mirrors/ZLUDA
1
0
Fork 0
mirror of https://github.com/vosen/ZLUDA.git synced 2025-04-20 00:14:45 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Rewrite surface implementation to more accurately support unofficial CUDA semantics (#203)

Browse source 
This fixes black screen in some CompuBench tests (TV-L1 Optical Flow) and other apps that use CUDA surfaces incorrectly
This commit is contained in:
Andrzej Janik 2024-04-14 02:39:34 +02:00 • committed by GitHub
parent 774f4bcb37
commit 5d5f7cca75
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
8 changed files with 543 additions and 273 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

4
README.md
View file

@ -215,10 +215,6 @@ Performance is currently much lower than the native HIP backend, see the discuss
This is a ROCm/HIP bug. Currently, CompuBench tests have to be run one at a time.
- Some tests output black screen.
This is due to a bug (or an unintended hardware feature) in CompuBench that just happens to work on NVIDIA GPUs.
#### V-Ray Benchmark
- Currently, ZLUDA crashes when running V-Ray benchmark. Nonetheless, certain "lucky" older combinations of ZLUDA and ROCm/HIP are known to run V-Ray Benchmark successfully.

BIN
ptx/lib/zluda_ptx_impl.bc
View file

Binary file not shown.

654
ptx/lib/zluda_ptx_impl.cpp
View file

@ -11,6 +11,7 @@
// https://llvm.org/docs/AMDGPUUsage.html
#include <cstdint>
#include <bit>
#include <hip/hip_runtime.h>
#define HIP_NO_HALF
#include <hip/amd_detail/amd_hip_fp16.h>
@ -155,6 +156,399 @@ static __device__ float4::Native_vec_ __pack_to_float4(const T &t)
return result;
}
typedef uint32_t uint8 __attribute__((ext_vector_type(8)));
typedef uint32_t zluda_uint3 __attribute__((ext_vector_type(3)));
typedef uint8 CONSTANT_SPACE *surface_ptr;
template <typename To, typename From>
static __device__ To transmute(From f)
{
if constexpr (sizeof(To) == sizeof(From))
{
return std::bit_cast<To>(f);
}
else if constexpr (sizeof(To) > sizeof(From))
{
union
{
To t;
From f;
} u = {To{0}};
u.f = f;
return u.t;
}
else if constexpr (sizeof(To) < sizeof(From))
{
union
{
From f;
To t;
} u = {From{f}};
return u.t;
}
else
{
static_assert(sizeof(To) == 0);
}
}
enum class ImageGeometry
{
_1D,
_2D,
_3D,
A1D,
A2D
};
// clang-format off
template <ImageGeometry> struct Coordinates;
template <> struct Coordinates<ImageGeometry::_1D> { using type = uint1::Native_vec_; };
template <> struct Coordinates<ImageGeometry::_2D> { using type = uint2::Native_vec_; };
template <> struct Coordinates<ImageGeometry::_3D> { using type = uint4::Native_vec_; };
template <> struct Coordinates<ImageGeometry::A1D>
{
using type = uint2::Native_vec_; using arg_type = uint1::Native_vec_;
static __device__ type pack_layer(uint32_t layer, arg_type coord)
{
return type { coord.x, layer };
}
};
template <> struct Coordinates<ImageGeometry::A2D>
{
using type = zluda_uint3; using arg_type = uint2::Native_vec_;
static __device__ type pack_layer(uint32_t layer, arg_type coord)
{
return type { coord.x, coord.y, layer };
}
};
// clang-format on
template <typename T, ImageGeometry geo>
static __device__ void image_store_pck(T value, typename Coordinates<geo>::type coord, surface_ptr surface)
{
if constexpr (sizeof(T) <= sizeof(uint))
{
uint value_dword = transmute<uint>(value);
if constexpr (geo == ImageGeometry::_1D)
{
asm volatile("image_store_pck %0, %1, %2 dmask:0x1 dim:1D unorm" : : "v"(value_dword), "v"(coord.x), "s"(*surface) : "memory");
}
else if constexpr (geo == ImageGeometry::_2D)
{
asm volatile("image_store_pck %0, %1, %2 dmask:0x1 dim:2D unorm" : : "v"(value_dword), "v"(coord), "s"(*surface) : "memory");
}
else if constexpr (geo == ImageGeometry::_3D)
{
asm volatile("image_store_pck %0, %1, %2 dmask:0x1 dim:3D unorm" : : "v"(value_dword), "v"(transmute<zluda_uint3>(coord)), "s"(*surface) : "memory");
}
else if constexpr (geo == ImageGeometry::A1D)
{
asm volatile("image_store_pck %0, %1, %2 dmask:0x1 dim:1D_ARRAY unorm" : : "v"(value_dword), "v"(coord), "s"(*surface) : "memory");
}
else if constexpr (geo == ImageGeometry::A2D)
{
asm volatile("image_store_pck %0, %1, %2 dmask:0x1 dim:2D_ARRAY unorm" : : "v"(value_dword), "v"(coord), "s"(*surface) : "memory");
}
else
{
static_assert(sizeof(T) == 0, "Invalid geometry");
}
}
else if constexpr (sizeof(T) == sizeof(uint2::Native_vec_))
{
uint2::Native_vec_ value_dword2 = transmute<uint2::Native_vec_>(value);
if constexpr (geo == ImageGeometry::_1D)
{
asm volatile("image_store_pck %0, %1, %2 dmask:0x3 dim:1D unorm" : : "v"(value_dword2), "v"(coord.x), "s"(*surface) : "memory");
}
else if constexpr (geo == ImageGeometry::_2D)
{
asm volatile("image_store_pck %0, %1, %2 dmask:0x3 dim:2D unorm" : : "v"(value_dword2), "v"(coord), "s"(*surface) : "memory");
}
else if constexpr (geo == ImageGeometry::_3D)
{
asm volatile("image_store_pck %0, %1, %2 dmask:0x3 dim:3D unorm" : : "v"(value_dword2), "v"(transmute<zluda_uint3>(coord)), "s"(*surface) : "memory");
}
else if constexpr (geo == ImageGeometry::A1D)
{
asm volatile("image_store_pck %0, %1, %2 dmask:0x3 dim:1D_ARRAY unorm" : : "v"(value_dword2), "v"(coord), "s"(*surface) : "memory");
}
else if constexpr (geo == ImageGeometry::A2D)
{
asm volatile("image_store_pck %0, %1, %2 dmask:0x3 dim:2D_ARRAY unorm" : : "v"(value_dword2), "v"(coord), "s"(*surface) : "memory");
}
else
{
static_assert(sizeof(T) == 0, "Invalid geometry");
}
}
else if constexpr (sizeof(T) == sizeof(uint4::Native_vec_))
{
uint4::Native_vec_ value_dword4 = transmute<uint4::Native_vec_>(value);
if constexpr (geo == ImageGeometry::_1D)
{
asm volatile("image_store_pck %0, %1, %2 dmask:0xf dim:1D unorm" : : "v"(value_dword4), "v"(coord.x), "s"(*surface) : "memory");
}
else if constexpr (geo == ImageGeometry::_2D)
{
asm volatile("image_store_pck %0, %1, %2 dmask:0xf dim:2D unorm" : : "v"(value_dword4), "v"(coord), "s"(*surface) : "memory");
}
else if constexpr (geo == ImageGeometry::_3D)
{
asm volatile("image_store_pck %0, %1, %2 dmask:0xf dim:3D unorm" : : "v"(value_dword4), "v"(transmute<zluda_uint3>(coord)), "s"(*surface) : "memory");
}
else if constexpr (geo == ImageGeometry::A1D)
{
asm volatile("image_store_pck %0, %1, %2 dmask:0xf dim:1D_ARRAY unorm" : : "v"(value_dword4), "v"(coord), "s"(*surface) : "memory");
}
else if constexpr (geo == ImageGeometry::A2D)
{
asm volatile("image_store_pck %0, %1, %2 dmask:0xf dim:2D_ARRAY unorm" : : "v"(value_dword4), "v"(coord), "s"(*surface) : "memory");
}
else
{
static_assert(sizeof(T) == 0, "Invalid geometry");
}
}
else
{
static_assert(sizeof(T) == 0, "Invalid vector size");
}
}
template <typename T, ImageGeometry geo>
static __device__ T image_load_pck(typename Coordinates<geo>::type coord, surface_ptr surface)
{
if constexpr (sizeof(T) <= sizeof(uint))
{
uint data;
if constexpr (geo == ImageGeometry::_1D)
{
asm volatile("image_load_pck %0, %1, %2 dmask:0x1 dim:1D unorm\ns_waitcnt vmcnt(0)" : "=v"(data) : "v"(coord.x), "s"(*surface) : "memory");
}
else if constexpr (geo == ImageGeometry::_2D)
{
asm volatile("image_load_pck %0, %1, %2 dmask:0x1 dim:2D unorm\ns_waitcnt vmcnt(0)" : "=v"(data) : "v"(coord), "s"(*surface) : "memory");
}
else if constexpr (geo == ImageGeometry::_3D)
{
asm volatile("image_load_pck %0, %1, %2 dmask:0x1 dim:3D unorm\ns_waitcnt vmcnt(0)" : "=v"(data) : "v"(transmute<zluda_uint3>(coord)), "s"(*surface) : "memory");
}
else if constexpr (geo == ImageGeometry::A1D)
{
asm volatile("image_load_pck %0, %1, %2 dmask:0x1 dim:1D_ARRAY unorm\ns_waitcnt vmcnt(0)" : "=v"(data) : "v"(coord), "s"(*surface) : "memory");
}
else if constexpr (geo == ImageGeometry::A2D)
{
asm volatile("image_load_pck %0, %1, %2 dmask:0x1 dim:2D_ARRAY unorm\ns_waitcnt vmcnt(0)" : "=v"(data) : "v"(coord), "s"(*surface) : "memory");
}
else
{
static_assert(sizeof(ImageGeometry) == 0, "Invalid geometry");
}
return transmute<T>(data);
}
else if constexpr (sizeof(T) == sizeof(uint2::Native_vec_))
{
uint2::Native_vec_ data;
if constexpr (geo == ImageGeometry::_1D)
{
asm volatile("image_load_pck %0, %1, %2 dmask:0x3 dim:1D unorm\ns_waitcnt vmcnt(0)" : "=v"(data) : "v"(coord.x), "s"(*surface) : "memory");
}
else if constexpr (geo == ImageGeometry::_2D)
{
asm volatile("image_load_pck %0, %1, %2 dmask:0x3 dim:2D unorm\ns_waitcnt vmcnt(0)" : "=v"(data) : "v"(coord), "s"(*surface) : "memory");
}
else if constexpr (geo == ImageGeometry::_3D)
{
asm volatile("image_load_pck %0, %1, %2 dmask:0x3 dim:3D unorm\ns_waitcnt vmcnt(0)" : "=v"(data) : "v"(transmute<zluda_uint3>(coord)), "s"(*surface) : "memory");
}
else if constexpr (geo == ImageGeometry::A1D)
{
asm volatile("image_load_pck %0, %1, %2 dmask:0x3 dim:1D_ARRAY unorm\ns_waitcnt vmcnt(0)" : "=v"(data) : "v"(coord), "s"(*surface) : "memory");
}
else if constexpr (geo == ImageGeometry::A2D)
{
asm volatile("image_load_pck %0, %1, %2 dmask:0x3 dim:2D_ARRAY unorm\ns_waitcnt vmcnt(0)" : "=v"(data) : "v"(coord), "s"(*surface) : "memory");
}
else
{
static_assert(sizeof(ImageGeometry) == 0, "Invalid geometry");
}
return transmute<T>(data);
}
else if constexpr (sizeof(T) == sizeof(uint4::Native_vec_))
{
uint4::Native_vec_ data;
if constexpr (geo == ImageGeometry::_1D)
{
asm volatile("image_load_pck %0, %1, %2 dmask:0xf dim:1D unorm\ns_waitcnt vmcnt(0)" : "=v"(data) : "v"(coord.x), "s"(*surface) : "memory");
}
else if constexpr (geo == ImageGeometry::_2D)
{
asm volatile("image_load_pck %0, %1, %2 dmask:0xf dim:2D unorm\ns_waitcnt vmcnt(0)" : "=v"(data) : "v"(coord), "s"(*surface) : "memory");
}
else if constexpr (geo == ImageGeometry::_3D)
{
asm volatile("image_load_pck %0, %1, %2 dmask:0xf dim:3D unorm\ns_waitcnt vmcnt(0)" : "=v"(data) : "v"(transmute<zluda_uint3>(coord)), "s"(*surface) : "memory");
}
else if constexpr (geo == ImageGeometry::A1D)
{
asm volatile("image_load_pck %0, %1, %2 dmask:0xf dim:1D_ARRAY unorm\ns_waitcnt vmcnt(0)" : "=v"(data) : "v"(coord), "s"(*surface) : "memory");
}
else if constexpr (geo == ImageGeometry::A2D)
{
asm volatile("image_load_pck %0, %1, %2 dmask:0xf dim:2D_ARRAY unorm\ns_waitcnt vmcnt(0)" : "=v"(data) : "v"(coord), "s"(*surface) : "memory");
}
else
{
static_assert(sizeof(ImageGeometry) == 0, "Invalid geometry");
}
return transmute<T>(data);
}
else
{
static_assert(sizeof(T) == 0, "Invalid vector size");
}
}
template <ImageGeometry geo>
static __device__ uint4::Native_vec_ image_load_pck_full(typename Coordinates<geo>::type coord, surface_ptr surface)
{
uint4::Native_vec_ data;
if constexpr (geo == ImageGeometry::_1D)
{
asm volatile("image_load_pck %0, %1, %2 dmask:0xf dim:1D unorm\ns_waitcnt vmcnt(0)" : "=v"(data) : "v"(coord.x), "s"(*surface) : "memory");
}
else if constexpr (geo == ImageGeometry::_2D)
{
asm volatile("image_load_pck %0, %1, %2 dmask:0xf dim:2D unorm\ns_waitcnt vmcnt(0)" : "=v"(data) : "v"(coord), "s"(*surface) : "memory");
}
else if constexpr (geo == ImageGeometry::_3D)
{
asm volatile("image_load_pck %0, %1, %2 dmask:0xf dim:3D unorm\ns_waitcnt vmcnt(0)" : "=v"(data) : "v"(transmute<zluda_uint3>(coord)), "s"(*surface) : "memory");
}
else if constexpr (geo == ImageGeometry::A1D)
{
asm volatile("image_load_pck %0, %1, %2 dmask:0xf dim:1D_ARRAY unorm\ns_waitcnt vmcnt(0)" : "=v"(data) : "v"(coord), "s"(*surface) : "memory");
}
else if constexpr (geo == ImageGeometry::A2D)
{
asm volatile("image_load_pck %0, %1, %2 dmask:0xf dim:2D_ARRAY unorm\ns_waitcnt vmcnt(0)" : "=v"(data) : "v"(coord), "s"(*surface) : "memory");
}
else
{
static_assert(sizeof(ImageGeometry) == 0, "Invalid geometry");
}
return data;
}
template <typename T, ImageGeometry geo>
static __device__ void image_store_pck_full_with(uint4::Native_vec_ data, T value, typename Coordinates<geo>::type coord, surface_ptr surface)
{
// We avoid unions for types smaller than sizeof(uint32_t),
// because in those cases we get this garbage:
// ds_write_b128 v2, v[5:8]
// ds_write_b16 v2, v9
// ds_read_b128 v[5:8], v2
// tested with ROCm 5.7.1 on gfx1030
if constexpr (sizeof(T) == sizeof(uint8_t))
{
uint32_t x = uint32_t(std::bit_cast<uint8_t>(value));
uint32_t data_0 = ((data[0]) >> 8) << 8;
data[0] = data_0 | x;
}
else if constexpr (sizeof(T) == sizeof(uint16_t))
{
uint32_t x = uint32_t(std::bit_cast<uint16_t>(value));
uint32_t data_0 = ((data[0]) >> 16) << 16;
data[0] = data_0 | x;
}
else
{
union
{
uint4::Native_vec_ full_vec;
T value;
} u = {0};
u.full_vec = data;
u.value = value;
data = u.full_vec;
}
image_store_pck<uint4::Native_vec_, geo>(data, coord, surface);
}
constexpr auto IMAGE_RESERVED_TOP_BITS = 3;
static __device__ surface_ptr get_surface_pointer(uint64_t s)
{
return (surface_ptr)((s << IMAGE_RESERVED_TOP_BITS) >> IMAGE_RESERVED_TOP_BITS);
}
static __device__ surface_ptr get_surface_pointer(struct textureReference GLOBAL_SPACE *surf_ref)
{
return (surface_ptr)(surf_ref->textureObject);
}
static __device__ uint32_t x_coordinate_shift(uint64_t s)
{
return uint32_t(s >> (64 - IMAGE_RESERVED_TOP_BITS));
}
static __device__ uint32_t x_coordinate_shift(struct textureReference GLOBAL_SPACE *ptr)
{
uint32_t channels = uint32_t(ptr->numChannels);
uint32_t format_width = 0;
hipArray_Format format = ptr->format;
switch (format)
{
case hipArray_Format::HIP_AD_FORMAT_UNSIGNED_INT8:
case hipArray_Format::HIP_AD_FORMAT_SIGNED_INT8:
format_width = 1;
break;
case hipArray_Format::HIP_AD_FORMAT_UNSIGNED_INT16:
case hipArray_Format::HIP_AD_FORMAT_SIGNED_INT16:
case hipArray_Format::HIP_AD_FORMAT_HALF:
format_width = 2;
break;
case hipArray_Format::HIP_AD_FORMAT_UNSIGNED_INT32:
case hipArray_Format::HIP_AD_FORMAT_SIGNED_INT32:
case hipArray_Format::HIP_AD_FORMAT_FLOAT:
format_width = 4;
break;
default:
__builtin_unreachable();
}
return uint32_t(__builtin_ctz(format_width * channels));
}
template <typename T, ImageGeometry geo, typename Surface>
static __device__ T suld_b_zero(Surface surf_arg, typename Coordinates<geo>::type coord)
{
surface_ptr surface = get_surface_pointer(surf_arg);
uint32_t shift_x = x_coordinate_shift(surf_arg);
coord.x = coord.x >> shift_x;
return image_load_pck<T, geo>(coord, surface);
}
template <typename T, ImageGeometry geo, typename Surface>
static __device__ void sust_b_zero(Surface surf_arg, typename Coordinates<geo>::type coord, T data)
{
surface_ptr surface = get_surface_pointer(surf_arg);
uint32_t shift_x = x_coordinate_shift(surf_arg);
coord.x = coord.x >> shift_x;
if (shift_x <= __builtin_ctz(sizeof(T))) [[likely]]
{
image_store_pck<T, geo>(data, coord, surface);
}
else
{
uint4::Native_vec_ pixel = image_load_pck_full<geo>(coord, surface);
image_store_pck_full_with<T, geo>(pixel, data, coord, surface);
}
}
extern "C"
{
#define atomic_inc(NAME, SUCCESS, FAILURE, SCOPE, SPACE) \
@ -620,179 +1014,101 @@ extern "C"
suld_b_a2d_vec(_v4, b32, uint4);
// suld_b_a2d_vec(_v4, b64, ulong4);
#define sust_b_1d_vec(VEC, TYPE, HIP_TYPE) \
void FUNC(sust_b_1d##VEC##_##TYPE##_trap)(struct textureReference GLOBAL_SPACE * ptr, int1::Native_vec_ coord, HIP_TYPE::Native_vec_ data) \
{ \
hipTextureObject_t textureObject = ptr->textureObject; \
TEXTURE_OBJECT_PARAMETERS_INIT; \
(void)s; \
int byte_coord = __hipGetPixelAddr(coord.x, __ockl_image_channel_data_type_1D(i), __ockl_image_channel_order_1D(i)); \
HIP_TYPE hip_data; \
hip_data.data = data; \
auto tmp = __pack_to_float4(hip_data); \
__ockl_image_store_1D(i, byte_coord, tmp); \
} \
void FUNC(sust_b_indirect_1d##VEC##_##TYPE##_trap)(uint64_t serf_arg, int1::Native_vec_ coord, HIP_TYPE::Native_vec_ data) \
{ \
hipSurfaceObject_t surfObj = (hipSurfaceObject_t)serf_arg; \
HIP_TYPE hip_data; \
hip_data.data = data; \
surf1Dwrite(hip_data, surfObj, coord.x); \
#define SUST_B_ZERO(TYPE, GEOMETRY, HIP_TYPE) \
HIP_TYPE::Native_vec_ FUNC(suld_b_indirect_##TYPE##_zero)(uint64_t surf_arg, typename Coordinates<GEOMETRY>::type coord) \
{ \
return suld_b_zero<HIP_TYPE::Native_vec_, GEOMETRY>(surf_arg, coord); \
} \
void FUNC(sust_b_indirect_##TYPE##_zero)(uint64_t surf_arg, typename Coordinates<GEOMETRY>::type coord, HIP_TYPE::Native_vec_ data) \
{ \
sust_b_zero<HIP_TYPE::Native_vec_, GEOMETRY>(surf_arg, coord, data); \
} \
HIP_TYPE::Native_vec_ FUNC(suld_b_##TYPE##_zero)(struct textureReference GLOBAL_SPACE * ptr, typename Coordinates<GEOMETRY>::type coord) \
{ \
return suld_b_zero<HIP_TYPE::Native_vec_, GEOMETRY>(ptr, coord); \
} \
void FUNC(sust_b_##TYPE##_zero)(struct textureReference GLOBAL_SPACE * ptr, typename Coordinates<GEOMETRY>::type coord, HIP_TYPE::Native_vec_ data) \
{ \
sust_b_zero<HIP_TYPE::Native_vec_, GEOMETRY>(ptr, coord, data); \
}
sust_b_1d_vec(, b8, uchar1);
sust_b_1d_vec(, b16, ushort1);
sust_b_1d_vec(, b32, uint1);
// sust_b_1d_vec(, b64, ulong1);
sust_b_1d_vec(_v2, b8, uchar2);
sust_b_1d_vec(_v2, b16, ushort2);
sust_b_1d_vec(_v2, b32, uint2);
// sust_b_1d_vec(_v2, b64, ulong2);
sust_b_1d_vec(_v4, b8, uchar4);
sust_b_1d_vec(_v4, b16, ushort4);
sust_b_1d_vec(_v4, b32, uint4);
// sust_b_1d_vec(_v4, b64, ulong4);
#define sust_b_2d_vec(VEC, TYPE, HIP_TYPE) \
void FUNC(sust_b_2d##VEC##_##TYPE##_trap)(struct textureReference GLOBAL_SPACE * ptr, int2::Native_vec_ coord, HIP_TYPE::Native_vec_ data) \
{ \
hipTextureObject_t textureObject = ptr->textureObject; \
TEXTURE_OBJECT_PARAMETERS_INIT; \
(void)s; \
int byte_coord = __hipGetPixelAddr(coord.x, __ockl_image_channel_data_type_2D(i), __ockl_image_channel_order_2D(i)); \
HIP_TYPE hip_data; \
hip_data.data = data; \
auto tmp = __pack_to_float4(hip_data); \
__ockl_image_store_2D(i, int2(byte_coord, coord.y).data, tmp); \
} \
void FUNC(sust_b_indirect_2d##VEC##_##TYPE##_trap)(uint64_t serf_arg, int2::Native_vec_ coord, HIP_TYPE::Native_vec_ data) \
{ \
hipSurfaceObject_t surfObj = (hipSurfaceObject_t)serf_arg; \
HIP_TYPE hip_data; \
hip_data.data = data; \
surf2Dwrite(hip_data, surfObj, coord.x, coord.y); \
#define SUST_B_ZERO_ARRAY(TYPE, GEOMETRY, HIP_TYPE) \
HIP_TYPE::Native_vec_ FUNC(suld_b_indirect_##TYPE##_zero)(uint64_t surf_arg, uint32_t layer, typename Coordinates<GEOMETRY>::arg_type coord) \
{ \
auto coord_array = Coordinates<GEOMETRY>::pack_layer(layer, coord); \
return suld_b_zero<HIP_TYPE::Native_vec_, GEOMETRY>(surf_arg, coord_array); \
} \
void FUNC(sust_b_indirect_##TYPE##_zero)(uint64_t surf_arg, uint32_t layer, typename Coordinates<GEOMETRY>::arg_type coord, HIP_TYPE::Native_vec_ data) \
{ \
auto coord_array = Coordinates<GEOMETRY>::pack_layer(layer, coord); \
sust_b_zero<HIP_TYPE::Native_vec_, GEOMETRY>(surf_arg, coord_array, data); \
} \
HIP_TYPE::Native_vec_ FUNC(suld_b_##TYPE##_zero)(struct textureReference GLOBAL_SPACE * ptr, uint32_t layer, typename Coordinates<GEOMETRY>::arg_type coord) \
{ \
auto coord_array = Coordinates<GEOMETRY>::pack_layer(layer, coord); \
return suld_b_zero<HIP_TYPE::Native_vec_, GEOMETRY>(ptr, coord_array); \
} \
void FUNC(sust_b_##TYPE##_zero)(struct textureReference GLOBAL_SPACE * ptr, uint32_t layer, typename Coordinates<GEOMETRY>::arg_type coord, HIP_TYPE::Native_vec_ data) \
{ \
auto coord_array = Coordinates<GEOMETRY>::pack_layer(layer, coord); \
sust_b_zero<HIP_TYPE::Native_vec_, GEOMETRY>(ptr, coord_array, data); \
}
sust_b_2d_vec(, b8, uchar1);
sust_b_2d_vec(, b16, ushort1);
sust_b_2d_vec(, b32, uint1);
// sust_b_2d_vec(, b64, ulong1);
sust_b_2d_vec(_v2, b8, uchar2);
sust_b_2d_vec(_v2, b16, ushort2);
sust_b_2d_vec(_v2, b32, uint2);
// sust_b_2d_vec(_v2, b64, ulong2);
sust_b_2d_vec(_v4, b8, uchar4);
sust_b_2d_vec(_v4, b16, ushort4);
sust_b_2d_vec(_v4, b32, uint4);
// sust_b_2d_vec(_v4, b64, ulong4);
#define sust_b_3d_vec(VEC, TYPE, HIP_TYPE) \
void FUNC(sust_b_3d##VEC##_##TYPE##_trap)(struct textureReference GLOBAL_SPACE * ptr, int4::Native_vec_ coord, HIP_TYPE::Native_vec_ data) \
{ \
hipTextureObject_t textureObject = ptr->textureObject; \
TEXTURE_OBJECT_PARAMETERS_INIT; \
(void)s; \
int byte_coord = __hipGetPixelAddr(coord.x, __ockl_image_channel_data_type_3D(i), __ockl_image_channel_order_3D(i)); \
HIP_TYPE hip_data; \
hip_data.data = data; \
auto tmp = __pack_to_float4(hip_data); \
__ockl_image_store_3D(i, int4(byte_coord, coord.y, coord.z, 0).data, tmp); \
} \
void FUNC(sust_b_indirect_3d##VEC##_##TYPE##_trap)(uint64_t serf_arg, int4::Native_vec_ coord, HIP_TYPE::Native_vec_ data) \
{ \
hipSurfaceObject_t surfObj = (hipSurfaceObject_t)serf_arg; \
__HIP_SURFACE_OBJECT_PARAMETERS_INIT; \
int byte_coord = __hipGetPixelAddr(coord.x, __ockl_image_channel_data_type_3D(i), __ockl_image_channel_order_3D(i)); \
HIP_TYPE hip_data; \
hip_data.data = data; \
auto tmp = __pack_to_float4(hip_data); \
__ockl_image_store_3D(i, int4(byte_coord, coord.y, coord.z, 0).data, tmp); \
}
sust_b_3d_vec(, b8, uchar1);
sust_b_3d_vec(, b16, ushort1);
sust_b_3d_vec(, b32, uint1);
// sust_b_3d_vec(, b64, ulong1);
sust_b_3d_vec(_v2, b8, uchar2);
sust_b_3d_vec(_v2, b16, ushort2);
sust_b_3d_vec(_v2, b32, uint2);
// sust_b_3d_vec(_v2, b64, ulong2);
sust_b_3d_vec(_v4, b8, uchar4);
sust_b_3d_vec(_v4, b16, ushort4);
sust_b_3d_vec(_v4, b32, uint4);
// sust_b_3d_vec(_v4, b64, ulong4);
#define sust_b_a1d_vec(VEC, TYPE, HIP_TYPE) \
void FUNC(sust_b_a1d##VEC##_##TYPE##_trap)(struct textureReference GLOBAL_SPACE * ptr, uint layer, int x, HIP_TYPE::Native_vec_ data) \
{ \
hipTextureObject_t textureObject = ptr->textureObject; \
TEXTURE_OBJECT_PARAMETERS_INIT; \
(void)s; \
int byte_coord = __hipGetPixelAddr(x, __ockl_image_channel_data_type_1Da(i), __ockl_image_channel_order_1Da(i)); \
HIP_TYPE hip_data; \
hip_data.data = data; \
auto tmp = __pack_to_float4(hip_data); \
__ockl_image_store_1Da(i, int2(byte_coord, int(layer)).data, tmp); \
} \
void FUNC(sust_b_indirect_a1d##VEC##_##TYPE##_trap)(uint64_t serf_arg, uint layer, int x, HIP_TYPE::Native_vec_ data) \
{ \
hipSurfaceObject_t surfObj = (hipSurfaceObject_t)serf_arg; \
__HIP_SURFACE_OBJECT_PARAMETERS_INIT; \
int byte_coord = __hipGetPixelAddr(x, __ockl_image_channel_data_type_1Da(i), __ockl_image_channel_order_1Da(i)); \
HIP_TYPE hip_data; \
hip_data.data = data; \
auto tmp = __pack_to_float4(hip_data); \
__ockl_image_store_1Da(i, int2(byte_coord, int(layer)).data, tmp); \
}
sust_b_a1d_vec(, b8, uchar1);
sust_b_a1d_vec(, b16, ushort1);
sust_b_a1d_vec(, b32, uint1);
// sust_b_a1d_vec(, b64, ulong1);
sust_b_a1d_vec(_v2, b8, uchar2);
sust_b_a1d_vec(_v2, b16, ushort2);
sust_b_a1d_vec(_v2, b32, uint2);
// sust_b_a1d_vec(_v2, b64, ulong2);
sust_b_a1d_vec(_v4, b8, uchar4);
sust_b_a1d_vec(_v4, b16, ushort4);
sust_b_a1d_vec(_v4, b32, uint4);
// sust_b_a1d_vec(_v4, b64, ulong4);
#define sust_b_a2d_vec(VEC, TYPE, HIP_TYPE) \
void FUNC(sust_b_a2d##VEC##_##TYPE##_trap)(struct textureReference GLOBAL_SPACE * ptr, uint layer, int x, int y, HIP_TYPE::Native_vec_ data) \
{ \
hipTextureObject_t textureObject = ptr->textureObject; \
TEXTURE_OBJECT_PARAMETERS_INIT; \
(void)s; \
int byte_coord = __hipGetPixelAddr(x, __ockl_image_channel_data_type_2Da(i), __ockl_image_channel_order_2Da(i)); \
HIP_TYPE hip_data; \
hip_data.data = data; \
auto tmp = __pack_to_float4(hip_data); \
__ockl_image_store_2Da(i, int4(byte_coord, y, int(layer), 0).data, tmp); \
} \
void FUNC(sust_b_indirect_a2d##VEC##_##TYPE##_trap)(uint64_t serf_arg, uint layer, int x, int y, HIP_TYPE::Native_vec_ data) \
{ \
hipSurfaceObject_t surfObj = (hipSurfaceObject_t)serf_arg; \
__HIP_SURFACE_OBJECT_PARAMETERS_INIT; \
int byte_coord = __hipGetPixelAddr(x, __ockl_image_channel_data_type_2Da(i), __ockl_image_channel_order_2Da(i)); \
HIP_TYPE hip_data; \
hip_data.data = data; \
auto tmp = __pack_to_float4(hip_data); \
__ockl_image_store_2Da(i, int4(byte_coord, y, int(layer), 0).data, tmp); \
}
sust_b_a2d_vec(, b8, uchar1);
sust_b_a2d_vec(, b16, ushort1);
sust_b_a2d_vec(, b32, uint1);
// sust_b_a2d_vec(, b64, ulong1);
sust_b_a2d_vec(_v2, b8, uchar2);
sust_b_a2d_vec(_v2, b16, ushort2);
sust_b_a2d_vec(_v2, b32, uint2);
// sust_b_a2d_vec(_v2, b64, ulong2);
sust_b_a2d_vec(_v4, b8, uchar4);
sust_b_a2d_vec(_v4, b16, ushort4);
sust_b_a2d_vec(_v4, b32, uint4);
// sust_b_a2d_vec(_v4, b64, ulong4);
SUST_B_ZERO(1d_b8, ImageGeometry::_1D, uchar1);
SUST_B_ZERO(1d_b16, ImageGeometry::_1D, ushort1);
SUST_B_ZERO(1d_b32, ImageGeometry::_1D, uint1);
SUST_B_ZERO(1d_b64, ImageGeometry::_1D, ulong1);
SUST_B_ZERO(1d_v2_b8, ImageGeometry::_1D, uchar2);
SUST_B_ZERO(1d_v2_b16, ImageGeometry::_1D, ushort2);
SUST_B_ZERO(1d_v2_b32, ImageGeometry::_1D, uint2);
SUST_B_ZERO(1d_v2_b64, ImageGeometry::_1D, ulong2);
SUST_B_ZERO(1d_v4_b8, ImageGeometry::_1D, uchar4);
SUST_B_ZERO(1d_v4_b16, ImageGeometry::_1D, ushort4);
SUST_B_ZERO(1d_v4_b32, ImageGeometry::_1D, uint4);
SUST_B_ZERO(2d_b8, ImageGeometry::_2D, uchar1);
SUST_B_ZERO(2d_b16, ImageGeometry::_2D, ushort1);
SUST_B_ZERO(2d_b32, ImageGeometry::_2D, uint1);
SUST_B_ZERO(2d_b64, ImageGeometry::_2D, ulong1);
SUST_B_ZERO(2d_v2_b8, ImageGeometry::_2D, uchar2);
SUST_B_ZERO(2d_v2_b16, ImageGeometry::_2D, ushort2);
SUST_B_ZERO(2d_v2_b32, ImageGeometry::_2D, uint2);
SUST_B_ZERO(2d_v2_b64, ImageGeometry::_2D, ulong2);
SUST_B_ZERO(2d_v4_b8, ImageGeometry::_2D, uchar4);
SUST_B_ZERO(2d_v4_b16, ImageGeometry::_2D, ushort4);
SUST_B_ZERO(2d_v4_b32, ImageGeometry::_2D, uint4);
SUST_B_ZERO(3d_b8, ImageGeometry::_3D, uchar1);
SUST_B_ZERO(3d_b16, ImageGeometry::_3D, ushort1);
SUST_B_ZERO(3d_b32, ImageGeometry::_3D, uint1);
SUST_B_ZERO(3d_b64, ImageGeometry::_3D, ulong1);
SUST_B_ZERO(3d_v2_b8, ImageGeometry::_3D, uchar2);
SUST_B_ZERO(3d_v2_b16, ImageGeometry::_3D, ushort2);
SUST_B_ZERO(3d_v2_b32, ImageGeometry::_3D, uint2);
SUST_B_ZERO(3d_v2_b64, ImageGeometry::_3D, ulong2);
SUST_B_ZERO(3d_v4_b8, ImageGeometry::_3D, uchar4);
SUST_B_ZERO(3d_v4_b16, ImageGeometry::_3D, ushort4);
SUST_B_ZERO(3d_v4_b32, ImageGeometry::_3D, uint4);
SUST_B_ZERO_ARRAY(a1d_b8, ImageGeometry::A1D, uchar1);
SUST_B_ZERO_ARRAY(a1d_b16, ImageGeometry::A1D, ushort1);
SUST_B_ZERO_ARRAY(a1d_b32, ImageGeometry::A1D, uint1);
SUST_B_ZERO_ARRAY(a1d_b64, ImageGeometry::A1D, ulong1);
SUST_B_ZERO_ARRAY(a1d_v2_b8, ImageGeometry::A1D, uchar2);
SUST_B_ZERO_ARRAY(a1d_v2_b16, ImageGeometry::A1D, ushort2);
SUST_B_ZERO_ARRAY(a1d_v2_b32, ImageGeometry::A1D, uint2);
SUST_B_ZERO_ARRAY(a1d_v2_b64, ImageGeometry::A1D, ulong2);
SUST_B_ZERO_ARRAY(a1d_v4_b8, ImageGeometry::A1D, uchar4);
SUST_B_ZERO_ARRAY(a1d_v4_b16, ImageGeometry::A1D, ushort4);
SUST_B_ZERO_ARRAY(a1d_v4_b32, ImageGeometry::A1D, uint4);
SUST_B_ZERO_ARRAY(a2d_b8, ImageGeometry::A2D, uchar1);
SUST_B_ZERO_ARRAY(a2d_b16, ImageGeometry::A2D, ushort1);
SUST_B_ZERO_ARRAY(a2d_b32, ImageGeometry::A2D, uint1);
SUST_B_ZERO_ARRAY(a2d_b64, ImageGeometry::A2D, ulong1);
SUST_B_ZERO_ARRAY(a2d_v2_b8, ImageGeometry::A2D, uchar2);
SUST_B_ZERO_ARRAY(a2d_v2_b16, ImageGeometry::A2D, ushort2);
SUST_B_ZERO_ARRAY(a2d_v2_b32, ImageGeometry::A2D, uint2);
SUST_B_ZERO_ARRAY(a2d_v2_b64, ImageGeometry::A2D, ulong2);
SUST_B_ZERO_ARRAY(a2d_v4_b8, ImageGeometry::A2D, uchar4);
SUST_B_ZERO_ARRAY(a2d_v4_b16, ImageGeometry::A2D, ushort4);
SUST_B_ZERO_ARRAY(a2d_v4_b32, ImageGeometry::A2D, uint4);
__device__ static inline bool is_upper_warp()
{

4
ptx/src/translate.rs
View file

@ -2934,7 +2934,7 @@ fn replace_instructions_with_builtins_impl<'input>(
vector,
"_",
suld.type_.to_ptx_name(),
"_trap",
"_zero",
]
.concat();
statements.push(instruction_to_fn_call(
@ -2955,7 +2955,7 @@ fn replace_instructions_with_builtins_impl<'input>(
vector,
"_",
sust.type_.to_ptx_name(),
"_trap",
"_zero",
]
.concat();
statements.push(instruction_to_fn_call(

2
zluda/src/cuda.rs
View file

@ -1245,7 +1245,7 @@ mod definitions {
pub(crate) unsafe fn cuSurfObjectDestroy(
surfObject: hipSurfaceObject_t,
) -> hipError_t {
hipDestroySurfaceObject(surfObject)
surface::destroy(surfObject)
}
pub(crate) unsafe fn cuTexObjectCreate(

136
zluda/src/impl/surface.rs
View file

@ -1,23 +1,65 @@
use super::hipfix;
use crate::hip_call_cuda;
use cuda_types::*;
use hip_runtime_sys::*;
use std::{mem, ptr};
use crate::hip_call_cuda;
use super::{hipfix, FromCuda};
// Same as in zluda_ptx_impl.cpp
const IMAGE_RESERVED_TOP_BITS: u32 = 3;
pub(crate) unsafe fn create(
p_surf_object: *mut hipSurfaceObject_t,
result: *mut hipSurfaceObject_t,
p_res_desc: *const CUDA_RESOURCE_DESC,
) -> Result<(), CUresult> {
if p_res_desc == ptr::null() {
return Err(CUresult::CUDA_ERROR_INVALID_VALUE);
}
let desc = to_surface_desc(*p_res_desc)?;
hip_call_cuda!(hipCreateSurfaceObject(p_surf_object, &desc));
// We need to check array format and channel count to set top bits of the surface object.
// HIP does not support non-Array sources anyway
if desc.resType != hipResourceType::hipResourceTypeArray {
return Err(CUresult::CUDA_ERROR_NOT_SUPPORTED);
}
let mut surf_obj = mem::zeroed();
hip_call_cuda!(hipCreateSurfaceObject(&mut surf_obj, &desc));
let top_reserved_bits = surf_obj as usize >> (usize::BITS - IMAGE_RESERVED_TOP_BITS);
if top_reserved_bits != 0 {
#[allow(unused_must_use)]
{
hipDestroySurfaceObject(surf_obj);
}
return Err(CUresult::CUDA_ERROR_UNKNOWN);
}
let format_size = format_size((&*desc.res.array.array).Format)?;
let channels = (&*desc.res.array.array).NumChannels;
let pixel_size = format_size * channels as usize;
let shift_amount =
(pixel_size.trailing_zeros() as usize) << (usize::BITS - IMAGE_RESERVED_TOP_BITS);
surf_obj = (surf_obj as usize | shift_amount) as _;
*result = surf_obj;
Ok(())
}
pub(crate) unsafe fn destroy(surf_object: hipSurfaceObject_t) -> hipError_t {
hipDestroySurfaceObject(
(((surf_object as usize) << IMAGE_RESERVED_TOP_BITS) >> IMAGE_RESERVED_TOP_BITS) as _,
)
}
pub(crate) fn format_size(f: hipArray_Format) -> Result<usize, CUresult> {
Ok(match f {
hipArray_Format::HIP_AD_FORMAT_UNSIGNED_INT8
| hipArray_Format::HIP_AD_FORMAT_SIGNED_INT8 => 1,
hipArray_Format::HIP_AD_FORMAT_UNSIGNED_INT16
| hipArray_Format::HIP_AD_FORMAT_SIGNED_INT16
| hipArray_Format::HIP_AD_FORMAT_HALF => 2,
hipArray_Format::HIP_AD_FORMAT_UNSIGNED_INT32
| hipArray_Format::HIP_AD_FORMAT_SIGNED_INT32
| hipArray_Format::HIP_AD_FORMAT_FLOAT => 4,
_ => return Err(CUresult::CUDA_ERROR_NOT_SUPPORTED),
})
}
unsafe fn to_surface_desc(res_desc: CUDA_RESOURCE_DESC) -> Result<hipResourceDesc, CUresult> {
let res_type = mem::transmute(res_desc.resType);
let res: hipResourceDesc__bindgen_ty_1 = match res_desc.resType {
@ -26,92 +68,10 @@ unsafe fn to_surface_desc(res_desc: CUDA_RESOURCE_DESC) -> Result<hipResourceDes
array: hipfix::array::get(res_desc.res.array.hArray),
},
},
CUresourcetype::CU_RESOURCE_TYPE_MIPMAPPED_ARRAY => hipResourceDesc__bindgen_ty_1 {
mipmap: hipResourceDesc__bindgen_ty_1__bindgen_ty_2 {
mipmap: mem::transmute(res_desc.res.mipmap.hMipmappedArray),
},
},
CUresourcetype::CU_RESOURCE_TYPE_LINEAR => hipResourceDesc__bindgen_ty_1 {
linear: hipResourceDesc__bindgen_ty_1__bindgen_ty_3 {
devPtr: res_desc.res.linear.devPtr.0,
desc: channel_format_desc(
FromCuda::from_cuda(res_desc.res.linear.format),
res_desc.res.linear.numChannels,
)?,
sizeInBytes: res_desc.res.linear.sizeInBytes,
},
},
CUresourcetype::CU_RESOURCE_TYPE_PITCH2D => hipResourceDesc__bindgen_ty_1 {
pitch2D: hipResourceDesc__bindgen_ty_1__bindgen_ty_4 {
devPtr: res_desc.res.pitch2D.devPtr.0,
desc: channel_format_desc(
FromCuda::from_cuda(res_desc.res.pitch2D.format),
res_desc.res.pitch2D.numChannels,
)?,
width: res_desc.res.pitch2D.width,
height: res_desc.res.pitch2D.height,
pitchInBytes: res_desc.res.pitch2D.pitchInBytes,
},
},
_ => todo!(),
_ => return Err(CUresult::CUDA_ERROR_NOT_SUPPORTED),
};
Ok(hipResourceDesc {
resType: res_type,
res,
})
}
fn channel_format_desc(
format: hipArray_Format,
num_channels: u32,
) -> Result<hipChannelFormatDesc, CUresult> {
let mut bits = match num_channels {
1 => (1, 0, 0, 0),
2 => (1, 1, 0, 0),
3 => (1, 1, 1, 0),
4 => (1, 1, 1, 1),
_ => return Err(CUresult::CUDA_ERROR_INVALID_VALUE),
};
let (kind, bit_width) = match format {
hipArray_Format::HIP_AD_FORMAT_UNSIGNED_INT8 => {
(hipChannelFormatKind::hipChannelFormatKindUnsigned, u8::BITS)
}
hipArray_Format::HIP_AD_FORMAT_UNSIGNED_INT16 => (
hipChannelFormatKind::hipChannelFormatKindUnsigned,
u16::BITS,
),
hipArray_Format::HIP_AD_FORMAT_UNSIGNED_INT32 => (
hipChannelFormatKind::hipChannelFormatKindUnsigned,
u32::BITS,
),
hipArray_Format::HIP_AD_FORMAT_SIGNED_INT8 => {
(hipChannelFormatKind::hipChannelFormatKindSigned, i8::BITS)
}
hipArray_Format::HIP_AD_FORMAT_SIGNED_INT16 => {
(hipChannelFormatKind::hipChannelFormatKindSigned, i16::BITS)
}
hipArray_Format::HIP_AD_FORMAT_SIGNED_INT32 => {
(hipChannelFormatKind::hipChannelFormatKindSigned, i32::BITS)
}
hipArray_Format::HIP_AD_FORMAT_HALF => (
hipChannelFormatKind::hipChannelFormatKindFloat,
mem::size_of::<u16>() as u32 * u8::BITS,
),
hipArray_Format::HIP_AD_FORMAT_FLOAT => (
hipChannelFormatKind::hipChannelFormatKindFloat,
mem::size_of::<f32>() as u32 * u8::BITS,
),
_ => return Err(CUresult::CUDA_ERROR_INVALID_VALUE),
};
bits.0 *= bit_width;
bits.1 *= bit_width;
bits.2 *= bit_width;
bits.3 *= bit_width;
Ok(hipChannelFormatDesc {
x: bits.0 as i32,
y: bits.0 as i32,
z: bits.0 as i32,
w: bits.0 as i32,
f: kind,
})
}

4
zluda/tests/kernel_suld.rs
View file

@ -340,10 +340,6 @@ unsafe fn kernel_suld_impl<
if mem::size_of::<Format>() * CHANNELS < mem::size_of::<SustType>() * SULD_N {
return;
}
// TODO: reenable those tests
if mem::size_of::<Format>() != mem::size_of::<SustType>() || CHANNELS != SULD_N {
return;
}
let mut rng = rand_chacha::ChaCha8Rng::seed_from_u64(seed);
let size = 4usize;
let random_size = rand::distributions::Uniform::<u32>::new(1, size as u32);

12
zluda/tests/kernel_sust.rs
View file

@ -312,7 +312,9 @@ unsafe fn byte_fill<T: Copy>(vec: &mut Vec<T>, value: u8) {
fn extend_bytes_with(slice: &[u8], elm: u8, desired_length: usize) -> Vec<u8> {
let mut result = slice.to_vec();
result.extend(std::iter::repeat(elm).take(desired_length - slice.len()));
if desired_length > slice.len() {
result.extend(std::iter::repeat(elm).take(desired_length - slice.len()));
}
result
}
@ -337,10 +339,6 @@ unsafe fn kernel_sust_impl<
if mem::size_of::<Format>() * CHANNELS < mem::size_of::<SustType>() * SUST_N {
return;
}
// TODO: reenable those tests
if mem::size_of::<Format>() != mem::size_of::<SustType>() || CHANNELS != SUST_N {
return;
}
let mut rng = rand_chacha::ChaCha8Rng::seed_from_u64(seed);
let size = 4usize;
let random_size = rand::distributions::Uniform::<u32>::new(1, size as u32);
@ -461,4 +459,8 @@ unsafe fn kernel_sust_impl<
assert_eq!(expected, &*observed);
let mut unused = mem::zeroed();
assert_eq!(cuda.cuCtxPopCurrent(&mut unused), CUresult::CUDA_SUCCESS);
assert_eq!(
cuda.cuDevicePrimaryCtxRelease_v2(CUdevice_v1(0)),
CUresult::CUDA_SUCCESS
);
}