| /external/XNNPACK/src/f32-igemm/gen/ |
| D | 5x8s4-wasmrelaxedsimd-fma.c | 117 const v128_t vb0123c0 = wasm_v128_load(w + 0); in xnn_f32_igemm_ukernel_5x8s4__wasmrelaxedsimd_fma() local
120 vacc0x0123 = __builtin_wasm_fma_f32x4(vacc0x0123, va0, vb0123c0); in xnn_f32_igemm_ukernel_5x8s4__wasmrelaxedsimd_fma()
121 vacc1x0123 = __builtin_wasm_fma_f32x4(vacc1x0123, va1, vb0123c0); in xnn_f32_igemm_ukernel_5x8s4__wasmrelaxedsimd_fma()
122 vacc2x0123 = __builtin_wasm_fma_f32x4(vacc2x0123, va2, vb0123c0); in xnn_f32_igemm_ukernel_5x8s4__wasmrelaxedsimd_fma()
123 vacc3x0123 = __builtin_wasm_fma_f32x4(vacc3x0123, va3, vb0123c0); in xnn_f32_igemm_ukernel_5x8s4__wasmrelaxedsimd_fma()
124 vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, va4, vb0123c0); in xnn_f32_igemm_ukernel_5x8s4__wasmrelaxedsimd_fma()
209 const v128_t vb0123c0 = wasm_v128_load(w + 0); in xnn_f32_igemm_ukernel_5x8s4__wasmrelaxedsimd_fma() local
212 …uiltin_wasm_fma_f32x4(vacc0x0123, wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0); in xnn_f32_igemm_ukernel_5x8s4__wasmrelaxedsimd_fma()
213 …uiltin_wasm_fma_f32x4(vacc1x0123, wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0); in xnn_f32_igemm_ukernel_5x8s4__wasmrelaxedsimd_fma()
214 …uiltin_wasm_fma_f32x4(vacc2x0123, wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0); in xnn_f32_igemm_ukernel_5x8s4__wasmrelaxedsimd_fma()
[all …]
|
| D | 6x8s4-wasmrelaxedsimd-fma.c | 130 const v128_t vb0123c0 = wasm_v128_load(w + 0); in xnn_f32_igemm_ukernel_6x8s4__wasmrelaxedsimd_fma() local
133 vacc0x0123 = __builtin_wasm_fma_f32x4(vacc0x0123, va0, vb0123c0); in xnn_f32_igemm_ukernel_6x8s4__wasmrelaxedsimd_fma()
134 vacc1x0123 = __builtin_wasm_fma_f32x4(vacc1x0123, va1, vb0123c0); in xnn_f32_igemm_ukernel_6x8s4__wasmrelaxedsimd_fma()
135 vacc2x0123 = __builtin_wasm_fma_f32x4(vacc2x0123, va2, vb0123c0); in xnn_f32_igemm_ukernel_6x8s4__wasmrelaxedsimd_fma()
136 vacc3x0123 = __builtin_wasm_fma_f32x4(vacc3x0123, va3, vb0123c0); in xnn_f32_igemm_ukernel_6x8s4__wasmrelaxedsimd_fma()
137 vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, va4, vb0123c0); in xnn_f32_igemm_ukernel_6x8s4__wasmrelaxedsimd_fma()
138 vacc5x0123 = __builtin_wasm_fma_f32x4(vacc5x0123, va5, vb0123c0); in xnn_f32_igemm_ukernel_6x8s4__wasmrelaxedsimd_fma()
235 const v128_t vb0123c0 = wasm_v128_load(w + 0); in xnn_f32_igemm_ukernel_6x8s4__wasmrelaxedsimd_fma() local
238 …uiltin_wasm_fma_f32x4(vacc0x0123, wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0); in xnn_f32_igemm_ukernel_6x8s4__wasmrelaxedsimd_fma()
239 …uiltin_wasm_fma_f32x4(vacc1x0123, wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0); in xnn_f32_igemm_ukernel_6x8s4__wasmrelaxedsimd_fma()
[all …]
|
| D | 6x8s4-relu-wasmrelaxedsimd-fma.c | 130 const v128_t vb0123c0 = wasm_v128_load(w + 0); in xnn_f32_igemm_relu_ukernel_6x8s4__wasmrelaxedsimd_fma() local
133 vacc0x0123 = __builtin_wasm_fma_f32x4(vacc0x0123, va0, vb0123c0); in xnn_f32_igemm_relu_ukernel_6x8s4__wasmrelaxedsimd_fma()
134 vacc1x0123 = __builtin_wasm_fma_f32x4(vacc1x0123, va1, vb0123c0); in xnn_f32_igemm_relu_ukernel_6x8s4__wasmrelaxedsimd_fma()
135 vacc2x0123 = __builtin_wasm_fma_f32x4(vacc2x0123, va2, vb0123c0); in xnn_f32_igemm_relu_ukernel_6x8s4__wasmrelaxedsimd_fma()
136 vacc3x0123 = __builtin_wasm_fma_f32x4(vacc3x0123, va3, vb0123c0); in xnn_f32_igemm_relu_ukernel_6x8s4__wasmrelaxedsimd_fma()
137 vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, va4, vb0123c0); in xnn_f32_igemm_relu_ukernel_6x8s4__wasmrelaxedsimd_fma()
138 vacc5x0123 = __builtin_wasm_fma_f32x4(vacc5x0123, va5, vb0123c0); in xnn_f32_igemm_relu_ukernel_6x8s4__wasmrelaxedsimd_fma()
235 const v128_t vb0123c0 = wasm_v128_load(w + 0); in xnn_f32_igemm_relu_ukernel_6x8s4__wasmrelaxedsimd_fma() local
238 …uiltin_wasm_fma_f32x4(vacc0x0123, wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0); in xnn_f32_igemm_relu_ukernel_6x8s4__wasmrelaxedsimd_fma()
239 …uiltin_wasm_fma_f32x4(vacc1x0123, wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0); in xnn_f32_igemm_relu_ukernel_6x8s4__wasmrelaxedsimd_fma()
[all …]
|
| D | 4x8s4-wasmrelaxedsimd-fma.c | 104 const v128_t vb0123c0 = wasm_v128_load(w + 0); in xnn_f32_igemm_ukernel_4x8s4__wasmrelaxedsimd_fma() local
107 vacc0x0123 = __builtin_wasm_fma_f32x4(vacc0x0123, va0, vb0123c0); in xnn_f32_igemm_ukernel_4x8s4__wasmrelaxedsimd_fma()
108 vacc1x0123 = __builtin_wasm_fma_f32x4(vacc1x0123, va1, vb0123c0); in xnn_f32_igemm_ukernel_4x8s4__wasmrelaxedsimd_fma()
109 vacc2x0123 = __builtin_wasm_fma_f32x4(vacc2x0123, va2, vb0123c0); in xnn_f32_igemm_ukernel_4x8s4__wasmrelaxedsimd_fma()
110 vacc3x0123 = __builtin_wasm_fma_f32x4(vacc3x0123, va3, vb0123c0); in xnn_f32_igemm_ukernel_4x8s4__wasmrelaxedsimd_fma()
183 const v128_t vb0123c0 = wasm_v128_load(w + 0); in xnn_f32_igemm_ukernel_4x8s4__wasmrelaxedsimd_fma() local
186 …uiltin_wasm_fma_f32x4(vacc0x0123, wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0); in xnn_f32_igemm_ukernel_4x8s4__wasmrelaxedsimd_fma()
187 …uiltin_wasm_fma_f32x4(vacc1x0123, wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0); in xnn_f32_igemm_ukernel_4x8s4__wasmrelaxedsimd_fma()
188 …uiltin_wasm_fma_f32x4(vacc2x0123, wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0); in xnn_f32_igemm_ukernel_4x8s4__wasmrelaxedsimd_fma()
189 …uiltin_wasm_fma_f32x4(vacc3x0123, wasm_v128_andnot(va3, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0); in xnn_f32_igemm_ukernel_4x8s4__wasmrelaxedsimd_fma()
|
| D | 4x8s4-relu-wasmrelaxedsimd-fma.c | 104 const v128_t vb0123c0 = wasm_v128_load(w + 0); in xnn_f32_igemm_relu_ukernel_4x8s4__wasmrelaxedsimd_fma() local
107 vacc0x0123 = __builtin_wasm_fma_f32x4(vacc0x0123, va0, vb0123c0); in xnn_f32_igemm_relu_ukernel_4x8s4__wasmrelaxedsimd_fma()
108 vacc1x0123 = __builtin_wasm_fma_f32x4(vacc1x0123, va1, vb0123c0); in xnn_f32_igemm_relu_ukernel_4x8s4__wasmrelaxedsimd_fma()
109 vacc2x0123 = __builtin_wasm_fma_f32x4(vacc2x0123, va2, vb0123c0); in xnn_f32_igemm_relu_ukernel_4x8s4__wasmrelaxedsimd_fma()
110 vacc3x0123 = __builtin_wasm_fma_f32x4(vacc3x0123, va3, vb0123c0); in xnn_f32_igemm_relu_ukernel_4x8s4__wasmrelaxedsimd_fma()
183 const v128_t vb0123c0 = wasm_v128_load(w + 0); in xnn_f32_igemm_relu_ukernel_4x8s4__wasmrelaxedsimd_fma() local
186 …uiltin_wasm_fma_f32x4(vacc0x0123, wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0); in xnn_f32_igemm_relu_ukernel_4x8s4__wasmrelaxedsimd_fma()
187 …uiltin_wasm_fma_f32x4(vacc1x0123, wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0); in xnn_f32_igemm_relu_ukernel_4x8s4__wasmrelaxedsimd_fma()
188 …uiltin_wasm_fma_f32x4(vacc2x0123, wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0); in xnn_f32_igemm_relu_ukernel_4x8s4__wasmrelaxedsimd_fma()
189 …uiltin_wasm_fma_f32x4(vacc3x0123, wasm_v128_andnot(va3, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0); in xnn_f32_igemm_relu_ukernel_4x8s4__wasmrelaxedsimd_fma()
|
| D | 5x8s4-minmax-wasmrelaxedsimd-fma.c | 119 const v128_t vb0123c0 = wasm_v128_load(w + 0); in xnn_f32_igemm_minmax_ukernel_5x8s4__wasmrelaxedsimd_fma() local
122 vacc0x0123 = __builtin_wasm_fma_f32x4(vacc0x0123, va0, vb0123c0); in xnn_f32_igemm_minmax_ukernel_5x8s4__wasmrelaxedsimd_fma()
123 vacc1x0123 = __builtin_wasm_fma_f32x4(vacc1x0123, va1, vb0123c0); in xnn_f32_igemm_minmax_ukernel_5x8s4__wasmrelaxedsimd_fma()
124 vacc2x0123 = __builtin_wasm_fma_f32x4(vacc2x0123, va2, vb0123c0); in xnn_f32_igemm_minmax_ukernel_5x8s4__wasmrelaxedsimd_fma()
125 vacc3x0123 = __builtin_wasm_fma_f32x4(vacc3x0123, va3, vb0123c0); in xnn_f32_igemm_minmax_ukernel_5x8s4__wasmrelaxedsimd_fma()
126 vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, va4, vb0123c0); in xnn_f32_igemm_minmax_ukernel_5x8s4__wasmrelaxedsimd_fma()
211 const v128_t vb0123c0 = wasm_v128_load(w + 0); in xnn_f32_igemm_minmax_ukernel_5x8s4__wasmrelaxedsimd_fma() local
214 …uiltin_wasm_fma_f32x4(vacc0x0123, wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0); in xnn_f32_igemm_minmax_ukernel_5x8s4__wasmrelaxedsimd_fma()
215 …uiltin_wasm_fma_f32x4(vacc1x0123, wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0); in xnn_f32_igemm_minmax_ukernel_5x8s4__wasmrelaxedsimd_fma()
216 …uiltin_wasm_fma_f32x4(vacc2x0123, wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0); in xnn_f32_igemm_minmax_ukernel_5x8s4__wasmrelaxedsimd_fma()
[all …]
|
| D | 5x8s4-relu-wasmsimd.c | 117 const v128_t vb0123c0 = wasm_v128_load(w + 0); in xnn_f32_igemm_relu_ukernel_5x8s4__wasmsimd() local
120 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c0)); in xnn_f32_igemm_relu_ukernel_5x8s4__wasmsimd()
121 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c0)); in xnn_f32_igemm_relu_ukernel_5x8s4__wasmsimd()
122 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c0)); in xnn_f32_igemm_relu_ukernel_5x8s4__wasmsimd()
123 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123c0)); in xnn_f32_igemm_relu_ukernel_5x8s4__wasmsimd()
124 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c0)); in xnn_f32_igemm_relu_ukernel_5x8s4__wasmsimd()
209 const v128_t vb0123c0 = wasm_v128_load(w + 0); in xnn_f32_igemm_relu_ukernel_5x8s4__wasmsimd() local
212 …4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0)); in xnn_f32_igemm_relu_ukernel_5x8s4__wasmsimd()
213 …4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0)); in xnn_f32_igemm_relu_ukernel_5x8s4__wasmsimd()
214 …4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0)); in xnn_f32_igemm_relu_ukernel_5x8s4__wasmsimd()
[all …]
|
| D | 5x8s4-wasmsimd.c | 117 const v128_t vb0123c0 = wasm_v128_load(w + 0); in xnn_f32_igemm_ukernel_5x8s4__wasmsimd() local
120 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c0)); in xnn_f32_igemm_ukernel_5x8s4__wasmsimd()
121 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c0)); in xnn_f32_igemm_ukernel_5x8s4__wasmsimd()
122 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c0)); in xnn_f32_igemm_ukernel_5x8s4__wasmsimd()
123 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123c0)); in xnn_f32_igemm_ukernel_5x8s4__wasmsimd()
124 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c0)); in xnn_f32_igemm_ukernel_5x8s4__wasmsimd()
209 const v128_t vb0123c0 = wasm_v128_load(w + 0); in xnn_f32_igemm_ukernel_5x8s4__wasmsimd() local
212 …4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0)); in xnn_f32_igemm_ukernel_5x8s4__wasmsimd()
213 …4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0)); in xnn_f32_igemm_ukernel_5x8s4__wasmsimd()
214 …4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0)); in xnn_f32_igemm_ukernel_5x8s4__wasmsimd()
[all …]
|
| D | 5x8s4-relu-wasmrelaxedsimd-fma.c | 117 const v128_t vb0123c0 = wasm_v128_load(w + 0); in xnn_f32_igemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma() local
120 vacc0x0123 = __builtin_wasm_fma_f32x4(vacc0x0123, va0, vb0123c0); in xnn_f32_igemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma()
121 vacc1x0123 = __builtin_wasm_fma_f32x4(vacc1x0123, va1, vb0123c0); in xnn_f32_igemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma()
122 vacc2x0123 = __builtin_wasm_fma_f32x4(vacc2x0123, va2, vb0123c0); in xnn_f32_igemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma()
123 vacc3x0123 = __builtin_wasm_fma_f32x4(vacc3x0123, va3, vb0123c0); in xnn_f32_igemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma()
124 vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, va4, vb0123c0); in xnn_f32_igemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma()
209 const v128_t vb0123c0 = wasm_v128_load(w + 0); in xnn_f32_igemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma() local
212 …uiltin_wasm_fma_f32x4(vacc0x0123, wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0); in xnn_f32_igemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma()
213 …uiltin_wasm_fma_f32x4(vacc1x0123, wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0); in xnn_f32_igemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma()
214 …uiltin_wasm_fma_f32x4(vacc2x0123, wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0); in xnn_f32_igemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma()
[all …]
|
| /external/XNNPACK/src/f32-gemm/gen/ |
| D | 5x8s4-wasmrelaxedsimd-fma.c | 92 const v128_t vb0123c0 = wasm_v128_load(w + 0); in xnn_f32_gemm_ukernel_5x8s4__wasmrelaxedsimd_fma() local
95 vacc0x0123 = __builtin_wasm_fma_f32x4(vacc0x0123, va0, vb0123c0); in xnn_f32_gemm_ukernel_5x8s4__wasmrelaxedsimd_fma()
96 vacc1x0123 = __builtin_wasm_fma_f32x4(vacc1x0123, va1, vb0123c0); in xnn_f32_gemm_ukernel_5x8s4__wasmrelaxedsimd_fma()
97 vacc2x0123 = __builtin_wasm_fma_f32x4(vacc2x0123, va2, vb0123c0); in xnn_f32_gemm_ukernel_5x8s4__wasmrelaxedsimd_fma()
98 vacc3x0123 = __builtin_wasm_fma_f32x4(vacc3x0123, va3, vb0123c0); in xnn_f32_gemm_ukernel_5x8s4__wasmrelaxedsimd_fma()
99 vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, va4, vb0123c0); in xnn_f32_gemm_ukernel_5x8s4__wasmrelaxedsimd_fma()
184 const v128_t vb0123c0 = wasm_v128_load(w + 0); in xnn_f32_gemm_ukernel_5x8s4__wasmrelaxedsimd_fma() local
187 …uiltin_wasm_fma_f32x4(vacc0x0123, wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0); in xnn_f32_gemm_ukernel_5x8s4__wasmrelaxedsimd_fma()
188 …uiltin_wasm_fma_f32x4(vacc1x0123, wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0); in xnn_f32_gemm_ukernel_5x8s4__wasmrelaxedsimd_fma()
189 …uiltin_wasm_fma_f32x4(vacc2x0123, wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0); in xnn_f32_gemm_ukernel_5x8s4__wasmrelaxedsimd_fma()
[all …]
|
| D | 5x8s4-relu-wasmrelaxedsimd-fma.c | 92 const v128_t vb0123c0 = wasm_v128_load(w + 0); in xnn_f32_gemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma() local
95 vacc0x0123 = __builtin_wasm_fma_f32x4(vacc0x0123, va0, vb0123c0); in xnn_f32_gemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma()
96 vacc1x0123 = __builtin_wasm_fma_f32x4(vacc1x0123, va1, vb0123c0); in xnn_f32_gemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma()
97 vacc2x0123 = __builtin_wasm_fma_f32x4(vacc2x0123, va2, vb0123c0); in xnn_f32_gemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma()
98 vacc3x0123 = __builtin_wasm_fma_f32x4(vacc3x0123, va3, vb0123c0); in xnn_f32_gemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma()
99 vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, va4, vb0123c0); in xnn_f32_gemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma()
184 const v128_t vb0123c0 = wasm_v128_load(w + 0); in xnn_f32_gemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma() local
187 …uiltin_wasm_fma_f32x4(vacc0x0123, wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0); in xnn_f32_gemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma()
188 …uiltin_wasm_fma_f32x4(vacc1x0123, wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0); in xnn_f32_gemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma()
189 …uiltin_wasm_fma_f32x4(vacc2x0123, wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0); in xnn_f32_gemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma()
[all …]
|
| D | 6x8s4-relu-wasmrelaxedsimd-fma.c | 102 const v128_t vb0123c0 = wasm_v128_load(w + 0); in xnn_f32_gemm_relu_ukernel_6x8s4__wasmrelaxedsimd_fma() local
105 vacc0x0123 = __builtin_wasm_fma_f32x4(vacc0x0123, va0, vb0123c0); in xnn_f32_gemm_relu_ukernel_6x8s4__wasmrelaxedsimd_fma()
106 vacc1x0123 = __builtin_wasm_fma_f32x4(vacc1x0123, va1, vb0123c0); in xnn_f32_gemm_relu_ukernel_6x8s4__wasmrelaxedsimd_fma()
107 vacc2x0123 = __builtin_wasm_fma_f32x4(vacc2x0123, va2, vb0123c0); in xnn_f32_gemm_relu_ukernel_6x8s4__wasmrelaxedsimd_fma()
108 vacc3x0123 = __builtin_wasm_fma_f32x4(vacc3x0123, va3, vb0123c0); in xnn_f32_gemm_relu_ukernel_6x8s4__wasmrelaxedsimd_fma()
109 vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, va4, vb0123c0); in xnn_f32_gemm_relu_ukernel_6x8s4__wasmrelaxedsimd_fma()
110 vacc5x0123 = __builtin_wasm_fma_f32x4(vacc5x0123, va5, vb0123c0); in xnn_f32_gemm_relu_ukernel_6x8s4__wasmrelaxedsimd_fma()
207 const v128_t vb0123c0 = wasm_v128_load(w + 0); in xnn_f32_gemm_relu_ukernel_6x8s4__wasmrelaxedsimd_fma() local
210 …uiltin_wasm_fma_f32x4(vacc0x0123, wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0); in xnn_f32_gemm_relu_ukernel_6x8s4__wasmrelaxedsimd_fma()
211 …uiltin_wasm_fma_f32x4(vacc1x0123, wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0); in xnn_f32_gemm_relu_ukernel_6x8s4__wasmrelaxedsimd_fma()
[all …]
|
| D | 6x8s4-wasmsimd.c | 102 const v128_t vb0123c0 = wasm_v128_load(w + 0); in xnn_f32_gemm_ukernel_6x8s4__wasmsimd() local
105 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c0)); in xnn_f32_gemm_ukernel_6x8s4__wasmsimd()
106 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c0)); in xnn_f32_gemm_ukernel_6x8s4__wasmsimd()
107 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c0)); in xnn_f32_gemm_ukernel_6x8s4__wasmsimd()
108 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123c0)); in xnn_f32_gemm_ukernel_6x8s4__wasmsimd()
109 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c0)); in xnn_f32_gemm_ukernel_6x8s4__wasmsimd()
110 vacc5x0123 = wasm_f32x4_add(vacc5x0123, wasm_f32x4_mul(va5, vb0123c0)); in xnn_f32_gemm_ukernel_6x8s4__wasmsimd()
207 const v128_t vb0123c0 = wasm_v128_load(w + 0); in xnn_f32_gemm_ukernel_6x8s4__wasmsimd() local
210 …4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0)); in xnn_f32_gemm_ukernel_6x8s4__wasmsimd()
211 …4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0)); in xnn_f32_gemm_ukernel_6x8s4__wasmsimd()
[all …]
|
| D | 6x8s4-wasmrelaxedsimd-fma.c | 102 const v128_t vb0123c0 = wasm_v128_load(w + 0); in xnn_f32_gemm_ukernel_6x8s4__wasmrelaxedsimd_fma() local
105 vacc0x0123 = __builtin_wasm_fma_f32x4(vacc0x0123, va0, vb0123c0); in xnn_f32_gemm_ukernel_6x8s4__wasmrelaxedsimd_fma()
106 vacc1x0123 = __builtin_wasm_fma_f32x4(vacc1x0123, va1, vb0123c0); in xnn_f32_gemm_ukernel_6x8s4__wasmrelaxedsimd_fma()
107 vacc2x0123 = __builtin_wasm_fma_f32x4(vacc2x0123, va2, vb0123c0); in xnn_f32_gemm_ukernel_6x8s4__wasmrelaxedsimd_fma()
108 vacc3x0123 = __builtin_wasm_fma_f32x4(vacc3x0123, va3, vb0123c0); in xnn_f32_gemm_ukernel_6x8s4__wasmrelaxedsimd_fma()
109 vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, va4, vb0123c0); in xnn_f32_gemm_ukernel_6x8s4__wasmrelaxedsimd_fma()
110 vacc5x0123 = __builtin_wasm_fma_f32x4(vacc5x0123, va5, vb0123c0); in xnn_f32_gemm_ukernel_6x8s4__wasmrelaxedsimd_fma()
207 const v128_t vb0123c0 = wasm_v128_load(w + 0); in xnn_f32_gemm_ukernel_6x8s4__wasmrelaxedsimd_fma() local
210 …uiltin_wasm_fma_f32x4(vacc0x0123, wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0); in xnn_f32_gemm_ukernel_6x8s4__wasmrelaxedsimd_fma()
211 …uiltin_wasm_fma_f32x4(vacc1x0123, wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0); in xnn_f32_gemm_ukernel_6x8s4__wasmrelaxedsimd_fma()
[all …]
|
| D | 6x8s4-minmax-wasmrelaxedsimd-fma.c | 104 const v128_t vb0123c0 = wasm_v128_load(w + 0); in xnn_f32_gemm_minmax_ukernel_6x8s4__wasmrelaxedsimd_fma() local
107 vacc0x0123 = __builtin_wasm_fma_f32x4(vacc0x0123, va0, vb0123c0); in xnn_f32_gemm_minmax_ukernel_6x8s4__wasmrelaxedsimd_fma()
108 vacc1x0123 = __builtin_wasm_fma_f32x4(vacc1x0123, va1, vb0123c0); in xnn_f32_gemm_minmax_ukernel_6x8s4__wasmrelaxedsimd_fma()
109 vacc2x0123 = __builtin_wasm_fma_f32x4(vacc2x0123, va2, vb0123c0); in xnn_f32_gemm_minmax_ukernel_6x8s4__wasmrelaxedsimd_fma()
110 vacc3x0123 = __builtin_wasm_fma_f32x4(vacc3x0123, va3, vb0123c0); in xnn_f32_gemm_minmax_ukernel_6x8s4__wasmrelaxedsimd_fma()
111 vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, va4, vb0123c0); in xnn_f32_gemm_minmax_ukernel_6x8s4__wasmrelaxedsimd_fma()
112 vacc5x0123 = __builtin_wasm_fma_f32x4(vacc5x0123, va5, vb0123c0); in xnn_f32_gemm_minmax_ukernel_6x8s4__wasmrelaxedsimd_fma()
209 const v128_t vb0123c0 = wasm_v128_load(w + 0); in xnn_f32_gemm_minmax_ukernel_6x8s4__wasmrelaxedsimd_fma() local
212 …uiltin_wasm_fma_f32x4(vacc0x0123, wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0); in xnn_f32_gemm_minmax_ukernel_6x8s4__wasmrelaxedsimd_fma()
213 …uiltin_wasm_fma_f32x4(vacc1x0123, wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0); in xnn_f32_gemm_minmax_ukernel_6x8s4__wasmrelaxedsimd_fma()
[all …]
|
| D | 4x8s4-wasmrelaxedsimd-fma.c | 82 const v128_t vb0123c0 = wasm_v128_load(w + 0); in xnn_f32_gemm_ukernel_4x8s4__wasmrelaxedsimd_fma() local
85 vacc0x0123 = __builtin_wasm_fma_f32x4(vacc0x0123, va0, vb0123c0); in xnn_f32_gemm_ukernel_4x8s4__wasmrelaxedsimd_fma()
86 vacc1x0123 = __builtin_wasm_fma_f32x4(vacc1x0123, va1, vb0123c0); in xnn_f32_gemm_ukernel_4x8s4__wasmrelaxedsimd_fma()
87 vacc2x0123 = __builtin_wasm_fma_f32x4(vacc2x0123, va2, vb0123c0); in xnn_f32_gemm_ukernel_4x8s4__wasmrelaxedsimd_fma()
88 vacc3x0123 = __builtin_wasm_fma_f32x4(vacc3x0123, va3, vb0123c0); in xnn_f32_gemm_ukernel_4x8s4__wasmrelaxedsimd_fma()
161 const v128_t vb0123c0 = wasm_v128_load(w + 0); in xnn_f32_gemm_ukernel_4x8s4__wasmrelaxedsimd_fma() local
164 …uiltin_wasm_fma_f32x4(vacc0x0123, wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0); in xnn_f32_gemm_ukernel_4x8s4__wasmrelaxedsimd_fma()
165 …uiltin_wasm_fma_f32x4(vacc1x0123, wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0); in xnn_f32_gemm_ukernel_4x8s4__wasmrelaxedsimd_fma()
166 …uiltin_wasm_fma_f32x4(vacc2x0123, wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0); in xnn_f32_gemm_ukernel_4x8s4__wasmrelaxedsimd_fma()
167 …uiltin_wasm_fma_f32x4(vacc3x0123, wasm_v128_andnot(va3, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0); in xnn_f32_gemm_ukernel_4x8s4__wasmrelaxedsimd_fma()
|
| D | 3x8s4-wasmrelaxedsimd-fma.c | 72 const v128_t vb0123c0 = wasm_v128_load(w + 0); in xnn_f32_gemm_ukernel_3x8s4__wasmrelaxedsimd_fma() local
75 vacc0x0123 = __builtin_wasm_fma_f32x4(vacc0x0123, va0, vb0123c0); in xnn_f32_gemm_ukernel_3x8s4__wasmrelaxedsimd_fma()
76 vacc1x0123 = __builtin_wasm_fma_f32x4(vacc1x0123, va1, vb0123c0); in xnn_f32_gemm_ukernel_3x8s4__wasmrelaxedsimd_fma()
77 vacc2x0123 = __builtin_wasm_fma_f32x4(vacc2x0123, va2, vb0123c0); in xnn_f32_gemm_ukernel_3x8s4__wasmrelaxedsimd_fma()
138 const v128_t vb0123c0 = wasm_v128_load(w + 0); in xnn_f32_gemm_ukernel_3x8s4__wasmrelaxedsimd_fma() local
141 …uiltin_wasm_fma_f32x4(vacc0x0123, wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0); in xnn_f32_gemm_ukernel_3x8s4__wasmrelaxedsimd_fma()
142 …uiltin_wasm_fma_f32x4(vacc1x0123, wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0); in xnn_f32_gemm_ukernel_3x8s4__wasmrelaxedsimd_fma()
143 …uiltin_wasm_fma_f32x4(vacc2x0123, wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0); in xnn_f32_gemm_ukernel_3x8s4__wasmrelaxedsimd_fma()
|
| D | 4x8s4-wasmsimd.c | 82 const v128_t vb0123c0 = wasm_v128_load(w + 0); in xnn_f32_gemm_ukernel_4x8s4__wasmsimd() local
85 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c0)); in xnn_f32_gemm_ukernel_4x8s4__wasmsimd()
86 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c0)); in xnn_f32_gemm_ukernel_4x8s4__wasmsimd()
87 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c0)); in xnn_f32_gemm_ukernel_4x8s4__wasmsimd()
88 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123c0)); in xnn_f32_gemm_ukernel_4x8s4__wasmsimd()
161 const v128_t vb0123c0 = wasm_v128_load(w + 0); in xnn_f32_gemm_ukernel_4x8s4__wasmsimd() local
164 …4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0)); in xnn_f32_gemm_ukernel_4x8s4__wasmsimd()
165 …4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0)); in xnn_f32_gemm_ukernel_4x8s4__wasmsimd()
166 …4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0)); in xnn_f32_gemm_ukernel_4x8s4__wasmsimd()
167 …4_add(vacc3x0123, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0)); in xnn_f32_gemm_ukernel_4x8s4__wasmsimd()
|
| D | 4x8s4-relu-wasmrelaxedsimd-fma.c | 82 const v128_t vb0123c0 = wasm_v128_load(w + 0); in xnn_f32_gemm_relu_ukernel_4x8s4__wasmrelaxedsimd_fma() local
85 vacc0x0123 = __builtin_wasm_fma_f32x4(vacc0x0123, va0, vb0123c0); in xnn_f32_gemm_relu_ukernel_4x8s4__wasmrelaxedsimd_fma()
86 vacc1x0123 = __builtin_wasm_fma_f32x4(vacc1x0123, va1, vb0123c0); in xnn_f32_gemm_relu_ukernel_4x8s4__wasmrelaxedsimd_fma()
87 vacc2x0123 = __builtin_wasm_fma_f32x4(vacc2x0123, va2, vb0123c0); in xnn_f32_gemm_relu_ukernel_4x8s4__wasmrelaxedsimd_fma()
88 vacc3x0123 = __builtin_wasm_fma_f32x4(vacc3x0123, va3, vb0123c0); in xnn_f32_gemm_relu_ukernel_4x8s4__wasmrelaxedsimd_fma()
161 const v128_t vb0123c0 = wasm_v128_load(w + 0); in xnn_f32_gemm_relu_ukernel_4x8s4__wasmrelaxedsimd_fma() local
164 …uiltin_wasm_fma_f32x4(vacc0x0123, wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0); in xnn_f32_gemm_relu_ukernel_4x8s4__wasmrelaxedsimd_fma()
165 …uiltin_wasm_fma_f32x4(vacc1x0123, wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0); in xnn_f32_gemm_relu_ukernel_4x8s4__wasmrelaxedsimd_fma()
166 …uiltin_wasm_fma_f32x4(vacc2x0123, wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0); in xnn_f32_gemm_relu_ukernel_4x8s4__wasmrelaxedsimd_fma()
167 …uiltin_wasm_fma_f32x4(vacc3x0123, wasm_v128_andnot(va3, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0); in xnn_f32_gemm_relu_ukernel_4x8s4__wasmrelaxedsimd_fma()
|
| D | 5x8s4-minmax-wasmrelaxedsimd.c | 94 const v128_t vb0123c0 = wasm_v128_load(w + 0); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmrelaxedsimd() local
97 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c0)); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmrelaxedsimd()
98 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c0)); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmrelaxedsimd()
99 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c0)); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmrelaxedsimd()
100 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123c0)); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmrelaxedsimd()
101 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c0)); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmrelaxedsimd()
186 const v128_t vb0123c0 = wasm_v128_load(w + 0); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmrelaxedsimd() local
189 …4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0)); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmrelaxedsimd()
190 …4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0)); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmrelaxedsimd()
191 …4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0)); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmrelaxedsimd()
[all …]
|
| D | 5x8s4-minmax-wasmsimd-arm.c | 94 const v128_t vb0123c0 = wasm_v128_load(w + 0); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmsimd_arm() local
97 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c0)); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmsimd_arm()
98 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c0)); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmsimd_arm()
99 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c0)); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmsimd_arm()
100 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123c0)); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmsimd_arm()
101 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c0)); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmsimd_arm()
186 const v128_t vb0123c0 = wasm_v128_load(w + 0); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmsimd_arm() local
189 …4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0)); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmsimd_arm()
190 …4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0)); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmsimd_arm()
191 …4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0)); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmsimd_arm()
[all …]
|
| D | 5x8s4-relu-wasmsimd.c | 92 const v128_t vb0123c0 = wasm_v128_load(w + 0); in xnn_f32_gemm_relu_ukernel_5x8s4__wasmsimd() local
95 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c0)); in xnn_f32_gemm_relu_ukernel_5x8s4__wasmsimd()
96 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c0)); in xnn_f32_gemm_relu_ukernel_5x8s4__wasmsimd()
97 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c0)); in xnn_f32_gemm_relu_ukernel_5x8s4__wasmsimd()
98 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123c0)); in xnn_f32_gemm_relu_ukernel_5x8s4__wasmsimd()
99 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c0)); in xnn_f32_gemm_relu_ukernel_5x8s4__wasmsimd()
184 const v128_t vb0123c0 = wasm_v128_load(w + 0); in xnn_f32_gemm_relu_ukernel_5x8s4__wasmsimd() local
187 …4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0)); in xnn_f32_gemm_relu_ukernel_5x8s4__wasmsimd()
188 …4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0)); in xnn_f32_gemm_relu_ukernel_5x8s4__wasmsimd()
189 …4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0)); in xnn_f32_gemm_relu_ukernel_5x8s4__wasmsimd()
[all …]
|
| /external/XNNPACK/src/f32-gemm/gen-inc/ |
| D | 5x8s4inc-minmax-wasmsimd-x86.c | 96 const v128_t vb0123c0 = wasm_v128_load(w + 0); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmsimd_x86() local
99 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c0)); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmsimd_x86()
100 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c0)); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmsimd_x86()
101 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c0)); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmsimd_x86()
102 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123c0)); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmsimd_x86()
103 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c0)); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmsimd_x86()
188 const v128_t vb0123c0 = wasm_v128_load(w + 0); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmsimd_x86() local
191 …4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0)); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmsimd_x86()
192 …4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0)); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmsimd_x86()
193 …4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0)); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmsimd_x86()
[all …]
|
| D | 5x8s4inc-minmax-wasmrelaxedsimd-fma.c | 96 const v128_t vb0123c0 = wasm_v128_load(w + 0); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmrelaxedsimd_fma() local
99 vacc0x0123 = __builtin_wasm_fma_f32x4(vacc0x0123, va0, vb0123c0); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmrelaxedsimd_fma()
100 vacc1x0123 = __builtin_wasm_fma_f32x4(vacc1x0123, va1, vb0123c0); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmrelaxedsimd_fma()
101 vacc2x0123 = __builtin_wasm_fma_f32x4(vacc2x0123, va2, vb0123c0); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmrelaxedsimd_fma()
102 vacc3x0123 = __builtin_wasm_fma_f32x4(vacc3x0123, va3, vb0123c0); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmrelaxedsimd_fma()
103 vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, va4, vb0123c0); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmrelaxedsimd_fma()
188 const v128_t vb0123c0 = wasm_v128_load(w + 0); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmrelaxedsimd_fma() local
191 …uiltin_wasm_fma_f32x4(vacc0x0123, wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmrelaxedsimd_fma()
192 …uiltin_wasm_fma_f32x4(vacc1x0123, wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmrelaxedsimd_fma()
193 …uiltin_wasm_fma_f32x4(vacc2x0123, wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmrelaxedsimd_fma()
[all …]
|
| D | 5x8s4inc-minmax-wasmsimd-arm.c | 96 const v128_t vb0123c0 = wasm_v128_load(w + 0); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmsimd_arm() local
99 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c0)); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmsimd_arm()
100 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c0)); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmsimd_arm()
101 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c0)); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmsimd_arm()
102 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123c0)); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmsimd_arm()
103 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c0)); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmsimd_arm()
188 const v128_t vb0123c0 = wasm_v128_load(w + 0); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmsimd_arm() local
191 …4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0)); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmsimd_arm()
192 …4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0)); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmsimd_arm()
193 …4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0)); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmsimd_arm()
[all …]
|