radv/gfx9: fix 3d image clears on compute queues
authorDave Airlie <airlied@redhat.com>
Thu, 28 Dec 2017 22:52:07 +0000 (08:52 +1000)
committerDave Airlie <airlied@redhat.com>
Thu, 28 Dec 2017 23:37:05 +0000 (09:37 +1000)
This fixes some of the broken:
dEQP-VK.synchronization.op.multi_queue.*64x64x8* tests.

Fixes: e38685cc62e 'Revert "radv: disable support for VEGA for now."'
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>
src/amd/vulkan/radv_meta_bufimage.c
src/amd/vulkan/radv_private.h

index b2dca80ad504e33b6bccdb24080838bb23780ac9..1696f85ac16f2532b39687b1509c97d35464ed35 100644 (file)
@@ -667,15 +667,16 @@ radv_device_finish_meta_itoi_state(struct radv_device *device)
 }
 
 static nir_shader *
-build_nir_cleari_compute_shader(struct radv_device *dev)
+build_nir_cleari_compute_shader(struct radv_device *dev, bool is_3d)
 {
        nir_builder b;
-       const struct glsl_type *img_type = glsl_sampler_type(GLSL_SAMPLER_DIM_2D,
+       enum glsl_sampler_dim dim = is_3d ? GLSL_SAMPLER_DIM_3D : GLSL_SAMPLER_DIM_2D;
+       const struct glsl_type *img_type = glsl_sampler_type(dim,
                                                             false,
                                                             false,
                                                             GLSL_TYPE_FLOAT);
        nir_builder_init_simple_shader(&b, NULL, MESA_SHADER_COMPUTE, NULL);
-       b.shader->info.name = ralloc_strdup(b.shader, "meta_cleari_cs");
+       b.shader->info.name = ralloc_strdup(b.shader, is_3d ? "meta_cleari_cs_3d" : "meta_cleari_cs");
        b.shader->info.cs.local_size[0] = 16;
        b.shader->info.cs.local_size[1] = 16;
        b.shader->info.cs.local_size[2] = 1;
@@ -696,12 +697,29 @@ build_nir_cleari_compute_shader(struct radv_device *dev)
 
        nir_intrinsic_instr *clear_val = nir_intrinsic_instr_create(b.shader, nir_intrinsic_load_push_constant);
        nir_intrinsic_set_base(clear_val, 0);
-       nir_intrinsic_set_range(clear_val, 16);
+       nir_intrinsic_set_range(clear_val, 20);
        clear_val->src[0] = nir_src_for_ssa(nir_imm_int(&b, 0));
        clear_val->num_components = 4;
        nir_ssa_dest_init(&clear_val->instr, &clear_val->dest, 4, 32, "clear_value");
        nir_builder_instr_insert(&b, &clear_val->instr);
 
+       nir_intrinsic_instr *layer = nir_intrinsic_instr_create(b.shader, nir_intrinsic_load_push_constant);
+       nir_intrinsic_set_base(layer, 0);
+       nir_intrinsic_set_range(layer, 20);
+       layer->src[0] = nir_src_for_ssa(nir_imm_int(&b, 16));
+       layer->num_components = 1;
+       nir_ssa_dest_init(&layer->instr, &layer->dest, 1, 32, "layer");
+       nir_builder_instr_insert(&b, &layer->instr);
+
+       nir_ssa_def *global_z = nir_iadd(&b, nir_channel(&b, global_id, 2), &layer->dest.ssa);
+
+       nir_ssa_def *comps[4];
+       comps[0] = nir_channel(&b, global_id, 0);
+       comps[1] = nir_channel(&b, global_id, 1);
+       comps[2] = global_z;
+       comps[3] = nir_imm_int(&b, 0);
+       global_id = nir_vec(&b, comps, 4);
+
        nir_intrinsic_instr *store = nir_intrinsic_instr_create(b.shader, nir_intrinsic_image_store);
        store->src[0] = nir_src_for_ssa(global_id);
        store->src[1] = nir_src_for_ssa(nir_ssa_undef(&b, 1, 32));
@@ -717,8 +735,10 @@ radv_device_init_meta_cleari_state(struct radv_device *device)
 {
        VkResult result;
        struct radv_shader_module cs = { .nir = NULL };
-
-       cs.nir = build_nir_cleari_compute_shader(device);
+       struct radv_shader_module cs_3d = { .nir = NULL };
+       cs.nir = build_nir_cleari_compute_shader(device, false);
+       if (device->physical_device->rad_info.chip_class >= GFX9)
+               cs_3d.nir = build_nir_cleari_compute_shader(device, true);
 
        /*
         * two descriptors one for the image being sampled
@@ -752,7 +772,7 @@ radv_device_init_meta_cleari_state(struct radv_device *device)
                .setLayoutCount = 1,
                .pSetLayouts = &device->meta_state.cleari.img_ds_layout,
                .pushConstantRangeCount = 1,
-               .pPushConstantRanges = &(VkPushConstantRange){VK_SHADER_STAGE_COMPUTE_BIT, 0, 16},
+               .pPushConstantRanges = &(VkPushConstantRange){VK_SHADER_STAGE_COMPUTE_BIT, 0, 20},
        };
 
        result = radv_CreatePipelineLayout(radv_device_to_handle(device),
@@ -786,10 +806,38 @@ radv_device_init_meta_cleari_state(struct radv_device *device)
        if (result != VK_SUCCESS)
                goto fail;
 
+
+       if (device->physical_device->rad_info.chip_class >= GFX9) {
+               /* compute shader */
+               VkPipelineShaderStageCreateInfo pipeline_shader_stage_3d = {
+                       .sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO,
+                       .stage = VK_SHADER_STAGE_COMPUTE_BIT,
+                       .module = radv_shader_module_to_handle(&cs_3d),
+                       .pName = "main",
+                       .pSpecializationInfo = NULL,
+               };
+
+               VkComputePipelineCreateInfo vk_pipeline_info_3d = {
+                       .sType = VK_STRUCTURE_TYPE_COMPUTE_PIPELINE_CREATE_INFO,
+                       .stage = pipeline_shader_stage_3d,
+                       .flags = 0,
+                       .layout = device->meta_state.cleari.img_p_layout,
+               };
+
+               result = radv_CreateComputePipelines(radv_device_to_handle(device),
+                                                    radv_pipeline_cache_to_handle(&device->meta_state.cache),
+                                                    1, &vk_pipeline_info_3d, NULL,
+                                                    &device->meta_state.cleari.pipeline_3d);
+               if (result != VK_SUCCESS)
+                       goto fail;
+
+               ralloc_free(cs_3d.nir);
+       }
        ralloc_free(cs.nir);
        return VK_SUCCESS;
 fail:
        ralloc_free(cs.nir);
+       ralloc_free(cs_3d.nir);
        return result;
 }
 
@@ -805,6 +853,8 @@ radv_device_finish_meta_cleari_state(struct radv_device *device)
                                        &state->alloc);
        radv_DestroyPipeline(radv_device_to_handle(device),
                             state->cleari.pipeline, &state->alloc);
+       radv_DestroyPipeline(radv_device_to_handle(device),
+                            state->cleari.pipeline_3d, &state->alloc);
 }
 
 void
@@ -1163,19 +1213,24 @@ radv_meta_clear_image_cs(struct radv_cmd_buffer *cmd_buffer,
        create_iview(cmd_buffer, dst, &dst_iview);
        cleari_bind_descriptors(cmd_buffer, &dst_iview);
 
+       if (device->physical_device->rad_info.chip_class >= GFX9 &&
+           dst->image->type == VK_IMAGE_TYPE_3D)
+               pipeline = cmd_buffer->device->meta_state.cleari.pipeline_3d;
+
        radv_CmdBindPipeline(radv_cmd_buffer_to_handle(cmd_buffer),
                             VK_PIPELINE_BIND_POINT_COMPUTE, pipeline);
 
-       unsigned push_constants[4] = {
+       unsigned push_constants[5] = {
                clear_color->uint32[0],
                clear_color->uint32[1],
                clear_color->uint32[2],
                clear_color->uint32[3],
+               dst->layer,
        };
 
        radv_CmdPushConstants(radv_cmd_buffer_to_handle(cmd_buffer),
                              device->meta_state.cleari.img_p_layout,
-                             VK_SHADER_STAGE_COMPUTE_BIT, 0, 16,
+                             VK_SHADER_STAGE_COMPUTE_BIT, 0, 20,
                              push_constants);
 
        radv_unaligned_dispatch(cmd_buffer, dst->image->info.width, dst->image->info.height, 1);
index a8270483014684340dbaee3315ca7f77b81f9b18..05ad480c5781979e85e644ad1d9c432600d96c26 100644 (file)
@@ -449,6 +449,7 @@ struct radv_meta_state {
                VkPipelineLayout                          img_p_layout;
                VkDescriptorSetLayout                     img_ds_layout;
                VkPipeline pipeline;
+               VkPipeline pipeline_3d;
        } cleari;
 
        struct {