/third_party/cmsis/CMSIS/DSP/Source/SupportFunctions/ |
D | arm_barycenter_f32.c | 62 float32_t *pOut; in arm_barycenter_f32() local 90 pOut = out; in arm_barycenter_f32() 99 outV = vld1q((const float32_t *) pOut); in arm_barycenter_f32() 108 vst1q(pOut, outV); in arm_barycenter_f32() 110 pOut += 4; in arm_barycenter_f32() 121 *pOut = *pOut + *pIn1++ * w1; in arm_barycenter_f32() 122 *pOut = *pOut + *pIn2++ * w2; in arm_barycenter_f32() 123 *pOut = *pOut + *pIn3++ * w3; in arm_barycenter_f32() 124 *pOut = *pOut + *pIn4++ * w4; in arm_barycenter_f32() 125 pOut++; in arm_barycenter_f32() [all …]
|
D | arm_barycenter_f16.c | 75 float16_t *pOut; in arm_barycenter_f16() local 103 pOut = out; in arm_barycenter_f16() 112 outV = vld1q((const float16_t *) pOut); in arm_barycenter_f16() 121 vst1q(pOut, outV); in arm_barycenter_f16() 123 pOut += 8; in arm_barycenter_f16() 134 *pOut = (_Float16)*pOut + (_Float16)*pIn1++ * (_Float16)w1; in arm_barycenter_f16() 135 *pOut = (_Float16)*pOut + (_Float16)*pIn2++ * (_Float16)w2; in arm_barycenter_f16() 136 *pOut = (_Float16)*pOut + (_Float16)*pIn3++ * (_Float16)w3; in arm_barycenter_f16() 137 *pOut = (_Float16)*pOut + (_Float16)*pIn4++ * (_Float16)w4; in arm_barycenter_f16() 138 pOut++; in arm_barycenter_f16() [all …]
|
/third_party/cmsis/CMSIS/DSP/Source/FilteringFunctions/ |
D | arm_fir_sparse_q31.c | 69 q31_t *pOut; /* Destination pointer */ in arm_fir_sparse_q31() local 105 pOut = pDst; in arm_fir_sparse_q31() 116 *pOut++ = (q31_t) (((q63_t) *px++ * coeff) >> 32); in arm_fir_sparse_q31() 118 *pOut++ = (q31_t) (((q63_t) *px++ * coeff) >> 32); in arm_fir_sparse_q31() 120 *pOut++ = (q31_t) (((q63_t) *px++ * coeff) >> 32); in arm_fir_sparse_q31() 122 *pOut++ = (q31_t) (((q63_t) *px++ * coeff) >> 32); in arm_fir_sparse_q31() 141 *pOut++ = (q31_t) (((q63_t) *px++ * coeff) >> 32); in arm_fir_sparse_q31() 176 pOut = pDst; in arm_fir_sparse_q31() 187 out = *pOut; in arm_fir_sparse_q31() 189 *pOut++ = (q31_t) (out); in arm_fir_sparse_q31() [all …]
|
D | arm_biquad_cascade_df1_f16.c | 61 float16_t *pOut = pDst; /* destination pointer */ in arm_biquad_cascade_df1_f16() local 136 vst1q(pOut, accVec); in arm_biquad_cascade_df1_f16() 137 pOut += 8; in arm_biquad_cascade_df1_f16() 211 *pOut++ = vgetq_lane(accVec, 0); in arm_biquad_cascade_df1_f16() 218 *pOut++ = vgetq_lane(accVec, 0); in arm_biquad_cascade_df1_f16() 219 *pOut++ = vgetq_lane(accVec, 1); in arm_biquad_cascade_df1_f16() 226 *pOut++ = vgetq_lane(accVec, 0); in arm_biquad_cascade_df1_f16() 227 *pOut++ = vgetq_lane(accVec, 1); in arm_biquad_cascade_df1_f16() 228 *pOut++ = vgetq_lane(accVec, 2); in arm_biquad_cascade_df1_f16() 236 *pOut++ = vgetq_lane(accVec, 0); in arm_biquad_cascade_df1_f16() [all …]
|
D | arm_fir_sparse_f32.c | 125 float32_t *pOut; /* Destination pointer */ in arm_fir_sparse_f32() local 158 pOut = pDst; in arm_fir_sparse_f32() 169 *pOut++ = *px++ * coeff; in arm_fir_sparse_f32() 171 *pOut++ = *px++ * coeff; in arm_fir_sparse_f32() 173 *pOut++ = *px++ * coeff; in arm_fir_sparse_f32() 175 *pOut++ = *px++ * coeff; in arm_fir_sparse_f32() 194 *pOut++ = *px++ * coeff; in arm_fir_sparse_f32() 229 pOut = pDst; in arm_fir_sparse_f32() 240 *pOut++ += *px++ * coeff; in arm_fir_sparse_f32() 242 *pOut++ += *px++ * coeff; in arm_fir_sparse_f32() [all …]
|
D | arm_biquad_cascade_df2T_f64.c | 157 float64_t *pOut = pDst; /* destination pointer */ in arm_biquad_cascade_df2T_f64() local 214 *pOut++ = acc0 ; in arm_biquad_cascade_df2T_f64() 245 *pOut++ = acc0; in arm_biquad_cascade_df2T_f64() 281 *pOut++ = acc0 ; in arm_biquad_cascade_df2T_f64() 294 pOut = pDst; in arm_biquad_cascade_df2T_f64() 311 float64_t *pOut = pDst; /* Destination pointer */ in arm_biquad_cascade_df2T_f64() local 358 *pOut++ = acc1; in arm_biquad_cascade_df2T_f64() 372 *pOut++ = acc1; in arm_biquad_cascade_df2T_f64() 385 *pOut++ = acc1; in arm_biquad_cascade_df2T_f64() 398 *pOut++ = acc1; in arm_biquad_cascade_df2T_f64() [all …]
|
D | arm_biquad_cascade_stereo_df2T_f32.c | 58 float32_t *pOut = pDst; /* destination pointer */ in arm_biquad_cascade_stereo_df2T_f32() local 135 *pOut++ = vgetq_lane(stateVec0, 0); in arm_biquad_cascade_stereo_df2T_f32() 136 *pOut++ = vgetq_lane(stateVec0, 1); in arm_biquad_cascade_stereo_df2T_f32() 172 pOut = pDst; in arm_biquad_cascade_stereo_df2T_f32() 190 float32_t *pOut = pDst; /* Destination pointer */ in arm_biquad_cascade_stereo_df2T_f32() local 233 *pOut++ = acc1a; in arm_biquad_cascade_stereo_df2T_f32() 234 *pOut++ = acc1b; in arm_biquad_cascade_stereo_df2T_f32() 249 *pOut++ = acc1a; in arm_biquad_cascade_stereo_df2T_f32() 250 *pOut++ = acc1b; in arm_biquad_cascade_stereo_df2T_f32() 265 *pOut++ = acc1a; in arm_biquad_cascade_stereo_df2T_f32() [all …]
|
/third_party/mesa3d/src/amd/addrlib/src/core/ |
D | addrlib2.cpp | 164 ADDR2_COMPUTE_SURFACE_INFO_OUTPUT* pOut ///< [out] output structure in ComputeSurfaceInfo() 172 (pOut->size != sizeof(ADDR2_COMPUTE_SURFACE_INFO_OUTPUT))) in ComputeSurfaceInfo() 251 returnCode = ComputeSurfaceInfoLinear(&localIn, pOut); in ComputeSurfaceInfo() 256 returnCode = ComputeSurfaceInfoTiled(&localIn, pOut); in ComputeSurfaceInfo() 261 pOut->bpp = localIn.bpp; in ComputeSurfaceInfo() 262 pOut->pixelPitch = pOut->pitch; in ComputeSurfaceInfo() 263 pOut->pixelHeight = pOut->height; in ComputeSurfaceInfo() 264 pOut->pixelMipChainPitch = pOut->mipChainPitch; in ComputeSurfaceInfo() 265 pOut->pixelMipChainHeight = pOut->mipChainHeight; in ComputeSurfaceInfo() 266 pOut->pixelBits = localIn.bpp; in ComputeSurfaceInfo() [all …]
|
D | addrlib1.cpp | 168 ADDR_COMPUTE_SURFACE_INFO_OUTPUT* pOut ///< [out] output structure in ComputeSurfaceInfo() 176 (pOut->size != sizeof(ADDR_COMPUTE_SURFACE_INFO_OUTPUT))) in ComputeSurfaceInfo() 225 pOut->height = pIn->height; in ComputeSurfaceInfo() 233 pOut->pixelBits = localIn.bpp; in ComputeSurfaceInfo() 234 pOut->numSamples = localIn.numSamples; in ComputeSurfaceInfo() 235 pOut->last2DLevel = FALSE; in ComputeSurfaceInfo() 236 pOut->tcCompatible = FALSE; in ComputeSurfaceInfo() 291 returnCode = PostComputeMipLevel(&localIn, pOut); in ComputeSurfaceInfo() 331 pOut->macroModeIndex = macroModeIndex; in ComputeSurfaceInfo() 357 returnCode = HwlComputeSurfaceInfo(&localIn, pOut); in ComputeSurfaceInfo() [all …]
|
D | addrlib1.h | 107 ADDR_COMPUTE_SURFACE_INFO_OUTPUT* pOut) const; 111 ADDR_COMPUTE_SURFACE_ADDRFROMCOORD_OUTPUT* pOut) const; 115 ADDR_COMPUTE_SURFACE_COORDFROMADDR_OUTPUT* pOut) const; 119 ADDR_COMPUTE_SLICESWIZZLE_OUTPUT* pOut) const; 123 ADDR_EXTRACT_BANKPIPE_SWIZZLE_OUTPUT* pOut) const; 127 ADDR_COMBINE_BANKPIPE_SWIZZLE_OUTPUT* pOut) const; 131 ADDR_COMPUTE_BASE_SWIZZLE_OUTPUT* pOut) const; 135 ADDR_COMPUTE_FMASK_INFO_OUTPUT* pOut); 139 ADDR_COMPUTE_FMASK_ADDRFROMCOORD_OUTPUT* pOut) const; 143 ADDR_COMPUTE_FMASK_COORDFROMADDR_OUTPUT* pOut) const; [all …]
|
D | addrlib2.h | 232 ADDR2_COMPUTE_SURFACE_INFO_OUTPUT* pOut) const; 236 ADDR2_COMPUTE_SURFACE_ADDRFROMCOORD_OUTPUT* pOut) const; 240 ADDR2_COMPUTE_SURFACE_COORDFROMADDR_OUTPUT* pOut) const; 245 ADDR2_COMPUTE_HTILE_INFO_OUTPUT* pOut) const; 249 ADDR2_COMPUTE_HTILE_ADDRFROMCOORD_OUTPUT* pOut); 253 ADDR2_COMPUTE_HTILE_COORDFROMADDR_OUTPUT* pOut); 258 ADDR2_COMPUTE_CMASK_INFO_OUTPUT* pOut) const; 262 ADDR2_COMPUTE_CMASK_ADDRFROMCOORD_OUTPUT* pOut); 266 ADDR2_COMPUTE_CMASK_COORDFROMADDR_OUTPUT* pOut) const; 271 ADDR2_COMPUTE_FMASK_INFO_OUTPUT* pOut); [all …]
|
/third_party/mesa3d/src/amd/addrlib/src/ |
D | addrinterface.cpp | 118 ADDR_COMPUTE_SURFACE_INFO_OUTPUT* pOut) ///< [out] surface parameters and alignments in AddrComputeSurfaceInfo() argument 126 returnCode = pLib->ComputeSurfaceInfo(pIn, pOut); in AddrComputeSurfaceInfo() 152 ADDR_COMPUTE_SURFACE_ADDRFROMCOORD_OUTPUT* pOut) ///< [out] surface address in AddrComputeSurfaceAddrFromCoord() argument 160 returnCode = pLib->ComputeSurfaceAddrFromCoord(pIn, pOut); in AddrComputeSurfaceAddrFromCoord() 184 ADDR_COMPUTE_SURFACE_COORDFROMADDR_OUTPUT* pOut) ///< [out] coordinates in AddrComputeSurfaceCoordFromAddr() argument 192 returnCode = pLib->ComputeSurfaceCoordFromAddr(pIn, pOut); in AddrComputeSurfaceCoordFromAddr() 222 ADDR_COMPUTE_HTILE_INFO_OUTPUT* pOut) ///< [out] Htile pitch, height and size in bytes in AddrComputeHtileInfo() argument 230 returnCode = pLib->ComputeHtileInfo(pIn, pOut); in AddrComputeHtileInfo() 254 ADDR_COMPUTE_HTILE_ADDRFROMCOORD_OUTPUT* pOut) ///< [out] Htile address in AddrComputeHtileAddrFromCoord() argument 262 returnCode = pLib->ComputeHtileAddrFromCoord(pIn, pOut); in AddrComputeHtileAddrFromCoord() [all …]
|
/third_party/mesa3d/src/amd/addrlib/src/gfx10/ |
D | gfx10addrlib.cpp | 162 ADDR2_COMPUTE_HTILE_INFO_OUTPUT* pOut ///< [out] output structure in HwlComputeHtileInfo() 184 pOut->pitch = PowTwoAlign(pIn->unalignedWidth, metaBlk.w); in HwlComputeHtileInfo() 185 pOut->height = PowTwoAlign(pIn->unalignedHeight, metaBlk.h); in HwlComputeHtileInfo() 186 pOut->baseAlign = Max(metaBlkSize, 1u << (m_pipesLog2 + 11u)); in HwlComputeHtileInfo() 187 pOut->metaBlkWidth = metaBlk.w; in HwlComputeHtileInfo() 188 pOut->metaBlkHeight = metaBlk.h; in HwlComputeHtileInfo() 209 if (pOut->pMipInfo != NULL) in HwlComputeHtileInfo() 211 pOut->pMipInfo[i].inMiptail = FALSE; in HwlComputeHtileInfo() 212 pOut->pMipInfo[i].offset = offset; in HwlComputeHtileInfo() 213 pOut->pMipInfo[i].sliceSize = mipSliceSize; in HwlComputeHtileInfo() [all …]
|
/third_party/mesa3d/src/amd/addrlib/src/gfx11/ |
D | gfx11addrlib.cpp | 163 ADDR2_COMPUTE_HTILE_INFO_OUTPUT* pOut ///< [out] output structure in HwlComputeHtileInfo() 185 pOut->pitch = PowTwoAlign(pIn->unalignedWidth, metaBlk.w); in HwlComputeHtileInfo() 186 pOut->height = PowTwoAlign(pIn->unalignedHeight, metaBlk.h); in HwlComputeHtileInfo() 187 pOut->baseAlign = Max(metaBlkSize, 1u << (m_pipesLog2 + 11u)); in HwlComputeHtileInfo() 188 pOut->metaBlkWidth = metaBlk.w; in HwlComputeHtileInfo() 189 pOut->metaBlkHeight = metaBlk.h; in HwlComputeHtileInfo() 210 if (pOut->pMipInfo != NULL) in HwlComputeHtileInfo() 212 pOut->pMipInfo[i].inMiptail = FALSE; in HwlComputeHtileInfo() 213 pOut->pMipInfo[i].offset = offset; in HwlComputeHtileInfo() 214 pOut->pMipInfo[i].sliceSize = mipSliceSize; in HwlComputeHtileInfo() [all …]
|
/third_party/mesa3d/src/amd/addrlib/src/r800/ |
D | ciaddrlib.cpp | 207 ADDR_COMPUTE_DCCINFO_OUTPUT* pOut) const in HwlComputeDccInfo() 240 pOut->dccRamSize = pIn->colorSurfSize >> 8; in HwlComputeDccInfo() 241 pOut->dccRamBaseAlign = pIn->tileInfo.banks * in HwlComputeDccInfo() 244 pOut->dccFastClearSize = dccFastClearSize; in HwlComputeDccInfo() 245 pOut->dccRamSizeAligned = TRUE; in HwlComputeDccInfo() 247 ADDR_ASSERT(IsPow2(pOut->dccRamBaseAlign)); in HwlComputeDccInfo() 249 if (0 == (pOut->dccRamSize & (pOut->dccRamBaseAlign - 1))) in HwlComputeDccInfo() 251 pOut->subLvlCompressible = TRUE; in HwlComputeDccInfo() 257 if (pOut->dccRamSize == pOut->dccFastClearSize) in HwlComputeDccInfo() 259 pOut->dccFastClearSize = PowTwoAlign(pOut->dccRamSize, dccRamSizeAlign); in HwlComputeDccInfo() [all …]
|
D | egbaddrlib.cpp | 86 ADDR_COMPUTE_SURFACE_INFO_OUTPUT* pOut ///< [out] output structure in DispatchComputeSurfaceInfo() 119 pOut->numSamples = numSamples; in DispatchComputeSurfaceInfo() 123 ADDR_ASSERT(pOut->pTileInfo); in DispatchComputeSurfaceInfo() 125 if (pOut->pTileInfo != NULL) in DispatchComputeSurfaceInfo() 127 pTileInfo = pOut->pTileInfo; in DispatchComputeSurfaceInfo() 153 pOut); in DispatchComputeSurfaceInfo() 173 valid = ComputeSurfaceInfoLinear(pIn, pOut, padDims); in DispatchComputeSurfaceInfo() 178 valid = ComputeSurfaceInfoMicroTiled(pIn, pOut, padDims, tileMode); in DispatchComputeSurfaceInfo() 193 valid = ComputeSurfaceInfoMacroTiled(pIn, pOut, padDims, tileMode); in DispatchComputeSurfaceInfo() 220 ADDR_COMPUTE_SURFACE_INFO_OUTPUT* pOut, ///< [out] Output structure in ComputeSurfaceInfoLinear() argument [all …]
|
/third_party/cmsis/CMSIS/DSP/Source/MatrixFunctions/ |
D | arm_mat_trans_q15.c | 104 q15_t *pOut = pDst->pData; /* output data matrix pointer */ in arm_mat_trans_q15() local 133 pOut = pDst->pData + i; in arm_mat_trans_q15() 147 *pOut = (q15_t) in; in arm_mat_trans_q15() 149 *pOut = (q15_t) ((in & (q31_t) 0xffff0000) >> 16); in arm_mat_trans_q15() 153 pOut += nRows; in arm_mat_trans_q15() 157 *pOut = (q15_t) ((in & (q31_t) 0xffff0000) >> 16); in arm_mat_trans_q15() 159 *pOut = (q15_t) in; in arm_mat_trans_q15() 163 pOut += nRows; in arm_mat_trans_q15() 170 *pOut = (q15_t) in; in arm_mat_trans_q15() 172 *pOut = (q15_t) ((in & (q31_t) 0xffff0000) >> 16); in arm_mat_trans_q15() [all …]
|
D | arm_mat_cmplx_mult_q31.c | 74 q31_t *pOut = pDst->pData; /* output data matrix pointer */ in arm_mat_cmplx_mult_q31_2x2_mve() local 99 pOut[0 * CMPLX_DIM * MATRIX_DIM2 + 0] = (q31_t) asrl(acc0, 31); in arm_mat_cmplx_mult_q31_2x2_mve() 100 pOut[0 * CMPLX_DIM * MATRIX_DIM2 + 1] = (q31_t) asrl(acc1, 31); in arm_mat_cmplx_mult_q31_2x2_mve() 101 pOut[1 * CMPLX_DIM * MATRIX_DIM2 + 0] = (q31_t) asrl(acc2, 31); in arm_mat_cmplx_mult_q31_2x2_mve() 102 pOut[1 * CMPLX_DIM * MATRIX_DIM2 + 1] = (q31_t) asrl(acc3, 31); in arm_mat_cmplx_mult_q31_2x2_mve() 117 pOut += CMPLX_DIM; in arm_mat_cmplx_mult_q31_2x2_mve() 119 pOut[0 * CMPLX_DIM * MATRIX_DIM2 + 0] = (q31_t) asrl(acc0, 31); in arm_mat_cmplx_mult_q31_2x2_mve() 120 pOut[0 * CMPLX_DIM * MATRIX_DIM2 + 1] = (q31_t) asrl(acc1, 31); in arm_mat_cmplx_mult_q31_2x2_mve() 121 pOut[1 * CMPLX_DIM * MATRIX_DIM2 + 0] = (q31_t) asrl(acc2, 31); in arm_mat_cmplx_mult_q31_2x2_mve() 122 pOut[1 * CMPLX_DIM * MATRIX_DIM2 + 1] = (q31_t) asrl(acc3, 31); in arm_mat_cmplx_mult_q31_2x2_mve() [all …]
|
D | arm_mat_mult_q31.c | 74 q31_t *pOut = pDst->pData; /* output data matrix pointer */ in arm_mat_mult_q31_2x2_mve() local 101 pOut[0 * MATRIX_DIM2] = (q31_t) acc0; in arm_mat_mult_q31_2x2_mve() 102 pOut[1 * MATRIX_DIM2] = (q31_t) acc1; in arm_mat_mult_q31_2x2_mve() 103 pOut++; in arm_mat_mult_q31_2x2_mve() 116 pOut[0 * MATRIX_DIM2] = (q31_t) acc0; in arm_mat_mult_q31_2x2_mve() 117 pOut[1 * MATRIX_DIM2] = (q31_t) acc1; in arm_mat_mult_q31_2x2_mve() 133 q31_t *pOut = pDst->pData; /* output data matrix pointer */ in arm_mat_mult_q31_3x3_mve() local 161 pOut[0 * MATRIX_DIM3] = (q31_t) acc0; in arm_mat_mult_q31_3x3_mve() 162 pOut[1 * MATRIX_DIM3] = (q31_t) acc1; in arm_mat_mult_q31_3x3_mve() 163 pOut[2 * MATRIX_DIM3] = (q31_t) acc2; in arm_mat_mult_q31_3x3_mve() [all …]
|
D | arm_mat_mult_q7.c | 67 q7_t *pOut = pDst->pData; /* output data matrix pointer */ in arm_mat_mult_q7_2x2_mve() local 87 pOut[0 * MATRIX_DIM] = (q7_t) __SSAT(acc0 >> 7, 8); in arm_mat_mult_q7_2x2_mve() 88 pOut[1 * MATRIX_DIM] = (q7_t) __SSAT(acc1 >> 7, 8); in arm_mat_mult_q7_2x2_mve() 89 pOut++; in arm_mat_mult_q7_2x2_mve() 99 pOut[0 * MATRIX_DIM] = (q7_t) __SSAT(acc0 >> 7, 8); in arm_mat_mult_q7_2x2_mve() 100 pOut[1 * MATRIX_DIM] = (q7_t) __SSAT(acc1 >> 7, 8); in arm_mat_mult_q7_2x2_mve() 116 q7_t *pOut = pDst->pData; /* output data matrix pointer */ in arm_mat_mult_q7_3x3_mve() local 140 pOut[0 * MATRIX_DIM] = (q7_t) __SSAT(acc0 >> 7, 8); in arm_mat_mult_q7_3x3_mve() 141 pOut[1 * MATRIX_DIM] = (q7_t) __SSAT(acc1 >> 7, 8); in arm_mat_mult_q7_3x3_mve() 142 pOut[2 * MATRIX_DIM] = (q7_t) __SSAT(acc2 >> 7, 8); in arm_mat_mult_q7_3x3_mve() [all …]
|
D | arm_mat_cmplx_mult_f32.c | 78 float32_t *pOut = pDst->pData; /* output data matrix pointer */ in arm_mat_cmplx_mult_f32_2x2_mve() local 103 pOut[0 * CMPLX_DIM * MATRIX_DIM2 + 0] = acc0[0] + acc0[2]; in arm_mat_cmplx_mult_f32_2x2_mve() 104 pOut[0 * CMPLX_DIM * MATRIX_DIM2 + 1] = acc0[1] + acc0[3]; in arm_mat_cmplx_mult_f32_2x2_mve() 105 pOut[1 * CMPLX_DIM * MATRIX_DIM2 + 0] = acc1[0] + acc1[2]; in arm_mat_cmplx_mult_f32_2x2_mve() 106 pOut[1 * CMPLX_DIM * MATRIX_DIM2 + 1] = acc1[1] + acc1[3]; in arm_mat_cmplx_mult_f32_2x2_mve() 107 pOut += CMPLX_DIM; in arm_mat_cmplx_mult_f32_2x2_mve() 124 pOut[0 * CMPLX_DIM * MATRIX_DIM2 + 0] = acc0[0] + acc0[2]; in arm_mat_cmplx_mult_f32_2x2_mve() 125 pOut[0 * CMPLX_DIM * MATRIX_DIM2 + 1] = acc0[1] + acc0[3]; in arm_mat_cmplx_mult_f32_2x2_mve() 126 pOut[1 * CMPLX_DIM * MATRIX_DIM2 + 0] = acc1[0] + acc1[2]; in arm_mat_cmplx_mult_f32_2x2_mve() 127 pOut[1 * CMPLX_DIM * MATRIX_DIM2 + 1] = acc1[1] + acc1[3]; in arm_mat_cmplx_mult_f32_2x2_mve() [all …]
|
D | arm_mat_scale_f32.c | 100 float32_t *pOut = pDst->pData; /* output data matrix pointer */ in arm_mat_scale_f32() local 123 vst1q(pOut, vecOut); in arm_mat_scale_f32() 124 pOut += 4; in arm_mat_scale_f32() 140 vstrwq_p(pOut, vecOut, p0); in arm_mat_scale_f32() 158 float32_t *pOut = pDst->pData; /* output data matrix pointer */ in arm_mat_scale_f32() local 190 vst1q_f32(pOut, res); in arm_mat_scale_f32() 194 pOut += 4U; in arm_mat_scale_f32() 208 *pOut++ = (*pIn++) * scale; in arm_mat_scale_f32() 228 float32_t *pOut = pDst->pData; /* Output data matrix pointer */ in arm_mat_scale_f32() local 260 *pOut++ = (*pIn++) * scale; in arm_mat_scale_f32() [all …]
|
/third_party/mesa3d/src/amd/addrlib/src/gfx9/ |
D | gfx9addrlib.cpp | 162 ADDR2_COMPUTE_HTILE_INFO_OUTPUT* pOut ///< [out] output structure in HwlComputeHtileInfo() 218 GetMetaMipInfo(pIn->numMipLevels, &metaBlkDim, FALSE, pOut->pMipInfo, in HwlComputeHtileInfo() 248 pOut->pitch = numMetaBlkX * metaBlkDim.w; in HwlComputeHtileInfo() 249 pOut->height = numMetaBlkY * metaBlkDim.h; in HwlComputeHtileInfo() 250 pOut->sliceSize = numMetaBlkX * numMetaBlkY * metaBlkSize; in HwlComputeHtileInfo() 252 pOut->metaBlkWidth = metaBlkDim.w; in HwlComputeHtileInfo() 253 pOut->metaBlkHeight = metaBlkDim.h; in HwlComputeHtileInfo() 254 pOut->metaBlkNumPerSlice = numMetaBlkX * numMetaBlkY; in HwlComputeHtileInfo() 256 pOut->baseAlign = align; in HwlComputeHtileInfo() 257 pOut->htileBytes = PowTwoAlign(pOut->sliceSize * numMetaBlkZ, align); in HwlComputeHtileInfo() [all …]
|
/third_party/cmsis/CMSIS/DSP/Include/dsp/ |
D | fast_math_functions.h | 242 float32_t * pOut) in arm_sqrt_f32() argument 248 *pOut = __sqrtf(in); in arm_sqrt_f32() 250 *pOut = sqrtf(in); in arm_sqrt_f32() 255 __ASM("VSQRT.F32 %0,%1" : "=t"(*pOut) : "t"(in)); in arm_sqrt_f32() 257 *pOut = sqrtf(in); in arm_sqrt_f32() 261 *pOut = _sqrtf(in); in arm_sqrt_f32() 263 *pOut = sqrtf(in); in arm_sqrt_f32() 266 __ASM("VSQRT.F32 %0,%1" : "=t"(*pOut) : "t"(in)); in arm_sqrt_f32() 268 *pOut = sqrtf(in); in arm_sqrt_f32() 271 *pOut = sqrtf(in); in arm_sqrt_f32() [all …]
|
/third_party/cmsis/CMSIS/DSP/Source/TransformFunctions/ |
D | arm_rfft_fast_f32.c | 35 float32_t * pOut) in stage_rfft_f32() argument 77 *pOut++ = 0.5f * ( t1a + t1b ); in stage_rfft_f32() 78 *pOut++ = 0.5f * ( t1a - t1b ); in stage_rfft_f32() 138 vst2q_f32(pOut, res); in stage_rfft_f32() 139 pOut += 8; in stage_rfft_f32() 181 *pOut++ = 0.5f * (xAR + xBR + p0 + p3 ); //xAR in stage_rfft_f32() 182 *pOut++ = 0.5f * (xAI - xBI + p1 - p2 ); //xAI in stage_rfft_f32() 194 float32_t * pOut) in merge_rfft_f32() argument 221 *pOut++ = 0.5f * ( xAR + xAI ); in merge_rfft_f32() 222 *pOut++ = 0.5f * ( xAR - xAI ); in merge_rfft_f32() [all …]
|