void TransformDrawEngine::SoftwareTransformAndDraw(
    int prim, u8 *decoded, LinkedShader *program, int vertexCount, u32 vertType, void *inds, int indexType, const DecVtxFormat &decVtxFormat, int maxIndex) {

    bool throughmode = (vertType & GE_VTYPE_THROUGH_MASK) != 0;
    bool lmode = gstate.isUsingSecondaryColor() && gstate.isLightingEnabled();

    // TODO: Split up into multiple draw calls for GLES 2.0 where you can't guarantee support for more than 0x10000 verts.

#if defined(MOBILE_DEVICE)
    if (vertexCount > 0x10000/3)
        vertexCount = 0x10000/3;
#endif

    float uscale = 1.0f;
    float vscale = 1.0f;
    bool scaleUV = false;
    if (throughmode) {
        uscale /= gstate_c.curTextureWidth;
        vscale /= gstate_c.curTextureHeight;
    } else {
        scaleUV = !g_Config.bPrescaleUV;
    }

    bool skinningEnabled = vertTypeIsSkinningEnabled(vertType);

    int w = gstate.getTextureWidth(0);
    int h = gstate.getTextureHeight(0);
    float widthFactor = (float) w / (float) gstate_c.curTextureWidth;
    float heightFactor = (float) h / (float) gstate_c.curTextureHeight;

    Lighter lighter(vertType);
    float fog_end = getFloat24(gstate.fog1);
    float fog_slope = getFloat24(gstate.fog2);

    VertexReader reader(decoded, decVtxFormat, vertType);
    for (int index = 0; index < maxIndex; index++) {
        reader.Goto(index);

        float v[3] = {0, 0, 0};
        float c0[4] = {1, 1, 1, 1};
        float c1[4] = {0, 0, 0, 0};
        float uv[3] = {0, 0, 1};
        float fogCoef = 1.0f;

        if (throughmode) {
            // Do not touch the coordinates or the colors. No lighting.
            reader.ReadPos(v);
            if (reader.hasColor0()) {
                reader.ReadColor0(c0);
                for (int j = 0; j < 4; j++) {
                    c1[j] = 0.0f;
                }
            } else {
                c0[0] = gstate.getMaterialAmbientR() / 255.f;
                c0[1] = gstate.getMaterialAmbientG() / 255.f;
                c0[2] = gstate.getMaterialAmbientB() / 255.f;
                c0[3] = gstate.getMaterialAmbientA() / 255.f;
            }

            if (reader.hasUV()) {
                reader.ReadUV(uv);

                uv[0] *= uscale;
                uv[1] *= vscale;
            }
            fogCoef = 1.0f;
            // Scale UV?
        } else {
            // We do software T&L for now
            float out[3], norm[3];
            float pos[3], nrm[3];
            Vec3f normal(0, 0, 1);
            reader.ReadPos(pos);
            if (reader.hasNormal())
                reader.ReadNrm(nrm);

            if (!skinningEnabled) {
                Vec3ByMatrix43(out, pos, gstate.worldMatrix);
                if (reader.hasNormal()) {
                    Norm3ByMatrix43(norm, nrm, gstate.worldMatrix);
                    normal = Vec3f(norm).Normalized();
                }
            } else {
                float weights[8];
                reader.ReadWeights(weights);
                // Skinning
                Vec3f psum(0,0,0);
                Vec3f nsum(0,0,0);
                for (int i = 0; i < vertTypeGetNumBoneWeights(vertType); i++) {
                    if (weights[i] != 0.0f) {
                        Vec3ByMatrix43(out, pos, gstate.boneMatrix+i*12);
                        Vec3f tpos(out);
                        psum += tpos * weights[i];
                        if (reader.hasNormal()) {
                            Norm3ByMatrix43(norm, nrm, gstate.boneMatrix+i*12);
                            Vec3f tnorm(norm);
                            nsum += tnorm * weights[i];
                        }
                    }
                }

                // Yes, we really must multiply by the world matrix too.
                Vec3ByMatrix43(out, psum.AsArray(), gstate.worldMatrix);
                if (reader.hasNormal()) {
                    Norm3ByMatrix43(norm, nsum.AsArray(), gstate.worldMatrix);
                    normal = Vec3f(norm).Normalized();
                }
            }

            // Perform lighting here if enabled. don't need to check through, it's checked above.
            float unlitColor[4] = {1, 1, 1, 1};
            if (reader.hasColor0()) {
                reader.ReadColor0(unlitColor);
            } else {
                unlitColor[0] = gstate.getMaterialAmbientR() / 255.f;
                unlitColor[1] = gstate.getMaterialAmbientG() / 255.f;
                unlitColor[2] = gstate.getMaterialAmbientB() / 255.f;
                unlitColor[3] = gstate.getMaterialAmbientA() / 255.f;
            }
            float litColor0[4];
            float litColor1[4];
            lighter.Light(litColor0, litColor1, unlitColor, out, normal);

            if (gstate.isLightingEnabled()) {
                // Don't ignore gstate.lmode - we should send two colors in that case
                for (int j = 0; j < 4; j++) {
                    c0[j] = litColor0[j];
                }
                if (lmode) {
                    // Separate colors
                    for (int j = 0; j < 4; j++) {
                        c1[j] = litColor1[j];
                    }
                } else {
                    // Summed color into c0
                    for (int j = 0; j < 4; j++) {
                        c0[j] = ((c0[j] + litColor1[j]) > 1.0f) ? 1.0f : (c0[j] + litColor1[j]);
                    }
                }
            } else {
                if (reader.hasColor0()) {
                    for (int j = 0; j < 4; j++) {
                        c0[j] = unlitColor[j];
                    }
                } else {
                    c0[0] = gstate.getMaterialAmbientR() / 255.f;
                    c0[1] = gstate.getMaterialAmbientG() / 255.f;
                    c0[2] = gstate.getMaterialAmbientB() / 255.f;
                    c0[3] = gstate.getMaterialAmbientA() / 255.f;
                }
                if (lmode) {
                    for (int j = 0; j < 4; j++) {
                        c1[j] = 0.0f;
                    }
                }
            }

            float ruv[2] = {0.0f, 0.0f};
            if (reader.hasUV())
                reader.ReadUV(ruv);

            // Perform texture coordinate generation after the transform and lighting - one style of UV depends on lights.
            switch (gstate.getUVGenMode()) {
            case GE_TEXMAP_TEXTURE_COORDS:	// UV mapping
            case GE_TEXMAP_UNKNOWN: // Seen in Riviera.  Unsure of meaning, but this works.
                // Texture scale/offset is only performed in this mode.
                if (scaleUV) {
                    uv[0] = ruv[0]*gstate_c.uv.uScale + gstate_c.uv.uOff;
                    uv[1] = ruv[1]*gstate_c.uv.vScale + gstate_c.uv.vOff;
                } else {
                    uv[0] = ruv[0];
                    uv[1] = ruv[1];
                }
                uv[2] = 1.0f;
                break;

            case GE_TEXMAP_TEXTURE_MATRIX:
            {
                // Projection mapping
                Vec3f source;
                switch (gstate.getUVProjMode())	{
                case GE_PROJMAP_POSITION: // Use model space XYZ as source
                    source = pos;
                    break;

                case GE_PROJMAP_UV: // Use unscaled UV as source
                    source = Vec3f(ruv[0], ruv[1], 0.0f);
                    break;

                case GE_PROJMAP_NORMALIZED_NORMAL: // Use normalized normal as source
                    if (reader.hasNormal()) {
                        source = Vec3f(norm).Normalized();
                    } else {
                        ERROR_LOG_REPORT(G3D, "Normal projection mapping without normal?");
                        source = Vec3f(0.0f, 0.0f, 1.0f);
                    }
                    break;

                case GE_PROJMAP_NORMAL: // Use non-normalized normal as source!
                    if (reader.hasNormal()) {
                        source = Vec3f(norm);
                    } else {
                        ERROR_LOG_REPORT(G3D, "Normal projection mapping without normal?");
                        source = Vec3f(0.0f, 0.0f, 1.0f);
                    }
                    break;
                }

                float uvw[3];
                Vec3ByMatrix43(uvw, &source.x, gstate.tgenMatrix);
                uv[0] = uvw[0];
                uv[1] = uvw[1];
                uv[2] = uvw[2];
            }
            break;

            case GE_TEXMAP_ENVIRONMENT_MAP:
                // Shade mapping - use two light sources to generate U and V.
            {
                Vec3f lightpos0 = Vec3f(gstate_c.lightpos[gstate.getUVLS0()]).Normalized();
                Vec3f lightpos1 = Vec3f(gstate_c.lightpos[gstate.getUVLS1()]).Normalized();

                uv[0] = (1.0f + Dot(lightpos0, normal))/2.0f;
                uv[1] = (1.0f - Dot(lightpos1, normal))/2.0f;
                uv[2] = 1.0f;
            }
            break;

            default:
                // Illegal
                ERROR_LOG_REPORT(G3D, "Impossible UV gen mode? %d", gstate.getUVGenMode());
                break;
            }

            uv[0] = uv[0] * widthFactor;
            uv[1] = uv[1] * heightFactor;

            // Transform the coord by the view matrix.
            Vec3ByMatrix43(v, out, gstate.viewMatrix);
            fogCoef = (v[2] + fog_end) * fog_slope;
        }

        // TODO: Write to a flexible buffer, we don't always need all four components.
        memcpy(&transformed[index].x, v, 3 * sizeof(float));
        transformed[index].fog = fogCoef;
        memcpy(&transformed[index].u, uv, 3 * sizeof(float));
        if (gstate_c.flipTexture) {
            transformed[index].v = 1.0f - transformed[index].v;
        }
        for (int i = 0; i < 4; i++) {
            transformed[index].color0[i] = c0[i] * 255.0f;
        }
        for (int i = 0; i < 3; i++) {
            transformed[index].color1[i] = c1[i] * 255.0f;
        }
    }

    // Here's the best opportunity to try to detect rectangles used to clear the screen, and
    // replace them with real OpenGL clears. This can provide a speedup on certain mobile chips.
    // Disabled for now - depth does not come out exactly the same.
    //
    // An alternative option is to simply ditch all the verts except the first and last to create a single
    // rectangle out of many. Quite a small optimization though.
    if (false && maxIndex > 1 && gstate.isModeClear() && prim == GE_PRIM_RECTANGLES && IsReallyAClear(maxIndex)) {
        u32 clearColor;
        memcpy(&clearColor, transformed[0].color0, 4);
        float clearDepth = transformed[0].z;
        const float col[4] = {
            ((clearColor & 0xFF)) / 255.0f,
            ((clearColor & 0xFF00) >> 8) / 255.0f,
            ((clearColor & 0xFF0000) >> 16) / 255.0f,
            ((clearColor & 0xFF000000) >> 24) / 255.0f,
        };

        bool colorMask = gstate.isClearModeColorMask();
        bool alphaMask = gstate.isClearModeAlphaMask();
        glstate.colorMask.set(colorMask, colorMask, colorMask, alphaMask);
        if (alphaMask) {
            glstate.stencilTest.set(true);
            // Clear stencil
            // TODO: extract the stencilValue properly, see below
            int stencilValue = 0;
            glstate.stencilFunc.set(GL_ALWAYS, stencilValue, 255);
        } else {
            // Don't touch stencil
            glstate.stencilTest.set(false);
        }
        glstate.scissorTest.set(false);
        bool depthMask = gstate.isClearModeDepthMask();

        int target = 0;
        if (colorMask || alphaMask) target |= GL_COLOR_BUFFER_BIT | GL_STENCIL_BUFFER_BIT;
        if (depthMask) target |= GL_DEPTH_BUFFER_BIT;

        glClearColor(col[0], col[1], col[2], col[3]);
#ifdef USING_GLES2
        glClearDepthf(clearDepth);
#else
        glClearDepth(clearDepth);
#endif
        glClearStencil(0);  // TODO - take from alpha?
        glClear(target);
        return;
    }
void SoftwareTransform(
	int prim, int vertexCount, u32 vertType, u16 *&inds, int indexType,
	const DecVtxFormat &decVtxFormat, int &maxIndex, TransformedVertex *&drawBuffer, int &numTrans, bool &drawIndexed, const SoftwareTransformParams *params, SoftwareTransformResult *result) {
	u8 *decoded = params->decoded;
	FramebufferManagerCommon *fbman = params->fbman;
	TextureCacheCommon *texCache = params->texCache;
	TransformedVertex *transformed = params->transformed;
	TransformedVertex *transformedExpanded = params->transformedExpanded;
	float ySign = 1.0f;
	bool throughmode = (vertType & GE_VTYPE_THROUGH_MASK) != 0;
	bool lmode = gstate.isUsingSecondaryColor() && gstate.isLightingEnabled();

	// TODO: Split up into multiple draw calls for GLES 2.0 where you can't guarantee support for more than 0x10000 verts.

#if defined(MOBILE_DEVICE)
	if (vertexCount > 0x10000/3)
		vertexCount = 0x10000/3;
#endif

	float uscale = 1.0f;
	float vscale = 1.0f;
	if (throughmode) {
		uscale /= gstate_c.curTextureWidth;
		vscale /= gstate_c.curTextureHeight;
	}

	bool skinningEnabled = vertTypeIsSkinningEnabled(vertType);

	const int w = gstate.getTextureWidth(0);
	const int h = gstate.getTextureHeight(0);
	float widthFactor = (float) w / (float) gstate_c.curTextureWidth;
	float heightFactor = (float) h / (float) gstate_c.curTextureHeight;

	Lighter lighter(vertType);
	float fog_end = getFloat24(gstate.fog1);
	float fog_slope = getFloat24(gstate.fog2);
	// Same fixup as in ShaderManager.cpp
	if (my_isinf(fog_slope)) {
		// not really sure what a sensible value might be.
		fog_slope = fog_slope < 0.0f ? -10000.0f : 10000.0f;
	}
	if (my_isnan(fog_slope)) {
		// Workaround for https://github.com/hrydgard/ppsspp/issues/5384#issuecomment-38365988
		// Just put the fog far away at a large finite distance.
		// Infinities and NaNs are rather unpredictable in shaders on many GPUs
		// so it's best to just make it a sane calculation.
		fog_end = 100000.0f;
		fog_slope = 1.0f;
	}

	VertexReader reader(decoded, decVtxFormat, vertType);
	if (throughmode) {
		for (int index = 0; index < maxIndex; index++) {
			// Do not touch the coordinates or the colors. No lighting.
			reader.Goto(index);
			// TODO: Write to a flexible buffer, we don't always need all four components.
			TransformedVertex &vert = transformed[index];
			reader.ReadPos(vert.pos);

			if (reader.hasColor0()) {
				reader.ReadColor0_8888(vert.color0);
			} else {
				vert.color0_32 = gstate.getMaterialAmbientRGBA();
			}

			if (reader.hasUV()) {
				reader.ReadUV(vert.uv);

				vert.u *= uscale;
				vert.v *= vscale;
			} else {
				vert.u = 0.0f;
				vert.v = 0.0f;
			}

			// Ignore color1 and fog, never used in throughmode anyway.
			// The w of uv is also never used (hardcoded to 1.0.)
		}
	} else {
		// Okay, need to actually perform the full transform.
		for (int index = 0; index < maxIndex; index++) {
			reader.Goto(index);

			float v[3] = {0, 0, 0};
			Vec4f c0 = Vec4f(1, 1, 1, 1);
			Vec4f c1 = Vec4f(0, 0, 0, 0);
			float uv[3] = {0, 0, 1};
			float fogCoef = 1.0f;

			// We do software T&L for now
			float out[3];
			float pos[3];
			Vec3f normal(0, 0, 1);
			Vec3f worldnormal(0, 0, 1);
			reader.ReadPos(pos);

			if (!skinningEnabled) {
				Vec3ByMatrix43(out, pos, gstate.worldMatrix);
				if (reader.hasNormal()) {
					reader.ReadNrm(normal.AsArray());
					if (gstate.areNormalsReversed()) {
						normal = -normal;
					}
					Norm3ByMatrix43(worldnormal.AsArray(), normal.AsArray(), gstate.worldMatrix);
					worldnormal = worldnormal.Normalized();
				}
			} else {
				float weights[8];
				reader.ReadWeights(weights);
				if (reader.hasNormal())
					reader.ReadNrm(normal.AsArray());

				// Skinning
				Vec3f psum(0, 0, 0);
				Vec3f nsum(0, 0, 0);
				for (int i = 0; i < vertTypeGetNumBoneWeights(vertType); i++) {
					if (weights[i] != 0.0f) {
						Vec3ByMatrix43(out, pos, gstate.boneMatrix+i*12);
						Vec3f tpos(out);
						psum += tpos * weights[i];
						if (reader.hasNormal()) {
							Vec3f norm;
							Norm3ByMatrix43(norm.AsArray(), normal.AsArray(), gstate.boneMatrix+i*12);
							nsum += norm * weights[i];
						}
					}
				}

				// Yes, we really must multiply by the world matrix too.
				Vec3ByMatrix43(out, psum.AsArray(), gstate.worldMatrix);
				if (reader.hasNormal()) {
					normal = nsum;
					if (gstate.areNormalsReversed()) {
						normal = -normal;
					}
					Norm3ByMatrix43(worldnormal.AsArray(), normal.AsArray(), gstate.worldMatrix);
					worldnormal = worldnormal.Normalized();
				}
			}

			// Perform lighting here if enabled. don't need to check through, it's checked above.
			Vec4f unlitColor = Vec4f(1, 1, 1, 1);
			if (reader.hasColor0()) {
				reader.ReadColor0(&unlitColor.x);
			} else {
				unlitColor = Vec4f::FromRGBA(gstate.getMaterialAmbientRGBA());
			}

			if (gstate.isLightingEnabled()) {
				float litColor0[4];
				float litColor1[4];
				lighter.Light(litColor0, litColor1, unlitColor.AsArray(), out, worldnormal);

				// Don't ignore gstate.lmode - we should send two colors in that case
				for (int j = 0; j < 4; j++) {
					c0[j] = litColor0[j];
				}
				if (lmode) {
					// Separate colors
					for (int j = 0; j < 4; j++) {
						c1[j] = litColor1[j];
					}
				} else {
					// Summed color into c0 (will clamp in ToRGBA().)
					for (int j = 0; j < 4; j++) {
						c0[j] += litColor1[j];
					}
				}
			} else {
				if (reader.hasColor0()) {
					for (int j = 0; j < 4; j++) {
						c0[j] = unlitColor[j];
					}
				} else {
					c0 = Vec4f::FromRGBA(gstate.getMaterialAmbientRGBA());
				}
				if (lmode) {
					// c1 is already 0.
				}
			}

			float ruv[2] = {0.0f, 0.0f};
			if (reader.hasUV())
				reader.ReadUV(ruv);

			// Perform texture coordinate generation after the transform and lighting - one style of UV depends on lights.
			switch (gstate.getUVGenMode()) {
			case GE_TEXMAP_TEXTURE_COORDS:	// UV mapping
			case GE_TEXMAP_UNKNOWN: // Seen in Riviera.  Unsure of meaning, but this works.
				// We always prescale in the vertex decoder now.
				uv[0] = ruv[0];
				uv[1] = ruv[1];
				uv[2] = 1.0f;
				break;

			case GE_TEXMAP_TEXTURE_MATRIX:
				{
					// Projection mapping
					Vec3f source;
					switch (gstate.getUVProjMode())	{
					case GE_PROJMAP_POSITION: // Use model space XYZ as source
						source = pos;
						break;

					case GE_PROJMAP_UV: // Use unscaled UV as source
						source = Vec3f(ruv[0], ruv[1], 0.0f);
						break;

					case GE_PROJMAP_NORMALIZED_NORMAL: // Use normalized normal as source
						source = normal.Normalized();
						if (!reader.hasNormal()) {
							ERROR_LOG_REPORT(G3D, "Normal projection mapping without normal?");
						}
						break;

					case GE_PROJMAP_NORMAL: // Use non-normalized normal as source!
						source = normal;
						if (!reader.hasNormal()) {
							ERROR_LOG_REPORT(G3D, "Normal projection mapping without normal?");
						}
						break;
					}

					float uvw[3];
					Vec3ByMatrix43(uvw, &source.x, gstate.tgenMatrix);
					uv[0] = uvw[0];
					uv[1] = uvw[1];
					uv[2] = uvw[2];
				}
				break;

			case GE_TEXMAP_ENVIRONMENT_MAP:
				// Shade mapping - use two light sources to generate U and V.
				{
					Vec3f lightpos0 = Vec3f(&lighter.lpos[gstate.getUVLS0() * 3]).Normalized();
					Vec3f lightpos1 = Vec3f(&lighter.lpos[gstate.getUVLS1() * 3]).Normalized();

					uv[0] = (1.0f + Dot(lightpos0, worldnormal))/2.0f;
					uv[1] = (1.0f + Dot(lightpos1, worldnormal))/2.0f;
					uv[2] = 1.0f;
				}
				break;

			default:
				// Illegal
				ERROR_LOG_REPORT(G3D, "Impossible UV gen mode? %d", gstate.getUVGenMode());
				break;
			}

			uv[0] = uv[0] * widthFactor;
			uv[1] = uv[1] * heightFactor;

			// Transform the coord by the view matrix.
			Vec3ByMatrix43(v, out, gstate.viewMatrix);
			fogCoef = (v[2] + fog_end) * fog_slope;

			// TODO: Write to a flexible buffer, we don't always need all four components.
			memcpy(&transformed[index].x, v, 3 * sizeof(float));
			transformed[index].fog = fogCoef;
			memcpy(&transformed[index].u, uv, 3 * sizeof(float));
			transformed[index].color0_32 = c0.ToRGBA();
			transformed[index].color1_32 = c1.ToRGBA();

			// The multiplication by the projection matrix is still performed in the vertex shader.
			// So is vertex depth rounding, to simulate the 16-bit depth buffer.
		}
	}

	// Here's the best opportunity to try to detect rectangles used to clear the screen, and
	// replace them with real clears. This can provide a speedup on certain mobile chips.
	//
	// An alternative option is to simply ditch all the verts except the first and last to create a single
	// rectangle out of many. Quite a small optimization though.
	// Experiment: Disable on PowerVR (see issue #6290)
	// TODO: This bleeds outside the play area in non-buffered mode. Big deal? Probably not.
	bool reallyAClear = false;
	if (maxIndex > 1 && prim == GE_PRIM_RECTANGLES && gstate.isModeClear()) {
		int scissorX2 = gstate.getScissorX2() + 1;
		int scissorY2 = gstate.getScissorY2() + 1;
		reallyAClear = IsReallyAClear(transformed, maxIndex, scissorX2, scissorY2);
	}
	if (reallyAClear && gl_extensions.gpuVendor != GPU_VENDOR_POWERVR) {  // && g_Config.iRenderingMode != FB_NON_BUFFERED_MODE) {
		// If alpha is not allowed to be separate, it must match for both depth/stencil and color.  Vulkan requires this.
		bool alphaMatchesColor = gstate.isClearModeColorMask() == gstate.isClearModeAlphaMask();
		bool depthMatchesStencil = gstate.isClearModeAlphaMask() == gstate.isClearModeDepthMask();
		if (params->allowSeparateAlphaClear || (alphaMatchesColor && depthMatchesStencil)) {
			result->color = transformed[1].color0_32;
			// Need to rescale from a [0, 1] float.  This is the final transformed value.
			result->depth = ToScaledDepth((s16)(int)(transformed[1].z * 65535.0f));
			result->action = SW_CLEAR;
			return;
		}
	}

	// This means we're using a framebuffer (and one that isn't big enough.)
	if (gstate_c.curTextureHeight < (u32)h && maxIndex >= 2) {
		// Even if not rectangles, this will detect if either of the first two are outside the framebuffer.
		// HACK: Adding one pixel margin to this detection fixes issues in Assassin's Creed : Bloodlines,
		// while still keeping BOF working (see below).
		const float invTexH = 1.0f / gstate_c.curTextureHeight; // size of one texel.
		bool tlOutside;
		bool tlAlmostOutside;
		bool brOutside;
		// If we're outside heightFactor, then v must be wrapping or clamping.  Avoid this workaround.
		// If we're <= 1.0f, we're inside the framebuffer (workaround not needed.)
		// We buffer that 1.0f a little more with a texel to avoid some false positives.
		tlOutside = transformed[0].v <= heightFactor && transformed[0].v > 1.0f + invTexH;
		brOutside = transformed[1].v <= heightFactor && transformed[1].v > 1.0f + invTexH;
		// Careful: if br is outside, but tl is well inside, this workaround still doesn't make sense.
		// We go with halfway, since we overestimate framebuffer heights sometimes but not by much.
		tlAlmostOutside = transformed[0].v <= heightFactor && transformed[0].v >= 0.5f;
		if (tlOutside || (brOutside && tlAlmostOutside)) {
			// Okay, so we're texturing from outside the framebuffer, but inside the texture height.
			// Breath of Fire 3 does this to access a render surface at an offset.
			const u32 bpp = fbman->GetTargetFormat() == GE_FORMAT_8888 ? 4 : 2;
			const u32 prevH = texCache->AttachedDrawingHeight();
			const u32 fb_size = bpp * fbman->GetTargetStride() * prevH;
			const u32 prevYOffset = gstate_c.curTextureYOffset;
			if (texCache->SetOffsetTexture(fb_size)) {
				const float oldWidthFactor = widthFactor;
				const float oldHeightFactor = heightFactor;
				widthFactor = (float) w / (float) gstate_c.curTextureWidth;
				heightFactor = (float) h / (float) gstate_c.curTextureHeight;

				// We've already baked in the old gstate_c.curTextureYOffset, so correct.
				const float yDiff = (float) (prevH + prevYOffset - gstate_c.curTextureYOffset) / (float) h;
				for (int index = 0; index < maxIndex; ++index) {
					transformed[index].u *= widthFactor / oldWidthFactor;
					// Inverse it back to scale to the new FBO, and add 1.0f to account for old FBO.
					transformed[index].v = (transformed[index].v / oldHeightFactor - yDiff) * heightFactor;
				}
			}
		}
	}

	// Step 2: expand rectangles.
	drawBuffer = transformed;
	numTrans = 0;
	drawIndexed = false;

	if (prim != GE_PRIM_RECTANGLES) {
		// We can simply draw the unexpanded buffer.
		numTrans = vertexCount;
		drawIndexed = true;
	} else {
		bool useBufferedRendering = g_Config.iRenderingMode != FB_NON_BUFFERED_MODE;
		if (useBufferedRendering)
			ySign = -ySign;

		float flippedMatrix[16];
		if (!throughmode) {
			memcpy(&flippedMatrix, gstate.projMatrix, 16 * sizeof(float));

			const bool invertedY = useBufferedRendering ? (gstate_c.vpHeight < 0) : (gstate_c.vpHeight > 0);
			if (invertedY) {
				flippedMatrix[1] = -flippedMatrix[1];
				flippedMatrix[5] = -flippedMatrix[5];
				flippedMatrix[9] = -flippedMatrix[9];
				flippedMatrix[13] = -flippedMatrix[13];
			}
			const bool invertedX = gstate_c.vpWidth < 0;
			if (invertedX) {
				flippedMatrix[0] = -flippedMatrix[0];
				flippedMatrix[4] = -flippedMatrix[4];
				flippedMatrix[8] = -flippedMatrix[8];
				flippedMatrix[12] = -flippedMatrix[12];
			}
		}

		//rectangles always need 2 vertices, disregard the last one if there's an odd number
		vertexCount = vertexCount & ~1;
		numTrans = 0;
		drawBuffer = transformedExpanded;
		TransformedVertex *trans = &transformedExpanded[0];
		const u16 *indsIn = (const u16 *)inds;
		u16 *newInds = inds + vertexCount;
		u16 *indsOut = newInds;
		maxIndex = 4 * vertexCount;
		for (int i = 0; i < vertexCount; i += 2) {
			const TransformedVertex &transVtxTL = transformed[indsIn[i + 0]];
			const TransformedVertex &transVtxBR = transformed[indsIn[i + 1]];

			// We have to turn the rectangle into two triangles, so 6 points.
			// This is 4 verts + 6 indices.

			// bottom right
			trans[0] = transVtxBR;

			// top right
			trans[1] = transVtxBR;
			trans[1].y = transVtxTL.y;
			trans[1].v = transVtxTL.v;

			// top left
			trans[2] = transVtxBR;
			trans[2].x = transVtxTL.x;
			trans[2].y = transVtxTL.y;
			trans[2].u = transVtxTL.u;
			trans[2].v = transVtxTL.v;

			// bottom left
			trans[3] = transVtxBR;
			trans[3].x = transVtxTL.x;
			trans[3].u = transVtxTL.u;

			// That's the four corners. Now process UV rotation.
			if (throughmode)
				RotateUVThrough(trans);
			else
				RotateUV(trans, flippedMatrix, ySign);

			// Triangle: BR-TR-TL
			indsOut[0] = i * 2 + 0;
			indsOut[1] = i * 2 + 1;
			indsOut[2] = i * 2 + 2;
			// Triangle: BL-BR-TL
			indsOut[3] = i * 2 + 3;
			indsOut[4] = i * 2 + 0;
			indsOut[5] = i * 2 + 2;
			trans += 4;
			indsOut += 6;

			numTrans += 6;
		}
		inds = newInds;
		drawIndexed = true;

		// We don't know the color until here, so we have to do it now, instead of in StateMapping.
		// Might want to reconsider the order of things later...
		if (gstate.isModeClear() && gstate.isClearModeAlphaMask()) {
			result->setStencil = true;
			if (vertexCount > 1) {
				// Take the bottom right alpha value of the first rect as the stencil value.
				// Technically, each rect could individually fill its stencil, but most of the
				// time they use the same one.
				result->stencilValue = transformed[indsIn[1]].color0[3];
			} else {
				result->stencilValue = 0;
			}
		}
	}

	result->action = SW_DRAW_PRIMITIVES;
}
void TransformAndDrawPrim(void *verts, void *inds, int prim, int vertexCount, LinkedShader *program, float *customUV, int forceIndexType)
{
	// First, decode the verts and apply morphing
	VertexDecoder dec;
	dec.SetVertexType(gstate.vertType);
	dec.DecodeVerts(decoded, verts, inds, prim, vertexCount);

	bool useTexCoord = false;

	// Check if anything needs updating
	if (gstate.textureChanged)
	{
		if (gstate.textureMapEnable && !(gstate.clearmode & 1))
		{
			PSPSetTexture();
			useTexCoord = true;
		}
	}

	// Then, transform and draw in one big swoop (urgh!)
	// need to move this to the shader.
	
	// We're gonna have to keep software transforming RECTANGLES, unless we use a geom shader which we can't on OpenGL ES 2.0.
	// Usually, though, these primitives don't use lighting etc so it's no biggie performance wise, but it would be nice to get rid of
	// this code.

	// Actually, if we find the camera-relative right and down vectors, it might even be possible to add the extra points in pre-transformed
	// space and thus make decent use of hardware transform.

	// Actually again, single quads could be drawn more efficiently using GL_TRIANGLE_STRIP, no need to duplicate verts as for
	// GL_TRIANGLES. Still need to sw transform to compute the extra two corners though.
	
	// Temporary storage for RECTANGLES emulation
	float v2[3] = {0};
	float uv2[2] = {0};

	int numTrans = 0;
	TransformedVertex *trans = &transformed[0];

	// TODO: Could use glDrawElements in some cases, see below.


	// TODO: Split up into multiple draw calls for Android where you can't guarantee support for more than 0x10000 verts.
	int i = 0;

#ifdef ANDROID
	if (vertexCount > 0x10000/3)
		vertexCount = 0x10000/3;
#endif

	for (int i = 0; i < vertexCount; i++)
	{	
		int indexType = (gstate.vertType & GE_VTYPE_IDX_MASK);
		if (forceIndexType != -1) {
			indexType = forceIndexType;
		}

		int index;
		if (indexType == GE_VTYPE_IDX_8BIT)
		{
			index = ((u8*)inds)[i];
		} 
		else if (indexType == GE_VTYPE_IDX_16BIT)
		{
			index = ((u16*)inds)[i];
		}
		else
		{
			index = i;
		}

		float v[3] = {0,0,0};
		float c[4] = {1,1,1,1};
		float uv[2] = {0,0};

		if (gstate.vertType & GE_VTYPE_THROUGH_MASK)
		{
			// Do not touch the coordinates or the colors. No lighting.
			for (int j=0; j<3; j++)
				v[j] = decoded[index].pos[j];
			// TODO : check if has color
			for (int j=0; j<4; j++)
				c[j] = decoded[index].color[j];
			// TODO : check if has uv
			for (int j=0; j<2; j++)
				uv[j] = decoded[index].uv[j];

			//Rescale UV?
		}
		else
		{
			//We do software T&L for now
			float out[3], norm[3];
			if ((gstate.vertType & GE_VTYPE_WEIGHT_MASK) == GE_VTYPE_WEIGHT_NONE)
			{
				Vec3ByMatrix43(out, decoded[index].pos, gstate.worldMatrix);
				Norm3ByMatrix43(norm, decoded[index].normal, gstate.worldMatrix);
			}
			else
			{
				Vec3 psum(0,0,0);
				Vec3 nsum(0,0,0);
				int nweights = (gstate.vertType & GE_VTYPE_WEIGHT_MASK) >> GE_VTYPE_WEIGHT_SHIFT;
				for (int i = 0; i < nweights; i++)
				{
					Vec3ByMatrix43(out, decoded[index].pos, gstate.boneMatrix+i*12);
					Norm3ByMatrix43(norm, decoded[index].normal, gstate.boneMatrix+i*12);
					Vec3 tpos(out), tnorm(norm);
					psum += tpos*decoded[index].weights[i];
					nsum += tnorm*decoded[index].weights[i];
				}
				nsum.Normalize();
				psum.Write(out);
				nsum.Write(norm);
			}

			// Perform lighting here if enabled. don't need to check through, it's checked above.
			float dots[4] = {0,0,0,0};
			if (program->a_color0 != -1)
			{
				//c[1] = norm[1];
				float litColor[4] = {0,0,0,0};
				Light(litColor, decoded[index].color, out, norm, dots);
				if (gstate.lightingEnable & 1)
				{
					memcpy(c, litColor, sizeof(litColor));
				}
				else
				{
					// no lighting? copy the color.
					for (int j=0; j<4; j++)
						c[j] = decoded[index].color[j];
				}
			}
			else
			{
				// no color in the fragment program???
				for (int j=0; j<4; j++)
					c[j] = decoded[index].color[j];
			}

			if (customUV) {
				uv[0] = customUV[index * 2 + 0]*gstate.uScale + gstate.uOff;
				uv[1] = customUV[index * 2 + 1]*gstate.vScale + gstate.vOff;
			} else {
				// Perform texture coordinate generation after the transform and lighting - one style of UV depends on lights.
				switch (gstate.texmapmode & 0x3)
				{
				case 0:	// UV mapping
					// Texture scale/offset is only performed in this mode.
					uv[0] = decoded[index].uv[0]*gstate.uScale + gstate.uOff;
					uv[1] = decoded[index].uv[1]*gstate.vScale + gstate.vOff;
					break;
				case 1:
					{
						// Projection mapping
						Vec3 source;
						switch ((gstate.texmapmode >> 8) & 0x3)
						{
						case 0: // Use model space XYZ as source
							source = decoded[index].pos;
							break;
						case 1: // Use unscaled UV as source
							source = Vec3(decoded[index].uv[0], decoded[index].uv[1], 0.0f);
							break;
						case 2: // Use normalized normal as source
							source = Vec3(norm).Normalized();
							break;
						case 3: // Use non-normalized normal as source!
							source = Vec3(norm);
							break;
						}
						float uvw[3];
						Vec3ByMatrix43(uvw, &source.x, gstate.tgenMatrix);
						uv[0] = uvw[0];
						uv[1] = uvw[1];
					}
					break;
				case 2:
					// Shade mapping
					{
						int lightsource1 = gstate.texshade & 0x3;
						int lightsource2 = (gstate.texshade >> 8) & 0x3;
						uv[0] = dots[lightsource1];
						uv[1] = dots[lightsource2];
					}
					break;
				case 3:
					// Illegal
					break;
				}
			}
			// Transform the coord by the view matrix. Should this be done before or after texcoord generation?
			Vec3ByMatrix43(v, out, gstate.viewMatrix);
		}


		// We need to tesselate axis-aligned rectangles, as they're only specified by two coordinates.
		if (prim == GE_PRIM_RECTANGLES)
		{
			if ((i & 1) == 0)
			{
				// Save this vertex so we can generate when we get the next one. Color is taken from the last vertex.
				memcpy(v2, v, sizeof(float)*3);
				memcpy(uv2,uv,sizeof(float)*2);
			}
			else
			{
				// We have to turn the rectangle into two triangles, so 6 points. Sigh.

				// top left
				trans->x = v[0]; trans->y = v[1];
				trans->z = v[2]; 
				trans->uv[0] = uv[0]; trans->uv[1] = uv[1];
				memcpy(trans->color, c, 4*sizeof(float));
				trans++;

				// top right
				trans->x = v2[0]; trans->y = v[1];
				trans->z = v[2]; 
				trans->uv[0] = uv2[0]; trans->uv[1] = uv[1];
				memcpy(trans->color, c, 4*sizeof(float));
				trans++;

				// bottom right
				trans->x = v2[0]; trans->y = v2[1];
				trans->z = v[2]; 
				trans->uv[0] = uv2[0]; trans->uv[1] = uv2[1];
				memcpy(trans->color, c, 4*sizeof(float));
				trans++;

				// bottom left
				trans->x = v[0]; trans->y = v2[1];
				trans->z = v[2]; 
				trans->uv[0] = uv[0]; trans->uv[1] = uv2[1];
				memcpy(trans->color, c, 4*sizeof(float));
				trans++;

				// top left
				trans->x = v[0]; trans->y = v[1];
				trans->z = v[2]; 
				trans->uv[0] = uv[0]; trans->uv[1] = uv[1];
				memcpy(trans->color, c, 4*sizeof(float));
				trans++;

				// bottom right
				trans->x = v2[0]; trans->y = v2[1];
				trans->z = v[2]; 
				trans->uv[0] = uv2[0]; trans->uv[1] = uv2[1];
				memcpy(trans->color, c, 4*sizeof(float));
				trans++;

				numTrans += 6;
			}
		}
		else
		{
			memcpy(&trans->x, v, 3*sizeof(float));
			memcpy(trans->color, c, 4*sizeof(float));
			memcpy(trans->uv, uv, 2*sizeof(float));
			trans++;
			numTrans++;
		}
	}

	glEnableVertexAttribArray(program->a_position);
	if (useTexCoord && program->a_texcoord != -1) glEnableVertexAttribArray(program->a_texcoord);
	if (program->a_color0 != -1) glEnableVertexAttribArray(program->a_color0);
	const int vertexSize = sizeof(*trans);
	glVertexAttribPointer(program->a_position, 3, GL_FLOAT, GL_FALSE, vertexSize, transformed);
	if (useTexCoord && program->a_texcoord != -1) glVertexAttribPointer(program->a_texcoord, 2, GL_FLOAT, GL_FALSE, vertexSize, ((uint8_t*)transformed) + 3 * 4);	
	if (program->a_color0 != -1) glVertexAttribPointer(program->a_color0, 4, GL_FLOAT, GL_FALSE, vertexSize, ((uint8_t*)transformed) + 5 * 4);
	// NOTICE_LOG(G3D,"DrawPrimitive: %i", numTrans);
	glDrawArrays(glprim[prim], 0, numTrans);
	glDisableVertexAttribArray(program->a_position);
	if (useTexCoord && program->a_texcoord != -1) glDisableVertexAttribArray(program->a_texcoord);
	if (program->a_color0 != -1) glDisableVertexAttribArray(program->a_color0);

	/*
	if (((gstate.vertType ) & GE_VTYPE_IDX_MASK) == GE_VTYPE_IDX_8BIT)
	{
		glDrawElements(glprim, vertexCount, GL_UNSIGNED_BYTE, inds);
	} 
	else if (((gstate.vertType ) & GE_VTYPE_IDX_MASK) == GE_VTYPE_IDX_16BIT)
	{
		glDrawElements(glprim, vertexCount, GL_UNSIGNED_SHORT, inds);
	}
	else
	{*/

}
Exemple #4
0
// This normalizes a set of vertices in any format to SimpleVertex format, by processing away morphing AND skinning.
// The rest of the transform pipeline like lighting will go as normal, either hardware or software.
// The implementation is initially a bit inefficient but shouldn't be a big deal.
// An intermediate buffer of not-easy-to-predict size is stored at bufPtr.
u32 TransformDrawEngine::NormalizeVertices(u8 *outPtr, u8 *bufPtr, const u8 *inPtr, VertexDecoder *dec, int lowerBound, int upperBound, u32 vertType) {
	// First, decode the vertices into a GPU compatible format. This step can be eliminated but will need a separate
	// implementation of the vertex decoder.
	dec->DecodeVerts(bufPtr, inPtr, lowerBound, upperBound);

	// OK, morphing eliminated but bones still remain to be taken care of.
	// Let's do a partial software transform where we only do skinning.

	VertexReader reader(bufPtr, dec->GetDecVtxFmt(), vertType);

	SimpleVertex *sverts = (SimpleVertex *)outPtr;	

	const u8 defaultColor[4] = {
		(u8)gstate.getMaterialAmbientR(),
		(u8)gstate.getMaterialAmbientG(),
		(u8)gstate.getMaterialAmbientB(),
		(u8)gstate.getMaterialAmbientA(),
	};

	// Let's have two separate loops, one for non skinning and one for skinning.
	if (!g_Config.bSoftwareSkinning && (vertType & GE_VTYPE_WEIGHT_MASK) != GE_VTYPE_WEIGHT_NONE) {
		int numBoneWeights = vertTypeGetNumBoneWeights(vertType);
		for (int i = lowerBound; i <= upperBound; i++) {
			reader.Goto(i);
			SimpleVertex &sv = sverts[i];
			if (vertType & GE_VTYPE_TC_MASK) {
				reader.ReadUV(sv.uv);
			}

			if (vertType & GE_VTYPE_COL_MASK) {
				reader.ReadColor0_8888(sv.color);
			} else {
				memcpy(sv.color, defaultColor, 4);
			}

			float nrm[3], pos[3];
			float bnrm[3], bpos[3];

			if (vertType & GE_VTYPE_NRM_MASK) {
				// Normals are generated during tesselation anyway, not sure if any need to supply
				reader.ReadNrm(nrm);
			} else {
				nrm[0] = 0;
				nrm[1] = 0;
				nrm[2] = 1.0f;
			}
			reader.ReadPos(pos);

			// Apply skinning transform directly
			float weights[8];
			reader.ReadWeights(weights);
			// Skinning
			Vec3Packedf psum(0,0,0);
			Vec3Packedf nsum(0,0,0);
			for (int w = 0; w < numBoneWeights; w++) {
				if (weights[w] != 0.0f) {
					Vec3ByMatrix43(bpos, pos, gstate.boneMatrix+w*12);
					Vec3Packedf tpos(bpos);
					psum += tpos * weights[w];

					Norm3ByMatrix43(bnrm, nrm, gstate.boneMatrix+w*12);
					Vec3Packedf tnorm(bnrm);
					nsum += tnorm * weights[w];
				}
			}
			sv.pos = psum;
			sv.nrm = nsum;
		}
	} else {
		for (int i = lowerBound; i <= upperBound; i++) {
			reader.Goto(i);
			SimpleVertex &sv = sverts[i];
			if (vertType & GE_VTYPE_TC_MASK) {
				reader.ReadUV(sv.uv);
			} else {
				sv.uv[0] = 0;  // This will get filled in during tesselation
				sv.uv[1] = 0;
			}
			if (vertType & GE_VTYPE_COL_MASK) {
				reader.ReadColor0_8888(sv.color);
			} else {
				memcpy(sv.color, defaultColor, 4);
			}
			if (vertType & GE_VTYPE_NRM_MASK) {
				// Normals are generated during tesselation anyway, not sure if any need to supply
				reader.ReadNrm((float *)&sv.nrm);
			} else {
				sv.nrm.x = 0;
				sv.nrm.y = 0;
				sv.nrm.z = 1.0f;
			}
			reader.ReadPos((float *)&sv.pos);
		}
	}

	// Okay, there we are! Return the new type (but keep the index bits)
	return GE_VTYPE_TC_FLOAT | GE_VTYPE_COL_8888 | GE_VTYPE_NRM_FLOAT | GE_VTYPE_POS_FLOAT | (vertType & (GE_VTYPE_IDX_MASK | GE_VTYPE_THROUGH));
}