X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fvulkan%2Fanv_cmd_buffer.c;h=0966e7658bfb95d5118383a9cfb474f240d5a501;hb=ac0589b213f21f22dc330d2e8fb8dceda3960ddc;hp=28d9dd9d69426e64dbf45643d27857b4505b1e8e;hpb=2228ec0112cbc21f0bac986582ef706476094f0b;p=mesa.git diff --git a/src/vulkan/anv_cmd_buffer.c b/src/vulkan/anv_cmd_buffer.c index 28d9dd9d694..0966e7658bf 100644 --- a/src/vulkan/anv_cmd_buffer.c +++ b/src/vulkan/anv_cmd_buffer.c @@ -50,7 +50,7 @@ const struct anv_dynamic_state default_dynamic_state = { .depth_bias = { .bias = 0.0f, .clamp = 0.0f, - .slope_scaled = 0.0f, + .slope = 0.0f, }, .blend_constants = { 0.0f, 0.0f, 0.0f, 0.0f }, .depth_bounds = { @@ -111,46 +111,108 @@ anv_dynamic_state_copy(struct anv_dynamic_state *dest, } static void -anv_cmd_state_init(struct anv_cmd_state *state) +anv_cmd_state_reset(struct anv_cmd_buffer *cmd_buffer) { - memset(&state->state_vf, 0, sizeof(state->state_vf)); + struct anv_cmd_state *state = &cmd_buffer->state; + memset(&state->descriptors, 0, sizeof(state->descriptors)); memset(&state->push_constants, 0, sizeof(state->push_constants)); + memset(state->binding_tables, 0, sizeof(state->binding_tables)); + memset(state->samplers, 0, sizeof(state->samplers)); + + /* 0 isn't a valid config. This ensures that we always configure L3$. */ + cmd_buffer->state.current_l3_config = 0; state->dirty = ~0; state->vb_dirty = 0; state->descriptors_dirty = 0; state->push_constants_dirty = 0; state->pipeline = NULL; + state->restart_index = UINT32_MAX; state->dynamic = default_dynamic_state; + state->need_query_wa = true; + + if (state->attachments != NULL) { + anv_free(&cmd_buffer->pool->alloc, state->attachments); + state->attachments = NULL; + } state->gen7.index_buffer = NULL; } +/** + * Setup anv_cmd_state::attachments for vkCmdBeginRenderPass. + */ +void +anv_cmd_state_setup_attachments(struct anv_cmd_buffer *cmd_buffer, + const VkRenderPassBeginInfo *info) +{ + struct anv_cmd_state *state = &cmd_buffer->state; + ANV_FROM_HANDLE(anv_render_pass, pass, info->renderPass); + + anv_free(&cmd_buffer->pool->alloc, state->attachments); + + if (pass->attachment_count == 0) { + state->attachments = NULL; + return; + } + + state->attachments = anv_alloc(&cmd_buffer->pool->alloc, + pass->attachment_count * + sizeof(state->attachments[0]), + 8, VK_SYSTEM_ALLOCATION_SCOPE_OBJECT); + if (state->attachments == NULL) { + /* FIXME: Propagate VK_ERROR_OUT_OF_HOST_MEMORY to vkEndCommandBuffer */ + abort(); + } + + for (uint32_t i = 0; i < pass->attachment_count; ++i) { + struct anv_render_pass_attachment *att = &pass->attachments[i]; + VkImageAspectFlags clear_aspects = 0; + + if (anv_format_is_color(att->format)) { + /* color attachment */ + if (att->load_op == VK_ATTACHMENT_LOAD_OP_CLEAR) { + clear_aspects |= VK_IMAGE_ASPECT_COLOR_BIT; + } + } else { + /* depthstencil attachment */ + if (att->format->depth_format && + att->load_op == VK_ATTACHMENT_LOAD_OP_CLEAR) { + clear_aspects |= VK_IMAGE_ASPECT_DEPTH_BIT; + } + if (att->format->has_stencil && + att->stencil_load_op == VK_ATTACHMENT_LOAD_OP_CLEAR) { + clear_aspects |= VK_IMAGE_ASPECT_STENCIL_BIT; + } + } + + state->attachments[i].pending_clear_aspects = clear_aspects; + if (clear_aspects) { + assert(info->clearValueCount > i); + state->attachments[i].clear_value = info->pClearValues[i]; + } + } +} + static VkResult anv_cmd_buffer_ensure_push_constants_size(struct anv_cmd_buffer *cmd_buffer, - VkShaderStage stage, uint32_t size) + gl_shader_stage stage, uint32_t size) { struct anv_push_constants **ptr = &cmd_buffer->state.push_constants[stage]; if (*ptr == NULL) { - *ptr = anv_device_alloc(cmd_buffer->device, size, 8, - VK_SYSTEM_ALLOC_TYPE_INTERNAL); + *ptr = anv_alloc(&cmd_buffer->pool->alloc, size, 8, + VK_SYSTEM_ALLOCATION_SCOPE_OBJECT); if (*ptr == NULL) return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY); - (*ptr)->size = size; } else if ((*ptr)->size < size) { - void *new_data = anv_device_alloc(cmd_buffer->device, size, 8, - VK_SYSTEM_ALLOC_TYPE_INTERNAL); - if (new_data == NULL) + *ptr = anv_realloc(&cmd_buffer->pool->alloc, *ptr, size, 8, + VK_SYSTEM_ALLOCATION_SCOPE_OBJECT); + if (*ptr == NULL) return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY); - - memcpy(new_data, *ptr, (*ptr)->size); - anv_device_free(cmd_buffer->device, *ptr); - - *ptr = new_data; - (*ptr)->size = size; } + (*ptr)->size = size; return VK_SUCCESS; } @@ -160,23 +222,25 @@ anv_cmd_buffer_ensure_push_constants_size(struct anv_cmd_buffer *cmd_buffer, (offsetof(struct anv_push_constants, field) + \ sizeof(cmd_buffer->state.push_constants[0]->field))) -VkResult anv_CreateCommandBuffer( - VkDevice _device, - const VkCmdBufferCreateInfo* pCreateInfo, - VkCmdBuffer* pCmdBuffer) +static VkResult anv_create_cmd_buffer( + struct anv_device * device, + struct anv_cmd_pool * pool, + VkCommandBufferLevel level, + VkCommandBuffer* pCommandBuffer) { - ANV_FROM_HANDLE(anv_device, device, _device); - ANV_FROM_HANDLE(anv_cmd_pool, pool, pCreateInfo->cmdPool); struct anv_cmd_buffer *cmd_buffer; VkResult result; - cmd_buffer = anv_device_alloc(device, sizeof(*cmd_buffer), 8, - VK_SYSTEM_ALLOC_TYPE_API_OBJECT); + cmd_buffer = anv_alloc(&pool->alloc, sizeof(*cmd_buffer), 8, + VK_SYSTEM_ALLOCATION_SCOPE_OBJECT); if (cmd_buffer == NULL) return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY); cmd_buffer->_loader_data.loaderMagic = ICD_LOADER_MAGIC; cmd_buffer->device = device; + cmd_buffer->pool = pool; + cmd_buffer->level = level; + cmd_buffer->state.attachments = NULL; result = anv_cmd_buffer_init_batch_bo_chain(cmd_buffer); if (result != VK_SUCCESS) @@ -187,11 +251,6 @@ VkResult anv_CreateCommandBuffer( anv_state_stream_init(&cmd_buffer->dynamic_state_stream, &device->dynamic_state_block_pool); - cmd_buffer->level = pCreateInfo->level; - cmd_buffer->opt_flags = 0; - - anv_cmd_state_init(&cmd_buffer->state); - if (pool) { list_addtail(&cmd_buffer->pool_link, &pool->cmd_buffers); } else { @@ -201,40 +260,78 @@ VkResult anv_CreateCommandBuffer( list_inithead(&cmd_buffer->pool_link); } - *pCmdBuffer = anv_cmd_buffer_to_handle(cmd_buffer); + *pCommandBuffer = anv_cmd_buffer_to_handle(cmd_buffer); return VK_SUCCESS; - fail: anv_device_free(device, cmd_buffer); + fail: + anv_free(&cmd_buffer->pool->alloc, cmd_buffer); return result; } -void anv_DestroyCommandBuffer( +VkResult anv_AllocateCommandBuffers( VkDevice _device, - VkCmdBuffer _cmd_buffer) + const VkCommandBufferAllocateInfo* pAllocateInfo, + VkCommandBuffer* pCommandBuffers) { ANV_FROM_HANDLE(anv_device, device, _device); - ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, _cmd_buffer); + ANV_FROM_HANDLE(anv_cmd_pool, pool, pAllocateInfo->commandPool); + + VkResult result = VK_SUCCESS; + uint32_t i; + + for (i = 0; i < pAllocateInfo->commandBufferCount; i++) { + result = anv_create_cmd_buffer(device, pool, pAllocateInfo->level, + &pCommandBuffers[i]); + if (result != VK_SUCCESS) + break; + } + + if (result != VK_SUCCESS) + anv_FreeCommandBuffers(_device, pAllocateInfo->commandPool, + i, pCommandBuffers); + + return result; +} +static void +anv_cmd_buffer_destroy(struct anv_cmd_buffer *cmd_buffer) +{ list_del(&cmd_buffer->pool_link); anv_cmd_buffer_fini_batch_bo_chain(cmd_buffer); anv_state_stream_finish(&cmd_buffer->surface_state_stream); anv_state_stream_finish(&cmd_buffer->dynamic_state_stream); - anv_device_free(device, cmd_buffer); + + anv_free(&cmd_buffer->pool->alloc, cmd_buffer->state.attachments); + anv_free(&cmd_buffer->pool->alloc, cmd_buffer); +} + +void anv_FreeCommandBuffers( + VkDevice device, + VkCommandPool commandPool, + uint32_t commandBufferCount, + const VkCommandBuffer* pCommandBuffers) +{ + for (uint32_t i = 0; i < commandBufferCount; i++) { + ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, pCommandBuffers[i]); + + anv_cmd_buffer_destroy(cmd_buffer); + } } VkResult anv_ResetCommandBuffer( - VkCmdBuffer cmdBuffer, - VkCmdBufferResetFlags flags) + VkCommandBuffer commandBuffer, + VkCommandBufferResetFlags flags) { - ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, cmdBuffer); + ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer); + cmd_buffer->usage_flags = 0; + cmd_buffer->state.current_pipeline = UINT32_MAX; anv_cmd_buffer_reset_batch_bo_chain(cmd_buffer); - - anv_cmd_state_init(&cmd_buffer->state); + anv_cmd_state_reset(cmd_buffer); return VK_SUCCESS; } @@ -244,49 +341,73 @@ anv_cmd_buffer_emit_state_base_address(struct anv_cmd_buffer *cmd_buffer) { switch (cmd_buffer->device->info.gen) { case 7: - return gen7_cmd_buffer_emit_state_base_address(cmd_buffer); + if (cmd_buffer->device->info.is_haswell) + return gen7_cmd_buffer_emit_state_base_address(cmd_buffer); + else + return gen7_cmd_buffer_emit_state_base_address(cmd_buffer); case 8: return gen8_cmd_buffer_emit_state_base_address(cmd_buffer); + case 9: + return gen9_cmd_buffer_emit_state_base_address(cmd_buffer); default: unreachable("unsupported gen\n"); } } VkResult anv_BeginCommandBuffer( - VkCmdBuffer cmdBuffer, - const VkCmdBufferBeginInfo* pBeginInfo) + VkCommandBuffer commandBuffer, + const VkCommandBufferBeginInfo* pBeginInfo) { - ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, cmdBuffer); + ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer); + + /* If this is the first vkBeginCommandBuffer, we must *initialize* the + * command buffer's state. Otherwise, we must *reset* its state. In both + * cases we reset it. + * + * From the Vulkan 1.0 spec: + * + * If a command buffer is in the executable state and the command buffer + * was allocated from a command pool with the + * VK_COMMAND_POOL_CREATE_RESET_COMMAND_BUFFER_BIT flag set, then + * vkBeginCommandBuffer implicitly resets the command buffer, behaving + * as if vkResetCommandBuffer had been called with + * VK_COMMAND_BUFFER_RESET_RELEASE_RESOURCES_BIT not set. It then puts + * the command buffer in the recording state. + */ + anv_ResetCommandBuffer(commandBuffer, /*flags*/ 0); + + cmd_buffer->usage_flags = pBeginInfo->flags; - cmd_buffer->opt_flags = pBeginInfo->flags; + assert(cmd_buffer->level == VK_COMMAND_BUFFER_LEVEL_SECONDARY || + !(cmd_buffer->usage_flags & VK_COMMAND_BUFFER_USAGE_RENDER_PASS_CONTINUE_BIT)); - if (cmd_buffer->level == VK_CMD_BUFFER_LEVEL_SECONDARY) { + anv_cmd_buffer_emit_state_base_address(cmd_buffer); + + if (cmd_buffer->usage_flags & + VK_COMMAND_BUFFER_USAGE_RENDER_PASS_CONTINUE_BIT) { cmd_buffer->state.framebuffer = - anv_framebuffer_from_handle(pBeginInfo->framebuffer); + anv_framebuffer_from_handle(pBeginInfo->pInheritanceInfo->framebuffer); cmd_buffer->state.pass = - anv_render_pass_from_handle(pBeginInfo->renderPass); + anv_render_pass_from_handle(pBeginInfo->pInheritanceInfo->renderPass); struct anv_subpass *subpass = - &cmd_buffer->state.pass->subpasses[pBeginInfo->subpass]; + &cmd_buffer->state.pass->subpasses[pBeginInfo->pInheritanceInfo->subpass]; - anv_cmd_buffer_begin_subpass(cmd_buffer, subpass); + anv_cmd_buffer_set_subpass(cmd_buffer, subpass); } - anv_cmd_buffer_emit_state_base_address(cmd_buffer); - cmd_buffer->state.current_pipeline = UINT32_MAX; - return VK_SUCCESS; } VkResult anv_EndCommandBuffer( - VkCmdBuffer cmdBuffer) + VkCommandBuffer commandBuffer) { - ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, cmdBuffer); + ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer); struct anv_device *device = cmd_buffer->device; anv_cmd_buffer_end_batch_buffer(cmd_buffer); - if (cmd_buffer->level == VK_CMD_BUFFER_LEVEL_PRIMARY) { + if (cmd_buffer->level == VK_COMMAND_BUFFER_LEVEL_PRIMARY) { /* The algorithm used to compute the validate list is not threadsafe as * it uses the bo->index field. We have to lock the device around it. * Fortunately, the chances for contention here are probably very low. @@ -300,24 +421,24 @@ VkResult anv_EndCommandBuffer( } void anv_CmdBindPipeline( - VkCmdBuffer cmdBuffer, + VkCommandBuffer commandBuffer, VkPipelineBindPoint pipelineBindPoint, VkPipeline _pipeline) { - ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, cmdBuffer); + ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer); ANV_FROM_HANDLE(anv_pipeline, pipeline, _pipeline); switch (pipelineBindPoint) { case VK_PIPELINE_BIND_POINT_COMPUTE: cmd_buffer->state.compute_pipeline = pipeline; - cmd_buffer->state.compute_dirty |= ANV_CMD_BUFFER_PIPELINE_DIRTY; + cmd_buffer->state.compute_dirty |= ANV_CMD_DIRTY_PIPELINE; cmd_buffer->state.push_constants_dirty |= VK_SHADER_STAGE_COMPUTE_BIT; break; case VK_PIPELINE_BIND_POINT_GRAPHICS: cmd_buffer->state.pipeline = pipeline; cmd_buffer->state.vb_dirty |= pipeline->vb_used; - cmd_buffer->state.dirty |= ANV_CMD_BUFFER_PIPELINE_DIRTY; + cmd_buffer->state.dirty |= ANV_CMD_DIRTY_PIPELINE; cmd_buffer->state.push_constants_dirty |= pipeline->active_stages; /* Apply the dynamic state from the pipeline */ @@ -334,193 +455,212 @@ void anv_CmdBindPipeline( } void anv_CmdSetViewport( - VkCmdBuffer cmdBuffer, + VkCommandBuffer commandBuffer, + uint32_t firstViewport, uint32_t viewportCount, const VkViewport* pViewports) { - ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, cmdBuffer); + ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer); + + const uint32_t total_count = firstViewport + viewportCount; + if (cmd_buffer->state.dynamic.viewport.count < total_count); + cmd_buffer->state.dynamic.viewport.count = total_count; - cmd_buffer->state.dynamic.viewport.count = viewportCount; - memcpy(cmd_buffer->state.dynamic.viewport.viewports, + memcpy(cmd_buffer->state.dynamic.viewport.viewports + firstViewport, pViewports, viewportCount * sizeof(*pViewports)); - cmd_buffer->state.dirty |= ANV_DYNAMIC_VIEWPORT_DIRTY; + cmd_buffer->state.dirty |= ANV_CMD_DIRTY_DYNAMIC_VIEWPORT; } void anv_CmdSetScissor( - VkCmdBuffer cmdBuffer, + VkCommandBuffer commandBuffer, + uint32_t firstScissor, uint32_t scissorCount, const VkRect2D* pScissors) { - ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, cmdBuffer); + ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer); + + const uint32_t total_count = firstScissor + scissorCount; + if (cmd_buffer->state.dynamic.scissor.count < total_count); + cmd_buffer->state.dynamic.scissor.count = total_count; - cmd_buffer->state.dynamic.scissor.count = scissorCount; - memcpy(cmd_buffer->state.dynamic.scissor.scissors, + memcpy(cmd_buffer->state.dynamic.scissor.scissors + firstScissor, pScissors, scissorCount * sizeof(*pScissors)); - cmd_buffer->state.dirty |= ANV_DYNAMIC_SCISSOR_DIRTY; + cmd_buffer->state.dirty |= ANV_CMD_DIRTY_DYNAMIC_SCISSOR; } void anv_CmdSetLineWidth( - VkCmdBuffer cmdBuffer, + VkCommandBuffer commandBuffer, float lineWidth) { - ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, cmdBuffer); + ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer); cmd_buffer->state.dynamic.line_width = lineWidth; - - cmd_buffer->state.dirty |= ANV_DYNAMIC_LINE_WIDTH_DIRTY; + cmd_buffer->state.dirty |= ANV_CMD_DIRTY_DYNAMIC_LINE_WIDTH; } void anv_CmdSetDepthBias( - VkCmdBuffer cmdBuffer, - float depthBias, + VkCommandBuffer commandBuffer, + float depthBiasConstantFactor, float depthBiasClamp, - float slopeScaledDepthBias) + float depthBiasSlopeFactor) { - ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, cmdBuffer); + ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer); - cmd_buffer->state.dynamic.depth_bias.bias = depthBias; + cmd_buffer->state.dynamic.depth_bias.bias = depthBiasConstantFactor; cmd_buffer->state.dynamic.depth_bias.clamp = depthBiasClamp; - cmd_buffer->state.dynamic.depth_bias.slope_scaled = slopeScaledDepthBias; + cmd_buffer->state.dynamic.depth_bias.slope = depthBiasSlopeFactor; - cmd_buffer->state.dirty |= ANV_DYNAMIC_DEPTH_BIAS_DIRTY; + cmd_buffer->state.dirty |= ANV_CMD_DIRTY_DYNAMIC_DEPTH_BIAS; } void anv_CmdSetBlendConstants( - VkCmdBuffer cmdBuffer, - const float blendConst[4]) + VkCommandBuffer commandBuffer, + const float blendConstants[4]) { - ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, cmdBuffer); + ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer); memcpy(cmd_buffer->state.dynamic.blend_constants, - blendConst, sizeof(float) * 4); + blendConstants, sizeof(float) * 4); - cmd_buffer->state.dirty |= ANV_DYNAMIC_BLEND_CONSTANTS_DIRTY; + cmd_buffer->state.dirty |= ANV_CMD_DIRTY_DYNAMIC_BLEND_CONSTANTS; } void anv_CmdSetDepthBounds( - VkCmdBuffer cmdBuffer, + VkCommandBuffer commandBuffer, float minDepthBounds, float maxDepthBounds) { - ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, cmdBuffer); + ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer); cmd_buffer->state.dynamic.depth_bounds.min = minDepthBounds; cmd_buffer->state.dynamic.depth_bounds.max = maxDepthBounds; - cmd_buffer->state.dirty |= ANV_DYNAMIC_DEPTH_BOUNDS_DIRTY; + cmd_buffer->state.dirty |= ANV_CMD_DIRTY_DYNAMIC_DEPTH_BOUNDS; } void anv_CmdSetStencilCompareMask( - VkCmdBuffer cmdBuffer, + VkCommandBuffer commandBuffer, VkStencilFaceFlags faceMask, - uint32_t stencilCompareMask) + uint32_t compareMask) { - ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, cmdBuffer); + ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer); if (faceMask & VK_STENCIL_FACE_FRONT_BIT) - cmd_buffer->state.dynamic.stencil_compare_mask.front = stencilCompareMask; + cmd_buffer->state.dynamic.stencil_compare_mask.front = compareMask; if (faceMask & VK_STENCIL_FACE_BACK_BIT) - cmd_buffer->state.dynamic.stencil_compare_mask.back = stencilCompareMask; + cmd_buffer->state.dynamic.stencil_compare_mask.back = compareMask; - cmd_buffer->state.dirty |= ANV_DYNAMIC_STENCIL_COMPARE_MASK_DIRTY; + cmd_buffer->state.dirty |= ANV_CMD_DIRTY_DYNAMIC_STENCIL_COMPARE_MASK; } void anv_CmdSetStencilWriteMask( - VkCmdBuffer cmdBuffer, + VkCommandBuffer commandBuffer, VkStencilFaceFlags faceMask, - uint32_t stencilWriteMask) + uint32_t writeMask) { - ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, cmdBuffer); + ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer); if (faceMask & VK_STENCIL_FACE_FRONT_BIT) - cmd_buffer->state.dynamic.stencil_write_mask.front = stencilWriteMask; + cmd_buffer->state.dynamic.stencil_write_mask.front = writeMask; if (faceMask & VK_STENCIL_FACE_BACK_BIT) - cmd_buffer->state.dynamic.stencil_write_mask.back = stencilWriteMask; + cmd_buffer->state.dynamic.stencil_write_mask.back = writeMask; - cmd_buffer->state.dirty |= ANV_DYNAMIC_STENCIL_WRITE_MASK_DIRTY; + cmd_buffer->state.dirty |= ANV_CMD_DIRTY_DYNAMIC_STENCIL_WRITE_MASK; } void anv_CmdSetStencilReference( - VkCmdBuffer cmdBuffer, + VkCommandBuffer commandBuffer, VkStencilFaceFlags faceMask, - uint32_t stencilReference) + uint32_t reference) { - ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, cmdBuffer); + ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer); if (faceMask & VK_STENCIL_FACE_FRONT_BIT) - cmd_buffer->state.dynamic.stencil_reference.front = stencilReference; + cmd_buffer->state.dynamic.stencil_reference.front = reference; if (faceMask & VK_STENCIL_FACE_BACK_BIT) - cmd_buffer->state.dynamic.stencil_reference.back = stencilReference; + cmd_buffer->state.dynamic.stencil_reference.back = reference; - cmd_buffer->state.dirty |= ANV_DYNAMIC_STENCIL_REFERENCE_DIRTY; + cmd_buffer->state.dirty |= ANV_CMD_DIRTY_DYNAMIC_STENCIL_REFERENCE; } void anv_CmdBindDescriptorSets( - VkCmdBuffer cmdBuffer, + VkCommandBuffer commandBuffer, VkPipelineBindPoint pipelineBindPoint, VkPipelineLayout _layout, uint32_t firstSet, - uint32_t setCount, + uint32_t descriptorSetCount, const VkDescriptorSet* pDescriptorSets, uint32_t dynamicOffsetCount, const uint32_t* pDynamicOffsets) { - ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, cmdBuffer); + ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer); ANV_FROM_HANDLE(anv_pipeline_layout, layout, _layout); struct anv_descriptor_set_layout *set_layout; - assert(firstSet + setCount < MAX_SETS); + assert(firstSet + descriptorSetCount < MAX_SETS); uint32_t dynamic_slot = 0; - for (uint32_t i = 0; i < setCount; i++) { + for (uint32_t i = 0; i < descriptorSetCount; i++) { ANV_FROM_HANDLE(anv_descriptor_set, set, pDescriptorSets[i]); set_layout = layout->set[firstSet + i].layout; - if (cmd_buffer->state.descriptors[firstSet + i].set != set) { - cmd_buffer->state.descriptors[firstSet + i].set = set; + if (cmd_buffer->state.descriptors[firstSet + i] != set) { + cmd_buffer->state.descriptors[firstSet + i] = set; cmd_buffer->state.descriptors_dirty |= set_layout->shader_stages; } - if (set_layout->num_dynamic_buffers > 0) { - VkShaderStage s; - for_each_bit(s, set_layout->shader_stages) { - anv_cmd_buffer_ensure_push_constant_field(cmd_buffer, s, - dynamic_offsets); - uint32_t *offsets = - cmd_buffer->state.push_constants[s]->dynamic_offsets + - layout->set[firstSet + i].dynamic_offset_start; - - memcpy(offsets, pDynamicOffsets + dynamic_slot, - set_layout->num_dynamic_buffers * sizeof(*pDynamicOffsets)); - + if (set_layout->dynamic_offset_count > 0) { + anv_foreach_stage(s, set_layout->shader_stages) { + anv_cmd_buffer_ensure_push_constant_field(cmd_buffer, s, dynamic); + + struct anv_push_constants *push = + cmd_buffer->state.push_constants[s]; + + unsigned d = layout->set[firstSet + i].dynamic_offset_start; + const uint32_t *offsets = pDynamicOffsets + dynamic_slot; + struct anv_descriptor *desc = set->descriptors; + + for (unsigned b = 0; b < set_layout->binding_count; b++) { + if (set_layout->binding[b].dynamic_offset_index < 0) + continue; + + unsigned array_size = set_layout->binding[b].array_size; + for (unsigned j = 0; j < array_size; j++) { + uint32_t range = 0; + if (desc->buffer_view) + range = desc->buffer_view->range; + push->dynamic[d].offset = *(offsets++); + push->dynamic[d].range = range; + desc++; + d++; + } + } } cmd_buffer->state.push_constants_dirty |= set_layout->shader_stages; - - dynamic_slot += set_layout->num_dynamic_buffers; } } } void anv_CmdBindVertexBuffers( - VkCmdBuffer cmdBuffer, - uint32_t startBinding, + VkCommandBuffer commandBuffer, + uint32_t firstBinding, uint32_t bindingCount, const VkBuffer* pBuffers, const VkDeviceSize* pOffsets) { - ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, cmdBuffer); + ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer); struct anv_vertex_binding *vb = cmd_buffer->state.vertex_bindings; /* We have to defer setting up vertex buffer since we need the buffer * stride from the pipeline. */ - assert(startBinding + bindingCount < MAX_VBS); + assert(firstBinding + bindingCount < MAX_VBS); for (uint32_t i = 0; i < bindingCount; i++) { - vb[startBinding + i].buffer = anv_buffer_from_handle(pBuffers[i]); - vb[startBinding + i].offset = pOffsets[i]; - cmd_buffer->state.vb_dirty |= 1 << (startBinding + i); + vb[firstBinding + i].buffer = anv_buffer_from_handle(pBuffers[i]); + vb[firstBinding + i].offset = pOffsets[i]; + cmd_buffer->state.vb_dirty |= 1 << (firstBinding + i); } } @@ -534,30 +674,53 @@ add_surface_state_reloc(struct anv_cmd_buffer *cmd_buffer, const uint32_t dword = cmd_buffer->device->info.gen < 8 ? 1 : 8; - anv_reloc_list_add(&cmd_buffer->surface_relocs, cmd_buffer->device, + anv_reloc_list_add(&cmd_buffer->surface_relocs, &cmd_buffer->pool->alloc, state.offset + dword * 4, bo, offset); } +const struct anv_format * +anv_format_for_descriptor_type(VkDescriptorType type) +{ + switch (type) { + case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER: + case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC: + return anv_format_for_vk_format(VK_FORMAT_R32G32B32A32_SFLOAT); + + case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER: + case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC: + return anv_format_for_vk_format(VK_FORMAT_UNDEFINED); + + default: + unreachable("Invalid descriptor type"); + } +} + VkResult anv_cmd_buffer_emit_binding_table(struct anv_cmd_buffer *cmd_buffer, - VkShaderStage stage, struct anv_state *bt_state) + gl_shader_stage stage, + struct anv_state *bt_state) { struct anv_framebuffer *fb = cmd_buffer->state.framebuffer; struct anv_subpass *subpass = cmd_buffer->state.subpass; struct anv_pipeline_layout *layout; - uint32_t attachments, bias, state_offset; + uint32_t color_count, bias, state_offset; - if (stage == VK_SHADER_STAGE_COMPUTE) - layout = cmd_buffer->state.compute_pipeline->layout; - else + switch (stage) { + case MESA_SHADER_FRAGMENT: layout = cmd_buffer->state.pipeline->layout; - - if (stage == VK_SHADER_STAGE_FRAGMENT) { bias = MAX_RTS; - attachments = subpass->color_count; - } else { + color_count = subpass->color_count; + break; + case MESA_SHADER_COMPUTE: + layout = cmd_buffer->state.compute_pipeline->layout; + bias = 1; + color_count = 0; + break; + default: + layout = cmd_buffer->state.pipeline->layout; bias = 0; - attachments = 0; + color_count = 0; + break; } /* This is a little awkward: layout can be NULL but we still have to @@ -565,8 +728,10 @@ anv_cmd_buffer_emit_binding_table(struct anv_cmd_buffer *cmd_buffer, * targets. */ uint32_t surface_count = layout ? layout->stage[stage].surface_count : 0; - if (attachments + surface_count == 0) + if (color_count + surface_count == 0) { + *bt_state = (struct anv_state) { 0, }; return VK_SUCCESS; + } *bt_state = anv_cmd_buffer_alloc_binding_table(cmd_buffer, bias + surface_count, @@ -576,79 +741,146 @@ anv_cmd_buffer_emit_binding_table(struct anv_cmd_buffer *cmd_buffer, if (bt_state->map == NULL) return VK_ERROR_OUT_OF_DEVICE_MEMORY; - /* This is highly annoying. The Vulkan spec puts the depth-stencil - * attachments in with the color attachments. Unfortunately, thanks to - * other aspects of the API, we cana't really saparate them before this - * point. Therefore, we have to walk all of the attachments but only - * put the color attachments into the binding table. - */ - for (uint32_t a = 0; a < attachments; a++) { + for (uint32_t a = 0; a < color_count; a++) { const struct anv_image_view *iview = fb->attachments[subpass->color_attachments[a]]; + assert(iview->color_rt_surface_state.alloc_size); bt_map[a] = iview->color_rt_surface_state.offset + state_offset; add_surface_state_reloc(cmd_buffer, iview->color_rt_surface_state, iview->bo, iview->offset); } + if (stage == MESA_SHADER_COMPUTE && + cmd_buffer->state.compute_pipeline->cs_prog_data.uses_num_work_groups) { + struct anv_bo *bo = cmd_buffer->state.num_workgroups_bo; + uint32_t bo_offset = cmd_buffer->state.num_workgroups_offset; + + struct anv_state surface_state; + surface_state = + anv_cmd_buffer_alloc_surface_state(cmd_buffer); + + const struct anv_format *format = + anv_format_for_descriptor_type(VK_DESCRIPTOR_TYPE_STORAGE_BUFFER); + anv_fill_buffer_surface_state(cmd_buffer->device, surface_state, + format->surface_format, bo_offset, 12, 1); + + bt_map[0] = surface_state.offset + state_offset; + add_surface_state_reloc(cmd_buffer, surface_state, bo, bo_offset); + } + if (layout == NULL) - return VK_SUCCESS; + goto out; - for (uint32_t set = 0; set < layout->num_sets; set++) { - struct anv_descriptor_set_binding *d = &cmd_buffer->state.descriptors[set]; - struct anv_descriptor_set_layout *set_layout = layout->set[set].layout; - struct anv_descriptor_slot *surface_slots = - set_layout->stage[stage].surface_start; - - uint32_t start = bias + layout->set[set].stage[stage].surface_start; - - for (uint32_t b = 0; b < set_layout->stage[stage].surface_count; b++) { - struct anv_descriptor *desc = - &d->set->descriptors[surface_slots[b].index]; - - const struct anv_state *surface_state; - struct anv_bo *bo; - uint32_t bo_offset; - - switch (desc->type) { - case ANV_DESCRIPTOR_TYPE_EMPTY: - case ANV_DESCRIPTOR_TYPE_SAMPLER: - continue; - case ANV_DESCRIPTOR_TYPE_BUFFER_VIEW: - surface_state = &desc->buffer_view->surface_state; - bo = desc->buffer_view->bo; - bo_offset = desc->buffer_view->offset; - break; - case ANV_DESCRIPTOR_TYPE_IMAGE_VIEW: - surface_state = &desc->image_view->nonrt_surface_state; - bo = desc->image_view->bo; - bo_offset = desc->image_view->offset; - break; - } + if (layout->stage[stage].image_count > 0) { + VkResult result = + anv_cmd_buffer_ensure_push_constant_field(cmd_buffer, stage, images); + if (result != VK_SUCCESS) + return result; + + cmd_buffer->state.push_constants_dirty |= 1 << stage; + } + + uint32_t image = 0; + for (uint32_t s = 0; s < layout->stage[stage].surface_count; s++) { + struct anv_pipeline_binding *binding = + &layout->stage[stage].surface_to_descriptor[s]; + struct anv_descriptor_set *set = + cmd_buffer->state.descriptors[binding->set]; + struct anv_descriptor *desc = &set->descriptors[binding->offset]; + + struct anv_state surface_state; + struct anv_bo *bo; + uint32_t bo_offset; + + switch (desc->type) { + case VK_DESCRIPTOR_TYPE_SAMPLER: + /* Nothing for us to do here */ + continue; + + case VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER: + case VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE: + case VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT: + surface_state = desc->image_view->nonrt_surface_state; + assert(surface_state.alloc_size); + bo = desc->image_view->bo; + bo_offset = desc->image_view->offset; + break; - bt_map[start + b] = surface_state->offset + state_offset; - add_surface_state_reloc(cmd_buffer, *surface_state, bo, bo_offset); + case VK_DESCRIPTOR_TYPE_STORAGE_IMAGE: { + surface_state = desc->image_view->storage_surface_state; + assert(surface_state.alloc_size); + bo = desc->image_view->bo; + bo_offset = desc->image_view->offset; + + struct brw_image_param *image_param = + &cmd_buffer->state.push_constants[stage]->images[image++]; + + anv_image_view_fill_image_param(cmd_buffer->device, desc->image_view, + image_param); + image_param->surface_idx = bias + s; + break; } + + case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER: + case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER: + case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC: + case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC: + case VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER: + surface_state = desc->buffer_view->surface_state; + assert(surface_state.alloc_size); + bo = desc->buffer_view->bo; + bo_offset = desc->buffer_view->offset; + break; + + case VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER: + surface_state = desc->buffer_view->storage_surface_state; + assert(surface_state.alloc_size); + bo = desc->buffer_view->bo; + bo_offset = desc->buffer_view->offset; + + struct brw_image_param *image_param = + &cmd_buffer->state.push_constants[stage]->images[image++]; + + anv_buffer_view_fill_image_param(cmd_buffer->device, desc->buffer_view, + image_param); + image_param->surface_idx = bias + s; + break; + + default: + assert(!"Invalid descriptor type"); + continue; + } + + bt_map[bias + s] = surface_state.offset + state_offset; + add_surface_state_reloc(cmd_buffer, surface_state, bo, bo_offset); } + assert(image == layout->stage[stage].image_count); + + out: + if (!cmd_buffer->device->info.has_llc) + anv_state_clflush(*bt_state); return VK_SUCCESS; } VkResult anv_cmd_buffer_emit_samplers(struct anv_cmd_buffer *cmd_buffer, - VkShaderStage stage, struct anv_state *state) + gl_shader_stage stage, struct anv_state *state) { struct anv_pipeline_layout *layout; uint32_t sampler_count; - if (stage == VK_SHADER_STAGE_COMPUTE) + if (stage == MESA_SHADER_COMPUTE) layout = cmd_buffer->state.compute_pipeline->layout; else layout = cmd_buffer->state.pipeline->layout; sampler_count = layout ? layout->stage[stage].sampler_count : 0; - if (sampler_count == 0) + if (sampler_count == 0) { + *state = (struct anv_state) { 0, }; return VK_SUCCESS; + } uint32_t size = sampler_count * 16; *state = anv_cmd_buffer_alloc_dynamic_state(cmd_buffer, size, 32); @@ -656,127 +888,48 @@ anv_cmd_buffer_emit_samplers(struct anv_cmd_buffer *cmd_buffer, if (state->map == NULL) return VK_ERROR_OUT_OF_DEVICE_MEMORY; - for (uint32_t set = 0; set < layout->num_sets; set++) { - struct anv_descriptor_set_binding *d = &cmd_buffer->state.descriptors[set]; - struct anv_descriptor_set_layout *set_layout = layout->set[set].layout; - struct anv_descriptor_slot *sampler_slots = - set_layout->stage[stage].sampler_start; - - uint32_t start = layout->set[set].stage[stage].sampler_start; - - for (uint32_t b = 0; b < set_layout->stage[stage].sampler_count; b++) { - struct anv_descriptor *desc = - &d->set->descriptors[sampler_slots[b].index]; - - if (desc->type != ANV_DESCRIPTOR_TYPE_SAMPLER) - continue; + for (uint32_t s = 0; s < layout->stage[stage].sampler_count; s++) { + struct anv_pipeline_binding *binding = + &layout->stage[stage].sampler_to_descriptor[s]; + struct anv_descriptor_set *set = + cmd_buffer->state.descriptors[binding->set]; + struct anv_descriptor *desc = &set->descriptors[binding->offset]; - struct anv_sampler *sampler = desc->sampler; - - memcpy(state->map + (start + b) * 16, - sampler->state, sizeof(sampler->state)); - } - } + if (desc->type != VK_DESCRIPTOR_TYPE_SAMPLER && + desc->type != VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER) + continue; - return VK_SUCCESS; -} + struct anv_sampler *sampler = desc->sampler; -static VkResult -flush_descriptor_set(struct anv_cmd_buffer *cmd_buffer, VkShaderStage stage) -{ - struct anv_state surfaces = { 0, }, samplers = { 0, }; - VkResult result; + /* This can happen if we have an unfilled slot since TYPE_SAMPLER + * happens to be zero. + */ + if (sampler == NULL) + continue; - result = anv_cmd_buffer_emit_samplers(cmd_buffer, stage, &samplers); - if (result != VK_SUCCESS) - return result; - result = anv_cmd_buffer_emit_binding_table(cmd_buffer, stage, &surfaces); - if (result != VK_SUCCESS) - return result; - - static const uint32_t sampler_state_opcodes[] = { - [VK_SHADER_STAGE_VERTEX] = 43, - [VK_SHADER_STAGE_TESS_CONTROL] = 44, /* HS */ - [VK_SHADER_STAGE_TESS_EVALUATION] = 45, /* DS */ - [VK_SHADER_STAGE_GEOMETRY] = 46, - [VK_SHADER_STAGE_FRAGMENT] = 47, - [VK_SHADER_STAGE_COMPUTE] = 0, - }; - - static const uint32_t binding_table_opcodes[] = { - [VK_SHADER_STAGE_VERTEX] = 38, - [VK_SHADER_STAGE_TESS_CONTROL] = 39, - [VK_SHADER_STAGE_TESS_EVALUATION] = 40, - [VK_SHADER_STAGE_GEOMETRY] = 41, - [VK_SHADER_STAGE_FRAGMENT] = 42, - [VK_SHADER_STAGE_COMPUTE] = 0, - }; - - if (samplers.alloc_size > 0) { - anv_batch_emit(&cmd_buffer->batch, - GEN7_3DSTATE_SAMPLER_STATE_POINTERS_VS, - ._3DCommandSubOpcode = sampler_state_opcodes[stage], - .PointertoVSSamplerState = samplers.offset); + memcpy(state->map + (s * 16), + sampler->state, sizeof(sampler->state)); } - if (surfaces.alloc_size > 0) { - anv_batch_emit(&cmd_buffer->batch, - GEN7_3DSTATE_BINDING_TABLE_POINTERS_VS, - ._3DCommandSubOpcode = binding_table_opcodes[stage], - .PointertoVSBindingTable = surfaces.offset); - } + if (!cmd_buffer->device->info.has_llc) + anv_state_clflush(*state); return VK_SUCCESS; } -void -anv_flush_descriptor_sets(struct anv_cmd_buffer *cmd_buffer) -{ - VkShaderStage s; - VkShaderStageFlags dirty = cmd_buffer->state.descriptors_dirty & - cmd_buffer->state.pipeline->active_stages; - - VkResult result = VK_SUCCESS; - for_each_bit(s, dirty) { - result = flush_descriptor_set(cmd_buffer, s); - if (result != VK_SUCCESS) - break; - } - - if (result != VK_SUCCESS) { - assert(result == VK_ERROR_OUT_OF_DEVICE_MEMORY); - - result = anv_cmd_buffer_new_binding_table_block(cmd_buffer); - assert(result == VK_SUCCESS); - - /* Re-emit state base addresses so we get the new surface state base - * address before we start emitting binding tables etc. - */ - anv_cmd_buffer_emit_state_base_address(cmd_buffer); - - /* Re-emit all active binding tables */ - for_each_bit(s, cmd_buffer->state.pipeline->active_stages) { - result = flush_descriptor_set(cmd_buffer, s); - - /* It had better succeed this time */ - assert(result == VK_SUCCESS); - } - } - - cmd_buffer->state.descriptors_dirty &= ~cmd_buffer->state.pipeline->active_stages; -} - struct anv_state anv_cmd_buffer_emit_dynamic(struct anv_cmd_buffer *cmd_buffer, - uint32_t *a, uint32_t dwords, uint32_t alignment) + const void *data, uint32_t size, uint32_t alignment) { struct anv_state state; - state = anv_cmd_buffer_alloc_dynamic_state(cmd_buffer, - dwords * 4, alignment); - memcpy(state.map, a, dwords * 4); + state = anv_cmd_buffer_alloc_dynamic_state(cmd_buffer, size, alignment); + memcpy(state.map, data, size); + + if (!cmd_buffer->device->info.has_llc) + anv_state_clflush(state); - VG(VALGRIND_CHECK_MEM_IS_DEFINED(state.map, dwords * 4)); + VG(VALGRIND_CHECK_MEM_IS_DEFINED(state.map, size)); return state; } @@ -795,261 +948,167 @@ anv_cmd_buffer_merge_dynamic(struct anv_cmd_buffer *cmd_buffer, for (uint32_t i = 0; i < dwords; i++) p[i] = a[i] | b[i]; + if (!cmd_buffer->device->info.has_llc) + anv_state_clflush(state); + VG(VALGRIND_CHECK_MEM_IS_DEFINED(p, dwords * 4)); return state; } +/** + * @brief Setup the command buffer for recording commands inside the given + * subpass. + * + * This does not record all commands needed for starting the subpass. + * Starting the subpass may require additional commands. + * + * Note that vkCmdBeginRenderPass, vkCmdNextSubpass, and vkBeginCommandBuffer + * with VK_COMMAND_BUFFER_USAGE_RENDER_PASS_CONTINUE_BIT, all setup the + * command buffer for recording commands for some subpass. But only the first + * two, vkCmdBeginRenderPass and vkCmdNextSubpass, can start a subpass. + */ void -anv_cmd_buffer_begin_subpass(struct anv_cmd_buffer *cmd_buffer, - struct anv_subpass *subpass) +anv_cmd_buffer_set_subpass(struct anv_cmd_buffer *cmd_buffer, + struct anv_subpass *subpass) { switch (cmd_buffer->device->info.gen) { case 7: - gen7_cmd_buffer_begin_subpass(cmd_buffer, subpass); + gen7_cmd_buffer_set_subpass(cmd_buffer, subpass); break; case 8: - gen8_cmd_buffer_begin_subpass(cmd_buffer, subpass); + gen8_cmd_buffer_set_subpass(cmd_buffer, subpass); + break; + case 9: + gen9_cmd_buffer_set_subpass(cmd_buffer, subpass); break; default: unreachable("unsupported gen\n"); } } -static void -emit_viewport_state(struct anv_cmd_buffer *cmd_buffer, - uint32_t count, const VkViewport *viewports) -{ - struct anv_state sf_clip_state = - anv_cmd_buffer_alloc_dynamic_state(cmd_buffer, count * 64, 64); - struct anv_state cc_state = - anv_cmd_buffer_alloc_dynamic_state(cmd_buffer, count * 8, 32); - - for (uint32_t i = 0; i < count; i++) { - const VkViewport *vp = &viewports[i]; - - /* The gen7 state struct has just the matrix and guardband fields, the - * gen8 struct adds the min/max viewport fields. */ - struct GEN8_SF_CLIP_VIEWPORT sf_clip_viewport = { - .ViewportMatrixElementm00 = vp->width / 2, - .ViewportMatrixElementm11 = vp->height / 2, - .ViewportMatrixElementm22 = (vp->maxDepth - vp->minDepth) / 2, - .ViewportMatrixElementm30 = vp->originX + vp->width / 2, - .ViewportMatrixElementm31 = vp->originY + vp->height / 2, - .ViewportMatrixElementm32 = (vp->maxDepth + vp->minDepth) / 2, - .XMinClipGuardband = -1.0f, - .XMaxClipGuardband = 1.0f, - .YMinClipGuardband = -1.0f, - .YMaxClipGuardband = 1.0f, - .XMinViewPort = vp->originX, - .XMaxViewPort = vp->originX + vp->width - 1, - .YMinViewPort = vp->originY, - .YMaxViewPort = vp->originY + vp->height - 1, - }; - - struct GEN7_CC_VIEWPORT cc_viewport = { - .MinimumDepth = vp->minDepth, - .MaximumDepth = vp->maxDepth - }; - - GEN8_SF_CLIP_VIEWPORT_pack(NULL, sf_clip_state.map + i * 64, - &sf_clip_viewport); - GEN7_CC_VIEWPORT_pack(NULL, cc_state.map + i * 32, &cc_viewport); - } - - anv_batch_emit(&cmd_buffer->batch, - GEN8_3DSTATE_VIEWPORT_STATE_POINTERS_CC, - .CCViewportPointer = cc_state.offset); - anv_batch_emit(&cmd_buffer->batch, - GEN8_3DSTATE_VIEWPORT_STATE_POINTERS_SF_CLIP, - .SFClipViewportPointer = sf_clip_state.offset); -} - -void -anv_cmd_buffer_emit_viewport(struct anv_cmd_buffer *cmd_buffer) -{ - if (cmd_buffer->state.dynamic.viewport.count > 0) { - emit_viewport_state(cmd_buffer, cmd_buffer->state.dynamic.viewport.count, - cmd_buffer->state.dynamic.viewport.viewports); - } else { - /* If viewport count is 0, this is taken to mean "use the default" */ - emit_viewport_state(cmd_buffer, 1, - &(VkViewport) { - .originX = 0.0f, - .originY = 0.0f, - .width = cmd_buffer->state.framebuffer->width, - .height = cmd_buffer->state.framebuffer->height, - .minDepth = 0.0f, - .maxDepth = 1.0f, - }); - } -} - -static inline int64_t -clamp_int64(int64_t x, int64_t min, int64_t max) +struct anv_state +anv_cmd_buffer_push_constants(struct anv_cmd_buffer *cmd_buffer, + gl_shader_stage stage) { - if (x < min) - return min; - else if (x < max) - return x; - else - return max; -} + struct anv_push_constants *data = + cmd_buffer->state.push_constants[stage]; + struct brw_stage_prog_data *prog_data = + cmd_buffer->state.pipeline->prog_data[stage]; -static void -emit_scissor_state(struct anv_cmd_buffer *cmd_buffer, - uint32_t count, const VkRect2D *scissors) -{ - struct anv_state scissor_state = - anv_cmd_buffer_alloc_dynamic_state(cmd_buffer, count * 32, 32); - - for (uint32_t i = 0; i < count; i++) { - const VkRect2D *s = &scissors[i]; - - /* Since xmax and ymax are inclusive, we have to have xmax < xmin or - * ymax < ymin for empty clips. In case clip x, y, width height are all - * 0, the clamps below produce 0 for xmin, ymin, xmax, ymax, which isn't - * what we want. Just special case empty clips and produce a canonical - * empty clip. */ - static const struct GEN7_SCISSOR_RECT empty_scissor = { - .ScissorRectangleYMin = 1, - .ScissorRectangleXMin = 1, - .ScissorRectangleYMax = 0, - .ScissorRectangleXMax = 0 - }; - - const int max = 0xffff; - struct GEN7_SCISSOR_RECT scissor = { - /* Do this math using int64_t so overflow gets clamped correctly. */ - .ScissorRectangleYMin = clamp_int64(s->offset.y, 0, max), - .ScissorRectangleXMin = clamp_int64(s->offset.x, 0, max), - .ScissorRectangleYMax = clamp_int64((uint64_t) s->offset.y + s->extent.height - 1, 0, max), - .ScissorRectangleXMax = clamp_int64((uint64_t) s->offset.x + s->extent.width - 1, 0, max) - }; - - if (s->extent.width <= 0 || s->extent.height <= 0) { - GEN7_SCISSOR_RECT_pack(NULL, scissor_state.map + i * 32, - &empty_scissor); - } else { - GEN7_SCISSOR_RECT_pack(NULL, scissor_state.map + i * 32, &scissor); - } - } + /* If we don't actually have any push constants, bail. */ + if (data == NULL || prog_data->nr_params == 0) + return (struct anv_state) { .offset = 0 }; - anv_batch_emit(&cmd_buffer->batch, GEN8_3DSTATE_SCISSOR_STATE_POINTERS, - .ScissorRectPointer = scissor_state.offset); -} + struct anv_state state = + anv_cmd_buffer_alloc_dynamic_state(cmd_buffer, + prog_data->nr_params * sizeof(float), + 32 /* bottom 5 bits MBZ */); -void -anv_cmd_buffer_emit_scissor(struct anv_cmd_buffer *cmd_buffer) -{ - if (cmd_buffer->state.dynamic.scissor.count > 0) { - emit_scissor_state(cmd_buffer, cmd_buffer->state.dynamic.scissor.count, - cmd_buffer->state.dynamic.scissor.scissors); - } else { - /* Emit a default scissor based on the currently bound framebuffer */ - emit_scissor_state(cmd_buffer, 1, - &(VkRect2D) { - .offset = { .x = 0, .y = 0, }, - .extent = { - .width = cmd_buffer->state.framebuffer->width, - .height = cmd_buffer->state.framebuffer->height, - }, - }); + /* Walk through the param array and fill the buffer with data */ + uint32_t *u32_map = state.map; + for (unsigned i = 0; i < prog_data->nr_params; i++) { + uint32_t offset = (uintptr_t)prog_data->param[i]; + u32_map[i] = *(uint32_t *)((uint8_t *)data + offset); } -} -void anv_CmdSetEvent( - VkCmdBuffer cmdBuffer, - VkEvent event, - VkPipelineStageFlags stageMask) -{ - stub(); -} - -void anv_CmdResetEvent( - VkCmdBuffer cmdBuffer, - VkEvent event, - VkPipelineStageFlags stageMask) -{ - stub(); -} + if (!cmd_buffer->device->info.has_llc) + anv_state_clflush(state); -void anv_CmdWaitEvents( - VkCmdBuffer cmdBuffer, - uint32_t eventCount, - const VkEvent* pEvents, - VkPipelineStageFlags srcStageMask, - VkPipelineStageFlags destStageMask, - uint32_t memBarrierCount, - const void* const* ppMemBarriers) -{ - stub(); + return state; } struct anv_state -anv_cmd_buffer_push_constants(struct anv_cmd_buffer *cmd_buffer, - VkShaderStage stage) +anv_cmd_buffer_cs_push_constants(struct anv_cmd_buffer *cmd_buffer) { struct anv_push_constants *data = - cmd_buffer->state.push_constants[stage]; - struct brw_stage_prog_data *prog_data = - cmd_buffer->state.pipeline->prog_data[stage]; + cmd_buffer->state.push_constants[MESA_SHADER_COMPUTE]; + struct anv_pipeline *pipeline = cmd_buffer->state.compute_pipeline; + const struct brw_cs_prog_data *cs_prog_data = &pipeline->cs_prog_data; + const struct brw_stage_prog_data *prog_data = &cs_prog_data->base; + + const unsigned local_id_dwords = cs_prog_data->local_invocation_id_regs * 8; + const unsigned push_constant_data_size = + (local_id_dwords + prog_data->nr_params) * 4; + const unsigned reg_aligned_constant_size = ALIGN(push_constant_data_size, 32); + const unsigned param_aligned_count = + reg_aligned_constant_size / sizeof(uint32_t); /* If we don't actually have any push constants, bail. */ - if (data == NULL || prog_data->nr_params == 0) + if (reg_aligned_constant_size == 0) return (struct anv_state) { .offset = 0 }; + const unsigned threads = pipeline->cs_thread_width_max; + const unsigned total_push_constants_size = + reg_aligned_constant_size * threads; + const unsigned push_constant_alignment = + cmd_buffer->device->info.gen < 8 ? 32 : 64; + const unsigned aligned_total_push_constants_size = + ALIGN(total_push_constants_size, push_constant_alignment); struct anv_state state = anv_cmd_buffer_alloc_dynamic_state(cmd_buffer, - prog_data->nr_params * sizeof(float), - 32 /* bottom 5 bits MBZ */); + aligned_total_push_constants_size, + push_constant_alignment); /* Walk through the param array and fill the buffer with data */ uint32_t *u32_map = state.map; + + brw_cs_fill_local_id_payload(cs_prog_data, u32_map, threads, + reg_aligned_constant_size); + + /* Setup uniform data for the first thread */ for (unsigned i = 0; i < prog_data->nr_params; i++) { uint32_t offset = (uintptr_t)prog_data->param[i]; - u32_map[i] = *(uint32_t *)((uint8_t *)data + offset); + u32_map[local_id_dwords + i] = *(uint32_t *)((uint8_t *)data + offset); + } + + /* Copy uniform data from the first thread to every other thread */ + const size_t uniform_data_size = prog_data->nr_params * sizeof(uint32_t); + for (unsigned t = 1; t < threads; t++) { + memcpy(&u32_map[t * param_aligned_count + local_id_dwords], + &u32_map[local_id_dwords], + uniform_data_size); } + if (!cmd_buffer->device->info.has_llc) + anv_state_clflush(state); + return state; } void anv_CmdPushConstants( - VkCmdBuffer cmdBuffer, + VkCommandBuffer commandBuffer, VkPipelineLayout layout, VkShaderStageFlags stageFlags, - uint32_t start, - uint32_t length, - const void* values) + uint32_t offset, + uint32_t size, + const void* pValues) { - ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, cmdBuffer); - VkShaderStage stage; + ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer); - for_each_bit(stage, stageFlags) { + anv_foreach_stage(stage, stageFlags) { anv_cmd_buffer_ensure_push_constant_field(cmd_buffer, stage, client_data); - memcpy(cmd_buffer->state.push_constants[stage]->client_data + start, - values, length); + memcpy(cmd_buffer->state.push_constants[stage]->client_data + offset, + pValues, size); } cmd_buffer->state.push_constants_dirty |= stageFlags; } void anv_CmdExecuteCommands( - VkCmdBuffer cmdBuffer, - uint32_t cmdBuffersCount, - const VkCmdBuffer* pCmdBuffers) + VkCommandBuffer commandBuffer, + uint32_t commandBufferCount, + const VkCommandBuffer* pCmdBuffers) { - ANV_FROM_HANDLE(anv_cmd_buffer, primary, cmdBuffer); - - assert(primary->level == VK_CMD_BUFFER_LEVEL_PRIMARY); + ANV_FROM_HANDLE(anv_cmd_buffer, primary, commandBuffer); - anv_assert(primary->state.subpass == &primary->state.pass->subpasses[0]); + assert(primary->level == VK_COMMAND_BUFFER_LEVEL_PRIMARY); - for (uint32_t i = 0; i < cmdBuffersCount; i++) { + for (uint32_t i = 0; i < commandBufferCount; i++) { ANV_FROM_HANDLE(anv_cmd_buffer, secondary, pCmdBuffers[i]); - assert(secondary->level == VK_CMD_BUFFER_LEVEL_SECONDARY); + assert(secondary->level == VK_COMMAND_BUFFER_LEVEL_SECONDARY); anv_cmd_buffer_add_secondary(primary, secondary); } @@ -1057,17 +1116,23 @@ void anv_CmdExecuteCommands( VkResult anv_CreateCommandPool( VkDevice _device, - const VkCmdPoolCreateInfo* pCreateInfo, - VkCmdPool* pCmdPool) + const VkCommandPoolCreateInfo* pCreateInfo, + const VkAllocationCallbacks* pAllocator, + VkCommandPool* pCmdPool) { ANV_FROM_HANDLE(anv_device, device, _device); struct anv_cmd_pool *pool; - pool = anv_device_alloc(device, sizeof(*pool), 8, - VK_SYSTEM_ALLOC_TYPE_API_OBJECT); + pool = anv_alloc2(&device->alloc, pAllocator, sizeof(*pool), 8, + VK_SYSTEM_ALLOCATION_SCOPE_OBJECT); if (pool == NULL) return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY); + if (pAllocator) + pool->alloc = *pAllocator; + else + pool->alloc = device->alloc; + list_inithead(&pool->cmd_buffers); *pCmdPool = anv_cmd_pool_to_handle(pool); @@ -1077,26 +1142,35 @@ VkResult anv_CreateCommandPool( void anv_DestroyCommandPool( VkDevice _device, - VkCmdPool cmdPool) + VkCommandPool commandPool, + const VkAllocationCallbacks* pAllocator) { ANV_FROM_HANDLE(anv_device, device, _device); - ANV_FROM_HANDLE(anv_cmd_pool, pool, cmdPool); + ANV_FROM_HANDLE(anv_cmd_pool, pool, commandPool); - anv_ResetCommandPool(_device, cmdPool, 0); + anv_ResetCommandPool(_device, commandPool, 0); - anv_device_free(device, pool); + anv_free2(&device->alloc, pAllocator, pool); } VkResult anv_ResetCommandPool( VkDevice device, - VkCmdPool cmdPool, - VkCmdPoolResetFlags flags) + VkCommandPool commandPool, + VkCommandPoolResetFlags flags) { - ANV_FROM_HANDLE(anv_cmd_pool, pool, cmdPool); - + ANV_FROM_HANDLE(anv_cmd_pool, pool, commandPool); + + /* FIXME: vkResetCommandPool must not destroy its command buffers. The + * Vulkan 1.0 spec requires that it only reset them: + * + * Resetting a command pool recycles all of the resources from all of + * the command buffers allocated from the command pool back to the + * command pool. All command buffers that have been allocated from the + * command pool are put in the initial state. + */ list_for_each_entry_safe(struct anv_cmd_buffer, cmd_buffer, &pool->cmd_buffers, pool_link) { - anv_DestroyCommandBuffer(device, anv_cmd_buffer_to_handle(cmd_buffer)); + anv_cmd_buffer_destroy(cmd_buffer); } return VK_SUCCESS; @@ -1117,7 +1191,8 @@ anv_cmd_buffer_get_depth_stencil_view(const struct anv_cmd_buffer *cmd_buffer) const struct anv_image_view *iview = fb->attachments[subpass->depth_stencil_attachment]; - assert(anv_format_is_depth_or_stencil(iview->format)); + assert(iview->aspect_mask & (VK_IMAGE_ASPECT_DEPTH_BIT | + VK_IMAGE_ASPECT_STENCIL_BIT)); return iview; }