radv: use a compute shader for copying timestamp query results
authorSamuel Pitoiset <samuel.pitoiset@gmail.com>
Wed, 9 Oct 2019 11:34:52 +0000 (13:34 +0200)
committerSamuel Pitoiset <samuel.pitoiset@gmail.com>
Thu, 10 Oct 2019 11:23:22 +0000 (13:23 +0200)
When the timestamp is not ready (ie. UINT64_MAX), the availabily bit
should be zero. The previous code used to copy the timestamp value
as the availabily bit and that's completely wrong.

Because it's not that simple to emit a conditional with the CP, the
driver now uses a compute shader for copying timestamp query results.

Fixes dEQP-VK.pipeline.timestamp.misc_tests.reset_query_before_copy.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
src/amd/vulkan/radv_private.h
src/amd/vulkan/radv_query.c

index 4babe8b0359de5bf93843bd1748ca270d15b82a4..20a3cd389036d51a90e8fc5850b187877e8d04cb 100644 (file)
@@ -638,6 +638,7 @@ struct radv_meta_state {
                VkPipeline occlusion_query_pipeline;
                VkPipeline pipeline_statistics_query_pipeline;
                VkPipeline tfb_query_pipeline;
+               VkPipeline timestamp_query_pipeline;
        } query;
 
        struct {
index cafe02538649c3eb878c1fcbb77744a3dcffca49..8872bf1bbce7d07c054535cc060527cfb90ba66a 100644 (file)
@@ -741,12 +741,196 @@ build_tfb_query_shader(struct radv_device *device)
        return b.shader;
 }
 
+static nir_shader *
+build_timestamp_query_shader(struct radv_device *device)
+{
+       /* the shader this builds is roughly
+        *
+        * uint32_t src_stride = 8;
+        *
+        * location(binding = 0) buffer dst_buf;
+        * location(binding = 1) buffer src_buf;
+        *
+        * void main() {
+        *      uint64_t result = 0;
+        *      bool available = false;
+        *      uint64_t src_offset = src_stride * global_id.x;
+        *      uint64_t dst_offset = dst_stride * global_id.x;
+        *      uint64_t timestamp = src_buf[src_offset];
+        *      if (timestamp != TIMESTAMP_NOT_READY) {
+        *              result = timestamp;
+        *              available = true;
+        *      }
+        *      uint32_t result_size = flags & VK_QUERY_RESULT_64_BIT ? 8 : 4;
+        *      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, "timestamp_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_uint64_t_type(), "result");
+       nir_variable *available =
+               nir_local_variable_create(b.impl, glsl_bool_type(), "available");
+
+       nir_store_var(&b, result, nir_imm_int64(&b, 0), 0x1);
+       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, 8);
+       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 *load = nir_intrinsic_instr_create(b.shader, nir_intrinsic_load_ssbo);
+       load->src[0] = nir_src_for_ssa(&src_buf->dest.ssa);
+       load->src[1] = nir_src_for_ssa(input_base);
+       nir_ssa_dest_init(&load->instr, &load->dest, 2, 32, NULL);
+       load->num_components = 2;
+       nir_builder_instr_insert(&b, &load->instr);
+
+       /* Pack the timestamp. */
+       nir_ssa_def *timestamp;
+       timestamp = nir_pack_64_2x32(&b, nir_vec2(&b,
+                                                 nir_channel(&b, &load->dest.ssa, 0),
+                                                 nir_channel(&b, &load->dest.ssa, 1)));
+
+       /* Check if result is available. */
+       nir_ssa_def *result_is_available =
+               nir_i2b(&b, nir_ine(&b, timestamp,
+                                   nir_imm_int64(&b, TIMESTAMP_NOT_READY)));
+
+       /* Only store 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);
+
+       nir_store_var(&b, result, timestamp, 0x1);
+       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, 8),
+                         nir_imm_int(&b, 4));
+
+       /* 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, 0x1);
+       store->num_components = 1;
+       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, 0x1);
+       store->num_components = 1;
+       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;
+}
+
 static VkResult radv_device_init_meta_query_state_internal(struct radv_device *device)
 {
        VkResult result;
        struct radv_shader_module occlusion_cs = { .nir = NULL };
        struct radv_shader_module pipeline_statistics_cs = { .nir = NULL };
        struct radv_shader_module tfb_cs = { .nir = NULL };
+       struct radv_shader_module timestamp_cs = { .nir = NULL };
 
        mtx_lock(&device->meta_state.mtx);
        if (device->meta_state.query.pipeline_statistics_query_pipeline) {
@@ -756,6 +940,7 @@ static VkResult radv_device_init_meta_query_state_internal(struct radv_device *d
        occlusion_cs.nir = build_occlusion_query_shader(device);
        pipeline_statistics_cs.nir = build_pipeline_statistics_query_shader(device);
        tfb_cs.nir = build_tfb_query_shader(device);
+       timestamp_cs.nir = build_timestamp_query_shader(device);
 
        VkDescriptorSetLayoutCreateInfo occlusion_ds_create_info = {
                .sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_CREATE_INFO,
@@ -864,12 +1049,36 @@ static VkResult radv_device_init_meta_query_state_internal(struct radv_device *d
                                             radv_pipeline_cache_to_handle(&device->meta_state.cache),
                                             1, &tfb_pipeline_info, NULL,
                                             &device->meta_state.query.tfb_query_pipeline);
+       if (result != VK_SUCCESS)
+               goto fail;
+
+       VkPipelineShaderStageCreateInfo timestamp_pipeline_shader_stage = {
+               .sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO,
+               .stage = VK_SHADER_STAGE_COMPUTE_BIT,
+               .module = radv_shader_module_to_handle(&timestamp_cs),
+               .pName = "main",
+               .pSpecializationInfo = NULL,
+       };
+
+       VkComputePipelineCreateInfo timestamp_pipeline_info = {
+               .sType = VK_STRUCTURE_TYPE_COMPUTE_PIPELINE_CREATE_INFO,
+               .stage = timestamp_pipeline_shader_stage,
+               .flags = 0,
+               .layout = device->meta_state.query.p_layout,
+       };
+
+       result = radv_CreateComputePipelines(radv_device_to_handle(device),
+                                            radv_pipeline_cache_to_handle(&device->meta_state.cache),
+                                            1, &timestamp_pipeline_info, NULL,
+                                            &device->meta_state.query.timestamp_query_pipeline);
+
 fail:
        if (result != VK_SUCCESS)
                radv_device_finish_meta_query_state(device);
        ralloc_free(occlusion_cs.nir);
        ralloc_free(pipeline_statistics_cs.nir);
        ralloc_free(tfb_cs.nir);
+       ralloc_free(timestamp_cs.nir);
        mtx_unlock(&device->meta_state.mtx);
        return result;
 }
@@ -899,6 +1108,11 @@ void radv_device_finish_meta_query_state(struct radv_device *device)
                                     device->meta_state.query.occlusion_query_pipeline,
                                     &device->meta_state.alloc);
 
+       if (device->meta_state.query.timestamp_query_pipeline)
+               radv_DestroyPipeline(radv_device_to_handle(device),
+                                    device->meta_state.query.timestamp_query_pipeline,
+                                    &device->meta_state.alloc);
+
        if (device->meta_state.query.p_layout)
                radv_DestroyPipelineLayout(radv_device_to_handle(device),
                                           device->meta_state.query.p_layout,
@@ -1308,7 +1522,6 @@ void radv_CmdCopyQueryPoolResults(
        RADV_FROM_HANDLE(radv_query_pool, pool, queryPool);
        RADV_FROM_HANDLE(radv_buffer, dst_buffer, dstBuffer);
        struct radeon_cmdbuf *cs = cmd_buffer->cs;
-       unsigned elem_size = (flags & VK_QUERY_RESULT_64_BIT) ? 8 : 4;
        uint64_t va = radv_buffer_get_va(pool->bo);
        uint64_t dest_va = radv_buffer_get_va(dst_buffer->bo);
        dest_va += dst_buffer->offset + dstOffset;
@@ -1368,14 +1581,13 @@ void radv_CmdCopyQueryPoolResults(
                                  pool->availability_offset + 4 * firstQuery);
                break;
        case VK_QUERY_TYPE_TIMESTAMP:
-               for(unsigned i = 0; i < queryCount; ++i, dest_va += stride) {
-                       unsigned query = firstQuery + i;
-                       uint64_t local_src_va = va  + query * pool->stride;
-
-                       ASSERTED unsigned cdw_max = radeon_check_space(cmd_buffer->device->ws, cs, 19);
+               if (flags & VK_QUERY_RESULT_WAIT_BIT) {
+                       for(unsigned i = 0; i < queryCount; ++i, dest_va += stride) {
+                               unsigned query = firstQuery + i;
+                               uint64_t local_src_va = va  + query * pool->stride;
 
+                               radeon_check_space(cmd_buffer->device->ws, cs, 7);
 
-                       if (flags & VK_QUERY_RESULT_WAIT_BIT) {
                                /* Wait on the high 32 bits of the timestamp in
                                 * case the low part is 0xffffffff.
                                 */
@@ -1384,30 +1596,14 @@ void radv_CmdCopyQueryPoolResults(
                                                 TIMESTAMP_NOT_READY >> 32,
                                                 0xffffffff);
                        }
-                       if (flags & VK_QUERY_RESULT_WITH_AVAILABILITY_BIT) {
-                               uint64_t avail_dest_va = dest_va + elem_size;
-
-                               radeon_emit(cs, PKT3(PKT3_COPY_DATA, 4, 0));
-                               radeon_emit(cs, COPY_DATA_SRC_SEL(COPY_DATA_SRC_MEM) |
-                                               COPY_DATA_DST_SEL(COPY_DATA_DST_MEM_GRBM));
-                               radeon_emit(cs, local_src_va);
-                               radeon_emit(cs, local_src_va >> 32);
-                               radeon_emit(cs, avail_dest_va);
-                               radeon_emit(cs, avail_dest_va >> 32);
-                       }
-
-                       radeon_emit(cs, PKT3(PKT3_COPY_DATA, 4, 0));
-                       radeon_emit(cs, COPY_DATA_SRC_SEL(COPY_DATA_SRC_MEM) |
-                                       COPY_DATA_DST_SEL(COPY_DATA_DST_MEM_GRBM) |
-                                       ((flags & VK_QUERY_RESULT_64_BIT) ? COPY_DATA_COUNT_SEL : 0));
-                       radeon_emit(cs, local_src_va);
-                       radeon_emit(cs, local_src_va >> 32);
-                       radeon_emit(cs, dest_va);
-                       radeon_emit(cs, dest_va >> 32);
-
-
-                       assert(cs->cdw <= cdw_max);
                }
+
+               radv_query_shader(cmd_buffer, &cmd_buffer->device->meta_state.query.timestamp_query_pipeline,
+                                 pool->bo, dst_buffer->bo,
+                                 firstQuery * pool->stride,
+                                 dst_buffer->offset + dstOffset,
+                                 pool->stride, stride,
+                                 queryCount, flags, 0, 0);
                break;
        case VK_QUERY_TYPE_TRANSFORM_FEEDBACK_STREAM_EXT:
                if (flags & VK_QUERY_RESULT_WAIT_BIT) {