[Bf-blender-cvs] [51839c10d10] cycles-x: Cycles X: use less memory for float3 integrator state on GPU
Brecht Van Lommel
noreply at git.blender.org
Wed Jul 14 17:50:46 CEST 2021
Commit: 51839c10d109a6c84e4f107d264bc5cdfd0c228e
Author: Brecht Van Lommel
Date: Tue Jul 13 18:10:58 2021 +0200
Branches: cycles-x
https://developer.blender.org/rB51839c10d109a6c84e4f107d264bc5cdfd0c228e
Cycles X: use less memory for float3 integrator state on GPU
Allocate different device_only_memory size depending if the device is CPU
or GPU, since for GPU we don't align to 16 bytes for SSE.
Also adds some sanity checks and ensure float3 is not used in device_vector
since it's incompatible for sharing data between CPU and GPU.
Differential Revision: https://developer.blender.org/D11913
===================================================================
M intern/cycles/device/device_memory.cpp
M intern/cycles/device/device_memory.h
M intern/cycles/device/optix/device_impl.cpp
===================================================================
diff --git a/intern/cycles/device/device_memory.cpp b/intern/cycles/device/device_memory.cpp
index 80a05fc32fe..c4d45829b83 100644
--- a/intern/cycles/device/device_memory.cpp
+++ b/intern/cycles/device/device_memory.cpp
@@ -23,7 +23,7 @@ CCL_NAMESPACE_BEGIN
device_memory::device_memory(Device *device, const char *name, MemoryType type)
: data_type(device_type_traits<uchar>::data_type),
- data_elements(device_type_traits<uchar>::num_elements),
+ data_elements(device_type_traits<uchar>::num_elements_cpu),
data_size(0),
device_size(0),
data_width(0),
@@ -149,6 +149,11 @@ void device_memory::device_zero()
}
}
+bool device_memory::device_is_cpu()
+{
+ return (device->info.type == DEVICE_CPU);
+}
+
void device_memory::swap_device(Device *new_device,
size_t new_device_size,
device_ptr new_device_ptr)
diff --git a/intern/cycles/device/device_memory.h b/intern/cycles/device/device_memory.h
index 9906812c888..c51594b8580 100644
--- a/intern/cycles/device/device_memory.h
+++ b/intern/cycles/device/device_memory.h
@@ -53,7 +53,7 @@ enum DataType {
TYPE_UINT64,
};
-static inline size_t datatype_size(DataType datatype)
+static constexpr size_t datatype_size(DataType datatype)
{
switch (datatype) {
case TYPE_UNKNOWN:
@@ -81,112 +81,155 @@ static inline size_t datatype_size(DataType datatype)
template<typename T> struct device_type_traits {
static const DataType data_type = TYPE_UNKNOWN;
- static const int num_elements = sizeof(T);
+ static const int num_elements_cpu = sizeof(T);
+ static const int num_elements_gpu = sizeof(T);
};
template<> struct device_type_traits<uchar> {
static const DataType data_type = TYPE_UCHAR;
- static const int num_elements = 1;
+ static const int num_elements_cpu = 1;
+ static const int num_elements_gpu = 1;
+ static_assert(sizeof(uchar) == num_elements_cpu * datatype_size(data_type));
};
template<> struct device_type_traits<uchar2> {
static const DataType data_type = TYPE_UCHAR;
- static const int num_elements = 2;
+ static const int num_elements_cpu = 2;
+ static const int num_elements_gpu = 2;
+ static_assert(sizeof(uchar2) == num_elements_cpu * datatype_size(data_type));
};
template<> struct device_type_traits<uchar3> {
static const DataType data_type = TYPE_UCHAR;
- static const int num_elements = 3;
+ static const int num_elements_cpu = 3;
+ static const int num_elements_gpu = 3;
+ static_assert(sizeof(uchar3) == num_elements_cpu * datatype_size(data_type));
};
template<> struct device_type_traits<uchar4> {
static const DataType data_type = TYPE_UCHAR;
- static const int num_elements = 4;
+ static const int num_elements_cpu = 4;
+ static const int num_elements_gpu = 4;
+ static_assert(sizeof(uchar4) == num_elements_cpu * datatype_size(data_type));
};
template<> struct device_type_traits<uint> {
static const DataType data_type = TYPE_UINT;
- static const int num_elements = 1;
+ static const int num_elements_cpu = 1;
+ static const int num_elements_gpu = 1;
+ static_assert(sizeof(uint) == num_elements_cpu * datatype_size(data_type));
};
template<> struct device_type_traits<uint2> {
static const DataType data_type = TYPE_UINT;
- static const int num_elements = 2;
+ static const int num_elements_cpu = 2;
+ static const int num_elements_gpu = 2;
+ static_assert(sizeof(uint2) == num_elements_cpu * datatype_size(data_type));
};
template<> struct device_type_traits<uint3> {
static const DataType data_type = TYPE_UINT;
- static const int num_elements = 3;
+ static const int num_elements_cpu = 3;
+ static const int num_elements_gpu = 3;
+ static_assert(sizeof(uint3) == num_elements_cpu * datatype_size(data_type));
};
template<> struct device_type_traits<uint4> {
static const DataType data_type = TYPE_UINT;
- static const int num_elements = 4;
+ static const int num_elements_cpu = 4;
+ static const int num_elements_gpu = 4;
+ static_assert(sizeof(uint4) == num_elements_cpu * datatype_size(data_type));
};
template<> struct device_type_traits<int> {
static const DataType data_type = TYPE_INT;
- static const int num_elements = 1;
+ static const int num_elements_cpu = 1;
+ static const int num_elements_gpu = 1;
+ static_assert(sizeof(int) == num_elements_cpu * datatype_size(data_type));
};
template<> struct device_type_traits<int2> {
static const DataType data_type = TYPE_INT;
- static const int num_elements = 2;
+ static const int num_elements_cpu = 2;
+ static const int num_elements_gpu = 2;
+ static_assert(sizeof(int2) == num_elements_cpu * datatype_size(data_type));
};
template<> struct device_type_traits<int3> {
static const DataType data_type = TYPE_INT;
- static const int num_elements = 3;
+ static const int num_elements_cpu = 4;
+ static const int num_elements_gpu = 3;
+ static_assert(sizeof(int3) == num_elements_cpu * datatype_size(data_type));
};
template<> struct device_type_traits<int4> {
static const DataType data_type = TYPE_INT;
- static const int num_elements = 4;
+ static const int num_elements_cpu = 4;
+ static const int num_elements_gpu = 4;
+ static_assert(sizeof(int4) == num_elements_cpu * datatype_size(data_type));
};
template<> struct device_type_traits<float> {
static const DataType data_type = TYPE_FLOAT;
- static const int num_elements = 1;
+ static const int num_elements_cpu = 1;
+ static const int num_elements_gpu = 1;
+ static_assert(sizeof(float) == num_elements_cpu * datatype_size(data_type));
};
template<> struct device_type_traits<float2> {
static const DataType data_type = TYPE_FLOAT;
- static const int num_elements = 2;
+ static const int num_elements_cpu = 2;
+ static const int num_elements_gpu = 2;
+ static_assert(sizeof(float2) == num_elements_cpu * datatype_size(data_type));
};
template<> struct device_type_traits<float3> {
static const DataType data_type = TYPE_FLOAT;
- static const int num_elements = 4;
+ static const int num_elements_cpu = 4;
+ static const int num_elements_gpu = 3;
+ static_assert(sizeof(float3) == num_elements_cpu * datatype_size(data_type));
};
template<> struct device_type_traits<float4> {
static const DataType data_type = TYPE_FLOAT;
- static const int num_elements = 4;
+ static const int num_elements_cpu = 4;
+ static const int num_elements_gpu = 4;
+ static_assert(sizeof(float4) == num_elements_cpu * datatype_size(data_type));
};
template<> struct device_type_traits<half> {
static const DataType data_type = TYPE_HALF;
- static const int num_elements = 1;
+ static const int num_elements_cpu = 1;
+ static const int num_elements_gpu = 1;
+ static_assert(sizeof(half) == num_elements_cpu * datatype_size(data_type));
};
template<> struct device_type_traits<ushort4> {
static const DataType data_type = TYPE_UINT16;
- static const int num_elements = 4;
+ static const int num_elements_cpu = 4;
+ static const int num_elements_gpu = 4;
+ static_assert(sizeof(ushort4) == num_elements_cpu * datatype_size(data_type));
};
template<> struct device_type_traits<uint16_t> {
static const DataType data_type = TYPE_UINT16;
- static const int num_elements = 1;
+ static const int num_elements_cpu = 1;
+ static const int num_elements_gpu = 1;
+ static_assert(sizeof(uint16_t) == num_elements_cpu * datatype_size(data_type));
};
template<> struct device_type_traits<half4> {
static const DataType data_type = TYPE_HALF;
- static const int num_elements = 4;
+ static const int num_elements_cpu = 4;
+ static const int num_elements_gpu = 4;
+ static_assert(sizeof(half4) == num_elements_cpu * datatype_size(data_type));
};
template<> struct device_type_traits<uint64_t> {
static const DataType data_type = TYPE_UINT64;
- static const int num_elements = 1;
+ static const int num_elements_cpu = 1;
+ static const int num_elements_gpu = 1;
+ static_assert(sizeof(uint64_t) == num_elements_cpu * datatype_size(data_type));
};
/* Device Memory
@@ -256,6 +299,8 @@ class device_memory {
void device_copy_from(int y, int w, int h, int elem);
void device_zero();
+ bool device_is_cpu();
+
device_ptr original_device_ptr;
size_t original_device_size;
Device *original_device;
@@ -274,7 +319,9 @@ template<typename T> class device_only_memory : public device_memory {
: device_memory(device, name, allow_host_memory_fallback ? MEM_READ_WRITE : MEM_DEVICE_ONLY)
{
data_type = device_type_traits<T>::data_type;
- data_elements = max(device_type_traits<T>::num_elements, 1);
+ data_elements = max(device_is_cpu() ? device_type_traits<T>::num_elements_cpu :
+ device_type_traits<T>::num_elements_gpu,
+ 1);
}
device_only_memory(device_only_memory &&other) noexcept : device_memory(std::move(other))
@@ -330,11 +377,15 @@ template<typename T> class device_only_memory : public device_memory {
template<typename T> class device_vector : public device_memory {
public:
+ /* Can only use this for types that have the same size on CPU and GPU. */
+ static_assert(device_type_traits<T>::num_elements_cpu ==
+ device_type_traits<T>::num_elements_gpu);
+
device_vector(Device *device, const char *name, MemoryType type)
: device_memory(device, name, type)
{
data_type = device_type_traits<T>::data_type;
- data_elements = device_type_traits<T>::num_elements;
+ data_elements = device_type_traits<T>::num_elements_cpu;
modified = true;
need_realloc_ = true;
diff --git a/intern/cycles/device/optix/device_impl.cpp b/intern/cycles/device/optix/device_impl.cpp
index 4907be6570f..38ee936e7e8 100644
--- a/intern/cycles/device/optix/device_impl.cpp
+++ b/intern/cycles/device/optix/device_impl.cpp
@@ -1368,7 +1368,7 @@ void OptiXDevice::build_bvh(BVH *bvh, Progress &progress, bool refit)
memcpy(index_data.data(),
mesh->get_triangles().data(),
mesh->get_triangles().size() * sizeof(int));
- device_vector<float3> vertex_data(this, "optix temp vertex data", MEM_READ_ONLY);
+ device_vector<float4> verte
@@ Diff output truncated at 10240 characters. @@
More information about the Bf-blender-cvs
mailing list