black magic
|
|
452
|
+ y = _mm_shuffle_epi8( y, mask );
|
|
453
|
+
|
|
454
|
+ // for some reason, storing float in an unsigned char array works fine lol
|
|
455
|
+ _mm_store_ss( (float*)&dest[i], (__m128)y );
|
|
456
|
+
|
|
457
|
+ // store the current prefix sum in offset
|
|
458
|
+ offset = _mm_castps_si128( _mm_shuffle_ps( _mm_castsi128_ps( x ),
|
|
459
|
+ _mm_castsi128_ps( x ),
|
|
460
|
+ _MM_SHUFFLE( 3, 3, 3, 3 ) ) );
|
456
|
461
|
}
|
|
462
|
+
|
|
463
|
+
|
|
464
|
+ // FT20D12 valnew = 0;
|
|
465
|
+ // //float value = 0.0f;
|
|
466
|
+ // while ( dest < dest_end )
|
|
467
|
+ // {
|
|
468
|
+ // valnew += *source++;
|
|
469
|
+
|
|
470
|
+ // // printf("%d\n", *source);
|
|
471
|
+
|
|
472
|
+ // if(valnew > 0){
|
|
473
|
+ // int nnew = valnew >>4;
|
|
474
|
+
|
|
475
|
+ // if(nnew>255)nnew=255;
|
|
476
|
+ // *dest = (unsigned char)nnew;
|
|
477
|
+ // }else{
|
|
478
|
+ // *dest = 0;
|
|
479
|
+ // }
|
|
480
|
+ // dest++;
|
|
481
|
+ // }
|
457
|
482
|
|
458
|
483
|
free(worker->m_a);
|
459
|
484
|
return error;
|