Star983
1134
代码
代码
Issues99
Pull Requests285
Actions
项目讨论
Wiki
分析
Star983
1134
已合并
C API新增l0c2l1矩阵搬出接口 #5091
UFOunder创建于 8月14日
C API新增l0c2l1矩阵搬出接口 #5091
已合并
UFOunder创建于 8月14日
从UFOunder/asc-devkit_1828master合入到cann/asc-devkitmaster
共 4 个文件变更+121-0
Mimpl/c_api/instr_impl/npu_arch_3510/cube_datamove_impl.h+27-0
@@ -2591,6 +2591,33 @@ __aicore__ inline void asc_copy_l0c2l1_sync(
2591 eltwise_op, eltwise_antq_en, c0_pad_en, broadcast_en, enable_nz2dn);2591 eltwise_op, eltwise_antq_en, c0_pad_en, broadcast_en, enable_nz2dn);
2592}2592}
2593 2593 
2594+// half float
W
Wwuzhaolin8月14日

这个注释没有真实意义，建议删掉，添加有效注释

likedislike
U
UFOunder
8月14日 评论：
2595+__aicore__ inline void asc_copy_l0c2l1(
2596+ __cbuf__ half* dst, __cc__ float* src, uint16_t n_size, uint16_t m_size, uint32_t dst_stride, uint16_t src_stride,
2597+ uint8_t l2_cache_mode, uint8_t enable_clip_relu_pre, uint8_t unit_flag_mode, uint64_t quant_pre_mode,
W
Wwuzhaolin8月14日

如果是新增的， 是不是一把把uint8_t改枚举的东东干上去

likedislike
U
UFOunder
8月14日 评论：
2598+ uint8_t relu_pre_mode, bool enable_channel_split, bool enable_nz2nd, uint64_t quant_post, uint8_t relu_post,
2599+ bool clip_relu_post, uint8_t eltwise_op, bool eltwise_antq_en, bool c0_pad_en, bool broadcast_en, bool enable_nz2dn)
2600+{
2601+ asc_copy_l0c2l1_impl(
2602+ dst, src, n_size, m_size, dst_stride, src_stride, l2_cache_mode, enable_clip_relu_pre, unit_flag_mode,
2603+ quant_pre_mode, relu_pre_mode, enable_channel_split, enable_nz2nd, quant_post, relu_post, clip_relu_post,
2604+ eltwise_op, eltwise_antq_en, c0_pad_en, broadcast_en, enable_nz2dn);
2605+}
2606+ 
2607+// bfloat16_t float
2608+__aicore__ inline void asc_copy_l0c2l1(
2609+ __cbuf__ bfloat16_t* dst, __cc__ float* src, uint16_t n_size, uint16_t m_size, uint32_t dst_stride,
2610+ uint16_t src_stride, uint8_t l2_cache_mode, uint8_t enable_clip_relu_pre, uint8_t unit_flag_mode,
2611+ uint64_t quant_pre_mode, uint8_t relu_pre_mode, bool enable_channel_split, bool enable_nz2nd, uint64_t quant_post,
2612+ uint8_t relu_post, bool clip_relu_post, uint8_t eltwise_op, bool eltwise_antq_en, bool c0_pad_en, bool broadcast_en,
2613+ bool enable_nz2dn)
2614+{
2615+ asc_copy_l0c2l1_impl(
2616+ dst, src, n_size, m_size, dst_stride, src_stride, l2_cache_mode, enable_clip_relu_pre, unit_flag_mode,
2617+ quant_pre_mode, relu_pre_mode, enable_channel_split, enable_nz2nd, quant_post, relu_post, clip_relu_post,
2618+ eltwise_op, eltwise_antq_en, c0_pad_en, broadcast_en, enable_nz2dn);
2619+}
2620+ 
2594// ==========asc_copy_l0c2gm===========2621// ==========asc_copy_l0c2gm===========
2595// bfloat16_t float2622// bfloat16_t float
2596__aicore__ inline void asc_copy_l0c2gm(2623__aicore__ inline void asc_copy_l0c2gm(
Mimpl/c_api/instr_impl/npu_arch_3510/cube_datamove_impl/asc_copy_l0c2l1_impl.h+33-0
@@ -226,6 +226,39 @@ __aicore__ inline void asc_copy_l0c2l1_sync_impl(
226 asc_sync_post_process();226 asc_sync_post_process();
227}227}
228 228 
229+// half float
230+__aicore__ inline void asc_copy_l0c2l1_impl(
231+ __cbuf__ half* dst, __cc__ float* src, uint16_t n_size, uint16_t m_size, uint32_t dst_stride, uint16_t src_stride,
232+ uint8_t l2_cache_mode, uint8_t enable_clip_relu_pre, uint8_t unit_flag_mode, uint64_t quant_pre_mode,
233+ uint8_t relu_pre_mode, bool enable_channel_split, bool enable_nz2nd, uint64_t quant_post, uint8_t relu_post,
234+ bool clip_relu_post, uint8_t eltwise_op, bool eltwise_antq_en, bool c0_pad_en, bool broadcast_en, bool enable_nz2dn)
235+{
236+ if ASC_IS_AIC {
237+ copy_matrix_cc_to_cbuf(
238+ dst, src, 0, n_size, m_size, dst_stride, src_stride, l2_cache_mode, enable_clip_relu_pre, unit_flag_mode,
239+ static_cast<QuantMode_t>(quant_pre_mode), relu_pre_mode, enable_channel_split, enable_nz2nd,
240+ static_cast<QuantMode_post>(quant_post), relu_post, clip_relu_post, false, eltwise_op, eltwise_antq_en,
241+ false, c0_pad_en, false, broadcast_en, enable_nz2dn);
242+ }
243+}
244+ 
245+// bfloat16_t float
246+__aicore__ inline void asc_copy_l0c2l1_impl(
247+ __cbuf__ bfloat16_t* dst, __cc__ float* src, uint16_t n_size, uint16_t m_size, uint32_t dst_stride,
248+ uint16_t src_stride, uint8_t l2_cache_mode, uint8_t enable_clip_relu_pre, uint8_t unit_flag_mode,
249+ uint64_t quant_pre_mode, uint8_t relu_pre_mode, bool enable_channel_split, bool enable_nz2nd, uint64_t quant_post,
250+ uint8_t relu_post, bool clip_relu_post, uint8_t eltwise_op, bool eltwise_antq_en, bool c0_pad_en, bool broadcast_en,
251+ bool enable_nz2dn)
252+{
253+ if ASC_IS_AIC {
254+ copy_matrix_cc_to_cbuf(
255+ dst, src, 0, n_size, m_size, dst_stride, src_stride, l2_cache_mode, enable_clip_relu_pre, unit_flag_mode,
256+ static_cast<QuantMode_t>(quant_pre_mode), relu_pre_mode, enable_channel_split, enable_nz2nd,
257+ static_cast<QuantMode_post>(quant_post), relu_post, clip_relu_post, false, eltwise_op, eltwise_antq_en,
258+ false, c0_pad_en, false, broadcast_en, enable_nz2dn);
259+ }
260+}
261+ 
229#endif262#endif
230 263 
231#if defined(UNDEF_ASCENDC_C_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC)264#if defined(UNDEF_ASCENDC_C_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC)
Minclude/c_api/cube_datamove/cube_datamove.h+16-0
@@ -1886,6 +1886,22 @@ __aicore__ inline void asc_copy_l0c2l1_sync(
1886 uint8_t relu_post, bool clip_relu_post, uint8_t eltwise_op, bool eltwise_antq_en, bool c0_pad_en, bool broadcast_en,1886 uint8_t relu_post, bool clip_relu_post, uint8_t eltwise_op, bool eltwise_antq_en, bool c0_pad_en, bool broadcast_en,
1887 bool enable_nz2dn);1887 bool enable_nz2dn);
1888 1888 
1889+// half float
1890+__aicore__ inline void asc_copy_l0c2l1(
1891+ __cbuf__ half* dst, __cc__ float* src, uint16_t n_size, uint16_t m_size, uint32_t dst_stride, uint16_t src_stride,
1892+ uint8_t l2_cache_mode, uint8_t enable_clip_relu_pre, uint8_t unit_flag_mode, uint64_t quant_pre_mode,
1893+ uint8_t relu_pre_mode, bool enable_channel_split, bool enable_nz2nd, uint64_t quant_post, uint8_t relu_post,
1894+ bool clip_relu_post, uint8_t eltwise_op, bool eltwise_antq_en, bool c0_pad_en, bool broadcast_en,
1895+ bool enable_nz2dn);
1896+ 
1897+// bfloat16_t float
1898+__aicore__ inline void asc_copy_l0c2l1(
1899+ __cbuf__ bfloat16_t* dst, __cc__ float* src, uint16_t n_size, uint16_t m_size, uint32_t dst_stride,
1900+ uint16_t src_stride, uint8_t l2_cache_mode, uint8_t enable_clip_relu_pre, uint8_t unit_flag_mode,
1901+ uint64_t quant_pre_mode, uint8_t relu_pre_mode, bool enable_channel_split, bool enable_nz2nd, uint64_t quant_post,
1902+ uint8_t relu_post, bool clip_relu_post, uint8_t eltwise_op, bool eltwise_antq_en, bool c0_pad_en, bool broadcast_en,
1903+ bool enable_nz2dn);
1904+ 
1889// ==========asc_copy_l0c2gm===========1905// ==========asc_copy_l0c2gm===========
1890// bfloat16_t float1906// bfloat16_t float
1891__aicore__ inline void asc_copy_l0c2gm(1907__aicore__ inline void asc_copy_l0c2gm(
Mtests/api/c_api/npu_arch_3510/cube_datamove/test_copy_l0c2l1_instr.cpp+45-0
@@ -142,6 +142,44 @@
142 GlobalMockObject::verify(); \142 GlobalMockObject::verify(); \
143 }143 }
144 144 
145+class TestCubeDatamoveL0C2L1Full : public testing::Test {
146+protected:
147+ void SetUp() { g_coreType = C_API_AIC_TYPE; }
148+ void TearDown() { g_coreType = C_API_AIV_TYPE; }
149+};
150+ 
151+template <typename DstType>
152+void TestAscCopyL0C2L1FullParams()
153+{
154+ __cbuf__ DstType* dst = reinterpret_cast<__cbuf__ DstType*>(1);
155+ __cc__ float* src = reinterpret_cast<__cc__ float*>(2);
156+ uint16_t n_size = 3;
157+ uint16_t m_size = 4;
158+ uint32_t dst_stride = 6;
159+ uint16_t src_stride = 7;
160+ uint8_t l2_cache_mode = 0;
161+ uint8_t enable_clip_relu_pre = 8;
162+ uint8_t unit_flag_mode = 10;
163+ uint64_t quant_pre_mode = 11;
164+ uint8_t relu_pre_mode = 12;
165+ bool enable_channel_split = false;
166+ bool enable_nz2nd = false;
167+ uint64_t quant_post = 13;
168+ uint8_t relu_post = 14;
169+ bool clip_relu_post = true;
170+ uint8_t eltwise_op = 15;
171+ bool eltwise_antq_en = true;
172+ bool c0_pad_en = true;
173+ bool broadcast_en = false;
174+ bool enable_nz2dn = true;
175+ 
176+ asc_copy_l0c2l1(
177+ dst, src, n_size, m_size, dst_stride, src_stride, l2_cache_mode, enable_clip_relu_pre, unit_flag_mode,
178+ quant_pre_mode, relu_pre_mode, enable_channel_split, enable_nz2nd, quant_post, relu_post, clip_relu_post,
179+ eltwise_op, eltwise_antq_en, c0_pad_en, broadcast_en, enable_nz2dn);
180+ GlobalMockObject::verify();
181+}
182+ 
145#endif183#endif
146 184 
147TEST_CUBE_COMPUTE_FIXPIPE_L0C2L1_INSTR(L0C2L1, asc_copy_l0c2l1, copy_matrix_cc_to_cbuf, half, float);185TEST_CUBE_COMPUTE_FIXPIPE_L0C2L1_INSTR(L0C2L1, asc_copy_l0c2l1, copy_matrix_cc_to_cbuf, half, float);
@@ -156,3 +194,10 @@ TEST_CUBE_COMPUTE_FIXPIPE_L0C2L1_INSTR(L0C2L1, asc_copy_l0c2l1, copy_matrix_cc_t
156 194 
157TEST_CUBE_COMPUTE_FIXPIPE_L0C2L1_S4_INSTR(L0C2L1, asc_copy_l0c2l1, copy_matrix_cc_to_cbuf_s4, int4b_t, float);195TEST_CUBE_COMPUTE_FIXPIPE_L0C2L1_S4_INSTR(L0C2L1, asc_copy_l0c2l1, copy_matrix_cc_to_cbuf_s4, int4b_t, float);
158TEST_CUBE_COMPUTE_FIXPIPE_L0C2L1_S4_INSTR(L0C2L1, asc_copy_l0c2l1, copy_matrix_cc_to_cbuf_s4, int4b_t, int32_t);196TEST_CUBE_COMPUTE_FIXPIPE_L0C2L1_S4_INSTR(L0C2L1, asc_copy_l0c2l1, copy_matrix_cc_to_cbuf_s4, int4b_t, int32_t);
197+ 
198+TEST_F(TestCubeDatamoveL0C2L1Full, asc_copy_l0c2l1_half_float_full_params_Succ) { TestAscCopyL0C2L1FullParams<half>(); }
199+ 
200+TEST_F(TestCubeDatamoveL0C2L1Full, asc_copy_l0c2l1_bfloat16_t_float_full_params_Succ)
201+{
202+ TestAscCopyL0C2L1FullParams<bfloat16_t>();
203+}