reorganize and simplify the VP3 IDCT stuff

Originally committed as revision 3071 to svn://svn.ffmpeg.org/ffmpeg/trunk

3 files changed, 8 insertions, 141 deletions

diff --git a/libavcodec/i386/dsputil_mmx.c b/libavcodec/i386/dsputil_mmx.c
index 772c9c1f03..61bfc89ac5 100644
--- a/libavcodec/i386/dsputil_mmx.c
+++ b/libavcodec/i386/dsputil_mmx.c
@@ -2149,14 +2149,12 @@ void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx)
         /* VP3 optimized DSP functions */
         if (mm_flags & MM_SSE2) {
             c->vp3_dsp_init = vp3_dsp_init_sse2;
-            c->vp3_idct_put = vp3_idct_put_sse2;
-            c->vp3_idct_add = vp3_idct_add_sse2;
+            c->vp3_idct = vp3_idct_sse2;
         } else {
             c->vp3_dsp_init = vp3_dsp_init_mmx;
-            c->vp3_idct_put = vp3_idct_put_mmx;
-            c->vp3_idct_add = vp3_idct_add_mmx;
+            c->vp3_idct = vp3_idct_mmx;
         }
-        
+
 #ifdef CONFIG_ENCODERS
         c->get_pixels = get_pixels_mmx;
         c->diff_pixels = diff_pixels_mmx;
diff --git a/libavcodec/i386/vp3dsp_mmx.c b/libavcodec/i386/vp3dsp_mmx.c
index 76007a1d16..319e57f1bb 100644
--- a/libavcodec/i386/vp3dsp_mmx.c
+++ b/libavcodec/i386/vp3dsp_mmx.c
@@ -279,8 +279,8 @@ void vp3_dsp_init_mmx(void)
     idct_constants[46] = idct_constants[47] = IdctAdjustBeforeShift;
 }
 
-static void vp3_idct_mmx(int16_t *input_data, int16_t *dequant_matrix,
-     int16_t *output_data)
+void vp3_idct_mmx(int16_t *input_data, int16_t *dequant_matrix,
+    int coeff_count, int16_t *output_data)
 {
     /* eax = quantized input
      * ebx = dequantizer matrix
@@ -563,79 +563,3 @@ static void vp3_idct_mmx(int16_t *input_data, int16_t *dequant_matrix,
 #undef J
 
 }
-
-void vp3_idct_put_mmx(int16_t *input_data, int16_t *dequant_matrix,
-    int coeff_count, uint8_t *dest, int stride)
-{
-    int16_t transformed_data[64];
-    int16_t *op;
-    int i, j;
-    uint8_t vector128[8] = { 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80 };
-
-    vp3_idct_mmx(input_data, dequant_matrix, transformed_data);
-
-    /* place in final output */
-    op = transformed_data;
-    movq_m2r(*vector128, mm0);
-    for (i = 0; i < 8; i++) {
-#if 1
-        for (j = 0; j < 8; j++) {
-            if (*op < -128)
-                *dest = 0;
-            else if (*op > 127)
-                *dest = 255;
-            else
-                *dest = (uint8_t)(*op + 128);
-            op++;
-            dest++;
-        }
-        dest += (stride - 8);
-#else
-/* prototype optimization */
-        pxor_r2r(mm1, mm1);
-        packsswb_m2r(*(op + 4), mm1);
-        movq_r2r(mm1, mm2);
-        psrlq_i2r(32, mm2);
-        packsswb_m2r(*(op + 0), mm1);
-        op += 8;
-        por_r2r(mm2, mm1);
-        paddb_r2r(mm0, mm1);
-        movq_r2m(mm1, *dest);
-        dest += stride;
-#endif
-    }
-
-    /* be a good MMX citizen */
-    emms();
-}
-
-void vp3_idct_add_mmx(int16_t *input_data, int16_t *dequant_matrix,
-    int coeff_count, uint8_t *dest, int stride)
-{
-    int16_t transformed_data[64];
-    int16_t *op;
-    int i, j;
-    int16_t sample;
-
-    vp3_idct_mmx(input_data, dequant_matrix, transformed_data);
-
-    /* place in final output */
-    op = transformed_data;
-    for (i = 0; i < 8; i++) {
-        for (j = 0; j < 8; j++) {
-            sample = *dest + *op;
-            if (sample < 0)
-                *dest = 0;
-            else if (sample > 255)
-                *dest = 255;
-            else
-                *dest = (uint8_t)(sample & 0xFF);
-            op++;
-            dest++;
-        }
-        dest += (stride - 8);
-    }
-
-    /* be a good MMX citizen */
-    emms();
-}
diff --git a/libavcodec/i386/vp3dsp_sse2.c b/libavcodec/i386/vp3dsp_sse2.c
index c8f9158afb..6adfd2f9fa 100644
--- a/libavcodec/i386/vp3dsp_sse2.c
+++ b/libavcodec/i386/vp3dsp_sse2.c
@@ -799,11 +799,12 @@ static unsigned short __align16 SSE2_idct_data[7 * 8] =
 void vp3_dsp_init_sse2(void)
 {
     /* nop */
+av_log(NULL, AV_LOG_INFO, "Hey! SSE2!\n");
 }
 
 
-static void vp3_idct_sse2(int16_t *input_data, int16_t *dequant_matrix,
-     int16_t *output_data)
+void vp3_idct_sse2(int16_t *input_data, int16_t *dequant_matrix,
+    int coeff_count, int16_t *output_data)
 {
     unsigned char *input_bytes = (unsigned char *)input_data;
     unsigned char *dequant_matrix_bytes = (unsigned char *)dequant_matrix;
@@ -832,59 +833,3 @@ static void vp3_idct_sse2(int16_t *input_data, int16_t *dequant_matrix,
         
     SSE2_Column_IDCT();
 }
-
-
-void vp3_idct_put_sse2(int16_t *input_data, int16_t *dequant_matrix,
-    int coeff_count, uint8_t *dest, int stride)
-{
-    int16_t transformed_data[64];
-    int16_t *op;
-    int i, j;
-
-    vp3_idct_sse2(input_data, dequant_matrix, transformed_data);
-
-    /* place in final output */
-    op = transformed_data;
-    for (i = 0; i < 8; i++) {
-        for (j = 0; j < 8; j++) {
-            if (*op < -128)
-                *dest = 0;
-            else if (*op > 127)
-                *dest = 255;
-            else
-                *dest = (uint8_t)(*op + 128);
-            op++;
-            dest++;
-        }
-        dest += (stride - 8);
-    }
-}
-
-
-void vp3_idct_add_sse2(int16_t *input_data, int16_t *dequant_matrix,
-    int coeff_count, uint8_t *dest, int stride)
-{
-    int16_t transformed_data[64];
-    int16_t *op;
-    int i, j;
-    int16_t sample;
-
-    vp3_idct_sse2(input_data, dequant_matrix, transformed_data);
-
-    /* place in final output */
-    op = transformed_data;
-    for (i = 0; i < 8; i++) {
-        for (j = 0; j < 8; j++) {
-            sample = *dest + *op;
-            if (sample < 0)
-                *dest = 0;
-            else if (sample > 255)
-                *dest = 255;
-            else
-                *dest = (uint8_t)(sample & 0xFF);
-            op++;
-            dest++;
-        }
-        dest += (stride - 8);
-    }
-}