zq/fix_declaration_for_some_func (DeepLink-org#1141)

* fix_declaration_for_some_func * fix gen_output and diopi_functions for optimizer_funcs * reformat customized_test
DoorKickers · Apr 16, 2024 · 510359e · 510359e
1 parent b4126ce
commit 510359e
Show file tree

Hide file tree

Showing 14 changed files with 945 additions and 779 deletions.
diff --git a/diopi_test/python/conformance/customized_test.py b/diopi_test/python/conformance/customized_test.py
diff --git a/diopi_test/python/conformance/diopi_functions.py b/diopi_test/python/conformance/diopi_functions.py
diff --git a/diopi_test/python/conformance/gen_output.py b/diopi_test/python/conformance/gen_output.py
diff --git a/impl/ascend_npu/diopi_impl/functions_ext/adamw.cpp b/impl/ascend_npu/diopi_impl/functions_ext/adamw.cpp
@@ -12,7 +12,9 @@
 
 namespace OP_IMPL_NS {
 
-diopiError_t diopiAdamW(diopiContextHandle_t ctx, diopiTensorHandle_t input, diopiTensorHandle_t grad, diopiTensorHandle_t expAvg, diopiTensorHandle_t expAvgSq,
+diopiError_t diopiAdamW(diopiContextHandle_t ctx, diopiTensorHandle_t input, diopiConstTensorHandle_t grad, diopiTensorHandle_t expAvg,
+                        diopiTensorHandle_t expAvgSq,
+
                         diopiTensorHandle_t maxExpAvgSq, float lr, float beta1, float beta2, float eps, float weightDecay, int64_t step, bool amsgrad) {
     DIOPI_CHECK(amsgrad == false, "at present, ApplyAdamW only supports amsgrad false on ascend.");
     BEGIN_CALL_ACL_OP(input, grad, expAvg, expAvgSq, maxExpAvgSq);

diff --git a/impl/camb/device_configs.py b/impl/camb/device_configs.py
@@ -1110,15 +1110,15 @@
     'adadelta': dict(
         name=["adadelta"],
         atol_half=1e-1,
-        rtol_half=1e-3,
+        rtol_half=1e-1,
         atol=1e-1,
         rtol=1e-3,
         tensor_para=dict(
             args=[
                 {
                     # can't get correct result
                     "ins": ['param', 'param_grad'],
-                    "dtype": [Skip(np.float16)],
+                    "dtype": [Skip(np.float64)],
                 },
             ]
         ),

diff --git a/impl/camb/functions/adadelta.cpp b/impl/camb/functions/adadelta.cpp
@@ -4,7 +4,7 @@
 namespace impl {
 namespace camb {
 
-diopiError_t diopiAdadelta(diopiContextHandle_t ctx, diopiTensorHandle_t input, diopiTensorHandle_t grad, diopiTensorHandle_t squareAvg,
+diopiError_t diopiAdadelta(diopiContextHandle_t ctx, diopiTensorHandle_t input, diopiConstTensorHandle_t grad, diopiTensorHandle_t squareAvg,
                            diopiTensorHandle_t accDelta, float lr, float rho, float eps, float weightDecay) {
     cnnlHandle_t handle = cnnlHandlePool.get(ctx);
     DiopiTensor inputTensor = DiopiTensor(input);

diff --git a/impl/camb/functions/adam.cpp b/impl/camb/functions/adam.cpp
@@ -16,9 +16,9 @@ void bangAdamInternal(void* grad, void* m, void* v, void* vMax, void* variable,
                       float learningRateCorrection, int adamMode, float decay, float decayCorrection, cnrtDim3_t kDim, cnrtFunctionType_t kType,
                       cnrtQueue_t queue, cnrtDataType_t cnrtType, bool amsgrad);
 
-diopiError_t bangAdam(diopiContextHandle_t ctx, diopiTensorHandle_t input, diopiTensorHandle_t grad, diopiTensorHandle_t expAvg, diopiTensorHandle_t expAvgSq,
-                      diopiTensorHandle_t maxExpAvgSq, float lr, float beta1, float beta2, float eps, float weightDecay, int64_t step, bool amsgrad,
-                      int adamMode = 0) {
+diopiError_t bangAdam(diopiContextHandle_t ctx, diopiTensorHandle_t input, diopiConstTensorHandle_t grad, diopiTensorHandle_t expAvg,
+                      diopiTensorHandle_t expAvgSq, diopiTensorHandle_t maxExpAvgSq, float lr, float beta1, float beta2, float eps, float weightDecay,
+                      int64_t step, bool amsgrad, int adamMode = 0) {
     cnrtQueue_t queue = getStream(ctx);
     DiopiTensor inputTensor = DiopiTensor(input);
     DiopiTensor gradTensor = DiopiTensor(grad);
@@ -92,13 +92,16 @@ diopiError_t bangAdam(diopiContextHandle_t ctx, diopiTensorHandle_t input, diopi
     return diopiSuccess;
 }
 
-diopiError_t diopiAdam(diopiContextHandle_t ctx, diopiTensorHandle_t input, diopiTensorHandle_t grad, diopiTensorHandle_t expAvg, diopiTensorHandle_t expAvgSq,
-                       diopiTensorHandle_t maxExpAvgSq, float lr, float beta1, float beta2, float eps, float weightDecay, int64_t step, bool amsgrad) {
+diopiError_t diopiAdam(diopiContextHandle_t ctx, diopiTensorHandle_t input, diopiConstTensorHandle_t grad, diopiTensorHandle_t expAvg,
+                       diopiTensorHandle_t expAvgSq, diopiTensorHandle_t maxExpAvgSq, float lr, float beta1, float beta2, float eps, float weightDecay,
+                       int64_t step, bool amsgrad) {
     DIOPI_CALL(bangAdam(ctx, input, grad, expAvg, expAvgSq, maxExpAvgSq, lr, beta1, beta2, eps, weightDecay, step, amsgrad, 0));
     return diopiSuccess;
 }
 
-diopiError_t diopiAdamW(diopiContextHandle_t ctx, diopiTensorHandle_t input, diopiTensorHandle_t grad, diopiTensorHandle_t expAvg, diopiTensorHandle_t expAvgSq,
+diopiError_t diopiAdamW(diopiContextHandle_t ctx, diopiTensorHandle_t input, diopiConstTensorHandle_t grad, diopiTensorHandle_t expAvg,
+                        diopiTensorHandle_t expAvgSq,
+
                         diopiTensorHandle_t maxExpAvgSq, float lr, float beta1, float beta2, float eps, float weightDecay, int64_t step, bool amsgrad) {
     DIOPI_CALL(bangAdam(ctx, input, grad, expAvg, expAvgSq, maxExpAvgSq, lr, beta1, beta2, eps, weightDecay, step, amsgrad, 1));
     return diopiSuccess;

diff --git a/impl/camb_pytorch/functions.cpp b/impl/camb_pytorch/functions.cpp
@@ -1923,7 +1923,7 @@ diopiError_t diopiMaskedFillInpScalar(diopiContextHandle_t ctx, diopiTensorHandl
     return diopiSuccess;
 }
 
-diopiError_t diopiAdamW(diopiContextHandle_t ctx, diopiTensorHandle_t input, diopiTensorHandle_t grad, diopiTensorHandle_t exp_avg,
+diopiError_t diopiAdamW(diopiContextHandle_t ctx, diopiTensorHandle_t input, diopiConstTensorHandle_t grad, diopiTensorHandle_t exp_avg,
                         diopiTensorHandle_t exp_avg_sq, diopiTensorHandle_t max_exp_avg_sq, float lr, float beta1, float beta2, float eps, float weight_decay,
                         int64_t step, bool amsgrad) {
     camb::aten::setCurCtx(ctx);
@@ -1959,7 +1959,7 @@ diopiError_t diopiAdamW(diopiContextHandle_t ctx, diopiTensorHandle_t input, dio
     return diopiSuccess;
 }
 
-diopiError_t diopiAdam(diopiContextHandle_t ctx, diopiTensorHandle_t input, diopiTensorHandle_t grad, diopiTensorHandle_t exp_avg,
+diopiError_t diopiAdam(diopiContextHandle_t ctx, diopiTensorHandle_t input, diopiConstTensorHandle_t grad, diopiTensorHandle_t exp_avg,
                        diopiTensorHandle_t exp_avg_sq, diopiTensorHandle_t max_exp_avg_sq, float lr, float beta1, float beta2, float eps, float weight_decay,
                        int64_t step, bool amsgrad) {
     camb::aten::setCurCtx(ctx);
@@ -1998,7 +1998,7 @@ diopiError_t diopiAdam(diopiContextHandle_t ctx, diopiTensorHandle_t input, diop
     return diopiSuccess;
 }
 
-diopiError_t diopiAdadelta(diopiContextHandle_t ctx, diopiTensorHandle_t input, diopiTensorHandle_t grad, diopiTensorHandle_t square_avg,
+diopiError_t diopiAdadelta(diopiContextHandle_t ctx, diopiTensorHandle_t input, diopiConstTensorHandle_t grad, diopiTensorHandle_t square_avg,
                            diopiTensorHandle_t acc_delta, float lr, float rho, float eps, float weight_decay) {
     camb::aten::setCurCtx(ctx);
     auto atInput = camb::aten::buildATen(input);

diff --git a/impl/topsrider/functions/functions.cpp b/impl/topsrider/functions/functions.cpp
@@ -717,7 +717,7 @@ DIOPI_API diopiError_t diopiExpand(diopiContextHandle_t ctx, diopiTensorHandle_t
     return impl::tops::topsExpand(ctx, out, input);
 }
 
-DIOPI_API diopiError_t diopiAdam(diopiContextHandle_t ctx, diopiTensorHandle_t input, diopiTensorHandle_t grad, diopiTensorHandle_t exp_avg,
+DIOPI_API diopiError_t diopiAdam(diopiContextHandle_t ctx, diopiTensorHandle_t input, diopiConstTensorHandle_t grad, diopiTensorHandle_t exp_avg,
                                  diopiTensorHandle_t exp_avg_sq, diopiTensorHandle_t max_exp_avg_sq, float lr, float beta1, float beta2, float eps,
                                  float weight_decay, int64_t step, bool amsgrad) {
     TOPSOP_LOG();
@@ -739,7 +739,7 @@ DIOPI_API diopiError_t diopiAddInpScalar(diopiContextHandle_t ctx, diopiTensorHa
     return impl::tops::topsAddInpScalar(ctx, input, other, alpha);
 }
 
-DIOPI_API diopiError_t diopiAdamW(diopiContextHandle_t ctx, diopiTensorHandle_t input, diopiTensorHandle_t grad, diopiTensorHandle_t exp_avg,
+DIOPI_API diopiError_t diopiAdamW(diopiContextHandle_t ctx, diopiTensorHandle_t input, diopiConstTensorHandle_t grad, diopiTensorHandle_t exp_avg,
                                   diopiTensorHandle_t exp_avg_sq, diopiTensorHandle_t max_exp_avg_sq, float lr, float beta1, float beta2, float eps,
                                   float weight_decay, int64_t step, bool amsgrad) {
     TOPSOP_LOG();

diff --git a/impl/topsrider/ops.h b/impl/topsrider/ops.h
@@ -303,7 +303,7 @@ DIOPI_API diopiError_t topsSgd(diopiContextHandle_t ctx, diopiTensorHandle_t w,
 
 DIOPI_API diopiError_t topsExpand(diopiContextHandle_t ctx, diopiTensorHandle_t out, diopiConstTensorHandle_t input);
 
-DIOPI_API diopiError_t topsAdam(diopiContextHandle_t ctx, diopiTensorHandle_t input, diopiTensorHandle_t grad, diopiTensorHandle_t exp_avg,
+DIOPI_API diopiError_t topsAdam(diopiContextHandle_t ctx, diopiTensorHandle_t input, diopiConstTensorHandle_t grad, diopiTensorHandle_t exp_avg,
                                 diopiTensorHandle_t exp_avg_sq, diopiTensorHandle_t max_exp_avg_sq, float lr, float beta1, float beta2, float eps,
                                 float weight_decay, int64_t step, bool amsgrad);
 
@@ -321,7 +321,7 @@ DIOPI_API diopiError_t topsMaxAll(diopiContextHandle_t ctx, diopiTensorHandle_t
 
 DIOPI_API diopiError_t topsAddInpScalar(diopiContextHandle_t ctx, diopiTensorHandle_t input, const diopiScalar_t *other, const diopiScalar_t *alpha);
 
-DIOPI_API diopiError_t topsAdamW(diopiContextHandle_t ctx, diopiTensorHandle_t input, diopiTensorHandle_t grad, diopiTensorHandle_t exp_avg,
+DIOPI_API diopiError_t topsAdamW(diopiContextHandle_t ctx, diopiTensorHandle_t input, diopiConstTensorHandle_t grad, diopiTensorHandle_t exp_avg,
                                  diopiTensorHandle_t exp_avg_sq, diopiTensorHandle_t max_exp_avg_sq, float lr, float beta1, float beta2, float eps,
                                  float weight_decay, int64_t step, bool amsgrad);
 

diff --git a/impl/torch/functions/functions.cpp b/impl/torch/functions/functions.cpp
@@ -2730,7 +2730,7 @@ diopiError_t diopiMeshGrid(diopiContextHandle_t ctx, diopiTensorHandle_t* outs,
     return diopiSuccess;
 }
 
-diopiError_t diopiAdamW(diopiContextHandle_t ctx, diopiTensorHandle_t input, diopiTensorHandle_t grad, diopiTensorHandle_t exp_avg,
+diopiError_t diopiAdamW(diopiContextHandle_t ctx, diopiTensorHandle_t input, diopiConstTensorHandle_t grad, diopiTensorHandle_t exp_avg,
                         diopiTensorHandle_t exp_avg_sq, diopiTensorHandle_t max_exp_avg_sq, float lr, float beta1, float beta2, float eps, float weight_decay,
                         int64_t step, bool amsgrad) {
     impl::aten::setCurStream(ctx);
@@ -2759,7 +2759,7 @@ diopiError_t diopiAdamW(diopiContextHandle_t ctx, diopiTensorHandle_t input, dio
     return diopiSuccess;
 }
 
-diopiError_t diopiAdam(diopiContextHandle_t ctx, diopiTensorHandle_t input, diopiTensorHandle_t grad, diopiTensorHandle_t exp_avg,
+diopiError_t diopiAdam(diopiContextHandle_t ctx, diopiTensorHandle_t input, diopiConstTensorHandle_t grad, diopiTensorHandle_t exp_avg,
                        diopiTensorHandle_t exp_avg_sq, diopiTensorHandle_t max_exp_avg_sq, float lr, float beta1, float beta2, float eps, float weight_decay,
                        int64_t step, bool amsgrad) {
     impl::aten::setCurStream(ctx);
@@ -2791,7 +2791,7 @@ diopiError_t diopiAdam(diopiContextHandle_t ctx, diopiTensorHandle_t input, diop
     return diopiSuccess;
 }
 
-diopiError_t diopiAdadelta(diopiContextHandle_t ctx, diopiTensorHandle_t input, diopiTensorHandle_t grad, diopiTensorHandle_t square_avg,
+diopiError_t diopiAdadelta(diopiContextHandle_t ctx, diopiTensorHandle_t input, diopiConstTensorHandle_t grad, diopiTensorHandle_t square_avg,
                            diopiTensorHandle_t acc_delta, float lr, float rho, float eps, float weight_decay) {
     impl::aten::setCurStream(ctx);
     auto atInput = impl::aten::buildATen(input);
@@ -2813,7 +2813,7 @@ diopiError_t diopiAdadelta(diopiContextHandle_t ctx, diopiTensorHandle_t input,
     return diopiSuccess;
 }
 
-diopiError_t diopiRmsprop(diopiContextHandle_t ctx, diopiTensorHandle_t input, diopiTensorHandle_t grad, diopiTensorHandle_t square_avg,
+diopiError_t diopiRmsprop(diopiContextHandle_t ctx, diopiTensorHandle_t input, diopiConstTensorHandle_t grad, diopiTensorHandle_t square_avg,
                           diopiTensorHandle_t grad_avg, diopiTensorHandle_t momentum_buf, float lr, float alpha, float eps, float weight_decay, float momentum,
                           bool centered) {
     impl::aten::setCurStream(ctx);