gallium: add a cap to enable double rounding opcodes
[mesa.git] / src / gallium / drivers / r600 / r600_blit.c
index dc19b30bda2ca9b1aba9e47c7470955020ad933f..01262a59e906d745bf85c54443cb5c3dd29b0cb0 100644 (file)
@@ -21,8 +21,9 @@
  * USE OR OTHER DEALINGS IN THE SOFTWARE.
  */
 #include "r600_pipe.h"
+#include "compute_memory_pool.h"
+#include "evergreen_compute.h"
 #include "util/u_surface.h"
-#include "util/u_blitter.h"
 #include "util/u_format.h"
 #include "evergreend.h"
 
@@ -42,12 +43,11 @@ enum r600_blitter_op /* bitmask */
        R600_COPY_TEXTURE  = R600_SAVE_FRAGMENT_STATE | R600_SAVE_FRAMEBUFFER | R600_SAVE_TEXTURES |
                             R600_DISABLE_RENDER_COND,
 
-       R600_BLIT          = R600_SAVE_FRAGMENT_STATE | R600_SAVE_FRAMEBUFFER | R600_SAVE_TEXTURES |
-                            R600_DISABLE_RENDER_COND,
+       R600_BLIT          = R600_SAVE_FRAGMENT_STATE | R600_SAVE_FRAMEBUFFER | R600_SAVE_TEXTURES,
 
        R600_DECOMPRESS    = R600_SAVE_FRAGMENT_STATE | R600_SAVE_FRAMEBUFFER | R600_DISABLE_RENDER_COND,
 
-       R600_COLOR_RESOLVE = R600_SAVE_FRAGMENT_STATE | R600_SAVE_FRAMEBUFFER | R600_DISABLE_RENDER_COND
+       R600_COLOR_RESOLVE = R600_SAVE_FRAGMENT_STATE | R600_SAVE_FRAMEBUFFER
 };
 
 static void r600_blitter_begin(struct pipe_context *ctx, enum r600_blitter_op op)
@@ -386,106 +386,6 @@ static bool r600_decompress_subresource(struct pipe_context *ctx,
        return true;
 }
 
-static void r600_clear_buffer(struct pipe_context *ctx, struct pipe_resource *dst,
-                             unsigned offset, unsigned size, unsigned value);
-
-static void evergreen_set_clear_color(struct pipe_surface *cbuf,
-                                     const union pipe_color_union *color)
-{
-       unsigned *clear_value = ((struct r600_texture *)cbuf->texture)->color_clear_value;
-       union util_color uc;
-
-       memset(&uc, 0, sizeof(uc));
-
-       if (util_format_is_pure_uint(cbuf->format)) {
-               util_format_write_4ui(cbuf->format, color->ui, 0, &uc, 0, 0, 0, 1, 1);
-       } else if (util_format_is_pure_sint(cbuf->format)) {
-               util_format_write_4i(cbuf->format, color->i, 0, &uc, 0, 0, 0, 1, 1);
-       } else {
-               util_pack_color(color->f, cbuf->format, &uc);
-       }
-
-       memcpy(clear_value, &uc, 2 * sizeof(uint32_t));
-}
-
-static void evergreen_check_alloc_cmask(struct pipe_context *ctx,
-                                        struct pipe_surface *cbuf)
-{
-        struct r600_context *rctx = (struct r600_context *)ctx;
-        struct r600_texture *tex = (struct r600_texture *)cbuf->texture;
-        struct r600_surface *surf = (struct r600_surface *)cbuf;
-
-        if (tex->cmask_buffer)
-                return;
-
-        r600_texture_init_cmask(&rctx->screen->b, tex);
-
-        /* update colorbuffer state bits */
-        if (tex->cmask_buffer != NULL) {
-                uint64_t va = r600_resource_va(rctx->b.b.screen, &tex->cmask_buffer->b.b);
-                surf->cb_color_cmask = va >> 8;
-                surf->cb_color_cmask_slice = S_028C80_TILE_MAX(tex->cmask.slice_tile_max);
-                surf->cb_color_info |= S_028C70_FAST_CLEAR(1);
-        }
-}
-
-static void r600_try_fast_color_clear(struct r600_context *rctx, unsigned *buffers,
-                                     const union pipe_color_union *color)
-{
-       struct pipe_framebuffer_state *fb = &rctx->framebuffer.state;
-       int i;
-
-       for (i = 0; i < fb->nr_cbufs; i++) {
-               struct r600_texture *tex;
-               unsigned clear_bit = PIPE_CLEAR_COLOR0 << i;
-
-               if (!fb->cbufs[i])
-                       continue;
-
-               /* if this colorbuffer is not being cleared */
-               if (!(*buffers & clear_bit))
-                       continue;
-
-               tex = (struct r600_texture *)fb->cbufs[i]->texture;
-
-               /* 128-bit formats are unusupported */
-               if (util_format_get_blocksizebits(fb->cbufs[i]->format) > 64) {
-                       continue;
-               }
-
-               /* the clear is allowed if all layers are bound */
-               if (fb->cbufs[i]->u.tex.first_layer != 0 ||
-                   fb->cbufs[i]->u.tex.last_layer != util_max_layer(&tex->resource.b.b, 0)) {
-                       continue;
-               }
-
-               /* cannot clear mipmapped textures */
-               if (fb->cbufs[i]->texture->last_level != 0) {
-                       continue;
-               }
-
-               /* only supported on tiled surfaces */
-               if (tex->surface.level[0].mode < RADEON_SURF_MODE_1D) {
-                       continue;
-               }
-
-               /* ensure CMASK is enabled */
-               evergreen_check_alloc_cmask(&rctx->b.b, fb->cbufs[i]);
-               if (tex->cmask.size == 0) {
-                       continue;
-               }
-
-               /* Do the fast clear. */
-               evergreen_set_clear_color(fb->cbufs[i], color);
-               r600_clear_buffer(&rctx->b.b, &tex->cmask_buffer->b.b,
-                                 tex->cmask.offset, tex->cmask.size, 0);
-
-               tex->dirty_level_mask |= 1 << fb->cbufs[i]->u.tex.level;
-               rctx->framebuffer.atom.dirty = true;
-               *buffers &= ~clear_bit;
-       }
-}
-
 static void r600_clear(struct pipe_context *ctx, unsigned buffers,
                       const union pipe_color_union *color,
                       double depth, unsigned stencil)
@@ -494,7 +394,8 @@ static void r600_clear(struct pipe_context *ctx, unsigned buffers,
        struct pipe_framebuffer_state *fb = &rctx->framebuffer.state;
 
        if (buffers & PIPE_CLEAR_COLOR && rctx->b.chip_class >= EVERGREEN) {
-               r600_try_fast_color_clear(rctx, &buffers, color);
+               evergreen_do_fast_color_clear(&rctx->b, fb, &rctx->framebuffer.atom,
+                                             &buffers, color);
        }
 
        if (buffers & PIPE_CLEAR_COLOR) {
@@ -529,7 +430,9 @@ static void r600_clear(struct pipe_context *ctx, unsigned buffers,
                 * disable fast clear for texture array.
                 */
                /* Only use htile for first level */
-               if (rtex->htile_buffer && !level && rtex->surface.array_size == 1) {
+               if (rtex->htile_buffer && !level &&
+                   fb->zsbuf->u.tex.first_layer == 0 &&
+                   fb->zsbuf->u.tex.last_layer == util_max_layer(&rtex->resource.b.b, level)) {
                        if (rtex->depth_clear_value != depth) {
                                rtex->depth_clear_value = depth;
                                rctx->db_state.atom.dirty = true;
@@ -540,7 +443,8 @@ static void r600_clear(struct pipe_context *ctx, unsigned buffers,
        }
 
        r600_blitter_begin(ctx, R600_CLEAR);
-       util_blitter_clear(rctx->blitter, fb->width, fb->height, 1,
+       util_blitter_clear(rctx->blitter, fb->width, fb->height,
+                          util_framebuffer_get_num_layers(fb),
                           buffers, color, depth, stencil);
        r600_blitter_end(ctx);
 
@@ -599,6 +503,12 @@ static void r600_copy_buffer(struct pipe_context *ctx, struct pipe_resource *dst
        } else {
                util_resource_copy_region(ctx, dst, 0, dstx, 0, 0, src, 0, src_box);
        }
+
+       /* The index buffer (VGT) doesn't seem to see the result of the copying.
+        * Can we somehow flush the index buffer cache? Starting a new IB seems
+        * to do the trick. */
+       if (rctx->b.chip_class <= R700)
+               rctx->b.rings.gfx.flush(ctx, RADEON_FLUSH_ASYNC, NULL);
 }
 
 /**
@@ -606,33 +516,57 @@ static void r600_copy_buffer(struct pipe_context *ctx, struct pipe_resource *dst
  * into a single global resource (r600_screen::global_pool).  The means
  * they don't have their own cs_buf handle, so they cannot be passed
  * to r600_copy_buffer() and must be handled separately.
- *
- * XXX: It should be possible to implement this function using
- * r600_copy_buffer() by passing the memory_pool resource as both src
- * and dst and updating dstx and src_box to point to the correct offsets.
- * This would likely perform better than the current implementation.
  */
 static void r600_copy_global_buffer(struct pipe_context *ctx,
                                    struct pipe_resource *dst, unsigned
                                    dstx, struct pipe_resource *src,
                                    const struct pipe_box *src_box)
 {
-       struct pipe_box dst_box; struct pipe_transfer *src_pxfer,
-       *dst_pxfer;
-
-       u_box_1d(dstx, src_box->width, &dst_box);
-       void *src_ptr = ctx->transfer_map(ctx, src, 0, PIPE_TRANSFER_READ,
-                                         src_box, &src_pxfer);
-       void *dst_ptr = ctx->transfer_map(ctx, dst, 0, PIPE_TRANSFER_WRITE,
-                                         &dst_box, &dst_pxfer);
-       memcpy(dst_ptr, src_ptr, src_box->width);
-
-       ctx->transfer_unmap(ctx, src_pxfer);
-       ctx->transfer_unmap(ctx, dst_pxfer);
+       struct r600_context *rctx = (struct r600_context*)ctx;
+       struct compute_memory_pool *pool = rctx->screen->global_pool;
+       struct pipe_box new_src_box = *src_box;
+
+       if (src->bind & PIPE_BIND_GLOBAL) {
+               struct r600_resource_global *rsrc =
+                       (struct r600_resource_global *)src;
+               struct compute_memory_item *item = rsrc->chunk;
+
+               if (is_item_in_pool(item)) {
+                       new_src_box.x += 4 * item->start_in_dw;
+                       src = (struct pipe_resource *)pool->bo;
+               } else {
+                       if (item->real_buffer == NULL) {
+                               item->real_buffer = (struct r600_resource*)
+                                       r600_compute_buffer_alloc_vram(pool->screen,
+                                                                      item->size_in_dw * 4);
+                       }
+                       src = (struct pipe_resource*)item->real_buffer;
+               }
+       }
+       if (dst->bind & PIPE_BIND_GLOBAL) {
+               struct r600_resource_global *rdst =
+                       (struct r600_resource_global *)dst;
+               struct compute_memory_item *item = rdst->chunk;
+
+               if (is_item_in_pool(item)) {
+                       dstx += 4 * item->start_in_dw;
+                       dst = (struct pipe_resource *)pool->bo;
+               } else {
+                       if (item->real_buffer == NULL) {
+                               item->real_buffer = (struct r600_resource*)
+                                       r600_compute_buffer_alloc_vram(pool->screen,
+                                                                      item->size_in_dw * 4);
+                       }
+                       dst = (struct pipe_resource*)item->real_buffer;
+               }
+       }
+
+       r600_copy_buffer(ctx, dst, dstx, src, &new_src_box);
 }
 
 static void r600_clear_buffer(struct pipe_context *ctx, struct pipe_resource *dst,
-                             unsigned offset, unsigned size, unsigned value)
+                             unsigned offset, unsigned size, unsigned value,
+                             bool is_framebuffer)
 {
        struct r600_context *rctx = (struct r600_context*)ctx;
 
@@ -657,23 +591,13 @@ static void r600_clear_buffer(struct pipe_context *ctx, struct pipe_resource *ds
        }
 }
 
-static bool util_format_is_subsampled_2x1_32bpp(enum pipe_format format)
-{
-       const struct util_format_description *desc = util_format_description(format);
-
-       return desc->layout == UTIL_FORMAT_LAYOUT_SUBSAMPLED &&
-              desc->block.width == 2 &&
-              desc->block.height == 1 &&
-              desc->block.bits == 32;
-}
-
-static void r600_resource_copy_region(struct pipe_context *ctx,
-                                     struct pipe_resource *dst,
-                                     unsigned dst_level,
-                                     unsigned dstx, unsigned dsty, unsigned dstz,
-                                     struct pipe_resource *src,
-                                     unsigned src_level,
-                                     const struct pipe_box *src_box)
+void r600_resource_copy_region(struct pipe_context *ctx,
+                              struct pipe_resource *dst,
+                              unsigned dst_level,
+                              unsigned dstx, unsigned dsty, unsigned dstz,
+                              struct pipe_resource *src,
+                              unsigned src_level,
+                              const struct pipe_box *src_box)
 {
        struct r600_context *rctx = (struct r600_context *)ctx;
        struct pipe_surface *dst_view, dst_templ;
@@ -741,7 +665,7 @@ static void r600_resource_copy_region(struct pipe_context *ctx,
 
                src_force_level = src_level;
        } else if (!util_blitter_is_copy_supported(rctx->blitter, dst, src)) {
-               if (util_format_is_subsampled_2x1_32bpp(src->format)) {
+               if (util_format_is_subsampled_422(src->format)) {
 
                        src_templ.format = PIPE_FORMAT_R8G8B8A8_UINT;
                        dst_templ.format = PIPE_FORMAT_R8G8B8A8_UINT;
@@ -882,8 +806,10 @@ static bool do_hardware_msaa_resolve(struct pipe_context *ctx,
            info->src.box.width == dst_width &&
            info->src.box.height == dst_height &&
            info->src.box.depth == 1 &&
-           dst->surface.level[info->dst.level].mode >= RADEON_SURF_MODE_1D) {
-               r600_blitter_begin(ctx, R600_COLOR_RESOLVE);
+           dst->surface.level[info->dst.level].mode >= RADEON_SURF_MODE_1D &&
+           (!dst->cmask.size || !dst->dirty_level_mask) /* dst cannot be fast-cleared */) {
+               r600_blitter_begin(ctx, R600_COLOR_RESOLVE |
+                                  (info->render_condition_enable ? 0 : R600_DISABLE_RENDER_COND));
                util_blitter_custom_resolve_color(rctx->blitter,
                                                  info->dst.resource, info->dst.level,
                                                  info->dst.box.z,
@@ -915,7 +841,12 @@ static void r600_blit(struct pipe_context *ctx,
                return; /* error */
        }
 
-       r600_blitter_begin(ctx, R600_BLIT);
+       if (rctx->screen->b.debug_flags & DBG_FORCE_DMA &&
+           util_try_blit_via_copy_region(ctx, info))
+               return;
+
+       r600_blitter_begin(ctx, R600_BLIT |
+                          (info->render_condition_enable ? 0 : R600_DISABLE_RENDER_COND));
        util_blitter_blit(rctx->blitter, info);
        r600_blitter_end(ctx);
 }
@@ -929,7 +860,7 @@ static void r600_flush_resource(struct pipe_context *ctx,
 
        if (!rtex->is_depth && rtex->cmask.size) {
                r600_blit_decompress_color(ctx, rtex, 0, res->last_level,
-                                          0, res->array_size - 1);
+                                          0, util_max_layer(res, 0));
        }
 }