[GPU/WGF] Use ByteAddressBuffer in precompiled shaders
Drivers and hardware may provide more optimal paths for untyped buffers compared to typed. Also, ByteAddressBuffer may be bound via a root descriptor in Direct3D 12, greatly simplifying the binding logic. ByteAddressBuffer also doesn't have the 128 * 2^20 element count limit that typed and structured buffers have, and doesn't require the structure stride to be specified at resource creation time in Direct3D 11.
This commit is contained in:
Triang3l
@@ -281,70 +281,70 @@ float4_xe float_x4_xe(float x) { return float4_xe(x, x, x, x); }
|
||||
#endif // SHADING_LANGUAGE_*_XE
|
||||
#ifndef entry_outputs_begin_xe
|
||||
#define entry_outputs_begin_xe
|
||||
#endif // !entry_outputs_begin_xe
|
||||
#endif
|
||||
#ifndef entry_out_position_xe
|
||||
#define entry_out_position_xe
|
||||
#endif // !entry_out_position_xe
|
||||
#endif
|
||||
#ifndef entry_outputs_end_stage_inputs_begin_xe
|
||||
#define entry_outputs_end_stage_inputs_begin_xe
|
||||
#endif // !entry_outputs_end_stage_inputs_begin_xe
|
||||
#endif
|
||||
#ifndef entry_in_stage_vertex_xe
|
||||
#define entry_in_stage_vertex_xe(type, name, index, semantic) \
|
||||
entry_in_stage_xe(type, name, index, semantic)
|
||||
#endif // !entry_in_stage_vertex_xe
|
||||
#endif
|
||||
#ifndef entry_stage_inputs_end_bindings_begin_vertex_xe
|
||||
#define entry_stage_inputs_end_bindings_begin_vertex_xe
|
||||
#endif // !entry_stage_inputs_end_bindings_begin_vertex_xe
|
||||
#endif
|
||||
#ifndef entry_stage_inputs_end_bindings_begin_pixel_xe
|
||||
#define entry_stage_inputs_end_bindings_begin_pixel_xe
|
||||
#endif // !entry_stage_inputs_end_bindings_begin_pixel_xe
|
||||
#endif
|
||||
#ifndef entry_bindings_begin_compute_xe
|
||||
#define entry_bindings_begin_compute_xe
|
||||
#endif // !entry_bindings_begin_compute_xe
|
||||
#endif
|
||||
#ifndef entry_binding_next_xe
|
||||
#define entry_binding_next_xe
|
||||
#endif // !entry_binding_next_xe
|
||||
#endif
|
||||
#ifndef entry_bindings_end_inputs_begin_xe
|
||||
#define entry_bindings_end_inputs_begin_xe
|
||||
#endif // !entry_bindings_end_inputs_begin_xe
|
||||
#endif
|
||||
#ifndef entry_bindings_empty_end_inputs_begin_xe
|
||||
#define entry_bindings_empty_end_inputs_begin_xe
|
||||
#endif // !entry_bindings_empty_end_inputs_begin_xe
|
||||
#endif
|
||||
#ifndef entry_bindings_end_inputs_begin_compute_xe
|
||||
#define entry_bindings_end_inputs_begin_compute_xe
|
||||
#endif // !entry_bindings_end_inputs_begin_compute_xe
|
||||
#endif
|
||||
#ifndef entry_input_next_xe
|
||||
#define entry_input_next_xe
|
||||
#endif // !entry_input_next_xe
|
||||
#endif
|
||||
#ifndef entry_in_stage_inputs_xe
|
||||
#define entry_in_stage_inputs_xe
|
||||
#endif // !entry_in_stage_inputs_xe
|
||||
#endif
|
||||
#ifndef entry_in_vertex_id_xe
|
||||
#define entry_in_vertex_id_xe
|
||||
#endif // !entry_in_vertex_id_xe
|
||||
#endif
|
||||
#ifndef entry_in_pixel_coord_xe
|
||||
#define entry_in_pixel_coord_xe
|
||||
#endif // !entry_in_pixel_coord_xe
|
||||
#endif
|
||||
#ifndef entry_in_group_id_xe
|
||||
#define entry_in_group_id_xe
|
||||
#endif // !entry_in_group_id_xe
|
||||
#endif
|
||||
#ifndef entry_in_local_thread_id_xe
|
||||
#define entry_in_local_thread_id_xe
|
||||
#endif // !entry_in_local_thread_id_xe
|
||||
#endif
|
||||
#ifndef entry_in_global_thread_id_xe
|
||||
#define entry_in_global_thread_id_xe
|
||||
#endif // !entry_in_global_thread_id_xe
|
||||
#endif
|
||||
#ifndef entry_in_local_thread_index_xe
|
||||
#define entry_in_local_thread_index_xe
|
||||
#endif // !entry_in_local_thread_index_xe
|
||||
#endif
|
||||
#ifndef entry_code_end_xe
|
||||
#define entry_code_end_xe \
|
||||
entry_return_xe \
|
||||
}
|
||||
#endif // !entry_code_end_xe
|
||||
#endif
|
||||
#ifndef entry_code_end_compute_xe
|
||||
#define entry_code_end_compute_xe }
|
||||
#endif // !entry_code_end_compute_xe
|
||||
#endif
|
||||
|
||||
// NDC_DIRECTION_Y_XE, assuming a positive viewport height, is:
|
||||
// * 1.0 if +out_position_xe.y is towards +in_pixel_coord_xe.y,
|
||||
@@ -353,7 +353,7 @@ float4_xe float_x4_xe(float x) { return float4_xe(x, x, x, x); }
|
||||
#define NDC_DIRECTION_Y_XE 1.0f
|
||||
#else
|
||||
#define NDC_DIRECTION_Y_XE -1.0f
|
||||
#endif // SHADING_LANGUAGE_GLSL_XE
|
||||
#endif
|
||||
|
||||
#if SHADING_LANGUAGE_GLSL_XE
|
||||
// GLSL requires just const for declaring a constant in the global scope.
|
||||
@@ -461,140 +461,214 @@ float4_xe float_x4_xe(float x) { return float4_xe(x, x, x, x); }
|
||||
#endif // SHADING_LANGUAGE_*_XE
|
||||
#ifndef const_buffer_binding_xe
|
||||
#define const_buffer_binding_xe(name, msl_buffer)
|
||||
#endif // !const_buffer_binding_xe
|
||||
#endif
|
||||
#ifndef push_const_binding_xe
|
||||
#define push_const_binding_xe(msl_buffer)
|
||||
#endif // !push_const_binding_xe
|
||||
#endif
|
||||
|
||||
// Declarations of typed storage buffers and dword buffers (_declare) must be
|
||||
// outside the entry point, but their bindings (_binding) must also be specified
|
||||
// in the entry point binding declarations.
|
||||
// Byte buffers represent raw data accessible as `uint` vectors, addressed with
|
||||
// a byte offset with some alignment requirement, and implemented as storage
|
||||
// buffers on GLSL, byte address buffers on HLSL, and buffer pointers on MSL.
|
||||
//
|
||||
// An array buffer is a buffer limited to 1/2/4-component vectors of 32-bit
|
||||
// integers and floats, a typed buffer on Direct3D, but a storage buffer (as
|
||||
// opposed to a texel buffer, which has a very small minimum requirement for the
|
||||
// maximum size) on Vulkan.
|
||||
// These underlying binding types were chosen primarily because of the high
|
||||
// range limits they offer, as well as the ability for the compiler and the GPU
|
||||
// to take more optimal access paths as it's known that no type conversion would
|
||||
// be necessary.
|
||||
//
|
||||
// A UInt vector buffer is a buffer containing 32-bit values, but loading or
|
||||
// storing may be done for 2, 3 or 4 consecutive values that are still
|
||||
// 32-bit-aligned, and depending on the language and hardware support, access
|
||||
// of multiple elements may or may not be compiled into a single hardware
|
||||
// instruction instead of separate accesses of individual elements. Each index
|
||||
// value corresponds to a 32-bit element. Implementations for languages without
|
||||
// native support must use functions, not macros, for adding the component
|
||||
// offset to the index to avoid evaluating the address multiple times.
|
||||
// In Vulkan, the minimum required `maxStorageBufferRange` is 128 MB, while only
|
||||
// 65536 is required for `maxTexelBufferElements`.
|
||||
//
|
||||
// In Direct3D 11 and 12, the range limit for typed buffers is 2^27 elements
|
||||
// (`REQ_BUFFER_RESOURCE_TEXEL_COUNT_2_TO_EXP`), and according to "Summary of
|
||||
// Changes in this Chapter from D3D10 to D3D11.3" in the chapter 21 "System
|
||||
// Limits on Various Resources" in the Direct3D 11.3 Functional Specification,
|
||||
// this limit also applies to structured buffers. However, the entire resource
|
||||
// (up to 2 GB - `REQ_RESOURCE_SIZE_IN_MEGABYTES_EXPRESSION_C_TERM` - in most
|
||||
// cases) must be accessible via a byte address buffer.
|
||||
//
|
||||
// Also, byte address buffers were chosen over structured buffers because 8-byte
|
||||
// and 16-byte accesses are common, and the same buffer may need to be accessed
|
||||
// with different granularities, however, in Direct3D 11, one buffer can't be
|
||||
// accessed with different structure byte strides.
|
||||
//
|
||||
// Bounds checking is not guaranteed for byte buffers (may be bound via root
|
||||
// descriptors on Direct3D 12, and on MSL they're bound as pointers).
|
||||
//
|
||||
// For 4-aligned byte buffers, loads larger than the alignment of the buffer are
|
||||
// supported, with the `u` load expecting the address to be aligned to the
|
||||
// buffer alignment, and the `a` load expecting it to be aligned to the size of
|
||||
// the loaded data.
|
||||
//
|
||||
// Declarations of byte buffers (`_declare`) must be outside the entry point,
|
||||
// but their bindings (`_binding`) must also be specified in the entry point
|
||||
// function signature. This is necessary so byte buffers can be declared and
|
||||
// used in header files and in functions declared in them, as they're
|
||||
// implemented as buffer blocks in GLSL.
|
||||
//
|
||||
// The `_impl` definitions are for internal use in this file.
|
||||
|
||||
#if SHADING_LANGUAGE_GLSL_XE
|
||||
// Binding declarations.
|
||||
#define array_buffer_declare_xe(value_type, name, glsl_set, glsl_binding, \
|
||||
hlsl_t, hlsl_t_space) \
|
||||
layout(std430, glsl_set, glsl_binding) \
|
||||
readonly buffer name##_xe_block { \
|
||||
value_type data[]; \
|
||||
#define byte_buffer_declare_xe_impl(value_type, name, glsl_set, \
|
||||
glsl_binding, hlsl_t, hlsl_t_space) \
|
||||
layout(std430, glsl_set, glsl_binding) \
|
||||
readonly buffer name##_xe_block { \
|
||||
value_type data[]; \
|
||||
} name;
|
||||
#define array_buffer_wo_declare_xe(value_type, name, glsl_set, glsl_binding, \
|
||||
hlsl_u, hlsl_u_space) \
|
||||
layout(std430, glsl_set, glsl_binding) \
|
||||
writeonly buffer name##_xe_block { \
|
||||
value_type data[]; \
|
||||
#define byte_buffer_wo_declare_xe_impl(value_type, name, glsl_set, \
|
||||
glsl_binding, hlsl_t, hlsl_t_space) \
|
||||
layout(std430, glsl_set, glsl_binding) \
|
||||
writeonly buffer name##_xe_block { \
|
||||
value_type data[]; \
|
||||
} name;
|
||||
#define uint_vector_buffer_declare_xe(name, glsl_set, glsl_binding, hlsl_t, \
|
||||
hlsl_t_space) \
|
||||
layout(std430, glsl_set, glsl_binding) \
|
||||
readonly buffer name##_xe_block { \
|
||||
uint data[]; \
|
||||
} name; \
|
||||
uint2_xe name##_xe_load2(uint position) { \
|
||||
return uint2_xe(name.data[position], name.data[position + 1u]); \
|
||||
} \
|
||||
uint3_xe name##_xe_load3(uint position) { \
|
||||
return uint3_xe(name.data[position], name.data[position + 1u], \
|
||||
name.data[position + 2u]); \
|
||||
} \
|
||||
uint4_xe name##_xe_load4(uint position) { \
|
||||
return uint4_xe(name.data[position], name.data[position + 1u], \
|
||||
name.data[position + 2u], name.data[position + 3u]); \
|
||||
}
|
||||
// Loading and storing.
|
||||
#define array_buffer_load_xe(name, position) ((name).data[uint(position)])
|
||||
#define array_buffer_store_xe(name, position, value) \
|
||||
((name).data[uint(position)] = (value))
|
||||
#define uint_vector_buffer_load1_xe(name, position) \
|
||||
((name).data[uint(position)])
|
||||
#define uint_vector_buffer_load2_xe(name, position) \
|
||||
name##_xe_load2(uint(position))
|
||||
#define uint_vector_buffer_load3_xe(name, position) \
|
||||
name##_xe_load3(uint(position))
|
||||
#define uint_vector_buffer_load4_xe(name, position) \
|
||||
name##_xe_load4(uint(position))
|
||||
|
||||
#define byte_buffer_align4_load4_xe(name, byte_address) \
|
||||
((name).data[uint(byte_address) >> 2])
|
||||
#define byte_buffer_align4_load8u_xe(name, byte_address) \
|
||||
uint2_xe((name).data[uint(byte_address) >> 2], \
|
||||
(name).data[(uint(byte_address) >> 2) + 1u])
|
||||
#define byte_buffer_align4_load16u_xe(name, byte_address) \
|
||||
uint4_xe((name).data[uint(byte_address) >> 2], \
|
||||
(name).data[(uint(byte_address) >> 2) + 1u], \
|
||||
(name).data[(uint(byte_address) >> 2) + 2u], \
|
||||
(name).data[(uint(byte_address) >> 2) + 3u])
|
||||
|
||||
#define byte_buffer_align8_load8_xe(name, byte_address) \
|
||||
((name).data[uint(byte_address) >> 3])
|
||||
|
||||
#define byte_buffer_align16_load16_xe(name, byte_address) \
|
||||
((name).data[uint(byte_address) >> 4])
|
||||
|
||||
#define byte_buffer_align4_store4_xe(name, byte_address, value) \
|
||||
((name).data[uint(byte_address) >> 2] = (value))
|
||||
#define byte_buffer_align8_store8_xe(name, byte_address, value) \
|
||||
((name).data[uint(byte_address) >> 3] = (value))
|
||||
#define byte_buffer_align16_store16_xe(name, byte_address, value) \
|
||||
((name).data[uint(byte_address) >> 4] = (value))
|
||||
|
||||
#elif SHADING_LANGUAGE_HLSL_XE
|
||||
// Binding declarations.
|
||||
#define array_buffer_declare_xe(value_type, name, glsl_set, glsl_binding, \
|
||||
hlsl_t, hlsl_t_space) \
|
||||
Buffer<value_type> name : register(hlsl_t, hlsl_t_space);
|
||||
#define array_buffer_wo_declare_xe(value_type, name, glsl_set, glsl_binding, \
|
||||
hlsl_u, hlsl_u_space) \
|
||||
RWBuffer<value_type> name : register(hlsl_u, hlsl_u_space);
|
||||
#define uint_vector_buffer_declare_xe(name, glsl_set, glsl_binding, hlsl_t, \
|
||||
hlsl_t_space) \
|
||||
#define byte_buffer_declare_xe_impl(value_type, name, glsl_set, \
|
||||
glsl_binding, hlsl_t, hlsl_t_space) \
|
||||
ByteAddressBuffer name : register(hlsl_t, hlsl_t_space);
|
||||
// Loading and storing.
|
||||
#define array_buffer_load_xe(name, position) ((name)[uint(position)])
|
||||
#define array_buffer_store_xe(name, position, value) \
|
||||
((name)[uint(position)] = (value))
|
||||
#define uint_vector_buffer_load1_xe(name, position) \
|
||||
((name).Load(int(position) << 2))
|
||||
#define uint_vector_buffer_load2_xe(name, position) \
|
||||
((name).Load2(int(position) << 2))
|
||||
#define uint_vector_buffer_load3_xe(name, position) \
|
||||
((name).Load3(int(position) << 2))
|
||||
#define uint_vector_buffer_load4_xe(name, position) \
|
||||
((name).Load4(int(position) << 2))
|
||||
#define byte_buffer_wo_declare_xe_impl(value_type, name, glsl_set, \
|
||||
glsl_binding, hlsl_t, hlsl_t_space) \
|
||||
RWByteAddressBuffer name : register(hlsl_t, hlsl_t_space);
|
||||
|
||||
#define byte_buffer_align4_load4_xe(name, byte_address) \
|
||||
((name).Load(uint(byte_address)))
|
||||
#define byte_buffer_align4_load8u_xe(name, byte_address) \
|
||||
((name).Load2(uint(byte_address)))
|
||||
#define byte_buffer_align4_load16u_xe(name, byte_address) \
|
||||
((name).Load4(uint(byte_address)))
|
||||
|
||||
#define byte_buffer_align8_load8_xe(name, byte_address) \
|
||||
((name).Load2(uint(byte_address)))
|
||||
|
||||
#define byte_buffer_align16_load16_xe(name, byte_address) \
|
||||
((name).Load4(uint(byte_address)))
|
||||
|
||||
#define byte_buffer_align4_store4_xe(name, byte_address, value) \
|
||||
((name).Store(uint(byte_address), value))
|
||||
#define byte_buffer_align8_store8_xe(name, byte_address, value) \
|
||||
((name).Store2(uint(byte_address), value))
|
||||
#define byte_buffer_align16_store16_xe(name, byte_address, value) \
|
||||
((name).Store4(uint(byte_address), value))
|
||||
|
||||
#elif SHADING_LANGUAGE_MSL_XE
|
||||
// Binding declarations.
|
||||
#define array_buffer_binding_xe(value_type, name, msl_buffer) \
|
||||
const device value_type* name [[msl_buffer]]
|
||||
#define array_buffer_wo_binding_xe(value_type, name, msl_buffer) \
|
||||
device value_type* name [[msl_buffer]]
|
||||
#define uint_vector_buffer_binding_xe(name, msl_buffer) \
|
||||
const device uint* name [[msl_buffer]]
|
||||
// Loading and storing.
|
||||
#define array_buffer_load_xe(name, position) ((name)[size_t(position)])
|
||||
#define array_buffer_store_xe(name, position, value) \
|
||||
((name)[size_t(position)] = (value))
|
||||
#define uint_vector_buffer_load1_xe(name, position) ((name)[size_t(position)])
|
||||
#define uint_vector_buffer_load2_xe(name, position) \
|
||||
#define byte_buffer_binding_xe(name, msl_buffer) \
|
||||
const device char* name [[msl_buffer]]
|
||||
#define byte_buffer_wo_binding_xe(name, msl_buffer) \
|
||||
device char* name [[msl_buffer]]
|
||||
|
||||
#define byte_buffer_align4_load4_xe(name, byte_address) \
|
||||
(*reinterpret_cast<const device uint*>(&(name)[(byte_address)]))
|
||||
#define byte_buffer_align4_load8u_xe(name, byte_address) \
|
||||
uint2_xe(*reinterpret_cast<const device packed_uint2*>( \
|
||||
&((name)[size_t(position)])))
|
||||
#define uint_vector_buffer_load3_xe(name, position) \
|
||||
uint3_xe(*reinterpret_cast<const device packed_uint3*>( \
|
||||
&((name)[size_t(position)])))
|
||||
#define uint_vector_buffer_load4_xe(name, position) \
|
||||
&(name)[(byte_address)]))
|
||||
#define byte_buffer_align4_load8a_xe(name, byte_address) \
|
||||
(*reinterpret_cast<const device uint2_xe*>(&(name)[(byte_address)]))
|
||||
#define byte_buffer_align4_load16u_xe(name, byte_address) \
|
||||
uint4_xe(*reinterpret_cast<const device packed_uint4*>( \
|
||||
&((name)[size_t(position)])))
|
||||
&(name)[(byte_address)]))
|
||||
#define byte_buffer_align4_load16a_xe(name, byte_address) \
|
||||
(*reinterpret_cast<const device uint4_xe*>(&(name)[(byte_address)]))
|
||||
|
||||
#define byte_buffer_align8_load8_xe(name, byte_address) \
|
||||
(*reinterpret_cast<const device uint2_xe*>(&(name)[(byte_address)]))
|
||||
|
||||
#define byte_buffer_align16_load16_xe(name, byte_address) \
|
||||
(*reinterpret_cast<const device uint4_xe*>(&(name)[(byte_address)]))
|
||||
|
||||
#define byte_buffer_align4_store4_xe(name, byte_address, value) \
|
||||
(*reinterpret_cast<device uint*>(&(name)[(byte_address)]) = (value))
|
||||
#define byte_buffer_align8_store8_xe(name, byte_address, value) \
|
||||
(*reinterpret_cast<device uint2_xe*>(&(name)[(byte_address)]) = (value))
|
||||
#define byte_buffer_align16_store16_xe(name, byte_address, value) \
|
||||
(*reinterpret_cast<device uint4_xe*>(&(name)[(byte_address)]) = (value))
|
||||
|
||||
#else
|
||||
#error Storage buffers not defined for the target language.
|
||||
#error Byte buffers not defined for the target language.
|
||||
#endif // SHADING_LANGUAGE_*_XE
|
||||
#ifndef array_buffer_declare_xe
|
||||
#define array_buffer_declare_xe(value_type, name, glsl_set, glsl_binding, \
|
||||
|
||||
#ifndef byte_buffer_declare_xe_impl
|
||||
#define byte_buffer_declare_xe_impl(value_type, name, glsl_set, \
|
||||
glsl_binding, hlsl_t, hlsl_t_space)
|
||||
#endif
|
||||
#ifndef byte_buffer_binding_xe
|
||||
#define byte_buffer_binding_xe(name, msl_buffer)
|
||||
#endif
|
||||
#ifndef byte_buffer_wo_declare_xe_impl
|
||||
#define byte_buffer_wo_declare_xe_impl(value_type, name, glsl_set, \
|
||||
glsl_binding, hlsl_t, hlsl_t_space)
|
||||
#endif
|
||||
#ifndef byte_buffer_wo_binding_xe
|
||||
#define byte_buffer_wo_binding_xe(name, msl_buffer)
|
||||
#endif
|
||||
|
||||
#ifndef byte_buffer_align4_declare_xe
|
||||
#define byte_buffer_align4_declare_xe(name, glsl_set, glsl_binding, hlsl_t, \
|
||||
hlsl_t_space) \
|
||||
byte_buffer_declare_xe_impl(uint, name, glsl_set, glsl_binding, hlsl_t, \
|
||||
hlsl_t_space)
|
||||
#endif
|
||||
#ifndef byte_buffer_align4_wo_declare_xe
|
||||
#define byte_buffer_align4_wo_declare_xe(name, glsl_set, glsl_binding, \
|
||||
hlsl_t, hlsl_t_space) \
|
||||
byte_buffer_wo_declare_xe_impl(uint, name, glsl_set, glsl_binding, \
|
||||
hlsl_t, hlsl_t_space)
|
||||
#endif
|
||||
|
||||
#ifndef byte_buffer_align8_declare_xe
|
||||
#define byte_buffer_align8_declare_xe(name, glsl_set, glsl_binding, hlsl_t, \
|
||||
hlsl_t_space) \
|
||||
byte_buffer_declare_xe_impl(uint2_xe, name, glsl_set, glsl_binding, \
|
||||
hlsl_t, hlsl_t_space)
|
||||
#endif // !array_buffer_declare_xe
|
||||
#ifndef array_buffer_wo_declare_xe
|
||||
#define array_buffer_wo_declare_xe(value_type, name, glsl_set, glsl_binding, \
|
||||
hlsl_u, hlsl_u_space)
|
||||
#endif // !array_buffer_wo_declare_xe
|
||||
#ifndef array_buffer_binding_xe
|
||||
#define array_buffer_binding_xe(value_type, name, msl_buffer)
|
||||
#endif // !array_buffer_binding_xe
|
||||
#ifndef array_buffer_wo_binding_xe
|
||||
#define array_buffer_wo_binding_xe(value_type, name, msl_buffer)
|
||||
#endif // !array_buffer_wo_binding_xe
|
||||
#ifndef uint_vector_buffer_declare_xe
|
||||
#define uint_vector_buffer_declare_xe(name, glsl_set, glsl_binding, hlsl_t, \
|
||||
hlsl_t_space)
|
||||
#endif // !uint_vector_buffer_declare_xe
|
||||
#ifndef uint_vector_buffer_binding_xe
|
||||
#define uint_vector_buffer_binding_xe(name, msl_buffer)
|
||||
#endif // !uint_vector_buffer_binding_xe
|
||||
#endif
|
||||
#ifndef byte_buffer_align8_wo_declare_xe
|
||||
#define byte_buffer_align8_wo_declare_xe(name, glsl_set, glsl_binding, \
|
||||
hlsl_t, hlsl_t_space) \
|
||||
byte_buffer_wo_declare_xe_impl(uint2_xe, name, glsl_set, glsl_binding, \
|
||||
hlsl_t, hlsl_t_space)
|
||||
#endif
|
||||
|
||||
#ifndef byte_buffer_align16_declare_xe
|
||||
#define byte_buffer_align16_declare_xe(name, glsl_set, glsl_binding, hlsl_t, \
|
||||
hlsl_t_space) \
|
||||
byte_buffer_declare_xe_impl(uint4_xe, name, glsl_set, glsl_binding, \
|
||||
hlsl_t, hlsl_t_space)
|
||||
#endif
|
||||
#ifndef byte_buffer_align16_wo_declare_xe
|
||||
#define byte_buffer_align16_wo_declare_xe(name, glsl_set, glsl_binding, \
|
||||
hlsl_t, hlsl_t_space) \
|
||||
byte_buffer_wo_declare_xe_impl(uint4_xe, name, glsl_set, glsl_binding, \
|
||||
hlsl_t, hlsl_t_space)
|
||||
#endif
|
||||
|
||||
#ifndef byte_buffer_align4_load8a_xe
|
||||
#define byte_buffer_align4_load8a_xe byte_buffer_align4_load8u_xe
|
||||
#endif
|
||||
#ifndef byte_buffer_align4_load16a_xe
|
||||
#define byte_buffer_align4_load16a_xe byte_buffer_align4_load16u_xe
|
||||
#endif
|
||||
|
||||
// Buffer, texture, sampler and image bindings must be in the entry point
|
||||
// bindings declaration.
|
||||
@@ -741,7 +815,7 @@ float4_xe float_x4_xe(float x) { return float4_xe(x, x, x, x); }
|
||||
#if !COMBINED_TEXTURE_SAMPLER_XE
|
||||
#ifndef sampler_2d_xe
|
||||
#define sampler_2d_xe texture_2d_xe
|
||||
#endif // !sampler_2d_xe
|
||||
#endif
|
||||
#ifndef sampler_xe
|
||||
#define sampler_xe(sampler_type, name, glsl_set, glsl_binding, hlsl_t, \
|
||||
hlsl_t_space, hlsl_s, hlsl_s_space, msl_texture, \
|
||||
@@ -751,32 +825,32 @@ float4_xe float_x4_xe(float x) { return float4_xe(x, x, x, x); }
|
||||
entry_binding_next_xe \
|
||||
sampler_state_xe(name##_xe_sampler, glsl_set, glsl_binding, hlsl_s, \
|
||||
hlsl_s_space, msl_sampler)
|
||||
#endif // !sampler_xe
|
||||
#endif
|
||||
#ifndef sample_comb_lod_2d_xe
|
||||
#define sample_comb_lod_2d_xe(texture_sampler_name, position, lod) \
|
||||
sample_sep_lod_2d_xe(texture_sampler_name, \
|
||||
texture_sampler_name##_xe_sampler, position, lod)
|
||||
#endif // !sample_comb_lod_2d_xe
|
||||
#endif
|
||||
#ifndef gather_comb_2d_r_xe
|
||||
#define gather_comb_2d_r_xe(texture_sampler_name, position) \
|
||||
gather_sep_2d_r_xe(texture_sampler_name, \
|
||||
texture_sampler_name##_xe_sampler, position)
|
||||
#endif // !gather_comb_2d_r_xe
|
||||
#endif
|
||||
#ifndef gather_comb_2d_g_xe
|
||||
#define gather_comb_2d_g_xe(texture_sampler_name, position) \
|
||||
gather_sep_2d_g_xe(texture_sampler_name, \
|
||||
texture_sampler_name##_xe_sampler, position)
|
||||
#endif // !gather_comb_2d_g_xe
|
||||
#endif
|
||||
#ifndef gather_comb_2d_b_xe
|
||||
#define gather_comb_2d_b_xe(texture_sampler_name, position) \
|
||||
gather_sep_2d_b_xe(texture_sampler_name, \
|
||||
texture_sampler_name##_xe_sampler, position)
|
||||
#endif // !gather_comb_2d_b_xe
|
||||
#endif
|
||||
#ifndef gather_comb_2d_a_xe
|
||||
#define gather_comb_2d_a_xe(texture_sampler_name, position) \
|
||||
gather_sep_2d_a_xe(texture_sampler_name, \
|
||||
texture_sampler_name##_xe_sampler, position)
|
||||
#endif // !gather_comb_2d_a_xe
|
||||
#endif
|
||||
#endif // !COMBINED_TEXTURE_SAMPLER_XE
|
||||
|
||||
// Passing bindings to functions, and also output and input / output parameters.
|
||||
@@ -795,55 +869,59 @@ float4_xe float_x4_xe(float x) { return float4_xe(x, x, x, x); }
|
||||
#define param_next_after_const_buffer_xe ,
|
||||
#define param_push_consts_xe constant push_const_struct_xe& push_consts_xe
|
||||
#define param_next_after_push_consts_xe ,
|
||||
#define param_uint_vector_buffer_xe(name) const device uint* name
|
||||
#define param_next_after_uint_vector_buffer_xe ,
|
||||
#define param_byte_buffer_xe(name) const device char* name
|
||||
#define param_byte_buffer_wo_xe(name) device char* name
|
||||
#define param_next_after_byte_buffer_xe ,
|
||||
// Call arguments.
|
||||
#define pass_const_buffer_xe(name) (name)
|
||||
#define pass_next_after_const_buffer_xe ,
|
||||
#define pass_push_consts_xe push_consts_xe
|
||||
#define pass_next_after_push_consts_xe ,
|
||||
#define pass_uint_vector_buffer_xe(name) (name)
|
||||
#define pass_next_after_uint_vector_buffer_xe ,
|
||||
#define pass_byte_buffer_xe(name) (name)
|
||||
#define pass_next_after_byte_buffer_xe ,
|
||||
#endif // SHADING_LANGUAGE_*_XE
|
||||
|
||||
// Prototype parameters.
|
||||
#ifndef param_const_buffer_xe
|
||||
#define param_const_buffer_xe(name)
|
||||
#endif // !param_const_buffer_xe
|
||||
#endif
|
||||
#ifndef param_next_after_const_buffer_xe
|
||||
#define param_next_after_const_buffer_xe
|
||||
#endif // !param_next_after_const_buffer_xe
|
||||
#endif
|
||||
#ifndef param_push_consts_xe
|
||||
#define param_push_consts_xe
|
||||
#endif // !param_push_consts_xe
|
||||
#endif
|
||||
#ifndef param_next_after_push_consts_xe
|
||||
#define param_next_after_push_consts_xe
|
||||
#endif // !param_next_after_push_consts_xe
|
||||
#ifndef param_uint_vector_buffer_xe
|
||||
#define param_uint_vector_buffer_xe(name)
|
||||
#endif // !param_uint_vector_buffer_xe
|
||||
#ifndef param_next_after_uint_vector_buffer_xe
|
||||
#define param_next_after_uint_vector_buffer_xe
|
||||
#endif // !param_next_after_uint_vector_buffer_xe
|
||||
#endif
|
||||
#ifndef param_byte_buffer_xe
|
||||
#define param_byte_buffer_xe(name)
|
||||
#endif
|
||||
#ifndef param_byte_buffer_wo_xe
|
||||
#define param_byte_buffer_wo_xe(name)
|
||||
#endif
|
||||
#ifndef param_next_after_byte_buffer_xe
|
||||
#define param_next_after_byte_buffer_xe
|
||||
#endif
|
||||
// Call arguments.
|
||||
#ifndef pass_const_buffer_xe
|
||||
#define pass_const_buffer_xe(name)
|
||||
#endif // !pass_const_buffer_xe
|
||||
#endif
|
||||
#ifndef pass_next_after_const_buffer_xe
|
||||
#define pass_next_after_const_buffer_xe
|
||||
#endif // !pass_next_after_const_buffer_xe
|
||||
#endif
|
||||
#ifndef pass_push_consts_xe
|
||||
#define pass_push_consts_xe
|
||||
#endif // !pass_push_consts_xe
|
||||
#endif
|
||||
#ifndef pass_next_after_push_consts_xe
|
||||
#define pass_next_after_push_consts_xe
|
||||
#endif // !pass_next_after_push_consts_xe
|
||||
#ifndef pass_uint_vector_buffer_xe
|
||||
#define pass_uint_vector_buffer_xe(name)
|
||||
#endif // !pass_uint_vector_buffer_xe
|
||||
#ifndef pass_next_after_uint_vector_buffer_xe
|
||||
#define pass_next_after_uint_vector_buffer_xe
|
||||
#endif // !pass_next_after_uint_vector_buffer_xe
|
||||
#endif
|
||||
#ifndef pass_byte_buffer_xe
|
||||
#define pass_byte_buffer_xe(name)
|
||||
#endif
|
||||
#ifndef pass_next_after_byte_buffer_xe
|
||||
#define pass_next_after_byte_buffer_xe
|
||||
#endif
|
||||
|
||||
// Attributes.
|
||||
|
||||
@@ -860,16 +938,16 @@ float4_xe float_x4_xe(float x) { return float4_xe(x, x, x, x); }
|
||||
#endif // SHADING_LANGUAGE_*_XE
|
||||
#ifndef unroll_xe
|
||||
#define unroll_xe
|
||||
#endif // !unroll_xe
|
||||
#endif
|
||||
#ifndef dont_unroll
|
||||
#define dont_unroll
|
||||
#endif // !dont_unroll
|
||||
#endif
|
||||
#ifndef flatten_xe
|
||||
#define flatten_xe
|
||||
#endif // !flatten_xe
|
||||
#endif
|
||||
#ifndef dont_flatten_xe
|
||||
#define dont_flatten_xe
|
||||
#endif // !dont_flatten_xe
|
||||
#endif
|
||||
|
||||
// Function aliases.
|
||||
|
||||
|
||||
Reference in New Issue
Block a user