Accelerated double backward pass for convolution. (#95)

* Making progress on double backward pass.

* More progress on double backward setup.

* Custom double backward pass is compiling.

* Wrote the A kernel.

* Atomic double backward pass accelerated.

* More convolution double backward tests passing.

* Deterministic double backward A kernel written.

* Double backward deterministic is compiling, but we need to check the output.

* Double backward A kernel is working.

* Gradient of L1 has an error, need to fix.

* All tests passing.

* Updated conv test fixture to download graph again.

Author: vbharadwaj-bk

openequivariance/extension/convolution.hpp

@@ -97,41 +97,50 @@ class __attribute__ ((visibility ("default"))) JITConvImpl : public ConvolutionI
     JIT_IMPL jit;
     KernelLaunchConfig forward_config; 
     KernelLaunchConfig backward_config;
+    KernelLaunchConfig double_backward_config;
     bool is_uvw; 
 
     JITConvImpl(
         std::string jit_kernel,
         KernelLaunchConfig forward_config_i,
         KernelLaunchConfig backward_config_i,
+        KernelLaunchConfig double_backward_config_i,
         bool is_uvw_i) :
             jit(jit_kernel),
             forward_config(forward_config_i),  
             backward_config(backward_config_i),
+            double_backward_config(double_backward_config_i),
             is_uvw(is_uvw_i) {
 
-        vector<string> kernels = {"forward", "backward", "fixup_forward", "fixup_backward"};
+        vector<string> kernels = {"forward", "backward", "fixup_forward", "fixup_backward", "double_backward_A", "double_backward_B", "fixup_double_backwardB"};
 
         int opt_level = 3;
         #ifdef HIP_BACKEND
         if(is_uvw) {
             opt_level = 1;
         }
         #endif 
-        jit.compile(kernels, {{}, {}, {}, {}}, opt_level); 
+        jit.compile(kernels, {{}, {}, {}, {}, {}, {}, {}}, opt_level); 
 
         if(forward_config.smem > 0) {
             jit.set_max_smem(0, forward_config.smem);
+            jit.set_max_smem(4, forward_config.smem);
         }
 
         if(backward_config.smem > 0) {
             jit.set_max_smem(1, backward_config.smem);
         }
+
+        if(double_backward_config.smem > 0) {
+            jit.set_max_smem(5, double_backward_config.smem);
+        }
     }
 
     JITConvImpl(
             std::string jit_kernel,
             std::unordered_map<string, int64_t> fwd_dict, 
             std::unordered_map<string, int64_t> bwd_dict,
+            std::unordered_map<string, int64_t> dbl_bwd_dict,
             std::unordered_map<string, int64_t> kernel_dims 
     ) : JITConvImpl(
             jit_kernel,
@@ -145,6 +154,11 @@ class __attribute__ ((visibility ("default"))) JITConvImpl : public ConvolutionI
                 bwd_dict["num_threads"],
                 bwd_dict["smem"]
             ),
+            KernelLaunchConfig(
+                dbl_bwd_dict["num_blocks"],
+                dbl_bwd_dict["num_threads"],
+                dbl_bwd_dict["smem"]
+            ),
             kernel_dims["is_uvw"] == 1) { }
 
     void exec_conv(
@@ -201,5 +215,40 @@ class __attribute__ ((visibility ("default"))) JITConvImpl : public ConvolutionI
         }
     }
 
+    void double_backward(
+            void* L1_in, void* L2_in, void* W, void* L3_grad, 
+            void* L1_dgrad, void* L2_dgrad, void* w_dgrad, 
+            void* L1_grad, void* L2_grad, void* W_grad, void* L3_dgrad, 
+            void* rows, void* cols,
+            uint64_t nnz, uint64_t node_count,
+            void* wspace, void* transpose_perm) {
+
+        ConvData conv_data = {rows, cols, nnz, node_count};
+        void* args[] = { 
+            &L1_in, &L2_in, &W, &L3_grad, &L1_dgrad, &L2_dgrad, &w_dgrad, 
+            &L1_grad, &L2_grad, &W_grad, &L3_dgrad, &conv_data, &wspace, &transpose_perm
+        };
+
+        jit.execute(4, args, forward_config);
+        if(reinterpret_cast<uint64_t>(wspace) != 0) {
+            void *fixup_args[] = {&wspace, &L3_dgrad};    
+            KernelLaunchConfig fixup_config;
+            fixup_config.num_blocks = forward_config.num_blocks;
+            fixup_config.num_threads = forward_config.num_threads;
+            fixup_config.smem = 0;
+            jit.execute(2, fixup_args, fixup_config);
+        }
+
+        jit.execute(5, args, double_backward_config);
+        if(reinterpret_cast<uint64_t>(wspace) != 0) {
+            void *fixup_args[] = {&wspace, &L1_grad};
+            KernelLaunchConfig fixup_config;
+            fixup_config.num_blocks = double_backward_config.num_blocks;
+            fixup_config.num_threads = double_backward_config.num_threads;
+            fixup_config.smem = 0;
+            jit.execute(6, fixup_args, fixup_config);
+        }
+    }
+
     ~JITConvImpl() = default; 
 };

openequivariance/extension/generic_module.cpp

@@ -49,6 +49,7 @@ PYBIND11_MODULE(generic_module, m) {
         .def("backward_rawptrs", &ConvolutionImpl::backward_rawptrs);
     py::class_<JITConvImpl<JITKernel>, ConvolutionImpl>(m, "JITConvImpl")
         .def(py::init<  std::string, 
+                        std::unordered_map<string, int64_t>, 
                         std::unordered_map<string, int64_t>, 
                         std::unordered_map<string, int64_t>, 
                         std::unordered_map<string, int64_t>>());

openequivariance/extension/torch_tp_jit.cpp

@@ -215,25 +215,32 @@ tuple<torch::Tensor, torch::Tensor, torch::Tensor, torch::Tensor> jit_tp_double_
 
 class TorchJITConv : public torch::CustomClassHolder {
 public:
-    Map_t fwd_dict, bwd_dict, kernel_dims;
+    Map_t fwd_dict, bwd_dict, dbl_bwd_dict, kernel_dims;
     JITConvImpl<JITKernel> internal;
     int64_t L3_dim;
 
-    TorchJITConv(string kernel_plaintext, Map_t fwd_dict_i, Map_t bwd_dict_i, Map_t kernel_dims_i) :
+    TorchJITConv(string kernel_plaintext, Map_t fwd_dict_i, Map_t bwd_dict_i, Map_t dbl_bwd_dict_i, Map_t kernel_dims_i) :
         fwd_dict(fwd_dict_i.copy()),
         bwd_dict(bwd_dict_i.copy()),
+        dbl_bwd_dict(bwd_dict_i.copy()),
         kernel_dims(kernel_dims_i.copy()),
         internal(kernel_plaintext,
                 to_map(fwd_dict_i),
                 to_map(bwd_dict_i),
+                to_map(dbl_bwd_dict_i),
                 to_map(kernel_dims_i)
             ),
         L3_dim(kernel_dims.at("L3_dim")) { }
 
-    tuple<tuple<string, string>, tuple<string, Map_t>, tuple<string, Map_t>, tuple<string, Map_t>> __obj_flatten__() {
+    tuple<tuple<string, string>, 
+        tuple<string, Map_t>, 
+        tuple<string, Map_t>, 
+        tuple<string, Map_t>, 
+        tuple<string, Map_t>> __obj_flatten__() {
         return tuple(tuple("kernel_plaintext", internal.jit.kernel_plaintext),
             tuple("fwd_config", fwd_dict),
             tuple("bwd_config", bwd_dict),
+            tuple("dbl_bwd_config", dbl_bwd_dict),
             tuple("kernel_dims", kernel_dims));
     }
 
@@ -347,6 +354,55 @@ tuple<torch::Tensor, torch::Tensor, torch::Tensor> jit_conv_backward(
     return tuple(L1_grad, L2_grad, W_grad);
 }
 
+tuple<torch::Tensor, torch::Tensor, torch::Tensor, torch::Tensor> jit_conv_double_backward(
+        const c10::intrusive_ptr<TorchJITConv> &jit_instance,
+        const torch::Tensor &L1_in, 
+        const torch::Tensor &L2_in, 
+        const torch::Tensor &W, 
+        const torch::Tensor &L3_grad, 
+        const torch::Tensor &L1_dgrad, 
+        const torch::Tensor &L2_dgrad, 
+        const torch::Tensor &W_dgrad, 
+        const torch::Tensor &rows,
+        const torch::Tensor &cols,
+        const torch::Tensor &workspace,
+        const torch::Tensor &transpose_perm) {
+
+    int64_t nnz = rows.sizes()[0];
+    int64_t node_count = L1_in.sizes()[0];
+    torch::Tensor L1_grad = torch::zeros(L1_in.sizes(), L1_in.options());
+    torch::Tensor L2_grad = torch::empty(L2_in.sizes(), L2_in.options());
+    torch::Tensor W_grad = torch::empty(W.sizes(), W.options());
+    torch::Tensor L3_dgrad = torch::zeros(L3_grad.sizes(), L3_grad.options());
+
+    torch::Tensor L1_in_contig = L1_in.contiguous();
+    torch::Tensor L2_in_contig = L2_in.contiguous();
+    torch::Tensor W_contig = W.contiguous();
+    torch::Tensor L3_grad_contig = L3_grad.contiguous();
+    torch::Tensor L1_dgrad_contig = L1_dgrad.contiguous();
+    torch::Tensor L2_dgrad_contig = L2_dgrad.contiguous();
+    torch::Tensor W_dgrad_contig = W_dgrad.contiguous();
+
+    torch::Tensor rows_contig = rows.contiguous();
+    torch::Tensor cols_contig = cols.contiguous();
+    torch::Tensor workspace_contig = workspace.contiguous();
+    torch::Tensor transpose_perm_contig = transpose_perm.contiguous();
+
+    jit_instance->internal.double_backward(
+            data_ptr(L1_in_contig), data_ptr(L2_in_contig),
+            data_ptr(W_contig), data_ptr(L3_grad_contig),
+            data_ptr(L1_dgrad_contig), data_ptr(L2_dgrad_contig),
+            data_ptr(W_dgrad_contig),
+            data_ptr(L1_grad), data_ptr(L2_grad),
+            data_ptr(W_grad), data_ptr(L3_dgrad),
+            data_ptr(rows_contig), data_ptr(cols_contig),
+            nnz, node_count,
+            data_ptr(workspace_contig), data_ptr(transpose_perm_contig)
+    );
+
+    return tuple(L1_grad, L2_grad, W_grad, L3_dgrad);
+}
+
 // =========================================================== 
 
 TORCH_LIBRARY_FRAGMENT(torch_tp_jit, m) { 
@@ -376,7 +432,7 @@ TORCH_LIBRARY_FRAGMENT(torch_tp_jit, m) {
 
 
     m.class_<TorchJITConv>("TorchJITConv")
-        .def(torch::init<string, Map_t, Map_t, Map_t>())
+        .def(torch::init<string, Map_t, Map_t, Map_t, Map_t>())
         .def("__obj_flatten__", &TorchJITConv::__obj_flatten__)
         .def("exec_conv_rawptrs", &TorchJITConv::exec_conv_rawptrs)
         .def("backward_rawptrs", &TorchJITConv::backward_rawptrs)
@@ -386,27 +442,28 @@ TORCH_LIBRARY_FRAGMENT(torch_tp_jit, m) {
         .def_pickle(
             // __getstate__
             [](const c10::intrusive_ptr<TorchJITConv>& self)
-                -> tuple<string, Map_t, Map_t, Map_t> {
-                return tuple(self->internal.jit.kernel_plaintext, self->fwd_dict, self->bwd_dict, self->kernel_dims);
+                -> tuple<string, Map_t, Map_t, Map_t, Map_t> {
+                return tuple(self->internal.jit.kernel_plaintext, self->fwd_dict, self->bwd_dict, self->dbl_bwd_dict, self->kernel_dims);
             },
             // __setstate__
-            [](tuple<string, Map_t, Map_t, Map_t> state)
+            [](tuple<string, Map_t, Map_t, Map_t, Map_t> state)
                 -> c10::intrusive_ptr<TorchJITConv> {
-                return c10::make_intrusive<TorchJITConv>(get<0>(state), get<1>(state), get<2>(state), get<3>(state));
+                return c10::make_intrusive<TorchJITConv>(get<0>(state), get<1>(state), get<2>(state), get<3>(state), get<4>(state));
             });
 
     m.def("jit_conv_forward(__torch__.torch.classes.torch_tp_jit.TorchJITConv jit, Tensor L1_in, Tensor L2_in, Tensor W, Tensor rows, Tensor cols, Tensor workspace, Tensor transpose_perm) -> Tensor");
     m.def("jit_conv_backward(__torch__.torch.classes.torch_tp_jit.TorchJITConv jit, Tensor L1_in, Tensor L2_in, Tensor W, Tensor L3_grad, Tensor rows, Tensor cols, Tensor workspace, Tensor transpose_perm) -> (Tensor, Tensor, Tensor)");
+    m.def("jit_conv_double_backward(__torch__.torch.classes.torch_tp_jit.TorchJITConv jit, Tensor L1_in, Tensor L2_in, Tensor W, Tensor L3_grad, Tensor L1_dgrad, Tensor L2_dgrad, Tensor W_dgrad, Tensor rows, Tensor cols, Tensor workspace, Tensor transpose_perm) -> (Tensor, Tensor, Tensor, Tensor)");
 };
 
-
 TORCH_LIBRARY_IMPL(torch_tp_jit, CUDA, m) { 
     m.impl("jit_tp_forward", &jit_tp_forward);
     m.impl("jit_tp_backward", &jit_tp_backward);
     m.impl("jit_tp_double_backward", &jit_tp_double_backward);
 
     m.impl("jit_conv_forward", &jit_conv_forward);
     m.impl("jit_conv_backward", &jit_conv_backward);
+    m.impl("jit_conv_double_backward", &jit_conv_double_backward);
 };
 
 PYBIND11_MODULE(torch_tp_jit, m) {}

openequivariance/implementations/ComputationSchedule.py

@@ -276,7 +276,7 @@ def __init__(self,
 
         smem_limit -= 1
         self.memory_per_warp = smem_limit // warps_per_block
-        self.memory_per_warp -= self.memory_per_warp % 4
+        self.memory_per_warp -= self.memory_per_warp % 8
 
         # =====================================================================
         # Shared memory partitioning functions 

openequivariance/implementations/convolution/LoopUnrollConv.py

@@ -49,8 +49,21 @@ def generate_backward_schedule(warps_per_block):
                     warp_size=dp.warpsize,
                     include_scratch=self.is_uvw,
                     stream_weights=self.is_uvw)
+            
+        def generate_double_backward_schedule(warps_per_block):
+            self.double_backward_schedule = ComputationSchedule(self.config, 
+                    smem_limit=dp.maxSharedMemPerBlock, 
+                    warps_per_block=warps_per_block,
+                    warp_size=dp.warpsize,
+                    block_count=dp.multiprocessorCount,
+                    direction = "double_backward",
+                    irrep_dtype = config.irrep_dtype,
+                    weight_dtype = config.weight_dtype,
+                    include_scratch=self.is_uvw,
+                    stream_weights=self.is_uvw,
+                    schedule_type=3)
 
-        scheduler_generators = [generate_forward_schedule, generate_backward_schedule]
+        scheduler_generators = [generate_forward_schedule, generate_backward_schedule, generate_double_backward_schedule]
 
         for generate_schedule in scheduler_generators:
             warp_count = 6
@@ -72,35 +85,47 @@ def generate_backward_schedule(warps_per_block):
                 for key in segment.L1Map.storeback_procedure:
                     segment.L1Map.storeback_procedure[key] = "atomic_accumulate"
 
+            for segment in self.double_backward_schedule.segments:
+                for key in segment.L1Map.storeback_procedure:
+                    segment.L1Map.storeback_procedure[key] = "atomic_accumulate"
+
+
         idx_type_map = {np.int32: "int", np.int64: "long"}
 
         if self.torch_op:
             self.setup_torch_module()
 
         self.forward_workspace_offset = None
         self.backward_workspace_offset = None
+        self.double_backwardB_offset = None
 
         workspace_size = 1
         if deterministic:
             destination_index_bytes = 32 # Add extra to account for padding
             workspace_size = max(
                 (self.forward_schedule.L3.dim * np.dtype(config.irrep_dtype).itemsize + destination_index_bytes) * self.forward_schedule.total_warps,
-                (self.backward_schedule.L1.dim * np.dtype(config.irrep_dtype).itemsize + destination_index_bytes) * self.backward_schedule.total_warps)
+                (self.backward_schedule.L1.dim * np.dtype(config.irrep_dtype).itemsize + destination_index_bytes) * self.backward_schedule.total_warps,
+                (self.double_backward_schedule.L1.dim * np.dtype(config.irrep_dtype).itemsize + destination_index_bytes) * self.double_backward_schedule.total_warps
+            )
 
             self.forward_workspace_offset = self.forward_schedule.L3.dim * np.dtype(config.irrep_dtype).itemsize * self.forward_schedule.total_warps
             self.backward_workspace_offset = self.backward_schedule.L1.dim * np.dtype(config.irrep_dtype).itemsize * self.backward_schedule.total_warps
+            self.double_backwardB_offset = self.double_backward_schedule.L1.dim * np.dtype(config.irrep_dtype).itemsize * self.double_backward_schedule.total_warps
 
             self.forward_workspace_offset = (self.forward_workspace_offset + 7) // 8 * 8
             self.backward_workspace_offset = (self.backward_workspace_offset + 7) // 8 * 8
+            self.double_backwardB_offset = (self.double_backwardB_offset + 7) // 8 * 8
 
         self.allocate_workspace(workspace_size)
 
         self.jit_kernel = template.render(
             forward_schedule=self.forward_schedule,
             backward_schedule=self.backward_schedule,
+            double_backward_schedule=self.double_backward_schedule,
             idx_type=idx_type_map[idx_dtype],
             forward_workspace_offset=self.forward_workspace_offset,
-            backward_workspace_offset=self.backward_workspace_offset)
+            backward_workspace_offset=self.backward_workspace_offset,
+            double_backwardB_offset=self.double_backwardB_offset)
         self.jit_kernel = postprocess_kernel(self.jit_kernel)
 
         if self.torch_op and extlib.TORCH_COMPILE:
@@ -115,6 +140,7 @@ def generate_backward_schedule(warps_per_block):
         self.internal = internal_cls(self.jit_kernel,
                 vars(self.forward_schedule.launch_config), 
                 vars(self.backward_schedule.launch_config),
+                vars(self.double_backward_schedule.launch_config),
                 {"L3_dim": self.L3.dim,
                  "is_uvw": int(self.is_uvw)})
         logger.info("Kernel compiled!")
@@ -137,9 +163,10 @@ def register_torch_fakes(cls):
         class TorchJITConv:
             def __init__(self, kernel_plaintext: str, 
                         fwd_config: dict[str, int], 
-                        bwd_config: dict[str, int], 
+                        bwd_config: dict[str, int],
+                        dbl_bwd_config: dict[str, int], 
                         kernel_dims: dict[str, int]) -> None:
-                self.kernel_plaintext, self.fwd_config, self.bwd_config, self.kernel_dims = kernel_plaintext, fwd_config, bwd_config, kernel_dims
+                self.kernel_plaintext, self.fwd_config, self.bwd_config, self.dbl_bwd_config, self.kernel_dims = kernel_plaintext, fwd_config, bwd_config, dbl_bwd_config, kernel_dims
 
             @classmethod
             def __obj_unflatten__(cls, flattened_product):
@@ -149,7 +176,7 @@ def __len__(self):
                 return 0
 
             def __setstate__(self, state):
-                self.kernel_plaintext, self.fwd_config, self.bwd_config, self.kernel_dims = state 
+                self.kernel_plaintext, self.fwd_config, self.bwd_config, self.dbl_bwd_config, self.kernel_dims = state 
 
         @torch.library.register_fake("torch_tp_jit::jit_conv_forward")
         def fake_forward(jit, L1_in, L2_in, W, rows, cols, workspace_buffer, sender_perm):
@@ -163,6 +190,7 @@ def fake_backward(jit, L1_in, L2_in, W, L3_grad, rows, cols, workspace_buffer, s
     def register_autograd(cls):
         forward_op = torch.ops.torch_tp_jit.jit_conv_forward
         backward_op = torch.ops.torch_tp_jit.jit_conv_backward
+        double_backward_op = torch.ops.torch_tp_jit.jit_conv_double_backward
 
         def setup_context(ctx, inputs, output):
             ctx.jit, ctx.L1_in, ctx.L2_in, ctx.W, ctx.rows, ctx.cols, ctx.workspace_buffer, ctx.sender_perm = inputs
@@ -178,17 +206,8 @@ def setup_context_double_backward(ctx, inputs, output):
             ctx.inputs = inputs
 
         def double_backward(ctx, E, F, G):
-            jit, A, B, C, D, rows, cols, wspace, sender_perm = ctx.jit, ctx.L1_in, ctx.L2_in, ctx.grad_output, ctx.W, ctx.rows, ctx.cols, ctx.workspace_buffer, ctx.sender_perm
-
-            op1 = backward_op(jit, E, F, D, C, rows, cols, wspace, sender_perm)
-            op2 = backward_op(jit, A, B, G, C, rows, cols, wspace, sender_perm)
-            op3 = forward_op(jit, E, B, D, rows, cols, wspace, sender_perm)
-            op4 = backward_op(jit, E, B, D, C, rows, cols, wspace, sender_perm) # op4 and op5 could be combined with op3 and op6
-            op5 = backward_op(jit, A, F, D, C, rows, cols, wspace, sender_perm)
-            op6 = forward_op(jit, A, F, D, rows, cols, wspace, sender_perm)
-            op7 = forward_op(jit, A, B, G, rows, cols, wspace, sender_perm)
-
-            return None, op1[0] + op2[0], op1[1] + op2[1], op4[2] + op5[2], (op3 + op6 + op7), None, None, None, None
+            result = double_backward_op(ctx.jit, ctx.L1_in, ctx.L2_in, ctx.W, ctx.grad_output, E, F, G, ctx.rows, ctx.cols, ctx.workspace_buffer, ctx.sender_perm)
+            return None, result[0], result[1], result[2], result[3], None, None, None, None
 
         torch.library.register_autograd("torch_tp_jit::jit_conv_backward", double_backward, setup_context=setup_context_double_backward)