fix(gpu): fix scalar mul with 1 block

zama-ai · Jan 27, 2025 · 88126f7 · 88126f7
1 parent ae0dff9
commit 88126f7
Show file tree

Hide file tree

Showing 3 changed files with 16 additions and 9 deletions.
diff --git a/backends/tfhe-cuda-backend/cuda/include/integer/integer_utilities.h b/backends/tfhe-cuda-backend/cuda/include/integer/integer_utilities.h
@@ -4454,9 +4454,16 @@ template <typename Torus> struct int_scalar_mul_buffer {
                         num_ciphertext_bits * num_radix_blocks * lwe_size_bytes,
                         streams[0], gpu_indexes[0]);
 
-      logical_scalar_shift_buffer = new int_logical_scalar_shift_buffer<Torus>(
-          streams, gpu_indexes, gpu_count, LEFT_SHIFT, params, num_radix_blocks,
-          allocate_gpu_memory, all_shifted_buffer);
+      if (num_ciphertext_bits * num_radix_blocks >= num_radix_blocks + 2)
+        logical_scalar_shift_buffer =
+            new int_logical_scalar_shift_buffer<Torus>(
+                streams, gpu_indexes, gpu_count, LEFT_SHIFT, params,
+                num_radix_blocks, allocate_gpu_memory, all_shifted_buffer);
+      else
+        logical_scalar_shift_buffer =
+            new int_logical_scalar_shift_buffer<Torus>(
+                streams, gpu_indexes, gpu_count, LEFT_SHIFT, params,
+                num_radix_blocks, allocate_gpu_memory);
 
       sum_ciphertexts_vec_mem = new int_sum_ciphertexts_vec_memory<Torus>(
           streams, gpu_indexes, gpu_count, params, num_radix_blocks,

diff --git a/backends/tfhe-cuda-backend/cuda/src/integer/scalar_mul.cuh b/backends/tfhe-cuda-backend/cuda/src/integer/scalar_mul.cuh
@@ -47,9 +47,6 @@ __host__ void host_integer_scalar_mul_radix(
     void *const *bsks, T *const *ksks, uint32_t input_lwe_dimension,
     uint32_t message_modulus, uint32_t num_radix_blocks, uint32_t num_scalars) {
 
-  if (num_radix_blocks == 0 | num_scalars == 0)
-    return;
-
   // lwe_size includes the presence of the body
   // whereas lwe_dimension is the number of elements in the mask
   uint32_t lwe_size = input_lwe_dimension + 1;

diff --git a/tfhe/src/integer/gpu/server_key/radix/scalar_mul.rs b/tfhe/src/integer/gpu/server_key/radix/scalar_mul.rs
@@ -79,7 +79,9 @@ impl CudaServerKey {
             return;
         }
 
-        if scalar == Scalar::ONE {
+        let ciphertext = ct.as_mut();
+        let num_blocks = ciphertext.d_blocks.lwe_ciphertext_count().0;
+        if scalar == Scalar::ONE || num_blocks == 0 {
             return;
         }
 
@@ -89,8 +91,6 @@ impl CudaServerKey {
             self.unchecked_scalar_left_shift_assign_async(ct, scalar.ilog2() as u64, streams);
             return;
         }
-        let ciphertext = ct.as_mut();
-        let num_blocks = ciphertext.d_blocks.lwe_ciphertext_count().0;
         let msg_bits = self.message_modulus.0.ilog2() as usize;
         let decomposer = BlockDecomposer::with_early_stop_at_zero(scalar, 1).iter_as::<u8>();
 
@@ -106,6 +106,9 @@ impl CudaServerKey {
         let decomposed_scalar = BlockDecomposer::with_early_stop_at_zero(scalar, 1)
             .iter_as::<u64>()
             .collect::<Vec<_>>();
+        if decomposed_scalar.is_empty() {
+            return;
+        }
 
         match &self.bootstrapping_key {
             CudaBootstrappingKey::Classic(d_bsk) => {