From 5ca309ecb0c04031fb1540b371cabf340f3f5087 Mon Sep 17 00:00:00 2001
From: laurent <laurent.mazare@gmail.com>
Date: Fri, 23 Jun 2023 18:40:15 +0100
Subject: [PATCH] Optimize the unary cuda kernels for the contiguous case.

---
 kernels/src/affine.cu      | 24 +++++++++++++++---------
 kernels/src/cuda_utils.cuh | 16 ++++++++++++++++
 kernels/src/unary.cu       | 16 ++++++++++++----
 3 files changed, 43 insertions(+), 13 deletions(-)

diff --git a/kernels/src/affine.cu b/kernels/src/affine.cu
index 63835940..986c94a5 100644
--- a/kernels/src/affine.cu
+++ b/kernels/src/affine.cu
@@ -6,22 +6,28 @@ extern "C" __global__ void FN_NAME(  \
     const size_t numel,  \
     const size_t num_dims, \
     const size_t *info, \
-    const TYPENAME *x, \
-    TYPENAME *y, \
+    const TYPENAME *inp, \
+    TYPENAME *out, \
     const TYPENAME mul, \
     const TYPENAME add \
 ) {  \
     const size_t *dims = info; \
     const size_t *strides = info + num_dims; \
-    unsigned int i = blockIdx.x * blockDim.x + threadIdx.x;  \
-    if (i >= numel) {  \
-        return;  \
-    }  \
-    unsigned strided_i = get_strided_index(i, num_dims, dims, strides); \
-    y[strided_i] = x[i] * mul + add; \
+    if (is_contiguous(num_dims, dims, strides)) { \
+        for (unsigned int i = blockIdx.x * blockDim.x + threadIdx.x; i < numel; i += blockDim.x * gridDim.x) { \
+            TYPENAME x = inp ? inp[i] : out[i]; \
+            out[i] = x * mul + add; \
+        } \
+    } \
+    else { \
+        for (unsigned int i = blockIdx.x * blockDim.x + threadIdx.x; i < numel; i += blockDim.x * gridDim.x) { \
+            unsigned strided_i = get_strided_index(i, num_dims, dims, strides); \
+            TYPENAME x = inp ? inp[strided_i] : out[i]; \
+            out[i] = x * mul + add; \
+        } \
+    } \
 } \
 
 AFFINE_OP(float, affine_f32)
 AFFINE_OP(double, affine_f64)
 AFFINE_OP(uint32_t, affine_u32)
-
diff --git a/kernels/src/cuda_utils.cuh b/kernels/src/cuda_utils.cuh
index 580ed11a..a7c2a9f6 100644
--- a/kernels/src/cuda_utils.cuh
+++ b/kernels/src/cuda_utils.cuh
@@ -1,6 +1,22 @@
 #include "cuda_fp16.h"
 #include "compatibility.cuh"
 
+__device__ bool is_contiguous(
+    const size_t num_dims,
+    const size_t *dims,
+    const size_t *strides
+) {
+    size_t acc = 1;
+    for (unsigned int d = 0; d < num_dims; d++) {
+        unsigned int dim_idx = num_dims - 1 - d;
+        if (acc != strides[dim_idx]) {
+            return false;
+        }
+        acc *= dims[dim_idx];
+    }
+    return true;
+}
+
 __device__ unsigned int get_strided_index(
     unsigned int idx,
     const size_t num_dims,
diff --git a/kernels/src/unary.cu b/kernels/src/unary.cu
index 408f92a8..ef672f63 100644
--- a/kernels/src/unary.cu
+++ b/kernels/src/unary.cu
@@ -10,10 +10,18 @@ extern "C" __global__ void FN_NAME( \
 ) { \
     const size_t *dims = info; \
     const size_t *strides = info + num_dims; \
-    for (unsigned int i = blockIdx.x * blockDim.x + threadIdx.x; i < numel; i += blockDim.x * gridDim.x) { \
-        unsigned strided_i = get_strided_index(i, num_dims, dims, strides); \
-        TYPENAME x = inp ? inp[strided_i] : out[i]; \
-        out[i] = FUNC; \
+    if (is_contiguous(num_dims, dims, strides)) { \
+        for (unsigned int i = blockIdx.x * blockDim.x + threadIdx.x; i < numel; i += blockDim.x * gridDim.x) { \
+            TYPENAME x = inp ? inp[i] : out[i]; \
+            out[i] = FUNC; \
+        } \
+    } \
+    else { \
+        for (unsigned int i = blockIdx.x * blockDim.x + threadIdx.x; i < numel; i += blockDim.x * gridDim.x) { \
+            unsigned strided_i = get_strided_index(i, num_dims, dims, strides); \
+            TYPENAME x = inp ? inp[strided_i] : out[i]; \
+            out[i] = FUNC; \
+        } \
     } \
 } \