Merge pull request #2382 from jackchentwkh/vsh_cpu_2

Implement HLE D3DDevice_RunVertexStateShader()

.gitmodules

@@ -39,3 +39,6 @@
 	path = import/libusb
 	url = https://github.com/libusb/libusb
 	shallow = true
+[submodule "import/nv2a_vsh_cpu"]
+	path = import/nv2a_vsh_cpu
+	url = https://github.com/abaire/nv2a_vsh_cpu.git

CMakeLists.txt

@@ -39,6 +39,9 @@ add_subdirectory("${CMAKE_CURRENT_LIST_DIR}/projects/imgui")
 
 add_subdirectory("${CMAKE_CURRENT_LIST_DIR}/projects/libusb")
 
+set(nv2a_vsh_cpu_UNIT_TEST OFF)
+add_subdirectory("${CMAKE_CURRENT_LIST_DIR}/import/nv2a_vsh_cpu" EXCLUDE_FROM_ALL)
+
 # Split the files into group for which project is likely
 # going to be used for both header and source files.
 # Then move only specific project files into their
@@ -437,6 +440,10 @@ set_target_properties(cxbx cxbxr-ldr cxbxr-emu misc-batch SDL2 subhook libXbSymb
  PROPERTIES FOLDER Cxbx-Reloaded
 )
 
+set_target_properties(nv2a_vsh_emulator nv2a_vsh_disassembler nv2a_vsh_cpu
+ PROPERTIES FOLDER Cxbx-Reloaded/nv2a_vsh
+)
+
 if ("${CMAKE_CXX_COMPILER_ID}" STREQUAL "MSVC")
  # Configure startup project
  set_property(DIRECTORY "${CMAKE_CURRENT_LIST_DIR}" PROPERTY VS_STARTUP_PROJECT cxbx)

import/nv2a_vsh_cpu

@@ -0,0 +1 @@
+Subproject commit ead0af5dee49e408c005151393f8a7dbcd27026c

projects/cxbxr-emu/CMakeLists.txt

@@ -171,6 +171,7 @@ target_link_libraries(cxbxr-emu
  SDL2
  imgui
  libusb
+ nv2a_vsh_emulator
 
  ${WINS_LIB}
 )

src/core/hle/D3D8/Direct3D9/Direct3D9.cpp

@@ -78,6 +78,9 @@
 #include <thread>
 
 #include <wrl/client.h>
+
+#include "nv2a_vsh_emulator.h"
+
 using namespace Microsoft::WRL;
 
 XboxRenderStateConverter XboxRenderStates;
@@ -8660,6 +8663,8 @@ xbox::void_xt WINAPI xbox::EMUPATCH(D3DDevice_SetVertexShaderInput)
 		XB_TRMP(D3DDevice_SetVertexShaderInput)(Handle, StreamCount, pStreamInputs);
 }
 
+extern xbox::dword_xt* GetCxbxVertexShaderSlotPtr(const DWORD SlotIndexAddress); // tmp glue
+
 // ******************************************************************
 // * patch: D3DDevice_RunVertexStateShader
 // ******************************************************************
@@ -8676,8 +8681,41 @@ xbox::void_xt WINAPI xbox::EMUPATCH(D3DDevice_RunVertexStateShader)
 
 	// If pData is assigned, pData[0..3] is pushed towards nv2a transform data registers
 	// then sends the nv2a a command to launch the vertex shader function located at Address
+	if (Address >= NV2A_MAX_TRANSFORM_PROGRAM_LENGTH) {
+		LOG_TEST_CASE("Address out of bounds");
+		return;
+	}
 
-    LOG_UNIMPLEMENTED(); 
+	NV2AState* dev = g_NV2A->GetDeviceState();
+	PGRAPHState* pg = &(dev->pgraph);
+
+	Nv2aVshProgram program = {}; // Note: This nulls program.steps
+	// TODO : Retain program globally and perform nv2a_vsh_parse_program only when
+	//        the address-range we're about to emulate was modified since last parse.
+	// TODO : As a suggestion for this, parse all NV2A_MAX_TRANSFORM_PROGRAM_LENGTH slots,
+	//        and here just point program.steps to global vsh_program_steps[Address].
+	Nv2aVshParseResult result = nv2a_vsh_parse_program(
+		&program, // Note : program.steps will be malloc'ed
+		GetCxbxVertexShaderSlotPtr(Address), // TODO : At some point, use pg->program_data[Address] here instead
+		NV2A_MAX_TRANSFORM_PROGRAM_LENGTH - Address);
+	if (result != NV2AVPR_SUCCESS) {
+		LOG_TEST_CASE("nv2a_vsh_parse_program failed");
+		// TODO : Dump Nv2aVshParseResult as string and program for debugging purposes
+		return;
+	}
+
+	Nv2aVshCPUXVSSExecutionState state_linkage;
+	Nv2aVshExecutionState state = nv2a_vsh_emu_initialize_xss_execution_state(
+		&state_linkage, (float*)pg->vsh_constants); // Note : This wil memset(state_linkage, 0)
+	if (pData != nullptr)
+		//if pData != nullptr, then it contains v0.xyzw, we shall copy the binary content directly.
+		memcpy(state_linkage.input_regs, pData, sizeof(state_linkage.input_regs));
+
+	nv2a_vsh_emu_execute_track_context_writes(&state, &program, pg->vsh_constants_dirty);
+	// Note: Above emulation's primary purpose is to update pg->vsh_constants and pg->vsh_constants_dirty
+	// therefor, nothing else needs to be done here, other than to cleanup
+
+	nv2a_vsh_program_destroy(&program); // Note: program.steps will be free'ed
 }
 
 // ******************************************************************