From 2aa7fd011a4baff52dceb00a71b3674f819df8fc Mon Sep 17 00:00:00 2001 From: Sang-Hoon Park Date: Wed, 18 Sep 2019 13:39:00 +0100 Subject: COMPMID-2601 [CL] add mixed precision support to PoolingLayer * PoolingLayerInfo is updated with a new flag. * CL Kernel is updated to use FP32 accumulation. * CL pooling layer testscases are added for mixed precision. * Reference pooling layer is updated to use FP32 accumulation. Change-Id: I4ab2167cc7f86c86293cf50a0ca5119c04dc9c7e Signed-off-by: Sang-Hoon Park Reviewed-on: https://review.mlplatform.org/c/1973 Comments-Addressed: Arm Jenkins Tested-by: Arm Jenkins Reviewed-by: VidhyaSudhan Loganathan --- src/core/CL/cl_kernels/pooling_layer.cl | 294 ++++++++++++++++---------------- 1 file changed, 151 insertions(+), 143 deletions(-) (limited to 'src/core/CL/cl_kernels/pooling_layer.cl') diff --git a/src/core/CL/cl_kernels/pooling_layer.cl b/src/core/CL/cl_kernels/pooling_layer.cl index 6b2da0b87f..c8b5e07b47 100644 --- a/src/core/CL/cl_kernels/pooling_layer.cl +++ b/src/core/CL/cl_kernels/pooling_layer.cl @@ -38,7 +38,7 @@ #define DIV_OP(x, y) (x * (1.f / y)) #define SQRT_OP(x) sqrt((x)) -#define DIV_OP_NHWC(x, y) (x * (VEC_DATA_TYPE(float, 8))(1.f / y)) +#define DIV_OP_NHWC(x, y) (x * (VEC_DATA_TYPE(ACC_DATA_TYPE, 8))(1.f / y)) #if STRIDE_X == 1 #define POOLING3x3(res, input, output) POOLING3x3_STRIDE1(res, input, output) @@ -48,121 +48,129 @@ #define POOLING3x3(res, input, output) POOLING3x3_STRIDE3(res, input, output) #endif /* STRIDE_X == 3 */ -#define POOLING3x3_STRIDE1(res, input, output) \ - ({ \ - VEC_DATA_TYPE(DATA_TYPE, 4) \ - data00 = vload4(0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 0, 0)); \ - VEC_DATA_TYPE(DATA_TYPE, 2) \ - data01 = vload2(0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 0, 0) + 4); \ - VEC_DATA_TYPE(DATA_TYPE, 4) \ - data10 = vload4(0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 1, 0)); \ - VEC_DATA_TYPE(DATA_TYPE, 2) \ - data11 = vload2(0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 1, 0) + 4); \ - VEC_DATA_TYPE(DATA_TYPE, 4) \ - data20 = vload4(0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 2, 0)); \ - VEC_DATA_TYPE(DATA_TYPE, 2) \ - data21 = vload2(0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 2, 0) + 4); \ - data00 = POW2_OP(data00, 4); \ - data01 = POW2_OP(data01, 2); \ - data10 = POW2_OP(data10, 4); \ - data11 = POW2_OP(data11, 2); \ - data20 = POW2_OP(data20, 4); \ - data21 = POW2_OP(data21, 2); \ +#if defined(FP_MIXED_PRECISION) +#define CONVERT_TO_ACC_DATA_TYPE(x, n) CONVERT(x, VEC_DATA_TYPE(ACC_DATA_TYPE, n)) +#define VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(n, offset, ptr) \ + CONVERT_TO_ACC_DATA_TYPE(vload##n(offset, ptr), n) +#else /* defined(FP_MIXED_PRECISION) */ +#define VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(n, offset, ptr) vload##n(offset, ptr) +#endif /* defined(FP_MIXED_PRECISION) */ + +#define POOLING3x3_STRIDE1(res, input, output) \ + ({ \ + VEC_DATA_TYPE(ACC_DATA_TYPE, 4) \ + data00 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(4, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 0, 0)); \ + VEC_DATA_TYPE(ACC_DATA_TYPE, 2) \ + data01 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(2, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 0, 0) + 4); \ + VEC_DATA_TYPE(ACC_DATA_TYPE, 4) \ + data10 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(4, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 1, 0)); \ + VEC_DATA_TYPE(ACC_DATA_TYPE, 2) \ + data11 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(2, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 1, 0) + 4); \ + VEC_DATA_TYPE(ACC_DATA_TYPE, 4) \ + data20 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(4, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 2, 0)); \ + VEC_DATA_TYPE(ACC_DATA_TYPE, 2) \ + data21 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(2, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 2, 0) + 4); \ + data00 = POW2_OP(data00, 4); \ + data01 = POW2_OP(data01, 2); \ + data10 = POW2_OP(data10, 4); \ + data11 = POW2_OP(data11, 2); \ + data20 = POW2_OP(data20, 4); \ + data21 = POW2_OP(data21, 2); \ \ - VEC_DATA_TYPE(DATA_TYPE, 8) \ - values00 = (VEC_DATA_TYPE(DATA_TYPE, 8))(data00.s01212323); \ - VEC_DATA_TYPE(DATA_TYPE, 4) \ - values01 = (VEC_DATA_TYPE(DATA_TYPE, 4))(data01.s0, data00.s3, data01.s01); \ - VEC_DATA_TYPE(DATA_TYPE, 8) \ - values10 = (VEC_DATA_TYPE(DATA_TYPE, 8))(data10.s01212323); \ - VEC_DATA_TYPE(DATA_TYPE, 4) \ - values11 = (VEC_DATA_TYPE(DATA_TYPE, 4))(data11.s0, data10.s3, data11.s01); \ - VEC_DATA_TYPE(DATA_TYPE, 8) \ - values20 = (VEC_DATA_TYPE(DATA_TYPE, 8))(data20.s01212323); \ - VEC_DATA_TYPE(DATA_TYPE, 4) \ - values21 = (VEC_DATA_TYPE(DATA_TYPE, 4))(data21.s0, data20.s3, data21.s01); \ + VEC_DATA_TYPE(ACC_DATA_TYPE, 8) \ + values00 = (VEC_DATA_TYPE(ACC_DATA_TYPE, 8))(data00.s01212323); \ + VEC_DATA_TYPE(ACC_DATA_TYPE, 4) \ + values01 = (VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(data01.s0, data00.s3, data01.s01); \ + VEC_DATA_TYPE(ACC_DATA_TYPE, 8) \ + values10 = (VEC_DATA_TYPE(ACC_DATA_TYPE, 8))(data10.s01212323); \ + VEC_DATA_TYPE(ACC_DATA_TYPE, 4) \ + values11 = (VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(data11.s0, data10.s3, data11.s01); \ + VEC_DATA_TYPE(ACC_DATA_TYPE, 8) \ + values20 = (VEC_DATA_TYPE(ACC_DATA_TYPE, 8))(data20.s01212323); \ + VEC_DATA_TYPE(ACC_DATA_TYPE, 4) \ + values21 = (VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(data21.s0, data20.s3, data21.s01); \ \ - values00 = POOL_OP(values00, values10); \ - values01 = POOL_OP(values01, values11); \ - values00 = POOL_OP(values00, values20); \ - values01 = POOL_OP(values01, values21); \ + values00 = POOL_OP(values00, values10); \ + values01 = POOL_OP(values01, values11); \ + values00 = POOL_OP(values00, values20); \ + values01 = POOL_OP(values01, values21); \ \ - res = POOL_OP((VEC_DATA_TYPE(DATA_TYPE, 4))(values00.s036, values01.s1), (VEC_DATA_TYPE(DATA_TYPE, 4))(values00.s147, values01.s2)); \ - res = POOL_OP(res, (VEC_DATA_TYPE(DATA_TYPE, 4))(values00.s25, values01.s03)); \ + res = POOL_OP((VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(values00.s036, values01.s1), (VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(values00.s147, values01.s2)); \ + res = POOL_OP(res, (VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(values00.s25, values01.s03)); \ }) -#define POOLING3x3_STRIDE2(res, input, output) \ - ({ \ - VEC_DATA_TYPE(DATA_TYPE, 8) \ - data00 = vload8(0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 0, 0)); \ - DATA_TYPE data01 = *((__global DATA_TYPE *)tensor3D_offset(&input, 0, 0, 0) + 8); \ - VEC_DATA_TYPE(DATA_TYPE, 8) \ - data10 = vload8(0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 1, 0)); \ - DATA_TYPE data11 = *((__global DATA_TYPE *)tensor3D_offset(&input, 0, 1, 0) + 8); \ - VEC_DATA_TYPE(DATA_TYPE, 8) \ - data20 = vload8(0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 2, 0)); \ - DATA_TYPE data21 = *((__global DATA_TYPE *)tensor3D_offset(&input, 0, 2, 0) + 8); \ - data00 = POW2_OP(data00, 8); \ - data01 = POW2_OP(data01, 1); \ - data10 = POW2_OP(data10, 8); \ - data11 = POW2_OP(data11, 1); \ - data20 = POW2_OP(data20, 8); \ - data21 = POW2_OP(data21, 1); \ +#define POOLING3x3_STRIDE2(res, input, output) \ + ({ \ + VEC_DATA_TYPE(ACC_DATA_TYPE, 8) \ + data00 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(8, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 0, 0)); \ + ACC_DATA_TYPE data01 = (ACC_DATA_TYPE)(*((__global DATA_TYPE *)tensor3D_offset(&input, 0, 0, 0) + 8)); \ + VEC_DATA_TYPE(ACC_DATA_TYPE, 8) \ + data10 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(8, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 1, 0)); \ + ACC_DATA_TYPE data11 = (ACC_DATA_TYPE)(*((__global DATA_TYPE *)tensor3D_offset(&input, 0, 1, 0) + 8)); \ + VEC_DATA_TYPE(ACC_DATA_TYPE, 8) \ + data20 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(8, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 2, 0)); \ + ACC_DATA_TYPE data21 = (ACC_DATA_TYPE)(*((__global DATA_TYPE *)tensor3D_offset(&input, 0, 2, 0) + 8)); \ + data00 = POW2_OP(data00, 8); \ + data01 = POW2_OP(data01, 1); \ + data10 = POW2_OP(data10, 8); \ + data11 = POW2_OP(data11, 1); \ + data20 = POW2_OP(data20, 8); \ + data21 = POW2_OP(data21, 1); \ \ - VEC_DATA_TYPE(DATA_TYPE, 8) \ - values00 = (VEC_DATA_TYPE(DATA_TYPE, 8))(data00.s01223445); \ - VEC_DATA_TYPE(DATA_TYPE, 4) \ - values01 = (VEC_DATA_TYPE(DATA_TYPE, 4))(data00.s667, data01); \ - VEC_DATA_TYPE(DATA_TYPE, 8) \ - values10 = (VEC_DATA_TYPE(DATA_TYPE, 8))(data10.s01223445); \ - VEC_DATA_TYPE(DATA_TYPE, 4) \ - values11 = (VEC_DATA_TYPE(DATA_TYPE, 4))(data10.s667, data11); \ - VEC_DATA_TYPE(DATA_TYPE, 8) \ - values20 = (VEC_DATA_TYPE(DATA_TYPE, 8))(data20.s01223445); \ - VEC_DATA_TYPE(DATA_TYPE, 4) \ - values21 = (VEC_DATA_TYPE(DATA_TYPE, 4))(data20.s667, data21); \ + VEC_DATA_TYPE(ACC_DATA_TYPE, 8) \ + values00 = (VEC_DATA_TYPE(ACC_DATA_TYPE, 8))(data00.s01223445); \ + VEC_DATA_TYPE(ACC_DATA_TYPE, 4) \ + values01 = (VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(data00.s667, data01); \ + VEC_DATA_TYPE(ACC_DATA_TYPE, 8) \ + values10 = (VEC_DATA_TYPE(ACC_DATA_TYPE, 8))(data10.s01223445); \ + VEC_DATA_TYPE(ACC_DATA_TYPE, 4) \ + values11 = (VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(data10.s667, data11); \ + VEC_DATA_TYPE(ACC_DATA_TYPE, 8) \ + values20 = (VEC_DATA_TYPE(ACC_DATA_TYPE, 8))(data20.s01223445); \ + VEC_DATA_TYPE(ACC_DATA_TYPE, 4) \ + values21 = (VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(data20.s667, data21); \ \ - values00 = POOL_OP(values00, values10); \ - values01 = POOL_OP(values01, values11); \ - values00 = POOL_OP(values00, values20); \ - values01 = POOL_OP(values01, values21); \ + values00 = POOL_OP(values00, values10); \ + values01 = POOL_OP(values01, values11); \ + values00 = POOL_OP(values00, values20); \ + values01 = POOL_OP(values01, values21); \ \ - res = POOL_OP((VEC_DATA_TYPE(DATA_TYPE, 4))(values00.s036, values01.s1), (VEC_DATA_TYPE(DATA_TYPE, 4))(values00.s147, values01.s2)); \ - res = POOL_OP(res, (VEC_DATA_TYPE(DATA_TYPE, 4))(values00.s25, values01.s03)); \ + res = POOL_OP((VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(values00.s036, values01.s1), (VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(values00.s147, values01.s2)); \ + res = POOL_OP(res, (VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(values00.s25, values01.s03)); \ }) -#define POOLING3x3_STRIDE3(res, input, output) \ - ({ \ - VEC_DATA_TYPE(DATA_TYPE, 8) \ - data00 = vload8(0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 0, 0)); \ - VEC_DATA_TYPE(DATA_TYPE, 4) \ - data01 = vload4(0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 0, 0) + 8); \ - VEC_DATA_TYPE(DATA_TYPE, 8) \ - data10 = vload8(0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 1, 0)); \ - VEC_DATA_TYPE(DATA_TYPE, 4) \ - data11 = vload4(0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 1, 0) + 8); \ - VEC_DATA_TYPE(DATA_TYPE, 8) \ - data20 = vload8(0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 2, 0)); \ - VEC_DATA_TYPE(DATA_TYPE, 4) \ - data21 = vload4(0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 2, 0) + 8); \ - data00 = POW2_OP(data00, 8); \ - data01 = POW2_OP(data01, 4); \ - data10 = POW2_OP(data10, 8); \ - data11 = POW2_OP(data11, 4); \ - data20 = POW2_OP(data20, 8); \ - data21 = POW2_OP(data21, 4); \ +#define POOLING3x3_STRIDE3(res, input, output) \ + ({ \ + VEC_DATA_TYPE(ACC_DATA_TYPE, 8) \ + data00 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(8, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 0, 0)); \ + VEC_DATA_TYPE(ACC_DATA_TYPE, 4) \ + data01 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(4, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 0, 0) + 8); \ + VEC_DATA_TYPE(ACC_DATA_TYPE, 8) \ + data10 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(8, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 1, 0)); \ + VEC_DATA_TYPE(ACC_DATA_TYPE, 4) \ + data11 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(4, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 1, 0) + 8); \ + VEC_DATA_TYPE(ACC_DATA_TYPE, 8) \ + data20 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(8, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 2, 0)); \ + VEC_DATA_TYPE(ACC_DATA_TYPE, 4) \ + data21 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(4, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 2, 0) + 8); \ + data00 = POW2_OP(data00, 8); \ + data01 = POW2_OP(data01, 4); \ + data10 = POW2_OP(data10, 8); \ + data11 = POW2_OP(data11, 4); \ + data20 = POW2_OP(data20, 8); \ + data21 = POW2_OP(data21, 4); \ \ - data00 = POOL_OP(data00, data10); \ - data01 = POOL_OP(data01, data11); \ - data00 = POOL_OP(data00, data20); \ - data01 = POOL_OP(data01, data21); \ + data00 = POOL_OP(data00, data10); \ + data01 = POOL_OP(data01, data11); \ + data00 = POOL_OP(data00, data20); \ + data01 = POOL_OP(data01, data21); \ \ - res = POOL_OP((VEC_DATA_TYPE(DATA_TYPE, 4))(data00.s036, data01.s1), (VEC_DATA_TYPE(DATA_TYPE, 4))(data00.s147, data01.s2)); \ - res = POOL_OP(res, (VEC_DATA_TYPE(DATA_TYPE, 4))(data00.s25, data01.s03)); \ + res = POOL_OP((VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(data00.s036, data01.s1), (VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(data00.s147, data01.s2)); \ + res = POOL_OP(res, (VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(data00.s25, data01.s03)); \ }) -DATA_TYPE calculate_avg_scale(const int pool_size_x, const int pool_size_y, const int upper_bound_w, const int upper_bound_h, - const int pad_x, const int pad_y, const int stride_x, const int stride_y) +ACC_DATA_TYPE calculate_avg_scale(const int pool_size_x, const int pool_size_y, const int upper_bound_w, const int upper_bound_h, + const int pad_x, const int pad_y, const int stride_x, const int stride_y) { int start_x = get_global_id(0) * stride_x - pad_x; int start_y = get_global_id(1) * stride_y - pad_y; @@ -210,10 +218,10 @@ __kernel void pooling_layer_2( Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output); // Load data - VEC_DATA_TYPE(DATA_TYPE, 2) - data0 = vload2(0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 0, 0)); - VEC_DATA_TYPE(DATA_TYPE, 2) - data1 = vload2(0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 1, 0)); + VEC_DATA_TYPE(ACC_DATA_TYPE, 2) + data0 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(2, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 0, 0)); + VEC_DATA_TYPE(ACC_DATA_TYPE, 2) + data1 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(2, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 1, 0)); #if defined(POOL_L2) // Raise to power of 2 for L2 Pooling @@ -222,8 +230,8 @@ __kernel void pooling_layer_2( #endif /* defined(POOL_L2) */ // Perform calculations - data0 = POOL_OP(data0, data1); - DATA_TYPE res = POOL_OP(data0.s0, data0.s1); + data0 = POOL_OP(data0, data1); + ACC_DATA_TYPE res = POOL_OP(data0.s0, data0.s1); #if defined(POOL_AVG) || defined(POOL_L2) // Divide by pool region in case of average or l2 pooling @@ -236,7 +244,7 @@ __kernel void pooling_layer_2( #endif /* defined(POOL_L2) */ // Store result - *(__global DATA_TYPE *)output.ptr = res; + *(__global DATA_TYPE *)output.ptr = (DATA_TYPE)res; } /** Performs a pooling function of pool size equal to 3 @@ -274,12 +282,12 @@ __kernel void pooling_layer_3( Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output); // Load data - VEC_DATA_TYPE(DATA_TYPE, 3) - data0 = vload3(0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 0, 0)); - VEC_DATA_TYPE(DATA_TYPE, 3) - data1 = vload3(0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 1, 0)); - VEC_DATA_TYPE(DATA_TYPE, 3) - data2 = vload3(0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 2, 0)); + VEC_DATA_TYPE(ACC_DATA_TYPE, 3) + data0 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(3, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 0, 0)); + VEC_DATA_TYPE(ACC_DATA_TYPE, 3) + data1 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(3, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 1, 0)); + VEC_DATA_TYPE(ACC_DATA_TYPE, 3) + data2 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(3, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 2, 0)); #if defined(POOL_L2) // Raise to power of 2 for L2 Pooling @@ -289,9 +297,9 @@ __kernel void pooling_layer_3( #endif /* defined(POOL_L2) */ // Perform calculations - data0 = POOL_OP(data0, data1); - data0 = POOL_OP(data0, data2); - DATA_TYPE res = POOL_OP(POOL_OP(data0.s0, data0.s1), data0.s2); + data0 = POOL_OP(data0, data1); + data0 = POOL_OP(data0, data2); + ACC_DATA_TYPE res = POOL_OP(POOL_OP(data0.s0, data0.s1), data0.s2); #if defined(POOL_AVG) || defined(POOL_L2) // Divide by pool region in case of average pooling @@ -304,7 +312,7 @@ __kernel void pooling_layer_3( #endif /* defined(POOL_L2) */ // Store result - *(__global DATA_TYPE *)output.ptr = res; + *(__global DATA_TYPE *)output.ptr = (DATA_TYPE)res; } #if defined(POOLING3x3) @@ -312,7 +320,7 @@ __kernel void pooling_layer_3( #define CONVERT_OP(data_type) convert_##data_type##4 #define CONVERT_VECTOR4(data_type) CONVERT_OP(data_type) -VEC_DATA_TYPE(DATA_TYPE, 4) +VEC_DATA_TYPE(ACC_DATA_TYPE, 4) calculate_avg_scale4(const int pool_size, const int upper_bound_w, const int upper_bound_h, const int pad_x, const int pad_y, const int stride_x, const int stride_y) { @@ -324,7 +332,7 @@ calculate_avg_scale4(const int pool_size, const int upper_bound_w, const int upp start_x = max((int4)0, start_x); start_y = max(0, start_y); #endif /* defined(EXCLUDE_PADDING) */ - return (VEC_DATA_TYPE(DATA_TYPE, 4))(1.f) / CONVERT_VECTOR4(DATA_TYPE)(((int4)(end_y - start_y)) * (end_x - start_x)); + return (VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(1.f) / CONVERT_VECTOR4(ACC_DATA_TYPE)(((int4)(end_y - start_y)) * (end_x - start_x)); } /** Performs an optimized pooling function of pool size equal to 3 when the stride_x is less equal than 3 @@ -361,7 +369,7 @@ __kernel void pooling_layer_optimized_3( Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT(input); Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output); - VEC_DATA_TYPE(DATA_TYPE, 4) + VEC_DATA_TYPE(ACC_DATA_TYPE, 4) res; // Perform pooling 3x3 for 4 output elements @@ -377,7 +385,7 @@ __kernel void pooling_layer_optimized_3( res = SQRT_OP(res); #endif /* defined(POOL_L2) */ - vstore4(res, 0, (__global DATA_TYPE *)output.ptr); + vstore4(CONVERT(res, VEC_DATA_TYPE(DATA_TYPE, 4)), 0, (__global DATA_TYPE *)output.ptr); } #endif // defined(POOLING3x3) @@ -431,9 +439,9 @@ __kernel void pooling_layer_MxN_nchw( Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT(input); Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output); - VEC_DATA_TYPE(DATA_TYPE, 8) - vdata = INITIAL_VALUE; - DATA_TYPE sdata = INITIAL_VALUE; + VEC_DATA_TYPE(ACC_DATA_TYPE, 8) + vdata = INITIAL_VALUE; + ACC_DATA_TYPE sdata = INITIAL_VALUE; // Load data for(int y = 0; y < POOL_SIZE_Y; y++) @@ -441,8 +449,8 @@ __kernel void pooling_layer_MxN_nchw( int x = 0; for(; x <= ((int)POOL_SIZE_X - 8); x += 8) { - VEC_DATA_TYPE(DATA_TYPE, 8) - data0 = vload8(0, (__global DATA_TYPE *)tensor3D_offset(&input, x, y, 0)); + VEC_DATA_TYPE(ACC_DATA_TYPE, 8) + data0 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(8, 0, (__global DATA_TYPE *)tensor3D_offset(&input, x, y, 0)); #if defined(POOL_L2) // Raise to power of 2 for L2 Pooling data0 *= data0; @@ -453,7 +461,7 @@ __kernel void pooling_layer_MxN_nchw( // Leftover for(; x < (int)POOL_SIZE_X; ++x) { - DATA_TYPE data0 = *((__global DATA_TYPE *)tensor3D_offset(&input, x, y, 0)); + ACC_DATA_TYPE data0 = (ACC_DATA_TYPE)(*((__global DATA_TYPE *)tensor3D_offset(&input, x, y, 0))); #if defined(POOL_L2) // Raise to power of 2 for L2 Pooling data0 *= data0; @@ -463,12 +471,12 @@ __kernel void pooling_layer_MxN_nchw( } // Reduce result - VEC_DATA_TYPE(DATA_TYPE, 4) + VEC_DATA_TYPE(ACC_DATA_TYPE, 4) reduce4 = POOL_OP(vdata.s0123, vdata.s4567); - VEC_DATA_TYPE(DATA_TYPE, 2) - reduce2 = POOL_OP(reduce4.s01, reduce4.s23); - DATA_TYPE res = POOL_OP(reduce2.s0, reduce2.s1); - res = POOL_OP(res, sdata); + VEC_DATA_TYPE(ACC_DATA_TYPE, 2) + reduce2 = POOL_OP(reduce4.s01, reduce4.s23); + ACC_DATA_TYPE res = POOL_OP(reduce2.s0, reduce2.s1); + res = POOL_OP(res, sdata); #if defined(POOL_AVG) || defined(POOL_L2) // Divide by pool region in case of average pooling @@ -481,12 +489,12 @@ __kernel void pooling_layer_MxN_nchw( #endif /* defined(POOL_L2) */ // Store result - *(__global DATA_TYPE *)output.ptr = res; + *(__global DATA_TYPE *)output.ptr = (DATA_TYPE)res; } #endif // defined(POOL_SIZE_X) && defined(POOL_SIZE_Y) -float calculate_avg_scale_nhwc(const int pool_size_x, const int pool_size_y, int upper_bound_w, int upper_bound_h, - const int pad_x, const int pad_y, const int stride_x, const int stride_y) +ACC_DATA_TYPE calculate_avg_scale_nhwc(const int pool_size_x, const int pool_size_y, int upper_bound_w, int upper_bound_h, + const int pad_x, const int pad_y, const int stride_x, const int stride_y) { int start_x = get_global_id(1) * stride_x - pad_x; #if defined(DST_DEPTH) @@ -553,7 +561,7 @@ __kernel void pooling_layer_MxN_nhwc( Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output); #endif /* defined(DST_DEPTH) */ - VEC_DATA_TYPE(float, 8) + VEC_DATA_TYPE(ACC_DATA_TYPE, 8) vdata = INITIAL_VALUE; const int idx_width = get_global_id(1) * STRIDE_X; @@ -572,18 +580,18 @@ __kernel void pooling_layer_MxN_nhwc( x1 = select(x1, PAD_X - idx_width - 1, y != y1); #if defined(DST_DEPTH) - VEC_DATA_TYPE(DATA_TYPE, 8) - data0 = vload8(0, (__global DATA_TYPE *)tensor4D_offset(&input, 0, x1 - PAD_X, y1 - PAD_Y, 0)); + VEC_DATA_TYPE(ACC_DATA_TYPE, 8) + data0 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(8, 0, (__global DATA_TYPE *)tensor4D_offset(&input, 0, x1 - PAD_X, y1 - PAD_Y, 0)); #else /* defined(DST_DEPTH) */ - VEC_DATA_TYPE(DATA_TYPE, 8) - data0 = vload8(0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, x1 - PAD_X, y1 - PAD_Y)); + VEC_DATA_TYPE(ACC_DATA_TYPE, 8) + data0 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(8, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, x1 - PAD_X, y1 - PAD_Y)); #endif /* defined(DST_DEPTH) */ #if defined(POOL_L2) // Raise to power of 2 for L2 Pooling data0 *= data0; #endif /* defined(POOL_L2) */ - vdata = POOL_OP(vdata, CONVERT(data0, float8)); + vdata = POOL_OP(vdata, CONVERT(data0, VEC_DATA_TYPE(ACC_DATA_TYPE, 8))); } } -- cgit v1.2.1