Files
test/source/blender/gpu/metal/mtl_texture.hh
Thomas Dinges cc8ea6ac67 Metal: MTLShader and MTLShaderGenerator implementation.
Full support for translation and compilation of shaders in Metal, using
GPUShaderCreateInfo. Includes render pipeline state creation and management,
enabling all standard GPU viewport rendering features in Metal.

Authored by Apple: Michael Parkin-White, Marco Giordano

Ref T96261

Reviewed By: fclem

Maniphest Tasks: T96261

Differential Revision: https://developer.blender.org/D15563
2022-09-01 22:28:40 +02:00

584 lines
19 KiB
C++

/* SPDX-License-Identifier: GPL-2.0-or-later */
/** \file
* \ingroup gpu
*/
#pragma once
#include <Cocoa/Cocoa.h>
#include <Metal/Metal.h>
#include <QuartzCore/QuartzCore.h>
#include "BLI_assert.h"
#include "MEM_guardedalloc.h"
#include "gpu_texture_private.hh"
#include "BLI_map.hh"
#include "GPU_texture.h"
#include <mutex>
#include <thread>
@class CAMetalLayer;
@class MTLCommandQueue;
@class MTLRenderPipelineState;
struct GPUFrameBuffer;
/* Texture Update system structs. */
struct TextureUpdateRoutineSpecialisation {
/* The METAL type of data in input array, e.g. half, float, short, int */
std::string input_data_type;
/* The type of the texture data texture2d<T,..>, e.g. T=float, half, int etc. */
std::string output_data_type;
/* Number of image channels provided in input texture data array (min=1, max=4). */
int component_count_input;
/* Number of channels the destination texture has (min=1, max=4). */
int component_count_output;
bool operator==(const TextureUpdateRoutineSpecialisation &other) const
{
return ((input_data_type == other.input_data_type) &&
(output_data_type == other.output_data_type) &&
(component_count_input == other.component_count_input) &&
(component_count_output == other.component_count_output));
}
uint64_t hash() const
{
blender::DefaultHash<std::string> 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));
}
};
/* Type of data is being written to the depth target:
* 0 = floating point (0.0 - 1.0)
* 1 = 24 bit integer (0 - 2^24)
* 2 = 32 bit integer (0 - 2^32) */
typedef enum {
MTL_DEPTH_UPDATE_MODE_FLOAT = 0,
MTL_DEPTH_UPDATE_MODE_INT24 = 1,
MTL_DEPTH_UPDATE_MODE_INT32 = 2
} DepthTextureUpdateMode;
struct DepthTextureUpdateRoutineSpecialisation {
DepthTextureUpdateMode data_mode;
bool operator==(const DepthTextureUpdateRoutineSpecialisation &other) const
{
return ((data_mode == other.data_mode));
}
uint64_t hash() const
{
return (uint64_t)(this->data_mode);
}
};
/* Texture Read system structs. */
struct TextureReadRoutineSpecialisation {
std::string input_data_type;
std::string output_data_type;
int component_count_input;
int component_count_output;
/* Format for depth data.
* 0 = Not a Depth format,
* 1 = FLOAT DEPTH,
* 2 = 24Bit Integer Depth,
* 4 = 32bit Unsigned-Integer Depth. */
int depth_format_mode;
bool operator==(const TextureReadRoutineSpecialisation &other) const
{
return ((input_data_type == other.input_data_type) &&
(output_data_type == other.output_data_type) &&
(component_count_input == other.component_count_input) &&
(component_count_output == other.component_count_output) &&
(depth_format_mode == other.depth_format_mode));
}
uint64_t hash() const
{
blender::DefaultHash<std::string> 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 +
(this->depth_format_mode << 28)));
}
};
namespace blender::gpu {
class MTLContext;
class MTLVertBuf;
/* Metal Texture internal implementation. */
static const int MTL_MAX_MIPMAP_COUNT = 15; /* Max: 16384x16384 */
static const int MTL_MAX_FBO_ATTACHED = 16;
/* Samplers */
struct MTLSamplerState {
eGPUSamplerState state;
/* Mip min and mip max on sampler state always the same.
* Level range now controlled with textureView to be consistent with GL baseLevel. */
bool operator==(const MTLSamplerState &other) const
{
/* Add other parameters as needed. */
return (this->state == other.state);
}
operator uint() const
{
return (uint)state;
}
operator uint64_t() const
{
return (uint64_t)state;
}
};
const MTLSamplerState DEFAULT_SAMPLER_STATE = {GPU_SAMPLER_DEFAULT /*, 0, 9999*/};
class MTLTexture : public Texture {
friend class MTLContext;
friend class MTLStateManager;
friend class MTLFrameBuffer;
private:
/* Where the textures data comes from. */
enum {
MTL_TEXTURE_MODE_DEFAULT, /* Texture is self-initialized (Standard). */
MTL_TEXTURE_MODE_EXTERNAL, /* Texture source from external id<MTLTexture> handle */
MTL_TEXTURE_MODE_VBO, /* Texture source initialized from VBO */
MTL_TEXTURE_MODE_TEXTURE_VIEW /* Texture is a view into an existing texture. */
} resource_mode_;
/* 'baking' refers to the generation of GPU-backed resources. This flag ensures GPU resources are
* ready. Baking is generally deferred until as late as possible, to ensure all associated
* resource state has been specified up-front. */
bool is_baked_;
MTLTextureDescriptor *texture_descriptor_;
id<MTLTexture> texture_;
MTLTextureUsage usage_;
/* Texture Storage. */
id<MTLBuffer> texture_buffer_;
uint aligned_w_ = 0;
/* Blit Frame-buffer. */
GPUFrameBuffer *blit_fb_ = nullptr;
uint blit_fb_slice_ = 0;
uint blit_fb_mip_ = 0;
/* Texture view properties */
/* In Metal, we use texture views to either limit mipmap ranges,
* , apply a swizzle mask, or both.
*
* We apply the mip limit in the view rather than in the sampler, as
* certain effects and functionality such as textureSize rely on the base level
* being modified.
*
* Texture views can also point to external textures, rather than the owned
* texture if MTL_TEXTURE_MODE_TEXTURE_VIEW is used.
* If this mode is used, source_texture points to a GPUTexture from which
* we pull their texture handle as a root.
*/
const GPUTexture *source_texture_ = nullptr;
enum TextureViewDirtyState {
TEXTURE_VIEW_NOT_DIRTY = 0,
TEXTURE_VIEW_SWIZZLE_DIRTY = (1 << 0),
TEXTURE_VIEW_MIP_DIRTY = (1 << 1)
};
id<MTLTexture> mip_swizzle_view_;
char tex_swizzle_mask_[4];
MTLTextureSwizzleChannels mtl_swizzle_mask_;
bool mip_range_dirty_ = false;
int mip_texture_base_level_ = 0;
int mip_texture_max_level_ = 1000;
int mip_texture_base_layer_ = 0;
int texture_view_dirty_flags_ = TEXTURE_VIEW_NOT_DIRTY;
/* Max mip-maps for currently allocated texture resource. */
int mtl_max_mips_ = 1;
/* VBO. */
MTLVertBuf *vert_buffer_;
id<MTLBuffer> vert_buffer_mtl_;
int vert_buffer_offset_;
/* Core parameters and sub-resources. */
eGPUTextureUsage gpu_image_usage_flags_;
/* Whether the texture's properties or state has changed (e.g. mipmap range), and re-baking of
* GPU resource is required. */
bool is_dirty_;
bool is_bound_;
public:
MTLTexture(const char *name);
MTLTexture(const char *name,
eGPUTextureFormat format,
eGPUTextureType type,
id<MTLTexture> metal_texture);
~MTLTexture();
void update_sub(
int mip, int offset[3], int extent[3], eGPUDataFormat type, const void *data) override;
void generate_mipmap() override;
void copy_to(Texture *dst) override;
void clear(eGPUDataFormat format, const void *data) override;
void swizzle_set(const char swizzle_mask[4]) override;
void stencil_texture_mode_set(bool use_stencil) override{
/* TODO(Metal): implement. */
};
void mip_range_set(int min, int max) override;
void *read(int mip, eGPUDataFormat type) override;
/* Remove once no longer required -- will just return 0 for now in MTL path*/
uint gl_bindcode_get() const override;
bool texture_is_baked();
const char *get_name()
{
return name_;
}
protected:
bool init_internal() override;
bool init_internal(GPUVertBuf *vbo) override;
bool init_internal(const GPUTexture *src,
int mip_offset,
int layer_offset) override; /* Texture View */
private:
/* Common Constructor, default initialization. */
void mtl_texture_init();
/* Post-construction and member initialization, prior to baking.
* Called during init_internal */
void prepare_internal();
/* Generate Metal GPU resources and upload data if needed */
void ensure_baked();
/* Delete associated Metal GPU resources. */
void reset();
void ensure_mipmaps(int miplvl);
/* Flags a given mip level as being used. */
void add_subresource(uint level);
void read_internal(int mip,
int x_off,
int y_off,
int z_off,
int width,
int height,
int depth,
eGPUDataFormat desired_output_format,
int num_output_components,
int debug_data_size,
void *r_data);
void bake_mip_swizzle_view();
id<MTLTexture> get_metal_handle();
id<MTLTexture> get_metal_handle_base();
MTLSamplerState get_sampler_state();
void blit(id<MTLBlitCommandEncoder> blit_encoder,
uint src_x_offset,
uint src_y_offset,
uint src_z_offset,
uint src_slice,
uint src_mip,
gpu::MTLTexture *dest,
uint dst_x_offset,
uint dst_y_offset,
uint dst_z_offset,
uint dst_slice,
uint dst_mip,
uint width,
uint height,
uint depth);
void blit(gpu::MTLTexture *dest,
uint src_x_offset,
uint src_y_offset,
uint dst_x_offset,
uint dst_y_offset,
uint src_mip,
uint dst_mip,
uint dst_slice,
int width,
int height);
GPUFrameBuffer *get_blit_framebuffer(uint dst_slice, uint dst_mip);
MEM_CXX_CLASS_ALLOC_FUNCS("gpu::MTLTexture")
/* Texture Update function Utilities. */
/* Metal texture updating does not provide the same range of functionality for type conversion
* and format compatibility as are available in OpenGL. To achieve the same level of
* functionality, we need to instead use compute kernels to perform texture data conversions
* where appropriate.
* There are a number of different inputs which affect permutations and thus require different
* shaders and PSOs, such as:
* - Texture format
* - Texture type (e.g. 2D, 3D, 2D Array, Depth etc;)
* - Source data format and component count (e.g. floating point)
*
* MECHANISM:
*
* blender::map<INPUT DEFINES STRUCT, compute PSO> update_2d_array_kernel_psos;
* - Generate compute shader with configured kernel below with variable parameters depending
* on input/output format configurations. Do not need to keep source or descriptors around,
* just PSO, as same input defines will always generate the same code.
*
* - IF datatype IS an exact match e.g. :
* - Per-component size matches (e.g. GPU_DATA_UBYTE)
* OR GPU_DATA_10_11_11_REV && GPU_R11G11B10 (equiv)
* OR D24S8 and GPU_DATA_UINT_24_8
* We can use BLIT ENCODER.
*
* OTHERWISE TRIGGER COMPUTE:
* - Compute sizes will vary. Threads per grid WILL match 'extent'.
* Dimensions will vary depending on texture type.
* - Will use setBytes with 'TextureUpdateParams' struct to pass in useful member params.
*/
struct TextureUpdateParams {
int mip_index;
int extent[3]; /* Width, Height, Slice on 2D Array tex*/
int offset[3]; /* Width, Height, Slice on 2D Array tex*/
uint unpack_row_length; /* Number of pixels between bytes in input data */
};
id<MTLComputePipelineState> texture_update_1d_get_kernel(
TextureUpdateRoutineSpecialisation specialization);
id<MTLComputePipelineState> texture_update_1d_array_get_kernel(
TextureUpdateRoutineSpecialisation specialization);
id<MTLComputePipelineState> texture_update_2d_get_kernel(
TextureUpdateRoutineSpecialisation specialization);
id<MTLComputePipelineState> texture_update_2d_array_get_kernel(
TextureUpdateRoutineSpecialisation specialization);
id<MTLComputePipelineState> texture_update_3d_get_kernel(
TextureUpdateRoutineSpecialisation specialization);
id<MTLComputePipelineState> mtl_texture_update_impl(
TextureUpdateRoutineSpecialisation specialization_params,
blender::Map<TextureUpdateRoutineSpecialisation, id<MTLComputePipelineState>>
&specialization_cache,
eGPUTextureType texture_type);
/* Depth Update Utilities */
/* Depth texture updates are not directly supported with Blit operations, similarly, we cannot
* use a compute shader to write to depth, so we must instead render to a depth target.
* These processes use vertex/fragment shaders to render texture data from an intermediate
* source, in order to prime the depth buffer*/
GPUShader *depth_2d_update_sh_get(DepthTextureUpdateRoutineSpecialisation specialization);
void update_sub_depth_2d(
int mip, int offset[3], int extent[3], eGPUDataFormat type, const void *data);
/* Texture Read function utilities -- Follows a similar mechanism to the updating routines */
struct TextureReadParams {
int mip_index;
int extent[3]; /* Width, Height, Slice on 2D Array tex*/
int offset[3]; /* Width, Height, Slice on 2D Array tex*/
};
id<MTLComputePipelineState> texture_read_1d_get_kernel(
TextureReadRoutineSpecialisation specialization);
id<MTLComputePipelineState> texture_read_1d_array_get_kernel(
TextureReadRoutineSpecialisation specialization);
id<MTLComputePipelineState> texture_read_2d_get_kernel(
TextureReadRoutineSpecialisation specialization);
id<MTLComputePipelineState> texture_read_2d_array_get_kernel(
TextureReadRoutineSpecialisation specialization);
id<MTLComputePipelineState> texture_read_3d_get_kernel(
TextureReadRoutineSpecialisation specialization);
id<MTLComputePipelineState> mtl_texture_read_impl(
TextureReadRoutineSpecialisation specialization_params,
blender::Map<TextureReadRoutineSpecialisation, id<MTLComputePipelineState>>
&specialization_cache,
eGPUTextureType texture_type);
/* fullscreen blit utilities. */
GPUShader *fullscreen_blit_sh_get();
};
/* Utility */
MTLPixelFormat gpu_texture_format_to_metal(eGPUTextureFormat tex_format);
int get_mtl_format_bytesize(MTLPixelFormat tex_format);
int get_mtl_format_num_components(MTLPixelFormat tex_format);
bool mtl_format_supports_blending(MTLPixelFormat format);
/* The type used to define the per-component data in the input buffer. */
inline std::string tex_data_format_to_msl_type_str(eGPUDataFormat type)
{
switch (type) {
case GPU_DATA_FLOAT:
return "float";
case GPU_DATA_HALF_FLOAT:
return "half";
case GPU_DATA_INT:
return "int";
case GPU_DATA_UINT:
return "uint";
case GPU_DATA_UBYTE:
return "uchar";
case GPU_DATA_UINT_24_8:
return "uint"; /* Problematic type - but will match alignment. */
case GPU_DATA_10_11_11_REV:
return "float"; /* Problematic type - each component will be read as a float. */
default:
BLI_assert(false);
break;
}
return "";
}
/* The type T which goes into texture2d<T, access>. */
inline std::string tex_data_format_to_msl_texture_template_type(eGPUDataFormat type)
{
switch (type) {
case GPU_DATA_FLOAT:
return "float";
case GPU_DATA_HALF_FLOAT:
return "half";
case GPU_DATA_INT:
return "int";
case GPU_DATA_UINT:
return "uint";
case GPU_DATA_UBYTE:
return "ushort";
case GPU_DATA_UINT_24_8:
return "uint"; /* Problematic type. */
case GPU_DATA_10_11_11_REV:
return "float"; /* Problematic type. */
default:
BLI_assert(false);
break;
}
return "";
}
/* Determine whether format is writable or not. Use mtl_format_get_writeable_view_format(..) for
* these. */
inline bool mtl_format_is_writable(MTLPixelFormat format)
{
switch (format) {
case MTLPixelFormatRGBA8Unorm_sRGB:
case MTLPixelFormatBGRA8Unorm_sRGB:
case MTLPixelFormatDepth16Unorm:
case MTLPixelFormatDepth32Float:
case MTLPixelFormatDepth32Float_Stencil8:
case MTLPixelFormatBGR10A2Unorm:
case MTLPixelFormatDepth24Unorm_Stencil8:
return false;
default:
return true;
}
return true;
}
/* For the cases where a texture format is unwritable, we can create a texture view of a similar
* format */
inline MTLPixelFormat mtl_format_get_writeable_view_format(MTLPixelFormat format)
{
switch (format) {
case MTLPixelFormatRGBA8Unorm_sRGB:
return MTLPixelFormatRGBA8Unorm;
case MTLPixelFormatBGRA8Unorm_sRGB:
return MTLPixelFormatBGRA8Unorm;
case MTLPixelFormatDepth16Unorm:
return MTLPixelFormatR16Unorm;
case MTLPixelFormatDepth32Float:
return MTLPixelFormatR32Float;
case MTLPixelFormatDepth32Float_Stencil8:
/* return MTLPixelFormatRG32Float; */
/* No alternative mirror format. This should not be used for
* manual data upload */
return MTLPixelFormatInvalid;
case MTLPixelFormatBGR10A2Unorm:
/* return MTLPixelFormatBGRA8Unorm; */
/* No alternative mirror format. This should not be used for
* manual data upload */
return MTLPixelFormatInvalid;
case MTLPixelFormatDepth24Unorm_Stencil8:
/* No direct format, but we'll just mirror the bytes -- Uint
* should ensure bytes are not re-normalized or manipulated */
/* return MTLPixelFormatR32Uint; */
return MTLPixelFormatInvalid;
default:
return format;
}
return format;
}
/* Returns the associated engine data type with a given texture:
* Definitely not complete, edit according to the METAL specification. */
inline eGPUDataFormat to_mtl_internal_data_format(eGPUTextureFormat tex_format)
{
switch (tex_format) {
case GPU_RGBA8:
case GPU_RGBA32F:
case GPU_RGBA16F:
case GPU_RGBA16:
case GPU_RG8:
case GPU_RG32F:
case GPU_RG16F:
case GPU_RG16:
case GPU_R8:
case GPU_R32F:
case GPU_R16F:
case GPU_R16:
case GPU_RGB16F:
case GPU_DEPTH_COMPONENT24:
case GPU_DEPTH_COMPONENT16:
case GPU_DEPTH_COMPONENT32F:
case GPU_SRGB8_A8:
return GPU_DATA_FLOAT;
case GPU_DEPTH24_STENCIL8:
case GPU_DEPTH32F_STENCIL8:
return GPU_DATA_UINT_24_8;
case GPU_RGBA8UI:
case GPU_RGBA32UI:
case GPU_RGBA16UI:
case GPU_RG8UI:
case GPU_RG32UI:
case GPU_R8UI:
case GPU_R16UI:
case GPU_RG16UI:
case GPU_R32UI:
return GPU_DATA_UINT;
case GPU_R8I:
case GPU_RG8I:
case GPU_R16I:
case GPU_R32I:
case GPU_RG16I:
case GPU_RGBA8I:
case GPU_RGBA32I:
case GPU_RGBA16I:
case GPU_RG32I:
return GPU_DATA_INT;
case GPU_R11F_G11F_B10F:
return GPU_DATA_10_11_11_REV;
default:
BLI_assert(false && "Texture not yet handled");
return GPU_DATA_FLOAT;
}
}
} // namespace blender::gpu