COMPMID-477 - Optimized CLDirectConvolution1x1 for Bifrost

	    - Fixed bug in CLDirectConvolution3x3

Change-Id: Iaf34ef44f0b7bc02e66f3eb4452ff7a90ef83523
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/86725
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Reviewed-by: Moritz Pflanzer <moritz.pflanzer@arm.com>

4 files changed, 135 insertions, 6 deletions

diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index 03c9903d6d..dd549f035b 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -148,6 +148,7 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "derivative", "derivative.cl" },
     { "dilate", "dilate.cl" },
     { "direct_convolution1x1", "direct_convolution1x1.cl" },
+    { "direct_convolution1x1_f32_bifrost", "direct_convolution1x1.cl" },
     { "direct_convolution3x3", "direct_convolution3x3.cl" },
     { "direct_convolution3x3_f32_bifrost", "direct_convolution3x3.cl" },
     { "direct_convolution5x5", "direct_convolution5x5.cl" },
diff --git a/src/core/CL/cl_kernels/direct_convolution1x1.cl b/src/core/CL/cl_kernels/direct_convolution1x1.cl
index 2aa999a80f..fb516ddd6e 100644
--- a/src/core/CL/cl_kernels/direct_convolution1x1.cl
+++ b/src/core/CL/cl_kernels/direct_convolution1x1.cl
@@ -41,6 +41,8 @@ MULQ_SAT_IMPL(qs32x8, qs32x8)
 
 #endif /* FIXED_POINT_POSITION */
 
+#if defined(DATA_TYPE) && defined(DATA_SIZE) && defined(STRIDE_X) && defined(WEIGHTS_DEPTH)
+
 #if STRIDE_X == 3
 #define INPUT_PIXEL_STR(data_size) extract_input_stride3_##data_size
 #define INPUT_PIXEL(data_size) INPUT_PIXEL_STR(data_size)
@@ -165,7 +167,6 @@ inline VEC_DATA_TYPE(DATA_TYPE, 8) extract_input_stride3_8(__global const DATA_T
  * @param[in]  biases_offset_first_element_in_bytes  The offset of the first element in the biases tensor
  * @param[in]  weights_stride_w                      Stride of the weights tensor in the 4th dimension
  */
-#if defined(DATA_TYPE) && defined(DATA_SIZE) && defined(STRIDE_X) && defined(WEIGHTS_DEPTH)
 __kernel void direct_convolution1x1(
     TENSOR3D_DECLARATION(src),
     TENSOR3D_DECLARATION(dst),
@@ -207,3 +208,122 @@ __kernel void direct_convolution1x1(
     vstore8(CONVERT_SAT(pixels, VEC_DATA_TYPE(DATA_TYPE, 8)), 0, (__global DATA_TYPE *)dst.ptr);
 }
 #endif // defined(DATA_TYPE) && defined(DATA_SIZE) && defined(STRIDE_X) && defined(WEIGHTS_DEPTH)
+
+#if defined(WEIGHTS_DEPTH)
+
+#define CONVOLUTION1x1_BIFROST(acc, src, weight_value) \
+    ({                                                 \
+        acc.s0 = mad(src.s0, weight_value, acc.s0);    \
+        acc.s1 = mad(src.s1, weight_value, acc.s1);    \
+        acc.s2 = mad(src.s2, weight_value, acc.s2);    \
+        acc.s3 = mad(src.s3, weight_value, acc.s3);    \
+    })
+
+/** An optimized direct convolution 1x1 OpenCL kernel for Bifrost architectures when the data type is F32
+ *
+ * @note This OpenCL kernel works only with stride_x and stride_y equal to 1
+ * @note The third dimensions of the weights tensors must be passed at compile time using -DWEIGHTS_DEPTH
+ * @note In case biases, -DHAS_BIAS must to be passed at compile
+ *
+ * @param[in]  src_ptr                               Pointer to the source tensor. Supported data types: F32
+ * @param[in]  src_stride_x                          Stride of the source tensor in X dimension (in bytes)
+ * @param[in]  src_step_x                            src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  src_stride_y                          Stride of the source tensor in Y dimension (in bytes)
+ * @param[in]  src_step_y                            src_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  src_stride_z                          Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  src_step_z                            src_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  src_offset_first_element_in_bytes     The offset of the first element in the source tensor
+ * @param[out] dst_ptr                               Pointer to the destination tensor. Supported data types: same as @p src_ptr
+ * @param[in]  dst_stride_x                          Stride of the destination tensor in X dimension (in bytes)
+ * @param[in]  dst_step_x                            dst_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  dst_stride_y                          Stride of the destination tensor in Y dimension (in bytes)
+ * @param[in]  dst_step_y                            dst_stride_y * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  dst_stride_z                          Stride of the destination tensor in Z dimension (in bytes)
+ * @param[in]  dst_step_z                            dst_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes     The offset of the first element in the destination tensor
+ * @param[out] weights_ptr                           Pointer to the weights tensor. Supported data types: same as @p weights_ptr
+ * @param[in]  weights_stride_x                      Stride of the weights tensor in X dimension (in bytes)
+ * @param[in]  weights_step_x                        weights_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  weights_stride_y                      Stride of the weights tensor in Y dimension (in bytes)
+ * @param[in]  weights_step_y                        weights_stride_y * number of elements along y processed per workitem(in bytes)
+ * @param[in]  weights_stride_z                      Stride of the weights tensor in Z dimension (in bytes)
+ * @param[in]  weights_step_z                        weights_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  weights_offset_first_element_in_bytes The offset of the first element in the weights tensor
+ * @param[in]  biases_ptr                            Pointer to the biases tensor. Same as @p src_ptr
+ * @param[in]  biases_stride_x                       Stride of the biases tensor in X dimension (in bytes)
+ * @param[in]  biases_step_x                         biases_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  biases_offset_first_element_in_bytes  The offset of the first element in the biases tensor
+ * @param[in]  weights_stride_w                      Stride of the weights tensor in the 4th dimension
+ */
+__kernel void direct_convolution1x1_f32_bifrost(
+    TENSOR3D_DECLARATION(src),
+    TENSOR3D_DECLARATION(dst),
+    TENSOR3D_DECLARATION(weights),
+#ifdef HAS_BIAS
+    VECTOR_DECLARATION(biases),
+#endif /* defined(HAS_BIAS) */
+    unsigned int weights_stride_w)
+{
+    // Get the kernel index
+    const int kernel_index = get_global_id(2);
+
+    Image    src = CONVERT_TO_IMAGE_STRUCT(src);
+    Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst);
+
+    float4 acc0 = 0.0f;
+    float4 acc1 = 0.0f;
+    float4 acc2 = 0.0f;
+    float4 acc3 = 0.0f;
+
+    __global uchar *weights_addr = (__global uchar *)(weights_ptr + weights_offset_first_element_in_bytes + kernel_index * weights_stride_w);
+    __global uchar *src_addr     = (__global uchar *)offset(&src, 0, 0);
+
+    for(ushort d = 0; d < (ushort)WEIGHTS_DEPTH; ++d)
+    {
+        // Load the weights
+        float weight = *((__global float *)weights_addr);
+
+        // Load values from row0 of input tensor
+        float4 src0 = vload4(0, (__global float *)(src_addr + 0 * src_stride_y));
+        float4 src1 = vload4(0, (__global float *)(src_addr + 1 * src_stride_y));
+        float4 src2 = vload4(0, (__global float *)(src_addr + 2 * src_stride_y));
+        float4 src3 = vload4(0, (__global float *)(src_addr + 3 * src_stride_y));
+
+        CONVOLUTION1x1_BIFROST(acc0, src0, weight);
+        CONVOLUTION1x1_BIFROST(acc1, src1, weight);
+        CONVOLUTION1x1_BIFROST(acc2, src2, weight);
+        CONVOLUTION1x1_BIFROST(acc3, src3, weight);
+
+        src_addr += src_stride_z;
+        weights_addr += weights_stride_z;
+    }
+
+#ifdef HAS_BIAS
+    Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases);
+
+    float bias = (float) * ((__global float *)(vector_offset(&biases, kernel_index)));
+
+    acc0.s0 += bias;
+    acc0.s1 += bias;
+    acc0.s2 += bias;
+    acc0.s3 += bias;
+    acc1.s0 += bias;
+    acc1.s1 += bias;
+    acc1.s2 += bias;
+    acc1.s3 += bias;
+    acc2.s0 += bias;
+    acc2.s1 += bias;
+    acc2.s2 += bias;
+    acc2.s3 += bias;
+    acc3.s0 += bias;
+    acc3.s1 += bias;
+    acc3.s2 += bias;
+    acc3.s3 += bias;
+#endif /* defined(HAS_BIAS) */
+
+    vstore4(acc0, 0, (__global float *)(dst.ptr + 0 * dst_stride_y));
+    vstore4(acc1, 0, (__global float *)(dst.ptr + 1 * dst_stride_y));
+    vstore4(acc2, 0, (__global float *)(dst.ptr + 2 * dst_stride_y));
+    vstore4(acc3, 0, (__global float *)(dst.ptr + 3 * dst_stride_y));
+}
+#endif // defined(WEIGHTS_DEPTH)
\ No newline at end of file
diff --git a/src/core/CL/cl_kernels/direct_convolution3x3.cl b/src/core/CL/cl_kernels/direct_convolution3x3.cl
index 26f24e187b..1a228dd74c 100644
--- a/src/core/CL/cl_kernels/direct_convolution3x3.cl
+++ b/src/core/CL/cl_kernels/direct_convolution3x3.cl
@@ -288,11 +288,11 @@ __kernel void direct_convolution3x3_f32_bifrost(
 #ifdef HAS_BIAS
     Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases);
 
-    float4 bias = (float4) * ((__global float *)(vector_offset(&biases, kernel_index)));
+    float bias = (float) * ((__global float *)(vector_offset(&biases, kernel_index)));
 
-    pixels0 += bias;
-    pixels1 += bias;
-    pixels2 += bias;
+    pixels0 += (float4)bias;
+    pixels1 += (float4)bias;
+    pixels2 += (float4)bias;
 #endif /* defined(HAS_BIAS) */
 
     vstore4(pixels0, 0, (__global float *)(dst.ptr + 0 * dst_stride_y));
diff --git a/src/core/CL/kernels/CLDirectConvolutionLayerKernel.cpp b/src/core/CL/kernels/CLDirectConvolutionLayerKernel.cpp
index 265c5074c5..75e6d5e971 100644
--- a/src/core/CL/kernels/CLDirectConvolutionLayerKernel.cpp
+++ b/src/core/CL/kernels/CLDirectConvolutionLayerKernel.cpp
@@ -108,7 +108,7 @@ void CLDirectConvolutionLayerKernel::configure(const ICLTensor *input, const ICL
         options.emplace("-DHAS_BIAS");
     }
 
-    if((gpu_target == GPUTarget::BIFROST) && (kernel_size <= 5) && (kernel_size != 1) && (_conv_stride_x == 1) && (_conv_stride_y == 1) && (input->info()->data_type() == DataType::F32))
+    if((gpu_target == GPUTarget::BIFROST) && (kernel_size <= 5) && (_conv_stride_x == 1) && (_conv_stride_y == 1) && (input->info()->data_type() == DataType::F32))
     {
         options.emplace("-DWEIGHTS_DEPTH=" + support::cpp11::to_string(_weights->info()->dimension(2)));
 
@@ -125,6 +125,14 @@ void CLDirectConvolutionLayerKernel::configure(const ICLTensor *input, const ICL
 
         switch(kernel_size)
         {
+            case 1:
+            {
+                num_elems_read_per_iteration_x    = 4;
+                num_elems_read_per_iteration_y    = 4;
+                num_elems_written_per_iteration_x = 4;
+                num_elems_written_per_iteration_y = 4;
+                break;
+            }
             case 3:
             {
                 num_elems_read_per_iteration_x    = 6;