+static nir_shader *
+build_tfb_query_shader(struct radv_device *device)
+{
+ /* the shader this builds is roughly
+ *
+ * uint32_t src_stride = 32;
+ *
+ * location(binding = 0) buffer dst_buf;
+ * location(binding = 1) buffer src_buf;
+ *
+ * void main() {
+ * uint64_t result[2] = {};
+ * bool available = false;
+ * uint64_t src_offset = src_stride * global_id.x;
+ * uint64_t dst_offset = dst_stride * global_id.x;
+ * uint64_t *src_data = src_buf[src_offset];
+ * uint32_t avail = (src_data[0] >> 32) &
+ * (src_data[1] >> 32) &
+ * (src_data[2] >> 32) &
+ * (src_data[3] >> 32);
+ * if (avail & 0x80000000) {
+ * result[0] = src_data[3] - src_data[1];
+ * result[1] = src_data[2] - src_data[0];
+ * available = true;
+ * }
+ * uint32_t result_size = flags & VK_QUERY_RESULT_64_BIT ? 16 : 8;
+ * if ((flags & VK_QUERY_RESULT_PARTIAL_BIT) || available) {
+ * if (flags & VK_QUERY_RESULT_64_BIT) {
+ * dst_buf[dst_offset] = result;
+ * } else {
+ * dst_buf[dst_offset] = (uint32_t)result;
+ * }
+ * }
+ * if (flags & VK_QUERY_RESULT_WITH_AVAILABILITY_BIT) {
+ * dst_buf[dst_offset + result_size] = available;
+ * }
+ * }
+ */
+ nir_builder b;
+ nir_builder_init_simple_shader(&b, NULL, MESA_SHADER_COMPUTE, NULL);
+ b.shader->info.name = ralloc_strdup(b.shader, "tfb_query");
+ b.shader->info.cs.local_size[0] = 64;
+ b.shader->info.cs.local_size[1] = 1;
+ b.shader->info.cs.local_size[2] = 1;
+
+ /* Create and initialize local variables. */
+ nir_variable *result =
+ nir_local_variable_create(b.impl,
+ glsl_vector_type(GLSL_TYPE_UINT64, 2),
+ "result");
+ nir_variable *available =
+ nir_local_variable_create(b.impl, glsl_bool_type(), "available");
+
+ nir_store_var(&b, result,
+ nir_vec2(&b, nir_imm_int64(&b, 0),
+ nir_imm_int64(&b, 0)), 0x3);
+ nir_store_var(&b, available, nir_imm_false(&b), 0x1);
+
+ nir_ssa_def *flags = radv_load_push_int(&b, 0, "flags");
+
+ /* Load resources. */
+ nir_intrinsic_instr *dst_buf = nir_intrinsic_instr_create(b.shader,
+ nir_intrinsic_vulkan_resource_index);
+ dst_buf->src[0] = nir_src_for_ssa(nir_imm_int(&b, 0));
+ dst_buf->num_components = 1;
+ nir_intrinsic_set_desc_set(dst_buf, 0);
+ nir_intrinsic_set_binding(dst_buf, 0);
+ nir_ssa_dest_init(&dst_buf->instr, &dst_buf->dest, dst_buf->num_components, 32, NULL);
+ nir_builder_instr_insert(&b, &dst_buf->instr);
+
+ nir_intrinsic_instr *src_buf = nir_intrinsic_instr_create(b.shader,
+ nir_intrinsic_vulkan_resource_index);
+ src_buf->src[0] = nir_src_for_ssa(nir_imm_int(&b, 0));
+ src_buf->num_components = 1;
+ nir_intrinsic_set_desc_set(src_buf, 0);
+ nir_intrinsic_set_binding(src_buf, 1);
+ nir_ssa_dest_init(&src_buf->instr, &src_buf->dest, src_buf->num_components, 32, NULL);
+ nir_builder_instr_insert(&b, &src_buf->instr);
+
+ /* Compute global ID. */
+ nir_ssa_def *invoc_id = nir_load_local_invocation_id(&b);
+ nir_ssa_def *wg_id = nir_load_work_group_id(&b);
+ nir_ssa_def *block_size = nir_imm_ivec4(&b,
+ b.shader->info.cs.local_size[0],
+ b.shader->info.cs.local_size[1],
+ b.shader->info.cs.local_size[2], 0);
+ nir_ssa_def *global_id = nir_iadd(&b, nir_imul(&b, wg_id, block_size), invoc_id);
+ global_id = nir_channel(&b, global_id, 0); // We only care about x here.
+
+ /* Compute src/dst strides. */
+ nir_ssa_def *input_stride = nir_imm_int(&b, 32);
+ nir_ssa_def *input_base = nir_imul(&b, input_stride, global_id);
+ nir_ssa_def *output_stride = radv_load_push_int(&b, 4, "output_stride");
+ nir_ssa_def *output_base = nir_imul(&b, output_stride, global_id);
+
+ /* Load data from the query pool. */
+ nir_intrinsic_instr *load1 = nir_intrinsic_instr_create(b.shader, nir_intrinsic_load_ssbo);
+ load1->src[0] = nir_src_for_ssa(&src_buf->dest.ssa);
+ load1->src[1] = nir_src_for_ssa(input_base);
+ nir_ssa_dest_init(&load1->instr, &load1->dest, 4, 32, NULL);
+ load1->num_components = 4;
+ nir_builder_instr_insert(&b, &load1->instr);
+
+ nir_intrinsic_instr *load2 = nir_intrinsic_instr_create(b.shader, nir_intrinsic_load_ssbo);
+ load2->src[0] = nir_src_for_ssa(&src_buf->dest.ssa);
+ load2->src[1] = nir_src_for_ssa(nir_iadd(&b, input_base, nir_imm_int(&b, 16)));
+ nir_ssa_dest_init(&load2->instr, &load2->dest, 4, 32, NULL);
+ load2->num_components = 4;
+ nir_builder_instr_insert(&b, &load2->instr);
+
+ /* Check if result is available. */
+ nir_ssa_def *avails[2];
+ avails[0] = nir_iand(&b, nir_channel(&b, &load1->dest.ssa, 1),
+ nir_channel(&b, &load1->dest.ssa, 3));
+ avails[1] = nir_iand(&b, nir_channel(&b, &load2->dest.ssa, 1),
+ nir_channel(&b, &load2->dest.ssa, 3));
+ nir_ssa_def *result_is_available =
+ nir_i2b(&b, nir_iand(&b, nir_iand(&b, avails[0], avails[1]),
+ nir_imm_int(&b, 0x80000000)));
+
+ /* Only compute result if available. */
+ nir_if *available_if = nir_if_create(b.shader);
+ available_if->condition = nir_src_for_ssa(result_is_available);
+ nir_cf_node_insert(b.cursor, &available_if->cf_node);
+
+ b.cursor = nir_after_cf_list(&available_if->then_list);
+
+ /* Pack values. */
+ nir_ssa_def *packed64[4];
+ packed64[0] = nir_pack_64_2x32(&b, nir_vec2(&b,
+ nir_channel(&b, &load1->dest.ssa, 0),
+ nir_channel(&b, &load1->dest.ssa, 1)));
+ packed64[1] = nir_pack_64_2x32(&b, nir_vec2(&b,
+ nir_channel(&b, &load1->dest.ssa, 2),
+ nir_channel(&b, &load1->dest.ssa, 3)));
+ packed64[2] = nir_pack_64_2x32(&b, nir_vec2(&b,
+ nir_channel(&b, &load2->dest.ssa, 0),
+ nir_channel(&b, &load2->dest.ssa, 1)));
+ packed64[3] = nir_pack_64_2x32(&b, nir_vec2(&b,
+ nir_channel(&b, &load2->dest.ssa, 2),
+ nir_channel(&b, &load2->dest.ssa, 3)));
+
+ /* Compute result. */
+ nir_ssa_def *num_primitive_written =
+ nir_isub(&b, packed64[3], packed64[1]);
+ nir_ssa_def *primitive_storage_needed =
+ nir_isub(&b, packed64[2], packed64[0]);
+
+ nir_store_var(&b, result,
+ nir_vec2(&b, num_primitive_written,
+ primitive_storage_needed), 0x3);
+ nir_store_var(&b, available, nir_imm_true(&b), 0x1);
+
+ b.cursor = nir_after_cf_node(&available_if->cf_node);
+
+ /* Determine if result is 64 or 32 bit. */
+ nir_ssa_def *result_is_64bit =
+ nir_test_flag(&b, flags, VK_QUERY_RESULT_64_BIT);
+ nir_ssa_def *result_size =
+ nir_bcsel(&b, result_is_64bit, nir_imm_int(&b, 16),
+ nir_imm_int(&b, 8));
+
+ /* Store the result if complete or partial results have been requested. */
+ nir_if *store_if = nir_if_create(b.shader);
+ store_if->condition =
+ nir_src_for_ssa(nir_ior(&b, nir_test_flag(&b, flags, VK_QUERY_RESULT_PARTIAL_BIT),
+ nir_load_var(&b, available)));
+ nir_cf_node_insert(b.cursor, &store_if->cf_node);
+
+ b.cursor = nir_after_cf_list(&store_if->then_list);
+
+ /* Store result. */
+ nir_if *store_64bit_if = nir_if_create(b.shader);
+ store_64bit_if->condition = nir_src_for_ssa(result_is_64bit);
+ nir_cf_node_insert(b.cursor, &store_64bit_if->cf_node);
+
+ b.cursor = nir_after_cf_list(&store_64bit_if->then_list);
+
+ nir_intrinsic_instr *store = nir_intrinsic_instr_create(b.shader, nir_intrinsic_store_ssbo);
+ store->src[0] = nir_src_for_ssa(nir_load_var(&b, result));
+ store->src[1] = nir_src_for_ssa(&dst_buf->dest.ssa);
+ store->src[2] = nir_src_for_ssa(output_base);
+ nir_intrinsic_set_write_mask(store, 0x3);
+ store->num_components = 2;
+ nir_builder_instr_insert(&b, &store->instr);
+
+ b.cursor = nir_after_cf_list(&store_64bit_if->else_list);
+
+ store = nir_intrinsic_instr_create(b.shader, nir_intrinsic_store_ssbo);
+ store->src[0] = nir_src_for_ssa(nir_u2u32(&b, nir_load_var(&b, result)));
+ store->src[1] = nir_src_for_ssa(&dst_buf->dest.ssa);
+ store->src[2] = nir_src_for_ssa(output_base);
+ nir_intrinsic_set_write_mask(store, 0x3);
+ store->num_components = 2;
+ nir_builder_instr_insert(&b, &store->instr);
+
+ b.cursor = nir_after_cf_node(&store_64bit_if->cf_node);
+
+ b.cursor = nir_after_cf_node(&store_if->cf_node);
+
+ /* Store the availability bit if requested. */
+ nir_if *availability_if = nir_if_create(b.shader);
+ availability_if->condition =
+ nir_src_for_ssa(nir_test_flag(&b, flags, VK_QUERY_RESULT_WITH_AVAILABILITY_BIT));
+ nir_cf_node_insert(b.cursor, &availability_if->cf_node);
+
+ b.cursor = nir_after_cf_list(&availability_if->then_list);
+
+ store = nir_intrinsic_instr_create(b.shader, nir_intrinsic_store_ssbo);
+ store->src[0] = nir_src_for_ssa(nir_b2i32(&b, nir_load_var(&b, available)));
+ store->src[1] = nir_src_for_ssa(&dst_buf->dest.ssa);
+ store->src[2] = nir_src_for_ssa(nir_iadd(&b, result_size, output_base));
+ nir_intrinsic_set_write_mask(store, 0x1);
+ store->num_components = 1;
+ nir_builder_instr_insert(&b, &store->instr);
+
+ b.cursor = nir_after_cf_node(&availability_if->cf_node);
+
+ return b.shader;
+}
+