radeon: implement pipe_context::bind_sampler_states()
[mesa.git] / src / gallium / drivers / r300 / r300_blit.c
index 9fff3700ed3f5dbc369e91d7e6744f2b4f1a0ca4..4ec68ae6e1c1256210cb05a5fa39a625f1bd2683 100644 (file)
 #include "r300_context.h"
 #include "r300_emit.h"
 #include "r300_texture.h"
+#include "r300_reg.h"
 
 #include "util/u_format.h"
+#include "util/u_half.h"
 #include "util/u_pack_color.h"
 #include "util/u_surface.h"
 
@@ -66,6 +68,7 @@ static void r300_blitter_begin(struct r300_context* r300, enum r300_blitter_op o
     util_blitter_save_vertex_shader(r300->blitter, r300->vs_state.state);
     util_blitter_save_viewport(r300->blitter, &r300->viewport);
     util_blitter_save_scissor(r300->blitter, r300->scissor_state.state);
+    util_blitter_save_sample_mask(r300->blitter, *(unsigned*)r300->sample_mask.state);
     util_blitter_save_vertex_buffer_slot(r300->blitter, r300->vertex_buffer);
     util_blitter_save_vertex_elements(r300->blitter, r300->velems);
 
@@ -133,7 +136,8 @@ static boolean r300_cbzb_clear_allowed(struct r300_context *r300,
     return r300_surface(fb->cbufs[0])->cbzb_allowed;
 }
 
-static boolean r300_fast_zclear_allowed(struct r300_context *r300)
+static boolean r300_fast_zclear_allowed(struct r300_context *r300,
+                                        unsigned clear_buffers)
 {
     struct pipe_framebuffer_state *fb =
         (struct pipe_framebuffer_state*)r300->fb_state.state;
@@ -173,6 +177,26 @@ static uint32_t r300_hiz_clear_value(double depth)
     return r | (r << 8) | (r << 16) | (r << 24);
 }
 
+static void r300_set_clear_color(struct r300_context *r300,
+                                 const union pipe_color_union *color)
+{
+    struct pipe_framebuffer_state *fb =
+        (struct pipe_framebuffer_state*)r300->fb_state.state;
+    union util_color uc;
+
+    memset(&uc, 0, sizeof(uc));
+    util_pack_color(color->f, fb->cbufs[0]->format, &uc);
+
+    if (fb->cbufs[0]->format == PIPE_FORMAT_R16G16B16A16_FLOAT ||
+        fb->cbufs[0]->format == PIPE_FORMAT_R16G16B16X16_FLOAT) {
+        /* (0,1,2,3) maps to (B,G,R,A) */
+        r300->color_clear_value_gb = uc.h[0] | ((uint32_t)uc.h[1] << 16);
+        r300->color_clear_value_ar = uc.h[2] | ((uint32_t)uc.h[3] << 16);
+    } else {
+        r300->color_clear_value = uc.ui;
+    }
+}
+
 DEBUG_GET_ONCE_BOOL_OPTION(hyperz, "RADEON_HYPERZ", FALSE)
 
 /* Clear currently bound buffers. */
@@ -235,13 +259,20 @@ static void r300_clear(struct pipe_context* pipe,
     uint32_t height = fb->height;
     uint32_t hyperz_dcv = hyperz->zb_depthclearvalue;
 
-    /* Enable fast Z clear.
+    /* Use fast Z clear.
      * The zbuffer must be in micro-tiled mode, otherwise it locks up. */
     if (buffers & PIPE_CLEAR_DEPTHSTENCIL) {
         boolean zmask_clear, hiz_clear;
 
-        zmask_clear = r300_fast_zclear_allowed(r300);
-        hiz_clear = r300_hiz_clear_allowed(r300);
+        /* If both depth and stencil are present, they must be cleared together. */
+        if (fb->zsbuf->texture->format == PIPE_FORMAT_S8_UINT_Z24_UNORM &&
+            (buffers & PIPE_CLEAR_DEPTHSTENCIL) != PIPE_CLEAR_DEPTHSTENCIL) {
+            zmask_clear = FALSE;
+            hiz_clear = FALSE;
+        } else {
+            zmask_clear = r300_fast_zclear_allowed(r300, buffers);
+            hiz_clear = r300_hiz_clear_allowed(r300);
+        }
 
         /* If we need Hyper-Z. */
         if (zmask_clear || hiz_clear) {
@@ -260,28 +291,64 @@ static void r300_clear(struct pipe_context* pipe,
 
             /* Setup Hyper-Z clears. */
             if (r300->hyperz_enabled) {
-                DBG(r300, DBG_HYPERZ, "r300: Clear memory: %s%s\n",
-                    zmask_clear ? "ZMASK " : "", hiz_clear ? "HIZ" : "");
-
                 if (zmask_clear) {
                     hyperz_dcv = hyperz->zb_depthclearvalue =
                         r300_depth_clear_value(fb->zsbuf->format, depth, stencil);
 
                     r300_mark_atom_dirty(r300, &r300->zmask_clear);
+                    r300_mark_atom_dirty(r300, &r300->gpu_flush);
                     buffers &= ~PIPE_CLEAR_DEPTHSTENCIL;
                 }
 
                 if (hiz_clear) {
                     r300->hiz_clear_value = r300_hiz_clear_value(depth);
                     r300_mark_atom_dirty(r300, &r300->hiz_clear);
+                    r300_mark_atom_dirty(r300, &r300->gpu_flush);
                 }
                 r300->num_z_clears++;
             }
         }
     }
 
+    /* Use fast color clear for an AA colorbuffer.
+     * The CMASK is shared between all colorbuffers, so we use it
+     * if there is only one colorbuffer bound. */
+    if ((buffers & PIPE_CLEAR_COLOR) && fb->nr_cbufs == 1 &&
+        r300_resource(fb->cbufs[0]->texture)->tex.cmask_dwords) {
+        /* Try to obtain the access to the CMASK if we don't have one. */
+        if (!r300->cmask_access) {
+            r300->cmask_access =
+                r300->rws->cs_request_feature(r300->cs,
+                                              RADEON_FID_R300_CMASK_ACCESS,
+                                              TRUE);
+        }
+
+        /* Setup the clear. */
+        if (r300->cmask_access) {
+            /* Pair the resource with the CMASK to avoid other resources
+             * accessing it. */
+            if (!r300->screen->cmask_resource) {
+                pipe_mutex_lock(r300->screen->cmask_mutex);
+                /* Double checking (first unlocked, then locked). */
+                if (!r300->screen->cmask_resource) {
+                    /* Don't reference this, so that the texture can be
+                     * destroyed while set in cmask_resource.
+                     * Then in texture_destroy, we set cmask_resource to NULL. */
+                    r300->screen->cmask_resource = fb->cbufs[0]->texture;
+                }
+                pipe_mutex_unlock(r300->screen->cmask_mutex);
+            }
+
+            if (r300->screen->cmask_resource == fb->cbufs[0]->texture) {
+                r300_set_clear_color(r300, color);
+                r300_mark_atom_dirty(r300, &r300->cmask_clear);
+                r300_mark_atom_dirty(r300, &r300->gpu_flush);
+                buffers &= ~PIPE_CLEAR_COLOR;
+            }
+        }
+    }
     /* Enable CBZB clear. */
-    if (r300_cbzb_clear_allowed(r300, buffers)) {
+    else if (r300_cbzb_clear_allowed(r300, buffers)) {
         struct r300_surface *surf = r300_surface(fb->cbufs[0]);
 
         hyperz->zb_depthclearvalue =
@@ -296,22 +363,24 @@ static void r300_clear(struct pipe_context* pipe,
 
     /* Clear. */
     if (buffers) {
-        enum pipe_format cformat = fb->nr_cbufs ? fb->cbufs[0]->format : PIPE_FORMAT_NONE;
         /* Clear using the blitter. */
         r300_blitter_begin(r300, R300_CLEAR);
         util_blitter_clear(r300->blitter,
                            width,
                            height,
-                           fb->nr_cbufs,
-                           buffers, cformat, color, depth, stencil);
+                           buffers, color, depth, stencil);
         r300_blitter_end(r300);
-    } else if (r300->zmask_clear.dirty || r300->hiz_clear.dirty) {
+    } else if (r300->zmask_clear.dirty ||
+               r300->hiz_clear.dirty ||
+               r300->cmask_clear.dirty) {
         /* Just clear zmask and hiz now, this does not use the standard draw
          * procedure. */
         /* Calculate zmask_clear and hiz_clear atom sizes. */
         unsigned dwords =
+            r300->gpu_flush.size +
             (r300->zmask_clear.dirty ? r300->zmask_clear.size : 0) +
             (r300->hiz_clear.dirty ? r300->hiz_clear.size : 0) +
+            (r300->cmask_clear.dirty ? r300->cmask_clear.size : 0) +
             r300_get_num_cs_end_dwords(r300);
 
         /* Reserve CS space. */
@@ -320,6 +389,9 @@ static void r300_clear(struct pipe_context* pipe,
         }
 
         /* Emit clear packets. */
+        r300_emit_gpu_flush(r300, r300->gpu_flush.size, r300->gpu_flush.state);
+        r300->gpu_flush.dirty = FALSE;
+
         if (r300->zmask_clear.dirty) {
             r300_emit_zmask_clear(r300, r300->zmask_clear.size,
                                   r300->zmask_clear.state);
@@ -330,6 +402,11 @@ static void r300_clear(struct pipe_context* pipe,
                                 r300->hiz_clear.state);
             r300->hiz_clear.dirty = FALSE;
         }
+        if (r300->cmask_clear.dirty) {
+            r300_emit_cmask_clear(r300, r300->cmask_clear.size,
+                                  r300->cmask_clear.state);
+            r300->cmask_clear.dirty = FALSE;
+        }
     } else {
         assert(0);
     }
@@ -478,6 +555,11 @@ static void r300_resource_copy_region(struct pipe_context *pipe,
         return;
     }
 
+    /* Can't read MSAA textures. */
+    if (src->nr_samples > 1 || dst->nr_samples > 1) {
+        return;
+    }
+
     /* The code below changes the texture format so that the copy can be done
      * on hardware. E.g. depth-stencil surfaces are copied as RGBA
      * colorbuffers. */
@@ -595,20 +677,144 @@ static void r300_resource_copy_region(struct pipe_context *pipe,
     pipe_sampler_view_reference(&src_view, NULL);
 }
 
+static boolean r300_is_simple_msaa_resolve(const struct pipe_blit_info *info)
+{
+    unsigned dst_width = u_minify(info->dst.resource->width0, info->dst.level);
+    unsigned dst_height = u_minify(info->dst.resource->height0, info->dst.level);
+
+    return info->dst.resource->format == info->src.resource->format &&
+           info->dst.resource->format == info->dst.format &&
+           info->src.resource->format == info->src.format &&
+           !info->scissor_enable &&
+           info->mask == PIPE_MASK_RGBA &&
+           dst_width == info->src.resource->width0 &&
+           dst_height == info->src.resource->height0 &&
+           info->dst.box.x == 0 &&
+           info->dst.box.y == 0 &&
+           info->dst.box.width == dst_width &&
+           info->dst.box.height == dst_height &&
+           info->src.box.x == 0 &&
+           info->src.box.y == 0 &&
+           info->src.box.width == dst_width &&
+           info->src.box.height == dst_height &&
+           (r300_resource(info->dst.resource)->tex.microtile != RADEON_LAYOUT_LINEAR ||
+            r300_resource(info->dst.resource)->tex.macrotile[info->dst.level] != RADEON_LAYOUT_LINEAR);
+}
+
+static void r300_simple_msaa_resolve(struct pipe_context *pipe,
+                                     struct pipe_resource *dst,
+                                     unsigned dst_level,
+                                     unsigned dst_layer,
+                                     struct pipe_resource *src,
+                                     enum pipe_format format)
+{
+    struct r300_context *r300 = r300_context(pipe);
+    struct r300_surface *srcsurf, *dstsurf;
+    struct pipe_surface surf_tmpl;
+    struct r300_aa_state *aa = (struct r300_aa_state*)r300->aa_state.state;
+
+    memset(&surf_tmpl, 0, sizeof(surf_tmpl));
+    surf_tmpl.format = format;
+    srcsurf = r300_surface(pipe->create_surface(pipe, src, &surf_tmpl));
+
+    surf_tmpl.format = format;
+    surf_tmpl.u.tex.level = dst_level;
+    surf_tmpl.u.tex.first_layer =
+    surf_tmpl.u.tex.last_layer = dst_layer;
+    dstsurf = r300_surface(pipe->create_surface(pipe, dst, &surf_tmpl));
+
+    /* COLORPITCH should contain the tiling info of the resolve buffer.
+     * The tiling of the AA buffer isn't programmable anyway. */
+    srcsurf->pitch &= ~(R300_COLOR_TILE(1) | R300_COLOR_MICROTILE(3));
+    srcsurf->pitch |= dstsurf->pitch & (R300_COLOR_TILE(1) | R300_COLOR_MICROTILE(3));
+
+    /* Enable AA resolve. */
+    aa->dest = dstsurf;
+    r300->aa_state.size = 8;
+    r300_mark_atom_dirty(r300, &r300->aa_state);
+
+    /* Resolve the surface. */
+    r300_blitter_begin(r300, R300_CLEAR_SURFACE);
+    util_blitter_custom_color(r300->blitter, &srcsurf->base, NULL);
+    r300_blitter_end(r300);
+
+    /* Disable AA resolve. */
+    aa->dest = NULL;
+    r300->aa_state.size = 4;
+    r300_mark_atom_dirty(r300, &r300->aa_state);
+
+    pipe_surface_reference((struct pipe_surface**)&srcsurf, NULL);
+    pipe_surface_reference((struct pipe_surface**)&dstsurf, NULL);
+}
+
+static void r300_msaa_resolve(struct pipe_context *pipe,
+                              const struct pipe_blit_info *info)
+{
+    struct r300_context *r300 = r300_context(pipe);
+    struct pipe_screen *screen = pipe->screen;
+    struct pipe_resource *tmp, templ;
+    struct pipe_blit_info blit;
+
+    assert(info->src.level == 0);
+    assert(info->src.box.z == 0);
+    assert(info->src.box.depth == 1);
+    assert(info->dst.box.depth == 1);
+
+    if (r300_is_simple_msaa_resolve(info)) {
+        r300_simple_msaa_resolve(pipe, info->dst.resource, info->dst.level,
+                                 info->dst.box.z, info->src.resource,
+                                 info->src.format);
+        return;
+    }
+
+    /* resolve into a temporary texture, then blit */
+    memset(&templ, 0, sizeof(templ));
+    templ.target = PIPE_TEXTURE_2D;
+    templ.format = info->src.resource->format;
+    templ.width0 = info->src.resource->width0;
+    templ.height0 = info->src.resource->height0;
+    templ.depth0 = 1;
+    templ.array_size = 1;
+    templ.usage = PIPE_USAGE_STATIC;
+    templ.flags = R300_RESOURCE_FORCE_MICROTILING;
+
+    tmp = screen->resource_create(screen, &templ);
+
+    /* resolve */
+    r300_simple_msaa_resolve(pipe, tmp, 0, 0, info->src.resource,
+                             info->src.format);
+
+    /* blit */
+    blit = *info;
+    blit.src.resource = tmp;
+    blit.src.box.z = 0;
+
+    r300_blitter_begin(r300, R300_BLIT);
+    util_blitter_blit(r300->blitter, &blit);
+    r300_blitter_end(r300);
+
+    pipe_resource_reference(&tmp, NULL);
+}
+
 static void r300_blit(struct pipe_context *pipe,
-                      const struct pipe_blit_info *blit_info)
+                      const struct pipe_blit_info *blit)
 {
     struct r300_context *r300 = r300_context(pipe);
     struct pipe_framebuffer_state *fb =
         (struct pipe_framebuffer_state*)r300->fb_state.state;
-    struct pipe_blit_info info = *blit_info;
+    struct pipe_blit_info info = *blit;
 
-    /* Decompress ZMASK. */
-    if (r300->zmask_in_use && !r300->locked_zbuffer) {
-        if (fb->zsbuf->texture == info.src.resource ||
-            fb->zsbuf->texture == info.dst.resource) {
-            r300_decompress_zmask(r300);
-        }
+    /* MSAA resolve. */
+    if (info.src.resource->nr_samples > 1 &&
+        info.dst.resource->nr_samples <= 1 &&
+        !util_format_is_depth_or_stencil(info.src.resource->format)) {
+        r300_msaa_resolve(pipe, &info);
+        return;
+    }
+
+    /* Can't read MSAA textures. */
+    if (info.src.resource->nr_samples > 1) {
+        return;
     }
 
     /* Blit a combined depth-stencil resource as color.
@@ -616,12 +822,29 @@ static void r300_blit(struct pipe_context *pipe,
     if ((info.mask & PIPE_MASK_S) &&
         info.src.format == PIPE_FORMAT_S8_UINT_Z24_UNORM &&
         info.dst.format == PIPE_FORMAT_S8_UINT_Z24_UNORM) {
-        info.src.format = PIPE_FORMAT_B8G8R8A8_UNORM;
-        info.dst.format = PIPE_FORMAT_B8G8R8A8_UNORM;
-        if (info.mask & PIPE_MASK_Z) {
-            info.mask = PIPE_MASK_RGBA; /* depth+stencil */
+        if (info.dst.resource->nr_samples > 1) {
+            /* Cannot do that with MSAA buffers. */
+            info.mask &= ~PIPE_MASK_S;
+            if (!(info.mask & PIPE_MASK_Z)) {
+                return;
+            }
         } else {
-            info.mask = PIPE_MASK_B; /* stencil only */
+            /* Single-sample buffer. */
+            info.src.format = PIPE_FORMAT_B8G8R8A8_UNORM;
+            info.dst.format = PIPE_FORMAT_B8G8R8A8_UNORM;
+            if (info.mask & PIPE_MASK_Z) {
+                info.mask = PIPE_MASK_RGBA; /* depth+stencil */
+            } else {
+                info.mask = PIPE_MASK_B; /* stencil only */
+            }
+        }
+    }
+
+    /* Decompress ZMASK. */
+    if (r300->zmask_in_use && !r300->locked_zbuffer) {
+        if (fb->zsbuf->texture == info.src.resource ||
+            fb->zsbuf->texture == info.dst.resource) {
+            r300_decompress_zmask(r300);
         }
     }
 
@@ -630,6 +853,11 @@ static void r300_blit(struct pipe_context *pipe,
     r300_blitter_end(r300);
 }
 
+static void r300_flush_resource(struct pipe_context *ctx,
+                               struct pipe_resource *resource)
+{
+}
+
 void r300_init_blit_functions(struct r300_context *r300)
 {
     r300->context.clear = r300_clear;
@@ -637,4 +865,5 @@ void r300_init_blit_functions(struct r300_context *r300)
     r300->context.clear_depth_stencil = r300_clear_depth_stencil;
     r300->context.resource_copy_region = r300_resource_copy_region;
     r300->context.blit = r300_blit;
+    r300->context.flush_resource = r300_flush_resource;
 }