radv: rename decompress/resummarize depth/stencil functions
[mesa.git] / src / amd / vulkan / radv_meta_decompress.c
index 262dfd67b3c14b3c0d9aed61281131e489c8186f..8fc264defe55a96c24202e7d09cc316796440339 100644 (file)
 #include "radv_private.h"
 #include "sid.h"
 
+enum radv_depth_op {
+       DEPTH_DECOMPRESS,
+       DEPTH_RESUMMARIZE,
+};
+
+enum radv_depth_decompress {
+       DECOMPRESS_DEPTH_STENCIL,
+       DECOMPRESS_DEPTH,
+       DECOMPRESS_STENCIL,
+};
+
 static VkResult
 create_pass(struct radv_device *device,
            uint32_t samples,
@@ -67,7 +78,27 @@ create_pass(struct radv_device *device,
                                                       .preserveAttachmentCount = 0,
                                                       .pPreserveAttachments = NULL,
                                               },
-                                                               .dependencyCount = 0,
+                                                       .dependencyCount = 2,
+                                                       .pDependencies = (VkSubpassDependency[]) {
+                                                               {
+                                                                       .srcSubpass = VK_SUBPASS_EXTERNAL,
+                                                                       .dstSubpass = 0,
+                                                                       .srcStageMask = VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT,
+                                                                       .dstStageMask = VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT,
+                                                                       .srcAccessMask = 0,
+                                                                       .dstAccessMask = 0,
+                                                                       .dependencyFlags = 0
+                                                               },
+                                                               {
+                                                                       .srcSubpass = 0,
+                                                                       .dstSubpass = VK_SUBPASS_EXTERNAL,
+                                                                       .srcStageMask = VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT,
+                                                                       .dstStageMask = VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT,
+                                                                       .srcAccessMask = 0,
+                                                                       .dstAccessMask = 0,
+                                                                       .dependencyFlags = 0
+                                                               }
+                                                       },
                                                                   },
                                       alloc,
                                       pass);
@@ -98,15 +129,16 @@ create_pipeline(struct radv_device *device,
                uint32_t samples,
                VkRenderPass pass,
                VkPipelineLayout layout,
-               VkPipeline *decompress_pipeline,
-               VkPipeline *resummarize_pipeline)
+               enum radv_depth_op op,
+               enum radv_depth_decompress decompress,
+               VkPipeline *pipeline)
 {
        VkResult result;
        VkDevice device_h = radv_device_to_handle(device);
        struct radv_shader_module vs_module = {0};
 
        mtx_lock(&device->meta_state.mtx);
-       if (*decompress_pipeline) {
+       if (*pipeline) {
                mtx_unlock(&device->meta_state.mtx);
                return VK_SUCCESS;
        }
@@ -207,34 +239,20 @@ create_pipeline(struct radv_device *device,
                .subpass = 0,
        };
 
-       result = radv_graphics_pipeline_create(device_h,
-                                              radv_pipeline_cache_to_handle(&device->meta_state.cache),
-                                              &pipeline_create_info,
-                                              &(struct radv_graphics_pipeline_create_info) {
-                                                       .use_rectlist = true,
-                                                       .db_flush_depth_inplace = true,
-                                                       .db_flush_stencil_inplace = true,
-                                              },
-                                              &device->meta_state.alloc,
-                                              decompress_pipeline);
-       if (result != VK_SUCCESS)
-               goto cleanup;
+       struct radv_graphics_pipeline_create_info extra = {
+               .use_rectlist = true,
+               .db_flush_depth_inplace = decompress == DECOMPRESS_DEPTH_STENCIL ||
+                                         decompress == DECOMPRESS_DEPTH,
+               .db_flush_stencil_inplace = decompress == DECOMPRESS_DEPTH_STENCIL ||
+                                           decompress == DECOMPRESS_STENCIL,
+               .db_resummarize = op == DEPTH_RESUMMARIZE,
+       };
 
        result = radv_graphics_pipeline_create(device_h,
                                               radv_pipeline_cache_to_handle(&device->meta_state.cache),
-                                              &pipeline_create_info,
-                                              &(struct radv_graphics_pipeline_create_info) {
-                                                       .use_rectlist = true,
-                                                       .db_flush_depth_inplace = true,
-                                                       .db_flush_stencil_inplace = true,
-                                                       .db_resummarize = true,
-                                              },
+                                              &pipeline_create_info, &extra,
                                               &device->meta_state.alloc,
-                                              resummarize_pipeline);
-       if (result != VK_SUCCESS)
-               goto cleanup;
-
-       goto cleanup;
+                                              pipeline);
 
 cleanup:
        ralloc_free(fs_module.nir);
@@ -256,9 +274,12 @@ radv_device_finish_meta_depth_decomp_state(struct radv_device *device)
                radv_DestroyPipelineLayout(radv_device_to_handle(device),
                                           state->depth_decomp[i].p_layout,
                                           &state->alloc);
-               radv_DestroyPipeline(radv_device_to_handle(device),
-                                    state->depth_decomp[i].decompress_pipeline,
-                                    &state->alloc);
+
+               for (uint32_t j = 0; j < NUM_DEPTH_DECOMPRESS_PIPELINES; j++) {
+                       radv_DestroyPipeline(radv_device_to_handle(device),
+                                            state->depth_decomp[i].decompress_pipeline[j],
+                                            &state->alloc);
+               }
                radv_DestroyPipeline(radv_device_to_handle(device),
                                     state->depth_decomp[i].resummarize_pipeline,
                                     &state->alloc);
@@ -295,10 +316,22 @@ radv_device_init_meta_depth_decomp_state(struct radv_device *device, bool on_dem
                if (on_demand)
                        continue;
 
+               for (uint32_t j = 0; j < NUM_DEPTH_DECOMPRESS_PIPELINES; j++) {
+                       res = create_pipeline(device, vs_module_h, samples,
+                                             state->depth_decomp[i].pass,
+                                             state->depth_decomp[i].p_layout,
+                                             DEPTH_DECOMPRESS,
+                                             j,
+                                             &state->depth_decomp[i].decompress_pipeline[j]);
+                       if (res != VK_SUCCESS)
+                               goto fail;
+               }
+
                res = create_pipeline(device, vs_module_h, samples,
                                      state->depth_decomp[i].pass,
                                      state->depth_decomp[i].p_layout,
-                                     &state->depth_decomp[i].decompress_pipeline,
+                                     DEPTH_RESUMMARIZE,
+                                     0, /* unused */
                                      &state->depth_decomp[i].resummarize_pipeline);
                if (res != VK_SUCCESS)
                        goto fail;
@@ -315,27 +348,47 @@ cleanup:
        return res;
 }
 
-enum radv_depth_op {
-       DEPTH_DECOMPRESS,
-       DEPTH_RESUMMARIZE,
-};
-
 static VkPipeline *
 radv_get_depth_pipeline(struct radv_cmd_buffer *cmd_buffer,
-                       struct radv_image *image, enum radv_depth_op op)
+                       struct radv_image *image,
+                       const VkImageSubresourceRange *subresourceRange,
+                       enum radv_depth_op op)
 {
        struct radv_meta_state *state = &cmd_buffer->device->meta_state;
        uint32_t samples = image->info.samples;
        uint32_t samples_log2 = ffs(samples) - 1;
+       enum radv_depth_decompress decompress;
        VkPipeline *pipeline;
 
-       if (!state->depth_decomp[samples_log2].decompress_pipeline) {
+       if (subresourceRange->aspectMask == VK_IMAGE_ASPECT_DEPTH_BIT) {
+               decompress = DECOMPRESS_DEPTH;
+       } else if (subresourceRange->aspectMask == VK_IMAGE_ASPECT_STENCIL_BIT) {
+               decompress = DECOMPRESS_STENCIL;
+       } else {
+               decompress = DECOMPRESS_DEPTH_STENCIL;
+       }
+
+       if (!state->depth_decomp[samples_log2].decompress_pipeline[decompress]) {
                VkResult ret;
 
+               for (uint32_t i = 0; i < NUM_DEPTH_DECOMPRESS_PIPELINES; i++) {
+                       ret = create_pipeline(cmd_buffer->device, VK_NULL_HANDLE, samples,
+                                             state->depth_decomp[samples_log2].pass,
+                                             state->depth_decomp[samples_log2].p_layout,
+                                             DEPTH_DECOMPRESS,
+                                             i,
+                                             &state->depth_decomp[samples_log2].decompress_pipeline[i]);
+                       if (ret != VK_SUCCESS) {
+                               cmd_buffer->record_result = ret;
+                               return NULL;
+                       }
+               }
+
                ret = create_pipeline(cmd_buffer->device, VK_NULL_HANDLE, samples,
                                      state->depth_decomp[samples_log2].pass,
                                      state->depth_decomp[samples_log2].p_layout,
-                                     &state->depth_decomp[samples_log2].decompress_pipeline,
+                                     DEPTH_RESUMMARIZE,
+                                     0, /* unused */
                                      &state->depth_decomp[samples_log2].resummarize_pipeline);
                if (ret != VK_SUCCESS) {
                        cmd_buffer->record_result = ret;
@@ -345,7 +398,7 @@ radv_get_depth_pipeline(struct radv_cmd_buffer *cmd_buffer,
 
        switch (op) {
        case DEPTH_DECOMPRESS:
-               pipeline = &state->depth_decomp[samples_log2].decompress_pipeline;
+               pipeline = &state->depth_decomp[samples_log2].decompress_pipeline[decompress];
                break;
        case DEPTH_RESUMMARIZE:
                pipeline = &state->depth_decomp[samples_log2].resummarize_pipeline;
@@ -361,7 +414,7 @@ static void
 radv_process_depth_image_layer(struct radv_cmd_buffer *cmd_buffer,
                               struct radv_image *image,
                               const VkImageSubresourceRange *range,
-                              int layer)
+                              int level, int layer)
 {
        struct radv_device *device = cmd_buffer->device;
        struct radv_meta_state *state = &device->meta_state;
@@ -369,8 +422,8 @@ radv_process_depth_image_layer(struct radv_cmd_buffer *cmd_buffer,
        struct radv_image_view iview;
        uint32_t width, height;
 
-       width = radv_minify(image->info.width, range->baseMipLevel);
-       height = radv_minify(image->info.height, range->baseMipLevel);
+       width = radv_minify(image->info.width, range->baseMipLevel + level);
+       height = radv_minify(image->info.height, range->baseMipLevel + level);
 
        radv_image_view_init(&iview, device,
                             &(VkImageViewCreateInfo) {
@@ -380,7 +433,7 @@ radv_process_depth_image_layer(struct radv_cmd_buffer *cmd_buffer,
                                        .format = image->vk_format,
                                        .subresourceRange = {
                                                .aspectMask = VK_IMAGE_ASPECT_DEPTH_BIT,
-                                               .baseMipLevel = range->baseMipLevel,
+                                               .baseMipLevel = range->baseMipLevel + level,
                                                .levelCount = 1,
                                                .baseArrayLayer = range->baseArrayLayer + layer,
                                                .layerCount = 1,
@@ -401,45 +454,42 @@ radv_process_depth_image_layer(struct radv_cmd_buffer *cmd_buffer,
                                        .layers = 1
                               }, &cmd_buffer->pool->alloc, &fb_h);
 
-       radv_CmdBeginRenderPass(radv_cmd_buffer_to_handle(cmd_buffer),
-                               &(VkRenderPassBeginInfo) {
-                                       .sType = VK_STRUCTURE_TYPE_RENDER_PASS_BEGIN_INFO,
-                                       .renderPass = state->depth_decomp[samples_log2].pass,
-                                       .framebuffer = fb_h,
-                                       .renderArea = {
-                                               .offset = {
-                                                       0,
-                                                       0,
+       radv_cmd_buffer_begin_render_pass(cmd_buffer,
+                                         &(VkRenderPassBeginInfo) {
+                                               .sType = VK_STRUCTURE_TYPE_RENDER_PASS_BEGIN_INFO,
+                                               .renderPass = state->depth_decomp[samples_log2].pass,
+                                               .framebuffer = fb_h,
+                                               .renderArea = {
+                                                       .offset = {
+                                                               0,
+                                                               0,
+                                                       },
+                                                       .extent = {
+                                                               width,
+                                                               height,
+                                                       }
                                                },
-                                               .extent = {
-                                                       width,
-                                                       height,
-                                               }
-                                       },
-                                       .clearValueCount = 0,
-                                       .pClearValues = NULL,
-                               },
-                               VK_SUBPASS_CONTENTS_INLINE);
+                                               .clearValueCount = 0,
+                                               .pClearValues = NULL,
+                                       });
+       radv_cmd_buffer_set_subpass(cmd_buffer,
+                                   &cmd_buffer->state.pass->subpasses[0]);
 
        radv_CmdDraw(radv_cmd_buffer_to_handle(cmd_buffer), 3, 1, 0, 0);
-       radv_CmdEndRenderPass(radv_cmd_buffer_to_handle(cmd_buffer));
+       radv_cmd_buffer_end_render_pass(cmd_buffer);
 
        radv_DestroyFramebuffer(radv_device_to_handle(device), fb_h,
                                &cmd_buffer->pool->alloc);
 }
 
-static void radv_process_depth_image_inplace(struct radv_cmd_buffer *cmd_buffer,
-                                            struct radv_image *image,
-                                            VkImageSubresourceRange *subresourceRange,
-                                            struct radv_sample_locations_state *sample_locs,
-                                            enum radv_depth_op op)
+static void radv_process_depth_stencil(struct radv_cmd_buffer *cmd_buffer,
+                                      struct radv_image *image,
+                                      const VkImageSubresourceRange *subresourceRange,
+                                      struct radv_sample_locations_state *sample_locs,
+                                      enum radv_depth_op op)
 {
        struct radv_meta_saved_state saved_state;
        VkCommandBuffer cmd_buffer_h = radv_cmd_buffer_to_handle(cmd_buffer);
-       uint32_t width = radv_minify(image->info.width,
-                                    subresourceRange->baseMipLevel);
-       uint32_t height = radv_minify(image->info.height,
-                                    subresourceRange->baseMipLevel);
        VkPipeline *pipeline;
 
        if (!radv_image_has_htile(image))
@@ -450,25 +500,12 @@ static void radv_process_depth_image_inplace(struct radv_cmd_buffer *cmd_buffer,
                       RADV_META_SAVE_SAMPLE_LOCATIONS |
                       RADV_META_SAVE_PASS);
 
-       pipeline = radv_get_depth_pipeline(cmd_buffer, image, op);
+       pipeline = radv_get_depth_pipeline(cmd_buffer, image,
+                                          subresourceRange, op);
 
        radv_CmdBindPipeline(radv_cmd_buffer_to_handle(cmd_buffer),
                             VK_PIPELINE_BIND_POINT_GRAPHICS, *pipeline);
 
-       radv_CmdSetViewport(cmd_buffer_h, 0, 1, &(VkViewport) {
-               .x = 0,
-               .y = 0,
-               .width = width,
-               .height = height,
-               .minDepth = 0.0f,
-               .maxDepth = 1.0f
-       });
-
-       radv_CmdSetScissor(cmd_buffer_h, 0, 1, &(VkRect2D) {
-               .offset = { 0, 0 },
-               .extent = { width, height },
-       });
-
        if (sample_locs) {
                assert(image->flags & VK_IMAGE_CREATE_SAMPLE_LOCATIONS_COMPATIBLE_DEPTH_BIT_EXT);
 
@@ -484,30 +521,65 @@ static void radv_process_depth_image_inplace(struct radv_cmd_buffer *cmd_buffer,
                });
        }
 
-       for (uint32_t s = 0; s < radv_get_layerCount(image, subresourceRange); s++) {
-               radv_process_depth_image_layer(cmd_buffer, image,
-                                              subresourceRange, s);
+       for (uint32_t l = 0; l < radv_get_levelCount(image, subresourceRange); ++l) {
+               uint32_t width =
+                       radv_minify(image->info.width,
+                                   subresourceRange->baseMipLevel + l);
+               uint32_t height =
+                       radv_minify(image->info.height,
+                                   subresourceRange->baseMipLevel + l);
+
+               radv_CmdSetViewport(cmd_buffer_h, 0, 1,
+                                   &(VkViewport) {
+                                       .x = 0,
+                                       .y = 0,
+                                       .width = width,
+                                       .height = height,
+                                       .minDepth = 0.0f,
+                                       .maxDepth = 1.0f
+                                   });
+
+               radv_CmdSetScissor(cmd_buffer_h, 0, 1,
+                                  &(VkRect2D) {
+                                       .offset = { 0, 0 },
+                                       .extent = { width, height },
+                                  });
+
+               for (uint32_t s = 0; s < radv_get_layerCount(image, subresourceRange); s++) {
+                       radv_process_depth_image_layer(cmd_buffer, image,
+                                                      subresourceRange, l, s);
+               }
        }
 
        radv_meta_restore(&saved_state, cmd_buffer);
 }
 
-void radv_decompress_depth_image_inplace(struct radv_cmd_buffer *cmd_buffer,
-                                        struct radv_image *image,
-                                        VkImageSubresourceRange *subresourceRange,
-                                        struct radv_sample_locations_state *sample_locs)
+void radv_decompress_depth_stencil(struct radv_cmd_buffer *cmd_buffer,
+                                  struct radv_image *image,
+                                  const VkImageSubresourceRange *subresourceRange,
+                                  struct radv_sample_locations_state *sample_locs)
 {
+       struct radv_barrier_data barrier = {};
+
+       barrier.layout_transitions.depth_stencil_expand = 1;
+       radv_describe_layout_transition(cmd_buffer, &barrier);
+
        assert(cmd_buffer->queue_family_index == RADV_QUEUE_GENERAL);
-       radv_process_depth_image_inplace(cmd_buffer, image, subresourceRange,
-                                        sample_locs, DEPTH_DECOMPRESS);
+       radv_process_depth_stencil(cmd_buffer, image, subresourceRange,
+                                  sample_locs, DEPTH_DECOMPRESS);
 }
 
-void radv_resummarize_depth_image_inplace(struct radv_cmd_buffer *cmd_buffer,
-                                        struct radv_image *image,
-                                        VkImageSubresourceRange *subresourceRange,
-                                        struct radv_sample_locations_state *sample_locs)
+void radv_resummarize_depth_stencil(struct radv_cmd_buffer *cmd_buffer,
+                                   struct radv_image *image,
+                                   const VkImageSubresourceRange *subresourceRange,
+                                   struct radv_sample_locations_state *sample_locs)
 {
+       struct radv_barrier_data barrier = {};
+
+       barrier.layout_transitions.depth_stencil_resummarize = 1;
+       radv_describe_layout_transition(cmd_buffer, &barrier);
+
        assert(cmd_buffer->queue_family_index == RADV_QUEUE_GENERAL);
-       radv_process_depth_image_inplace(cmd_buffer, image, subresourceRange,
-                                        sample_locs, DEPTH_RESUMMARIZE);
+       radv_process_depth_stencil(cmd_buffer, image, subresourceRange,
+                                  sample_locs, DEPTH_RESUMMARIZE);
 }