DevVGA-SVGA3d-win-d3d9.cpp@ 105745

Last change on this file since 105745 was 98103, checked in by vboxsync, 23 months ago
Copyright year updates by scm.
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 50.8 KB

Line
1	/* $Id: DevVGA-SVGA3d-win-d3d9.cpp 98103 2023-01-17 14:15:46Z vboxsync $ */
2	/** @file
3	* DevVMWare - VMWare SVGA device Direct3D 9 backend.
4	*/
5
6	/*
7	* Copyright (C) 2019-2023 Oracle and/or its affiliates.
8	*
9	* This file is part of VirtualBox base platform packages, as
10	* available from https://www.virtualbox.org.
11	*
12	* This program is free software; you can redistribute it and/or
13	* modify it under the terms of the GNU General Public License
14	* as published by the Free Software Foundation, in version 3 of the
15	* License.
16	*
17	* This program is distributed in the hope that it will be useful, but
18	* WITHOUT ANY WARRANTY; without even the implied warranty of
19	* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
20	* General Public License for more details.
21	*
22	* You should have received a copy of the GNU General Public License
23	* along with this program; if not, see <https://www.gnu.org/licenses>.
24	*
25	* SPDX-License-Identifier: GPL-3.0-only
26	*/
27
28	#define LOG_GROUP LOG_GROUP_DEV_VMSVGA
29	#include <VBox/vmm/pdmdev.h>
30
31	#include <iprt/assert.h>
32
33	#include "DevVGA-SVGA.h"
34	#include "DevVGA-SVGA3d-internal.h"
35
36
37	typedef enum D3D9TextureType
38	{
39	D3D9TextureType_Texture,
40	D3D9TextureType_Bounce,
41	D3D9TextureType_Emulated
42	} D3D9TextureType;
43
44	DECLINLINE(D3DCUBEMAP_FACES) vmsvga3dCubemapFaceFromIndex(uint32_t iFace)
45	{
46	D3DCUBEMAP_FACES Face;
47	switch (iFace)
48	{
49	case 0: Face = D3DCUBEMAP_FACE_POSITIVE_X; break;
50	case 1: Face = D3DCUBEMAP_FACE_NEGATIVE_X; break;
51	case 2: Face = D3DCUBEMAP_FACE_POSITIVE_Y; break;
52	case 3: Face = D3DCUBEMAP_FACE_NEGATIVE_Y; break;
53	case 4: Face = D3DCUBEMAP_FACE_POSITIVE_Z; break;
54	default:
55	case 5: Face = D3DCUBEMAP_FACE_NEGATIVE_Z; break;
56	}
57	return Face;
58	}
59
60	IDirect3DTexture9 *D3D9GetTexture(PVMSVGA3DSURFACE pSurface,
61	D3D9TextureType enmType)
62	{
63	IDirect3DTexture9 *p;
64	switch (enmType)
65	{
66	default: AssertFailed();
67	RT_FALL_THRU();
68	case D3D9TextureType_Texture: p = pSurface->u.pTexture; break;
69	case D3D9TextureType_Bounce: p = pSurface->bounce.pTexture; break;
70	case D3D9TextureType_Emulated: p = pSurface->emulated.pTexture; break;
71	}
72	return p;
73	}
74
75	IDirect3DCubeTexture9 *D3D9GetCubeTexture(PVMSVGA3DSURFACE pSurface,
76	D3D9TextureType enmType)
77	{
78	IDirect3DCubeTexture9 *p;
79	switch (enmType)
80	{
81	default: AssertFailed();
82	RT_FALL_THRU();
83	case D3D9TextureType_Texture: p = pSurface->u.pCubeTexture; break;
84	case D3D9TextureType_Bounce: p = pSurface->bounce.pCubeTexture; break;
85	case D3D9TextureType_Emulated: p = pSurface->emulated.pCubeTexture; break;
86	}
87	return p;
88	}
89
90	IDirect3DVolumeTexture9 *D3D9GetVolumeTexture(PVMSVGA3DSURFACE pSurface,
91	D3D9TextureType enmType)
92	{
93	IDirect3DVolumeTexture9 *p;
94	switch (enmType)
95	{
96	default: AssertFailed();
97	RT_FALL_THRU();
98	case D3D9TextureType_Texture: p = pSurface->u.pVolumeTexture; break;
99	case D3D9TextureType_Bounce: p = pSurface->bounce.pVolumeTexture; break;
100	case D3D9TextureType_Emulated: p = pSurface->emulated.pVolumeTexture; break;
101	}
102	return p;
103	}
104
105	HRESULT D3D9GetTextureLevel(PVMSVGA3DSURFACE pSurface,
106	D3D9TextureType enmType,
107	uint32_t uFace,
108	uint32_t uMipmap,
109	IDirect3DSurface9 **ppD3DSurface)
110	{
111	HRESULT hr;
112
113	if (pSurface->enmD3DResType == VMSVGA3D_D3DRESTYPE_CUBE_TEXTURE)
114	{
115	Assert(pSurface->cFaces == 6);
116
117	IDirect3DCubeTexture9 *p = D3D9GetCubeTexture(pSurface, enmType);
118	if (p)
119	{
120	D3DCUBEMAP_FACES const FaceType = vmsvga3dCubemapFaceFromIndex(uFace);
121	hr = p->GetCubeMapSurface(FaceType, uMipmap, ppD3DSurface);
122	AssertMsg(hr == D3D_OK, ("GetCubeMapSurface failed with %x\n", hr));
123	}
124	else
125	{
126	AssertFailed();
127	hr = E_INVALIDARG;
128	}
129	}
130	else if (pSurface->enmD3DResType == VMSVGA3D_D3DRESTYPE_TEXTURE)
131	{
132	Assert(pSurface->cFaces == 1);
133	Assert(uFace == 0);
134
135	IDirect3DTexture9 *p = D3D9GetTexture(pSurface, enmType);
136	if (p)
137	{
138	hr = p->GetSurfaceLevel(uMipmap, ppD3DSurface);
139	AssertMsg(hr == D3D_OK, ("GetSurfaceLevel failed with %x\n", hr));
140	}
141	else
142	{
143	AssertFailed();
144	hr = E_INVALIDARG;
145	}
146	}
147	else
148	{
149	AssertMsgFailed(("No surface level for type %d\n", pSurface->enmD3DResType));
150	hr = E_INVALIDARG;
151	}
152
153	return hr;
154	}
155
156
157	static HRESULT d3dCopyToVertexBuffer(IDirect3DVertexBuffer9 pVB, const void pvSrc, int cbSrc)
158	{
159	HRESULT hr = D3D_OK;
160	void *pvDst = 0;
161	hr = pVB->Lock(0, 0, &pvDst, 0);
162	if (SUCCEEDED(hr))
163	{
164	memcpy(pvDst, pvSrc, cbSrc);
165	hr = pVB->Unlock();
166	}
167	return hr;
168	}
169
170	struct Vertex
171	{
172	float x, y; /* The vertex position. */
173	float u, v; /* Texture coordinates. */
174	};
175
176	typedef struct D3D9ConversionParameters
177	{
178	DWORD const paVS; / Vertex shader code. */
179	DWORD const paPS; / Pixel shader code. */
180	} D3D9ConversionParameters;
181
182	/** Select conversion shaders.
183	* @param d3dfmtFrom Source texture format.
184	* @param d3dfmtTo Target texture format.
185	* @param pResult Where the tore pointers to the shader code.
186	*/
187	static HRESULT d3d9SelectConversion(D3DFORMAT d3dfmtFrom, D3DFORMAT d3dfmtTo,
188	D3D9ConversionParameters *pResult)
189	{
190	/*
191	* The shader code has been obtained from the hex listing file (hexdump.txt) produced by fxc HLSL compiler:
192	* fxc.exe /Op /Tfx_2_0 /Fxhexdump.txt shader.fx
193	*
194	* The vertex shader code is the same for all convestion variants.
195	*
196	* For example here is the complete code for aPSCodeSwapRB:
197
198	uniform extern float4 gTextureInfo; // .xy = (width, height) in pixels, .zw = (1/width, 1/height)
199	uniform extern texture gTexSource;
200	sampler sSource = sampler_state
201	{
202	Texture = <gTexSource>;
203	};
204
205	struct VS_INPUT
206	{
207	float2 Position : POSITION; // In pixels.
208	float2 TexCoord : TEXCOORD0;
209	};
210
211	struct VS_OUTPUT
212	{
213	float4 Position : POSITION; // Normalized.
214	float2 TexCoord : TEXCOORD0;
215	};
216
217	VS_OUTPUT VS(VS_INPUT In)
218	{
219	VS_OUTPUT Output;
220
221	// Position is in pixels, i.e [0; width - 1]. Top, left is 0,0.
222	// Convert to the normalized coords in the -1;1 range
223	float4 Position;
224	Position.x = 2.0f * In.Position.x / (gTextureInfo.x - 1.0f) - 1.0f;
225	Position.y = -2.0f * In.Position.y / (gTextureInfo.y - 1.0f) + 1.0f;
226	Position.z = 0.0f; // Not used.
227	Position.w = 1.0f; // It is a point.
228
229	Output.Position = Position;
230	Output.TexCoord = In.TexCoord;
231
232	return Output;
233	}
234
235	struct PS_OUTPUT
236	{
237	float4 Color : COLOR0;
238	};
239
240	PS_OUTPUT PS(VS_OUTPUT In)
241	{
242	PS_OUTPUT Output;
243
244	float2 texCoord = In.TexCoord;
245
246	float4 texColor = tex2D(sSource, texCoord);
247
248	Output.Color = texColor.bgra; // Swizzle rgba -> bgra
249
250	return Output;
251	}
252
253	technique RenderScene
254	{
255	pass P0
256	{
257	VertexShader = compile vs_2_0 VS();
258	PixelShader = compile ps_2_0 PS();
259	}
260	}
261	*/
262
263	static DWORD const aVSCode[] =
264	{
265	0xfffe0200, // vs_2_0
266	0x05000051, 0xa00f0001, 0xbf800000, 0xc0000000, 0x3f800000, 0x00000000, // def c1, -1, -2, 1, 0
267	0x0200001f, 0x80000000, 0x900f0000, // dcl_position v0
268	0x0200001f, 0x80000005, 0x900f0001, // dcl_texcoord v1
269	0x03000002, 0x80010000, 0x90000000, 0x90000000, // add r0.x, v0.x, v0.x
270	0x02000001, 0x80010001, 0xa0000001, // mov r1.x, c1.x
271	0x03000002, 0x80060000, 0x80000001, 0xa0d00000, // add r0.yz, r1.x, c0.xxyw
272	0x02000006, 0x80020000, 0x80550000, // rcp r0.y, r0.y
273	0x02000006, 0x80040000, 0x80aa0000, // rcp r0.z, r0.z
274	0x04000004, 0xc0010000, 0x80000000, 0x80550000, 0xa0000001, // mad oPos.x, r0.x, r0.y, c1.x
275	0x03000005, 0x80010000, 0x90550000, 0xa0550001, // mul r0.x, v0.y, c1.y
276	0x04000004, 0xc0020000, 0x80000000, 0x80aa0000, 0xa0aa0001, // mad oPos.y, r0.x, r0.z, c1.z
277	0x02000001, 0xc00c0000, 0xa0b40001, // mov oPos.zw, c1.xywz
278	0x02000001, 0xe0030000, 0x90e40001, // mov oT0.xy, v1
279	0x0000ffff
280	};
281
282	/*
283	* Swap R and B components. Converts D3DFMT_A8R8G8B8 <-> D3DFMT_A8B8G8R8.
284	*/
285	static DWORD const aPSCodeSwapRB[] =
286	{
287	0xffff0200, // ps_2_0
288	0x0200001f, 0x80000000, 0xb0030000, // dcl t0.xy
289	0x0200001f, 0x90000000, 0xa00f0800, // dcl_2d s0
290	0x03000042, 0x800f0000, 0xb0e40000, 0xa0e40800, // texld r0, t0, s0
291	0x02000001, 0x80090001, 0x80d20000, // mov r1.xw, r0.zxyw
292	0x02000001, 0x80040001, 0x80000000, // mov r1.z, r0.x
293	0x02000001, 0x80020001, 0x80550000, // mov r1.y, r0.y
294	0x02000001, 0x800f0800, 0x80e40001, // mov oC0, r1
295	0x0000ffff
296	};
297
298	/* YUY2 to RGB:
299
300	// YUY2 if not defined
301	// #define UYVY
302
303	static const float3x3 yuvCoeffs =
304	{
305	1.164383f, 1.164383f, 1.164383f,
306	0.0f, -0.391762f, 2.017232f,
307	1.596027f, -0.812968f, 0.0f
308	};
309
310	PS_OUTPUT PS(VS_OUTPUT In)
311	{
312	PS_OUTPUT Output;
313
314	// 4 bytes of an YUV macropixel contain 2 pixels in X for the target.
315	// I.e. each YUV texture pixel is sampled twice: for both even and odd target pixels.
316
317	// In.TexCoord are in [0;1] range for the target.
318	float2 texCoord = In.TexCoord;
319
320	// Convert to the target coords in pixels: xPixel = TexCoord.x * Width.
321	float xTargetPixel = texCoord.x * gTextureInfo.x;
322
323	// Source texture is half width, i.e. it contains data in pixels [0; width / 2 - 1].
324	float xSourcePixel = xTargetPixel / 2.0f;
325
326	// Remainder is about 0.25 for even pixels and about 0.75 for odd pixels.
327	float remainder = xSourcePixel - trunc(xSourcePixel);
328
329	// Back to the normalized coords: texCoord.x = xPixel / Width.
330	texCoord.x = xSourcePixel * gTextureInfo.z;
331
332	// Fetch YUV
333	float4 texColor = tex2D(sSource, texCoord);
334
335	// Get YUV components.
336	#ifdef UYVY
337	float u = texColor.b;
338	float y0 = texColor.g;
339	float v = texColor.r;
340	float y1 = texColor.a;
341	#else // YUY2
342	float y0 = texColor.b;
343	float u = texColor.g;
344	float y1 = texColor.r;
345	float v = texColor.a;
346	#endif
347
348	// Get y0 for even x coordinates and y1 for odd ones.
349	float y = remainder < 0.5f ? y0 : y1;
350
351	// Make a vector for easier calculation.
352	float3 yuv = float3(y, u, v);
353
354	// Convert YUV to RGB:
355	// https://docs.microsoft.com/en-us/windows/win32/medfound/recommended-8-bit-yuv-formats-for-video-rendering#converting-8-bit-yuv-to-rgb888
356	//
357	// For 8bit [0;255] when Y = [16;235], U,V = [16;239]:
358	//
359	// C = Y - 16
360	// D = U - 128
361	// E = V - 128
362	//
363	// R = 1.164383 * C + 1.596027 * E
364	// G = 1.164383 * C - 0.391762 * D - 0.812968 * E
365	// B = 1.164383 * C + 2.017232 * D
366	//
367	// For shader values [0;1.0] when Y = [16/255;235/255], U,V = [16/255;239/255]:
368	//
369	// C = Y - 0.0627
370	// D = U - 0.5020
371	// E = V - 0.5020
372	//
373	// R = 1.164383 * C + 1.596027 * E
374	// G = 1.164383 * C - 0.391762 * D - 0.812968 * E
375	// B = 1.164383 * C + 2.017232 * D
376	//
377	yuv -= float3(0.0627f, 0.502f, 0.502f);
378	float3 bgr = mul(yuv, yuvCoeffs);
379
380	// Clamp to [0;1]
381	bgr = saturate(bgr);
382
383	// Return RGBA
384	Output.Color = float4(bgr, 1.0f);
385
386	return Output;
387	}
388	*/
389	static DWORD const aPSCodeYUY2toRGB[] =
390	{
391	0xffff0200, // ps_2_0
392	0x05000051, 0xa00f0001, 0x3f000000, 0x00000000, 0x3f800000, 0xbf000000, // def c1, 0.5, 0, 1, -0.5
393	0x05000051, 0xa00f0002, 0xbd8068dc, 0xbf008312, 0xbf008312, 0x00000000, // def c2, -0.0627000034, -0.501999974, -0.501999974, 0
394	0x05000051, 0xa00f0003, 0x3f950a81, 0x00000000, 0x3fcc4a9d, 0x00000000, // def c3, 1.16438305, 0, 1.59602702, 0
395	0x05000051, 0xa00f0004, 0x3f950a81, 0xbec89507, 0xbf501eac, 0x00000000, // def c4, 1.16438305, -0.391761988, -0.812968016, 0
396	0x05000051, 0xa00f0005, 0x3f950a81, 0x40011a54, 0x00000000, 0x00000000, // def c5, 1.16438305, 2.01723194, 0, 0
397	0x0200001f, 0x80000000, 0xb0030000, // dcl t0.xy
398	0x0200001f, 0x90000000, 0xa00f0800, // dcl_2d s0
399	0x03000005, 0x80080000, 0xb0000000, 0xa0000000, // mul r0.w, t0.x, c0.x
400	0x03000005, 0x80010000, 0x80ff0000, 0xa0000001, // mul r0.x, r0.w, c1.x
401	0x02000013, 0x80020000, 0x80000000, // frc r0.y, r0.x
402	0x04000058, 0x80040000, 0x81550000, 0xa0550001, 0xa0aa0001, // cmp r0.z, -r0.y, c1.y, c1.z
403	0x03000002, 0x80020000, 0x80000000, 0x81550000, // add r0.y, r0.x, -r0.y
404	0x03000005, 0x80010001, 0x80000000, 0xa0aa0000, // mul r1.x, r0.x, c0.z
405	0x04000058, 0x80010000, 0x80ff0000, 0xa0550001, 0x80aa0000, // cmp r0.x, r0.w, c1.y, r0.z
406	0x03000002, 0x80010000, 0x80000000, 0x80550000, // add r0.x, r0.x, r0.y
407	0x04000004, 0x80010000, 0x80ff0000, 0xa0000001, 0x81000000, // mad r0.x, r0.w, c1.x, -r0.x
408	0x03000002, 0x80010000, 0x80000000, 0xa0ff0001, // add r0.x, r0.x, c1.w
409	0x02000001, 0x80020001, 0xb0550000, // mov r1.y, t0.y
410	0x03000042, 0x800f0001, 0x80e40001, 0xa0e40800, // texld r1, r1, s0
411	0x04000058, 0x80010001, 0x80000000, 0x80000001, 0x80aa0001, // cmp r1.x, r0.x, r1.x, r1.z
412	0x02000001, 0x80040001, 0x80ff0001, // mov r1.z, r1.w
413	0x03000002, 0x80070000, 0x80e40001, 0xa0e40002, // add r0.xyz, r1, c2
414	0x03000008, 0x80110001, 0x80e40000, 0xa0e40003, // dp3_sat r1.x, r0, c3
415	0x03000008, 0x80120001, 0x80e40000, 0xa0e40004, // dp3_sat r1.y, r0, c4
416	0x0400005a, 0x80140001, 0x80e40000, 0xa0e40005, 0xa0aa0005, // dp2add_sat r1.z, r0, c5, c5.z
417	0x02000001, 0x80080001, 0xa0aa0001, // mov r1.w, c1.z
418	0x02000001, 0x800f0800, 0x80e40001, // mov oC0, r1
419	0x0000ffff
420	};
421
422	/* UYVY to RGB is same as YUY2 above except for the order of yuv components:
423
424	// YUY2 if not defined
425	#define UYVY
426	...
427	*/
428	static DWORD const aPSCodeUYVYtoRGB[] =
429	{
430	0xffff0200, // ps_2_0
431	0x05000051, 0xa00f0001, 0x3f000000, 0x00000000, 0x3f800000, 0xbf000000, // def c1, 0.5, 0, 1, -0.5
432	0x05000051, 0xa00f0002, 0xbd8068dc, 0xbf008312, 0xbf008312, 0x00000000, // def c2, -0.0627000034, -0.501999974, -0.501999974, 0
433	0x05000051, 0xa00f0003, 0x3f950a81, 0x00000000, 0x3fcc4a9d, 0x00000000, // def c3, 1.16438305, 0, 1.59602702, 0
434	0x05000051, 0xa00f0004, 0x3f950a81, 0xbec89507, 0xbf501eac, 0x00000000, // def c4, 1.16438305, -0.391761988, -0.812968016, 0
435	0x05000051, 0xa00f0005, 0x3f950a81, 0x40011a54, 0x00000000, 0x00000000, // def c5, 1.16438305, 2.01723194, 0, 0
436	0x0200001f, 0x80000000, 0xb0030000, // dcl t0.xy
437	0x0200001f, 0x90000000, 0xa00f0800, // dcl_2d s0
438	0x03000005, 0x80080000, 0xb0000000, 0xa0000000, // mul r0.w, t0.x, c0.x
439	0x03000005, 0x80010000, 0x80ff0000, 0xa0000001, // mul r0.x, r0.w, c1.x
440	0x02000013, 0x80020000, 0x80000000, // frc r0.y, r0.x
441	0x04000058, 0x80040000, 0x81550000, 0xa0550001, 0xa0aa0001, // cmp r0.z, -r0.y, c1.y, c1.z
442	0x03000002, 0x80020000, 0x80000000, 0x81550000, // add r0.y, r0.x, -r0.y
443	0x03000005, 0x80010001, 0x80000000, 0xa0aa0000, // mul r1.x, r0.x, c0.z
444	0x04000058, 0x80010000, 0x80ff0000, 0xa0550001, 0x80aa0000, // cmp r0.x, r0.w, c1.y, r0.z
445	0x03000002, 0x80010000, 0x80000000, 0x80550000, // add r0.x, r0.x, r0.y
446	0x04000004, 0x80010000, 0x80ff0000, 0xa0000001, 0x81000000, // mad r0.x, r0.w, c1.x, -r0.x
447	0x03000002, 0x80010000, 0x80000000, 0xa0ff0001, // add r0.x, r0.x, c1.w
448	0x02000001, 0x80020001, 0xb0550000, // mov r1.y, t0.y
449	0x03000042, 0x800f0001, 0x80e40001, 0xa0e40800, // texld r1, r1, s0
450	0x04000058, 0x80010000, 0x80000000, 0x80ff0001, 0x80550001, // cmp r0.x, r0.x, r1.w, r1.y
451	0x02000001, 0x80060000, 0x80c90001, // mov r0.yz, r1.yzxw
452	0x03000002, 0x80070000, 0x80e40000, 0xa0e40002, // add r0.xyz, r0, c2
453	0x03000008, 0x80110001, 0x80e40000, 0xa0e40003, // dp3_sat r1.x, r0, c3
454	0x03000008, 0x80120001, 0x80e40000, 0xa0e40004, // dp3_sat r1.y, r0, c4
455	0x0400005a, 0x80140001, 0x80e40000, 0xa0e40005, 0xa0aa0005, // dp2add_sat r1.z, r0, c5, c5.z
456	0x02000001, 0x80080001, 0xa0aa0001, // mov r1.w, c1.z
457	0x02000001, 0x800f0800, 0x80e40001, // mov oC0, r1
458	0x0000ffff
459	};
460
461	/* RGB to YUY2.
462	* UYVY is not defined.
463
464	static const float3x3 bgrCoeffs =
465	{
466	0.0977f, 0.4375f, -0.0703f,
467	0.5039f, -0.2891f, -0.3672f,
468	0.2578f, -0.1484f, 0.4375f
469	};
470
471	static const float3 yuvShift = { 0.0647f, 0.5039f, 0.5039f };
472
473	PS_OUTPUT PS(VS_OUTPUT In)
474	{
475	PS_OUTPUT Output;
476
477	// 4 bytes of an YUV macropixel contain 2 source pixels in X.
478	// I.e. each YUV texture target pixel is computed from 2 source pixels.
479	// The target texture pixels are located in the [0; width / 2 - 1] range.
480
481	// In.TexCoord are in [0;1] range, applicable both to the source and the target textures.
482	float2 texCoordDst = In.TexCoord;
483
484	// Convert to the target coords in pixels: xPixel = TexCoord.x * Width.
485	float xTargetPixel = texCoordDst.x * gTextureInfo.x;
486
487	float4 bgraOutputPixel;
488	if (xTargetPixel < gTextureInfo.x / 2.0f)
489	{
490	// Target texture is half width, i.e. it contains data in pixels [0; width / 2 - 1].
491	// Compute the source texture coords for the pixels which will be used to compute the target pixel.
492	float2 texCoordSrc = texCoordDst;
493	texCoordSrc.x *= 2.0f;
494
495	// Even pixel. Fetch two BGRA source pixels.
496	float4 texColor0 = tex2D(sSource, texCoordSrc);
497
498	// Advance one pixel (+ 1/Width)
499	texCoordSrc.x += gTextureInfo.z;
500	float4 texColor1 = tex2D(sSource, texCoordSrc);
501
502	// Compute y0, u, y1, v components
503	// https://docs.microsoft.com/en-us/windows/win32/medfound/recommended-8-bit-yuv-formats-for-video-rendering#converting-rgb888-to-yuv-444
504	//
505	// For R,G,B and Y,U,V in [0;255]
506	// Y = ( ( 66 * R + 129 * G + 25 * B + 128) >> 8) + 16
507	// U = ( ( -38 * R - 74 * G + 112 * B + 128) >> 8) + 128
508	// V = ( ( 112 * R - 94 * G - 18 * B + 128) >> 8) + 128
509	//
510	// For r,g,b and y,u,v in [0;1.0]
511	// y = 0.2578 * r + 0.5039 * g + 0.0977 * b + 0.0647
512	// u = -0.1484 * r - 0.2891 * g + 0.4375 * b + 0.5039
513	// v = 0.4375 * r - 0.3672 * g - 0.0703 * b + 0.5039
514
515	float3 yuv0 = mul(texColor0.bgr, bgrCoeffs);
516	yuv0 -= yuvShift;
517
518	float3 yuv1 = mul(texColor1.bgr, bgrCoeffs);
519	yuv1 -= yuvShift;
520
521	float y0 = yuv0.b;
522	float u = (yuv0.g + yuv1.g) / 2.0f;
523	float y1 = yuv1.b;
524	float v = (yuv0.r + yuv1.r) / 2.0f;
525
526	#ifdef UYVY
527	bgraOutputPixel = float4(u, y0, v, y1);
528	#else // YUY2
529	bgraOutputPixel = float4(y0, u, y1, v);
530	#endif
531	}
532	else
533	{
534	// [width / 2; width - 1] pixels are not used. Set to something.
535	bgraOutputPixel = float4(0.0f, 0.0f, 0.0f, 0.0f);
536	}
537
538	// Clamp to [0;1]
539	bgraOutputPixel = saturate(bgraOutputPixel);
540
541	// Return RGBA
542	Output.Color = bgraOutputPixel;
543
544	return Output;
545	}
546	*/
547	static DWORD const aPSCodeRGBtoYUY2[] =
548	{
549	0xffff0200, // ps_2_0
550	0x05000051, 0xa00f0001, 0xbd84816f, 0xbf00ff97, 0xbf00ff97, 0x00000000, // def c1, -0.0647, -0.503899992, -0.503899992, 0
551	0x05000051, 0xa00f0002, 0xbe80ff97, 0x00000000, 0xbd04816f, 0x00000000, // def c2, -0.251949996, 0, -0.03235, 0
552	0x05000051, 0xa00f0003, 0x3dc816f0, 0x3f00ff97, 0x3e83fe5d, 0x00000000, // def c3, 0.0976999998, 0.503899992, 0.257800013, 0
553	0x05000051, 0xa00f0004, 0x3ee00000, 0xbe9404ea, 0xbe17f62b, 0x00000000, // def c4, 0.4375, -0.289099991, -0.148399994, 0
554	0x05000051, 0xa00f0005, 0xbd8ff972, 0xbebc01a3, 0x3ee00000, 0x00000000, // def c5, -0.0702999979, -0.367199987, 0.4375, 0
555	0x05000051, 0xa00f0006, 0x3f000000, 0x40000000, 0x3f800000, 0xbf00ff97, // def c6, 0.5, 2, 1, -0.503899992
556	0x05000051, 0xa00f0007, 0x3f000000, 0x3f800000, 0x3f000000, 0x00000000, // def c7, 0.5, 1, 0.5, 0
557	0x0200001f, 0x80000000, 0xb0030000, // dcl t0.xy
558	0x0200001f, 0x90000000, 0xa00f0800, // dcl_2d s0
559	0x03000005, 0x80030000, 0xb0e40000, 0xa0c90006, // mul r0.xy, t0, c6.yzxw
560	0x02000001, 0x80030001, 0xa0e40006, // mov r1.xy, c6
561	0x04000004, 0x80010002, 0xb0000000, 0x80550001, 0xa0aa0000, // mad r2.x, t0.x, r1.y, c0.z
562	0x02000001, 0x80020002, 0xb0550000, // mov r2.y, t0.y
563	0x03000042, 0x800f0000, 0x80e40000, 0xa0e40800, // texld r0, r0, s0
564	0x03000042, 0x800f0002, 0x80e40002, 0xa0e40800, // texld r2, r2, s0
565	0x03000005, 0x80080000, 0x80aa0000, 0xa0000003, // mul r0.w, r0.z, c3.x
566	0x04000004, 0x80080000, 0x80550000, 0xa0550003, 0x80ff0000, // mad r0.w, r0.y, c3.y, r0.w
567	0x04000004, 0x80010003, 0x80000000, 0xa0aa0003, 0x80ff0000, // mad r3.x, r0.x, c3.z, r0.w
568	0x03000005, 0x80080000, 0x80aa0000, 0xa0000004, // mul r0.w, r0.z, c4.x
569	0x04000004, 0x80080000, 0x80550000, 0xa0550004, 0x80ff0000, // mad r0.w, r0.y, c4.y, r0.w
570	0x04000004, 0x80020003, 0x80000000, 0xa0aa0004, 0x80ff0000, // mad r3.y, r0.x, c4.z, r0.w
571	0x03000005, 0x80080002, 0x80aa0000, 0xa0000005, // mul r2.w, r0.z, c5.x
572	0x04000004, 0x80080002, 0x80550000, 0xa0550005, 0x80ff0002, // mad r2.w, r0.y, c5.y, r2.w
573	0x04000004, 0x80040003, 0x80000000, 0xa0aa0005, 0x80ff0002, // mad r3.z, r0.x, c5.z, r2.w
574	0x03000002, 0x80070000, 0x80e40003, 0xa0e40001, // add r0.xyz, r3, c1
575	0x02000001, 0x80080000, 0xa0ff0006, // mov r0.w, c6.w
576	0x03000005, 0x80080002, 0x80aa0002, 0xa0000003, // mul r2.w, r2.z, c3.x
577	0x04000004, 0x80080002, 0x80550002, 0xa0550003, 0x80ff0002, // mad r2.w, r2.y, c3.y, r2.w
578	0x04000004, 0x80040003, 0x80000002, 0xa0aa0003, 0x80ff0002, // mad r3.z, r2.x, c3.z, r2.w
579	0x03000005, 0x80080002, 0x80aa0002, 0xa0000004, // mul r2.w, r2.z, c4.x
580	0x04000004, 0x80080002, 0x80550002, 0xa0550004, 0x80ff0002, // mad r2.w, r2.y, c4.y, r2.w
581	0x04000004, 0x80010003, 0x80000002, 0xa0aa0004, 0x80ff0002, // mad r3.x, r2.x, c4.z, r2.w
582	0x03000005, 0x80080003, 0x80aa0002, 0xa0000005, // mul r3.w, r2.z, c5.x
583	0x04000004, 0x80080003, 0x80550002, 0xa0550005, 0x80ff0003, // mad r3.w, r2.y, c5.y, r3.w
584	0x04000004, 0x80020003, 0x80000002, 0xa0aa0005, 0x80ff0003, // mad r3.y, r2.x, c5.z, r3.w
585	0x03000002, 0x80050002, 0x80c90000, 0x80e40003, // add r2.xz, r0.yzxw, r3
586	0x03000002, 0x80020002, 0x80ff0000, 0x80550003, // add r2.y, r0.w, r3.y
587	0x02000001, 0x80110000, 0x80aa0000, // mov_sat r0.x, r0.z
588	0x02000001, 0x80070003, 0xa0e40007, // mov r3.xyz, c7
589	0x04000004, 0x80160000, 0x80d20002, 0x80d20003, 0xa0d20002, // mad_sat r0.yz, r2.zxyw, r3.zxyw, c2.zxyw
590	0x04000004, 0x80180000, 0x80aa0002, 0x80aa0003, 0xa0aa0002, // mad_sat r0.w, r2.z, r3.z, c2.z
591	0x03000005, 0x80010001, 0x80000001, 0xa0000000, // mul r1.x, r1.x, c0.x
592	0x04000004, 0x80010001, 0xb0000000, 0xa0000000, 0x81000001, // mad r1.x, t0.x, c0.x, -r1.x
593	0x04000058, 0x800f0000, 0x80000001, 0xa0ff0003, 0x80e40000, // cmp r0, r1.x, c3.w, r0
594	0x02000001, 0x800f0800, 0x80e40000, // mov oC0, r0
595	0x0000ffff
596	};
597
598	/* RGB to UYVY is same as YUY2 above except for the order of yuv components.
599	* UYVY is defined.
600	*/
601	static DWORD const aPSCodeRGBtoUYVY[] =
602	{
603	0xffff0200, // ps_2_0
604	0x05000051, 0xa00f0001, 0xbd84816f, 0xbf00ff97, 0xbf00ff97, 0x00000000, // def c1, -0.0647, -0.503899992, -0.503899992, 0
605	0x05000051, 0xa00f0002, 0xbe80ff97, 0xbd04816f, 0x00000000, 0x00000000, // def c2, -0.251949996, -0.03235, 0, 0
606	0x05000051, 0xa00f0003, 0x3dc816f0, 0x3f00ff97, 0x3e83fe5d, 0x00000000, // def c3, 0.0976999998, 0.503899992, 0.257800013, 0
607	0x05000051, 0xa00f0004, 0x3ee00000, 0xbe9404ea, 0xbe17f62b, 0x00000000, // def c4, 0.4375, -0.289099991, -0.148399994, 0
608	0x05000051, 0xa00f0005, 0xbd8ff972, 0xbebc01a3, 0x3ee00000, 0x00000000, // def c5, -0.0702999979, -0.367199987, 0.4375, 0
609	0x05000051, 0xa00f0006, 0x3f000000, 0x40000000, 0x3f800000, 0xbf00ff97, // def c6, 0.5, 2, 1, -0.503899992
610	0x05000051, 0xa00f0007, 0x3f000000, 0x3f000000, 0x3f800000, 0x00000000, // def c7, 0.5, 0.5, 1, 0
611	0x0200001f, 0x80000000, 0xb0030000, // dcl t0.xy
612	0x0200001f, 0x90000000, 0xa00f0800, // dcl_2d s0
613	0x03000005, 0x80030000, 0xb0e40000, 0xa0c90006, // mul r0.xy, t0, c6.yzxw
614	0x02000001, 0x80030001, 0xa0e40006, // mov r1.xy, c6
615	0x04000004, 0x80010002, 0xb0000000, 0x80550001, 0xa0aa0000, // mad r2.x, t0.x, r1.y, c0.z
616	0x02000001, 0x80020002, 0xb0550000, // mov r2.y, t0.y
617	0x03000042, 0x800f0000, 0x80e40000, 0xa0e40800, // texld r0, r0, s0
618	0x03000042, 0x800f0002, 0x80e40002, 0xa0e40800, // texld r2, r2, s0
619	0x03000005, 0x80080000, 0x80aa0000, 0xa0000003, // mul r0.w, r0.z, c3.x
620	0x04000004, 0x80080000, 0x80550000, 0xa0550003, 0x80ff0000, // mad r0.w, r0.y, c3.y, r0.w
621	0x04000004, 0x80010003, 0x80000000, 0xa0aa0003, 0x80ff0000, // mad r3.x, r0.x, c3.z, r0.w
622	0x03000005, 0x80080000, 0x80aa0000, 0xa0000004, // mul r0.w, r0.z, c4.x
623	0x04000004, 0x80080000, 0x80550000, 0xa0550004, 0x80ff0000, // mad r0.w, r0.y, c4.y, r0.w
624	0x04000004, 0x80020003, 0x80000000, 0xa0aa0004, 0x80ff0000, // mad r3.y, r0.x, c4.z, r0.w
625	0x03000005, 0x80080002, 0x80aa0000, 0xa0000005, // mul r2.w, r0.z, c5.x
626	0x04000004, 0x80080002, 0x80550000, 0xa0550005, 0x80ff0002, // mad r2.w, r0.y, c5.y, r2.w
627	0x04000004, 0x80040003, 0x80000000, 0xa0aa0005, 0x80ff0002, // mad r3.z, r0.x, c5.z, r2.w
628	0x03000002, 0x80070000, 0x80e40003, 0xa0e40001, // add r0.xyz, r3, c1
629	0x02000001, 0x80080000, 0xa0ff0006, // mov r0.w, c6.w
630	0x03000005, 0x80080002, 0x80aa0002, 0xa0000003, // mul r2.w, r2.z, c3.x
631	0x04000004, 0x80080002, 0x80550002, 0xa0550003, 0x80ff0002, // mad r2.w, r2.y, c3.y, r2.w
632	0x04000004, 0x80020003, 0x80000002, 0xa0aa0003, 0x80ff0002, // mad r3.y, r2.x, c3.z, r2.w
633	0x03000005, 0x80080002, 0x80aa0002, 0xa0000004, // mul r2.w, r2.z, c4.x
634	0x04000004, 0x80080002, 0x80550002, 0xa0550004, 0x80ff0002, // mad r2.w, r2.y, c4.y, r2.w
635	0x04000004, 0x80010003, 0x80000002, 0xa0aa0004, 0x80ff0002, // mad r3.x, r2.x, c4.z, r2.w
636	0x03000005, 0x80080003, 0x80aa0002, 0xa0000005, // mul r3.w, r2.z, c5.x
637	0x04000004, 0x80080003, 0x80550002, 0xa0550005, 0x80ff0003, // mad r3.w, r2.y, c5.y, r3.w
638	0x04000004, 0x80040003, 0x80000002, 0xa0aa0005, 0x80ff0003, // mad r3.z, r2.x, c5.z, r3.w
639	0x03000002, 0x80010002, 0x80550000, 0x80000003, // add r2.x, r0.y, r3.x
640	0x03000002, 0x80020002, 0x80000000, 0x80550003, // add r2.y, r0.x, r3.y
641	0x03000002, 0x80040002, 0x80ff0000, 0x80aa0003, // add r2.z, r0.w, r3.z
642	0x02000001, 0x80120000, 0x80aa0000, // mov_sat r0.y, r0.z
643	0x02000001, 0x80070003, 0xa0e40007, // mov r3.xyz, c7
644	0x04000004, 0x80110000, 0x80000002, 0x80000003, 0xa0000002, // mad_sat r0.x, r2.x, r3.x, c2.x
645	0x04000004, 0x80140000, 0x80550002, 0x80550003, 0xa0550002, // mad_sat r0.z, r2.y, r3.y, c2.y
646	0x04000004, 0x80180000, 0x80aa0002, 0x80aa0003, 0xa0aa0002, // mad_sat r0.w, r2.z, r3.z, c2.z
647	0x03000005, 0x80010001, 0x80000001, 0xa0000000, // mul r1.x, r1.x, c0.x
648	0x04000004, 0x80010001, 0xb0000000, 0xa0000000, 0x81000001, // mad r1.x, t0.x, c0.x, -r1.x
649	0x04000058, 0x800f0000, 0x80000001, 0xa0ff0003, 0x80e40000, // cmp r0, r1.x, c3.w, r0
650	0x02000001, 0x800f0800, 0x80e40000, // mov oC0, r0
651	0x0000ffff
652	};
653
654	switch (d3dfmtFrom)
655	{
656	/*
657	* Emulated to ARGB
658	*/
659	case D3DFMT_A8B8G8R8:
660	{
661	if (d3dfmtTo == D3DFMT_A8R8G8B8)
662	{
663	pResult->paVS = aVSCode;
664	pResult->paPS = aPSCodeSwapRB;
665
666	return D3D_OK;
667	}
668	} break;
669	case D3DFMT_UYVY:
670	{
671	if (d3dfmtTo == D3DFMT_A8R8G8B8)
672	{
673	pResult->paVS = aVSCode;
674	pResult->paPS = aPSCodeUYVYtoRGB;
675
676	return D3D_OK;
677	}
678	} break;
679	case D3DFMT_YUY2:
680	{
681	if (d3dfmtTo == D3DFMT_A8R8G8B8)
682	{
683	pResult->paVS = aVSCode;
684	pResult->paPS = aPSCodeYUY2toRGB;
685
686	return D3D_OK;
687	}
688	} break;
689
690	/*
691	* ARGB to emulated.
692	*/
693	case D3DFMT_A8R8G8B8:
694	{
695	if (d3dfmtTo == D3DFMT_A8B8G8R8)
696	{
697	pResult->paVS = aVSCode;
698	pResult->paPS = aPSCodeSwapRB;
699
700	return D3D_OK;
701	}
702	else if (d3dfmtTo == D3DFMT_UYVY)
703	{
704	pResult->paVS = aVSCode;
705	pResult->paPS = aPSCodeRGBtoUYVY;
706
707	return D3D_OK;
708	}
709	else if (d3dfmtTo == D3DFMT_YUY2)
710	{
711	pResult->paVS = aVSCode;
712	pResult->paPS = aPSCodeRGBtoYUY2;
713
714	return D3D_OK;
715	}
716	} break;
717
718	default:
719	break;
720	}
721
722	return E_NOTIMPL;
723	}
724
725	class D3D9Conversion
726	{
727	private:
728	IDirect3DDevice9Ex *mpDevice;
729
730	/* State objects. */
731	IDirect3DVertexBuffer9 *mpVB;
732	IDirect3DVertexDeclaration9 *mpVertexDecl;
733	IDirect3DVertexShader9 *mpVS;
734	IDirect3DPixelShader9 *mpPS;
735
736	D3D9ConversionParameters mParameters;
737
738	typedef struct D3DSamplerState
739	{
740	D3DSAMPLERSTATETYPE Type;
741	DWORD Value;
742	} D3DSamplerState;
743
744	/* Saved context. */
745	struct
746	{
747	DWORD dwCull;
748	DWORD dwZEnable;
749	IDirect3DSurface9 *pRT;
750	IDirect3DVertexShader9 *pVS;
751	IDirect3DPixelShader9 *pPS;
752	IDirect3DBaseTexture9 *pTexture;
753	float aVSConstantData[4];
754	float aPSConstantData[4];
755	D3DSamplerState aSamplerState[3];
756	} mSaved;
757
758	void destroyConversion();
759	HRESULT saveContextState();
760	HRESULT restoreContextState(PVMSVGA3DCONTEXT pContext);
761	HRESULT initConversion();
762	HRESULT setConversionState(IDirect3DTexture9 *pSourceTexture,
763	uint32_t cWidth,
764	uint32_t cHeight);
765
766	public:
767	enum Direction
768	{
769	FromEmulated,
770	ToEmulated,
771	};
772
773	D3D9Conversion(IDirect3DDevice9Ex *pDevice);
774	~D3D9Conversion();
775
776	HRESULT SelectConversion(D3DFORMAT d3dfmtFrom, D3DFORMAT d3dfmtTo);
777	HRESULT ConvertTexture(PVMSVGA3DCONTEXT pContext,
778	PVMSVGA3DSURFACE pSurface,
779	Direction enmDirection);
780	};
781
782	D3D9Conversion::D3D9Conversion(IDirect3DDevice9Ex *pDevice)
783	:
784	mpDevice(pDevice),
785	mpVB(0),
786	mpVertexDecl(0),
787	mpVS(0),
788	mpPS(0)
789	{
790	mParameters.paVS = 0;
791	mParameters.paPS = 0;
792	mSaved.dwCull = D3DCULL_NONE;
793	mSaved.dwZEnable = D3DZB_FALSE;
794	mSaved.pRT = 0;
795	mSaved.pVS = 0;
796	mSaved.pPS = 0;
797	mSaved.pTexture = 0;
798	RT_ZERO(mSaved.aVSConstantData);
799	RT_ZERO(mSaved.aPSConstantData);
800	mSaved.aSamplerState[0].Type = D3DSAMP_MAGFILTER;
801	mSaved.aSamplerState[0].Value = D3DTEXF_POINT;
802	mSaved.aSamplerState[1].Type = D3DSAMP_MINFILTER;
803	mSaved.aSamplerState[1].Value = D3DTEXF_POINT;
804	mSaved.aSamplerState[2].Type = D3DSAMP_MIPFILTER;
805	mSaved.aSamplerState[2].Value = D3DTEXF_NONE;
806	}
807
808	D3D9Conversion::~D3D9Conversion()
809	{
810	destroyConversion();
811	}
812
813	void D3D9Conversion::destroyConversion()
814	{
815	D3D_RELEASE(mpVertexDecl);
816	D3D_RELEASE(mpVB);
817	D3D_RELEASE(mpVS);
818	D3D_RELEASE(mpPS);
819	}
820
821	HRESULT D3D9Conversion::saveContextState()
822	{
823	/*
824	* This is probably faster than
825	* IDirect3DStateBlock9 *mpStateBlock;
826	* mpDevice->CreateStateBlock(D3DSBT_ALL, &mpStateBlock);
827	*/
828	HRESULT hr = mpDevice->GetRenderState(D3DRS_CULLMODE, &mSaved.dwCull);
829	AssertReturn(hr == D3D_OK, hr);
830
831	hr = mpDevice->GetRenderState(D3DRS_ZENABLE, &mSaved.dwZEnable);
832	AssertReturn(hr == D3D_OK, hr);
833
834	hr = mpDevice->GetRenderTarget(0, &mSaved.pRT);
835	AssertReturn(hr == D3D_OK, hr);
836
837	hr = mpDevice->GetVertexShader(&mSaved.pVS);
838	AssertReturn(hr == D3D_OK, hr);
839
840	hr = mpDevice->GetPixelShader(&mSaved.pPS);
841	AssertReturn(hr == D3D_OK, hr);
842
843	hr = mpDevice->GetTexture(0, &mSaved.pTexture);
844	AssertReturn(hr == D3D_OK, hr);
845
846	hr = mpDevice->GetVertexShaderConstantF(0, mSaved.aVSConstantData, 1);
847	AssertReturn(hr == D3D_OK, hr);
848
849	hr = mpDevice->GetPixelShaderConstantF(0, mSaved.aPSConstantData, 1);
850	AssertReturn(hr == D3D_OK, hr);
851
852	for (uint32_t i = 0; i < RT_ELEMENTS(mSaved.aSamplerState); ++i)
853	{
854	hr = mpDevice->GetSamplerState(0, mSaved.aSamplerState[i].Type, &mSaved.aSamplerState[i].Value);
855	AssertReturn(hr == D3D_OK, hr);
856	}
857
858	return hr;
859	}
860
861	HRESULT D3D9Conversion::restoreContextState(PVMSVGA3DCONTEXT pContext)
862	{
863	HRESULT hr = mpDevice->SetRenderState(D3DRS_CULLMODE, mSaved.dwCull);
864	Assert(hr == D3D_OK);
865
866	hr = mpDevice->SetRenderState(D3DRS_ZENABLE, mSaved.dwZEnable);
867	Assert(hr == D3D_OK);
868
869	hr = mpDevice->SetRenderTarget(0, mSaved.pRT);
870	D3D_RELEASE(mSaved.pRT); /* GetRenderTarget increases the internal reference count. */
871	Assert(hr == D3D_OK);
872
873	hr = mpDevice->SetVertexDeclaration(pContext->d3dState.pVertexDecl);
874	Assert(hr == D3D_OK);
875
876	hr = mpDevice->SetVertexShader(mSaved.pVS);
877	Assert(hr == D3D_OK);
878
879	hr = mpDevice->SetPixelShader(mSaved.pPS);
880	Assert(hr == D3D_OK);
881
882	hr = mpDevice->SetTexture(0, mSaved.pTexture);
883	D3D_RELEASE(mSaved.pTexture); /* GetTexture increases the internal reference count. */
884	Assert(hr == D3D_OK);
885
886	hr = mpDevice->SetVertexShaderConstantF(0, mSaved.aVSConstantData, 1);
887	Assert(hr == D3D_OK);
888
889	hr = mpDevice->SetPixelShaderConstantF(0, mSaved.aPSConstantData, 1);
890	Assert(hr == D3D_OK);
891
892	for (uint32_t i = 0; i < RT_ELEMENTS(mSaved.aSamplerState); ++i)
893	{
894	hr = mpDevice->SetSamplerState(0, mSaved.aSamplerState[i].Type, mSaved.aSamplerState[i].Value);
895	Assert(hr == D3D_OK);
896	}
897
898	return hr;
899	}
900
901	HRESULT D3D9Conversion::initConversion()
902	{
903	static D3DVERTEXELEMENT9 const aVertexElements[] =
904	{
905	{0, 0, D3DDECLTYPE_FLOAT2, D3DDECLMETHOD_DEFAULT, D3DDECLUSAGE_POSITION, 0},
906	{0, 8, D3DDECLTYPE_FLOAT2, D3DDECLMETHOD_DEFAULT, D3DDECLUSAGE_TEXCOORD, 0},
907	D3DDECL_END()
908	};
909
910	HRESULT hr = mpDevice->CreateVertexDeclaration(aVertexElements, &mpVertexDecl);
911	AssertReturn(hr == D3D_OK, hr);
912
913	hr = mpDevice->CreateVertexBuffer(6 * sizeof(Vertex),
914	0, /* D3DUSAGE_* */
915	0, /* FVF */
916	D3DPOOL_DEFAULT,
917	&mpVB,
918	0);
919	AssertReturn(hr == D3D_OK, hr);
920
921	hr = mpDevice->CreateVertexShader(mParameters.paVS, &mpVS);
922	AssertReturn(hr == D3D_OK, hr);
923
924	hr = mpDevice->CreatePixelShader(mParameters.paPS, &mpPS);
925	AssertReturn(hr == D3D_OK, hr);
926
927	return hr;
928	}
929
930	HRESULT D3D9Conversion::setConversionState(IDirect3DTexture9 *pSourceTexture,
931	uint32_t cWidth,
932	uint32_t cHeight)
933	{
934	/* Subtract 0.5 to line up the pixel centers with texels
935	* https://docs.microsoft.com/en-us/windows/win32/direct3d9/directly-mapping-texels-to-pixels
936	*/
937	float const xLeft = -0.5f;
938	float const xRight = (float)(cWidth - 1) - 0.5f;
939	float const yTop = -0.5f;
940	float const yBottom = (float)(cHeight - 1) - 0.5f;
941
942	Vertex const aVertices[] =
943	{
944	{ xLeft, yTop, 0.0f, 0.0f},
945	{ xRight, yTop, 1.0f, 0.0f},
946	{ xRight, yBottom, 1.0f, 1.0f},
947
948	{ xLeft, yTop, 0.0f, 0.0f},
949	{ xRight, yBottom, 1.0f, 1.0f},
950	{ xLeft, yBottom, 0.0f, 1.0f},
951	};
952
953	HRESULT hr = d3dCopyToVertexBuffer(mpVB, aVertices, sizeof(aVertices));
954	AssertReturn(hr == D3D_OK, hr);
955
956	/* No need to save the stream source, because vmsvga3dDrawPrimitives always sets it. */
957	hr = mpDevice->SetStreamSource(0, mpVB, 0, sizeof(Vertex));
958	AssertReturn(hr == D3D_OK, hr);
959
960	/* Stored in pContext->d3dState.pVertexDecl. */
961	hr = mpDevice->SetVertexDeclaration(mpVertexDecl);
962	AssertReturn(hr == D3D_OK, hr);
963
964	/* Saved by saveContextState. */
965	hr = mpDevice->SetVertexShader(mpVS);
966	AssertReturn(hr == D3D_OK, hr);
967
968	/* Saved by saveContextState. */
969	hr = mpDevice->SetPixelShader(mpPS);
970	AssertReturn(hr == D3D_OK, hr);
971
972	/* Saved by saveContextState. */
973	hr = mpDevice->SetRenderState(D3DRS_CULLMODE, D3DCULL_NONE);
974	AssertReturn(hr == D3D_OK, hr);
975
976	/* Saved by saveContextState. */
977	hr = mpDevice->SetRenderState(D3DRS_ZENABLE, D3DZB_FALSE);
978	AssertReturn(hr == D3D_OK, hr);
979
980	/* Saved by saveContextState. */
981	hr = mpDevice->SetTexture(0, pSourceTexture);
982	AssertReturn(hr == D3D_OK, hr);
983
984	float aTextureInfo[4];
985	aTextureInfo[0] = (float)cWidth;
986	aTextureInfo[1] = (float)cHeight;
987	aTextureInfo[2] = 1.0f / (float)cWidth; /* Pixel width in texture coords. */
988	aTextureInfo[3] = 1.0f / (float)cHeight; /* Pixel height in texture coords. */
989
990	/* Saved by saveContextState. */
991	hr = mpDevice->SetVertexShaderConstantF(0, aTextureInfo, 1);
992	AssertReturn(hr == D3D_OK, hr);
993
994	/* Saved by saveContextState. */
995	hr = mpDevice->SetPixelShaderConstantF(0, aTextureInfo, 1);
996	AssertReturn(hr == D3D_OK, hr);
997
998	/* Saved by saveContextState. */
999	hr = mpDevice->SetSamplerState(0, D3DSAMP_MAGFILTER, D3DTEXF_POINT);
1000	AssertReturn(hr == D3D_OK, hr);
1001	hr = mpDevice->SetSamplerState(0, D3DSAMP_MINFILTER, D3DTEXF_POINT);
1002	AssertReturn(hr == D3D_OK, hr);
1003	hr = mpDevice->SetSamplerState(0, D3DSAMP_MIPFILTER, D3DTEXF_NONE);
1004	AssertReturn(hr == D3D_OK, hr);
1005
1006	return hr;
1007	}
1008
1009
1010	HRESULT D3D9Conversion::SelectConversion(D3DFORMAT d3dfmtFrom, D3DFORMAT d3dfmtTo)
1011	{
1012	/** @todo d3d9SelectConversion should be a member. Move the code here? */
1013	HRESULT hr = d3d9SelectConversion(d3dfmtFrom, d3dfmtTo, &mParameters);
1014	return hr;
1015	}
1016
1017	HRESULT D3D9Conversion::ConvertTexture(PVMSVGA3DCONTEXT pContext,
1018	PVMSVGA3DSURFACE pSurface,
1019	Direction enmDirection)
1020	{
1021	IDirect3DTexture9 *pSourceTexture;
1022	IDirect3DTexture9 *pTargetTexture;
1023	D3D9TextureType enmTargetType;
1024
1025	if (enmDirection == FromEmulated)
1026	{
1027	pSourceTexture = pSurface->emulated.pTexture;
1028	pTargetTexture = pSurface->u.pTexture;
1029	enmTargetType = D3D9TextureType_Texture;
1030	}
1031	else if (enmDirection == ToEmulated)
1032	{
1033	pSourceTexture = pSurface->u.pTexture;
1034	pTargetTexture = pSurface->emulated.pTexture;
1035	enmTargetType = D3D9TextureType_Emulated;
1036	}
1037	else
1038	AssertFailedReturn(E_INVALIDARG);
1039
1040	AssertPtrReturn(pSourceTexture, E_INVALIDARG);
1041	AssertPtrReturn(pTargetTexture, E_INVALIDARG);
1042
1043	HRESULT hr = saveContextState();
1044	if (hr == D3D_OK)
1045	{
1046	hr = initConversion();
1047	if (hr == D3D_OK)
1048	{
1049	uint32_t const cWidth = pSurface->paMipmapLevels[0].mipmapSize.width;
1050	uint32_t const cHeight = pSurface->paMipmapLevels[0].mipmapSize.height;
1051
1052	hr = setConversionState(pSourceTexture, cWidth, cHeight);
1053	if (hr == D3D_OK)
1054	{
1055	hr = mpDevice->BeginScene();
1056	Assert(hr == D3D_OK);
1057	if (hr == D3D_OK)
1058	{
1059	for (DWORD iFace = 0; iFace < pSurface->cFaces; ++iFace)
1060	{
1061	DWORD const cMipLevels = pTargetTexture->GetLevelCount();
1062	for (DWORD iMipmap = 0; iMipmap < cMipLevels && hr == D3D_OK; ++iMipmap)
1063	{
1064	IDirect3DSurface9 *pRT = NULL;
1065	hr = D3D9GetTextureLevel(pSurface, enmTargetType, iFace, iMipmap, &pRT);
1066	Assert(hr == D3D_OK);
1067	if (hr == D3D_OK)
1068	{
1069	hr = mpDevice->SetRenderTarget(0, pRT);
1070	Assert(hr == D3D_OK);
1071	if (hr == D3D_OK)
1072	{
1073	hr = mpDevice->DrawPrimitive(D3DPT_TRIANGLELIST, 0, 2);
1074	Assert(hr == D3D_OK);
1075	}
1076
1077	D3D_RELEASE(pRT);
1078	}
1079	}
1080	}
1081
1082	hr = mpDevice->EndScene();
1083	Assert(hr == D3D_OK);
1084	}
1085	}
1086	}
1087
1088	hr = restoreContextState(pContext);
1089	}
1090
1091	destroyConversion();
1092
1093	return hr;
1094	}
1095
1096
1097	HRESULT D3D9UpdateTexture(PVMSVGA3DCONTEXT pContext,
1098	PVMSVGA3DSURFACE pSurface)
1099	{
1100	HRESULT hr = S_OK;
1101
1102	if (pSurface->formatD3D == pSurface->d3dfmtRequested)
1103	{
1104	hr = pContext->pDevice->UpdateTexture(pSurface->bounce.pTexture, pSurface->u.pTexture);
1105	}
1106	else
1107	{
1108	if (pSurface->enmD3DResType == VMSVGA3D_D3DRESTYPE_TEXTURE)
1109	{
1110	hr = pContext->pDevice->UpdateTexture(pSurface->bounce.pTexture, pSurface->emulated.pTexture);
1111	if (hr == D3D_OK)
1112	{
1113	D3D9Conversion conv(pContext->pDevice);
1114	hr = conv.SelectConversion(pSurface->d3dfmtRequested, pSurface->formatD3D);
1115	if (hr == D3D_OK)
1116	{
1117	hr = conv.ConvertTexture(pContext, pSurface, D3D9Conversion::FromEmulated);
1118	}
1119	}
1120	}
1121	else
1122	{
1123	/** @todo Cubemaps and maybe volume textures. */
1124	hr = pContext->pDevice->UpdateTexture(pSurface->bounce.pTexture, pSurface->u.pTexture);
1125	}
1126	}
1127
1128	return hr;
1129	}
1130
1131	HRESULT D3D9GetSurfaceLevel(PVMSVGA3DSURFACE pSurface,
1132	uint32_t uFace,
1133	uint32_t uMipmap,
1134	bool fBounce,
1135	IDirect3DSurface9 **ppD3DSurface)
1136	{
1137	HRESULT hr;
1138
1139	if ( pSurface->enmD3DResType == VMSVGA3D_D3DRESTYPE_CUBE_TEXTURE
1140	\|\| pSurface->enmD3DResType == VMSVGA3D_D3DRESTYPE_TEXTURE)
1141	{
1142	D3D9TextureType enmType;
1143	if (fBounce)
1144	enmType = D3D9TextureType_Bounce;
1145	else if (pSurface->formatD3D != pSurface->d3dfmtRequested)
1146	enmType = D3D9TextureType_Emulated;
1147	else
1148	enmType = D3D9TextureType_Texture;
1149
1150	hr = D3D9GetTextureLevel(pSurface, enmType, uFace, uMipmap, ppD3DSurface);
1151	}
1152	else if (pSurface->enmD3DResType == VMSVGA3D_D3DRESTYPE_SURFACE)
1153	{
1154	pSurface->u.pSurface->AddRef();
1155	*ppD3DSurface = pSurface->u.pSurface;
1156	hr = S_OK;
1157	}
1158	else
1159	{
1160	AssertMsgFailed(("No surface for type %d\n", pSurface->enmD3DResType));
1161	hr = E_INVALIDARG;
1162	}
1163
1164	return hr;
1165	}
1166
1167	/** Copy the texture content to the bounce texture.
1168	*/
1169	HRESULT D3D9GetRenderTargetData(PVMSVGA3DCONTEXT pContext,
1170	PVMSVGA3DSURFACE pSurface,
1171	uint32_t uFace,
1172	uint32_t uMipmap)
1173	{
1174	HRESULT hr;
1175
1176	/* Get the corresponding bounce texture surface. */
1177	IDirect3DSurface9 *pDst = NULL;
1178	hr = D3D9GetSurfaceLevel(pSurface, uFace, uMipmap, true, &pDst);
1179	AssertReturn(hr == D3D_OK, hr);
1180
1181	/* Get the actual texture surface, emulated or actual. */
1182	IDirect3DSurface9 *pSrc = NULL;
1183	hr = D3D9GetSurfaceLevel(pSurface, uFace, uMipmap, false, &pSrc);
1184	AssertReturnStmt(hr == D3D_OK,
1185	D3D_RELEASE(pDst), hr);
1186
1187	Assert(pDst != pSrc);
1188
1189	if (pSurface->formatD3D == pSurface->d3dfmtRequested)
1190	{
1191	hr = pContext->pDevice->GetRenderTargetData(pSrc, pDst);
1192	AssertMsg(hr == D3D_OK, ("GetRenderTargetData failed with %x\n", hr));
1193	}
1194	else
1195	{
1196	D3D9Conversion conv(pContext->pDevice);
1197	hr = conv.SelectConversion(pSurface->formatD3D, pSurface->d3dfmtRequested);
1198	if (hr == D3D_OK)
1199	{
1200	hr = conv.ConvertTexture(pContext, pSurface, D3D9Conversion::ToEmulated);
1201	}
1202
1203	if (hr == D3D_OK)
1204	{
1205	hr = pContext->pDevice->GetRenderTargetData(pSrc, pDst);
1206	AssertMsg(hr == D3D_OK, ("GetRenderTargetData failed with %x\n", hr));
1207	}
1208	}
1209
1210	D3D_RELEASE(pSrc);
1211	D3D_RELEASE(pDst);
1212	return hr;
1213	}
1214
1215	D3DFORMAT D3D9GetActualFormat(PVMSVGA3DSTATE pState, D3DFORMAT d3dfmtRequested)
1216	{
1217	RT_NOREF(pState);
1218
1219	switch (d3dfmtRequested)
1220	{
1221	case D3DFMT_UYVY:
1222	if (!pState->fSupportedFormatUYVY) return D3DFMT_A8R8G8B8;
1223	break;
1224	case D3DFMT_YUY2:
1225	if (!pState->fSupportedFormatYUY2) return D3DFMT_A8R8G8B8;
1226	break;
1227	case D3DFMT_A8B8G8R8:
1228	if (!pState->fSupportedFormatA8B8G8R8) return D3DFMT_A8R8G8B8;
1229	break;
1230	default:
1231	break;
1232	}
1233
1234	/* Use the requested format. No emulation required. */
1235	return d3dfmtRequested;
1236	}
1237
1238	bool D3D9CheckDeviceFormat(IDirect3D9 *pD3D9, DWORD Usage, D3DRESOURCETYPE RType, D3DFORMAT CheckFormat)
1239	{
1240	HRESULT hr = pD3D9->CheckDeviceFormat(D3DADAPTER_DEFAULT,
1241	D3DDEVTYPE_HAL,
1242	D3DFMT_X8R8G8B8, /* assume standard 32-bit display mode */
1243	Usage,
1244	RType,
1245	CheckFormat);
1246	return (hr == D3D_OK);
1247	}

Note: See TracBrowser for help on using the repository browser.

source: vbox/trunk/src/VBox/Devices/Graphics/DevVGA-SVGA3d-win-d3d9.cpp@ 105745

Download in other formats: