cuda : try cublasGemmStridedBatchedEx

Author: ggerganov

ggml-cuda.cu

@@ -7134,8 +7134,21 @@ static void ggml_cuda_mul_mat_mat_batched_cublas(const ggml_tensor * src0, const
         }
     }
 #else
-    // use cublasGemmBatchedEx
-    {
+    if (r2 == 1 && r3 == 1 && src0->nb[2]*src0->ne[2] == src0->nb[3] && src1->nb[2]*src1->ne[2] == src1->nb[3]) {
+        // there is no broadcast and src0, src1 are contiguous across dims 2, 3
+        // use cublasGemmStridedBatchedEx
+        CUBLAS_CHECK(
+        cublasGemmStridedBatchedEx(g_cublas_handles[id], CUBLAS_OP_T, CUBLAS_OP_N,
+                ne01, ne11, ne10,
+                &alpha_f16, (char *) src0_as_f16, CUDA_R_16F, nb01/sizeof(half),  ne02*src0->nb[2],   // strideA
+                            (char *) src1_as_f16, CUDA_R_16F, nb11/sizeof(float), ne12*src1->nb[2]/2, // strideB
+                &beta_f16,  (char *)     dst_f16, CUDA_R_16F, ne01,               ne12* dst->nb[2]/2, // strideC
+                ne13,
+                CUBLAS_COMPUTE_16F,
+                CUBLAS_GEMM_DEFAULT_TENSOR_OP));
+        printf("cublasGemmStridedBatchedEx\n");
+    } else {
+        // use cublasGemmBatchedEx
         const int ne23 = ne12*ne13;
 
         // TODO: avoid this alloc