| /external/XNNPACK/src/f32-igemm/gen/ |
| D | 5x8s4-wasmrelaxedsimd-fma.c | 113 v128_t va4 = wasm_v128_load(a4); in xnn_f32_igemm_ukernel_5x8s4__wasmrelaxedsimd_fma() local
124 vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, va4, vb0123c0); in xnn_f32_igemm_ukernel_5x8s4__wasmrelaxedsimd_fma()
129 vacc4x4567 = __builtin_wasm_fma_f32x4(vacc4x4567, va4, vb4567c0); in xnn_f32_igemm_ukernel_5x8s4__wasmrelaxedsimd_fma()
135 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_igemm_ukernel_5x8s4__wasmrelaxedsimd_fma()
144 vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, va4, vb0123c1); in xnn_f32_igemm_ukernel_5x8s4__wasmrelaxedsimd_fma()
149 vacc4x4567 = __builtin_wasm_fma_f32x4(vacc4x4567, va4, vb4567c1); in xnn_f32_igemm_ukernel_5x8s4__wasmrelaxedsimd_fma()
155 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_igemm_ukernel_5x8s4__wasmrelaxedsimd_fma()
164 vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, va4, vb0123c2); in xnn_f32_igemm_ukernel_5x8s4__wasmrelaxedsimd_fma()
169 vacc4x4567 = __builtin_wasm_fma_f32x4(vacc4x4567, va4, vb4567c2); in xnn_f32_igemm_ukernel_5x8s4__wasmrelaxedsimd_fma()
175 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_igemm_ukernel_5x8s4__wasmrelaxedsimd_fma()
[all …]
|
| D | 5x8s4-minmax-wasmrelaxedsimd-fma.c | 115 v128_t va4 = wasm_v128_load(a4); in xnn_f32_igemm_minmax_ukernel_5x8s4__wasmrelaxedsimd_fma() local
126 vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, va4, vb0123c0); in xnn_f32_igemm_minmax_ukernel_5x8s4__wasmrelaxedsimd_fma()
131 vacc4x4567 = __builtin_wasm_fma_f32x4(vacc4x4567, va4, vb4567c0); in xnn_f32_igemm_minmax_ukernel_5x8s4__wasmrelaxedsimd_fma()
137 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_igemm_minmax_ukernel_5x8s4__wasmrelaxedsimd_fma()
146 vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, va4, vb0123c1); in xnn_f32_igemm_minmax_ukernel_5x8s4__wasmrelaxedsimd_fma()
151 vacc4x4567 = __builtin_wasm_fma_f32x4(vacc4x4567, va4, vb4567c1); in xnn_f32_igemm_minmax_ukernel_5x8s4__wasmrelaxedsimd_fma()
157 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_igemm_minmax_ukernel_5x8s4__wasmrelaxedsimd_fma()
166 vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, va4, vb0123c2); in xnn_f32_igemm_minmax_ukernel_5x8s4__wasmrelaxedsimd_fma()
171 vacc4x4567 = __builtin_wasm_fma_f32x4(vacc4x4567, va4, vb4567c2); in xnn_f32_igemm_minmax_ukernel_5x8s4__wasmrelaxedsimd_fma()
177 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_igemm_minmax_ukernel_5x8s4__wasmrelaxedsimd_fma()
[all …]
|
| D | 5x8s4-relu-wasmsimd.c | 113 v128_t va4 = wasm_v128_load(a4); in xnn_f32_igemm_relu_ukernel_5x8s4__wasmsimd() local
124 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c0)); in xnn_f32_igemm_relu_ukernel_5x8s4__wasmsimd()
129 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c0)); in xnn_f32_igemm_relu_ukernel_5x8s4__wasmsimd()
135 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_igemm_relu_ukernel_5x8s4__wasmsimd()
144 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c1)); in xnn_f32_igemm_relu_ukernel_5x8s4__wasmsimd()
149 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c1)); in xnn_f32_igemm_relu_ukernel_5x8s4__wasmsimd()
155 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_igemm_relu_ukernel_5x8s4__wasmsimd()
164 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c2)); in xnn_f32_igemm_relu_ukernel_5x8s4__wasmsimd()
169 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c2)); in xnn_f32_igemm_relu_ukernel_5x8s4__wasmsimd()
175 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_igemm_relu_ukernel_5x8s4__wasmsimd()
[all …]
|
| D | 5x8s4-wasmsimd.c | 113 v128_t va4 = wasm_v128_load(a4); in xnn_f32_igemm_ukernel_5x8s4__wasmsimd() local
124 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c0)); in xnn_f32_igemm_ukernel_5x8s4__wasmsimd()
129 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c0)); in xnn_f32_igemm_ukernel_5x8s4__wasmsimd()
135 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_igemm_ukernel_5x8s4__wasmsimd()
144 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c1)); in xnn_f32_igemm_ukernel_5x8s4__wasmsimd()
149 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c1)); in xnn_f32_igemm_ukernel_5x8s4__wasmsimd()
155 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_igemm_ukernel_5x8s4__wasmsimd()
164 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c2)); in xnn_f32_igemm_ukernel_5x8s4__wasmsimd()
169 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c2)); in xnn_f32_igemm_ukernel_5x8s4__wasmsimd()
175 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_igemm_ukernel_5x8s4__wasmsimd()
[all …]
|
| D | 5x8s4-relu-wasmrelaxedsimd-fma.c | 113 v128_t va4 = wasm_v128_load(a4); in xnn_f32_igemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma() local
124 vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, va4, vb0123c0); in xnn_f32_igemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma()
129 vacc4x4567 = __builtin_wasm_fma_f32x4(vacc4x4567, va4, vb4567c0); in xnn_f32_igemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma()
135 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_igemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma()
144 vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, va4, vb0123c1); in xnn_f32_igemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma()
149 vacc4x4567 = __builtin_wasm_fma_f32x4(vacc4x4567, va4, vb4567c1); in xnn_f32_igemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma()
155 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_igemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma()
164 vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, va4, vb0123c2); in xnn_f32_igemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma()
169 vacc4x4567 = __builtin_wasm_fma_f32x4(vacc4x4567, va4, vb4567c2); in xnn_f32_igemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma()
175 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_igemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma()
[all …]
|
| D | 5x8s4-minmax-wasmsimd-arm.c | 115 v128_t va4 = wasm_v128_load(a4); in xnn_f32_igemm_minmax_ukernel_5x8s4__wasmsimd_arm() local
126 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c0)); in xnn_f32_igemm_minmax_ukernel_5x8s4__wasmsimd_arm()
131 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c0)); in xnn_f32_igemm_minmax_ukernel_5x8s4__wasmsimd_arm()
137 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_igemm_minmax_ukernel_5x8s4__wasmsimd_arm()
146 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c1)); in xnn_f32_igemm_minmax_ukernel_5x8s4__wasmsimd_arm()
151 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c1)); in xnn_f32_igemm_minmax_ukernel_5x8s4__wasmsimd_arm()
157 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_igemm_minmax_ukernel_5x8s4__wasmsimd_arm()
166 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c2)); in xnn_f32_igemm_minmax_ukernel_5x8s4__wasmsimd_arm()
171 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c2)); in xnn_f32_igemm_minmax_ukernel_5x8s4__wasmsimd_arm()
177 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_igemm_minmax_ukernel_5x8s4__wasmsimd_arm()
[all …]
|
| D | 6x8s4-wasmrelaxedsimd-fma.c | 124 v128_t va4 = wasm_v128_load(a4); in xnn_f32_igemm_ukernel_6x8s4__wasmrelaxedsimd_fma() local
137 vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, va4, vb0123c0); in xnn_f32_igemm_ukernel_6x8s4__wasmrelaxedsimd_fma()
143 vacc4x4567 = __builtin_wasm_fma_f32x4(vacc4x4567, va4, vb4567c0); in xnn_f32_igemm_ukernel_6x8s4__wasmrelaxedsimd_fma()
150 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_igemm_ukernel_6x8s4__wasmrelaxedsimd_fma()
160 vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, va4, vb0123c1); in xnn_f32_igemm_ukernel_6x8s4__wasmrelaxedsimd_fma()
166 vacc4x4567 = __builtin_wasm_fma_f32x4(vacc4x4567, va4, vb4567c1); in xnn_f32_igemm_ukernel_6x8s4__wasmrelaxedsimd_fma()
173 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_igemm_ukernel_6x8s4__wasmrelaxedsimd_fma()
183 vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, va4, vb0123c2); in xnn_f32_igemm_ukernel_6x8s4__wasmrelaxedsimd_fma()
189 vacc4x4567 = __builtin_wasm_fma_f32x4(vacc4x4567, va4, vb4567c2); in xnn_f32_igemm_ukernel_6x8s4__wasmrelaxedsimd_fma()
196 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_igemm_ukernel_6x8s4__wasmrelaxedsimd_fma()
[all …]
|
| D | 5x8s4-minmax-wasmrelaxedsimd.c | 115 v128_t va4 = wasm_v128_load(a4); in xnn_f32_igemm_minmax_ukernel_5x8s4__wasmrelaxedsimd() local
126 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c0)); in xnn_f32_igemm_minmax_ukernel_5x8s4__wasmrelaxedsimd()
131 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c0)); in xnn_f32_igemm_minmax_ukernel_5x8s4__wasmrelaxedsimd()
137 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_igemm_minmax_ukernel_5x8s4__wasmrelaxedsimd()
146 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c1)); in xnn_f32_igemm_minmax_ukernel_5x8s4__wasmrelaxedsimd()
151 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c1)); in xnn_f32_igemm_minmax_ukernel_5x8s4__wasmrelaxedsimd()
157 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_igemm_minmax_ukernel_5x8s4__wasmrelaxedsimd()
166 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c2)); in xnn_f32_igemm_minmax_ukernel_5x8s4__wasmrelaxedsimd()
171 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c2)); in xnn_f32_igemm_minmax_ukernel_5x8s4__wasmrelaxedsimd()
177 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_igemm_minmax_ukernel_5x8s4__wasmrelaxedsimd()
[all …]
|
| /external/XNNPACK/src/f32-gemm/gen/ |
| D | 5x8s4-wasmrelaxedsimd-fma.c | 88 v128_t va4 = wasm_v128_load(a4); in xnn_f32_gemm_ukernel_5x8s4__wasmrelaxedsimd_fma() local
99 vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, va4, vb0123c0); in xnn_f32_gemm_ukernel_5x8s4__wasmrelaxedsimd_fma()
104 vacc4x4567 = __builtin_wasm_fma_f32x4(vacc4x4567, va4, vb4567c0); in xnn_f32_gemm_ukernel_5x8s4__wasmrelaxedsimd_fma()
110 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_gemm_ukernel_5x8s4__wasmrelaxedsimd_fma()
119 vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, va4, vb0123c1); in xnn_f32_gemm_ukernel_5x8s4__wasmrelaxedsimd_fma()
124 vacc4x4567 = __builtin_wasm_fma_f32x4(vacc4x4567, va4, vb4567c1); in xnn_f32_gemm_ukernel_5x8s4__wasmrelaxedsimd_fma()
130 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_gemm_ukernel_5x8s4__wasmrelaxedsimd_fma()
139 vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, va4, vb0123c2); in xnn_f32_gemm_ukernel_5x8s4__wasmrelaxedsimd_fma()
144 vacc4x4567 = __builtin_wasm_fma_f32x4(vacc4x4567, va4, vb4567c2); in xnn_f32_gemm_ukernel_5x8s4__wasmrelaxedsimd_fma()
150 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_gemm_ukernel_5x8s4__wasmrelaxedsimd_fma()
[all …]
|
| D | 5x8s4-relu-wasmrelaxedsimd-fma.c | 88 v128_t va4 = wasm_v128_load(a4); in xnn_f32_gemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma() local
99 vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, va4, vb0123c0); in xnn_f32_gemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma()
104 vacc4x4567 = __builtin_wasm_fma_f32x4(vacc4x4567, va4, vb4567c0); in xnn_f32_gemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma()
110 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_gemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma()
119 vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, va4, vb0123c1); in xnn_f32_gemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma()
124 vacc4x4567 = __builtin_wasm_fma_f32x4(vacc4x4567, va4, vb4567c1); in xnn_f32_gemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma()
130 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_gemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma()
139 vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, va4, vb0123c2); in xnn_f32_gemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma()
144 vacc4x4567 = __builtin_wasm_fma_f32x4(vacc4x4567, va4, vb4567c2); in xnn_f32_gemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma()
150 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_gemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma()
[all …]
|
| D | 5x8s4-minmax-wasmrelaxedsimd.c | 90 v128_t va4 = wasm_v128_load(a4); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmrelaxedsimd() local
101 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c0)); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmrelaxedsimd()
106 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c0)); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmrelaxedsimd()
112 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmrelaxedsimd()
121 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c1)); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmrelaxedsimd()
126 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c1)); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmrelaxedsimd()
132 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmrelaxedsimd()
141 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c2)); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmrelaxedsimd()
146 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c2)); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmrelaxedsimd()
152 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmrelaxedsimd()
[all …]
|
| D | 5x8s4-minmax-wasmsimd-arm.c | 90 v128_t va4 = wasm_v128_load(a4); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmsimd_arm() local
101 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c0)); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmsimd_arm()
106 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c0)); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmsimd_arm()
112 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmsimd_arm()
121 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c1)); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmsimd_arm()
126 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c1)); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmsimd_arm()
132 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmsimd_arm()
141 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c2)); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmsimd_arm()
146 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c2)); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmsimd_arm()
152 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmsimd_arm()
[all …]
|
| D | 5x8s4-relu-wasmsimd.c | 88 v128_t va4 = wasm_v128_load(a4); in xnn_f32_gemm_relu_ukernel_5x8s4__wasmsimd() local
99 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c0)); in xnn_f32_gemm_relu_ukernel_5x8s4__wasmsimd()
104 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c0)); in xnn_f32_gemm_relu_ukernel_5x8s4__wasmsimd()
110 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_gemm_relu_ukernel_5x8s4__wasmsimd()
119 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c1)); in xnn_f32_gemm_relu_ukernel_5x8s4__wasmsimd()
124 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c1)); in xnn_f32_gemm_relu_ukernel_5x8s4__wasmsimd()
130 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_gemm_relu_ukernel_5x8s4__wasmsimd()
139 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c2)); in xnn_f32_gemm_relu_ukernel_5x8s4__wasmsimd()
144 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c2)); in xnn_f32_gemm_relu_ukernel_5x8s4__wasmsimd()
150 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_gemm_relu_ukernel_5x8s4__wasmsimd()
[all …]
|
| D | 5x8s4-minmax-wasmsimd-x86.c | 90 v128_t va4 = wasm_v128_load(a4); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmsimd_x86() local
101 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c0)); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmsimd_x86()
106 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c0)); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmsimd_x86()
112 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmsimd_x86()
121 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c1)); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmsimd_x86()
126 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c1)); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmsimd_x86()
132 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmsimd_x86()
141 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c2)); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmsimd_x86()
146 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c2)); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmsimd_x86()
152 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmsimd_x86()
[all …]
|
| D | 5x8s4-minmax-wasmrelaxedsimd-fma.c | 90 v128_t va4 = wasm_v128_load(a4); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmrelaxedsimd_fma() local
101 vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, va4, vb0123c0); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmrelaxedsimd_fma()
106 vacc4x4567 = __builtin_wasm_fma_f32x4(vacc4x4567, va4, vb4567c0); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmrelaxedsimd_fma()
112 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmrelaxedsimd_fma()
121 vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, va4, vb0123c1); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmrelaxedsimd_fma()
126 vacc4x4567 = __builtin_wasm_fma_f32x4(vacc4x4567, va4, vb4567c1); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmrelaxedsimd_fma()
132 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmrelaxedsimd_fma()
141 vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, va4, vb0123c2); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmrelaxedsimd_fma()
146 vacc4x4567 = __builtin_wasm_fma_f32x4(vacc4x4567, va4, vb4567c2); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmrelaxedsimd_fma()
152 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmrelaxedsimd_fma()
[all …]
|
| D | 5x8s4-wasmsimd.c | 88 v128_t va4 = wasm_v128_load(a4); in xnn_f32_gemm_ukernel_5x8s4__wasmsimd() local
99 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c0)); in xnn_f32_gemm_ukernel_5x8s4__wasmsimd()
104 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c0)); in xnn_f32_gemm_ukernel_5x8s4__wasmsimd()
110 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_gemm_ukernel_5x8s4__wasmsimd()
119 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c1)); in xnn_f32_gemm_ukernel_5x8s4__wasmsimd()
124 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c1)); in xnn_f32_gemm_ukernel_5x8s4__wasmsimd()
130 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_gemm_ukernel_5x8s4__wasmsimd()
139 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c2)); in xnn_f32_gemm_ukernel_5x8s4__wasmsimd()
144 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c2)); in xnn_f32_gemm_ukernel_5x8s4__wasmsimd()
150 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_gemm_ukernel_5x8s4__wasmsimd()
[all …]
|
| D | 5x8s4-minmax-sse.c | 88 __m128 va4 = _mm_loadu_ps(a4); in xnn_f32_gemm_minmax_ukernel_5x8s4__sse() local
99 vacc4x0123 = _mm_add_ps(vacc4x0123, _mm_mul_ps(va4, vb0123c0)); in xnn_f32_gemm_minmax_ukernel_5x8s4__sse()
104 vacc4x4567 = _mm_add_ps(vacc4x4567, _mm_mul_ps(va4, vb4567c0)); in xnn_f32_gemm_minmax_ukernel_5x8s4__sse()
110 va4 = _mm_shuffle_ps(va4, va4, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_gemm_minmax_ukernel_5x8s4__sse()
119 vacc4x0123 = _mm_add_ps(vacc4x0123, _mm_mul_ps(va4, vb0123c1)); in xnn_f32_gemm_minmax_ukernel_5x8s4__sse()
124 vacc4x4567 = _mm_add_ps(vacc4x4567, _mm_mul_ps(va4, vb4567c1)); in xnn_f32_gemm_minmax_ukernel_5x8s4__sse()
130 va4 = _mm_shuffle_ps(va4, va4, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_gemm_minmax_ukernel_5x8s4__sse()
139 vacc4x0123 = _mm_add_ps(vacc4x0123, _mm_mul_ps(va4, vb0123c2)); in xnn_f32_gemm_minmax_ukernel_5x8s4__sse()
144 vacc4x4567 = _mm_add_ps(vacc4x4567, _mm_mul_ps(va4, vb4567c2)); in xnn_f32_gemm_minmax_ukernel_5x8s4__sse()
150 va4 = _mm_shuffle_ps(va4, va4, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_gemm_minmax_ukernel_5x8s4__sse()
[all …]
|
| D | 6x8s4-relu-wasmrelaxedsimd-fma.c | 96 v128_t va4 = wasm_v128_load(a4); in xnn_f32_gemm_relu_ukernel_6x8s4__wasmrelaxedsimd_fma() local
109 vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, va4, vb0123c0); in xnn_f32_gemm_relu_ukernel_6x8s4__wasmrelaxedsimd_fma()
115 vacc4x4567 = __builtin_wasm_fma_f32x4(vacc4x4567, va4, vb4567c0); in xnn_f32_gemm_relu_ukernel_6x8s4__wasmrelaxedsimd_fma()
122 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_gemm_relu_ukernel_6x8s4__wasmrelaxedsimd_fma()
132 vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, va4, vb0123c1); in xnn_f32_gemm_relu_ukernel_6x8s4__wasmrelaxedsimd_fma()
138 vacc4x4567 = __builtin_wasm_fma_f32x4(vacc4x4567, va4, vb4567c1); in xnn_f32_gemm_relu_ukernel_6x8s4__wasmrelaxedsimd_fma()
145 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_gemm_relu_ukernel_6x8s4__wasmrelaxedsimd_fma()
155 vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, va4, vb0123c2); in xnn_f32_gemm_relu_ukernel_6x8s4__wasmrelaxedsimd_fma()
161 vacc4x4567 = __builtin_wasm_fma_f32x4(vacc4x4567, va4, vb4567c2); in xnn_f32_gemm_relu_ukernel_6x8s4__wasmrelaxedsimd_fma()
168 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_gemm_relu_ukernel_6x8s4__wasmrelaxedsimd_fma()
[all …]
|
| D | 6x8s4-wasmsimd.c | 96 v128_t va4 = wasm_v128_load(a4); in xnn_f32_gemm_ukernel_6x8s4__wasmsimd() local
109 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c0)); in xnn_f32_gemm_ukernel_6x8s4__wasmsimd()
115 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c0)); in xnn_f32_gemm_ukernel_6x8s4__wasmsimd()
122 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_gemm_ukernel_6x8s4__wasmsimd()
132 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c1)); in xnn_f32_gemm_ukernel_6x8s4__wasmsimd()
138 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c1)); in xnn_f32_gemm_ukernel_6x8s4__wasmsimd()
145 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_gemm_ukernel_6x8s4__wasmsimd()
155 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c2)); in xnn_f32_gemm_ukernel_6x8s4__wasmsimd()
161 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c2)); in xnn_f32_gemm_ukernel_6x8s4__wasmsimd()
168 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_gemm_ukernel_6x8s4__wasmsimd()
[all …]
|
| D | 6x8s4-wasmrelaxedsimd-fma.c | 96 v128_t va4 = wasm_v128_load(a4); in xnn_f32_gemm_ukernel_6x8s4__wasmrelaxedsimd_fma() local
109 vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, va4, vb0123c0); in xnn_f32_gemm_ukernel_6x8s4__wasmrelaxedsimd_fma()
115 vacc4x4567 = __builtin_wasm_fma_f32x4(vacc4x4567, va4, vb4567c0); in xnn_f32_gemm_ukernel_6x8s4__wasmrelaxedsimd_fma()
122 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_gemm_ukernel_6x8s4__wasmrelaxedsimd_fma()
132 vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, va4, vb0123c1); in xnn_f32_gemm_ukernel_6x8s4__wasmrelaxedsimd_fma()
138 vacc4x4567 = __builtin_wasm_fma_f32x4(vacc4x4567, va4, vb4567c1); in xnn_f32_gemm_ukernel_6x8s4__wasmrelaxedsimd_fma()
145 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_gemm_ukernel_6x8s4__wasmrelaxedsimd_fma()
155 vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, va4, vb0123c2); in xnn_f32_gemm_ukernel_6x8s4__wasmrelaxedsimd_fma()
161 vacc4x4567 = __builtin_wasm_fma_f32x4(vacc4x4567, va4, vb4567c2); in xnn_f32_gemm_ukernel_6x8s4__wasmrelaxedsimd_fma()
168 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_gemm_ukernel_6x8s4__wasmrelaxedsimd_fma()
[all …]
|
| /external/XNNPACK/src/f32-gemm/gen-inc/ |
| D | 5x8s4inc-minmax-wasmsimd-x86.c | 92 v128_t va4 = wasm_v128_load(a4); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmsimd_x86() local
103 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c0)); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmsimd_x86()
108 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c0)); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmsimd_x86()
114 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmsimd_x86()
123 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c1)); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmsimd_x86()
128 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c1)); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmsimd_x86()
134 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmsimd_x86()
143 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c2)); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmsimd_x86()
148 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c2)); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmsimd_x86()
154 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmsimd_x86()
[all …]
|
| D | 5x8s4inc-minmax-wasmrelaxedsimd-fma.c | 92 v128_t va4 = wasm_v128_load(a4); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmrelaxedsimd_fma() local
103 vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, va4, vb0123c0); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmrelaxedsimd_fma()
108 vacc4x4567 = __builtin_wasm_fma_f32x4(vacc4x4567, va4, vb4567c0); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmrelaxedsimd_fma()
114 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmrelaxedsimd_fma()
123 vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, va4, vb0123c1); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmrelaxedsimd_fma()
128 vacc4x4567 = __builtin_wasm_fma_f32x4(vacc4x4567, va4, vb4567c1); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmrelaxedsimd_fma()
134 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmrelaxedsimd_fma()
143 vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, va4, vb0123c2); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmrelaxedsimd_fma()
148 vacc4x4567 = __builtin_wasm_fma_f32x4(vacc4x4567, va4, vb4567c2); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmrelaxedsimd_fma()
154 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmrelaxedsimd_fma()
[all …]
|
| D | 5x8s4inc-minmax-wasmsimd-arm.c | 92 v128_t va4 = wasm_v128_load(a4); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmsimd_arm() local
103 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c0)); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmsimd_arm()
108 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c0)); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmsimd_arm()
114 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmsimd_arm()
123 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c1)); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmsimd_arm()
128 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c1)); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmsimd_arm()
134 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmsimd_arm()
143 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c2)); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmsimd_arm()
148 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c2)); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmsimd_arm()
154 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmsimd_arm()
[all …]
|
| D | 5x8s4inc-minmax-relaxedwasmsimd.c | 92 v128_t va4 = wasm_v128_load(a4); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmsimd_x86() local
103 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c0)); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmsimd_x86()
108 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c0)); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmsimd_x86()
114 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmsimd_x86()
123 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c1)); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmsimd_x86()
128 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c1)); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmsimd_x86()
134 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmsimd_x86()
143 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c2)); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmsimd_x86()
148 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c2)); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmsimd_x86()
154 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmsimd_x86()
[all …]
|
| D | 5x8s4inc-minmax-wasmrelaxedsimd.c | 92 v128_t va4 = wasm_v128_load(a4); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmrelaxedsimd() local
103 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c0)); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmrelaxedsimd()
108 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c0)); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmrelaxedsimd()
114 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmrelaxedsimd()
123 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c1)); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmrelaxedsimd()
128 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c1)); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmrelaxedsimd()
134 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmrelaxedsimd()
143 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c2)); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmrelaxedsimd()
148 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c2)); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmrelaxedsimd()
154 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmrelaxedsimd()
[all …]
|