static VkResult
compute_pipeline_create(
VkDevice _device,
struct anv_pipeline_cache * cache,
const VkComputePipelineCreateInfo* pCreateInfo,
const VkAllocationCallbacks* pAllocator,
VkPipeline* pPipeline)
{
ANV_FROM_HANDLE(anv_device, device, _device);
const struct anv_physical_device *physical_device =
&device->instance->physicalDevice;
const struct gen_device_info *devinfo = &physical_device->info;
struct anv_pipeline *pipeline;
VkResult result;
assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_COMPUTE_PIPELINE_CREATE_INFO);
pipeline = vk_alloc2(&device->alloc, pAllocator, sizeof(*pipeline), 8,
VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
if (pipeline == NULL)
return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
pipeline->device = device;
pipeline->layout = anv_pipeline_layout_from_handle(pCreateInfo->layout);
pipeline->blend_state.map = NULL;
result = anv_reloc_list_init(&pipeline->batch_relocs,
pAllocator ? pAllocator : &device->alloc);
if (result != VK_SUCCESS) {
vk_free2(&device->alloc, pAllocator, pipeline);
return result;
}
pipeline->batch.next = pipeline->batch.start = pipeline->batch_data;
pipeline->batch.end = pipeline->batch.start + sizeof(pipeline->batch_data);
pipeline->batch.relocs = &pipeline->batch_relocs;
/* When we free the pipeline, we detect stages based on the NULL status
* of various prog_data pointers. Make them NULL by default.
*/
memset(pipeline->shaders, 0, sizeof(pipeline->shaders));
pipeline->vs_simd8 = NO_KERNEL;
pipeline->vs_vec4 = NO_KERNEL;
pipeline->gs_kernel = NO_KERNEL;
pipeline->active_stages = 0;
pipeline->needs_data_cache = false;
assert(pCreateInfo->stage.stage == VK_SHADER_STAGE_COMPUTE_BIT);
ANV_FROM_HANDLE(anv_shader_module, module, pCreateInfo->stage.module);
result = anv_pipeline_compile_cs(pipeline, cache, pCreateInfo, module,
pCreateInfo->stage.pName,
pCreateInfo->stage.pSpecializationInfo);
if (result != VK_SUCCESS) {
vk_free2(&device->alloc, pAllocator, pipeline);
return result;
}
const struct brw_cs_prog_data *cs_prog_data = get_cs_prog_data(pipeline);
anv_pipeline_setup_l3_config(pipeline, cs_prog_data->base.total_shared > 0);
uint32_t group_size = cs_prog_data->local_size[0] *
cs_prog_data->local_size[1] * cs_prog_data->local_size[2];
uint32_t remainder = group_size & (cs_prog_data->simd_size - 1);
if (remainder > 0)
pipeline->cs_right_mask = ~0u >> (32 - remainder);
else
VkResult
genX(compute_pipeline_create)(
VkDevice _device,
struct anv_pipeline_cache * cache,
const VkComputePipelineCreateInfo* pCreateInfo,
const VkAllocationCallbacks* pAllocator,
VkPipeline* pPipeline)
{
ANV_FROM_HANDLE(anv_device, device, _device);
struct anv_pipeline *pipeline;
VkResult result;
assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_COMPUTE_PIPELINE_CREATE_INFO);
pipeline = anv_alloc2(&device->alloc, pAllocator, sizeof(*pipeline), 8,
VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
if (pipeline == NULL)
return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
pipeline->device = device;
pipeline->layout = anv_pipeline_layout_from_handle(pCreateInfo->layout);
pipeline->blend_state.map = NULL;
result = anv_reloc_list_init(&pipeline->batch_relocs,
pAllocator ? pAllocator : &device->alloc);
if (result != VK_SUCCESS) {
anv_free2(&device->alloc, pAllocator, pipeline);
return result;
}
pipeline->batch.next = pipeline->batch.start = pipeline->batch_data;
pipeline->batch.end = pipeline->batch.start + sizeof(pipeline->batch_data);
pipeline->batch.relocs = &pipeline->batch_relocs;
/* When we free the pipeline, we detect stages based on the NULL status
* of various prog_data pointers. Make them NULL by default.
*/
memset(pipeline->prog_data, 0, sizeof(pipeline->prog_data));
memset(pipeline->scratch_start, 0, sizeof(pipeline->scratch_start));
memset(pipeline->bindings, 0, sizeof(pipeline->bindings));
pipeline->vs_simd8 = NO_KERNEL;
pipeline->vs_vec4 = NO_KERNEL;
pipeline->gs_kernel = NO_KERNEL;
pipeline->active_stages = 0;
pipeline->total_scratch = 0;
assert(pCreateInfo->stage.stage == VK_SHADER_STAGE_COMPUTE_BIT);
ANV_FROM_HANDLE(anv_shader_module, module, pCreateInfo->stage.module);
anv_pipeline_compile_cs(pipeline, cache, pCreateInfo, module,
pCreateInfo->stage.pName,
pCreateInfo->stage.pSpecializationInfo);
pipeline->use_repclear = false;
const struct brw_cs_prog_data *cs_prog_data = get_cs_prog_data(pipeline);
const struct brw_stage_prog_data *prog_data = &cs_prog_data->base;
unsigned local_id_dwords = cs_prog_data->local_invocation_id_regs * 8;
unsigned push_constant_data_size =
(prog_data->nr_params + local_id_dwords) * 4;
unsigned reg_aligned_constant_size = ALIGN(push_constant_data_size, 32);
unsigned push_constant_regs = reg_aligned_constant_size / 32;
uint32_t group_size = cs_prog_data->local_size[0] *
cs_prog_data->local_size[1] * cs_prog_data->local_size[2];
pipeline->cs_thread_width_max =
DIV_ROUND_UP(group_size, cs_prog_data->simd_size);
uint32_t remainder = group_size & (cs_prog_data->simd_size - 1);
if (remainder > 0)
pipeline->cs_right_mask = ~0u >> (32 - remainder);
else