r600g/compute: Fix bug in compute memory pool
[mesa.git] / src / gallium / drivers / r600 / r600_blit.c
index 3e0d89c06f00731d39660b19d6a08242f2d42b1a..60dda28e1845671f15967f029c32c6e2e9ed0a97 100644 (file)
@@ -58,8 +58,8 @@ static void r600_blitter_begin(struct pipe_context *ctx, enum r600_blitter_op op
        util_blitter_save_vertex_buffer_slot(rctx->blitter, rctx->vertex_buffer_state.vb);
        util_blitter_save_vertex_elements(rctx->blitter, rctx->vertex_fetch_shader.cso);
        util_blitter_save_vertex_shader(rctx->blitter, rctx->vs_shader);
-       util_blitter_save_so_targets(rctx->blitter, rctx->num_so_targets,
-                                    (struct pipe_stream_output_target**)rctx->so_targets);
+       util_blitter_save_so_targets(rctx->blitter, rctx->b.streamout.num_targets,
+                                    (struct pipe_stream_output_target**)rctx->b.streamout.targets);
        util_blitter_save_rasterizer(rctx->blitter, rctx->rasterizer_state.cso);
 
        if (op & R600_SAVE_FRAGMENT_STATE) {
@@ -88,6 +88,7 @@ static void r600_blitter_begin(struct pipe_context *ctx, enum r600_blitter_op op
        if ((op & R600_DISABLE_RENDER_COND) && rctx->current_render_cond) {
            util_blitter_save_render_condition(rctx->blitter,
                                               rctx->current_render_cond,
+                                              rctx->current_render_cond_cond,
                                               rctx->current_render_cond_mode);
         }
 }
@@ -126,13 +127,13 @@ void r600_blit_decompress_depth(struct pipe_context *ctx,
        /* XXX Decompressing MSAA depth textures is broken on R6xx.
         * There is also a hardlock if CMASK and FMASK are not present.
         * Just skip this until we find out how to fix it. */
-       if (rctx->chip_class == R600 && max_sample > 0) {
+       if (rctx->b.chip_class == R600 && max_sample > 0) {
                texture->dirty_level_mask = 0;
                return;
        }
 
-       if (rctx->family == CHIP_RV610 || rctx->family == CHIP_RV630 ||
-           rctx->family == CHIP_RV620 || rctx->family == CHIP_RV635)
+       if (rctx->b.family == CHIP_RV610 || rctx->b.family == CHIP_RV630 ||
+           rctx->b.family == CHIP_RV620 || rctx->b.family == CHIP_RV635)
                depth = 0.0f;
        else
                depth = 1.0f;
@@ -170,9 +171,6 @@ void r600_blit_decompress_depth(struct pipe_context *ctx,
                                zsurf = ctx->create_surface(ctx, &texture->resource.b.b, &surf_tmpl);
 
                                surf_tmpl.format = flushed_depth_texture->resource.b.b.format;
-                               surf_tmpl.u.tex.level = level;
-                               surf_tmpl.u.tex.first_layer = layer;
-                               surf_tmpl.u.tex.last_layer = layer;
                                cbsurf = ctx->create_surface(ctx,
                                                &flushed_depth_texture->resource.b.b, &surf_tmpl);
 
@@ -229,12 +227,12 @@ static void r600_blit_decompress_depth_in_place(struct r600_context *rctx,
                        surf_tmpl.u.tex.first_layer = layer;
                        surf_tmpl.u.tex.last_layer = layer;
 
-                       zsurf = rctx->context.create_surface(&rctx->context, &texture->resource.b.b, &surf_tmpl);
+                       zsurf = rctx->b.b.create_surface(&rctx->b.b, &texture->resource.b.b, &surf_tmpl);
 
-                       r600_blitter_begin(&rctx->context, R600_DECOMPRESS);
+                       r600_blitter_begin(&rctx->b.b, R600_DECOMPRESS);
                        util_blitter_custom_depth_stencil(rctx->blitter, zsurf, NULL, ~0,
                                                          rctx->custom_dsa_flush, 1.0f);
-                       r600_blitter_end(&rctx->context);
+                       r600_blitter_end(&rctx->b.b);
 
                        pipe_surface_reference(&zsurf, NULL);
                }
@@ -269,13 +267,13 @@ void r600_decompress_depth_textures(struct r600_context *rctx,
                tex = (struct r600_texture *)view->texture;
                assert(tex->is_depth && !tex->is_flushing_texture);
 
-               if (rctx->chip_class >= EVERGREEN ||
+               if (rctx->b.chip_class >= EVERGREEN ||
                    r600_can_read_depth(tex)) {
                        r600_blit_decompress_depth_in_place(rctx, tex,
                                                   view->u.tex.first_level, view->u.tex.last_level,
                                                   0, util_max_layer(&tex->resource.b.b, view->u.tex.first_level));
                } else {
-                       r600_blit_decompress_depth(&rctx->context, tex, NULL,
+                       r600_blit_decompress_depth(&rctx->b.b, tex, NULL,
                                                   view->u.tex.first_level, view->u.tex.last_level,
                                                   0, util_max_layer(&tex->resource.b.b, view->u.tex.first_level),
                                                   0, u_max_sample(&tex->resource.b.b));
@@ -290,29 +288,10 @@ static void r600_blit_decompress_color(struct pipe_context *ctx,
 {
        struct r600_context *rctx = (struct r600_context *)ctx;
        unsigned layer, level, checked_last_layer, max_layer;
-       void *blend_decompress;
 
        if (!rtex->dirty_level_mask)
                return;
 
-       switch (rctx->screen->msaa_texture_support) {
-       case MSAA_TEXTURE_DECOMPRESSED:
-               blend_decompress = rctx->custom_blend_decompress;
-               break;
-       case MSAA_TEXTURE_COMPRESSED:
-               /* XXX the 2x and 4x cases are broken. */
-               if (rtex->resource.b.b.nr_samples == 8)
-                       blend_decompress = rctx->custom_blend_fmask_decompress;
-               else
-                       blend_decompress = rctx->custom_blend_decompress;
-               break;
-       case MSAA_TEXTURE_SAMPLE_ZERO:
-       default:
-               /* Nothing to do. */
-               rtex->dirty_level_mask = 0;
-               return;
-       }
-
        for (level = first_level; level <= last_level; level++) {
                if (!(rtex->dirty_level_mask & (1 << level)))
                        continue;
@@ -332,7 +311,7 @@ static void r600_blit_decompress_color(struct pipe_context *ctx,
                        cbsurf = ctx->create_surface(ctx, &rtex->resource.b.b, &surf_tmpl);
 
                        r600_blitter_begin(ctx, R600_DECOMPRESS);
-                       util_blitter_custom_color(rctx->blitter, cbsurf, blend_decompress);
+                       util_blitter_custom_color(rctx->blitter, cbsurf, rctx->custom_blend_decompress);
                        r600_blitter_end(ctx);
 
                        pipe_surface_reference(&cbsurf, NULL);
@@ -364,7 +343,7 @@ void r600_decompress_color_textures(struct r600_context *rctx,
                tex = (struct r600_texture *)view->texture;
                assert(tex->cmask_size && tex->fmask_size);
 
-               r600_blit_decompress_color(&rctx->context, tex,
+               r600_blit_decompress_color(&rctx->b.b, tex,
                                           view->u.tex.first_level, view->u.tex.last_level,
                                           0, util_max_layer(&tex->resource.b.b, view->u.tex.first_level));
        }
@@ -383,7 +362,7 @@ static bool r600_decompress_subresource(struct pipe_context *ctx,
        struct r600_texture *rtex = (struct r600_texture*)tex;
 
        if (rtex->is_depth && !rtex->is_flushing_texture) {
-               if (rctx->chip_class >= EVERGREEN ||
+               if (rctx->b.chip_class >= EVERGREEN ||
                    r600_can_read_depth(rtex)) {
                        r600_blit_decompress_depth_in_place(rctx, rtex,
                                                   level, level,
@@ -431,6 +410,60 @@ static boolean is_simple_msaa_resolve(const struct pipe_blit_info *info)
                dst_tile_mode >= RADEON_SURF_MODE_1D;
 }
 
+static void r600_clear_buffer(struct pipe_context *ctx, struct pipe_resource *dst,
+                             unsigned offset, unsigned size, unsigned char 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 bool can_fast_clear_color(struct pipe_context *ctx)
+{
+       struct r600_context *rctx = (struct r600_context *)ctx;
+       struct pipe_framebuffer_state *fb = &rctx->framebuffer.state;
+       int i;
+
+       if (rctx->b.chip_class < EVERGREEN) {
+               return false;
+       }
+
+       for (i = 0; i < fb->nr_cbufs; i++) {
+               struct r600_texture *tex = (struct r600_texture *)fb->cbufs[i]->texture;
+
+               if (tex->cmask_size == 0) {
+                       return false;
+               }
+
+               /* 128-bit formats are unuspported */
+               if (util_format_get_blocksizebits(fb->cbufs[i]->format) > 64) {
+                       return false;
+               }
+
+               /* 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)) {
+                       return false;
+               }
+       }
+
+       return true;
+}
+
 static void r600_clear(struct pipe_context *ctx, unsigned buffers,
                       const union pipe_color_union *color,
                       double depth, unsigned stencil)
@@ -438,6 +471,26 @@ static void r600_clear(struct pipe_context *ctx, unsigned buffers,
        struct r600_context *rctx = (struct r600_context *)ctx;
        struct pipe_framebuffer_state *fb = &rctx->framebuffer.state;
 
+       /* fast color clear on AA framebuffers (EG+) */
+       if ((buffers & PIPE_CLEAR_COLOR) && can_fast_clear_color(ctx)) {
+               int i;
+
+               for (i = 0; i < fb->nr_cbufs; i++) {
+                       struct r600_texture *tex = (struct r600_texture *)fb->cbufs[i]->texture;
+
+                       evergreen_set_clear_color(fb->cbufs[i], color);
+                       r600_clear_buffer(ctx, fb->cbufs[i]->texture,
+                                       tex->cmask_offset, tex->cmask_size, 0);
+                       tex->dirty_level_mask |= 1 << fb->cbufs[i]->u.tex.level;
+               }
+
+               rctx->framebuffer.atom.dirty = true;
+
+               buffers &= ~PIPE_CLEAR_COLOR;
+               if (!buffers)
+                       return;
+       }
+
        /* if hyperz enabled just clear hyperz */
        if (fb->zsbuf && (buffers & PIPE_CLEAR_DEPTH)) {
                struct r600_texture *rtex;
@@ -462,8 +515,7 @@ 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,
-                          fb->nr_cbufs, buffers, fb->nr_cbufs ? fb->cbufs[0]->format : PIPE_FORMAT_NONE,
-                          color, depth, stencil);
+                          buffers, color, depth, stencil);
        r600_blitter_end(ctx);
 
        /* disable fast clear */
@@ -508,21 +560,70 @@ void r600_copy_buffer(struct pipe_context *ctx, struct pipe_resource *dst, unsig
 {
        struct r600_context *rctx = (struct r600_context*)ctx;
 
-       /* CP DMA doesn't work on R600 (flushing seems to be unreliable). */
-       if (rctx->screen->info.drm_minor >= 27 && rctx->chip_class >= R700) {
+       if (rctx->screen->has_cp_dma) {
                r600_cp_dma_copy_buffer(rctx, dst, dstx, src, src_box->x, src_box->width);
        }
        else if (rctx->screen->has_streamout &&
                 /* Require 4-byte alignment. */
                 dstx % 4 == 0 && src_box->x % 4 == 0 && src_box->width % 4 == 0) {
+
+               /* Flush both resources. */
+               r600_flag_resource_cache_flush(rctx, src);
+               r600_flag_resource_cache_flush(rctx, dst);
+
                r600_blitter_begin(ctx, R600_COPY_BUFFER);
                util_blitter_copy_buffer(rctx->blitter, dst, dstx, src, src_box->x, src_box->width);
                r600_blitter_end(ctx);
+
+               /* Flush the dst in case the 3D engine has been prefetching the resource. */
+               r600_flag_resource_cache_flush(rctx, dst);
        } else {
                util_resource_copy_region(ctx, dst, 0, dstx, 0, 0, src, 0, src_box);
        }
 }
 
+static void r600_clear_buffer(struct pipe_context *ctx, struct pipe_resource *dst,
+                             unsigned offset, unsigned size, unsigned char value)
+{
+       struct r600_context *rctx = (struct r600_context*)ctx;
+       uint32_t v = value;
+
+       if (rctx->screen->has_cp_dma &&
+           rctx->b.chip_class >= EVERGREEN &&
+           offset % 4 == 0 && size % 4 == 0) {
+               uint32_t clear_value = v | (v << 8) | (v << 16) | (v << 24);
+
+               evergreen_cp_dma_clear_buffer(rctx, dst, offset, size, clear_value);
+       } else if (rctx->screen->has_streamout && offset % 4 == 0 && size % 4 == 0) {
+               union pipe_color_union clear_value;
+
+               clear_value.ui[0] = v | (v << 8) | (v << 16) | (v << 24);
+
+               r600_flag_resource_cache_flush(rctx, dst);
+
+               r600_blitter_begin(ctx, R600_DISABLE_RENDER_COND);
+               util_blitter_clear_buffer(rctx->blitter, dst, offset, size,
+                                         1, &clear_value);
+               r600_blitter_end(ctx);
+
+               /* Flush again in case the 3D engine has been prefetching the resource. */
+               r600_flag_resource_cache_flush(rctx, dst);
+       } else {
+               char *map = r600_buffer_mmap_sync_with_rings(rctx, r600_resource(dst),
+                                                            PIPE_TRANSFER_WRITE);
+               memset(map + offset, value, size);
+       }
+}
+
+void r600_screen_clear_buffer(struct r600_screen *rscreen, struct pipe_resource *dst,
+                             unsigned offset, unsigned size, unsigned char value)
+{
+       pipe_mutex_lock(rscreen->aux_context_lock);
+       r600_clear_buffer(rscreen->aux_context, dst, offset, size, value);
+       rscreen->aux_context->flush(rscreen->aux_context, NULL, 0);
+       pipe_mutex_unlock(rscreen->aux_context_lock);
+}
+
 static bool util_format_is_subsampled_2x1_32bpp(enum pipe_format format)
 {
        const struct util_format_description *desc = util_format_description(format);
@@ -546,7 +647,6 @@ static void r600_resource_copy_region(struct pipe_context *ctx,
        struct pipe_sampler_view src_templ, *src_view;
        unsigned dst_width, dst_height, src_width0, src_height0, src_widthFL, src_heightFL;
        struct pipe_box sbox, dstbox;
-       bool copy_all_samples;
 
        /* Handle buffers first. */
        if (dst->target == PIPE_BUFFER && src->target == PIPE_BUFFER) {
@@ -650,7 +750,7 @@ static void r600_resource_copy_region(struct pipe_context *ctx,
 
        dst_view = r600_create_surface_custom(ctx, dst, &dst_templ, dst_width, dst_height);
 
-       if (rctx->chip_class >= EVERGREEN) {
+       if (rctx->b.chip_class >= EVERGREEN) {
                src_view = evergreen_create_sampler_view_custom(ctx, src, &src_templ,
                                                                src_width0, src_height0);
        } else {
@@ -658,8 +758,6 @@ static void r600_resource_copy_region(struct pipe_context *ctx,
                                                           src_widthFL, src_heightFL);
        }
 
-       copy_all_samples = rctx->screen->msaa_texture_support != MSAA_TEXTURE_SAMPLE_ZERO;
-
         u_box_3d(dstx, dsty, dstz, abs(src_box->width), abs(src_box->height),
                  abs(src_box->depth), &dstbox);
 
@@ -668,7 +766,7 @@ static void r600_resource_copy_region(struct pipe_context *ctx,
        util_blitter_blit_generic(rctx->blitter, dst_view, &dstbox,
                                  src_view, src_box, src_width0, src_height0,
                                  PIPE_MASK_RGBAZS, PIPE_TEX_FILTER_NEAREST, NULL,
-                                 copy_all_samples);
+                                 TRUE);
        r600_blitter_end(ctx);
 
        pipe_surface_reference(&dst_view, NULL);
@@ -689,7 +787,7 @@ static enum pipe_format int_to_norm_format(enum pipe_format format)
        REPLACE_FORMAT_SIGN(B10G10R10A2, U);
        REPLACE_FORMAT(R8);
        REPLACE_FORMAT(R8G8);
-       REPLACE_FORMAT(R8G8B8);
+       REPLACE_FORMAT(R8G8B8X8);
        REPLACE_FORMAT(R8G8B8A8);
        REPLACE_FORMAT(A8);
        REPLACE_FORMAT(I8);
@@ -697,7 +795,7 @@ static enum pipe_format int_to_norm_format(enum pipe_format format)
        REPLACE_FORMAT(L8A8);
        REPLACE_FORMAT(R16);
        REPLACE_FORMAT(R16G16);
-       REPLACE_FORMAT(R16G16B16);
+       REPLACE_FORMAT(R16G16B16X16);
        REPLACE_FORMAT(R16G16B16A16);
        REPLACE_FORMAT(A16);
        REPLACE_FORMAT(I16);
@@ -719,7 +817,7 @@ static void r600_msaa_color_resolve(struct pipe_context *ctx,
        struct pipe_resource *tmp, templ;
        struct pipe_blit_info blit;
        unsigned sample_mask =
-               rctx->chip_class == CAYMAN ? ~0 :
+               rctx->b.chip_class == CAYMAN ? ~0 :
                ((1ull << MAX2(1, info->src.resource->nr_samples)) - 1);
 
        assert(info->src.level == 0);
@@ -804,9 +902,9 @@ static void r600_blit(struct pipe_context *ctx,
 
 void r600_init_blit_functions(struct r600_context *rctx)
 {
-       rctx->context.clear = r600_clear;
-       rctx->context.clear_render_target = r600_clear_render_target;
-       rctx->context.clear_depth_stencil = r600_clear_depth_stencil;
-       rctx->context.resource_copy_region = r600_resource_copy_region;
-       rctx->context.blit = r600_blit;
+       rctx->b.b.clear = r600_clear;
+       rctx->b.b.clear_render_target = r600_clear_render_target;
+       rctx->b.b.clear_depth_stencil = r600_clear_depth_stencil;
+       rctx->b.b.resource_copy_region = r600_resource_copy_region;
+       rctx->b.b.blit = r600_blit;
 }