radv/gfx9: fix buffer to image for 3d images on compute queues

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>

1 files changed, 47 insertions, 15 deletions

diff --git a/src/amd/vulkan/radv_meta_bufimage.c b/src/amd/vulkan/radv_meta_bufimage.c
index 1696f85ac16..5bcc1e62dba 100644
--- a/src/amd/vulkan/radv_meta_bufimage.c
+++ b/src/amd/vulkan/radv_meta_bufimage.c
@@ -259,19 +259,20 @@ radv_device_finish_meta_itob_state(struct radv_device *device)
 }
 
 static nir_shader *
-build_nir_btoi_compute_shader(struct radv_device *dev)
+build_nir_btoi_compute_shader(struct radv_device *dev, bool is_3d)
 {
 	nir_builder b;
+	enum glsl_sampler_dim dim = is_3d ? GLSL_SAMPLER_DIM_3D : GLSL_SAMPLER_DIM_2D;
 	const struct glsl_type *buf_type = glsl_sampler_type(GLSL_SAMPLER_DIM_BUF,
 							     false,
 							     false,
 							     GLSL_TYPE_FLOAT);
-	const struct glsl_type *img_type = glsl_sampler_type(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_btoi_cs");
+	b.shader->info.name = ralloc_strdup(b.shader, is_3d ? "meta_btoi_cs_3d" : "meta_btoi_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;
@@ -296,16 +297,16 @@ build_nir_btoi_compute_shader(struct radv_device *dev)
 
 	nir_intrinsic_instr *offset = nir_intrinsic_instr_create(b.shader, nir_intrinsic_load_push_constant);
 	nir_intrinsic_set_base(offset, 0);
-	nir_intrinsic_set_range(offset, 12);
+	nir_intrinsic_set_range(offset, 16);
 	offset->src[0] = nir_src_for_ssa(nir_imm_int(&b, 0));
-	offset->num_components = 2;
-	nir_ssa_dest_init(&offset->instr, &offset->dest, 2, 32, "offset");
+	offset->num_components = is_3d ? 3 : 2;
+	nir_ssa_dest_init(&offset->instr, &offset->dest, is_3d ? 3 : 2, 32, "offset");
 	nir_builder_instr_insert(&b, &offset->instr);
 
 	nir_intrinsic_instr *stride = nir_intrinsic_instr_create(b.shader, nir_intrinsic_load_push_constant);
 	nir_intrinsic_set_base(stride, 0);
-	nir_intrinsic_set_range(stride, 12);
-	stride->src[0] = nir_src_for_ssa(nir_imm_int(&b, 8));
+	nir_intrinsic_set_range(stride, 16);
+	stride->src[0] = nir_src_for_ssa(nir_imm_int(&b, 12));
 	stride->num_components = 1;
 	nir_ssa_dest_init(&stride->instr, &stride->dest, 1, 32, "stride");
 	nir_builder_instr_insert(&b, &stride->instr);
@@ -353,9 +354,10 @@ radv_device_init_meta_btoi_state(struct radv_device *device)
 {
 	VkResult result;
 	struct radv_shader_module cs = { .nir = NULL };
-
-	cs.nir = build_nir_btoi_compute_shader(device);
-
+	struct radv_shader_module cs_3d = { .nir = NULL };
+	cs.nir = build_nir_btoi_compute_shader(device, false);
+	if (device->physical_device->rad_info.chip_class >= GFX9)
+		cs_3d.nir = build_nir_btoi_compute_shader(device, true);
 	/*
 	 * two descriptors one for the image being sampled
 	 * one for the buffer being written.
@@ -395,7 +397,7 @@ radv_device_init_meta_btoi_state(struct radv_device *device)
 		.setLayoutCount = 1,
 		.pSetLayouts = &device->meta_state.btoi.img_ds_layout,
 		.pushConstantRangeCount = 1,
-		.pPushConstantRanges = &(VkPushConstantRange){VK_SHADER_STAGE_COMPUTE_BIT, 0, 12},
+		.pPushConstantRanges = &(VkPushConstantRange){VK_SHADER_STAGE_COMPUTE_BIT, 0, 16},
 	};
 
 	result = radv_CreatePipelineLayout(radv_device_to_handle(device),
@@ -429,9 +431,33 @@ radv_device_init_meta_btoi_state(struct radv_device *device)
 	if (result != VK_SUCCESS)
 		goto fail;
 
+	if (device->physical_device->rad_info.chip_class >= GFX9) {
+		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.btoi.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.btoi.pipeline_3d);
+		ralloc_free(cs_3d.nir);
+	}
 	ralloc_free(cs.nir);
+
 	return VK_SUCCESS;
 fail:
+	ralloc_free(cs_3d.nir);
 	ralloc_free(cs.nir);
 	return result;
 }
@@ -448,6 +474,8 @@ radv_device_finish_meta_btoi_state(struct radv_device *device)
 					&state->alloc);
 	radv_DestroyPipeline(radv_device_to_handle(device),
 			     state->btoi.pipeline, &state->alloc);
+	radv_DestroyPipeline(radv_device_to_handle(device),
+			     state->btoi.pipeline_3d, &state->alloc);
 }
 
 static nir_shader *
@@ -1070,18 +1098,22 @@ radv_meta_buffer_to_image_cs(struct radv_cmd_buffer *cmd_buffer,
 	create_iview(cmd_buffer, dst, &dst_view);
 	btoi_bind_descriptors(cmd_buffer, &src_view, &dst_view);
 
+	if (device->physical_device->rad_info.chip_class >= GFX9 &&
+	    dst->image->type == VK_IMAGE_TYPE_3D)
+		pipeline = cmd_buffer->device->meta_state.btoi.pipeline_3d;
 	radv_CmdBindPipeline(radv_cmd_buffer_to_handle(cmd_buffer),
 			     VK_PIPELINE_BIND_POINT_COMPUTE, pipeline);
 
 	for (unsigned r = 0; r < num_rects; ++r) {
-		unsigned push_constants[3] = {
+		unsigned push_constants[4] = {
 			rects[r].dst_x,
 			rects[r].dst_y,
-			src->pitch
+			dst->layer,
+			src->pitch,
 		};
 		radv_CmdPushConstants(radv_cmd_buffer_to_handle(cmd_buffer),
 				      device->meta_state.btoi.img_p_layout,
-				      VK_SHADER_STAGE_COMPUTE_BIT, 0, 12,
+				      VK_SHADER_STAGE_COMPUTE_BIT, 0, 16,
 				      push_constants);
 
 		radv_unaligned_dispatch(cmd_buffer, rects[r].width, rects[r].height, 1);