diff options
Diffstat (limited to 'source/blender/gpu/metal')
| -rw-r--r-- | source/blender/gpu/metal/mtl_backend.mm | 3 | ||||
| -rw-r--r-- | source/blender/gpu/metal/mtl_batch.hh | 115 | ||||
| -rw-r--r-- | source/blender/gpu/metal/mtl_batch.mm | 998 | ||||
| -rw-r--r-- | source/blender/gpu/metal/mtl_context.mm | 10 | ||||
| -rw-r--r-- | source/blender/gpu/metal/mtl_drawlist.hh | 58 | ||||
| -rw-r--r-- | source/blender/gpu/metal/mtl_drawlist.mm | 284 | ||||
| -rw-r--r-- | source/blender/gpu/metal/mtl_immediate.mm | 3 | ||||
| -rw-r--r-- | source/blender/gpu/metal/mtl_pso_descriptor_state.hh | 13 | ||||
| -rw-r--r-- | source/blender/gpu/metal/mtl_shader_generator.hh | 2 | ||||
| -rw-r--r-- | source/blender/gpu/metal/mtl_shader_interface.mm | 10 | ||||
| -rw-r--r-- | source/blender/gpu/metal/mtl_texture.hh | 4 | ||||
| -rw-r--r-- | source/blender/gpu/metal/mtl_texture.mm | 97 | ||||
| -rw-r--r-- | source/blender/gpu/metal/mtl_texture_util.mm | 2 |
13 files changed, 1533 insertions, 66 deletions
diff --git a/source/blender/gpu/metal/mtl_backend.mm b/source/blender/gpu/metal/mtl_backend.mm index 2ca1fd3f3d0..240951c1ebd 100644 --- a/source/blender/gpu/metal/mtl_backend.mm +++ b/source/blender/gpu/metal/mtl_backend.mm @@ -47,13 +47,11 @@ Context *MTLBackend::context_alloc(void *ghost_window, void *ghost_context) Batch *MTLBackend::batch_alloc() { - /* TODO(Metal): Full MTLBatch implementation. */ return new MTLBatch(); }; DrawList *MTLBackend::drawlist_alloc(int list_length) { - /* TODO(Metal): Full MTLDrawList implementation. */ return new MTLDrawList(list_length); }; @@ -420,6 +418,7 @@ void MTLBackend::capabilities_init(MTLContext *ctx) GCaps.depth_blitting_workaround = false; GCaps.use_main_context_workaround = false; GCaps.broken_amd_driver = false; + GCaps.clear_viewport_workaround = true; /* Metal related workarounds. */ /* Minimum per-vertex stride is 4 bytes in Metal. diff --git a/source/blender/gpu/metal/mtl_batch.hh b/source/blender/gpu/metal/mtl_batch.hh index 236367bf5a4..9e179e662b5 100644 --- a/source/blender/gpu/metal/mtl_batch.hh +++ b/source/blender/gpu/metal/mtl_batch.hh @@ -10,31 +10,126 @@ #pragma once #include "MEM_guardedalloc.h" - #include "gpu_batch_private.hh" +#include "mtl_index_buffer.hh" +#include "mtl_primitive.hh" +#include "mtl_shader.hh" +#include "mtl_vertex_buffer.hh" + +namespace blender::gpu { + +class MTLContext; +class MTLShaderInterface; + +#define GPU_VAO_STATIC_LEN 64 -namespace blender { -namespace gpu { +struct VertexBufferID { + uint32_t id : 16; + uint32_t is_instance : 15; + uint32_t used : 1; +}; -/* Pass-through MTLBatch. TODO(Metal): Implement. */ class MTLBatch : public Batch { + + /* Vertex Bind-state Caching for a given shader interface used with the Batch. */ + struct VertexDescriptorShaderInterfacePair { + MTLVertexDescriptor vertex_descriptor{}; + const ShaderInterface *interface = nullptr; + uint16_t attr_mask{}; + int num_buffers{}; + VertexBufferID bufferIds[GPU_BATCH_VBO_MAX_LEN] = {}; + /* Cache life index compares a cache entry with the active MTLBatch state. + * This is initially set to the cache life index of MTLBatch. If the batch has been modified, + * this index is incremented to cheaply invalidate existing cache entries. */ + uint32_t cache_life_index = 0; + }; + + class MTLVertexDescriptorCache { + + private: + MTLBatch *batch_; + + VertexDescriptorShaderInterfacePair cache_[GPU_VAO_STATIC_LEN] = {}; + MTLContext *cache_context_ = nullptr; + uint32_t cache_life_index_ = 0; + + public: + MTLVertexDescriptorCache(MTLBatch *batch) : batch_(batch){}; + VertexDescriptorShaderInterfacePair *find(const ShaderInterface *interface); + bool insert(VertexDescriptorShaderInterfacePair &data); + + private: + void vertex_descriptor_cache_init(MTLContext *ctx); + void vertex_descriptor_cache_clear(); + void vertex_descriptor_cache_ensure(); + }; + + private: + MTLShader *active_shader_ = nullptr; + bool shader_in_use_ = false; + MTLVertexDescriptorCache vao_cache = {this}; + + /* Topology emulation. */ + gpu::MTLBuffer *emulated_topology_buffer_ = nullptr; + GPUPrimType emulated_topology_type_; + uint32_t topology_buffer_input_v_count_ = 0; + uint32_t topology_buffer_output_v_count_ = 0; + public: - void draw(int v_first, int v_count, int i_first, int i_count) override - { - } + MTLBatch(){}; + ~MTLBatch(){}; + void draw(int v_first, int v_count, int i_first, int i_count) override; void draw_indirect(GPUStorageBuf *indirect_buf, intptr_t offset) override { + /* TODO(Metal): Support indirect draw commands. */ } - void multi_draw_indirect(GPUStorageBuf *indirect_buf, int count, intptr_t offset, intptr_t stride) override { + /* TODO(Metal): Support indirect draw commands. */ + } + + /* Returns an initialized RenderComandEncoder for drawing if all is good. + * Otherwise, nil. */ + id<MTLRenderCommandEncoder> bind(uint v_first, uint v_count, uint i_first, uint i_count); + void unbind(); + + /* Convenience getters. */ + MTLIndexBuf *elem_() const + { + return static_cast<MTLIndexBuf *>(unwrap(elem)); + } + MTLVertBuf *verts_(const int index) const + { + return static_cast<MTLVertBuf *>(unwrap(verts[index])); } + MTLVertBuf *inst_(const int index) const + { + return static_cast<MTLVertBuf *>(unwrap(inst[index])); + } + MTLShader *active_shader_get() const + { + return active_shader_; + } + + private: + void shader_bind(); + void draw_advanced(int v_first, int v_count, int i_first, int i_count); + int prepare_vertex_binding(MTLVertBuf *verts, + MTLRenderPipelineStateDescriptor &desc, + const MTLShaderInterface *interface, + uint16_t &attr_mask, + bool instanced); + + id<MTLBuffer> get_emulated_toplogy_buffer(GPUPrimType &in_out_prim_type, uint32_t &v_count); + + void prepare_vertex_descriptor_and_bindings( + MTLVertBuf **buffers, int &num_buffers, int v_first, int v_count, int i_first, int i_count); + MEM_CXX_CLASS_ALLOC_FUNCS("MTLBatch"); }; -} // namespace gpu -} // namespace blender +} // namespace blender::gpu diff --git a/source/blender/gpu/metal/mtl_batch.mm b/source/blender/gpu/metal/mtl_batch.mm new file mode 100644 index 00000000000..988fb9b793b --- /dev/null +++ b/source/blender/gpu/metal/mtl_batch.mm @@ -0,0 +1,998 @@ +/* SPDX-License-Identifier: GPL-2.0-or-later */ + +/** \file + * \ingroup gpu + * + * Metal implementation of GPUBatch. + */ + +#include "BLI_assert.h" +#include "BLI_span.hh" + +#include "BKE_global.h" + +#include "GPU_common.h" +#include "gpu_batch_private.hh" +#include "gpu_shader_private.hh" + +#include "mtl_batch.hh" +#include "mtl_context.hh" +#include "mtl_debug.hh" +#include "mtl_index_buffer.hh" +#include "mtl_shader.hh" +#include "mtl_vertex_buffer.hh" + +#include <string> + +namespace blender::gpu { + +/* -------------------------------------------------------------------- */ +/** \name Creation & Deletion + * \{ */ +void MTLBatch::draw(int v_first, int v_count, int i_first, int i_count) +{ + if (this->flag & GPU_BATCH_INVALID) { + this->shader_in_use_ = false; + } + this->draw_advanced(v_first, v_count, i_first, i_count); +} + +void MTLBatch::shader_bind() +{ + if (active_shader_ && active_shader_->is_valid()) { + active_shader_->bind(); + shader_in_use_ = true; + } +} + +void MTLBatch::MTLVertexDescriptorCache::vertex_descriptor_cache_init(MTLContext *ctx) +{ + BLI_assert(ctx != nullptr); + this->vertex_descriptor_cache_clear(); + cache_context_ = ctx; +} + +void MTLBatch::MTLVertexDescriptorCache::vertex_descriptor_cache_clear() +{ + cache_life_index_++; + cache_context_ = nullptr; +} + +void MTLBatch::MTLVertexDescriptorCache::vertex_descriptor_cache_ensure() +{ + if (this->cache_context_ != nullptr) { + + /* Invalidate vertex descriptor bindings cache if batch has changed. */ + if (batch_->flag & GPU_BATCH_DIRTY) { + batch_->flag &= ~GPU_BATCH_DIRTY; + this->vertex_descriptor_cache_clear(); + } + } + + /* Initialize cache if not ready. */ + if (cache_context_ == nullptr) { + this->vertex_descriptor_cache_init(MTLContext::get()); + } +} + +MTLBatch::VertexDescriptorShaderInterfacePair *MTLBatch::MTLVertexDescriptorCache::find( + const ShaderInterface *interface) +{ + this->vertex_descriptor_cache_ensure(); + for (int i = 0; i < GPU_VAO_STATIC_LEN; ++i) { + if (cache_[i].interface == interface && cache_[i].cache_life_index == cache_life_index_) { + return &cache_[i]; + } + } + return nullptr; +} + +bool MTLBatch::MTLVertexDescriptorCache::insert( + MTLBatch::VertexDescriptorShaderInterfacePair &data) +{ + vertex_descriptor_cache_ensure(); + for (int i = 0; i < GPU_VAO_STATIC_LEN; ++i) { + if (cache_[i].interface == nullptr || cache_[i].cache_life_index != cache_life_index_) { + cache_[i] = data; + cache_[i].cache_life_index = cache_life_index_; + return true; + } + } + return false; +} + +int MTLBatch::prepare_vertex_binding(MTLVertBuf *verts, + MTLRenderPipelineStateDescriptor &desc, + const MTLShaderInterface *interface, + uint16_t &attr_mask, + bool instanced) +{ + + const GPUVertFormat *format = &verts->format; + /* Whether the current vertex buffer has been added to the buffer layout descriptor. */ + bool buffer_added = false; + /* Per-vertex stride of current vertex buffer. */ + int buffer_stride = format->stride; + /* Buffer binding index of the vertex buffer once added to the buffer layout descriptor. */ + int buffer_index = -1; + int attribute_offset = 0; + + if (!active_shader_->get_uses_ssbo_vertex_fetch()) { + BLI_assert( + buffer_stride >= 4 && + "In Metal, Vertex buffer stride should be 4. SSBO Vertex fetch is not affected by this"); + } + + /* Iterate over GPUVertBuf vertex format and find attributes matching those in the active + * shader's interface. */ + for (uint32_t a_idx = 0; a_idx < format->attr_len; a_idx++) { + const GPUVertAttr *a = &format->attrs[a_idx]; + + if (format->deinterleaved) { + attribute_offset += ((a_idx == 0) ? 0 : format->attrs[a_idx - 1].size) * verts->vertex_len; + buffer_stride = a->size; + } + else { + attribute_offset = a->offset; + } + + /* Find attribute with the matching name. Attributes may have multiple compatible + * name aliases. */ + for (uint32_t n_idx = 0; n_idx < a->name_len; n_idx++) { + const char *name = GPU_vertformat_attr_name_get(format, a, n_idx); + const ShaderInput *input = interface->attr_get(name); + + if (input == nullptr || input->location == -1) { + /* Vertex/instance buffers provided have attribute data for attributes which are not needed + * by this particular shader. This shader only needs binding information for the attributes + * has in the shader interface. */ + MTL_LOG_WARNING( + "MTLBatch: Could not find attribute with name '%s' (defined in active vertex format) " + "in the shader interface for shader '%s'\n", + name, + interface->get_name()); + continue; + } + + /* Fetch metal attribute information. */ + const MTLShaderInputAttribute &mtl_attr = interface->get_attribute(input->location); + BLI_assert(mtl_attr.location >= 0); + /* Verify that the attribute location from the shader interface + * matches the attribute location returned. */ + BLI_assert(mtl_attr.location == input->location); + + /* Check if attribute is already present in the given slot. */ + if ((~attr_mask) & (1 << mtl_attr.location)) { + MTL_LOG_INFO( + " -- [Batch] Skipping attribute with input location %d (As one is already bound)\n", + mtl_attr.location); + } + else { + + /* Update attribute used-slot mask. */ + attr_mask &= ~(1 << mtl_attr.location); + + /* Add buffer layout entry in descriptor if it has not yet been added + * for current vertex buffer. */ + if (!buffer_added) { + buffer_index = desc.vertex_descriptor.num_vert_buffers; + desc.vertex_descriptor.buffer_layouts[buffer_index].step_function = + (instanced) ? MTLVertexStepFunctionPerInstance : MTLVertexStepFunctionPerVertex; + desc.vertex_descriptor.buffer_layouts[buffer_index].step_rate = 1; + desc.vertex_descriptor.buffer_layouts[buffer_index].stride = buffer_stride; + desc.vertex_descriptor.num_vert_buffers++; + buffer_added = true; + + MTL_LOG_INFO(" -- [Batch] Adding source %s buffer (Index: %d, Stride: %d)\n", + (instanced) ? "instance" : "vertex", + buffer_index, + buffer_stride); + } + else { + /* Ensure stride is correct for de-interleaved attributes. */ + desc.vertex_descriptor.buffer_layouts[buffer_index].stride = buffer_stride; + } + + /* Handle Matrix/Array vertex attribute types. + * Metal does not natively support these as attribute types, so we handle these cases + * by stacking together compatible types (e.g. 4xVec4 for Mat4) and combining + * the data in the shader. + * The generated Metal shader will contain a generated input binding, which reads + * in individual attributes and merges them into the desired type after vertex + * assembly. e.g. a Mat4 (Float4x4) will generate 4 Float4 attributes. */ + if (a->comp_len == 16 || a->comp_len == 12 || a->comp_len == 8) { + BLI_assert_msg( + a->comp_len == 16, + "only mat4 attributes currently supported -- Not ready to handle other long " + "component length attributes yet"); + + /* SSBO Vertex Fetch Attribute safety checks. */ + if (active_shader_->get_uses_ssbo_vertex_fetch()) { + /* When using SSBO vertex fetch, we do not need to expose split attributes, + * A matrix can be read directly as a whole block of contiguous data. */ + MTLSSBOAttribute ssbo_attr(mtl_attr.index, + buffer_index, + attribute_offset, + buffer_stride, + GPU_SHADER_ATTR_TYPE_MAT4, + instanced); + active_shader_->ssbo_vertex_fetch_bind_attribute(ssbo_attr); + desc.vertex_descriptor.ssbo_attributes[desc.vertex_descriptor.num_ssbo_attributes] = + ssbo_attr; + desc.vertex_descriptor.num_ssbo_attributes++; + } + else { + + /* Handle Mat4 attributes. */ + if (a->comp_len == 16) { + /* Debug safety checks. */ + BLI_assert_msg(mtl_attr.matrix_element_count == 4, + "mat4 type expected but there are fewer components"); + BLI_assert_msg(mtl_attr.size == 16, "Expecting subtype 'vec4' with 16 bytes"); + BLI_assert_msg( + mtl_attr.format == MTLVertexFormatFloat4, + "Per-attribute vertex format MUST be float4 for an input type of 'mat4'"); + + /* We have found the 'ROOT' attribute. A mat4 contains 4 consecutive float4 attribute + * locations we must map to. */ + for (int i = 0; i < a->comp_len / 4; i++) { + desc.vertex_descriptor.attributes[mtl_attr.location + i].format = + MTLVertexFormatFloat4; + /* Data is consecutive in the buffer for the whole matrix, each float4 will shift + * the offset by 16 bytes. */ + desc.vertex_descriptor.attributes[mtl_attr.location + i].offset = + attribute_offset + i * 16; + /* All source data for a matrix is in the same singular buffer. */ + desc.vertex_descriptor.attributes[mtl_attr.location + i].buffer_index = + buffer_index; + + /* Update total attribute account. */ + desc.vertex_descriptor.num_attributes = max_ii( + mtl_attr.location + i + 1, desc.vertex_descriptor.num_attributes); + MTL_LOG_INFO("-- Sub-Attrib Location: %d, offset: %d, buffer index: %d\n", + mtl_attr.location + i, + attribute_offset + i * 16, + buffer_index); + } + MTL_LOG_INFO( + "Float4x4 attribute type added for '%s' at attribute locations: %d to %d\n", + name, + mtl_attr.location, + mtl_attr.location + 3); + } + + /* Ensure we are not exceeding the attribute limit. */ + BLI_assert(desc.vertex_descriptor.num_attributes <= MTL_MAX_VERTEX_INPUT_ATTRIBUTES); + } + } + else { + + /* Handle Any required format conversions. + * NOTE(Metal): If there is a mis-match between the format of an attribute + * in the shader interface, and the specified format in the VertexBuffer VertexFormat, + * we need to perform a format conversion. + * + * The Metal API can perform certain conversions internally during vertex assembly: + * - Type Normalization e.g short2 to float2 between 0.0 to 1.0. + * - Type Truncation e.g. Float4 to Float2. + * - Type expansion e,g, Float3 to Float4 (Following 0,0,0,1 for assignment to empty + * elements). + * + * Certain conversion cannot be performed however, and in these cases, we need to + * instruct the shader to generate a specialized version with a conversion routine upon + * attribute read. + * - This handles cases such as conversion between types e.g. Integer to float without + * normalization. + * + * For more information on the supported and unsupported conversions, see: + * https://developer.apple.com/documentation/metal/mtlvertexattributedescriptor/1516081-format?language=objc + */ + MTLVertexFormat converted_format; + bool can_use_internal_conversion = mtl_convert_vertex_format( + mtl_attr.format, + (GPUVertCompType)a->comp_type, + a->comp_len, + (GPUVertFetchMode)a->fetch_mode, + &converted_format); + bool is_floating_point_format = (a->comp_type == GPU_COMP_F32); + + if (can_use_internal_conversion) { + desc.vertex_descriptor.attributes[mtl_attr.location].format = converted_format; + desc.vertex_descriptor.attributes[mtl_attr.location].format_conversion_mode = + is_floating_point_format ? (GPUVertFetchMode)GPU_FETCH_FLOAT : + (GPUVertFetchMode)GPU_FETCH_INT; + BLI_assert(converted_format != MTLVertexFormatInvalid); + } + else { + /* The internal implicit conversion is not supported. + * In this case, we need to handle conversion inside the shader. + * This is handled using `format_conversion_mode`. + * `format_conversion_mode` is assigned the blender-specified fetch mode (GPU_FETCH_*). + * This then controls how a given attribute is interpreted. The data will be read + * as specified and then converted appropriately to the correct form. + * + * e.g. if `GPU_FETCH_INT_TO_FLOAT` is specified, the specialized read-routine + * in the shader will read the data as an int, and cast this to floating point + * representation. (Rather than reading the source data as float). + * + * NOTE: Even if full conversion is not supported, we may still partially perform an + * implicit conversion where possible, such as vector truncation or expansion. */ + MTLVertexFormat converted_format; + bool can_convert = mtl_vertex_format_resize( + mtl_attr.format, a->comp_len, &converted_format); + desc.vertex_descriptor.attributes[mtl_attr.location].format = can_convert ? + converted_format : + mtl_attr.format; + desc.vertex_descriptor.attributes[mtl_attr.location].format_conversion_mode = + (GPUVertFetchMode)a->fetch_mode; + BLI_assert(desc.vertex_descriptor.attributes[mtl_attr.location].format != + MTLVertexFormatInvalid); + } + desc.vertex_descriptor.attributes[mtl_attr.location].offset = attribute_offset; + desc.vertex_descriptor.attributes[mtl_attr.location].buffer_index = buffer_index; + desc.vertex_descriptor.num_attributes = ((mtl_attr.location + 1) > + desc.vertex_descriptor.num_attributes) ? + (mtl_attr.location + 1) : + desc.vertex_descriptor.num_attributes; + + /* SSBO Vertex Fetch attribute bind. */ + if (active_shader_->get_uses_ssbo_vertex_fetch()) { + BLI_assert_msg(desc.vertex_descriptor.attributes[mtl_attr.location].format == + mtl_attr.format, + "SSBO Vertex Fetch does not support attribute conversion."); + + MTLSSBOAttribute ssbo_attr( + mtl_attr.index, + buffer_index, + attribute_offset, + buffer_stride, + MTLShader::ssbo_vertex_type_to_attr_type( + desc.vertex_descriptor.attributes[mtl_attr.location].format), + instanced); + + active_shader_->ssbo_vertex_fetch_bind_attribute(ssbo_attr); + desc.vertex_descriptor.ssbo_attributes[desc.vertex_descriptor.num_ssbo_attributes] = + ssbo_attr; + desc.vertex_descriptor.num_ssbo_attributes++; + } + + /* NOTE: We are setting num_attributes to be up to the maximum found index, because of + * this, it is possible that we may skip over certain attributes if they were not in the + * source GPUVertFormat. */ + MTL_LOG_INFO( + " -- Batch Attribute(%d): ORIG Shader Format: %d, ORIG Vert format: %d, Vert " + "components: %d, Fetch Mode %d --> FINAL FORMAT: %d\n", + mtl_attr.location, + (int)mtl_attr.format, + (int)a->comp_type, + (int)a->comp_len, + (int)a->fetch_mode, + (int)desc.vertex_descriptor.attributes[mtl_attr.location].format); + + MTL_LOG_INFO( + " -- [Batch] matching %s attribute '%s' (Attribute Index: %d, Buffer index: %d, " + "offset: %d)\n", + (instanced) ? "instance" : "vertex", + name, + mtl_attr.location, + buffer_index, + attribute_offset); + } + } + } + } + if (buffer_added) { + return buffer_index; + } + return -1; +} + +id<MTLRenderCommandEncoder> MTLBatch::bind(uint v_first, uint v_count, uint i_first, uint i_count) +{ + /* Setup draw call and render pipeline state here. Called by every draw, but setup here so that + * MTLDrawList only needs to perform setup a single time. */ + BLI_assert(this); + + /* Fetch Metal device. */ + MTLContext *ctx = MTLContext::get(); + if (!ctx) { + BLI_assert_msg(false, "No context available for rendering."); + return nil; + } + + /* Verify Shader. */ + active_shader_ = (shader) ? static_cast<MTLShader *>(unwrap(shader)) : nullptr; + + if (active_shader_ == nullptr || !active_shader_->is_valid()) { + /* Skip drawing if there is no valid Metal shader. + * This will occur if the path through which the shader is prepared + * is invalid (e.g. Python without create-info), or, the source shader uses a geometry pass. */ + BLI_assert_msg(false, "No valid Metal shader!"); + return nil; + } + + /* Check if using SSBO Fetch Mode. + * This is an alternative drawing mode to geometry shaders, wherein vertex buffers + * are bound as readable (random-access) GPU buffers and certain descriptor properties + * are passed using Shader uniforms. */ + bool uses_ssbo_fetch = active_shader_->get_uses_ssbo_vertex_fetch(); + + /* Prepare Vertex Descriptor and extract VertexBuffers to bind. */ + MTLVertBuf *buffers[GPU_BATCH_VBO_MAX_LEN] = {nullptr}; + int num_buffers = 0; + + /* Ensure Index Buffer is ready. */ + MTLIndexBuf *mtl_elem = static_cast<MTLIndexBuf *>(reinterpret_cast<IndexBuf *>(this->elem)); + if (mtl_elem != NULL) { + mtl_elem->upload_data(); + } + + /* Populate vertex descriptor with attribute binding information. + * The vertex descriptor and buffer layout descriptors describe + * how vertex data from bound vertex buffers maps to the + * shader's input. + * A unique vertex descriptor will result in a new PipelineStateObject + * being generated for the currently bound shader. */ + prepare_vertex_descriptor_and_bindings(buffers, num_buffers, v_first, v_count, i_first, i_count); + + /* Prepare Vertex Buffers - Run before RenderCommandEncoder in case BlitCommandEncoder buffer + * data operations are required. */ + for (int i = 0; i < num_buffers; i++) { + MTLVertBuf *buf_at_index = buffers[i]; + if (buf_at_index == NULL) { + BLI_assert_msg( + false, + "Total buffer count does not match highest buffer index, could be gaps in bindings"); + continue; + } + + MTLVertBuf *mtlvbo = static_cast<MTLVertBuf *>(reinterpret_cast<VertBuf *>(buf_at_index)); + mtlvbo->bind(); + } + + /* Ensure render pass is active and fetch active RenderCommandEncoder. */ + id<MTLRenderCommandEncoder> rec = ctx->ensure_begin_render_pass(); + + /* Fetch RenderPassState to enable resource binding for active pass. */ + MTLRenderPassState &rps = ctx->main_command_buffer.get_render_pass_state(); + + /* Debug Check: Ensure Frame-buffer instance is not dirty. */ + BLI_assert(!ctx->main_command_buffer.get_active_framebuffer()->get_dirty()); + + /* Bind Shader. */ + this->shader_bind(); + + /* GPU debug markers. */ + if (G.debug & G_DEBUG_GPU) { + [rec pushDebugGroup:[NSString stringWithFormat:@"batch_bind%@(shader: %s)", + this->elem ? @"(indexed)" : @"", + active_shader_->get_interface()->get_name()]]; + [rec insertDebugSignpost:[NSString + stringWithFormat:@"batch_bind%@(shader: %s)", + this->elem ? @"(indexed)" : @"", + active_shader_->get_interface()->get_name()]]; + } + + /* Ensure Context Render Pipeline State is fully setup and ready to execute the draw. */ + MTLPrimitiveType mtl_prim_type = gpu_prim_type_to_metal(this->prim_type); + if (!ctx->ensure_render_pipeline_state(mtl_prim_type)) { + printf("FAILED TO ENSURE RENDER PIPELINE STATE"); + BLI_assert(false); + + if (G.debug & G_DEBUG_GPU) { + [rec popDebugGroup]; + } + return nil; + } + + /*** Bind Vertex Buffers and Index Buffers **/ + + /* SSBO Vertex Fetch Buffer bindings. */ + if (uses_ssbo_fetch) { + + /* SSBO Vertex Fetch - Bind Index Buffer to appropriate slot -- if used. */ + id<MTLBuffer> idx_buffer = nil; + GPUPrimType final_prim_type = this->prim_type; + + if (mtl_elem != nullptr) { + + /* Fetch index buffer. This function can situationally return an optimized + * index buffer of a different primitive type. If this is the case, `final_prim_type` + * and `v_count` will be updated with the new format. + * NOTE: For indexed rendering, v_count represents the number of indices. */ + idx_buffer = mtl_elem->get_index_buffer(final_prim_type, v_count); + BLI_assert(idx_buffer != nil); + + /* Update uniforms for SSBO-vertex-fetch-mode indexed rendering to flag usage. */ + int &uniform_ssbo_index_mode_u16 = active_shader_->uni_ssbo_uses_index_mode_u16; + BLI_assert(uniform_ssbo_index_mode_u16 != -1); + int uses_index_mode_u16 = (mtl_elem->index_type_ == GPU_INDEX_U16) ? 1 : 0; + active_shader_->uniform_int(uniform_ssbo_index_mode_u16, 1, 1, &uses_index_mode_u16); + } + else { + idx_buffer = ctx->get_null_buffer(); + } + rps.bind_vertex_buffer(idx_buffer, 0, MTL_SSBO_VERTEX_FETCH_IBO_INDEX); + + /* Ensure all attributes are set */ + active_shader_->ssbo_vertex_fetch_bind_attributes_end(rec); + + /* Bind NULL Buffers for unused vertex data slots. */ + id<MTLBuffer> null_buffer = ctx->get_null_buffer(); + BLI_assert(null_buffer != nil); + for (int i = num_buffers; i < MTL_SSBO_VERTEX_FETCH_MAX_VBOS; i++) { + if (rps.cached_vertex_buffer_bindings[i].metal_buffer == nil) { + rps.bind_vertex_buffer(null_buffer, 0, i); + } + } + + /* Flag whether Indexed rendering is used or not. */ + int &uniform_ssbo_use_indexed = active_shader_->uni_ssbo_uses_indexed_rendering; + BLI_assert(uniform_ssbo_use_indexed != -1); + int uses_indexed_rendering = (mtl_elem != NULL) ? 1 : 0; + active_shader_->uniform_int(uniform_ssbo_use_indexed, 1, 1, &uses_indexed_rendering); + + /* Set SSBO-fetch-mode status uniforms. */ + BLI_assert(active_shader_->uni_ssbo_input_prim_type_loc != -1); + BLI_assert(active_shader_->uni_ssbo_input_vert_count_loc != -1); + GPU_shader_uniform_vector_int(reinterpret_cast<GPUShader *>(wrap(active_shader_)), + active_shader_->uni_ssbo_input_prim_type_loc, + 1, + 1, + (const int *)(&final_prim_type)); + GPU_shader_uniform_vector_int(reinterpret_cast<GPUShader *>(wrap(active_shader_)), + active_shader_->uni_ssbo_input_vert_count_loc, + 1, + 1, + (const int *)(&v_count)); + } + + /* Bind Vertex Buffers. */ + for (int i = 0; i < num_buffers; i++) { + MTLVertBuf *buf_at_index = buffers[i]; + if (buf_at_index == NULL) { + BLI_assert_msg( + false, + "Total buffer count does not match highest buffer index, could be gaps in bindings"); + continue; + } + /* Buffer handle. */ + MTLVertBuf *mtlvbo = static_cast<MTLVertBuf *>(reinterpret_cast<VertBuf *>(buf_at_index)); + mtlvbo->flag_used(); + + /* Fetch buffer from MTLVertexBuffer and bind. */ + id<MTLBuffer> mtl_buffer = mtlvbo->get_metal_buffer(); + + BLI_assert(mtl_buffer != nil); + rps.bind_vertex_buffer(mtl_buffer, 0, i); + } + + if (G.debug & G_DEBUG_GPU) { + [rec popDebugGroup]; + } + + /* Return Render Command Encoder used with setup. */ + return rec; +} + +void MTLBatch::unbind() +{ +} + +void MTLBatch::prepare_vertex_descriptor_and_bindings( + MTLVertBuf **buffers, int &num_buffers, int v_first, int v_count, int i_first, int i_count) +{ + + /* Here we populate the MTLContext vertex descriptor and resolve which buffers need to be bound. + */ + MTLStateManager *state_manager = static_cast<MTLStateManager *>( + MTLContext::get()->state_manager); + MTLRenderPipelineStateDescriptor &desc = state_manager->get_pipeline_descriptor(); + const MTLShaderInterface *interface = active_shader_->get_interface(); + uint16_t attr_mask = interface->get_enabled_attribute_mask(); + + /* Reset vertex descriptor to default state. */ + desc.reset_vertex_descriptor(); + + /* Fetch Vertex and Instance Buffers. */ + Span<MTLVertBuf *> mtl_verts(reinterpret_cast<MTLVertBuf **>(this->verts), + GPU_BATCH_VBO_MAX_LEN); + Span<MTLVertBuf *> mtl_inst(reinterpret_cast<MTLVertBuf **>(this->inst), + GPU_BATCH_INST_VBO_MAX_LEN); + + /* SSBO Vertex fetch also passes vertex descriptor information into the shader. */ + if (active_shader_->get_uses_ssbo_vertex_fetch()) { + active_shader_->ssbo_vertex_fetch_bind_attributes_begin(); + } + + /* Resolve Metal vertex buffer bindings. */ + /* Vertex Descriptors + * ------------------ + * Vertex Descriptors are required to generate a pipeline state, based on the current Batch's + * buffer bindings. These bindings are a unique matching, depending on what input attributes a + * batch has in its buffers, and those which are supported by the shader interface. + + * We iterate through the buffers and resolve which attributes satisfy the requirements of the + * currently bound shader. We cache this data, for a given Batch<->ShderInterface pairing in a + * VAO cache to avoid the need to recalculate this data. */ + bool buffer_is_instanced[GPU_BATCH_VBO_MAX_LEN] = {false}; + + VertexDescriptorShaderInterfacePair *descriptor = this->vao_cache.find(interface); + if (descriptor) { + desc.vertex_descriptor = descriptor->vertex_descriptor; + attr_mask = descriptor->attr_mask; + num_buffers = descriptor->num_buffers; + + for (int bid = 0; bid < GPU_BATCH_VBO_MAX_LEN; ++bid) { + if (descriptor->bufferIds[bid].used) { + if (descriptor->bufferIds[bid].is_instance) { + buffers[bid] = mtl_inst[descriptor->bufferIds[bid].id]; + buffer_is_instanced[bid] = true; + } + else { + buffers[bid] = mtl_verts[descriptor->bufferIds[bid].id]; + buffer_is_instanced[bid] = false; + } + } + } + + /* Use cached ssbo attribute binding data. */ + if (active_shader_->get_uses_ssbo_vertex_fetch()) { + BLI_assert(desc.vertex_descriptor.uses_ssbo_vertex_fetch); + for (int attr_id = 0; attr_id < desc.vertex_descriptor.num_ssbo_attributes; attr_id++) { + active_shader_->ssbo_vertex_fetch_bind_attribute( + desc.vertex_descriptor.ssbo_attributes[attr_id]); + } + } + } + else { + VertexDescriptorShaderInterfacePair pair{}; + pair.interface = interface; + + for (int i = 0; i < GPU_BATCH_VBO_MAX_LEN; ++i) { + pair.bufferIds[i].id = -1; + pair.bufferIds[i].is_instance = 0; + pair.bufferIds[i].used = 0; + } + /* NOTE: Attribute extraction order from buffer is the reverse of the OpenGL as we flag once an + * attribute is found, rather than pre-setting the mask. */ + /* Extract Instance attributes (These take highest priority). */ + for (int v = 0; v < GPU_BATCH_INST_VBO_MAX_LEN; v++) { + if (mtl_inst[v]) { + MTL_LOG_INFO(" -- [Batch] Checking bindings for bound instance buffer %p\n", mtl_inst[v]); + int buffer_ind = this->prepare_vertex_binding( + mtl_inst[v], desc, interface, attr_mask, true); + if (buffer_ind >= 0) { + buffers[buffer_ind] = mtl_inst[v]; + buffer_is_instanced[buffer_ind] = true; + + pair.bufferIds[buffer_ind].id = v; + pair.bufferIds[buffer_ind].used = 1; + pair.bufferIds[buffer_ind].is_instance = 1; + num_buffers = ((buffer_ind + 1) > num_buffers) ? (buffer_ind + 1) : num_buffers; + } + } + } + + /* Extract Vertex attributes (First-bound vertex buffer takes priority). */ + for (int v = 0; v < GPU_BATCH_VBO_MAX_LEN; v++) { + if (mtl_verts[v] != NULL) { + MTL_LOG_INFO(" -- [Batch] Checking bindings for bound vertex buffer %p\n", mtl_verts[v]); + int buffer_ind = this->prepare_vertex_binding( + mtl_verts[v], desc, interface, attr_mask, false); + if (buffer_ind >= 0) { + buffers[buffer_ind] = mtl_verts[v]; + buffer_is_instanced[buffer_ind] = false; + + pair.bufferIds[buffer_ind].id = v; + pair.bufferIds[buffer_ind].used = 1; + pair.bufferIds[buffer_ind].is_instance = 0; + num_buffers = ((buffer_ind + 1) > num_buffers) ? (buffer_ind + 1) : num_buffers; + } + } + } + + /* Add to VertexDescriptor cache */ + desc.vertex_descriptor.uses_ssbo_vertex_fetch = active_shader_->get_uses_ssbo_vertex_fetch(); + pair.attr_mask = attr_mask; + pair.vertex_descriptor = desc.vertex_descriptor; + pair.num_buffers = num_buffers; + if (!this->vao_cache.insert(pair)) { + printf( + "[Performance Warning] cache is full (Size: %d), vertex descriptor will not be cached\n", + GPU_VAO_STATIC_LEN); + } + } + +/* DEBUG: verify if our attribute bindings have been fully provided as expected. */ +#if MTL_DEBUG_SHADER_ATTRIBUTES == 1 + if (attr_mask != 0) { + for (uint16_t mask = 1, a = 0; a < 16; a++, mask <<= 1) { + if (attr_mask & mask) { + /* Fallback for setting default attributes, for missed slots. Attributes flagged with + * 'MTLVertexFormatInvalid' in the vertex descriptor are bound to a NULL buffer during PSO + * creation. */ + MTL_LOG_WARNING("MTLBatch: Missing expected attribute '%s' at index '%d' for shader: %s\n", + this->active_shader->interface->attributes[a].name, + a, + interface->name); + /* Ensure any assigned attribute has not been given an invalid format. This should not + * occur and may be the result of an unsupported attribute type conversion. */ + BLI_assert(desc.attributes[a].format == MTLVertexFormatInvalid); + } + } + } +#endif +} + +void MTLBatch::draw_advanced(int v_first, int v_count, int i_first, int i_count) +{ + +#if TRUST_NO_ONE + BLI_assert(v_count > 0 && i_count > 0); +#endif + + /* Setup RenderPipelineState for batch. */ + MTLContext *ctx = reinterpret_cast<MTLContext *>(GPU_context_active_get()); + id<MTLRenderCommandEncoder> rec = this->bind(v_first, v_count, i_first, i_count); + if (rec == nil) { + return; + } + + /* Fetch IndexBuffer and resolve primitive type. */ + MTLIndexBuf *mtl_elem = static_cast<MTLIndexBuf *>(reinterpret_cast<IndexBuf *>(this->elem)); + MTLPrimitiveType mtl_prim_type = gpu_prim_type_to_metal(this->prim_type); + + /* Render using SSBO Vertex Fetch. */ + if (active_shader_->get_uses_ssbo_vertex_fetch()) { + + /* Submit draw call with modified vertex count, which reflects vertices per primitive defined + * in the USE_SSBO_VERTEX_FETCH pragma. */ + int num_input_primitives = gpu_get_prim_count_from_type(v_count, this->prim_type); + int output_num_verts = num_input_primitives * + active_shader_->get_ssbo_vertex_fetch_output_num_verts(); + BLI_assert_msg( + mtl_vertex_count_fits_primitive_type( + output_num_verts, active_shader_->get_ssbo_vertex_fetch_output_prim_type()), + "Output Vertex count is not compatible with the requested output vertex primitive type"); + [rec drawPrimitives:active_shader_->get_ssbo_vertex_fetch_output_prim_type() + vertexStart:0 + vertexCount:output_num_verts + instanceCount:i_count + baseInstance:i_first]; + ctx->main_command_buffer.register_draw_counters(output_num_verts * i_count); + } + /* Perform regular draw. */ + else if (mtl_elem == NULL) { + + /* Primitive Type toplogy emulation. */ + if (mtl_needs_topology_emulation(this->prim_type)) { + + /* Generate index buffer for primitive types requiring emulation. */ + GPUPrimType emulated_prim_type = this->prim_type; + uint32_t emulated_v_count = v_count; + id<MTLBuffer> generated_index_buffer = this->get_emulated_toplogy_buffer(emulated_prim_type, + emulated_v_count); + BLI_assert(generated_index_buffer != nil); + + MTLPrimitiveType emulated_mtl_prim_type = gpu_prim_type_to_metal(emulated_prim_type); + + /* Temp: Disable culling for emulated primitive types. + * TODO(Metal): Support face winding in topology buffer. */ + [rec setCullMode:MTLCullModeNone]; + + if (generated_index_buffer != nil) { + BLI_assert(emulated_mtl_prim_type == MTLPrimitiveTypeTriangle || + emulated_mtl_prim_type == MTLPrimitiveTypeLine); + if (emulated_mtl_prim_type == MTLPrimitiveTypeTriangle) { + BLI_assert(emulated_v_count % 3 == 0); + } + if (emulated_mtl_prim_type == MTLPrimitiveTypeLine) { + BLI_assert(emulated_v_count % 2 == 0); + } + + /* Set depth stencil state (requires knowledge of primitive type). */ + ctx->ensure_depth_stencil_state(emulated_mtl_prim_type); + + [rec drawIndexedPrimitives:emulated_mtl_prim_type + indexCount:emulated_v_count + indexType:MTLIndexTypeUInt32 + indexBuffer:generated_index_buffer + indexBufferOffset:0 + instanceCount:i_count + baseVertex:v_first + baseInstance:i_first]; + } + else { + printf("[Note] Cannot draw batch -- Emulated Topology mode: %u not yet supported\n", + this->prim_type); + } + } + else { + /* Set depth stencil state (requires knowledge of primitive type). */ + ctx->ensure_depth_stencil_state(mtl_prim_type); + + /* Issue draw call. */ + [rec drawPrimitives:mtl_prim_type + vertexStart:v_first + vertexCount:v_count + instanceCount:i_count + baseInstance:i_first]; + } + ctx->main_command_buffer.register_draw_counters(v_count * i_count); + } + /* Perform indexed draw. */ + else { + + MTLIndexType index_type = MTLIndexBuf::gpu_index_type_to_metal(mtl_elem->index_type_); + uint32_t base_index = mtl_elem->index_base_; + uint32_t index_size = (mtl_elem->index_type_ == GPU_INDEX_U16) ? 2 : 4; + uint32_t v_first_ofs = ((v_first + mtl_elem->index_start_) * index_size); + BLI_assert_msg((v_first_ofs % index_size) == 0, + "Index offset is not 2/4-byte aligned as per METAL spec"); + + /* Fetch index buffer. May return an index buffer of a differing format, + * if index buffer optimization is used. In these cases, final_prim_type and + * index_count get updated with the new properties. */ + GPUPrimType final_prim_type = this->prim_type; + uint index_count = v_count; + + id<MTLBuffer> index_buffer = mtl_elem->get_index_buffer(final_prim_type, index_count); + mtl_prim_type = gpu_prim_type_to_metal(final_prim_type); + BLI_assert(index_buffer != nil); + + if (index_buffer != nil) { + + /* Set depth stencil state (requires knowledge of primitive type). */ + ctx->ensure_depth_stencil_state(mtl_prim_type); + + /* Issue draw call. */ + [rec drawIndexedPrimitives:mtl_prim_type + indexCount:index_count + indexType:index_type + indexBuffer:index_buffer + indexBufferOffset:v_first_ofs + instanceCount:i_count + baseVertex:base_index + baseInstance:i_first]; + ctx->main_command_buffer.register_draw_counters(index_count * i_count); + } + else { + BLI_assert_msg(false, "Index buffer does not have backing Metal buffer"); + } + } + + /* End of draw. */ + this->unbind(); +} + +/** \} */ + +/* -------------------------------------------------------------------- */ +/** \name Topology emulation and optimization + * \{ */ + +id<MTLBuffer> MTLBatch::get_emulated_toplogy_buffer(GPUPrimType &in_out_prim_type, + uint32_t &in_out_v_count) +{ + + BLI_assert(in_out_v_count > 0); + /* Determine emulated primitive types. */ + GPUPrimType input_prim_type = in_out_prim_type; + uint32_t v_count = in_out_v_count; + GPUPrimType output_prim_type; + switch (input_prim_type) { + case GPU_PRIM_POINTS: + case GPU_PRIM_LINES: + case GPU_PRIM_TRIS: + BLI_assert_msg(false, "Optimal primitive types should not reach here."); + return nil; + break; + case GPU_PRIM_LINES_ADJ: + case GPU_PRIM_TRIS_ADJ: + BLI_assert_msg(false, "Adjacency primitive types should not reach here."); + return nil; + break; + case GPU_PRIM_LINE_STRIP: + case GPU_PRIM_LINE_LOOP: + case GPU_PRIM_LINE_STRIP_ADJ: + output_prim_type = GPU_PRIM_LINES; + break; + case GPU_PRIM_TRI_STRIP: + case GPU_PRIM_TRI_FAN: + output_prim_type = GPU_PRIM_TRIS; + break; + default: + BLI_assert_msg(false, "Invalid primitive type."); + return nil; + } + + /* Check if topology buffer exists and is valid. */ + if (this->emulated_topology_buffer_ != nullptr && + (emulated_topology_type_ != input_prim_type || topology_buffer_input_v_count_ != v_count)) { + + /* Release existing topology buffer. */ + emulated_topology_buffer_->free(); + emulated_topology_buffer_ = nullptr; + } + + /* Generate new topology index buffer. */ + if (this->emulated_topology_buffer_ == nullptr) { + /* Calculate IB len. */ + uint32_t output_prim_count = 0; + switch (input_prim_type) { + case GPU_PRIM_LINE_STRIP: + case GPU_PRIM_LINE_STRIP_ADJ: + output_prim_count = v_count - 1; + break; + case GPU_PRIM_LINE_LOOP: + output_prim_count = v_count; + break; + case GPU_PRIM_TRI_STRIP: + case GPU_PRIM_TRI_FAN: + output_prim_count = v_count - 2; + break; + default: + BLI_assert_msg(false, "Cannot generate optimized topology buffer for other types."); + break; + } + uint32_t output_IB_elems = output_prim_count * ((output_prim_type == GPU_PRIM_TRIS) ? 3 : 2); + + /* Allocate buffer. */ + uint32_t buffer_bytes = output_IB_elems * 4; + BLI_assert(buffer_bytes > 0); + this->emulated_topology_buffer_ = MTLContext::get_global_memory_manager().allocate( + buffer_bytes, true); + + /* Populate. */ + uint32_t *data = (uint32_t *)this->emulated_topology_buffer_->get_host_ptr(); + BLI_assert(data != nullptr); + + /* TODO(Metal): Support inverse winding modes. */ + bool winding_clockwise = false; + UNUSED_VARS(winding_clockwise); + + switch (input_prim_type) { + /* Line Loop. */ + case GPU_PRIM_LINE_LOOP: { + int line = 0; + for (line = 0; line < output_prim_count - 1; line++) { + data[line * 3 + 0] = line + 0; + data[line * 3 + 1] = line + 1; + } + /* Closing line. */ + data[line * 2 + 0] = line + 0; + data[line * 2 + 1] = 0; + } break; + + /* Triangle Fan. */ + case GPU_PRIM_TRI_FAN: { + for (int triangle = 0; triangle < output_prim_count; triangle++) { + data[triangle * 3 + 0] = 0; /* Always 0 */ + data[triangle * 3 + 1] = triangle + 1; + data[triangle * 3 + 2] = triangle + 2; + } + } break; + + default: + BLI_assert_msg(false, "Other primitive types do not require emulation."); + return nil; + } + + /* Flush. */ + this->emulated_topology_buffer_->flush(); + /* Assign members relating to current cached IB. */ + topology_buffer_input_v_count_ = v_count; + topology_buffer_output_v_count_ = output_IB_elems; + emulated_topology_type_ = input_prim_type; + } + + /* Return. */ + in_out_v_count = topology_buffer_output_v_count_; + in_out_prim_type = output_prim_type; + return (emulated_topology_buffer_) ? emulated_topology_buffer_->get_metal_buffer() : nil; +} + +/** \} */ + +} // blender::gpu diff --git a/source/blender/gpu/metal/mtl_context.mm b/source/blender/gpu/metal/mtl_context.mm index ef66a1f2111..50576379f0d 100644 --- a/source/blender/gpu/metal/mtl_context.mm +++ b/source/blender/gpu/metal/mtl_context.mm @@ -995,19 +995,21 @@ bool MTLContext::ensure_uniform_buffer_bindings( if (ubo.buffer_index >= 0) { - const uint32_t buffer_index = ubo.buffer_index; + /* Uniform Buffer index offset by 1 as the first shader buffer binding slot is reserved for + * the uniform PushConstantBlock. */ + const uint32_t buffer_index = ubo.buffer_index + 1; int ubo_offset = 0; id<MTLBuffer> ubo_buffer = nil; int ubo_size = 0; bool bind_dummy_buffer = false; - if (this->pipeline_state.ubo_bindings[buffer_index].bound) { + if (this->pipeline_state.ubo_bindings[ubo_index].bound) { /* Fetch UBO global-binding properties from slot. */ ubo_offset = 0; - ubo_buffer = this->pipeline_state.ubo_bindings[buffer_index].ubo->get_metal_buffer( + ubo_buffer = this->pipeline_state.ubo_bindings[ubo_index].ubo->get_metal_buffer( &ubo_offset); - ubo_size = this->pipeline_state.ubo_bindings[buffer_index].ubo->get_size(); + ubo_size = this->pipeline_state.ubo_bindings[ubo_index].ubo->get_size(); /* Use dummy zero buffer if no buffer assigned -- this is an optimization to avoid * allocating zero buffers. */ diff --git a/source/blender/gpu/metal/mtl_drawlist.hh b/source/blender/gpu/metal/mtl_drawlist.hh index ed99c76faa7..47055f3d7f4 100644 --- a/source/blender/gpu/metal/mtl_drawlist.hh +++ b/source/blender/gpu/metal/mtl_drawlist.hh @@ -9,34 +9,50 @@ #pragma once -#pragma once - +#include "BLI_sys_types.h" +#include "GPU_batch.h" +#include "MEM_guardedalloc.h" #include "gpu_drawlist_private.hh" -namespace blender { -namespace gpu { +#include "mtl_batch.hh" +#include "mtl_context.hh" + +namespace blender::gpu { /** - * TODO(Metal): MTLDrawList Implementation. Included as temporary stub. - */ + * Implementation of Multi Draw Indirect using OpenGL. + **/ class MTLDrawList : public DrawList { + + private: + /** Batch for which we are recording commands for. */ + MTLBatch *batch_; + /** Mapped memory bounds. */ + void *data_; + /** Length of the mapped buffer (in byte). */ + size_t data_size_; + /** Current offset inside the mapped buffer (in byte). */ + size_t command_offset_; + /** Current number of command recorded inside the mapped buffer. */ + uint32_t command_len_; + /** Is UINT_MAX if not drawing indexed geom. Also Avoid dereferencing batch. */ + uint32_t base_index_; + /** Also Avoid dereferencing batch. */ + uint32_t v_first_, v_count_; + /** Length of whole the buffer (in byte). */ + uint32_t buffer_size_; + public: - MTLDrawList(int length) - { - } - ~MTLDrawList() - { - } - - void append(GPUBatch *batch, int i_first, int i_count) override - { - } - void submit() override - { - } + MTLDrawList(int length); + ~MTLDrawList(); + + void append(GPUBatch *batch, int i_first, int i_count) override; + void submit() override; + + private: + void init(); MEM_CXX_CLASS_ALLOC_FUNCS("MTLDrawList"); }; -} // namespace gpu -} // namespace blender +} // namespace blender::gpu diff --git a/source/blender/gpu/metal/mtl_drawlist.mm b/source/blender/gpu/metal/mtl_drawlist.mm new file mode 100644 index 00000000000..99194d2b72c --- /dev/null +++ b/source/blender/gpu/metal/mtl_drawlist.mm @@ -0,0 +1,284 @@ +/* SPDX-License-Identifier: GPL-2.0-or-later */ + +/** \file + * \ingroup gpu + * + * Implementation of Multi Draw Indirect using OpenGL. + * Fallback if the needed extensions are not supported. + */ + +#include "BLI_assert.h" + +#include "GPU_batch.h" +#include "mtl_common.hh" +#include "mtl_drawlist.hh" +#include "mtl_primitive.hh" + +using namespace blender::gpu; + +namespace blender::gpu { + +/* Indirect draw call structure for reference. */ +/* MTLDrawPrimitivesIndirectArguments -- + * https://developer.apple.com/documentation/metal/mtldrawprimitivesindirectarguments?language=objc + */ +/* struct MTLDrawPrimitivesIndirectArguments { + * uint32_t vertexCount; + * uint32_t instanceCount; + * uint32_t vertexStart; + * uint32_t baseInstance; +};*/ + +/* MTLDrawIndexedPrimitivesIndirectArguments -- + * https://developer.apple.com/documentation/metal/mtldrawindexedprimitivesindirectarguments?language=objc + */ +/* struct MTLDrawIndexedPrimitivesIndirectArguments { + * uint32_t indexCount; + * uint32_t instanceCount; + * uint32_t indexStart; + * uint32_t baseVertex; + * uint32_t baseInstance; +};*/ + +#define MDI_ENABLED (buffer_size_ != 0) +#define MDI_DISABLED (buffer_size_ == 0) +#define MDI_INDEXED (base_index_ != UINT_MAX) + +MTLDrawList::MTLDrawList(int length) +{ + BLI_assert(length > 0); + batch_ = nullptr; + command_len_ = 0; + base_index_ = 0; + command_offset_ = 0; + data_size_ = 0; + buffer_size_ = sizeof(MTLDrawIndexedPrimitivesIndirectArguments) * length; + data_ = (void *)MEM_mallocN(buffer_size_, __func__); +} + +MTLDrawList::~MTLDrawList() +{ + if (data_) { + MEM_freeN(data_); + data_ = nullptr; + } +} + +void MTLDrawList::init() +{ + MTLContext *ctx = reinterpret_cast<MTLContext *>(GPU_context_active_get()); + BLI_assert(ctx); + BLI_assert(MDI_ENABLED); + BLI_assert(data_ == nullptr); + UNUSED_VARS_NDEBUG(ctx); + + batch_ = nullptr; + command_len_ = 0; + BLI_assert(data_); + + command_offset_ = 0; +} + +void MTLDrawList::append(GPUBatch *gpu_batch, int i_first, int i_count) +{ + /* Fallback when MultiDrawIndirect is not supported/enabled. */ + MTLShader *shader = static_cast<MTLShader *>(unwrap(gpu_batch->shader)); + bool requires_ssbo = (shader->get_uses_ssbo_vertex_fetch()); + bool requires_emulation = mtl_needs_topology_emulation(gpu_batch->prim_type); + if (MDI_DISABLED || requires_ssbo || requires_emulation) { + GPU_batch_draw_advanced(gpu_batch, 0, 0, i_first, i_count); + return; + } + + if (data_ == nullptr) { + this->init(); + } + BLI_assert(data_); + + MTLBatch *mtl_batch = static_cast<MTLBatch *>(gpu_batch); + BLI_assert(mtl_batch); + if (mtl_batch != batch_) { + /* Submit existing calls. */ + this->submit(); + + /* Begin new batch. */ + batch_ = mtl_batch; + + /* Cached for faster access. */ + MTLIndexBuf *el = batch_->elem_(); + base_index_ = el ? el->index_base_ : UINT_MAX; + v_first_ = el ? el->index_start_ : 0; + v_count_ = el ? el->index_len_ : batch_->verts_(0)->vertex_len; + } + + if (v_count_ == 0) { + /* Nothing to draw. */ + return; + } + + if (MDI_INDEXED) { + MTLDrawIndexedPrimitivesIndirectArguments *cmd = + reinterpret_cast<MTLDrawIndexedPrimitivesIndirectArguments *>((char *)data_ + + command_offset_); + cmd->indexStart = v_first_; + cmd->indexCount = v_count_; + cmd->instanceCount = i_count; + cmd->baseVertex = base_index_; + cmd->baseInstance = i_first; + } + else { + MTLDrawPrimitivesIndirectArguments *cmd = + reinterpret_cast<MTLDrawPrimitivesIndirectArguments *>((char *)data_ + command_offset_); + cmd->vertexStart = v_first_; + cmd->vertexCount = v_count_; + cmd->instanceCount = i_count; + cmd->baseInstance = i_first; + } + + size_t command_size = MDI_INDEXED ? sizeof(MTLDrawIndexedPrimitivesIndirectArguments) : + sizeof(MTLDrawPrimitivesIndirectArguments); + + command_offset_ += command_size; + command_len_++; + + /* Check if we can fit at least one other command. */ + if (command_offset_ + command_size > buffer_size_) { + this->submit(); + } + + return; +} + +void MTLDrawList::submit() +{ + /* Metal does not support MDI from the host side, but we still benefit from only executing the + * batch bind a single time, rather than per-draw. + * NOTE(Metal): Consider using #MTLIndirectCommandBuffer to achieve similar behavior. */ + if (command_len_ == 0) { + return; + } + + /* Something's wrong if we get here without MDI support. */ + BLI_assert(MDI_ENABLED); + BLI_assert(data_); + + /* Host-side MDI Currently unsupported on Metal. */ + bool can_use_MDI = false; + + /* Verify context. */ + MTLContext *ctx = reinterpret_cast<MTLContext *>(GPU_context_active_get()); + BLI_assert(ctx); + + /* Execute indirect draw calls. */ + MTLShader *shader = static_cast<MTLShader *>(unwrap(batch_->shader)); + bool SSBO_MODE = (shader->get_uses_ssbo_vertex_fetch()); + if (SSBO_MODE) { + can_use_MDI = false; + BLI_assert(false); + return; + } + + /* Heuristic to determine whether using indirect drawing is more efficient. */ + size_t command_size = MDI_INDEXED ? sizeof(MTLDrawIndexedPrimitivesIndirectArguments) : + sizeof(MTLDrawPrimitivesIndirectArguments); + const bool is_finishing_a_buffer = (command_offset_ + command_size > buffer_size_); + can_use_MDI = can_use_MDI && (is_finishing_a_buffer || command_len_ > 2); + + /* Bind Batch to setup render pipeline state. */ + id<MTLRenderCommandEncoder> rec = batch_->bind(0, 0, 0, 0); + if (!rec) { + BLI_assert_msg(false, "A RenderCommandEncoder should always be available!\n"); + return; + } + + /* Common properties. */ + MTLPrimitiveType mtl_prim_type = gpu_prim_type_to_metal(batch_->prim_type); + + /* Execute multi-draw indirect. */ + if (can_use_MDI && false) { + /* Metal Doesn't support MDI -- Singular Indirect draw calls are supported, + * but Multi-draw is not. + * TODO(Metal): Consider using #IndirectCommandBuffers to provide similar + * behavior. */ + } + else { + + /* Execute draws manually. */ + if (MDI_INDEXED) { + MTLDrawIndexedPrimitivesIndirectArguments *cmd = + (MTLDrawIndexedPrimitivesIndirectArguments *)data_; + MTLIndexBuf *mtl_elem = static_cast<MTLIndexBuf *>( + reinterpret_cast<IndexBuf *>(batch_->elem)); + BLI_assert(mtl_elem); + MTLIndexType index_type = MTLIndexBuf::gpu_index_type_to_metal(mtl_elem->index_type_); + uint32_t index_size = (mtl_elem->index_type_ == GPU_INDEX_U16) ? 2 : 4; + uint32_t v_first_ofs = (mtl_elem->index_start_ * index_size); + uint32_t index_count = cmd->indexCount; + + /* Fetch index buffer. May return an index buffer of a differing format, + * if index buffer optimization is used. In these cases, mtl_prim_type and + * index_count get updated with the new properties. */ + GPUPrimType final_prim_type = batch_->prim_type; + id<MTLBuffer> index_buffer = mtl_elem->get_index_buffer(final_prim_type, index_count); + BLI_assert(index_buffer != nil); + + /* Final primitive type. */ + mtl_prim_type = gpu_prim_type_to_metal(final_prim_type); + + if (index_buffer != nil) { + + /* Set depth stencil state (requires knowledge of primitive type). */ + ctx->ensure_depth_stencil_state(mtl_prim_type); + + for (int i = 0; i < command_len_; i++, cmd++) { + [rec drawIndexedPrimitives:mtl_prim_type + indexCount:index_count + indexType:index_type + indexBuffer:index_buffer + indexBufferOffset:v_first_ofs + instanceCount:cmd->instanceCount + baseVertex:cmd->baseVertex + baseInstance:cmd->baseInstance]; + ctx->main_command_buffer.register_draw_counters(cmd->indexCount * cmd->instanceCount); + } + } + else { + BLI_assert_msg(false, "Index buffer does not have backing Metal buffer"); + } + } + else { + MTLDrawPrimitivesIndirectArguments *cmd = (MTLDrawPrimitivesIndirectArguments *)data_; + + /* Verify if topology emulation is required. */ + if (mtl_needs_topology_emulation(batch_->prim_type)) { + BLI_assert_msg(false, "topology emulation cases should use fallback."); + } + else { + + /* Set depth stencil state (requires knowledge of primitive type). */ + ctx->ensure_depth_stencil_state(mtl_prim_type); + + for (int i = 0; i < command_len_; i++, cmd++) { + [rec drawPrimitives:mtl_prim_type + vertexStart:cmd->vertexStart + vertexCount:cmd->vertexCount + instanceCount:cmd->instanceCount + baseInstance:cmd->baseInstance]; + ctx->main_command_buffer.register_draw_counters(cmd->vertexCount * cmd->instanceCount); + } + } + } + } + + /* Unbind batch. */ + batch_->unbind(); + + /* Reset command offsets. */ + command_len_ = 0; + command_offset_ = 0; + + /* Avoid keeping reference to the batch. */ + batch_ = nullptr; +} + +} // namespace blender::gpu diff --git a/source/blender/gpu/metal/mtl_immediate.mm b/source/blender/gpu/metal/mtl_immediate.mm index 4b63a3b1ce2..ee48bdd6ee1 100644 --- a/source/blender/gpu/metal/mtl_immediate.mm +++ b/source/blender/gpu/metal/mtl_immediate.mm @@ -99,6 +99,9 @@ void MTLImmediate::end() MTLRenderPipelineStateDescriptor &desc = state_manager->get_pipeline_descriptor(); const MTLShaderInterface *interface = active_mtl_shader->get_interface(); + /* Reset vertex descriptor to default state. */ + desc.reset_vertex_descriptor(); + desc.vertex_descriptor.num_attributes = interface->get_total_attributes(); desc.vertex_descriptor.num_vert_buffers = 1; diff --git a/source/blender/gpu/metal/mtl_pso_descriptor_state.hh b/source/blender/gpu/metal/mtl_pso_descriptor_state.hh index 198d309874b..04ceb5bdf03 100644 --- a/source/blender/gpu/metal/mtl_pso_descriptor_state.hh +++ b/source/blender/gpu/metal/mtl_pso_descriptor_state.hh @@ -243,6 +243,19 @@ struct MTLRenderPipelineStateDescriptor { return hash; } + + /* Reset the Vertex Descriptor to default. */ + void reset_vertex_descriptor() + { + vertex_descriptor.num_attributes = 0; + vertex_descriptor.num_vert_buffers = 0; + for (int i = 0; i < GPU_VERT_ATTR_MAX_LEN; i++) { + vertex_descriptor.attributes[i].format = MTLVertexFormatInvalid; + vertex_descriptor.attributes[i].offset = 0; + } + vertex_descriptor.uses_ssbo_vertex_fetch = false; + vertex_descriptor.num_ssbo_attributes = 0; + } }; } // namespace blender::gpu diff --git a/source/blender/gpu/metal/mtl_shader_generator.hh b/source/blender/gpu/metal/mtl_shader_generator.hh index 43890ca0170..63e2e6d5924 100644 --- a/source/blender/gpu/metal/mtl_shader_generator.hh +++ b/source/blender/gpu/metal/mtl_shader_generator.hh @@ -497,7 +497,7 @@ inline std::string get_stage_class_name(ShaderStage stage) inline bool is_builtin_type(std::string type) { /* Add Types as needed. */ - /* TODO(Metal): Consider replacing this with a switch and constexpr hash and switch. + /* TODO(Metal): Consider replacing this with a switch and `constexpr` hash and switch. * Though most efficient and maintainable approach to be determined. */ static std::map<std::string, eMTLDataType> glsl_builtin_types = { {"float", MTL_DATATYPE_FLOAT}, diff --git a/source/blender/gpu/metal/mtl_shader_interface.mm b/source/blender/gpu/metal/mtl_shader_interface.mm index 3703d5b5684..97a82345761 100644 --- a/source/blender/gpu/metal/mtl_shader_interface.mm +++ b/source/blender/gpu/metal/mtl_shader_interface.mm @@ -117,9 +117,7 @@ uint32_t MTLShaderInterface::add_uniform_block(uint32_t name_offset, MTLShaderUniformBlock &uni_block = ubos_[total_uniform_blocks_]; uni_block.name_offset = name_offset; - /* We offset the buffer binding index by one, as the first slot is reserved for push constant - * data. */ - uni_block.buffer_index = buffer_index + 1; + uni_block.buffer_index = buffer_index; uni_block.size = size; uni_block.current_offset = 0; uni_block.stage_mask = ShaderStage::BOTH; @@ -297,8 +295,10 @@ void MTLShaderInterface::prepare_common_shader_inputs() current_input->name_hash = BLI_hash_string(this->get_name_at_offset(shd_ubo.name_offset)); /* Location refers to the index in the ubos_ array. */ current_input->location = ubo_index; - /* Final binding location refers to the buffer binding index within the shader (Relative to - * MTL_uniform_buffer_base_index). */ + /* Binding location refers to the UBO bind slot in + * #MTLContextGlobalShaderPipelineState::ubo_bindings. The buffer bind index [[buffer(N)]] + * within the shader will apply an offset for bound vertex buffers and the default uniform + * PushConstantBlock. */ current_input->binding = shd_ubo.buffer_index; current_input++; } diff --git a/source/blender/gpu/metal/mtl_texture.hh b/source/blender/gpu/metal/mtl_texture.hh index ebc9eb2e00e..28b55306707 100644 --- a/source/blender/gpu/metal/mtl_texture.hh +++ b/source/blender/gpu/metal/mtl_texture.hh @@ -51,9 +51,9 @@ struct TextureUpdateRoutineSpecialisation { uint64_t hash() const { blender::DefaultHash<std::string> string_hasher; - return uint64_t(string_hasher( + return (uint64_t)string_hasher( this->input_data_type + this->output_data_type + - std::to_string((this->component_count_input << 8) + this->component_count_output))); + std::to_string((this->component_count_input << 8) + this->component_count_output)); } }; diff --git a/source/blender/gpu/metal/mtl_texture.mm b/source/blender/gpu/metal/mtl_texture.mm index 32029db6fd9..29dcc8d32ee 100644 --- a/source/blender/gpu/metal/mtl_texture.mm +++ b/source/blender/gpu/metal/mtl_texture.mm @@ -337,20 +337,6 @@ void gpu::MTLTexture::blit(gpu::MTLTexture *dst, GPU_batch_draw(quad); - /* TMP draw with IMM TODO(Metal): Remove this once GPUBatch is supported. */ - GPUVertFormat *imm_format = immVertexFormat(); - uint pos = GPU_vertformat_attr_add(imm_format, "pos", GPU_COMP_F32, 2, GPU_FETCH_FLOAT); - - immBindShader(shader); - immBegin(GPU_PRIM_TRI_STRIP, 4); - immVertex2f(pos, 1, 0); - immVertex2f(pos, 0, 0); - immVertex2f(pos, 1, 1); - immVertex2f(pos, 0, 1); - immEnd(); - immUnbindProgram(); - /**********************/ - /* restoring old pipeline state. */ GPU_depth_mask(depth_write_prev); GPU_stencil_write_mask_set(stencil_mask_prev); @@ -1472,10 +1458,82 @@ bool gpu::MTLTexture::init_internal() bool gpu::MTLTexture::init_internal(GPUVertBuf *vbo) { - /* Not a valid vertex buffer format, though verifying texture is not set as such - * as this is not supported on Apple Silicon. */ - BLI_assert_msg(this->format_ != GPU_DEPTH24_STENCIL8, - "Apple silicon does not support GPU_DEPTH24_S8"); + if (this->format_ == GPU_DEPTH24_STENCIL8) { + /* Apple Silicon requires GPU_DEPTH32F_STENCIL8 instead of GPU_DEPTH24_STENCIL8. */ + this->format_ = GPU_DEPTH32F_STENCIL8; + } + + MTLPixelFormat mtl_format = gpu_texture_format_to_metal(this->format_); + mtl_max_mips_ = 1; + mipmaps_ = 0; + this->mip_range_set(0, 0); + + /* Create texture from GPUVertBuf's buffer. */ + MTLVertBuf *mtl_vbo = static_cast<MTLVertBuf *>(unwrap(vbo)); + mtl_vbo->bind(); + mtl_vbo->flag_used(); + + /* Get Metal Buffer. */ + id<MTLBuffer> source_buffer = mtl_vbo->get_metal_buffer(); + BLI_assert(source_buffer); + + /* Verify size. */ + if (w_ <= 0) { + MTL_LOG_WARNING("Allocating texture buffer of width 0!\n"); + w_ = 1; + } + + /* Verify Texture and vertex buffer alignment. */ + int bytes_per_pixel = get_mtl_format_bytesize(mtl_format); + int bytes_per_row = bytes_per_pixel * w_; + + MTLContext *mtl_ctx = MTLContext::get(); + uint32_t align_requirement = static_cast<uint32_t>( + [mtl_ctx->device minimumLinearTextureAlignmentForPixelFormat:mtl_format]); + + /* Verify per-vertex size aligns with texture size. */ + const GPUVertFormat *format = GPU_vertbuf_get_format(vbo); + BLI_assert(bytes_per_pixel == format->stride && + "Pixel format stride MUST match the texture format stride -- These being different " + "is likely caused by Metal's VBO padding to a minimum of 4-bytes per-vertex"); + UNUSED_VARS_NDEBUG(format); + + /* Create texture descriptor. */ + BLI_assert(type_ == GPU_TEXTURE_BUFFER); + texture_descriptor_ = [[MTLTextureDescriptor alloc] init]; + texture_descriptor_.pixelFormat = mtl_format; + texture_descriptor_.textureType = MTLTextureTypeTextureBuffer; + texture_descriptor_.width = w_; + texture_descriptor_.height = 1; + texture_descriptor_.depth = 1; + texture_descriptor_.arrayLength = 1; + texture_descriptor_.mipmapLevelCount = mtl_max_mips_; + texture_descriptor_.usage = + MTLTextureUsageShaderRead | MTLTextureUsageShaderWrite | + MTLTextureUsagePixelFormatView; /* TODO(Metal): Optimize usage flags. */ + texture_descriptor_.storageMode = [source_buffer storageMode]; + texture_descriptor_.sampleCount = 1; + texture_descriptor_.cpuCacheMode = [source_buffer cpuCacheMode]; + texture_descriptor_.hazardTrackingMode = [source_buffer hazardTrackingMode]; + + texture_ = [source_buffer + newTextureWithDescriptor:texture_descriptor_ + offset:0 + bytesPerRow:ceil_to_multiple_u(bytes_per_row, align_requirement)]; + aligned_w_ = bytes_per_row / bytes_per_pixel; + + BLI_assert(texture_); + texture_.label = [NSString stringWithUTF8String:this->get_name()]; + is_baked_ = true; + is_dirty_ = false; + resource_mode_ = MTL_TEXTURE_MODE_VBO; + + /* Track Status. */ + vert_buffer_ = mtl_vbo; + vert_buffer_mtl_ = source_buffer; + /* Cleanup. */ + [texture_descriptor_ release]; + texture_descriptor_ = nullptr; return true; } @@ -1522,7 +1580,6 @@ bool gpu::MTLTexture::texture_is_baked() /* Prepare texture parameters after initialization, but before baking. */ void gpu::MTLTexture::prepare_internal() { - /* Derive implicit usage flags for Depth/Stencil attachments. */ if (format_flag_ & GPU_FORMAT_DEPTH || format_flag_ & GPU_FORMAT_STENCIL) { gpu_image_usage_flags_ |= GPU_TEXTURE_USAGE_ATTACHMENT; @@ -1687,7 +1744,7 @@ void gpu::MTLTexture::ensure_baked() /* Determine Resource Mode. */ resource_mode_ = MTL_TEXTURE_MODE_DEFAULT; - /* Create texture. */ + /* Standard texture allocation. */ texture_ = [ctx->device newTextureWithDescriptor:texture_descriptor_]; [texture_descriptor_ release]; diff --git a/source/blender/gpu/metal/mtl_texture_util.mm b/source/blender/gpu/metal/mtl_texture_util.mm index 5ed7659f260..33a62e2e3ef 100644 --- a/source/blender/gpu/metal/mtl_texture_util.mm +++ b/source/blender/gpu/metal/mtl_texture_util.mm @@ -34,7 +34,7 @@ MTLPixelFormat gpu_texture_format_to_metal(eGPUTextureFormat tex_format) { switch (tex_format) { - /* Formats texture & renderbuffer. */ + /* Formats texture & render-buffer. */ case GPU_RGBA8UI: return MTLPixelFormatRGBA8Uint; case GPU_RGBA8I: |