ShuriZma
/
desmume
mirror of https://github.com/TASEmulators/desmume.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

matrix.cpp: Remove saturation logic from MatrixMultiply(), as this causes characters to disappear during "Kingdom Hearts: Re-coded" character conversations. 

- This bug was found by enabling FIXED_POINT_MATH_FUNCTIONS_USE_ACCUMULATOR_SATURATE. Since this macro is disabled by default, this commit should not affect any normal operation.
This commit is contained in:
rogerman 2022-04-18 23:29:21 -07:00
parent 803c885ef2
commit 45a2f424e6
1 changed files with 32 additions and 23 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

55
desmume/src/matrix.cpp
View File

@ -26,8 +26,8 @@
// NDS matrix math functions uses 20.12 fixed-point for calculations. According to // NDS matrix math functions uses 20.12 fixed-point for calculations. According to
// GEM_TransformVertex(), dot product calculations use accumulation that goes beyond // GEM_TransformVertex(), dot product calculations use accumulation that goes beyond
// 32-bits and then saturates. Therefore, all fixed-point math functions will also // 32-bits and then saturates. Therefore, all fixed-point math functions (with the
// support that feature here. // exception of matrix-by-matrix multiplication,) will also support that feature here.
// //
// But for historical reasons, we can't enable this right away. Therefore, the scalar // But for historical reasons, we can't enable this right away. Therefore, the scalar
// function GEM_TransformVertex() will continue to be used for SetVertex() while these // function GEM_TransformVertex() will continue to be used for SetVertex() while these
@ -619,25 +619,30 @@ static FORCEINLINE void __mtx4_multiply_mtx4_fixed(s32 (&__restrict mtxA)[16], c
CACHE_ALIGN s32 a[16]; CACHE_ALIGN s32 a[16];
MatrixCopy(a, mtxA); MatrixCopy(a, mtxA);
mtxA[ 0] = ___s32_saturate_shiftdown_accum64_fixed( fx32_mul(a[ 0],mtxB[ 0]) + fx32_mul(a[ 4],mtxB[ 1]) + fx32_mul(a[ 8],mtxB[ 2]) + fx32_mul(a[12],mtxB[ 3]) ); // We can't saturate the accumulated results here because it breaks
mtxA[ 1] = ___s32_saturate_shiftdown_accum64_fixed( fx32_mul(a[ 1],mtxB[ 0]) + fx32_mul(a[ 5],mtxB[ 1]) + fx32_mul(a[ 9],mtxB[ 2]) + fx32_mul(a[13],mtxB[ 3]) ); // character conversations in "Kingdom Hears: Re-coded", causing the
mtxA[ 2] = ___s32_saturate_shiftdown_accum64_fixed( fx32_mul(a[ 2],mtxB[ 0]) + fx32_mul(a[ 6],mtxB[ 1]) + fx32_mul(a[10],mtxB[ 2]) + fx32_mul(a[14],mtxB[ 3]) ); // characters to disappear. Therefore, we will simply do a standalone
mtxA[ 3] = ___s32_saturate_shiftdown_accum64_fixed( fx32_mul(a[ 3],mtxB[ 0]) + fx32_mul(a[ 7],mtxB[ 1]) + fx32_mul(a[11],mtxB[ 2]) + fx32_mul(a[15],mtxB[ 3]) ); // shiftdown, and that's it.
mtxA[ 4] = ___s32_saturate_shiftdown_accum64_fixed( fx32_mul(a[ 0],mtxB[ 4]) + fx32_mul(a[ 4],mtxB[ 5]) + fx32_mul(a[ 8],mtxB[ 6]) + fx32_mul(a[12],mtxB[ 7]) ); mtxA[ 0] = sfx32_shiftdown( fx32_mul(a[ 0],mtxB[ 0]) + fx32_mul(a[ 4],mtxB[ 1]) + fx32_mul(a[ 8],mtxB[ 2]) + fx32_mul(a[12],mtxB[ 3]) );
mtxA[ 5] = ___s32_saturate_shiftdown_accum64_fixed( fx32_mul(a[ 1],mtxB[ 4]) + fx32_mul(a[ 5],mtxB[ 5]) + fx32_mul(a[ 9],mtxB[ 6]) + fx32_mul(a[13],mtxB[ 7]) ); mtxA[ 1] = sfx32_shiftdown( fx32_mul(a[ 1],mtxB[ 0]) + fx32_mul(a[ 5],mtxB[ 1]) + fx32_mul(a[ 9],mtxB[ 2]) + fx32_mul(a[13],mtxB[ 3]) );
mtxA[ 6] = ___s32_saturate_shiftdown_accum64_fixed( fx32_mul(a[ 2],mtxB[ 4]) + fx32_mul(a[ 6],mtxB[ 5]) + fx32_mul(a[10],mtxB[ 6]) + fx32_mul(a[14],mtxB[ 7]) ); mtxA[ 2] = sfx32_shiftdown( fx32_mul(a[ 2],mtxB[ 0]) + fx32_mul(a[ 6],mtxB[ 1]) + fx32_mul(a[10],mtxB[ 2]) + fx32_mul(a[14],mtxB[ 3]) );
mtxA[ 7] = ___s32_saturate_shiftdown_accum64_fixed( fx32_mul(a[ 3],mtxB[ 4]) + fx32_mul(a[ 7],mtxB[ 5]) + fx32_mul(a[11],mtxB[ 6]) + fx32_mul(a[15],mtxB[ 7]) ); mtxA[ 3] = sfx32_shiftdown( fx32_mul(a[ 3],mtxB[ 0]) + fx32_mul(a[ 7],mtxB[ 1]) + fx32_mul(a[11],mtxB[ 2]) + fx32_mul(a[15],mtxB[ 3]) );
mtxA[ 8] = ___s32_saturate_shiftdown_accum64_fixed( fx32_mul(a[ 0],mtxB[ 8]) + fx32_mul(a[ 4],mtxB[ 9]) + fx32_mul(a[ 8],mtxB[10]) + fx32_mul(a[12],mtxB[11]) ); mtxA[ 4] = sfx32_shiftdown( fx32_mul(a[ 0],mtxB[ 4]) + fx32_mul(a[ 4],mtxB[ 5]) + fx32_mul(a[ 8],mtxB[ 6]) + fx32_mul(a[12],mtxB[ 7]) );
mtxA[ 9] = ___s32_saturate_shiftdown_accum64_fixed( fx32_mul(a[ 1],mtxB[ 8]) + fx32_mul(a[ 5],mtxB[ 9]) + fx32_mul(a[ 9],mtxB[10]) + fx32_mul(a[13],mtxB[11]) ); mtxA[ 5] = sfx32_shiftdown( fx32_mul(a[ 1],mtxB[ 4]) + fx32_mul(a[ 5],mtxB[ 5]) + fx32_mul(a[ 9],mtxB[ 6]) + fx32_mul(a[13],mtxB[ 7]) );
mtxA[10] = ___s32_saturate_shiftdown_accum64_fixed( fx32_mul(a[ 2],mtxB[ 8]) + fx32_mul(a[ 6],mtxB[ 9]) + fx32_mul(a[10],mtxB[10]) + fx32_mul(a[14],mtxB[11]) ); mtxA[ 6] = sfx32_shiftdown( fx32_mul(a[ 2],mtxB[ 4]) + fx32_mul(a[ 6],mtxB[ 5]) + fx32_mul(a[10],mtxB[ 6]) + fx32_mul(a[14],mtxB[ 7]) );
mtxA[11] = ___s32_saturate_shiftdown_accum64_fixed( fx32_mul(a[ 3],mtxB[ 8]) + fx32_mul(a[ 7],mtxB[ 9]) + fx32_mul(a[11],mtxB[10]) + fx32_mul(a[15],mtxB[11]) ); mtxA[ 7] = sfx32_shiftdown( fx32_mul(a[ 3],mtxB[ 4]) + fx32_mul(a[ 7],mtxB[ 5]) + fx32_mul(a[11],mtxB[ 6]) + fx32_mul(a[15],mtxB[ 7]) );
mtxA[12] = ___s32_saturate_shiftdown_accum64_fixed( fx32_mul(a[ 0],mtxB[12]) + fx32_mul(a[ 4],mtxB[13]) + fx32_mul(a[ 8],mtxB[14]) + fx32_mul(a[12],mtxB[15]) ); mtxA[ 8] = sfx32_shiftdown( fx32_mul(a[ 0],mtxB[ 8]) + fx32_mul(a[ 4],mtxB[ 9]) + fx32_mul(a[ 8],mtxB[10]) + fx32_mul(a[12],mtxB[11]) );
mtxA[13] = ___s32_saturate_shiftdown_accum64_fixed( fx32_mul(a[ 1],mtxB[12]) + fx32_mul(a[ 5],mtxB[13]) + fx32_mul(a[ 9],mtxB[14]) + fx32_mul(a[13],mtxB[15]) ); mtxA[ 9] = sfx32_shiftdown( fx32_mul(a[ 1],mtxB[ 8]) + fx32_mul(a[ 5],mtxB[ 9]) + fx32_mul(a[ 9],mtxB[10]) + fx32_mul(a[13],mtxB[11]) );
mtxA[14] = ___s32_saturate_shiftdown_accum64_fixed( fx32_mul(a[ 2],mtxB[12]) + fx32_mul(a[ 6],mtxB[13]) + fx32_mul(a[10],mtxB[14]) + fx32_mul(a[14],mtxB[15]) ); mtxA[10] = sfx32_shiftdown( fx32_mul(a[ 2],mtxB[ 8]) + fx32_mul(a[ 6],mtxB[ 9]) + fx32_mul(a[10],mtxB[10]) + fx32_mul(a[14],mtxB[11]) );
mtxA[15] = ___s32_saturate_shiftdown_accum64_fixed( fx32_mul(a[ 3],mtxB[12]) + fx32_mul(a[ 7],mtxB[13]) + fx32_mul(a[11],mtxB[14]) + fx32_mul(a[15],mtxB[15]) ); mtxA[11] = sfx32_shiftdown( fx32_mul(a[ 3],mtxB[ 8]) + fx32_mul(a[ 7],mtxB[ 9]) + fx32_mul(a[11],mtxB[10]) + fx32_mul(a[15],mtxB[11]) );
mtxA[12] = sfx32_shiftdown( fx32_mul(a[ 0],mtxB[12]) + fx32_mul(a[ 4],mtxB[13]) + fx32_mul(a[ 8],mtxB[14]) + fx32_mul(a[12],mtxB[15]) );
mtxA[13] = sfx32_shiftdown( fx32_mul(a[ 1],mtxB[12]) + fx32_mul(a[ 5],mtxB[13]) + fx32_mul(a[ 9],mtxB[14]) + fx32_mul(a[13],mtxB[15]) );
mtxA[14] = sfx32_shiftdown( fx32_mul(a[ 2],mtxB[12]) + fx32_mul(a[ 6],mtxB[13]) + fx32_mul(a[10],mtxB[14]) + fx32_mul(a[14],mtxB[15]) );
mtxA[15] = sfx32_shiftdown( fx32_mul(a[ 3],mtxB[12]) + fx32_mul(a[ 7],mtxB[13]) + fx32_mul(a[11],mtxB[14]) + fx32_mul(a[15],mtxB[15]) );
} }
static FORCEINLINE void __mtx4_scale_vec3_fixed(s32 (&__restrict inoutMtx)[16], const s32 (&__restrict inVec)[4]) static FORCEINLINE void __mtx4_scale_vec3_fixed(s32 (&__restrict inoutMtx)[16], const s32 (&__restrict inVec)[4])
@ -832,12 +837,14 @@ static FORCEINLINE void __mtx4_multiply_mtx4_fixed_SSE4(s32 (&__restrict mtxA)[1
outVecLo = _mm_add_epi64( outVecLo, _mm_mul_epi32(rowLo[1], v[1]) );\ outVecLo = _mm_add_epi64( outVecLo, _mm_mul_epi32(rowLo[1], v[1]) );\
outVecLo = _mm_add_epi64( outVecLo, _mm_mul_epi32(rowLo[2], v[2]) );\ outVecLo = _mm_add_epi64( outVecLo, _mm_mul_epi32(rowLo[2], v[2]) );\
outVecLo = _mm_add_epi64( outVecLo, _mm_mul_epi32(rowLo[3], v[3]) );\ outVecLo = _mm_add_epi64( outVecLo, _mm_mul_epi32(rowLo[3], v[3]) );\
___s32_saturate_shiftdown_accum64_fixed_SSE4(outVecLo);\ outVecLo = _mm_srli_epi64(outVecLo, 12);\
outVecLo = _mm_shuffle_epi32(outVecLo, 0xD8);\
outVecHi = _mm_mul_epi32(rowHi[0], v[0]);\ outVecHi = _mm_mul_epi32(rowHi[0], v[0]);\
outVecHi = _mm_add_epi64( outVecHi, _mm_mul_epi32(rowHi[1], v[1]) );\ outVecHi = _mm_add_epi64( outVecHi, _mm_mul_epi32(rowHi[1], v[1]) );\
outVecHi = _mm_add_epi64( outVecHi, _mm_mul_epi32(rowHi[2], v[2]) );\ outVecHi = _mm_add_epi64( outVecHi, _mm_mul_epi32(rowHi[2], v[2]) );\
outVecHi = _mm_add_epi64( outVecHi, _mm_mul_epi32(rowHi[3], v[3]) );\ outVecHi = _mm_add_epi64( outVecHi, _mm_mul_epi32(rowHi[3], v[3]) );\
___s32_saturate_shiftdown_accum64_fixed_SSE4(outVecHi); outVecHi = _mm_srli_epi64(outVecHi, 12);\
outVecHi = _mm_shuffle_epi32(outVecHi, 0xD8);
CALCULATE_MATRIX_ROW_FIXED_SSE4(0); CALCULATE_MATRIX_ROW_FIXED_SSE4(0);
_mm_store_si128( (v128s32 *)(mtxA + 0), _mm_unpacklo_epi64(outVecLo, outVecHi) ); _mm_store_si128( (v128s32 *)(mtxA + 0), _mm_unpacklo_epi64(outVecLo, outVecHi) );
@ -971,7 +978,7 @@ static FORCEINLINE void ___s32_saturate_shiftdown_accum64_fixed_NEON(int64x2_t &
#endif // FIXED_POINT_MATH_FUNCTIONS_USE_ACCUMULATOR_SATURATE #endif // FIXED_POINT_MATH_FUNCTIONS_USE_ACCUMULATOR_SATURATE
inoutAccum = vshrq_n_s64(inoutAccum, 12); inoutAccum = vshrq_n_s64(inoutAccum, 12);
inoutAccum = vreinterpretq_s64_s32( vuzp1q_s32(inoutAccum, vdupq_n_s32(0)) ); inoutAccum = vreinterpretq_s64_s32( vuzp1q_s32(vreinterpretq_s32_s64(inoutAccum), vdupq_n_s32(0)) );
} }
static FORCEINLINE s32 __vec4_dotproduct_vec4_fixed_NEON(const s32 (&__restrict vecA)[4], const s32 (&__restrict vecB)[4]) static FORCEINLINE s32 __vec4_dotproduct_vec4_fixed_NEON(const s32 (&__restrict vecA)[4], const s32 (&__restrict vecB)[4])
@ -1059,12 +1066,14 @@ static FORCEINLINE void __mtx4_multiply_mtx4_fixed_NEON(s32 (&__restrict mtxA)[1
outVecLo = vmlal_s32( outVecLo, vget_low_s32(rowA.val[1]), v[1] );\ outVecLo = vmlal_s32( outVecLo, vget_low_s32(rowA.val[1]), v[1] );\
outVecLo = vmlal_s32( outVecLo, vget_low_s32(rowA.val[2]), v[2] );\ outVecLo = vmlal_s32( outVecLo, vget_low_s32(rowA.val[2]), v[2] );\
outVecLo = vmlal_s32( outVecLo, vget_low_s32(rowA.val[3]), v[3] );\ outVecLo = vmlal_s32( outVecLo, vget_low_s32(rowA.val[3]), v[3] );\
___s32_saturate_shiftdown_accum64_fixed_NEON(outVecLo);\ outVecLo = vshrq_n_s64(outVecLo, 12);\
outVecLo = vreinterpretq_s64_s32( vuzp1q_s32(vreinterpretq_s32_s64(outVecLo), vdupq_n_s32(0)) );\
outVecHi = vmull_s32( vget_high_s32(rowA.val[0]), v[0] );\ outVecHi = vmull_s32( vget_high_s32(rowA.val[0]), v[0] );\
outVecHi = vmlal_s32( outVecHi, vget_high_s32(rowA.val[1]), v[1] );\ outVecHi = vmlal_s32( outVecHi, vget_high_s32(rowA.val[1]), v[1] );\
outVecHi = vmlal_s32( outVecHi, vget_high_s32(rowA.val[2]), v[2] );\ outVecHi = vmlal_s32( outVecHi, vget_high_s32(rowA.val[2]), v[2] );\
outVecHi = vmlal_s32( outVecHi, vget_high_s32(rowA.val[3]), v[3] );\ outVecHi = vmlal_s32( outVecHi, vget_high_s32(rowA.val[3]), v[3] );\
___s32_saturate_shiftdown_accum64_fixed_NEON(outVecHi); outVecHi = vshrq_n_s64(outVecHi, 12);\
outVecHi = vreinterpretq_s64_s32( vuzp1q_s32(vreinterpretq_s32_s64(outVecHi), vdupq_n_s32(0)) );
CALCULATE_MATRIX_ROW_FIXED_NEON(0); CALCULATE_MATRIX_ROW_FIXED_NEON(0);
vst1q_s32( mtxA + 0, vreinterpretq_s32_s64(vzip1q_s64(outVecLo, outVecHi)) ); vst1q_s32( mtxA + 0, vreinterpretq_s32_s64(vzip1q_s64(outVecLo, outVecHi)) );