| /external/XNNPACK/src/f32-dwconv/gen/ |
| D | up8x9-wasmrelaxedsimd-fma.c | 154 const v128_t vk7x0123 = wasm_v128_load(w + 64); in xnn_f32_dwconv_ukernel_up8x9__wasmrelaxedsimd_fma() local
156 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi7x0123, vk7x0123); in xnn_f32_dwconv_ukernel_up8x9__wasmrelaxedsimd_fma()
226 const v128_t vk7x0123 = wasm_v128_load(w + 64); in xnn_f32_dwconv_ukernel_up8x9__wasmrelaxedsimd_fma() local
227 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi7x0123, vk7x0123); in xnn_f32_dwconv_ukernel_up8x9__wasmrelaxedsimd_fma()
275 const v128_t vk7x0123 = wasm_v128_load(w + 64); in xnn_f32_dwconv_ukernel_up8x9__wasmrelaxedsimd_fma() local
276 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi7x0123, vk7x0123); in xnn_f32_dwconv_ukernel_up8x9__wasmrelaxedsimd_fma()
|
| D | up8x9-minmax-wasmrelaxedsimd-fma.c | 156 const v128_t vk7x0123 = wasm_v128_load(w + 64); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmrelaxedsimd_fma() local
158 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi7x0123, vk7x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmrelaxedsimd_fma()
231 const v128_t vk7x0123 = wasm_v128_load(w + 64); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmrelaxedsimd_fma() local
232 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi7x0123, vk7x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmrelaxedsimd_fma()
281 const v128_t vk7x0123 = wasm_v128_load(w + 64); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmrelaxedsimd_fma() local
282 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi7x0123, vk7x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmrelaxedsimd_fma()
|
| D | up8x9-minmax-sse.c | 156 const __m128 vk7x0123 = _mm_load_ps(w + 64); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse() local
158 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse()
230 const __m128 vk7x0123 = _mm_load_ps(w + 64); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse() local
231 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse()
280 const __m128 vk7x0123 = _mm_load_ps(w + 64); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse() local
281 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse()
|
| D | up8x9-minmax-wasmsimd-arm.c | 156 const v128_t vk7x0123 = wasm_v128_load(w + 64); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm() local
158 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm()
231 const v128_t vk7x0123 = wasm_v128_load(w + 64); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm() local
232 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm()
281 const v128_t vk7x0123 = wasm_v128_load(w + 64); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm() local
282 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm()
|
| D | up8x9-minmax-wasmrelaxedsimd-acc2.c | 156 const v128_t vk7x0123 = wasm_v128_load(w + 64); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmrelaxedsimd_acc2() local
158 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmrelaxedsimd_acc2()
234 const v128_t vk7x0123 = wasm_v128_load(w + 64); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmrelaxedsimd_acc2() local
235 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmrelaxedsimd_acc2()
286 const v128_t vk7x0123 = wasm_v128_load(w + 64); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmrelaxedsimd_acc2() local
287 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmrelaxedsimd_acc2()
|
| D | up8x9-wasmsimd.c | 154 const v128_t vk7x0123 = wasm_v128_load(w + 64); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd() local
156 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd()
226 const v128_t vk7x0123 = wasm_v128_load(w + 64); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd() local
227 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd()
275 const v128_t vk7x0123 = wasm_v128_load(w + 64); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd() local
276 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd()
|
| D | up8x9-minmax-neonfma.c | 141 const float32x4_t vk7x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma() local
143 vacc0123p0 = vfmaq_f32(vacc0123p0, vi7x0123, vk7x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma()
195 const float32x4_t vk7x0123 = vld1q_f32(w + 60); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma() local
196 vacc0123p0 = vfmaq_f32(vacc0123p0, vi7x0123, vk7x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma()
241 const float32x4_t vk7x0123 = vld1q_f32(w + 64); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma() local
242 vacc0123p0 = vfmaq_f32(vacc0123p0, vi7x0123, vk7x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma()
|
| D | up8x9-minmax-neon-acc2.c | 141 const float32x4_t vk7x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2() local
143 vacc0123p1 = vmlaq_f32(vacc0123p1, vi7x0123, vk7x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2()
198 const float32x4_t vk7x0123 = vld1q_f32(w + 60); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2() local
199 vacc0123p1 = vmlaq_f32(vacc0123p1, vi7x0123, vk7x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2()
246 const float32x4_t vk7x0123 = vld1q_f32(w + 64); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2() local
247 vacc0123p1 = vmlaq_f32(vacc0123p1, vi7x0123, vk7x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2()
|
| D | up8x9-minmax-neon.c | 141 const float32x4_t vk7x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neon() local
143 vacc0123p0 = vmlaq_f32(vacc0123p0, vi7x0123, vk7x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon()
195 const float32x4_t vk7x0123 = vld1q_f32(w + 60); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon() local
196 vacc0123p0 = vmlaq_f32(vacc0123p0, vi7x0123, vk7x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon()
241 const float32x4_t vk7x0123 = vld1q_f32(w + 64); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon() local
242 vacc0123p0 = vmlaq_f32(vacc0123p0, vi7x0123, vk7x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon()
|
| D | up8x9-minmax-wasmsimd-x86-acc2.c | 156 const v128_t vk7x0123 = wasm_v128_load(w + 64); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2() local
158 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2()
234 const v128_t vk7x0123 = wasm_v128_load(w + 64); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2() local
235 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2()
286 const v128_t vk7x0123 = wasm_v128_load(w + 64); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2() local
287 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2()
|
| D | up8x9-minmax-wasmsimd-arm-acc2.c | 156 const v128_t vk7x0123 = wasm_v128_load(w + 64); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2() local
158 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2()
234 const v128_t vk7x0123 = wasm_v128_load(w + 64); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2() local
235 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2()
286 const v128_t vk7x0123 = wasm_v128_load(w + 64); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2() local
287 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2()
|
| D | up8x9-minmax-wasmsimd-x86.c | 156 const v128_t vk7x0123 = wasm_v128_load(w + 64); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86() local
158 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86()
231 const v128_t vk7x0123 = wasm_v128_load(w + 64); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86() local
232 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86()
281 const v128_t vk7x0123 = wasm_v128_load(w + 64); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86() local
282 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86()
|
| D | up8x9-minmax-wasmrelaxedsimd-fma-acc2.c | 156 const v128_t vk7x0123 = wasm_v128_load(w + 64); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmrelaxedsimd_fma_acc2() local
158 vacc0123p1 = __builtin_wasm_fma_f32x4(vacc0123p1, vi7x0123, vk7x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmrelaxedsimd_fma_acc2()
234 const v128_t vk7x0123 = wasm_v128_load(w + 64); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmrelaxedsimd_fma_acc2() local
235 vacc0123p1 = __builtin_wasm_fma_f32x4(vacc0123p1, vi7x0123, vk7x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmrelaxedsimd_fma_acc2()
286 const v128_t vk7x0123 = wasm_v128_load(w + 64); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmrelaxedsimd_fma_acc2() local
287 vacc0123p1 = __builtin_wasm_fma_f32x4(vacc0123p1, vi7x0123, vk7x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmrelaxedsimd_fma_acc2()
|
| D | up8x9-minmax-wasmrelaxedsimd.c | 156 const v128_t vk7x0123 = wasm_v128_load(w + 64); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmrelaxedsimd() local
158 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmrelaxedsimd()
231 const v128_t vk7x0123 = wasm_v128_load(w + 64); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmrelaxedsimd() local
232 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmrelaxedsimd()
281 const v128_t vk7x0123 = wasm_v128_load(w + 64); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmrelaxedsimd() local
282 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmrelaxedsimd()
|
| D | up8x9-minmax-neonfma-acc2.c | 141 const float32x4_t vk7x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2() local
143 vacc0123p1 = vfmaq_f32(vacc0123p1, vi7x0123, vk7x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2()
198 const float32x4_t vk7x0123 = vld1q_f32(w + 60); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2() local
199 vacc0123p1 = vfmaq_f32(vacc0123p1, vi7x0123, vk7x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2()
246 const float32x4_t vk7x0123 = vld1q_f32(w + 64); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2() local
247 vacc0123p1 = vfmaq_f32(vacc0123p1, vi7x0123, vk7x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2()
|
| D | up8x9-minmax-sse-acc2.c | 156 const __m128 vk7x0123 = _mm_load_ps(w + 64); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2() local
158 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2()
233 const __m128 vk7x0123 = _mm_load_ps(w + 64); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2() local
234 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2()
285 const __m128 vk7x0123 = _mm_load_ps(w + 64); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2() local
286 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2()
|
| D | up8x9-wasmsimd-acc2.c | 154 const v128_t vk7x0123 = wasm_v128_load(w + 64); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd_acc2() local
156 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd_acc2()
229 const v128_t vk7x0123 = wasm_v128_load(w + 64); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd_acc2() local
230 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd_acc2()
280 const v128_t vk7x0123 = wasm_v128_load(w + 64); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd_acc2() local
281 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd_acc2()
|
| D | up4x9-wasmrelaxedsimd-fma.c | 131 const v128_t vk7x0123 = wasm_v128_load(w + 32); in xnn_f32_dwconv_ukernel_up4x9__wasmrelaxedsimd_fma() local
132 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi7x0123, vk7x0123); in xnn_f32_dwconv_ukernel_up4x9__wasmrelaxedsimd_fma()
180 const v128_t vk7x0123 = wasm_v128_load(w + 32); in xnn_f32_dwconv_ukernel_up4x9__wasmrelaxedsimd_fma() local
181 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi7x0123, vk7x0123); in xnn_f32_dwconv_ukernel_up4x9__wasmrelaxedsimd_fma()
|
| D | up4x9-minmax-neon.c | 118 const float32x4_t vk7x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neon() local
119 vacc0123p0 = vmlaq_f32(vacc0123p0, vi7x0123, vk7x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__neon()
164 const float32x4_t vk7x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neon() local
165 vacc0123p0 = vmlaq_f32(vacc0123p0, vi7x0123, vk7x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__neon()
|
| D | up4x9-minmax-wasmsimd-arm.c | 133 const v128_t vk7x0123 = wasm_v128_load(w + 32); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm() local
134 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm()
184 const v128_t vk7x0123 = wasm_v128_load(w + 32); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm() local
185 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm()
|
| D | up4x9-minmax-wasmrelaxedsimd-fma.c | 133 const v128_t vk7x0123 = wasm_v128_load(w + 32); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmrelaxedsimd_fma() local
134 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi7x0123, vk7x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmrelaxedsimd_fma()
184 const v128_t vk7x0123 = wasm_v128_load(w + 32); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmrelaxedsimd_fma() local
185 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi7x0123, vk7x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmrelaxedsimd_fma()
|
| D | up4x9-minmax-wasmrelaxedsimd.c | 133 const v128_t vk7x0123 = wasm_v128_load(w + 32); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmrelaxedsimd() local
134 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmrelaxedsimd()
184 const v128_t vk7x0123 = wasm_v128_load(w + 32); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmrelaxedsimd() local
185 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmrelaxedsimd()
|
| D | up4x9-minmax-wasmrelaxedsimd-fma-acc2.c | 133 const v128_t vk7x0123 = wasm_v128_load(w + 32); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmrelaxedsimd_fma_acc2() local
134 vacc0123p1 = __builtin_wasm_fma_f32x4(vacc0123p1, vi7x0123, vk7x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmrelaxedsimd_fma_acc2()
186 const v128_t vk7x0123 = wasm_v128_load(w + 32); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmrelaxedsimd_fma_acc2() local
187 vacc0123p1 = __builtin_wasm_fma_f32x4(vacc0123p1, vi7x0123, vk7x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmrelaxedsimd_fma_acc2()
|
| D | up4x9-wasmsimd-acc2.c | 131 const v128_t vk7x0123 = wasm_v128_load(w + 32); in xnn_f32_dwconv_ukernel_up4x9__wasmsimd_acc2() local
132 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_ukernel_up4x9__wasmsimd_acc2()
182 const v128_t vk7x0123 = wasm_v128_load(w + 32); in xnn_f32_dwconv_ukernel_up4x9__wasmsimd_acc2() local
183 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_ukernel_up4x9__wasmsimd_acc2()
|
| D | up4x9-wasmsimd.c | 131 const v128_t vk7x0123 = wasm_v128_load(w + 32); in xnn_f32_dwconv_ukernel_up4x9__wasmsimd() local
132 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_ukernel_up4x9__wasmsimd()
180 const v128_t vk7x0123 = wasm_v128_load(w + 32); in xnn_f32_dwconv_ukernel_up4x9__wasmsimd() local
181 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_ukernel_up4x9__wasmsimd()
|