radeon/r200: add RV200 detiling + add an always tiled flag

passes readpix sanity on the M7.

Signed-off-by: Dave Airlie <airlied@redhat.com>

diff --git a/src/mesa/drivers/dri/radeon/radeon_chipset.h b/src/mesa/drivers/dri/radeon/radeon_chipset.h
index 10cf348fb5a4d93d52f8891f8054a9dfb6e23126..023c12c790d4e7358672a33939215554d89810c7 100644 (file)
--- a/src/mesa/drivers/dri/radeon/radeon_chipset.h
+++ b/src/mesa/drivers/dri/radeon/radeon_chipset.h
@@ -36,5 +36,6 @@ enum {
 #define RADEON_CHIPSET_TCL             (1 << 0)        /* tcl support - any radeon */
 #define RADEON_CHIPSET_BROKEN_STENCIL  (1 << 1)        /* r100 stencil bug */
 #define R200_CHIPSET_YCBCR_BROKEN      (1 << 2)        /* r200 ycbcr bug */
+#define RADEON_CHIPSET_DEPTH_ALWAYS_TILED (1 << 3)      /* M7 and R200s */
 
 #endif /* _RADEON_CHIPSET_H */

diff --git a/src/mesa/drivers/dri/radeon/radeon_fbo.c b/src/mesa/drivers/dri/radeon/radeon_fbo.c
index d8b3b075f14d920ea2332bc1c8c25b2b617b4606..314065703baffae516a6ae4fd0aa321c8b501589 100644 (file)
--- a/src/mesa/drivers/dri/radeon/radeon_fbo.c
+++ b/src/mesa/drivers/dri/radeon/radeon_fbo.c
@@ -70,8 +70,44 @@ radeon_delete_renderbuffer(struct gl_renderbuffer *rb)
   free(rrb);
 }
 
+#if defined(RADEON_R100)
+static GLuint get_depth_z32(const struct radeon_renderbuffer * rrb,
+                              GLint x, GLint y)
+{
+    GLuint ba, address = 0;
+
+    ba = (y >> 4) * (rrb->pitch >> 6) + (x >> 4);
+
+    address |= (x & 0x7) << 2;
+    address |= (y & 0x3) << 5;
+    address |= (((x & 0x10) >> 2) ^ (y & 0x4)) << 5;
+    address |= (ba & 3) << 8;
+    address |= (y & 0x8) << 7;
+    address |= (((x & 0x8) << 1) ^ (y & 0x10)) << 7;
+    address |= (ba & ~0x3) << 10;
+    return address;
+}
+
+static GLuint get_depth_z16(const struct radeon_renderbuffer * rrb,
+                              GLint x, GLint y)
+{
+    GLuint ba, address = 0;                   /* a[0]    = 0           */
+
+    ba = (y / 16) * (rrb->pitch >> 6) + (x / 32);
+
+    address |= (x & 0x7) << 1;                /* a[1..3] = x[0..2]     */
+    address |= (y & 0x7) << 4;                /* a[4..6] = y[0..2]     */
+    address |= (x & 0x8) << 4;                /* a[7]    = x[3]        */
+    address |= (ba & 0x3) << 8;               /* a[8..9] = ba[0..1]    */
+    address |= (y & 0x8) << 7;                /* a[10]   = y[3]        */
+    address |= ((x & 0x10) ^ (y & 0x10)) << 7;/* a[11]   = x[4] ^ y[4] */
+    address |= (ba & ~0x3) << 10;             /* a[12..] = ba[2..] */
+    return address;
+}
+#endif
+
 #if defined(RADEON_R200)
-static GLuint r200_depth_4byte(const struct radeon_renderbuffer * rrb,
+static GLuint get_depth_z32(const struct radeon_renderbuffer * rrb,
                                 GLint x, GLint y)
 {
     GLuint offset;
@@ -91,7 +127,7 @@ static GLuint r200_depth_4byte(const struct radeon_renderbuffer * rrb,
     return offset;
 }
 
-static GLuint r200_depth_2byte(const struct radeon_renderbuffer *rrb,
+static GLuint get_depth_z16(const struct radeon_renderbuffer *rrb,
                               GLint x, GLint y)
 {
    GLuint offset;
@@ -112,6 +148,7 @@ static GLuint r200_depth_2byte(const struct radeon_renderbuffer *rrb,
 
    return offset;
 }
+#endif
 
 static void
 radeon_map_renderbuffer_s8z24(struct gl_context *ctx,
@@ -132,14 +169,14 @@ radeon_map_renderbuffer_s8z24(struct gl_context *ctx,
 
     rrb->map_buffer = malloc(w * h * 4);
     ret = radeon_bo_map(rrb->bo, !!(mode & GL_MAP_WRITE_BIT));
-
+    assert(!ret);
     untiled_s8z24_map = rrb->map_buffer;
     tiled_s8z24_map = rrb->bo->ptr;
 
     for (uint32_t pix_y = 0; pix_y < h; ++ pix_y) {
        for (uint32_t pix_x = 0; pix_x < w; ++pix_x) {
            uint32_t flipped_y = y_flip * (int32_t)(y + pix_y) + y_bias;
-           uint32_t src_offset = r200_depth_4byte(rrb, x + pix_x, flipped_y);
+           uint32_t src_offset = get_depth_z32(rrb, x + pix_x, flipped_y);
            uint32_t dst_offset = pix_y * rrb->map_pitch + pix_x * rrb->cpp;
            untiled_s8z24_map[dst_offset/4] = tiled_s8z24_map[src_offset/4];
        }
@@ -170,6 +207,7 @@ radeon_map_renderbuffer_z16(struct gl_context *ctx,
 
     rrb->map_buffer = malloc(w * h * 2);
     ret = radeon_bo_map(rrb->bo, !!(mode & GL_MAP_WRITE_BIT));
+    assert(!ret);
 
     untiled_z16_map = rrb->map_buffer;
     tiled_z16_map = rrb->bo->ptr;
@@ -177,7 +215,7 @@ radeon_map_renderbuffer_z16(struct gl_context *ctx,
     for (uint32_t pix_y = 0; pix_y < h; ++ pix_y) {
        for (uint32_t pix_x = 0; pix_x < w; ++pix_x) {
            uint32_t flipped_y = y_flip * (int32_t)(y + pix_y) + y_bias;
-           uint32_t src_offset = r200_depth_2byte(rrb, x + pix_x, flipped_y);
+           uint32_t src_offset = get_depth_z16(rrb, x + pix_x, flipped_y);
            uint32_t dst_offset = pix_y * rrb->map_pitch + pix_x * rrb->cpp;
            untiled_z16_map[dst_offset/2] = tiled_z16_map[src_offset/2];
        }
@@ -188,7 +226,6 @@ radeon_map_renderbuffer_z16(struct gl_context *ctx,
     *out_map = rrb->map_buffer;
     *out_stride = rrb->map_pitch;
 }
-#endif
 
 static void
 radeon_map_renderbuffer(struct gl_context *ctx,
@@ -273,19 +310,19 @@ radeon_map_renderbuffer(struct gl_context *ctx,
       radeon_firevertices(rmesa);
    }
 
-#if defined(RADEON_R200)
-   if ((rb->Format == MESA_FORMAT_S8_Z24 || rb->Format == MESA_FORMAT_X8_Z24) && !rrb->has_surface) {
-     radeon_map_renderbuffer_s8z24(ctx, rb, x, y, w, h,
-                                  mode, out_map, out_stride);
-     return;
+   if ((rmesa->radeonScreen->chip_flags & RADEON_CHIPSET_DEPTH_ALWAYS_TILED) && !rrb->has_surface) {
+       if (rb->Format == MESA_FORMAT_S8_Z24 || rb->Format == MESA_FORMAT_X8_Z24) {
+          radeon_map_renderbuffer_s8z24(ctx, rb, x, y, w, h,
+                                        mode, out_map, out_stride);
+          return;
+       }
+       if (rb->Format == MESA_FORMAT_Z16) {
+          radeon_map_renderbuffer_z16(ctx, rb, x, y, w, h,
+                                      mode, out_map, out_stride);
+          return;
+       }
    }
 
-   if (rb->Format == MESA_FORMAT_Z16 && !rrb->has_surface) {
-      radeon_map_renderbuffer_z16(ctx, rb, x, y, w, h,
-                                 mode, out_map, out_stride);
-     return;
-   }
-#endif
    ret = radeon_bo_map(rrb->bo, !!(mode & GL_MAP_WRITE_BIT));
    assert(!ret);
 
@@ -307,12 +344,10 @@ radeon_map_renderbuffer(struct gl_context *ctx,
    *out_stride = flip_stride;
 }
 
-#if defined(RADEON_R200)
 static void
 radeon_unmap_renderbuffer_s8z24(struct gl_context *ctx,
                          struct gl_renderbuffer *rb)
 {
-   struct radeon_context *const rmesa = RADEON_CONTEXT(ctx);
    struct radeon_renderbuffer *rrb = radeon_renderbuffer(rb);
 
    if (!rrb->map_buffer)
@@ -331,7 +366,7 @@ radeon_unmap_renderbuffer_s8z24(struct gl_context *ctx,
        for (uint32_t pix_y = 0; pix_y < rrb->map_h; pix_y++) {
           for (uint32_t pix_x = 0; pix_x < rrb->map_w; pix_x++) {
               uint32_t flipped_y = y_flip * (int32_t)(pix_y + rrb->map_y) + y_bias;
-              uint32_t dst_offset = r200_depth_4byte(rrb, rrb->map_x + pix_x, flipped_y);
+              uint32_t dst_offset = get_depth_z32(rrb, rrb->map_x + pix_x, flipped_y);
               uint32_t src_offset = pix_y * rrb->map_pitch + pix_x * rrb->cpp;
               tiled_s8z24_map[dst_offset/4] = untiled_s8z24_map[src_offset/4];
           }
@@ -346,7 +381,6 @@ static void
 radeon_unmap_renderbuffer_z16(struct gl_context *ctx,
                              struct gl_renderbuffer *rb)
 {
-   struct radeon_context *const rmesa = RADEON_CONTEXT(ctx);
    struct radeon_renderbuffer *rrb = radeon_renderbuffer(rb);
 
    if (!rrb->map_buffer)
@@ -365,7 +399,7 @@ radeon_unmap_renderbuffer_z16(struct gl_context *ctx,
        for (uint32_t pix_y = 0; pix_y < rrb->map_h; pix_y++) {
           for (uint32_t pix_x = 0; pix_x < rrb->map_w; pix_x++) {
               uint32_t flipped_y = y_flip * (int32_t)(pix_y + rrb->map_y) + y_bias;
-              uint32_t dst_offset = r200_depth_2byte(rrb, rrb->map_x + pix_x, flipped_y);
+              uint32_t dst_offset = get_depth_z16(rrb, rrb->map_x + pix_x, flipped_y);
               uint32_t src_offset = pix_y * rrb->map_pitch + pix_x * rrb->cpp;
               tiled_z16_map[dst_offset/2] = untiled_z16_map[src_offset/2];
           }
@@ -375,7 +409,7 @@ radeon_unmap_renderbuffer_z16(struct gl_context *ctx,
    free(rrb->map_buffer);
    rrb->map_buffer = NULL;
 }
-#endif
+
 
 static void
 radeon_unmap_renderbuffer(struct gl_context *ctx,
@@ -385,17 +419,16 @@ radeon_unmap_renderbuffer(struct gl_context *ctx,
    struct radeon_renderbuffer *rrb = radeon_renderbuffer(rb);
    GLboolean ok;
 
-#ifdef RADEON_R200
-   if ((rb->Format == MESA_FORMAT_S8_Z24 || rb->Format == MESA_FORMAT_X8_Z24) && !rrb->has_surface) {
-       radeon_unmap_renderbuffer_s8z24(ctx, rb);
-       return;
-   }
-
-   if (rb->Format == MESA_FORMAT_Z16 && !rrb->has_surface) {
-       radeon_unmap_renderbuffer_z16(ctx, rb);
-       return;
+   if ((rmesa->radeonScreen->chip_flags & RADEON_CHIPSET_DEPTH_ALWAYS_TILED) && !rrb->has_surface) {
+       if (rb->Format == MESA_FORMAT_S8_Z24 || rb->Format == MESA_FORMAT_X8_Z24) {
+          radeon_unmap_renderbuffer_s8z24(ctx, rb);
+          return;
+       }
+       if (rb->Format == MESA_FORMAT_Z16) {
+          radeon_unmap_renderbuffer_z16(ctx, rb);
+          return;
+       }
    }
-#endif
 
    if (!rrb->map_bo) {
           if (rrb->bo)

diff --git a/src/mesa/drivers/dri/radeon/radeon_screen.c b/src/mesa/drivers/dri/radeon/radeon_screen.c
index 74b2ad15d3cb80d78fac9364c80e72bb5d0760eb..408d95b7e4a5c4688945c8e0b2c8d17ba6c97df7 100644 (file)
--- a/src/mesa/drivers/dri/radeon/radeon_screen.c
+++ b/src/mesa/drivers/dri/radeon/radeon_screen.c
@@ -425,7 +425,7 @@ static int radeon_set_screen_flags(radeonScreenPtr screen, int device_id)
    case PCI_CHIP_RADEON_LW:
    case PCI_CHIP_RADEON_LX:
       screen->chip_family = CHIP_FAMILY_RV200;
-      screen->chip_flags = RADEON_CHIPSET_TCL;
+      screen->chip_flags = RADEON_CHIPSET_TCL | RADEON_CHIPSET_DEPTH_ALWAYS_TILED;
       break;
 
 #elif defined(RADEON_R200)
@@ -434,7 +434,7 @@ static int radeon_set_screen_flags(radeonScreenPtr screen, int device_id)
    case PCI_CHIP_R200_QL:
    case PCI_CHIP_R200_QM:
       screen->chip_family = CHIP_FAMILY_R200;
-      screen->chip_flags = RADEON_CHIPSET_TCL;
+      screen->chip_flags = RADEON_CHIPSET_TCL | RADEON_CHIPSET_DEPTH_ALWAYS_TILED;
       break;
 
    case PCI_CHIP_RV250_If:
@@ -443,7 +443,7 @@ static int radeon_set_screen_flags(radeonScreenPtr screen, int device_id)
    case PCI_CHIP_RV250_Lf:
    case PCI_CHIP_RV250_Lg:
       screen->chip_family = CHIP_FAMILY_RV250;
-      screen->chip_flags = R200_CHIPSET_YCBCR_BROKEN | RADEON_CHIPSET_TCL;
+      screen->chip_flags = R200_CHIPSET_YCBCR_BROKEN | RADEON_CHIPSET_TCL | RADEON_CHIPSET_DEPTH_ALWAYS_TILED;
       break;
 
    case PCI_CHIP_RV280_4C6E:
@@ -455,7 +455,7 @@ static int radeon_set_screen_flags(radeonScreenPtr screen, int device_id)
    case PCI_CHIP_RV280_5C61:
    case PCI_CHIP_RV280_5C63:
       screen->chip_family = CHIP_FAMILY_RV280;
-      screen->chip_flags = RADEON_CHIPSET_TCL;
+      screen->chip_flags = RADEON_CHIPSET_TCL | RADEON_CHIPSET_DEPTH_ALWAYS_TILED;
       break;
 
    case PCI_CHIP_RS300_5834:
@@ -463,6 +463,7 @@ static int radeon_set_screen_flags(radeonScreenPtr screen, int device_id)
    case PCI_CHIP_RS350_7834:
    case PCI_CHIP_RS350_7835:
       screen->chip_family = CHIP_FAMILY_RS300;
+      screen->chip_flags = RADEON_CHIPSET_DEPTH_ALWAYS_TILED;
       break;
 #endif