| 97 | } // namespace |
| 98 | |
| 99 | std::vector<ComputeTaskDescriptorPtr> Softmax(int id, ValueId input_id, |
| 100 | ValueId output_id, |
| 101 | int channels_count) { |
| 102 | auto desc = std::make_shared<ComputeTaskDescriptor>(); |
| 103 | desc->id = id; |
| 104 | desc->is_linkable = false; |
| 105 | desc->shader_source = R"( |
| 106 | #include <metal_stdlib> |
| 107 | using namespace metal; |
| 108 | constant int src_channels = )"; |
| 109 | desc->shader_source += std::to_string(channels_count); |
| 110 | desc->shader_source += R"(; |
| 111 | $0 |
| 112 | kernel void ComputeFunction( |
| 113 | $1 |
| 114 | uint3 gid[[thread_position_in_grid]]) { |
| 115 | if (int(gid.x) >= size.x || int(gid.y) >= size.y) { |
| 116 | return; |
| 117 | } |
| 118 | float shift = 0.0f; |
| 119 | int remaining_channels = src_channels % 4; |
| 120 | |
| 121 | float sum = 0.0f; |
| 122 | for (int d = 0; d < src_channels / 4; ++d) { |
| 123 | int buffer_index = (d * size.y + gid.y) * size.x + gid.x; |
| 124 | sum += dot(float4(1.0f), exp(float4(input_buffer[buffer_index]) - shift)); |
| 125 | } |
| 126 | if (remaining_channels > 0) { |
| 127 | int buffer_index = ((src_channels / 4) * size.y + gid.y) * size.x + gid.x; |
| 128 | float4 last_element = float4(input_buffer[buffer_index]); |
| 129 | sum += exp(last_element.x - shift); |
| 130 | if (remaining_channels > 1) sum += exp(last_element.y - shift); |
| 131 | if (remaining_channels == 3) sum += exp(last_element.z - shift); |
| 132 | } |
| 133 | |
| 134 | for (int d = 0; d < (src_channels + 3) / 4; ++d) { |
| 135 | const int linear_index = (d * size.y + gid.y) * size.x + gid.x; |
| 136 | FLT4 value = FLT4(exp(float4(input_buffer[linear_index]) - shift) / sum); |
| 137 | $2 |
| 138 | output_buffer[linear_index] = value; |
| 139 | } |
| 140 | } |
| 141 | )"; |
| 142 | |
| 143 | desc->input_buffers = { |
| 144 | {input_id, "device FLT4* const input_buffer"}, |
| 145 | }; |
| 146 | |
| 147 | desc->output_buffer = {output_id, "device FLT4* output_buffer", |
| 148 | [input_id](const std::map<ValueId, BHWC>& buffers) { |
| 149 | return buffers.find(input_id)->second; |
| 150 | }}; |
| 151 | |
| 152 | desc->uniform_buffers = { |
| 153 | {"constant int2& size", |
| 154 | [output_id](const std::map<ValueId, BHWC>& buffers) { |
| 155 | const auto& dimension = buffers.find(output_id)->second; |
| 156 | std::vector<int> sizes{dimension.w, dimension.h}; |
no test coverage detected