Fix A8 masks

We were using the wrong operation for A8 masks, resulting in badness.
Also, clean up the mask blt to be much simpler.

3 files changed, 24 insertions, 128 deletions

diff --git a/src/amd_lx_exa.c b/src/amd_lx_exa.c
index 5f56f22..eec6a7c 100644
--- a/src/amd_lx_exa.c
+++ b/src/amd_lx_exa.c
@@ -408,9 +408,11 @@ struct blend_ops_t {
   /* PictOpDst */
   { CIMGP_ALPHA_TIMES_A, CIMGP_ALPHA_EQUALS_ONE, CIMGP_CHANNEL_A_DEST }, { },
   /* PictOpOver*/
-  { CIMGP_A_PLUS_BETA_B, CIMGP_CHANNEL_A_ALPHA, CIMGP_CHANNEL_A_SOURCE }, { },
+  { CIMGP_ALPHA_A_PLUS_BETA_B, CIMGP_CHANNEL_A_ALPHA, CIMGP_CHANNEL_A_SOURCE },
+  { },
   /* PictOpOverReverse */
-  { CIMGP_A_PLUS_BETA_B, CIMGP_CHANNEL_A_ALPHA, CIMGP_CHANNEL_A_DEST }, { },
+  { CIMGP_ALPHA_A_PLUS_BETA_B, CIMGP_CHANNEL_A_ALPHA, CIMGP_CHANNEL_A_DEST }, 
+  { },
   /* PictOpIn */
   { CIMGP_ALPHA_TIMES_A, CIMGP_CHANNEL_B_ALPHA, CIMGP_CHANNEL_A_SOURCE }, { },
   /* PictOpInReverse */
@@ -880,8 +882,8 @@ lx_do_composite_mask(PixmapPtr pxDst, unsigned long dstOffset,
   gp_set_bpp(lx_get_bpp_from_format(exaScratch.dstFormat->fmt));
   gp_set_solid_source (exaScratch.srcColor);
 
-  gp_blend_mask_blt(dstOffset, 0, width, height, data,
-		    exaScratch.srcPitch, opPtr->operation,
+  gp_blend_mask_blt(dstOffset, 0, width, height, maskOffset,
+		exaScratch.srcPitch, opPtr->operation,
 		    exaScratch.fourBpp);
 }
 
diff --git a/src/cim/cim_gp.c b/src/cim/cim_gp.c
index 6579f12..6aad40f 100644
--- a/src/cim/cim_gp.c
+++ b/src/cim/cim_gp.c
@@ -3399,14 +3399,11 @@ gp_restore_state(GP_SAVE_RESTORE * gp_state)
 void
 gp_blend_mask_blt(unsigned long dstoffset, unsigned long srcx,
     unsigned long width, unsigned long height,
-    unsigned char *data, long stride, int operation,
+    unsigned long dataoffset, long stride, int operation,
     int fourbpp)
 {
-    unsigned long indent, temp;
-    unsigned long total_dwords, size_dwords;
-    unsigned long dword_count, byte_count;
+    unsigned long indent;
     unsigned long size = ((width << 16) | height);
-    unsigned long ch3_offset, srcoffset;
     unsigned long base, depth_flag;
 
     base = ((gp3_fb_base << 24) + (dstoffset & 0xFFC00000)) |
@@ -3430,27 +3427,14 @@ gp_blend_mask_blt(unsigned long dstoffset, unsigned long srcx,
 
     if (fourbpp) {
         depth_flag = GP3_CH3_SRC_4BPP_ALPHA;
-        indent = (srcx >> 1);
-        srcoffset = (indent & ~3L);
-        indent &= 3;
-        ch3_offset = indent | ((srcx & 1) << 25);
-
-        temp = ((width + (srcx & 1) + 1) >> 1) + indent;
+        indent = (srcx >> 1) & 3;
+        dataoffset += indent | ((srcx & 1) << 25);
     } else {
         depth_flag = GP3_CH3_SRC_8BPP_ALPHA;
-        indent = srcx;
-        srcoffset = (indent & ~3L);
-        indent &= 3;
-        ch3_offset = indent;
-
-        temp = width + indent;
+        indent = srcx & 3;
+        dataoffset += indent;
     }
 
-    total_dwords = (temp + 3) >> 2;
-    size_dwords = (total_dwords << 2) + 8;
-    dword_count = (temp >> 2);
-    byte_count = (temp & 3);
-
     /* SET RASTER MODE REGISTER */
     /* Alpha blending will only apply to RGB when no alpha component is present. */
     /* As 8BPP is not supported for this routine, the only alpha-less mode is    */
@@ -3460,24 +3444,28 @@ gp_blend_mask_blt(unsigned long dstoffset, unsigned long srcx,
         WRITE_COMMAND32(GP3_BLT_RASTER_MODE,
             gp3_bpp |
             GP3_RM_ALPHA_TO_RGB |
-            ((unsigned long) operation << 20) | GP3_RM_SELECT_ALPHA_CHAN_3);
+            ((unsigned long) (operation << 20)) | GP3_RM_SELECT_ALPHA_CHAN_3);
     } else {
         WRITE_COMMAND32(GP3_BLT_RASTER_MODE,
             gp3_bpp |
-            GP3_RM_ALPHA_ALL |
-            ((unsigned long) operation << 20) | GP3_RM_SELECT_ALPHA_CHAN_3);
+            GP3_RM_ALPHA_ALL | ((unsigned long) (operation << 20)) |
+            GP3_RM_SELECT_ALPHA_CHAN_3);
     }
 
     /* WRITE ALL REMAINING REGISTERS */
 
     WRITE_COMMAND32(GP3_BLT_DST_OFFSET, (dstoffset & 0x3FFFFF));
-    WRITE_COMMAND32(GP3_BLT_CH3_OFFSET, ch3_offset);
+
+    /* Set the offset of the CH3 data in memory */
+    WRITE_COMMAND32(GP3_BLT_CH3_OFFSET, dataoffset & 0xFFFFFFF);
+
     WRITE_COMMAND32(GP3_BLT_WID_HEIGHT, size);
     WRITE_COMMAND32(GP3_BLT_CH3_WIDHI, size);
     WRITE_COMMAND32(GP3_BLT_BASE_OFFSET, base);
-    WRITE_COMMAND32(GP3_BLT_CH3_MODE_STR, GP3_CH3_C3EN |
-        GP3_CH3_HST_SRC_ENABLE |
-        depth_flag | ((gp3_blt_flags & CIMGP_BLTFLAGS_PRES_LUT) << 20));
+
+    WRITE_COMMAND32(GP3_BLT_CH3_MODE_STR, GP3_CH3_C3EN | (stride & 0xFFFF) |
+		    depth_flag | ((gp3_blt_flags & CIMGP_BLTFLAGS_PRES_LUT) << 20));
+
     WRITE_COMMAND32(GP3_BLT_MODE, GP3_BM_DST_REQ);
 
     /* START THE BLT */
@@ -3485,98 +3473,4 @@ gp_blend_mask_blt(unsigned long dstoffset, unsigned long srcx,
     WRITE_COMMAND32(GP3_BLT_CMD_HEADER, gp3_cmd_header);
     WRITE_GP32(GP3_CMD_WRITE, gp3_cmd_next);
     gp3_cmd_current = gp3_cmd_next;
-
-    /* WRITE DATA LINE BY LINE
-     * Each line will be created as a separate command buffer entry to allow
-     * line-by-line wrapping and to allow simultaneous rendering by the HW.
-     */
-
-    if (((total_dwords << 2) * height) <= GP3_BLT_1PASS_SIZE &&
-        (gp3_cmd_bottom - gp3_cmd_current) > (GP3_BLT_1PASS_SIZE + 72)) {
-        /* UPDATE THE COMMAND POINTER */
-
-        cim_cmd_ptr = cim_cmd_base_ptr + gp3_cmd_current;
-
-        /* CHECK IF A WRAP WILL BE NEEDED */
-
-        gp3_cmd_next = gp3_cmd_current + ((total_dwords << 2) * height) + 8;
-
-        if ((gp3_cmd_bottom - gp3_cmd_next) <= GP3_MAX_COMMAND_SIZE) {
-            gp3_cmd_next = gp3_cmd_top;
-
-            GP3_WAIT_WRAP(temp);
-            WRITE_COMMAND32(0,
-                GP3_DATA_LOAD_HDR_TYPE | GP3_DATA_LOAD_HDR_WRAP |
-                GP3_DATA_LOAD_HDR_ENABLE);
-        } else {
-            GP3_WAIT_PRIMITIVE(temp);
-            WRITE_COMMAND32(0,
-                GP3_DATA_LOAD_HDR_TYPE | GP3_DATA_LOAD_HDR_ENABLE);
-        }
-
-        /* WRITE DWORD COUNT */
-
-        WRITE_COMMAND32(4,
-            GP3_CH3_HOST_SOURCE_TYPE | (total_dwords * height));
-
-        while (height--) {
-            /* WRITE DATA */
-
-            WRITE_COMMAND_STRING32(8, data, srcoffset, dword_count);
-            WRITE_COMMAND_STRING8(8 + (dword_count << 2), data,
-                srcoffset + (dword_count << 2), byte_count);
-
-            srcoffset += stride;
-            cim_cmd_ptr += total_dwords << 2;
-        }
-
-        WRITE_GP32(GP3_CMD_WRITE, gp3_cmd_next);
-        gp3_cmd_current = gp3_cmd_next;
-    } else {
-        while (height--) {
-            /* UPDATE THE COMMAND POINTER
-             * The WRITE_COMMANDXX macros use a pointer to the current buffer
-             * space.  This is created by adding gp3_cmd_current to the base
-             * pointer.
-             */
-
-            cim_cmd_ptr = cim_cmd_base_ptr + gp3_cmd_current;
-
-            /* CHECK IF A WRAP WILL BE NEEDED */
-
-            gp3_cmd_next = gp3_cmd_current + size_dwords;
-            if ((gp3_cmd_bottom - gp3_cmd_next) <= GP3_MAX_COMMAND_SIZE) {
-                gp3_cmd_next = gp3_cmd_top;
-
-                /* WAIT FOR HARDWARE */
-
-                GP3_WAIT_WRAP(temp);
-                WRITE_COMMAND32(0,
-                    GP3_DATA_LOAD_HDR_TYPE | GP3_DATA_LOAD_HDR_WRAP |
-                    GP3_DATA_LOAD_HDR_ENABLE);
-            } else {
-                /* WAIT FOR AVAILABLE SPACE */
-
-                GP3_WAIT_PRIMITIVE(temp);
-                WRITE_COMMAND32(0,
-                    GP3_DATA_LOAD_HDR_TYPE | GP3_DATA_LOAD_HDR_ENABLE);
-            }
-
-            /* WRITE DWORD COUNT */
-
-            WRITE_COMMAND32(4, GP3_CH3_HOST_SOURCE_TYPE | total_dwords);
-
-            /* WRITE DATA */
-
-            WRITE_COMMAND_STRING32(8, data, srcoffset, dword_count);
-            WRITE_COMMAND_STRING8(8 + (dword_count << 2), data,
-                srcoffset + (dword_count << 2), byte_count);
-
-            /* UPDATE POINTERS */
-
-            srcoffset += stride;
-            WRITE_GP32(GP3_CMD_WRITE, gp3_cmd_next);
-            gp3_cmd_current = gp3_cmd_next;
-        }
-    }
 }
diff --git a/src/cim/cim_rtns.h b/src/cim/cim_rtns.h
index 6fecb4b..60bdcd1 100644
--- a/src/cim/cim_rtns.h
+++ b/src/cim/cim_rtns.h
@@ -126,7 +126,7 @@ extern "C"
         unsigned long width, unsigned long height, unsigned char *data,
         long stride, int fourbpp);
   void gp_blend_mask_blt(unsigned long dstoffset, unsigned long srcx,
-	unsigned long width, unsigned long height, unsigned char *data,
+	unsigned long width, unsigned long height, unsigned long,
         long stride, int operation, int fourbpp);
   void gp_masked_blt(unsigned long dstoffset, unsigned long width,
         unsigned long height, unsigned long mono_srcx,