| /cmsis-dsp-latest/PrivateInclude/ |
| D | arm_vec_filtering.h | 42 uint32_t k; \
50 k = count >> 2; \
52 while (k > 0U) \
66 k--; \
69 k = count % 0x4U; \
70 if (k > 0U) \
72 mve_pred16_t p0 = vctp32q(k); \
96 uint32_t k; \
101 k = count >> 2; \
103 while (k > 0U) \
[all …]
|
| /cmsis-dsp-latest/Source/MatrixFunctions/ |
| D | arm_mat_ldlt_f32.c | 87 int fullRank = 1, diag,k; in arm_mat_ldlt_f32() local
97 for(int k=0;k < n; k+=8) in arm_mat_ldlt_f32() local
102 vecP = vidupq_u16((uint16_t)k, 1); in arm_mat_ldlt_f32()
104 vstrhq_p(&pp[k], vecP, p0); in arm_mat_ldlt_f32()
110 for(k=0;k < n; k++) in arm_mat_ldlt_f32()
114 int j=k; in arm_mat_ldlt_f32()
117 for(int r=k;r<n;r++) in arm_mat_ldlt_f32()
126 if(j != k) in arm_mat_ldlt_f32()
128 SWAP_ROWS_F32(pl,0,k,j); in arm_mat_ldlt_f32()
129 SWAP_COLS_F32(pl,0,k,j); in arm_mat_ldlt_f32()
[all …]
|
| D | arm_mat_ldlt_f64.c | 87 int fullRank = 1, diag,k; in arm_mat_ldlt_f64() local
95 for(k=0;k < n; k++) in arm_mat_ldlt_f64()
97 pp[k] = k; in arm_mat_ldlt_f64()
101 for(k=0;k < n; k++) in arm_mat_ldlt_f64()
105 int r,j=k; in arm_mat_ldlt_f64()
108 for(r=k;r<n;r++) in arm_mat_ldlt_f64()
117 if(j != k) in arm_mat_ldlt_f64()
119 SWAP_ROWS_F64(pl,0,k,j); in arm_mat_ldlt_f64()
120 SWAP_COLS_F64(pl,0,k,j); in arm_mat_ldlt_f64()
124 pp[k] = j; in arm_mat_ldlt_f64()
[all …]
|
| D | arm_mat_cholesky_f64.c | 84 int i,j,k; in arm_mat_cholesky_f64() local
118 k=0; in arm_mat_cholesky_f64()
122 vecGi=vld1q_f64(&pG[i * n + k]); in arm_mat_cholesky_f64()
124 vecGj0=vld1q_f64(&pG[(j + 0) * n + k]); in arm_mat_cholesky_f64()
125 vecGj1=vld1q_f64(&pG[(j + 1) * n + k]); in arm_mat_cholesky_f64()
126 vecGj2=vld1q_f64(&pG[(j + 2) * n + k]); in arm_mat_cholesky_f64()
127 vecGj3=vld1q_f64(&pG[(j + 3) * n + k]); in arm_mat_cholesky_f64()
135 k+=2; in arm_mat_cholesky_f64()
149 sum0 = sum0 + pG[i * n + k] * pG[(j + 0) * n + k]; in arm_mat_cholesky_f64()
150 sum1 = sum1 + pG[i * n + k] * pG[(j + 1) * n + k]; in arm_mat_cholesky_f64()
[all …]
|
| D | arm_mat_cholesky_f32.c | 96 int i,j,k; in arm_mat_cholesky_f32() local
127 for(k=0; k < i ; k+=4) in arm_mat_cholesky_f32()
131 vecGi=vldrwq_z_f32(&pG[i * n + k],p0); in arm_mat_cholesky_f32()
133 vecGj0=vldrwq_z_f32(&pG[(j + 0) * n + k],p0); in arm_mat_cholesky_f32()
134 vecGj1=vldrwq_z_f32(&pG[(j + 1) * n + k],p0); in arm_mat_cholesky_f32()
135 vecGj2=vldrwq_z_f32(&pG[(j + 2) * n + k],p0); in arm_mat_cholesky_f32()
136 vecGj3=vldrwq_z_f32(&pG[(j + 3) * n + k],p0); in arm_mat_cholesky_f32()
158 for(k=0; k < i ; k+=4) in arm_mat_cholesky_f32()
162 vecGi=vldrwq_z_f32(&pG[i * n + k],p0); in arm_mat_cholesky_f32()
163 vecGj=vldrwq_z_f32(&pG[j * n + k],p0); in arm_mat_cholesky_f32()
[all …]
|
| D | arm_mat_solve_upper_triangular_f32.c | 77 int i,j,k,n,cols; in arm_mat_solve_upper_triangular_f32() local
100 for(k=n-1; k > i; k--) in arm_mat_solve_upper_triangular_f32()
102 vecX = vld1q_f32(&pX[cols*k+j]); in arm_mat_solve_upper_triangular_f32()
103 vecA = vfmsq(vecA,vdupq_n_f32(pUT[n*i + k]),vecX); in arm_mat_solve_upper_triangular_f32()
126 for(k=n-1; k > i; k--) in arm_mat_solve_upper_triangular_f32()
128 tmp -= ut_row[k] * pX[cols*k+j]; in arm_mat_solve_upper_triangular_f32()
174 int i,j,k,n,cols; in arm_mat_solve_upper_triangular_f32() local
197 for(k=n-1; k > i; k--) in arm_mat_solve_upper_triangular_f32()
199 vecX = vld1q_f32(&pX[cols*k+j]); in arm_mat_solve_upper_triangular_f32()
200 vecA = vfmsq_f32(vecA,vdupq_n_f32(pUT[n*i + k]),vecX); in arm_mat_solve_upper_triangular_f32()
[all …]
|
| D | arm_mat_solve_lower_triangular_f32.c | 84 int i,j,k,n,cols; in arm_mat_solve_lower_triangular_f32() local
108 for(k=0; k < i; k++) in arm_mat_solve_lower_triangular_f32()
110 vecX = vld1q_f32(&pX[cols*k+j]); in arm_mat_solve_lower_triangular_f32()
111 vecA = vfmsq(vecA,vdupq_n_f32(pLT[n*i + k]),vecX); in arm_mat_solve_lower_triangular_f32()
132 for(k=0; k < i; k++) in arm_mat_solve_lower_triangular_f32()
134 tmp -= lt_row[k] * pX[cols*k+j]; in arm_mat_solve_lower_triangular_f32()
185 int i,j,k,n,cols; in arm_mat_solve_lower_triangular_f32() local
209 for(k=0; k < i; k++) in arm_mat_solve_lower_triangular_f32()
211 vecX = vld1q_f32(&pX[cols*k+j]); in arm_mat_solve_lower_triangular_f32()
212 vecA = vfmsq_f32(vecA,vdupq_n_f32(pLT[n*i + k]),vecX); in arm_mat_solve_lower_triangular_f32()
[all …]
|
| D | arm_mat_solve_upper_triangular_f64.c | 74 int i,j,k,n,cols; in arm_mat_solve_upper_triangular_f64() local
97 for(k=n-1; k > i; k--) in arm_mat_solve_upper_triangular_f64()
99 vecX = vld1q_f64(&pX[cols*k+j]); in arm_mat_solve_upper_triangular_f64()
100 vecA = vfmsq_f64(vecA,vdupq_n_f64(pUT[n*i + k]),vecX); in arm_mat_solve_upper_triangular_f64()
123 for(k=n-1; k > i; k--) in arm_mat_solve_upper_triangular_f64()
125 tmp -= ut_row[k] * pX[cols*k+j]; in arm_mat_solve_upper_triangular_f64()
170 int i,j,k,n,cols; in arm_mat_solve_upper_triangular_f64() local
192 for(k=n-1; k > i; k--) in arm_mat_solve_upper_triangular_f64()
194 tmp -= ut_row[k] * pX[cols*k+j]; in arm_mat_solve_upper_triangular_f64()
|
| D | arm_mat_solve_lower_triangular_f16.c | 85 int i,j,k,n,cols; in arm_mat_solve_lower_triangular_f16() local
109 for(k=0; k < i; k++) in arm_mat_solve_lower_triangular_f16()
111 vecX = vld1q_f16(&pX[cols*k+j]); in arm_mat_solve_lower_triangular_f16()
112 vecA = vfmsq(vecA,vdupq_n_f16(pLT[n*i + k]),vecX); in arm_mat_solve_lower_triangular_f16()
133 for(k=0; k < i; k++) in arm_mat_solve_lower_triangular_f16()
135 tmp -= (_Float16)lt_row[k] * (_Float16)pX[cols*k+j]; in arm_mat_solve_lower_triangular_f16()
186 int i,j,k,n,cols; in arm_mat_solve_lower_triangular_f16() local
208 for(k=0; k < i; k++) in arm_mat_solve_lower_triangular_f16()
210 tmp -= (_Float16)lt_row[k] * (_Float16)pX[cols*k+j]; in arm_mat_solve_lower_triangular_f16()
|
| D | arm_mat_solve_lower_triangular_f64.c | 81 int i,j,k,n,cols; in arm_mat_solve_lower_triangular_f64() local
105 for(k=0; k < i; k++) in arm_mat_solve_lower_triangular_f64()
107 vecX = vld1q_f64(&pX[cols*k+j]); in arm_mat_solve_lower_triangular_f64()
108 vecA = vfmsq_f64(vecA,vdupq_n_f64(pLT[n*i + k]),vecX); in arm_mat_solve_lower_triangular_f64()
129 for(k=0; k < i; k++) in arm_mat_solve_lower_triangular_f64()
131 tmp -= lt_row[k] * pX[cols*k+j]; in arm_mat_solve_lower_triangular_f64()
182 int i,j,k,n,cols; in arm_mat_solve_lower_triangular_f64() local
204 for(k=0; k < i; k++) in arm_mat_solve_lower_triangular_f64()
206 tmp -= lt_row[k] * pX[cols*k+j]; in arm_mat_solve_lower_triangular_f64()
|
| /cmsis-dsp-latest/Source/FilteringFunctions/ |
| D | arm_correlate_f64.c | 64 uint32_t j, k, count, blkCnt; /* Loop counters */ in arm_correlate_f64() local
171 k = count >> 1U ; in arm_correlate_f64()
173 while(k > 0U) in arm_correlate_f64()
180 k--; in arm_correlate_f64()
183 k = count & 1 ; in arm_correlate_f64()
186 k = count; in arm_correlate_f64()
188 while (k > 0U) in arm_correlate_f64()
195 k--; in arm_correlate_f64()
253 k = srcBLen >> 1U ; in arm_correlate_f64()
254 while(k > 0U) in arm_correlate_f64()
[all …]
|
| D | arm_correlate_f16.c | 61 … uint32_t k; \
67 … k = count >> 3; \
68 … while (k > 0U) \
76 … k--; \
78 … k = count % 0x8U; \
82 mve_pred16_t p0 = vctp16q(k+1); \
88 p0 = vctp16q(k); \
100 … uint16_t k; \
105 … k = count >> 3; \
107 … while (k > 0U) \
[all …]
|
| D | arm_conv_fast_q15.c | 78 uint32_t j, k, count, blkCnt; /* Loop counters */ in arm_conv_fast_q15() local
157 k = count; in arm_conv_fast_q15()
159 while (k > 0U) in arm_conv_fast_q15()
165 k--; in arm_conv_fast_q15()
194 k = count >> 2U; in arm_conv_fast_q15()
198 while (k > 0U) in arm_conv_fast_q15()
207 k--; in arm_conv_fast_q15()
216 k = count % 0x4U; in arm_conv_fast_q15()
218 while (k > 0U) in arm_conv_fast_q15()
224 k--; in arm_conv_fast_q15()
[all …]
|
| D | arm_conv_partial_f32.c | 103 uint32_t j, k, count, blkCnt, check; in arm_conv_partial_f32() local
204 k = count >> 2U; in arm_conv_partial_f32()
206 while (k > 0U) in arm_conv_partial_f32()
221 k--; in arm_conv_partial_f32()
225 k = count % 0x4U; in arm_conv_partial_f32()
230 k = count; in arm_conv_partial_f32()
234 while (k > 0U) in arm_conv_partial_f32()
240 k--; in arm_conv_partial_f32()
313 k = srcBLen >> 2U; in arm_conv_partial_f32()
379 } while (--k); in arm_conv_partial_f32()
[all …]
|
| D | arm_conv_partial_q31.c | 76 uint32_t j, k, count, blkCnt, check; in arm_conv_partial_q31() local
177 k = count >> 2U; in arm_conv_partial_q31()
179 while (k > 0U) in arm_conv_partial_q31()
194 k--; in arm_conv_partial_q31()
198 k = count % 0x4U; in arm_conv_partial_q31()
203 k = count; in arm_conv_partial_q31()
207 while (k > 0U) in arm_conv_partial_q31()
213 k--; in arm_conv_partial_q31()
284 k = srcBLen / 3; in arm_conv_partial_q31()
337 } while (--k); in arm_conv_partial_q31()
[all …]
|
| D | arm_conv_partial_fast_q15.c | 73 uint32_t j, k, count, blkCnt, check; in arm_conv_partial_fast_q15() local
172 k = count; in arm_conv_partial_fast_q15()
174 while (k > 0U) in arm_conv_partial_fast_q15()
180 k--; in arm_conv_partial_fast_q15()
209 k = count >> 2U; in arm_conv_partial_fast_q15()
213 while (k > 0U) in arm_conv_partial_fast_q15()
222 k--; in arm_conv_partial_fast_q15()
231 k = count % 0x4U; in arm_conv_partial_fast_q15()
233 while (k > 0U) in arm_conv_partial_fast_q15()
239 k--; in arm_conv_partial_fast_q15()
[all …]
|
| D | arm_conv_f32.c | 271 uint32_t j, k, count, blkCnt; /* Loop counters */ in arm_conv_f32() local
360 k = count >> 2U; in arm_conv_f32()
367 k = count >> 2U; in arm_conv_f32()
372 while (k > 0U) in arm_conv_f32()
386 k--; in arm_conv_f32()
394 k = count & 3; in arm_conv_f32()
396 while (k > 0U) in arm_conv_f32()
411 k--; in arm_conv_f32()
415 k = count % 0x4U; in arm_conv_f32()
421 k = count; in arm_conv_f32()
[all …]
|
| D | arm_conv_partial_fast_q31.c | 73 uint32_t j, k, count, check, blkCnt; in arm_conv_partial_fast_q31() local
174 k = count >> 2U; in arm_conv_partial_fast_q31()
176 while (k > 0U) in arm_conv_partial_fast_q31()
195 k--; in arm_conv_partial_fast_q31()
199 k = count % 0x4U; in arm_conv_partial_fast_q31()
204 k = count; in arm_conv_partial_fast_q31()
208 while (k > 0U) in arm_conv_partial_fast_q31()
215 k--; in arm_conv_partial_fast_q31()
288 k = srcBLen >> 2U; in arm_conv_partial_fast_q31()
354 } while (--k); in arm_conv_partial_fast_q31()
[all …]
|
| D | arm_conv_fast_q31.c | 78 uint32_t j, k, count, blkCnt; /* Loop counters */ in arm_conv_fast_q31() local
152 k = count >> 2U; in arm_conv_fast_q31()
156 while (k > 0U) in arm_conv_fast_q31()
175 k--; in arm_conv_fast_q31()
180 k = count % 0x4U; in arm_conv_fast_q31()
182 while (k > 0U) in arm_conv_fast_q31()
189 k--; in arm_conv_fast_q31()
252 k = srcBLen >> 2U; in arm_conv_fast_q31()
322 } while (--k); in arm_conv_fast_q31()
326 k = srcBLen % 0x4U; in arm_conv_fast_q31()
[all …]
|
| D | arm_conv_partial_q15.c | 80 uint32_t j, k, count, blkCnt, check; in arm_conv_partial_q15() local
178 k = count; in arm_conv_partial_q15()
180 while (k > 0U) in arm_conv_partial_q15()
186 k--; in arm_conv_partial_q15()
215 k = count >> 2U; in arm_conv_partial_q15()
219 while (k > 0U) in arm_conv_partial_q15()
228 k--; in arm_conv_partial_q15()
237 k = count % 0x4U; in arm_conv_partial_q15()
239 while (k > 0U) in arm_conv_partial_q15()
245 k--; in arm_conv_partial_q15()
[all …]
|
| D | arm_correlate_fast_q15.c | 78 uint32_t j, k, count, blkCnt; /* Loop counters */ in arm_correlate_fast_q15() local
185 k = count >> 2U; in arm_correlate_fast_q15()
189 while (k > 0U) in arm_correlate_fast_q15()
197 k--; in arm_correlate_fast_q15()
202 k = count % 0x4U; in arm_correlate_fast_q15()
204 while (k > 0U) in arm_correlate_fast_q15()
211 k--; in arm_correlate_fast_q15()
276 k = srcBLen >> 2U; in arm_correlate_fast_q15()
326 } while (--k); in arm_correlate_fast_q15()
333 k = srcBLen % 0x4U; in arm_correlate_fast_q15()
[all …]
|
| D | arm_correlate_f32.c | 327 uint32_t j, k, count, blkCnt; /* Loop counters */ in arm_correlate_f32() local
440 k = count >> 2U; in arm_correlate_f32()
447 while (k > 0U) in arm_correlate_f32()
458 k--; in arm_correlate_f32()
464 k = count & 0x3; in arm_correlate_f32()
468 while (k > 0U) in arm_correlate_f32()
483 k--; in arm_correlate_f32()
487 k = count % 0x4U; in arm_correlate_f32()
493 k = count; in arm_correlate_f32()
497 while (k > 0U) in arm_correlate_f32()
[all …]
|
| D | arm_correlate_fast_q31.c | 80 uint32_t j, k, count, blkCnt; /* Loop counters */ in arm_correlate_fast_q31() local
176 k = count >> 2U; in arm_correlate_fast_q31()
180 while (k > 0U) in arm_correlate_fast_q31()
199 k--; in arm_correlate_fast_q31()
204 k = count % 0x4U; in arm_correlate_fast_q31()
206 while (k > 0U) in arm_correlate_fast_q31()
214 k--; in arm_correlate_fast_q31()
278 k = srcBLen >> 2U; in arm_correlate_fast_q31()
348 } while (--k); in arm_correlate_fast_q31()
352 k = srcBLen % 0x4U; in arm_correlate_fast_q31()
[all …]
|
| /cmsis-dsp-latest/dsppp/Include/dsppp/DSP/ |
| D | basic.hpp | 37 for(int k=0;k < (1<<DSP_UNROLL);k++) in _Fill() local
39 v.vector_store(i + k*nb_lanes,inner::vconst(val)); in _Fill()
69 for(int k=0;k<(1<<DSP_UNROLL);k++) in _Fill2D() local
71 v.matrix_store(row+k,col,inner::vconst(val)); in _Fill2D()
77 for(int k=0;k<(1<<DSP_UNROLL);k++) in _Fill2D() local
79 v(row+k,col) = val; in _Fill2D()
121 for(int k=0;k < (1<<U);k++) in eval() local
123 v.vector_store(i + k*nb_lanes,other.vector_op(i+k*nb_lanes)); in eval()
152 for(int k=0;k<(1<<DSP_UNROLL);k++) in eval2D() local
154 v.matrix_store(row+k,col,other.matrix_op(row+k,col)); in eval2D()
[all …]
|
| /cmsis-dsp-latest/dsppp/Include/dsppp/Scalar/ |
| D | basic.hpp | 42 for(int k=0;k < (1<<U);k++) in _Fill() local
44 v[i+k] = val; in _Fill()
84 for(int k=0;k<(1<<U);k++) in _Fill2D() local
86 v(row+k,col) = val; in _Fill2D()
125 for(int k=0;k < (1<<U);k++) in eval() local
127 v[i+k] = other[i+k]; in eval()
165 for(int k=0;k<(1<<U);k++) in eval2D() local
167 v(row+k,col) = other(row+k,col); in eval2D()
212 for(int k=0;k < (1<<U);k++) in _dot() local
214 acc = inner::mac(acc , a[i+k] , b[i+k]); in _dot()
|
