diff --git a/core/rend/soft/softrend.cpp b/core/rend/soft/softrend.cpp index d40d4ae1c..dca0ea260 100644 --- a/core/rend/soft/softrend.cpp +++ b/core/rend/soft/softrend.cpp @@ -38,11 +38,15 @@ struct softrend : Renderer DECL_ALIGN(32) u32 render_buffer[640 * 480 * 2 * 4]; //Color + depth DECL_ALIGN(32) u32 pixels[640 * 480 * 4]; - static __m128i _mm_load_scaled(int v, int s) + static __m128 _mm_load_scaled_float(float v, float s) { - return _mm_setr_epi32(v, v + s, v + s + s, v + s + s + s); + return _mm_setr_ps(v, v + s, v + s + s, v + s + s + s); } - static __m128i _mm_broadcast(int v) + static __m128 _mm_broadcast_float(float v) + { + return _mm_setr_ps(v, v, v, v); + } + static __m128i _mm_broadcast_int(int v) { __m128i rv = _mm_cvtsi32_si128(v); return _mm_shuffle_epi32(rv, 0); @@ -63,14 +67,14 @@ struct softrend : Renderer return _mm_cvtt_ss2si(_mm_load_ss(&x)); } - int mmin(int a, int b, int c, int d) + float mmin(float a, float b, float c, float d) { int rv = min(a, b); rv = min(c, rv); return max(d, rv); } - int mmax(int a, int b, int c, int d) + float mmax(float a, float b, float c, float d) { int rv = max(a, b); rv = max(c, rv); @@ -79,7 +83,7 @@ struct softrend : Renderer //i think this gives false positives ... //yup, if ANY of the 3 tests fail the ANY tests fails. - __forceinline void EvalHalfSpace(bool& all, bool& any, int cp, int sv, int lv) + __forceinline void EvalHalfSpace(bool& all, bool& any, float cp, float sv, float lv) { //bool a00 = C1 + DX12 * y0 - DY12 * x0 > 0; //bool a10 = C1 + DX12 * y0 - DY12 * x0 > qDY12; @@ -98,31 +102,31 @@ struct softrend : Renderer } //return true if any is positive - __forceinline bool EvalHalfSpaceFAny(int cp12, int cp23, int cp31) + __forceinline bool EvalHalfSpaceFAny(float cp12, float cp23, float cp31) { - int svt = cp12; //needed for ANY - svt |= cp23; - svt |= cp31; + bool svt = cp12 > 0; //needed for ANY + svt |= cp23 > 0; + svt |= cp31 > 0; - return svt>0; + return svt; } - __forceinline bool EvalHalfSpaceFAll(int cp12, int cp23, int cp31, int lv12, int lv23, int lv31) + __forceinline bool EvalHalfSpaceFAll(float cp12, float cp23, float cp31, float lv12, float lv23, float lv31) { - int lvt = cp12 - lv12; - lvt |= cp23 - lv23; - lvt |= cp31 - lv31; //needed for all + bool lvt = (cp12 - lv12) > 0; + lvt &= (cp23 - lv23) > 0; + lvt &= (cp31 - lv31) > 0; //needed for all - return lvt>0; + return lvt; } - __forceinline void PlaneMinMax(int& MIN, int& MAX, int DX, int DY, int q) + __forceinline void PlaneMinMax(float& MIN, float& MAX, float DX, float DY, float q) { - int q_fp = (q - 1) << 4; - int v1 = 0; - int v2 = q_fp*DY; - int v3 = -q_fp*DX; - int v4 = q_fp*(DY - DX); + float q_fp = (q - 1); + float v1 = 0; + float v2 = q_fp*DY; + float v3 = -q_fp*DX; + float v4 = q_fp*(DY - DX); MIN = min(v1, min(v2, min(v3, v4))); MAX = max(v1, max(v2, max(v3, v4))); @@ -334,13 +338,13 @@ struct softrend : Renderer // 28.4 fixed-point coordinates - const int Y1 = iround(16.0f * v1.y); - const int Y2 = iround(16.0f * v2.y); - const int Y3 = iround(16.0f * v3.y); + const float Y1 = v1.y;// iround(16.0f * v1.y); + const float Y2 = v2.y;// iround(16.0f * v2.y); + const float Y3 = v3.y;// iround(16.0f * v3.y); - const int X1 = iround(16.0f * v1.x); - const int X2 = iround(16.0f * v2.x); - const int X3 = iround(16.0f * v3.x); + const float X1 = v1.x;// iround(16.0f * v1.x); + const float X2 = v2.x;// iround(16.0f * v2.x); + const float X3 = v3.x;// iround(16.0f * v3.x); int sgn = 1; @@ -352,93 +356,93 @@ struct softrend : Renderer sgn = -1; } - const int DX12 = sgn*(X1 - X2); - const int DX23 = sgn*(X2 - X3); - const int DX31 = sgn*(X3 - X1); + const float DX12 = sgn*(X1 - X2); + const float DX23 = sgn*(X2 - X3); + const float DX31 = sgn*(X3 - X1); - const int DY12 = sgn*(Y1 - Y2); - const int DY23 = sgn*(Y2 - Y3); - const int DY31 = sgn*(Y3 - Y1); + const float DY12 = sgn*(Y1 - Y2); + const float DY23 = sgn*(Y2 - Y3); + const float DY31 = sgn*(Y3 - Y1); // Fixed-point deltas - const int FDX12 = DX12 << 4; - const int FDX23 = DX23 << 4; - const int FDX31 = DX31 << 4; + const float FDX12 = DX12;// << 4; + const float FDX23 = DX23;// << 4; + const float FDX31 = DX31;// << 4; - const int FDY12 = DY12 << 4; - const int FDY23 = DY23 << 4; - const int FDY31 = DY31 << 4; + const float FDY12 = DY12;// << 4; + const float FDY23 = DY23;// << 4; + const float FDY31 = DY31;// << 4; // Block size, standard 4x4 (must be power of two) const int q = 4; // Bounding rectangle - int minx = (mmin(X1, X2, X3, 0) + 0xF) >> 4; - int miny = (mmin(Y1, Y2, Y3, 0) + 0xF) >> 4; + int minx = iround(mmin(X1, X2, X3, 0) );// +0xF) >> 4; + int miny = iround(mmin(Y1, Y2, Y3, 0) );// +0xF) >> 4; // Start in corner of block minx &= ~(q - 1); miny &= ~(q - 1); - int spanx = ((mmax(X1, X2, X3, 640 << 4) + 0xF) >> 4) - minx; - int spany = ((mmax(Y1, Y2, Y3, 480 << 4) + 0xF) >> 4) - miny; + int spanx = iround(mmax(X1 + 0.5f, X2 + 0.5f, X3 + 0.5f, 640)) - minx; + int spany = iround(mmax(Y1 + 0.5f, Y2 + 0.5f, Y3 + 0.5f, 480)) - miny; // Half-edge constants - int C1 = DY12 * X1 - DX12 * Y1; - int C2 = DY23 * X2 - DX23 * Y2; - int C3 = DY31 * X3 - DX31 * Y3; + float C1 = DY12 * X1 - DX12 * Y1; + float C2 = DY23 * X2 - DX23 * Y2; + float C3 = DY31 * X3 - DX31 * Y3; // Correct for fill convention if (DY12 < 0 || (DY12 == 0 && DX12 > 0)) C1++; if (DY23 < 0 || (DY23 == 0 && DX23 > 0)) C2++; if (DY31 < 0 || (DY31 == 0 && DX31 > 0)) C3++; - int MAX_12, MAX_23, MAX_31, MIN_12, MIN_23, MIN_31; + float MAX_12, MAX_23, MAX_31, MIN_12, MIN_23, MIN_31; PlaneMinMax(MIN_12, MAX_12, DX12, DY12, q); PlaneMinMax(MIN_23, MAX_23, DX23, DY23, q); PlaneMinMax(MIN_31, MAX_31, DX31, DY31, q); - const int FDqX12 = FDX12 * q; - const int FDqX23 = FDX23 * q; - const int FDqX31 = FDX31 * q; + const float FDqX12 = FDX12 * q; + const float FDqX23 = FDX23 * q; + const float FDqX31 = FDX31 * q; - const int FDqY12 = FDY12 * q; - const int FDqY23 = FDY23 * q; - const int FDqY31 = FDY31 * q; + const float FDqY12 = FDY12 * q; + const float FDqY23 = FDY23 * q; + const float FDqY31 = FDY31 * q; - const int FDX12mq = FDX12 + FDY12*q; - const int FDX23mq = FDX23 + FDY23*q; - const int FDX31mq = FDX31 + FDY31*q; + const float FDX12mq = FDX12 + FDY12*q; + const float FDX23mq = FDX23 + FDY23*q; + const float FDX31mq = FDX31 + FDY31*q; - int hs12 = C1 + FDX12 * miny - FDY12 * minx + FDqY12 - MIN_12; - int hs23 = C2 + FDX23 * miny - FDY23 * minx + FDqY23 - MIN_23; - int hs31 = C3 + FDX31 * miny - FDY31 * minx + FDqY31 - MIN_31; + float hs12 = C1 + FDX12 * (miny+0.5f) - FDY12 * (minx+0.5f) + FDqY12 - MIN_12; + float hs23 = C2 + FDX23 * (miny+0.5f) - FDY23 * (minx+0.5f) + FDqY23 - MIN_23; + float hs31 = C3 + FDX31 * (miny+0.5f) - FDY31 * (minx+0.5f) + FDqY31 - MIN_31; MAX_12 -= MIN_12; MAX_23 -= MIN_23; MAX_31 -= MIN_31; - int C1_pm = MIN_12; - int C2_pm = MIN_23; - int C3_pm = MIN_31; + float C1_pm = MIN_12; + float C2_pm = MIN_23; + float C3_pm = MIN_31; u8* cb_y = (u8*)colorBuffer; cb_y += miny*stride + minx*(q * 4); ip.Setup(v1, v2, v3, minx, miny, q); - __m128 y_ps = _mm_cvtepi32_ps(_mm_broadcast(miny)); - __m128 minx_ps = _mm_cvtepi32_ps(_mm_load_scaled(minx - q, 1)); + __m128 y_ps = _mm_broadcast_float(miny); + __m128 minx_ps = _mm_load_scaled_float(minx - q, 1); static __declspec(align(16)) float ones_ps[4] = { 1, 1, 1, 1 }; static __declspec(align(16)) float q_ps[4] = { q, q, q, q }; // Loop through blocks for (int y = spany; y > 0; y -= q) { - int Xhs12 = hs12; - int Xhs23 = hs23; - int Xhs31 = hs31; + float Xhs12 = hs12; + float Xhs23 = hs23; + float Xhs31 = hs31; u8* cb_x = cb_y; __m128 x_ps = minx_ps; for (int x = spanx; x > 0; x -= q) @@ -447,7 +451,7 @@ struct softrend : Renderer Xhs23 -= FDqY23; Xhs31 -= FDqY31; x_ps = _mm_add_ps(x_ps, *(__m128*)q_ps); - + // Corners of block bool any = EvalHalfSpaceFAny(Xhs12, Xhs23, Xhs31); @@ -459,7 +463,7 @@ struct softrend : Renderer } bool all = EvalHalfSpaceFAll(Xhs12, Xhs23, Xhs31, MAX_12, MAX_23, MAX_31); - + // Accept whole block when totally covered if (all) { @@ -473,27 +477,32 @@ struct softrend : Renderer } else // Partially covered block { - int CY1 = C1_pm + Xhs12; - int CY2 = C2_pm + Xhs23; - int CY3 = C3_pm + Xhs31; + float CY1 = C1_pm + Xhs12; + float CY2 = C2_pm + Xhs23; + float CY3 = C3_pm + Xhs31; - __m128i pfdx12 = _mm_broadcast(FDX12); - __m128i pfdx23 = _mm_broadcast(FDX23); - __m128i pfdx31 = _mm_broadcast(FDX31); + __m128 pfdx12 = _mm_broadcast_float(FDX12); + __m128 pfdx23 = _mm_broadcast_float(FDX23); + __m128 pfdx31 = _mm_broadcast_float(FDX31); - __m128i pcy1 = _mm_load_scaled(CY1, -FDY12); - __m128i pcy2 = _mm_load_scaled(CY2, -FDY23); - __m128i pcy3 = _mm_load_scaled(CY3, -FDY31); + __m128 pcy1 = _mm_load_scaled_float(CY1, -FDY12); + __m128 pcy2 = _mm_load_scaled_float(CY2, -FDY23); + __m128 pcy3 = _mm_load_scaled_float(CY3, -FDY31); - __m128i pzero = _mm_setzero_si128(); + __m128 pzero = _mm_setzero_ps(); //bool ok=false; __m128 yl_ps = y_ps; for (int iy = q; iy > 0; iy--) { - __m128i a = _mm_cmpgt_epi32(_mm_or_si128(_mm_or_si128(pcy1, pcy2), pcy3), pzero); - int msk = _mm_movemask_ps(*(__m128*)&a); + __m128 mask1 = _mm_cmple_ps(pcy1, pzero); + __m128 mask2 = _mm_cmple_ps(pcy2, pzero); + __m128 mask3 = _mm_cmple_ps(pcy3, pzero); + __m128 summary = _mm_or_ps(mask3, _mm_or_ps(mask2, mask1)); + + __m128i a = _mm_cmpeq_epi32((__m128i&)summary, (__m128i&)pzero); + int msk = _mm_movemask_ps((__m128&)a); if (msk != 0) { PixelFlush(x_ps, yl_ps, cb_x, *(__m128*)&a); @@ -505,9 +514,9 @@ struct softrend : Renderer //CY1 += FDX12mq; //CY2 += FDX23mq; //CY3 += FDX31mq; - pcy1 = _mm_add_epi32(pcy1, pfdx12); - pcy2 = _mm_add_epi32(pcy2, pfdx23); - pcy3 = _mm_add_epi32(pcy3, pfdx31); + pcy1 = _mm_add_ps(pcy1, pfdx12); + pcy2 = _mm_add_ps(pcy2, pfdx23); + pcy3 = _mm_add_ps(pcy3, pfdx31); } /* if (!ok)