Chromium Code Reviews
chromiumcodereview-hr@appspot.gserviceaccount.com (chromiumcodereview-hr) | Please choose your nickname with Settings | Help | Chromium Project | Gerrit Changes | Sign out
(577)

Issues Search
    My Issues | Starred     Open | Closed | All

Unified Diff: third_party/WebKit/Source/platform/transforms/TransformationMatrix.cpp

Issue 2392493002: Add MSA (MIPS SIMD Arch) optimized matrix transforms functions (Closed)
Patch Set: Created 4 years, 2 months ago
Use n/p to move between diff chunks; N/P to move between comments. Draft comments are only viewable by you.
Jump to:
View side-by-side diff with in-line comments
Download patch
« no previous file with comments | « no previous file | no next file » | no next file with comments »
Expand Comments ('e') | Collapse Comments ('c') | Show Comments Hide Comments ('s')
Index: third_party/WebKit/Source/platform/transforms/TransformationMatrix.cpp
diff --git a/third_party/WebKit/Source/platform/transforms/TransformationMatrix.cpp b/third_party/WebKit/Source/platform/transforms/TransformationMatrix.cpp
index ab6cf39cece65fc727eb82d56c03a144a8fe0293..89a7abe515c912bcf4632490961aefb2b8ae9dc2 100644
--- a/third_party/WebKit/Source/platform/transforms/TransformationMatrix.cpp
+++ b/third_party/WebKit/Source/platform/transforms/TransformationMatrix.cpp
@@ -26,6 +26,9 @@
*/
#include "platform/transforms/TransformationMatrix.h"
+#if HAVE(MIPS_MSA_INTRINSICS)
+#include "platform/cpu/mips/CommonMacrosMSA.h"
+#endif
#include "platform/geometry/FloatBox.h"
#include "platform/geometry/FloatQuad.h"
@@ -344,6 +347,147 @@ static bool inverse(const TransformationMatrix::Matrix4& matrix,
: "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17",
"v18", "v19", "v20", "v21", "v22", "v23", "24", "25", "v26", "v27",
"v28", "v29", "v30");
+#elif HAVE(MIPS_MSA_INTRINSICS)
+ const double rDet = 1/det;
+ const double* mat = &(matrix[0][0]);
+ v2f64 mat0, mat1, mat2, mat3, mat4, mat5, mat6, mat7;
+ v2f64 rev2, rev3, rev4, rev5, rev6, rev7;
+ v2f64 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+ v2f64 det0, det1, det2, tmp8, tmp9, tmp10, tmp11;
+ const v2f64 rdet = COPY_DOUBLE_TO_VECTOR(rDet);
+ // mat0 mat1 --> m00 m01 m02 m03
+ // mat2 mat3 --> m10 m11 m12 m13
+ // mat4 mat5 --> m20 m21 m22 m23
+ // mat6 mat7 --> m30 m31 m32 m33
+ LD_DP8(mat, 2, mat0, mat1, mat2, mat3, mat4, mat5, mat6, mat7);
+
+ // Right half
+ rev3 = SLDI_D(mat3, mat3, 8); // m13 m12
+ rev5 = SLDI_D(mat5, mat5, 8); // m23 m22
+ rev7 = SLDI_D(mat7, mat7, 8); // m33 m32
+
+ // 2*2 Determinants
+ // for A00 & A01
+ tmp0 = mat5 * rev7;
+ tmp1 = mat3 * rev7;
+ tmp2 = mat3 * rev5;
+ // for A10 & A11
+ tmp3 = mat1 * rev7;
+ tmp4 = mat1 * rev5;
+ // for A20 & A21
+ tmp5 = mat1 * rev3;
+ // for A30 & A31
+ tmp6 = (v2f64) __msa_ilvr_d((v2i64) tmp1, (v2i64) tmp0);
+ tmp7 = (v2f64) __msa_ilvl_d((v2i64) tmp1, (v2i64) tmp0);
+ det0 = tmp6 - tmp7;
+ tmp6 = (v2f64) __msa_ilvr_d((v2i64) tmp3, (v2i64) tmp2);
+ tmp7 = (v2f64) __msa_ilvl_d((v2i64) tmp3, (v2i64) tmp2);
+ det1 = tmp6 - tmp7;
+ tmp6 = (v2f64) __msa_ilvr_d((v2i64) tmp5, (v2i64) tmp4);
+ tmp7 = (v2f64) __msa_ilvl_d((v2i64) tmp5, (v2i64) tmp4);
+ det2 = tmp6 - tmp7;
+
+ // Co-factors
+ tmp0 = mat0 * (v2f64) __msa_splati_d((v2i64) det0, 0);
+ tmp1 = mat0 * (v2f64) __msa_splati_d((v2i64) det0, 1);
+ tmp2 = mat0 * (v2f64) __msa_splati_d((v2i64) det1, 0);
+ tmp3 = mat2 * (v2f64) __msa_splati_d((v2i64) det0, 0);
+ tmp4 = mat2 * (v2f64) __msa_splati_d((v2i64) det1, 1);
+ tmp5 = mat2 * (v2f64) __msa_splati_d((v2i64) det2, 0);
+ tmp6 = mat4 * (v2f64) __msa_splati_d((v2i64) det0, 1);
+ tmp7 = mat4 * (v2f64) __msa_splati_d((v2i64) det1, 1);
+ tmp8 = mat4 * (v2f64) __msa_splati_d((v2i64) det2, 1);
+ tmp9 = mat6 * (v2f64) __msa_splati_d((v2i64) det1, 0);
+ tmp10 = mat6 * (v2f64) __msa_splati_d((v2i64) det2, 0);
+ tmp11 = mat6 * (v2f64) __msa_splati_d((v2i64) det2, 1);
+
+ tmp0 -= tmp7;
+ tmp1 -= tmp4;
+ tmp2 -= tmp5;
+ tmp3 -= tmp6;
+ tmp0 += tmp10;
+ tmp1 += tmp11;
+ tmp2 += tmp8;
+ tmp3 += tmp9;
+
+ // Multiply with 1/det
+ tmp0 *= rdet;
+ tmp1 *= rdet;
+ tmp2 *= rdet;
+ tmp3 *= rdet;
+
+ // Inverse: Upper half
+ result[0][0] = tmp3[1];
+ result[0][1] = -tmp0[1];
+ result[0][2] = tmp1[1];
+ result[0][3] = -tmp2[1];
+ result[1][0] = -tmp3[0];
+ result[1][1] = tmp0[0];
+ result[1][2] = -tmp1[0];
+ result[1][3] = tmp2[0];
+ // Left half
+ rev2 = SLDI_D(mat2, mat2, 8); // m13 m12
+ rev4 = SLDI_D(mat4, mat4, 8); // m23 m22
+ rev6 = SLDI_D(mat6, mat6, 8); // m33 m32
+
+ // 2*2 Determinants
+ // for A00 & A01
+ tmp0 = mat4 * rev6;
+ tmp1 = mat2 * rev6;
+ tmp2 = mat2 * rev4;
+ // for A10 & A11
+ tmp3 = mat0 * rev6;
+ tmp4 = mat0 * rev4;
+ // for A20 & A21
+ tmp5 = mat0 * rev2;
+ // for A30 & A31
+ tmp6 = (v2f64) __msa_ilvr_d((v2i64) tmp1, (v2i64) tmp0);
+ tmp7 = (v2f64) __msa_ilvl_d((v2i64) tmp1, (v2i64) tmp0);
+ det0 = tmp6 - tmp7;
+ tmp6 = (v2f64) __msa_ilvr_d((v2i64) tmp3, (v2i64) tmp2);
+ tmp7 = (v2f64) __msa_ilvl_d((v2i64) tmp3, (v2i64) tmp2);
+ det1 = tmp6 - tmp7;
+ tmp6 = (v2f64) __msa_ilvr_d((v2i64) tmp5, (v2i64) tmp4);
+ tmp7 = (v2f64) __msa_ilvl_d((v2i64) tmp5, (v2i64) tmp4);
+ det2 = tmp6 - tmp7;
+
+ // Co-factors
+ tmp0 = mat3 * (v2f64) __msa_splati_d((v2i64) det0, 0);
+ tmp1 = mat1 * (v2f64) __msa_splati_d((v2i64) det0, 1);
+ tmp2 = mat1 * (v2f64) __msa_splati_d((v2i64) det0, 0);
+ tmp3 = mat1 * (v2f64) __msa_splati_d((v2i64) det1, 0);
+ tmp4 = mat3 * (v2f64) __msa_splati_d((v2i64) det1, 1);
+ tmp5 = mat3 * (v2f64) __msa_splati_d((v2i64) det2, 0);
+ tmp6 = mat5 * (v2f64) __msa_splati_d((v2i64) det0, 1);
+ tmp7 = mat5 * (v2f64) __msa_splati_d((v2i64) det1, 1);
+ tmp8 = mat5 * (v2f64) __msa_splati_d((v2i64) det2, 1);
+ tmp9 = mat7 * (v2f64) __msa_splati_d((v2i64) det1, 0);
+ tmp10 = mat7 * (v2f64) __msa_splati_d((v2i64) det2, 0);
+ tmp11 = mat7 * (v2f64) __msa_splati_d((v2i64) det2, 1);
+ tmp0 -= tmp6;
+ tmp1 -= tmp4;
+ tmp2 -= tmp7;
+ tmp3 -= tmp5;
+ tmp0 += tmp9;
+ tmp1 += tmp11;
+ tmp2 += tmp10;
+ tmp3 += tmp8;
+
+ // Multiply with 1/det
+ tmp0 *= rdet;
+ tmp1 *= rdet;
+ tmp2 *= rdet;
+ tmp3 *= rdet;
+
+ // Inverse: Lower half
+ result[2][0] = tmp0[1];
+ result[2][1] = -tmp2[1];
+ result[2][2] = tmp1[1];
+ result[2][3] = -tmp3[1];
+ result[3][0] = -tmp0[0];
+ result[3][1] = tmp2[0];
+ result[3][2] = -tmp1[0];
+ result[3][3] = tmp3[0];
#else
// Calculate the adjoint matrix
adjoint(matrix, result);
@@ -1185,6 +1329,93 @@ TransformationMatrix& TransformationMatrix::multiply(
: "memory", "x9", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23",
"v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "v0", "v1",
"v2", "v3", "v4", "v5", "v6", "v7");
+#elif HAVE(MIPS_MSA_INTRINSICS)
+ v2f64 vleftM0, vleftM1, vleftM2, vleftM3, vleftM4, vleftM5, vleftM6, vleftM7;
+ v2f64 vRightM0, vRightM1, vRightM2, vRightM3, vRightM4, vRightM5, vRightM6, vRightM7;
+ v2f64 vTmpM0, vTmpM1, vTmpM2, vTmpM3;
+
+ vRightM0 = LD_DP(&(m_matrix[0][0]));
+ vRightM1 = LD_DP(&(m_matrix[0][2]));
+ vRightM2 = LD_DP(&(m_matrix[1][0]));
+ vRightM3 = LD_DP(&(m_matrix[1][2]));
+ vRightM4 = LD_DP(&(m_matrix[2][0]));
+ vRightM5 = LD_DP(&(m_matrix[2][2]));
+ vRightM6 = LD_DP(&(m_matrix[3][0]));
+ vRightM7 = LD_DP(&(m_matrix[3][2]));
+
+ vleftM0 = LD_DP(&(mat.m_matrix[0][0]));
+ vleftM2 = LD_DP(&(mat.m_matrix[0][2]));
+ vleftM4 = LD_DP(&(mat.m_matrix[1][0]));
+ vleftM6 = LD_DP(&(mat.m_matrix[1][2]));
+
+ vleftM1 = (v2f64)__msa_splati_d((v2i64)vleftM0, 1);
+ vleftM0 = (v2f64)__msa_splati_d((v2i64)vleftM0, 0);
+ vleftM3 = (v2f64)__msa_splati_d((v2i64)vleftM2, 1);
+ vleftM2 = (v2f64)__msa_splati_d((v2i64)vleftM2, 0);
+ vleftM5 = (v2f64)__msa_splati_d((v2i64)vleftM4, 1);
+ vleftM4 = (v2f64)__msa_splati_d((v2i64)vleftM4, 0);
+ vleftM7 = (v2f64)__msa_splati_d((v2i64)vleftM6, 1);
+ vleftM6 = (v2f64)__msa_splati_d((v2i64)vleftM6, 0);
+
+ vTmpM0 = vleftM0 * vRightM0;
+ vTmpM1 = vleftM0 * vRightM1;
+ vTmpM0 += vleftM1 * vRightM2;
+ vTmpM1 += vleftM1 * vRightM3;
+ vTmpM0 += vleftM2 * vRightM4;
+ vTmpM1 += vleftM2 * vRightM5;
+ vTmpM0 += vleftM3 * vRightM6;
+ vTmpM1 += vleftM3 * vRightM7;
+
+ vTmpM2 = vleftM4 * vRightM0;
+ vTmpM3 = vleftM4 * vRightM1;
+ vTmpM2 += vleftM5 * vRightM2;
+ vTmpM3 += vleftM5 * vRightM3;
+ vTmpM2 += vleftM6 * vRightM4;
+ vTmpM3 += vleftM6 * vRightM5;
+ vTmpM2 += vleftM7 * vRightM6;
+ vTmpM3 += vleftM7 * vRightM7;
+
+ vleftM0 = LD_DP(&(mat.m_matrix[2][0]));
+ vleftM2 = LD_DP(&(mat.m_matrix[2][2]));
+ vleftM4 = LD_DP(&(mat.m_matrix[3][0]));
+ vleftM6 = LD_DP(&(mat.m_matrix[3][2]));
+
+ ST_DP(vTmpM0, &(m_matrix[0][0]));
+ ST_DP(vTmpM1, &(m_matrix[0][2]));
+ ST_DP(vTmpM2, &(m_matrix[1][0]));
+ ST_DP(vTmpM3, &(m_matrix[1][2]));
+
+ vleftM1 = (v2f64)__msa_splati_d((v2i64)vleftM0, 1);
+ vleftM0 = (v2f64)__msa_splati_d((v2i64)vleftM0, 0);
+ vleftM3 = (v2f64)__msa_splati_d((v2i64)vleftM2, 1);
+ vleftM2 = (v2f64)__msa_splati_d((v2i64)vleftM2, 0);
+ vleftM5 = (v2f64)__msa_splati_d((v2i64)vleftM4, 1);
+ vleftM4 = (v2f64)__msa_splati_d((v2i64)vleftM4, 0);
+ vleftM7 = (v2f64)__msa_splati_d((v2i64)vleftM6, 1);
+ vleftM6 = (v2f64)__msa_splati_d((v2i64)vleftM6, 0);
+
+ vTmpM0 = vleftM0 * vRightM0;
+ vTmpM1 = vleftM0 * vRightM1;
+ vTmpM0 += vleftM1 * vRightM2;
+ vTmpM1 += vleftM1 * vRightM3;
+ vTmpM0 += vleftM2 * vRightM4;
+ vTmpM1 += vleftM2 * vRightM5;
+ vTmpM0 += vleftM3 * vRightM6;
+ vTmpM1 += vleftM3 * vRightM7;
+
+ vTmpM2 = vleftM4 * vRightM0;
+ vTmpM3 = vleftM4 * vRightM1;
+ vTmpM2 += vleftM5 * vRightM2;
+ vTmpM3 += vleftM5 * vRightM3;
+ vTmpM2 += vleftM6 * vRightM4;
+ vTmpM3 += vleftM6 * vRightM5;
+ vTmpM2 += vleftM7 * vRightM6;
+ vTmpM3 += vleftM7 * vRightM7;
+
+ ST_DP(vTmpM0, &(m_matrix[2][0]));
+ ST_DP(vTmpM1, &(m_matrix[2][2]));
+ ST_DP(vTmpM2, &(m_matrix[3][0]));
+ ST_DP(vTmpM3, &(m_matrix[3][2]));
#elif defined(TRANSFORMATION_MATRIX_USE_X86_64_SSE2)
// x86_64 has 16 XMM registers which is enough to do the multiplication fully in registers.
__m128d matrixBlockA = _mm_load_pd(&(m_matrix[0][0]));
« no previous file with comments | « no previous file | no next file » | no next file with comments »

Powered by Google App Engine
This is Rietveld 408576698