1.1 --- a/src/pvr2/yuv.c	Sun Jan 14 02:55:25 2007 +0000
     1.2 +++ b/src/pvr2/yuv.c	Mon Jan 15 08:30:50 2007 +0000
     1.3 @@ -1,5 +1,5 @@
     1.4  /**
     1.5 - * $Id: yuv.c,v 1.1 2007-01-14 02:55:25 nkeynes Exp $
     1.6 + * $Id: yuv.c,v 1.2 2007-01-14 11:43:00 nkeynes Exp $
     1.7   *
     1.8   * YUV420 and YUV422 decoding
     1.9   *
    1.10 @@ -16,46 +16,142 @@
    1.11   * GNU General Public License for more details.
    1.12   */
    1.13  #include "dream.h"
    1.14 +#include "asic.h"
    1.15 +#include "pvr2/pvr2.h"
    1.16  
    1.17  #define YUV420_BLOCK_SIZE 384
    1.18 +#define YUV422_BLOCK_SIZE 512
    1.19  
    1.20 -static inline uint16_t decode_yuv420_to_rgb565_pixel( uint8_t y, uint8_t u, uint8_t v )
    1.21 +#define FORMAT_YUV420 0
    1.22 +#define FORMAT_YUV422 1
    1.23 +
    1.24 +
    1.25 +static int yuv_block_size[2] = { YUV420_BLOCK_SIZE, YUV422_BLOCK_SIZE };
    1.26 +
    1.27 +struct yuv_state {
    1.28 +    uint32_t target;
    1.29 +    int width;
    1.30 +    int height;
    1.31 +    int input_format;
    1.32 +    char data[512];
    1.33 +    int data_length;
    1.34 +    int x, y;
    1.35 +} pvr2_yuv_state;
    1.36 +
    1.37 +/**
    1.38 + * Transformation table for yuv420.
    1.39 + */
    1.40 +uint16_t yuv420_lut[512] = { 0, 128, 64, 129, 1, 130, 65, 131, 2, 132, 66, 133, 3, 134, 67, 135, 4, 192, 68, 193, 5, 194, 69, 195, 6, 196, 70, 197, 7, 198, 71, 199,
    1.41 +			 0, 136, 64, 137, 1, 138, 65, 139, 2, 140, 66, 141, 3, 142, 67, 143, 4, 200, 68, 201, 5, 202, 69, 203, 6, 204, 70, 205, 7, 206, 71, 207,
    1.42 +			 8, 144, 72, 145, 9, 146, 73, 147, 10, 148, 74, 149, 11, 150, 75, 151, 12, 208, 76, 209, 13, 210, 77, 211, 14, 212, 78, 213, 15, 214, 79, 215,
    1.43 +			 8, 152, 72, 153, 9, 154, 73, 155, 10, 156, 74, 157, 11, 158, 75, 159, 12, 216, 76, 217, 13, 218, 77, 219, 14, 220, 78, 221, 15, 222, 79, 223,
    1.44 +			 16, 160, 80, 161, 17, 162, 81, 163, 18, 164, 82, 165, 19, 166, 83, 167, 20, 224, 84, 225, 21, 226, 85, 227, 22, 228, 86, 229, 23, 230, 87, 231,
    1.45 +			 16, 168, 80, 169, 17, 170, 81, 171, 18, 172, 82, 173, 19, 174, 83, 175, 20, 232, 84, 233, 21, 234, 85, 235, 22, 236, 86, 237, 23, 238, 87, 239,
    1.46 +			 24, 176, 88, 177, 25, 178, 89, 179, 26, 180, 90, 181, 27, 182, 91, 183, 28, 240, 92, 241, 29, 242, 93, 243, 30, 244, 94, 245, 31, 246, 95, 247,
    1.47 +			 24, 184, 88, 185, 25, 186, 89, 187, 26, 188, 90, 189, 27, 190, 91, 191, 28, 248, 92, 249, 29, 250, 93, 251, 30, 252, 94, 253, 31, 254, 95, 255,
    1.48 +			 32, 256, 96, 257, 33, 258, 97, 259, 34, 260, 98, 261, 35, 262, 99, 263, 36, 320, 100, 321, 37, 322, 101, 323, 38, 324, 102, 325, 39, 326, 103, 327,
    1.49 +			 32, 264, 96, 265, 33, 266, 97, 267, 34, 268, 98, 269, 35, 270, 99, 271, 36, 328, 100, 329, 37, 330, 101, 331, 38, 332, 102, 333, 39, 334, 103, 335,
    1.50 +			 40, 272, 104, 273, 41, 274, 105, 275, 42, 276, 106, 277, 43, 278, 107, 279, 44, 336, 108, 337, 45, 338, 109, 339, 46, 340, 110, 341, 47, 342, 111, 343,
    1.51 +			 40, 280, 104, 281, 41, 282, 105, 283, 42, 284, 106, 285, 43, 286, 107, 287, 44, 344, 108, 345, 45, 346, 109, 347, 46, 348, 110, 349, 47, 350, 111, 351,
    1.52 +			 48, 288, 112, 289, 49, 290, 113, 291, 50, 292, 114, 293, 51, 294, 115, 295, 52, 352, 116, 353, 53, 354, 117, 355, 54, 356, 118, 357, 55, 358, 119, 359,
    1.53 +			 48, 296, 112, 297, 49, 298, 113, 299, 50, 300, 114, 301, 51, 302, 115, 303, 52, 360, 116, 361, 53, 362, 117, 363, 54, 364, 118, 365, 55, 366, 119, 367,
    1.54 +			 56, 304, 120, 305, 57, 306, 121, 307, 58, 308, 122, 309, 59, 310, 123, 311, 60, 368, 124, 369, 61, 370, 125, 371, 62, 372, 126, 373, 63, 374, 127, 375,
    1.55 +			 56, 312, 120, 313, 57, 314, 121, 315, 58, 316, 122, 317, 59, 318, 123, 319, 60, 376, 124, 377, 61, 378, 125, 379, 62, 380, 126, 381, 63, 382, 127, 383 };
    1.56 +
    1.57 +
    1.58 +/**
    1.59 + * Input is 8x8 U, 8x8 V, 8x8 Y00, 8x8 Y01, 8x8 Y10, 8x8 Y11, 8 bits each,
    1.60 + * for a total of 384 bytes.
    1.61 + * Output is UVYV = 32 bits = 2 horizontal pixels, 8x16 = 512 bytes
    1.62 + */
    1.63 +void pvr2_decode_yuv420( char *dest, char *src )
    1.64  {
    1.65 -    
    1.66 +    int i;
    1.67 +    for( i=0; i<512; i++ ) {
    1.68 +	dest[i] = src[yuv420_lut[i]];
    1.69 +    }
    1.70 +}
    1.71 +
    1.72 +void pvr2_decode_yuv422( char *dest, char *src )
    1.73 +{
    1.74  
    1.75  }
    1.76  
    1.77  /**
    1.78 - * Convert a single 16x16 yuv420 block to rgb565.
    1.79 - * @param dest output memory location for this block
    1.80 - * @param src start of source block
    1.81 - * @param stride length of overall line in pixels (ie 16-bit words)
    1.82 + * Process a single macroblock of YUV data and write it out to 
    1.83 + * texture vram.
    1.84   */
    1.85 -static void decode_yuv420_to_rgb565_block( uint16_t *dest, uint8_t *src, uint32_t stride ) 
    1.86 +void pvr2_yuv_process_block( char *data )
    1.87  {
    1.88 -    uint8_t *up = *src;
    1.89 -    uint8_t *vp = u + 64;
    1.90 -    uint8_t *yp = v + 64;
    1.91 +    char output[512];
    1.92 +
    1.93 +    if( pvr2_yuv_state.input_format == FORMAT_YUV420 ) {
    1.94 +	pvr2_decode_yuv420( output, data );
    1.95 +    } else {
    1.96 +	pvr2_decode_yuv422( output, data );
    1.97 +    }
    1.98      
    1.99 -    for( int yb=0; yb<16; yb++ ) {
   1.100 -	for( int xb=0; xb<16; xb++ ) {
   1.101 -	    uint8_t y = *yp++;
   1.102 -	    uint8_t u = up[xb>>1 + (yb>>1)<<3];
   1.103 -	    uint8_t v = vp[xb>>1 + (yb>>1)<<3];
   1.104 -	    *dest++ = decode_yuv420_to_rgb565_block(y,u,v);
   1.105 +    uint32_t target = pvr2_yuv_state.target + 
   1.106 +	(pvr2_yuv_state.y * pvr2_yuv_state.width * 512) +
   1.107 +	(pvr2_yuv_state.x * 32);
   1.108 +
   1.109 +    pvr2_vram64_write_stride( target, output, 32, pvr2_yuv_state.width*32, 16 );
   1.110 +    if( ++pvr2_yuv_state.x >= pvr2_yuv_state.width ) {
   1.111 +	pvr2_yuv_state.x = 0;
   1.112 +	pvr2_yuv_state.y++;
   1.113 +	if( pvr2_yuv_state.y >= pvr2_yuv_state.height ) {
   1.114 +	    asic_event( EVENT_PVR_YUV_DONE );
   1.115  	}
   1.116 -	dest = dest + stride - 16;
   1.117 +    }
   1.118 +    
   1.119 +    MMIO_WRITE( PVR2, YUV_COUNT, MMIO_READ( PVR2, YUV_COUNT ) + 1 );
   1.120 +}
   1.121 +
   1.122 +/**
   1.123 + * Receive data from the SH4, usually via DMA. This method is mainly responsible
   1.124 + * for buffering the data into macroblock chunks and then passing it on to the
   1.125 + * real processing
   1.126 + */
   1.127 +void pvr2_yuv_write( char *data, uint32_t length )
   1.128 +{
   1.129 +    int block_size = yuv_block_size[pvr2_yuv_state.input_format];
   1.130 +
   1.131 +    if( pvr2_yuv_state.data_length != 0 ) { /* Append to existing data */
   1.132 +	int tmp = MIN( length, block_size - pvr2_yuv_state.data_length );
   1.133 +	memcpy( pvr2_yuv_state.data + pvr2_yuv_state.data_length, 
   1.134 +		data, tmp );
   1.135 +	pvr2_yuv_state.data_length += tmp;
   1.136 +	data += tmp;
   1.137 +	length -= tmp;
   1.138 +	if( pvr2_yuv_state.data_length == block_size ) {
   1.139 +	    pvr2_yuv_process_block( pvr2_yuv_state.data );
   1.140 +	}
   1.141 +    }
   1.142 +
   1.143 +    while( length >= block_size ) {
   1.144 +	pvr2_yuv_process_block( data );
   1.145 +	data += block_size;
   1.146 +	length -= block_size;
   1.147 +    }
   1.148 +	    
   1.149 +    if( length != 0 ) { /* Save the left over data */
   1.150 +	memcpy( pvr2_yuv_state.data, data, length );
   1.151 +	pvr2_yuv_state.data_length = length;
   1.152      }
   1.153  }
   1.154  
   1.155 -void decode_yuv420_to_rgb565( uint16_t *dest, uint8_t *src, int width, int height ) 
   1.156 +void pvr2_yuv_init( uint32_t target, uint32_t config )
   1.157  {
   1.158 -    uint16_t *p;
   1.159 -    for( int j=0; j<height; j++ ) {
   1.160 -	for( int i=0; i<width; i++ ) {
   1.161 -	    p = dest + (j<<5)*width + i<<5;
   1.162 -	    decode_yuv420_to_rgb565_block( p, src, width );
   1.163 -	    src += YUV420_BLOCK_SIZE;
   1.164 -	}
   1.165 +    pvr2_yuv_state.target = target;
   1.166 +    pvr2_yuv_state.width = (config & 0x3f) + 1;
   1.167 +    pvr2_yuv_state.height = ((config>>8) & 0x3f) +1;
   1.168 +    pvr2_yuv_state.x = 0;
   1.169 +    pvr2_yuv_state.y = 0;
   1.170 +    pvr2_yuv_state.data_length = 0;
   1.171 +    pvr2_yuv_state.input_format = (config & 0x01000000) ? FORMAT_YUV420 : FORMAT_YUV422;
   1.172 +    if( config & 0x00010000 ) {
   1.173 +	pvr2_yuv_state.height *= pvr2_yuv_state.width;
   1.174 +	pvr2_yuv_state.width = 1;
   1.175      }
   1.176 +    MMIO_WRITE( PVR2, YUV_COUNT, 0 );
   1.177  }