radeonsi: move all clear() code into si_clear.c
authorMarek Olšák <marek.olsak@amd.com>
Sat, 25 Nov 2017 19:36:35 +0000 (20:36 +0100)
committerMarek Olšák <marek.olsak@amd.com>
Wed, 29 Nov 2017 17:21:30 +0000 (18:21 +0100)
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
src/gallium/drivers/radeon/r600_pipe_common.h
src/gallium/drivers/radeon/r600_texture.c
src/gallium/drivers/radeonsi/Makefile.sources
src/gallium/drivers/radeonsi/meson.build
src/gallium/drivers/radeonsi/si_blit.c
src/gallium/drivers/radeonsi/si_clear.c [new file with mode: 0644]
src/gallium/drivers/radeonsi/si_cp_dma.c
src/gallium/drivers/radeonsi/si_pipe.c
src/gallium/drivers/radeonsi/si_pipe.h

index 2ece4095c4a89bfc899c54bcbbb9f167ce8e1198..36df0c4068a99daf60ce3cb2836d8e6fe65cd005 100644 (file)
@@ -735,6 +735,9 @@ void si_texture_get_fmask_info(struct r600_common_screen *rscreen,
                               struct r600_texture *rtex,
                               unsigned nr_samples,
                               struct r600_fmask_info *out);
+void si_texture_get_cmask_info(struct r600_common_screen *rscreen,
+                              struct r600_texture *rtex,
+                              struct r600_cmask_info *out);
 bool si_init_flushed_depth_texture(struct pipe_context *ctx,
                                   struct pipe_resource *texture,
                                   struct r600_texture **staging);
@@ -757,20 +760,14 @@ struct pipe_surface *si_create_surface_custom(struct pipe_context *pipe,
                                              unsigned width0, unsigned height0,
                                              unsigned width, unsigned height);
 unsigned si_translate_colorswap(enum pipe_format format, bool do_endian_swap);
+void vi_separate_dcc_try_enable(struct r600_common_context *rctx,
+                               struct r600_texture *tex);
 void vi_separate_dcc_start_query(struct pipe_context *ctx,
                                 struct r600_texture *tex);
 void vi_separate_dcc_stop_query(struct pipe_context *ctx,
                                struct r600_texture *tex);
 void vi_separate_dcc_process_and_reset_stats(struct pipe_context *ctx,
                                             struct r600_texture *tex);
-void vi_dcc_clear_level(struct r600_common_context *rctx,
-                       struct r600_texture *rtex,
-                       unsigned level, unsigned clear_value);
-void si_do_fast_color_clear(struct r600_common_context *rctx,
-                           struct pipe_framebuffer_state *fb,
-                           struct r600_atom *fb_state,
-                           unsigned *buffers, ubyte *dirty_cbufs,
-                           const union pipe_color_union *color);
 bool si_texture_disable_dcc(struct r600_common_context *rctx,
                            struct r600_texture *rtex);
 void si_init_screen_texture_functions(struct r600_common_screen *rscreen);
index e3658b465ae191dcb21c0c002441ebec0a595883..46962ebe01fa45c28a8f643818aeb2df0dcacdfb 100644 (file)
@@ -777,9 +777,9 @@ static void r600_texture_allocate_fmask(struct r600_common_screen *rscreen,
        rtex->size = rtex->fmask.offset + rtex->fmask.size;
 }
 
-static void si_texture_get_cmask_info(struct r600_common_screen *rscreen,
-                                     struct r600_texture *rtex,
-                                     struct r600_cmask_info *out)
+void si_texture_get_cmask_info(struct r600_common_screen *rscreen,
+                              struct r600_texture *rtex,
+                              struct r600_cmask_info *out)
 {
        unsigned pipe_interleave_bytes = rscreen->info.pipe_interleave_bytes;
        unsigned num_pipes = rscreen->info.num_tile_pipes;
@@ -842,37 +842,6 @@ static void r600_texture_allocate_cmask(struct r600_common_screen *rscreen,
        rtex->cb_color_info |= S_028C70_FAST_CLEAR(1);
 }
 
-static void r600_texture_alloc_cmask_separate(struct r600_common_screen *rscreen,
-                                             struct r600_texture *rtex)
-{
-       if (rtex->cmask_buffer)
-                return;
-
-       assert(rtex->cmask.size == 0);
-
-       si_texture_get_cmask_info(rscreen, rtex, &rtex->cmask);
-       if (!rtex->cmask.size)
-               return;
-
-       rtex->cmask_buffer = (struct r600_resource *)
-               si_aligned_buffer_create(&rscreen->b,
-                                          R600_RESOURCE_FLAG_UNMAPPABLE,
-                                          PIPE_USAGE_DEFAULT,
-                                          rtex->cmask.size,
-                                          rtex->cmask.alignment);
-       if (rtex->cmask_buffer == NULL) {
-               rtex->cmask.size = 0;
-               return;
-       }
-
-       /* update colorbuffer state bits */
-       rtex->cmask.base_address_reg = rtex->cmask_buffer->gpu_address >> 8;
-
-       rtex->cb_color_info |= S_028C70_FAST_CLEAR(1);
-
-       p_atomic_inc(&rscreen->compressed_colortex_counter);
-}
-
 static void r600_texture_get_htile_size(struct r600_common_screen *rscreen,
                                        struct r600_texture *rtex)
 {
@@ -1973,71 +1942,6 @@ static void r600_surface_destroy(struct pipe_context *pipe,
        FREE(surface);
 }
 
-static void r600_clear_texture(struct pipe_context *pipe,
-                              struct pipe_resource *tex,
-                              unsigned level,
-                              const struct pipe_box *box,
-                              const void *data)
-{
-       struct pipe_screen *screen = pipe->screen;
-       struct r600_texture *rtex = (struct r600_texture*)tex;
-       struct pipe_surface tmpl = {{0}};
-       struct pipe_surface *sf;
-       const struct util_format_description *desc =
-               util_format_description(tex->format);
-
-       tmpl.format = tex->format;
-       tmpl.u.tex.first_layer = box->z;
-       tmpl.u.tex.last_layer = box->z + box->depth - 1;
-       tmpl.u.tex.level = level;
-       sf = pipe->create_surface(pipe, tex, &tmpl);
-       if (!sf)
-               return;
-
-       if (rtex->is_depth) {
-               unsigned clear;
-               float depth;
-               uint8_t stencil = 0;
-
-               /* Depth is always present. */
-               clear = PIPE_CLEAR_DEPTH;
-               desc->unpack_z_float(&depth, 0, data, 0, 1, 1);
-
-               if (rtex->surface.has_stencil) {
-                       clear |= PIPE_CLEAR_STENCIL;
-                       desc->unpack_s_8uint(&stencil, 0, data, 0, 1, 1);
-               }
-
-               pipe->clear_depth_stencil(pipe, sf, clear, depth, stencil,
-                                         box->x, box->y,
-                                         box->width, box->height, false);
-       } else {
-               union pipe_color_union color;
-
-               /* pipe_color_union requires the full vec4 representation. */
-               if (util_format_is_pure_uint(tex->format))
-                       desc->unpack_rgba_uint(color.ui, 0, data, 0, 1, 1);
-               else if (util_format_is_pure_sint(tex->format))
-                       desc->unpack_rgba_sint(color.i, 0, data, 0, 1, 1);
-               else
-                       desc->unpack_rgba_float(color.f, 0, data, 0, 1, 1);
-
-               if (screen->is_format_supported(screen, tex->format,
-                                               tex->target, 0,
-                                               PIPE_BIND_RENDER_TARGET)) {
-                       pipe->clear_render_target(pipe, sf, &color,
-                                                 box->x, box->y,
-                                                 box->width, box->height, false);
-               } else {
-                       /* Software fallback - just for R9G9B9E5_FLOAT */
-                       util_clear_render_target(pipe, sf, &color,
-                                                box->x, box->y,
-                                                box->width, box->height);
-               }
-       }
-       pipe_surface_reference(&sf, NULL);
-}
-
 unsigned si_translate_colorswap(enum pipe_format format, bool do_endian_swap)
 {
        const struct util_format_description *desc = util_format_description(format);
@@ -2221,8 +2125,8 @@ static bool vi_should_enable_separate_dcc(struct r600_texture *tex)
 }
 
 /* Called by fast clear. */
-static void vi_separate_dcc_try_enable(struct r600_common_context *rctx,
-                                      struct r600_texture *tex)
+void vi_separate_dcc_try_enable(struct r600_common_context *rctx,
+                               struct r600_texture *tex)
 {
        /* The intent is to use this with shared displayable back buffers,
         * but it's not strictly limited only to them.
@@ -2336,446 +2240,6 @@ void vi_separate_dcc_process_and_reset_stats(struct pipe_context *ctx,
        }
 }
 
-/* FAST COLOR CLEAR */
-
-static void evergreen_set_clear_color(struct r600_texture *rtex,
-                                     enum pipe_format surface_format,
-                                     const union pipe_color_union *color)
-{
-       union util_color uc;
-
-       memset(&uc, 0, sizeof(uc));
-
-       if (rtex->surface.bpe == 16) {
-               /* DCC fast clear only:
-                *   CLEAR_WORD0 = R = G = B
-                *   CLEAR_WORD1 = A
-                */
-               assert(color->ui[0] == color->ui[1] &&
-                      color->ui[0] == color->ui[2]);
-               uc.ui[0] = color->ui[0];
-               uc.ui[1] = color->ui[3];
-       } else if (util_format_is_pure_uint(surface_format)) {
-               util_format_write_4ui(surface_format, color->ui, 0, &uc, 0, 0, 0, 1, 1);
-       } else if (util_format_is_pure_sint(surface_format)) {
-               util_format_write_4i(surface_format, color->i, 0, &uc, 0, 0, 0, 1, 1);
-       } else {
-               util_pack_color(color->f, surface_format, &uc);
-       }
-
-       memcpy(rtex->color_clear_value, &uc, 2 * sizeof(uint32_t));
-}
-
-static bool vi_get_fast_clear_parameters(enum pipe_format surface_format,
-                                        const union pipe_color_union *color,
-                                        uint32_t* reset_value,
-                                        bool* clear_words_needed)
-{
-       bool values[4] = {};
-       int i;
-       bool main_value = false;
-       bool extra_value = false;
-       int extra_channel;
-
-       /* This is needed to get the correct DCC clear value for luminance formats.
-        * 1) Get the linear format (because the next step can't handle L8_SRGB).
-        * 2) Convert luminance to red. (the real hw format for luminance)
-        */
-       surface_format = util_format_linear(surface_format);
-       surface_format = util_format_luminance_to_red(surface_format);
-
-       const struct util_format_description *desc = util_format_description(surface_format);
-
-       if (desc->block.bits == 128 &&
-           (color->ui[0] != color->ui[1] ||
-            color->ui[0] != color->ui[2]))
-               return false;
-
-       *clear_words_needed = true;
-       *reset_value = 0x20202020U;
-
-       /* If we want to clear without needing a fast clear eliminate step, we
-        * can set each channel to 0 or 1 (or 0/max for integer formats). We
-        * have two sets of flags, one for the last or first channel(extra) and
-        * one for the other channels(main).
-        */
-
-       if (surface_format == PIPE_FORMAT_R11G11B10_FLOAT ||
-           surface_format == PIPE_FORMAT_B5G6R5_UNORM ||
-           surface_format == PIPE_FORMAT_B5G6R5_SRGB ||
-           util_format_is_alpha(surface_format)) {
-               extra_channel = -1;
-       } else if (desc->layout == UTIL_FORMAT_LAYOUT_PLAIN) {
-               if(si_translate_colorswap(surface_format, false) <= 1)
-                       extra_channel = desc->nr_channels - 1;
-               else
-                       extra_channel = 0;
-       } else
-               return true;
-
-       for (i = 0; i < 4; ++i) {
-               int index = desc->swizzle[i] - PIPE_SWIZZLE_X;
-
-               if (desc->swizzle[i] < PIPE_SWIZZLE_X ||
-                   desc->swizzle[i] > PIPE_SWIZZLE_W)
-                       continue;
-
-               if (desc->channel[i].pure_integer &&
-                   desc->channel[i].type == UTIL_FORMAT_TYPE_SIGNED) {
-                       /* Use the maximum value for clamping the clear color. */
-                       int max = u_bit_consecutive(0, desc->channel[i].size - 1);
-
-                       values[i] = color->i[i] != 0;
-                       if (color->i[i] != 0 && MIN2(color->i[i], max) != max)
-                               return true;
-               } else if (desc->channel[i].pure_integer &&
-                          desc->channel[i].type == UTIL_FORMAT_TYPE_UNSIGNED) {
-                       /* Use the maximum value for clamping the clear color. */
-                       unsigned max = u_bit_consecutive(0, desc->channel[i].size);
-
-                       values[i] = color->ui[i] != 0U;
-                       if (color->ui[i] != 0U && MIN2(color->ui[i], max) != max)
-                               return true;
-               } else {
-                       values[i] = color->f[i] != 0.0F;
-                       if (color->f[i] != 0.0F && color->f[i] != 1.0F)
-                               return true;
-               }
-
-               if (index == extra_channel)
-                       extra_value = values[i];
-               else
-                       main_value = values[i];
-       }
-
-       for (int i = 0; i < 4; ++i)
-               if (values[i] != main_value &&
-                   desc->swizzle[i] - PIPE_SWIZZLE_X != extra_channel &&
-                   desc->swizzle[i] >= PIPE_SWIZZLE_X &&
-                   desc->swizzle[i] <= PIPE_SWIZZLE_W)
-                       return true;
-
-       *clear_words_needed = false;
-       if (main_value)
-               *reset_value |= 0x80808080U;
-
-       if (extra_value)
-               *reset_value |= 0x40404040U;
-       return true;
-}
-
-void vi_dcc_clear_level(struct r600_common_context *rctx,
-                       struct r600_texture *rtex,
-                       unsigned level, unsigned clear_value)
-{
-       struct pipe_resource *dcc_buffer;
-       uint64_t dcc_offset, clear_size;
-
-       assert(vi_dcc_enabled(rtex, level));
-
-       if (rtex->dcc_separate_buffer) {
-               dcc_buffer = &rtex->dcc_separate_buffer->b.b;
-               dcc_offset = 0;
-       } else {
-               dcc_buffer = &rtex->resource.b.b;
-               dcc_offset = rtex->dcc_offset;
-       }
-
-       if (rctx->chip_class >= GFX9) {
-               /* Mipmap level clears aren't implemented. */
-               assert(rtex->resource.b.b.last_level == 0);
-               /* MSAA needs a different clear size. */
-               assert(rtex->resource.b.b.nr_samples <= 1);
-               clear_size = rtex->surface.dcc_size;
-       } else {
-               unsigned num_layers = util_max_layer(&rtex->resource.b.b, level) + 1;
-
-               /* If this is 0, fast clear isn't possible. (can occur with MSAA) */
-               assert(rtex->surface.u.legacy.level[level].dcc_fast_clear_size);
-               /* Layered MSAA DCC fast clears need to clear dcc_fast_clear_size
-                * bytes for each layer. This is not currently implemented, and
-                * therefore MSAA DCC isn't even enabled with multiple layers.
-                */
-               assert(rtex->resource.b.b.nr_samples <= 1 || num_layers == 1);
-
-               dcc_offset += rtex->surface.u.legacy.level[level].dcc_offset;
-               clear_size = rtex->surface.u.legacy.level[level].dcc_fast_clear_size *
-                            num_layers;
-       }
-
-       rctx->clear_buffer(&rctx->b, dcc_buffer, dcc_offset, clear_size,
-                          clear_value, R600_COHERENCY_CB_META);
-}
-
-/* Set the same micro tile mode as the destination of the last MSAA resolve.
- * This allows hitting the MSAA resolve fast path, which requires that both
- * src and dst micro tile modes match.
- */
-static void si_set_optimal_micro_tile_mode(struct r600_common_screen *rscreen,
-                                          struct r600_texture *rtex)
-{
-       if (rtex->resource.b.is_shared ||
-           rtex->resource.b.b.nr_samples <= 1 ||
-           rtex->surface.micro_tile_mode == rtex->last_msaa_resolve_target_micro_mode)
-               return;
-
-       assert(rscreen->chip_class >= GFX9 ||
-              rtex->surface.u.legacy.level[0].mode == RADEON_SURF_MODE_2D);
-       assert(rtex->resource.b.b.last_level == 0);
-
-       if (rscreen->chip_class >= GFX9) {
-               /* 4K or larger tiles only. 0 is linear. 1-3 are 256B tiles. */
-               assert(rtex->surface.u.gfx9.surf.swizzle_mode >= 4);
-
-               /* If you do swizzle_mode % 4, you'll get:
-                *   0 = Depth
-                *   1 = Standard,
-                *   2 = Displayable
-                *   3 = Rotated
-                *
-                * Depth-sample order isn't allowed:
-                */
-               assert(rtex->surface.u.gfx9.surf.swizzle_mode % 4 != 0);
-
-               switch (rtex->last_msaa_resolve_target_micro_mode) {
-               case RADEON_MICRO_MODE_DISPLAY:
-                       rtex->surface.u.gfx9.surf.swizzle_mode &= ~0x3;
-                       rtex->surface.u.gfx9.surf.swizzle_mode += 2; /* D */
-                       break;
-               case RADEON_MICRO_MODE_THIN:
-                       rtex->surface.u.gfx9.surf.swizzle_mode &= ~0x3;
-                       rtex->surface.u.gfx9.surf.swizzle_mode += 1; /* S */
-                       break;
-               case RADEON_MICRO_MODE_ROTATED:
-                       rtex->surface.u.gfx9.surf.swizzle_mode &= ~0x3;
-                       rtex->surface.u.gfx9.surf.swizzle_mode += 3; /* R */
-                       break;
-               default: /* depth */
-                       assert(!"unexpected micro mode");
-                       return;
-               }
-       } else if (rscreen->chip_class >= CIK) {
-               /* These magic numbers were copied from addrlib. It doesn't use
-                * any definitions for them either. They are all 2D_TILED_THIN1
-                * modes with different bpp and micro tile mode.
-                */
-               switch (rtex->last_msaa_resolve_target_micro_mode) {
-               case RADEON_MICRO_MODE_DISPLAY:
-                       rtex->surface.u.legacy.tiling_index[0] = 10;
-                       break;
-               case RADEON_MICRO_MODE_THIN:
-                       rtex->surface.u.legacy.tiling_index[0] = 14;
-                       break;
-               case RADEON_MICRO_MODE_ROTATED:
-                       rtex->surface.u.legacy.tiling_index[0] = 28;
-                       break;
-               default: /* depth, thick */
-                       assert(!"unexpected micro mode");
-                       return;
-               }
-       } else { /* SI */
-               switch (rtex->last_msaa_resolve_target_micro_mode) {
-               case RADEON_MICRO_MODE_DISPLAY:
-                       switch (rtex->surface.bpe) {
-                       case 1:
-                            rtex->surface.u.legacy.tiling_index[0] = 10;
-                            break;
-                       case 2:
-                            rtex->surface.u.legacy.tiling_index[0] = 11;
-                            break;
-                       default: /* 4, 8 */
-                            rtex->surface.u.legacy.tiling_index[0] = 12;
-                            break;
-                       }
-                       break;
-               case RADEON_MICRO_MODE_THIN:
-                       switch (rtex->surface.bpe) {
-                       case 1:
-                                rtex->surface.u.legacy.tiling_index[0] = 14;
-                                break;
-                       case 2:
-                                rtex->surface.u.legacy.tiling_index[0] = 15;
-                                break;
-                       case 4:
-                                rtex->surface.u.legacy.tiling_index[0] = 16;
-                                break;
-                       default: /* 8, 16 */
-                                rtex->surface.u.legacy.tiling_index[0] = 17;
-                                break;
-                       }
-                       break;
-               default: /* depth, thick */
-                       assert(!"unexpected micro mode");
-                       return;
-               }
-       }
-
-       rtex->surface.micro_tile_mode = rtex->last_msaa_resolve_target_micro_mode;
-
-       p_atomic_inc(&rscreen->dirty_tex_counter);
-}
-
-void si_do_fast_color_clear(struct r600_common_context *rctx,
-                                  struct pipe_framebuffer_state *fb,
-                                  struct r600_atom *fb_state,
-                                  unsigned *buffers, ubyte *dirty_cbufs,
-                                  const union pipe_color_union *color)
-{
-       int i;
-
-       /* This function is broken in BE, so just disable this path for now */
-#ifdef PIPE_ARCH_BIG_ENDIAN
-       return;
-#endif
-
-       if (rctx->render_cond)
-               return;
-
-       for (i = 0; i < fb->nr_cbufs; i++) {
-               struct r600_texture *tex;
-               unsigned clear_bit = PIPE_CLEAR_COLOR0 << i;
-               unsigned level = fb->cbufs[i]->u.tex.level;
-
-               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;
-
-               /* 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.is_linear) {
-                       continue;
-               }
-
-               /* shared textures can't use fast clear without an explicit flush,
-                * because there is no way to communicate the clear color among
-                * all clients
-                */
-               if (tex->resource.b.is_shared &&
-                   !(tex->resource.external_usage & PIPE_HANDLE_USAGE_EXPLICIT_FLUSH))
-                       continue;
-
-               /* fast color clear with 1D tiling doesn't work on old kernels and CIK */
-               if (rctx->chip_class == CIK &&
-                   tex->surface.u.legacy.level[0].mode == RADEON_SURF_MODE_1D &&
-                   rctx->screen->info.drm_major == 2 &&
-                   rctx->screen->info.drm_minor < 38) {
-                       continue;
-               }
-
-               /* Fast clear is the most appropriate place to enable DCC for
-                * displayable surfaces.
-                */
-               if (rctx->chip_class >= VI &&
-                   !(rctx->screen->debug_flags & DBG(NO_DCC_FB))) {
-                       vi_separate_dcc_try_enable(rctx, tex);
-
-                       /* RB+ isn't supported with a CMASK clear only on Stoney,
-                        * so all clears are considered to be hypothetically slow
-                        * clears, which is weighed when determining whether to
-                        * enable separate DCC.
-                        */
-                       if (tex->dcc_gather_statistics &&
-                           rctx->family == CHIP_STONEY)
-                               tex->num_slow_clears++;
-               }
-
-               /* Try to clear DCC first, otherwise try CMASK. */
-               if (vi_dcc_enabled(tex, 0)) {
-                       uint32_t reset_value;
-                       bool clear_words_needed, cleared_cmask = false;
-
-                       if (rctx->screen->debug_flags & DBG(NO_DCC_CLEAR))
-                               continue;
-
-                       /* This can only occur with MSAA. */
-                       if (rctx->chip_class == VI &&
-                           !tex->surface.u.legacy.level[level].dcc_fast_clear_size)
-                               continue;
-
-                       if (!vi_get_fast_clear_parameters(fb->cbufs[i]->format,
-                                                         color, &reset_value,
-                                                         &clear_words_needed))
-                               continue;
-
-                       /* DCC fast clear with MSAA should clear CMASK to 0xC. */
-                       if (tex->resource.b.b.nr_samples >= 2 && tex->cmask.size) {
-                               /* TODO: This doesn't work with MSAA. */
-                               if (clear_words_needed)
-                                       continue;
-
-                               rctx->clear_buffer(&rctx->b, &tex->cmask_buffer->b.b,
-                                                  tex->cmask.offset, tex->cmask.size,
-                                                  0xCCCCCCCC, R600_COHERENCY_CB_META);
-                               cleared_cmask = true;
-                       }
-
-                       vi_dcc_clear_level(rctx, tex, 0, reset_value);
-
-                       if (clear_words_needed || cleared_cmask) {
-                               bool need_compressed_update = !tex->dirty_level_mask;
-
-                               tex->dirty_level_mask |= 1 << level;
-
-                               if (need_compressed_update)
-                                       p_atomic_inc(&rctx->screen->compressed_colortex_counter);
-                       }
-                       tex->separate_dcc_dirty = true;
-               } else {
-                       /* 128-bit formats are unusupported */
-                       if (tex->surface.bpe > 8) {
-                               continue;
-                       }
-
-                       /* RB+ doesn't work with CMASK fast clear on Stoney. */
-                       if (rctx->family == CHIP_STONEY)
-                               continue;
-
-                       /* ensure CMASK is enabled */
-                       r600_texture_alloc_cmask_separate(rctx->screen, tex);
-                       if (tex->cmask.size == 0) {
-                               continue;
-                       }
-
-                       /* Do the fast clear. */
-                       rctx->clear_buffer(&rctx->b, &tex->cmask_buffer->b.b,
-                                          tex->cmask.offset, tex->cmask.size, 0,
-                                          R600_COHERENCY_CB_META);
-
-                       bool need_compressed_update = !tex->dirty_level_mask;
-
-                       tex->dirty_level_mask |= 1 << level;
-
-                       if (need_compressed_update)
-                               p_atomic_inc(&rctx->screen->compressed_colortex_counter);
-               }
-
-               /* We can change the micro tile mode before a full clear. */
-               si_set_optimal_micro_tile_mode(rctx->screen, tex);
-
-               evergreen_set_clear_color(tex, fb->cbufs[i]->format, color);
-
-               if (dirty_cbufs)
-                       *dirty_cbufs |= 1 << i;
-               rctx->set_atom_dirty(rctx, fb_state, true);
-               *buffers &= ~clear_bit;
-       }
-}
-
 static struct pipe_memory_object *
 r600_memobj_from_handle(struct pipe_screen *screen,
                        struct winsys_handle *whandle,
@@ -2922,5 +2386,4 @@ void si_init_context_texture_functions(struct r600_common_context *rctx)
 {
        rctx->b.create_surface = r600_create_surface;
        rctx->b.surface_destroy = r600_surface_destroy;
-       rctx->b.clear_texture = r600_clear_texture;
 }
index 8dd33f60a8bff203c1ad791ff9f70231236ad344..18cd7d14c18c23e824a142e6dc39783e368afc60 100644 (file)
@@ -6,6 +6,7 @@ C_SOURCES := \
        cik_sdma.c \
        driinfo_radeonsi.h \
        si_blit.c \
+       si_clear.c \
        si_compute.c \
        si_compute.h \
        si_cp_dma.c \
index 974004db67f8d935d5e5fe6a67155f320b782e50..a04ea4c03fbfb5b83593b8d95f44286dc4ab06c4 100644 (file)
@@ -22,6 +22,7 @@ files_libradeonsi = files(
   'cik_sdma.c',
   'driinfo_radeonsi.h',
   'si_blit.c',
+  'si_clear.c',
   'si_compute.c',
   'si_compute.h',
   'si_cp_dma.c',
index abf25b605e800b9de8593096f9d317cff64dad76..e80bf5c24ef65aa8c160275b13881179f0dba3b6 100644 (file)
 #include "util/u_log.h"
 #include "util/u_surface.h"
 
-enum si_blitter_op /* bitmask */
-{
-       SI_SAVE_TEXTURES      = 1,
-       SI_SAVE_FRAMEBUFFER   = 2,
-       SI_SAVE_FRAGMENT_STATE = 4,
-       SI_DISABLE_RENDER_COND = 8,
-
-       SI_CLEAR         = SI_SAVE_FRAGMENT_STATE,
-
-       SI_CLEAR_SURFACE = SI_SAVE_FRAMEBUFFER | SI_SAVE_FRAGMENT_STATE,
-
+enum {
        SI_COPY          = SI_SAVE_FRAMEBUFFER | SI_SAVE_TEXTURES |
                           SI_SAVE_FRAGMENT_STATE | SI_DISABLE_RENDER_COND,
 
@@ -50,7 +40,7 @@ enum si_blitter_op /* bitmask */
        SI_COLOR_RESOLVE = SI_SAVE_FRAMEBUFFER | SI_SAVE_FRAGMENT_STATE
 };
 
-static void si_blitter_begin(struct pipe_context *ctx, enum si_blitter_op op)
+void si_blitter_begin(struct pipe_context *ctx, enum si_blitter_op op)
 {
        struct si_context *sctx = (struct si_context *)ctx;
 
@@ -87,7 +77,7 @@ static void si_blitter_begin(struct pipe_context *ctx, enum si_blitter_op op)
                sctx->b.render_cond_force_off = true;
 }
 
-static void si_blitter_end(struct pipe_context *ctx)
+void si_blitter_end(struct pipe_context *ctx)
 {
        struct si_context *sctx = (struct si_context *)ctx;
 
@@ -819,155 +809,6 @@ void si_decompress_textures(struct si_context *sctx, unsigned shader_mask)
        si_check_render_feedback(sctx);
 }
 
-static void si_clear(struct pipe_context *ctx, unsigned buffers,
-                    const union pipe_color_union *color,
-                    double depth, unsigned stencil)
-{
-       struct si_context *sctx = (struct si_context *)ctx;
-       struct pipe_framebuffer_state *fb = &sctx->framebuffer.state;
-       struct pipe_surface *zsbuf = fb->zsbuf;
-       struct r600_texture *zstex =
-               zsbuf ? (struct r600_texture*)zsbuf->texture : NULL;
-
-       if (buffers & PIPE_CLEAR_COLOR) {
-               si_do_fast_color_clear(&sctx->b, fb,
-                                             &sctx->framebuffer.atom, &buffers,
-                                             &sctx->framebuffer.dirty_cbufs,
-                                             color);
-               if (!buffers)
-                       return; /* all buffers have been fast cleared */
-       }
-
-       if (buffers & PIPE_CLEAR_COLOR) {
-               int i;
-
-               /* These buffers cannot use fast clear, make sure to disable expansion. */
-               for (i = 0; i < fb->nr_cbufs; i++) {
-                       struct r600_texture *tex;
-
-                       /* If not clearing this buffer, skip. */
-                       if (!(buffers & (PIPE_CLEAR_COLOR0 << i)))
-                               continue;
-
-                       if (!fb->cbufs[i])
-                               continue;
-
-                       tex = (struct r600_texture *)fb->cbufs[i]->texture;
-                       if (tex->fmask.size == 0)
-                               tex->dirty_level_mask &= ~(1 << fb->cbufs[i]->u.tex.level);
-               }
-       }
-
-       if (zstex &&
-           r600_htile_enabled(zstex, zsbuf->u.tex.level) &&
-           zsbuf->u.tex.first_layer == 0 &&
-           zsbuf->u.tex.last_layer == util_max_layer(&zstex->resource.b.b, 0)) {
-               /* TC-compatible HTILE only supports depth clears to 0 or 1. */
-               if (buffers & PIPE_CLEAR_DEPTH &&
-                   (!zstex->tc_compatible_htile ||
-                    depth == 0 || depth == 1)) {
-                       /* Need to disable EXPCLEAR temporarily if clearing
-                        * to a new value. */
-                       if (!zstex->depth_cleared || zstex->depth_clear_value != depth) {
-                               sctx->db_depth_disable_expclear = true;
-                       }
-
-                       zstex->depth_clear_value = depth;
-                       sctx->framebuffer.dirty_zsbuf = true;
-                       si_mark_atom_dirty(sctx, &sctx->framebuffer.atom); /* updates DB_DEPTH_CLEAR */
-                       sctx->db_depth_clear = true;
-                       si_mark_atom_dirty(sctx, &sctx->db_render_state);
-               }
-
-               /* TC-compatible HTILE only supports stencil clears to 0. */
-               if (buffers & PIPE_CLEAR_STENCIL &&
-                   (!zstex->tc_compatible_htile || stencil == 0)) {
-                       stencil &= 0xff;
-
-                       /* Need to disable EXPCLEAR temporarily if clearing
-                        * to a new value. */
-                       if (!zstex->stencil_cleared || zstex->stencil_clear_value != stencil) {
-                               sctx->db_stencil_disable_expclear = true;
-                       }
-
-                       zstex->stencil_clear_value = stencil;
-                       sctx->framebuffer.dirty_zsbuf = true;
-                       si_mark_atom_dirty(sctx, &sctx->framebuffer.atom); /* updates DB_STENCIL_CLEAR */
-                       sctx->db_stencil_clear = true;
-                       si_mark_atom_dirty(sctx, &sctx->db_render_state);
-               }
-
-               /* TODO: Find out what's wrong here. Fast depth clear leads to
-                * corruption in ARK: Survival Evolved, but that may just be
-                * a coincidence and the root cause is elsewhere.
-                *
-                * The corruption can be fixed by putting the DB flush before
-                * or after the depth clear. (surprisingly)
-                *
-                * https://bugs.freedesktop.org/show_bug.cgi?id=102955 (apitrace)
-                *
-                * This hack decreases back-to-back ClearDepth performance.
-                */
-               if (sctx->screen->clear_db_cache_before_clear) {
-                       sctx->b.flags |= SI_CONTEXT_FLUSH_AND_INV_DB;
-               }
-       }
-
-       si_blitter_begin(ctx, SI_CLEAR);
-       util_blitter_clear(sctx->blitter, fb->width, fb->height,
-                          util_framebuffer_get_num_layers(fb),
-                          buffers, color, depth, stencil);
-       si_blitter_end(ctx);
-
-       if (sctx->db_depth_clear) {
-               sctx->db_depth_clear = false;
-               sctx->db_depth_disable_expclear = false;
-               zstex->depth_cleared = true;
-               si_mark_atom_dirty(sctx, &sctx->db_render_state);
-       }
-
-       if (sctx->db_stencil_clear) {
-               sctx->db_stencil_clear = false;
-               sctx->db_stencil_disable_expclear = false;
-               zstex->stencil_cleared = true;
-               si_mark_atom_dirty(sctx, &sctx->db_render_state);
-       }
-}
-
-static void si_clear_render_target(struct pipe_context *ctx,
-                                  struct pipe_surface *dst,
-                                  const union pipe_color_union *color,
-                                  unsigned dstx, unsigned dsty,
-                                  unsigned width, unsigned height,
-                                  bool render_condition_enabled)
-{
-       struct si_context *sctx = (struct si_context *)ctx;
-
-       si_blitter_begin(ctx, SI_CLEAR_SURFACE |
-                        (render_condition_enabled ? 0 : SI_DISABLE_RENDER_COND));
-       util_blitter_clear_render_target(sctx->blitter, dst, color,
-                                        dstx, dsty, width, height);
-       si_blitter_end(ctx);
-}
-
-static void si_clear_depth_stencil(struct pipe_context *ctx,
-                                  struct pipe_surface *dst,
-                                  unsigned clear_flags,
-                                  double depth,
-                                  unsigned stencil,
-                                  unsigned dstx, unsigned dsty,
-                                  unsigned width, unsigned height,
-                                  bool render_condition_enabled)
-{
-       struct si_context *sctx = (struct si_context *)ctx;
-
-       si_blitter_begin(ctx, SI_CLEAR_SURFACE |
-                        (render_condition_enabled ? 0 : SI_DISABLE_RENDER_COND));
-       util_blitter_clear_depth_stencil(sctx->blitter, dst, clear_flags, depth, stencil,
-                                        dstx, dsty, width, height);
-       si_blitter_end(ctx);
-}
-
 /* Helper for decompressing a portion of a color or depth resource before
  * blitting if any decompression is needed.
  * The driver doesn't decompress resources automatically while u_blitter is
@@ -1300,7 +1141,7 @@ static bool do_hardware_msaa_resolve(struct pipe_context *ctx,
                            info->dst.resource->last_level != 0)
                                goto resolve_to_temp;
 
-                       vi_dcc_clear_level(&sctx->b, dst, info->dst.level,
+                       vi_dcc_clear_level(sctx, dst, info->dst.level,
                                           0xFFFFFFFF);
                        dst->dirty_level_mask &= ~(1 << info->dst.level);
                }
@@ -1544,10 +1385,7 @@ static void si_pipe_clear_buffer(struct pipe_context *ctx,
 
 void si_init_blit_functions(struct si_context *sctx)
 {
-       sctx->b.b.clear = si_clear;
        sctx->b.b.clear_buffer = si_pipe_clear_buffer;
-       sctx->b.b.clear_render_target = si_clear_render_target;
-       sctx->b.b.clear_depth_stencil = si_clear_depth_stencil;
        sctx->b.b.resource_copy_region = si_resource_copy_region;
        sctx->b.b.blit = si_blit;
        sctx->b.b.flush_resource = si_flush_resource;
diff --git a/src/gallium/drivers/radeonsi/si_clear.c b/src/gallium/drivers/radeonsi/si_clear.c
new file mode 100644 (file)
index 0000000..d99947f
--- /dev/null
@@ -0,0 +1,725 @@
+/*
+ * Copyright 2017 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * on the rights to use, copy, modify, merge, publish, distribute, sub
+ * license, and/or sell copies of the Software, and to permit persons to whom
+ * the Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
+ * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
+ * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
+ * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
+ * USE OR OTHER DEALINGS IN THE SOFTWARE.
+ */
+
+#include "si_pipe.h"
+#include "sid.h"
+
+#include "util/u_format.h"
+#include "util/u_pack_color.h"
+#include "util/u_surface.h"
+
+enum {
+       SI_CLEAR         = SI_SAVE_FRAGMENT_STATE,
+       SI_CLEAR_SURFACE = SI_SAVE_FRAMEBUFFER | SI_SAVE_FRAGMENT_STATE,
+};
+
+static void si_alloc_separate_cmask(struct si_screen *sscreen,
+                                   struct r600_texture *rtex)
+{
+       if (rtex->cmask_buffer)
+                return;
+
+       assert(rtex->cmask.size == 0);
+
+       si_texture_get_cmask_info(&sscreen->b, rtex, &rtex->cmask);
+       if (!rtex->cmask.size)
+               return;
+
+       rtex->cmask_buffer = (struct r600_resource *)
+               si_aligned_buffer_create(&sscreen->b.b,
+                                        R600_RESOURCE_FLAG_UNMAPPABLE,
+                                        PIPE_USAGE_DEFAULT,
+                                        rtex->cmask.size,
+                                        rtex->cmask.alignment);
+       if (rtex->cmask_buffer == NULL) {
+               rtex->cmask.size = 0;
+               return;
+       }
+
+       /* update colorbuffer state bits */
+       rtex->cmask.base_address_reg = rtex->cmask_buffer->gpu_address >> 8;
+
+       rtex->cb_color_info |= S_028C70_FAST_CLEAR(1);
+
+       p_atomic_inc(&sscreen->b.compressed_colortex_counter);
+}
+
+static void si_set_clear_color(struct r600_texture *rtex,
+                              enum pipe_format surface_format,
+                              const union pipe_color_union *color)
+{
+       union util_color uc;
+
+       memset(&uc, 0, sizeof(uc));
+
+       if (rtex->surface.bpe == 16) {
+               /* DCC fast clear only:
+                *   CLEAR_WORD0 = R = G = B
+                *   CLEAR_WORD1 = A
+                */
+               assert(color->ui[0] == color->ui[1] &&
+                      color->ui[0] == color->ui[2]);
+               uc.ui[0] = color->ui[0];
+               uc.ui[1] = color->ui[3];
+       } else if (util_format_is_pure_uint(surface_format)) {
+               util_format_write_4ui(surface_format, color->ui, 0, &uc, 0, 0, 0, 1, 1);
+       } else if (util_format_is_pure_sint(surface_format)) {
+               util_format_write_4i(surface_format, color->i, 0, &uc, 0, 0, 0, 1, 1);
+       } else {
+               util_pack_color(color->f, surface_format, &uc);
+       }
+
+       memcpy(rtex->color_clear_value, &uc, 2 * sizeof(uint32_t));
+}
+
+static bool vi_get_fast_clear_parameters(enum pipe_format surface_format,
+                                        const union pipe_color_union *color,
+                                        uint32_t* reset_value,
+                                        bool* clear_words_needed)
+{
+       bool values[4] = {};
+       int i;
+       bool main_value = false;
+       bool extra_value = false;
+       int extra_channel;
+
+       /* This is needed to get the correct DCC clear value for luminance formats.
+        * 1) Get the linear format (because the next step can't handle L8_SRGB).
+        * 2) Convert luminance to red. (the real hw format for luminance)
+        */
+       surface_format = util_format_linear(surface_format);
+       surface_format = util_format_luminance_to_red(surface_format);
+
+       const struct util_format_description *desc = util_format_description(surface_format);
+
+       if (desc->block.bits == 128 &&
+           (color->ui[0] != color->ui[1] ||
+            color->ui[0] != color->ui[2]))
+               return false;
+
+       *clear_words_needed = true;
+       *reset_value = 0x20202020U;
+
+       /* If we want to clear without needing a fast clear eliminate step, we
+        * can set each channel to 0 or 1 (or 0/max for integer formats). We
+        * have two sets of flags, one for the last or first channel(extra) and
+        * one for the other channels(main).
+        */
+
+       if (surface_format == PIPE_FORMAT_R11G11B10_FLOAT ||
+           surface_format == PIPE_FORMAT_B5G6R5_UNORM ||
+           surface_format == PIPE_FORMAT_B5G6R5_SRGB ||
+           util_format_is_alpha(surface_format)) {
+               extra_channel = -1;
+       } else if (desc->layout == UTIL_FORMAT_LAYOUT_PLAIN) {
+               if (si_translate_colorswap(surface_format, false) <= 1)
+                       extra_channel = desc->nr_channels - 1;
+               else
+                       extra_channel = 0;
+       } else
+               return true;
+
+       for (i = 0; i < 4; ++i) {
+               int index = desc->swizzle[i] - PIPE_SWIZZLE_X;
+
+               if (desc->swizzle[i] < PIPE_SWIZZLE_X ||
+                   desc->swizzle[i] > PIPE_SWIZZLE_W)
+                       continue;
+
+               if (desc->channel[i].pure_integer &&
+                   desc->channel[i].type == UTIL_FORMAT_TYPE_SIGNED) {
+                       /* Use the maximum value for clamping the clear color. */
+                       int max = u_bit_consecutive(0, desc->channel[i].size - 1);
+
+                       values[i] = color->i[i] != 0;
+                       if (color->i[i] != 0 && MIN2(color->i[i], max) != max)
+                               return true;
+               } else if (desc->channel[i].pure_integer &&
+                          desc->channel[i].type == UTIL_FORMAT_TYPE_UNSIGNED) {
+                       /* Use the maximum value for clamping the clear color. */
+                       unsigned max = u_bit_consecutive(0, desc->channel[i].size);
+
+                       values[i] = color->ui[i] != 0U;
+                       if (color->ui[i] != 0U && MIN2(color->ui[i], max) != max)
+                               return true;
+               } else {
+                       values[i] = color->f[i] != 0.0F;
+                       if (color->f[i] != 0.0F && color->f[i] != 1.0F)
+                               return true;
+               }
+
+               if (index == extra_channel)
+                       extra_value = values[i];
+               else
+                       main_value = values[i];
+       }
+
+       for (int i = 0; i < 4; ++i)
+               if (values[i] != main_value &&
+                   desc->swizzle[i] - PIPE_SWIZZLE_X != extra_channel &&
+                   desc->swizzle[i] >= PIPE_SWIZZLE_X &&
+                   desc->swizzle[i] <= PIPE_SWIZZLE_W)
+                       return true;
+
+       *clear_words_needed = false;
+       if (main_value)
+               *reset_value |= 0x80808080U;
+
+       if (extra_value)
+               *reset_value |= 0x40404040U;
+       return true;
+}
+
+void vi_dcc_clear_level(struct si_context *sctx,
+                       struct r600_texture *rtex,
+                       unsigned level, unsigned clear_value)
+{
+       struct pipe_resource *dcc_buffer;
+       uint64_t dcc_offset, clear_size;
+
+       assert(vi_dcc_enabled(rtex, level));
+
+       if (rtex->dcc_separate_buffer) {
+               dcc_buffer = &rtex->dcc_separate_buffer->b.b;
+               dcc_offset = 0;
+       } else {
+               dcc_buffer = &rtex->resource.b.b;
+               dcc_offset = rtex->dcc_offset;
+       }
+
+       if (sctx->b.chip_class >= GFX9) {
+               /* Mipmap level clears aren't implemented. */
+               assert(rtex->resource.b.b.last_level == 0);
+               /* MSAA needs a different clear size. */
+               assert(rtex->resource.b.b.nr_samples <= 1);
+               clear_size = rtex->surface.dcc_size;
+       } else {
+               unsigned num_layers = util_max_layer(&rtex->resource.b.b, level) + 1;
+
+               /* If this is 0, fast clear isn't possible. (can occur with MSAA) */
+               assert(rtex->surface.u.legacy.level[level].dcc_fast_clear_size);
+               /* Layered MSAA DCC fast clears need to clear dcc_fast_clear_size
+                * bytes for each layer. This is not currently implemented, and
+                * therefore MSAA DCC isn't even enabled with multiple layers.
+                */
+               assert(rtex->resource.b.b.nr_samples <= 1 || num_layers == 1);
+
+               dcc_offset += rtex->surface.u.legacy.level[level].dcc_offset;
+               clear_size = rtex->surface.u.legacy.level[level].dcc_fast_clear_size *
+                            num_layers;
+       }
+
+       si_clear_buffer(&sctx->b.b, dcc_buffer, dcc_offset, clear_size,
+                       clear_value, R600_COHERENCY_CB_META);
+}
+
+/* Set the same micro tile mode as the destination of the last MSAA resolve.
+ * This allows hitting the MSAA resolve fast path, which requires that both
+ * src and dst micro tile modes match.
+ */
+static void si_set_optimal_micro_tile_mode(struct si_screen *sscreen,
+                                          struct r600_texture *rtex)
+{
+       if (rtex->resource.b.is_shared ||
+           rtex->resource.b.b.nr_samples <= 1 ||
+           rtex->surface.micro_tile_mode == rtex->last_msaa_resolve_target_micro_mode)
+               return;
+
+       assert(sscreen->b.chip_class >= GFX9 ||
+              rtex->surface.u.legacy.level[0].mode == RADEON_SURF_MODE_2D);
+       assert(rtex->resource.b.b.last_level == 0);
+
+       if (sscreen->b.chip_class >= GFX9) {
+               /* 4K or larger tiles only. 0 is linear. 1-3 are 256B tiles. */
+               assert(rtex->surface.u.gfx9.surf.swizzle_mode >= 4);
+
+               /* If you do swizzle_mode % 4, you'll get:
+                *   0 = Depth
+                *   1 = Standard,
+                *   2 = Displayable
+                *   3 = Rotated
+                *
+                * Depth-sample order isn't allowed:
+                */
+               assert(rtex->surface.u.gfx9.surf.swizzle_mode % 4 != 0);
+
+               switch (rtex->last_msaa_resolve_target_micro_mode) {
+               case RADEON_MICRO_MODE_DISPLAY:
+                       rtex->surface.u.gfx9.surf.swizzle_mode &= ~0x3;
+                       rtex->surface.u.gfx9.surf.swizzle_mode += 2; /* D */
+                       break;
+               case RADEON_MICRO_MODE_THIN:
+                       rtex->surface.u.gfx9.surf.swizzle_mode &= ~0x3;
+                       rtex->surface.u.gfx9.surf.swizzle_mode += 1; /* S */
+                       break;
+               case RADEON_MICRO_MODE_ROTATED:
+                       rtex->surface.u.gfx9.surf.swizzle_mode &= ~0x3;
+                       rtex->surface.u.gfx9.surf.swizzle_mode += 3; /* R */
+                       break;
+               default: /* depth */
+                       assert(!"unexpected micro mode");
+                       return;
+               }
+       } else if (sscreen->b.chip_class >= CIK) {
+               /* These magic numbers were copied from addrlib. It doesn't use
+                * any definitions for them either. They are all 2D_TILED_THIN1
+                * modes with different bpp and micro tile mode.
+                */
+               switch (rtex->last_msaa_resolve_target_micro_mode) {
+               case RADEON_MICRO_MODE_DISPLAY:
+                       rtex->surface.u.legacy.tiling_index[0] = 10;
+                       break;
+               case RADEON_MICRO_MODE_THIN:
+                       rtex->surface.u.legacy.tiling_index[0] = 14;
+                       break;
+               case RADEON_MICRO_MODE_ROTATED:
+                       rtex->surface.u.legacy.tiling_index[0] = 28;
+                       break;
+               default: /* depth, thick */
+                       assert(!"unexpected micro mode");
+                       return;
+               }
+       } else { /* SI */
+               switch (rtex->last_msaa_resolve_target_micro_mode) {
+               case RADEON_MICRO_MODE_DISPLAY:
+                       switch (rtex->surface.bpe) {
+                       case 1:
+                            rtex->surface.u.legacy.tiling_index[0] = 10;
+                            break;
+                       case 2:
+                            rtex->surface.u.legacy.tiling_index[0] = 11;
+                            break;
+                       default: /* 4, 8 */
+                            rtex->surface.u.legacy.tiling_index[0] = 12;
+                            break;
+                       }
+                       break;
+               case RADEON_MICRO_MODE_THIN:
+                       switch (rtex->surface.bpe) {
+                       case 1:
+                                rtex->surface.u.legacy.tiling_index[0] = 14;
+                                break;
+                       case 2:
+                                rtex->surface.u.legacy.tiling_index[0] = 15;
+                                break;
+                       case 4:
+                                rtex->surface.u.legacy.tiling_index[0] = 16;
+                                break;
+                       default: /* 8, 16 */
+                                rtex->surface.u.legacy.tiling_index[0] = 17;
+                                break;
+                       }
+                       break;
+               default: /* depth, thick */
+                       assert(!"unexpected micro mode");
+                       return;
+               }
+       }
+
+       rtex->surface.micro_tile_mode = rtex->last_msaa_resolve_target_micro_mode;
+
+       p_atomic_inc(&sscreen->b.dirty_tex_counter);
+}
+
+static void si_do_fast_color_clear(struct si_context *sctx,
+                                  struct pipe_framebuffer_state *fb,
+                                  struct r600_atom *fb_state,
+                                  unsigned *buffers, ubyte *dirty_cbufs,
+                                  const union pipe_color_union *color)
+{
+       int i;
+
+       /* This function is broken in BE, so just disable this path for now */
+#ifdef PIPE_ARCH_BIG_ENDIAN
+       return;
+#endif
+
+       if (sctx->b.render_cond)
+               return;
+
+       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;
+
+               unsigned level = fb->cbufs[i]->u.tex.level;
+               tex = (struct r600_texture *)fb->cbufs[i]->texture;
+
+               /* 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.is_linear) {
+                       continue;
+               }
+
+               /* shared textures can't use fast clear without an explicit flush,
+                * because there is no way to communicate the clear color among
+                * all clients
+                */
+               if (tex->resource.b.is_shared &&
+                   !(tex->resource.external_usage & PIPE_HANDLE_USAGE_EXPLICIT_FLUSH))
+                       continue;
+
+               /* fast color clear with 1D tiling doesn't work on old kernels and CIK */
+               if (sctx->b.chip_class == CIK &&
+                   tex->surface.u.legacy.level[0].mode == RADEON_SURF_MODE_1D &&
+                   sctx->screen->b.info.drm_major == 2 &&
+                   sctx->screen->b.info.drm_minor < 38) {
+                       continue;
+               }
+
+               /* Fast clear is the most appropriate place to enable DCC for
+                * displayable surfaces.
+                */
+               if (sctx->b.chip_class >= VI &&
+                   !(sctx->screen->b.debug_flags & DBG(NO_DCC_FB))) {
+                       vi_separate_dcc_try_enable(&sctx->b, tex);
+
+                       /* RB+ isn't supported with a CMASK clear only on Stoney,
+                        * so all clears are considered to be hypothetically slow
+                        * clears, which is weighed when determining whether to
+                        * enable separate DCC.
+                        */
+                       if (tex->dcc_gather_statistics &&
+                           sctx->b.family == CHIP_STONEY)
+                               tex->num_slow_clears++;
+               }
+
+               /* Try to clear DCC first, otherwise try CMASK. */
+               if (vi_dcc_enabled(tex, 0)) {
+                       uint32_t reset_value;
+                       bool clear_words_needed, cleared_cmask = false;
+
+                       if (sctx->screen->b.debug_flags & DBG(NO_DCC_CLEAR))
+                               continue;
+
+                       /* This can only occur with MSAA. */
+                       if (sctx->b.chip_class == VI &&
+                           !tex->surface.u.legacy.level[level].dcc_fast_clear_size)
+                               continue;
+
+                       if (!vi_get_fast_clear_parameters(fb->cbufs[i]->format,
+                                                         color, &reset_value,
+                                                         &clear_words_needed))
+                               continue;
+
+                       /* DCC fast clear with MSAA should clear CMASK to 0xC. */
+                       if (tex->resource.b.b.nr_samples >= 2 && tex->cmask.size) {
+                               /* TODO: This doesn't work with MSAA. */
+                               if (clear_words_needed)
+                                       continue;
+
+                               si_clear_buffer(&sctx->b.b, &tex->cmask_buffer->b.b,
+                                               tex->cmask.offset, tex->cmask.size,
+                                               0xCCCCCCCC, R600_COHERENCY_CB_META);
+                               cleared_cmask = true;
+                       }
+
+                       vi_dcc_clear_level(sctx, tex, 0, reset_value);
+
+                       if (clear_words_needed || cleared_cmask) {
+                               bool need_compressed_update = !tex->dirty_level_mask;
+
+                               tex->dirty_level_mask |= 1 << level;
+
+                               if (need_compressed_update)
+                                       p_atomic_inc(&sctx->screen->b.compressed_colortex_counter);
+                       }
+                       tex->separate_dcc_dirty = true;
+               } else {
+                       /* 128-bit formats are unusupported */
+                       if (tex->surface.bpe > 8) {
+                               continue;
+                       }
+
+                       /* RB+ doesn't work with CMASK fast clear on Stoney. */
+                       if (sctx->b.family == CHIP_STONEY)
+                               continue;
+
+                       /* ensure CMASK is enabled */
+                       si_alloc_separate_cmask(sctx->screen, tex);
+                       if (tex->cmask.size == 0) {
+                               continue;
+                       }
+
+                       /* Do the fast clear. */
+                       si_clear_buffer(&sctx->b.b, &tex->cmask_buffer->b.b,
+                                       tex->cmask.offset, tex->cmask.size, 0,
+                                       R600_COHERENCY_CB_META);
+
+                       bool need_compressed_update = !tex->dirty_level_mask;
+
+                       tex->dirty_level_mask |= 1 << level;
+
+                       if (need_compressed_update)
+                               p_atomic_inc(&sctx->screen->b.compressed_colortex_counter);
+               }
+
+               /* We can change the micro tile mode before a full clear. */
+               si_set_optimal_micro_tile_mode(sctx->screen, tex);
+
+               si_set_clear_color(tex, fb->cbufs[i]->format, color);
+
+               if (dirty_cbufs)
+                       *dirty_cbufs |= 1 << i;
+               sctx->b.set_atom_dirty(&sctx->b, fb_state, true);
+               *buffers &= ~clear_bit;
+       }
+}
+
+static void si_clear(struct pipe_context *ctx, unsigned buffers,
+                    const union pipe_color_union *color,
+                    double depth, unsigned stencil)
+{
+       struct si_context *sctx = (struct si_context *)ctx;
+       struct pipe_framebuffer_state *fb = &sctx->framebuffer.state;
+       struct pipe_surface *zsbuf = fb->zsbuf;
+       struct r600_texture *zstex =
+               zsbuf ? (struct r600_texture*)zsbuf->texture : NULL;
+
+       if (buffers & PIPE_CLEAR_COLOR) {
+               si_do_fast_color_clear(sctx, fb,
+                                      &sctx->framebuffer.atom, &buffers,
+                                      &sctx->framebuffer.dirty_cbufs,
+                                      color);
+               if (!buffers)
+                       return; /* all buffers have been fast cleared */
+       }
+
+       if (buffers & PIPE_CLEAR_COLOR) {
+               int i;
+
+               /* These buffers cannot use fast clear, make sure to disable expansion. */
+               for (i = 0; i < fb->nr_cbufs; i++) {
+                       struct r600_texture *tex;
+
+                       /* If not clearing this buffer, skip. */
+                       if (!(buffers & (PIPE_CLEAR_COLOR0 << i)))
+                               continue;
+
+                       if (!fb->cbufs[i])
+                               continue;
+
+                       tex = (struct r600_texture *)fb->cbufs[i]->texture;
+                       if (tex->fmask.size == 0)
+                               tex->dirty_level_mask &= ~(1 << fb->cbufs[i]->u.tex.level);
+               }
+       }
+
+       if (zstex &&
+           r600_htile_enabled(zstex, zsbuf->u.tex.level) &&
+           zsbuf->u.tex.first_layer == 0 &&
+           zsbuf->u.tex.last_layer == util_max_layer(&zstex->resource.b.b, 0)) {
+               /* TC-compatible HTILE only supports depth clears to 0 or 1. */
+               if (buffers & PIPE_CLEAR_DEPTH &&
+                   (!zstex->tc_compatible_htile ||
+                    depth == 0 || depth == 1)) {
+                       /* Need to disable EXPCLEAR temporarily if clearing
+                        * to a new value. */
+                       if (!zstex->depth_cleared || zstex->depth_clear_value != depth) {
+                               sctx->db_depth_disable_expclear = true;
+                       }
+
+                       zstex->depth_clear_value = depth;
+                       sctx->framebuffer.dirty_zsbuf = true;
+                       si_mark_atom_dirty(sctx, &sctx->framebuffer.atom); /* updates DB_DEPTH_CLEAR */
+                       sctx->db_depth_clear = true;
+                       si_mark_atom_dirty(sctx, &sctx->db_render_state);
+               }
+
+               /* TC-compatible HTILE only supports stencil clears to 0. */
+               if (buffers & PIPE_CLEAR_STENCIL &&
+                   (!zstex->tc_compatible_htile || stencil == 0)) {
+                       stencil &= 0xff;
+
+                       /* Need to disable EXPCLEAR temporarily if clearing
+                        * to a new value. */
+                       if (!zstex->stencil_cleared || zstex->stencil_clear_value != stencil) {
+                               sctx->db_stencil_disable_expclear = true;
+                       }
+
+                       zstex->stencil_clear_value = stencil;
+                       sctx->framebuffer.dirty_zsbuf = true;
+                       si_mark_atom_dirty(sctx, &sctx->framebuffer.atom); /* updates DB_STENCIL_CLEAR */
+                       sctx->db_stencil_clear = true;
+                       si_mark_atom_dirty(sctx, &sctx->db_render_state);
+               }
+
+               /* TODO: Find out what's wrong here. Fast depth clear leads to
+                * corruption in ARK: Survival Evolved, but that may just be
+                * a coincidence and the root cause is elsewhere.
+                *
+                * The corruption can be fixed by putting the DB flush before
+                * or after the depth clear. (surprisingly)
+                *
+                * https://bugs.freedesktop.org/show_bug.cgi?id=102955 (apitrace)
+                *
+                * This hack decreases back-to-back ClearDepth performance.
+                */
+               if (sctx->screen->clear_db_cache_before_clear) {
+                       sctx->b.flags |= SI_CONTEXT_FLUSH_AND_INV_DB;
+               }
+       }
+
+       si_blitter_begin(ctx, SI_CLEAR);
+       util_blitter_clear(sctx->blitter, fb->width, fb->height,
+                          util_framebuffer_get_num_layers(fb),
+                          buffers, color, depth, stencil);
+       si_blitter_end(ctx);
+
+       if (sctx->db_depth_clear) {
+               sctx->db_depth_clear = false;
+               sctx->db_depth_disable_expclear = false;
+               zstex->depth_cleared = true;
+               si_mark_atom_dirty(sctx, &sctx->db_render_state);
+       }
+
+       if (sctx->db_stencil_clear) {
+               sctx->db_stencil_clear = false;
+               sctx->db_stencil_disable_expclear = false;
+               zstex->stencil_cleared = true;
+               si_mark_atom_dirty(sctx, &sctx->db_render_state);
+       }
+}
+
+static void si_clear_render_target(struct pipe_context *ctx,
+                                  struct pipe_surface *dst,
+                                  const union pipe_color_union *color,
+                                  unsigned dstx, unsigned dsty,
+                                  unsigned width, unsigned height,
+                                  bool render_condition_enabled)
+{
+       struct si_context *sctx = (struct si_context *)ctx;
+
+       si_blitter_begin(ctx, SI_CLEAR_SURFACE |
+                        (render_condition_enabled ? 0 : SI_DISABLE_RENDER_COND));
+       util_blitter_clear_render_target(sctx->blitter, dst, color,
+                                        dstx, dsty, width, height);
+       si_blitter_end(ctx);
+}
+
+static void si_clear_depth_stencil(struct pipe_context *ctx,
+                                  struct pipe_surface *dst,
+                                  unsigned clear_flags,
+                                  double depth,
+                                  unsigned stencil,
+                                  unsigned dstx, unsigned dsty,
+                                  unsigned width, unsigned height,
+                                  bool render_condition_enabled)
+{
+       struct si_context *sctx = (struct si_context *)ctx;
+
+       si_blitter_begin(ctx, SI_CLEAR_SURFACE |
+                        (render_condition_enabled ? 0 : SI_DISABLE_RENDER_COND));
+       util_blitter_clear_depth_stencil(sctx->blitter, dst, clear_flags, depth, stencil,
+                                        dstx, dsty, width, height);
+       si_blitter_end(ctx);
+}
+
+static void si_clear_texture(struct pipe_context *pipe,
+                            struct pipe_resource *tex,
+                            unsigned level,
+                            const struct pipe_box *box,
+                            const void *data)
+{
+       struct pipe_screen *screen = pipe->screen;
+       struct r600_texture *rtex = (struct r600_texture*)tex;
+       struct pipe_surface tmpl = {{0}};
+       struct pipe_surface *sf;
+       const struct util_format_description *desc =
+               util_format_description(tex->format);
+
+       tmpl.format = tex->format;
+       tmpl.u.tex.first_layer = box->z;
+       tmpl.u.tex.last_layer = box->z + box->depth - 1;
+       tmpl.u.tex.level = level;
+       sf = pipe->create_surface(pipe, tex, &tmpl);
+       if (!sf)
+               return;
+
+       if (rtex->is_depth) {
+               unsigned clear;
+               float depth;
+               uint8_t stencil = 0;
+
+               /* Depth is always present. */
+               clear = PIPE_CLEAR_DEPTH;
+               desc->unpack_z_float(&depth, 0, data, 0, 1, 1);
+
+               if (rtex->surface.has_stencil) {
+                       clear |= PIPE_CLEAR_STENCIL;
+                       desc->unpack_s_8uint(&stencil, 0, data, 0, 1, 1);
+               }
+
+               si_clear_depth_stencil(pipe, sf, clear, depth, stencil,
+                                      box->x, box->y,
+                                      box->width, box->height, false);
+       } else {
+               union pipe_color_union color;
+
+               /* pipe_color_union requires the full vec4 representation. */
+               if (util_format_is_pure_uint(tex->format))
+                       desc->unpack_rgba_uint(color.ui, 0, data, 0, 1, 1);
+               else if (util_format_is_pure_sint(tex->format))
+                       desc->unpack_rgba_sint(color.i, 0, data, 0, 1, 1);
+               else
+                       desc->unpack_rgba_float(color.f, 0, data, 0, 1, 1);
+
+               if (screen->is_format_supported(screen, tex->format,
+                                               tex->target, 0,
+                                               PIPE_BIND_RENDER_TARGET)) {
+                       si_clear_render_target(pipe, sf, &color,
+                                              box->x, box->y,
+                                              box->width, box->height, false);
+               } else {
+                       /* Software fallback - just for R9G9B9E5_FLOAT */
+                       util_clear_render_target(pipe, sf, &color,
+                                                box->x, box->y,
+                                                box->width, box->height);
+               }
+       }
+       pipe_surface_reference(&sf, NULL);
+}
+
+void si_init_clear_functions(struct si_context *sctx)
+{
+       sctx->b.b.clear = si_clear;
+       sctx->b.b.clear_render_target = si_clear_render_target;
+       sctx->b.b.clear_depth_stencil = si_clear_depth_stencil;
+       sctx->b.b.clear_texture = si_clear_texture;
+}
index 4ac0c34bcd2c82464df3ea1f39a495d3723f0726..ca8262649dcdc607064077aff9d1b1ffeea57b39 100644 (file)
@@ -203,9 +203,9 @@ static void si_cp_dma_prepare(struct si_context *sctx, struct pipe_resource *dst
                *packet_flags |= CP_DMA_SYNC;
 }
 
-static void si_clear_buffer(struct pipe_context *ctx, struct pipe_resource *dst,
-                           uint64_t offset, uint64_t size, unsigned value,
-                           enum r600_coherency coher)
+void si_clear_buffer(struct pipe_context *ctx, struct pipe_resource *dst,
+                    uint64_t offset, uint64_t size, unsigned value,
+                    enum r600_coherency coher)
 {
        struct si_context *sctx = (struct si_context*)ctx;
        struct radeon_winsys *ws = sctx->b.ws;
index 6e433a3170826d0edfadd667f6e98da8631ba8dc..f0a9ad197afab470b87eb726fe842e2668cfe71c 100644 (file)
@@ -209,6 +209,7 @@ static struct pipe_context *si_create_context(struct pipe_screen *screen,
        if (sscreen->b.info.drm_major == 3)
                sctx->b.b.get_device_reset_status = si_amdgpu_get_reset_status;
 
+       si_init_clear_functions(sctx);
        si_init_blit_functions(sctx);
        si_init_compute_functions(sctx);
        si_init_cp_dma_functions(sctx);
index 751441df1bc180ee5bccba418fd16d0fb1b38d44..820e5b2d85afd5d375bb698fd7628e939c3fd1e0 100644 (file)
@@ -558,6 +558,16 @@ struct si_context {
 void cik_init_sdma_functions(struct si_context *sctx);
 
 /* si_blit.c */
+enum si_blitter_op /* bitmask */
+{
+       SI_SAVE_TEXTURES      = 1,
+       SI_SAVE_FRAMEBUFFER   = 2,
+       SI_SAVE_FRAGMENT_STATE = 4,
+       SI_DISABLE_RENDER_COND = 8,
+};
+
+void si_blitter_begin(struct pipe_context *ctx, enum si_blitter_op op);
+void si_blitter_end(struct pipe_context *ctx);
 void si_init_blit_functions(struct si_context *sctx);
 void si_decompress_textures(struct si_context *sctx, unsigned shader_mask);
 void si_resource_copy_region(struct pipe_context *ctx,
@@ -568,6 +578,12 @@ void si_resource_copy_region(struct pipe_context *ctx,
                             unsigned src_level,
                             const struct pipe_box *src_box);
 
+/* si_clear.c */
+void vi_dcc_clear_level(struct si_context *sctx,
+                       struct r600_texture *rtex,
+                       unsigned level, unsigned clear_value);
+void si_init_clear_functions(struct si_context *sctx);
+
 /* si_cp_dma.c */
 #define SI_CPDMA_SKIP_CHECK_CS_SPACE   (1 << 0) /* don't call need_cs_space */
 #define SI_CPDMA_SKIP_SYNC_AFTER       (1 << 1) /* don't wait for DMA after the copy */
@@ -580,6 +596,9 @@ void si_resource_copy_region(struct pipe_context *ctx,
                           SI_CPDMA_SKIP_GFX_SYNC | \
                           SI_CPDMA_SKIP_BO_LIST_UPDATE)
 
+void si_clear_buffer(struct pipe_context *ctx, struct pipe_resource *dst,
+                    uint64_t offset, uint64_t size, unsigned value,
+                    enum r600_coherency coher);
 void si_copy_buffer(struct si_context *sctx,
                    struct pipe_resource *dst, struct pipe_resource *src,
                    uint64_t dst_offset, uint64_t src_offset, unsigned size,