L1 projections

src/qonnx/analysis/inference_cost.py

@@ -201,10 +201,10 @@ def inference_cost_upsample(model, node, discount_sparsity):
     return ret
 
 
-def inference_cost(model, discount_sparsity=True):
+def inference_cost(model, discount_sparsity=True, cost_breakdown=False):
     "Ensure all nodes have unique names prior to calling this analysis pass."
 
-    node_costs = {}
+    ret, node_costs, nodes_per_optype = {}, {}, {}
     zero_cost_ops = [
         "MaxPool",
         "AveragePool",
@@ -240,13 +240,24 @@ def inference_cost(model, discount_sparsity=True):
         if node.op_type in inference_cost_fxn_map.keys():
             node_cost = inference_cost_fxn_map[node.op_type](model, node, discount_sparsity)
             node_costs[node.name] = node_cost
+            if node.op_type not in nodes_per_optype.keys():
+                new_optype = {}
+                new_optype[node.name] = node_cost
+                nodes_per_optype[node.op_type] = new_optype
+            else:
+                nodes_per_optype[node.op_type][node.name] = node_cost
         elif node.op_type in zero_cost_ops:
             continue
         else:
             unsupported_ops.add(node.op_type)
-
-    ret = aggregate_dict_keys(node_costs)
-    ret["unsupported"] = unsupported_ops
-    ret["discount_sparsity"] = discount_sparsity
-
+    total = aggregate_dict_keys(node_costs)
+    total["unsupported"] = unsupported_ops
+    total["discount_sparsity"] = discount_sparsity
+    ret["total_cost"] = total
+    if cost_breakdown:
+        optype_cost = {}
+        for optype, resources in nodes_per_optype.items():
+            optype_cost[optype] = aggregate_dict_keys(resources)
+        ret["optype_cost"] = optype_cost
+        ret["node_cost"] = node_costs
     return ret

src/qonnx/analysis/l0_resource_estimates.py

@@ -0,0 +1,227 @@
+# Copyright (c) 2024 Advanced Micro Devices, Inc.
+# # All rights reserved.
+# #
+# # Redistribution and use in source and binary forms, with or without
+# # modification, are permitted provided that the following conditions are met:
+# #
+# # * Redistributions of source code must retain the above copyright notice, this
+# #   list of conditions and the following disclaimer.
+# #
+# # * Redistributions in binary form must reproduce the above copyright notice,
+# #   this list of conditions and the following disclaimer in the documentation
+# #   and/or other materials provided with the distribution.
+# #
+# # * Neither the name of qonnx nor the names of its
+# #   contributors may be used to endorse or promote products derived from
+# #   this software without specific prior written permission.
+# #
+# # THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+# # AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+# # IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+# # DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
+# # FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+# # DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+# # SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+# # CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+# # OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# # OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+from qonnx.core.datatype import DataType
+
+"""DSP Type: a) None:
+                    For Fixed Points and floating point
+                        1) When dsp_type is None. All operations will be processed using LUTs.
+                        2) LUTs are calculated using: 1.1*b_width1*b_width2
+                        2) Example:
+                            a) op_mac_Int4_Int2: 1.1*4*2 = 8.8 LUTs.
+                            b) op_mac_Int8_INT8: 1.1*8*8 = 70.4 LUTs.
+                            c) op_mac_Int8_FLOAT16: 1.1*8*16 = 140.8 LUTs
+                            d) op_mac_FLOAT16_FLOAT16: 1.1*16*16 = 281.6 LUTs.
+
+             b) DSP48:
+                    For Fixed Points
+                        1) Everything less than 4 will be promoted to 4. For ex: INT2 will use the same resources as INT4.
+                        2) INT4: One dsp48 + 200 LUTs can accomodate 4 (4*4) bit mac.
+                                So, no of dsp's from mac's can be calculated as (0.25).mac_count + (200*0.5)*mac_count LUTs.
+                        3) Everything between 5 and 8 will be promoted to 8, Ex: INT6 will use the same resources as INT8.
+                        4) INT88: One dsp48 + 200 LUTs can accomodate 2 (8*8) bit mac. So,
+                                 no of dsp's from mac's can be calculated as (0.5).mac_count + (200*0.25)*mac_count LUTs.
+                    For Floating Points
+                        1) FLOAT32: 2 dsp + 700 LUT can accomodate 1 mac count.
+                        2) FLOAT16: 1 dsp + 400 LUT can accomodate 1 mac count.
+            c) DSP58:
+                    For Fixed Points
+                        1) INT8: One dsp58 can accomodate 3 (8*8) bit mac.
+                           So, no of dsp's from mac's can be calculated as (0.33)*mac_count.
+                        2) INT4: One dsp58 can accomodate 4 (4*4) bit mac.
+                           So, no of dsp's from mac's can be calculated as (0.25)*mac_count.
+                        3) INT16: 1 mac count requires 1 dsp.
+                    For Floating Points
+                        1) FLOAT32:  1 mac count requires 1 dsp.
+                        2) FLOAT16: 1 mac count requires 1 dsp.
+    Mapping strategy for On-Chip Memory (bits_per_res):
+                        a) 1 "BRAM" can accomodate 36*1024 = 36864 bits.
+                        b) 1 "URAM" can accomodate 288*1024 = 294912 bits.
+                        c) 1 LUT can accomodate 64 bits.
+"""
+resource_table = {
+    "FLOAT32": {"NONE": (0, 1100), "DSP48": (2, 700), "DSP58": (1, 0)},
+    "FLOAT16": {"NONE": (0, 1100), "DSP48": (1, 400), "DSP58": (1, 0)},
+    "INT32": {"NONE": (0, 1100), "DSP48": (1, 0), "DSP58": (1, 0)},
+    "INT16": {"NONE": (0, 282), "DSP48": (1, 0), "DSP58": (1, 0)},
+    "INT8": {"NONE": (0, 71), "DSP48": (0.5, 100), "DSP58": (0.33, 0)},
+    "INT4": {"NONE": (0, 18), "DSP48": (0.25, 50), "DSP58": (0.25, 0)},
+}
+
+bits_per_res = {"BRAM": 36864, "URAM": 294912, "LUT": 64}
+
+
+def ocm_resources(num_mem_bits, d_factor):
+    """Provides an estimate about the number of urams and brams required for the
+       on-chip memory depending upon the distribution factor.
+    Args:
+        num_mem_bits (int): Number of memory bits.
+        d_factor (float): Distribution factor between 0 and 1.
+                         To distribute memory between BRAM and URAM.
+    Returns:
+        A dictionary for ocm resources containing memory requirements for luts, brams and urams
+    """
+    if d_factor is None:
+        luts_req = num_mem_bits / bits_per_res["LUT"]  # neither bram nor uram.
+        ocm_res = {"LUT": luts_req}
+    elif d_factor == 1:  # everything in uram.
+        uram_req = num_mem_bits / bits_per_res["URAM"]  # URAM: 288kbit/URAM
+        ocm_res = {"URAM": uram_req}
+    elif d_factor == 0:  # everything in bram 36kbit/BRAM
+        bram_req = num_mem_bits / bits_per_res["BRAM"]
+        ocm_res = {"BRAM": bram_req}
+    else:  # both bram and uram.
+        uram_por, bram_por = d_factor, 1 - d_factor
+        bram_req = (bram_por * num_mem_bits) / bits_per_res["BRAM"]
+        uram_req = (uram_por * num_mem_bits) / bits_per_res["URAM"]
+        ocm_res = {"BRAM": bram_req, "URAM": uram_req}
+    return ocm_res
+
+
+def promoting_datatype(dtype, b_width):
+    """Datatype promoting criterion. Only used when DSPs are used for processing.
+    Args:
+        dtype (str): conatining "INT" or "FLOAT".
+        b_width (int): precision of the respective datatype.
+    Returns:
+        Returns promoted datatype and precision value."""
+
+    if "INT" in dtype:
+        promoted_dtype = "INT"
+        if b_width <= 4:
+            promoted_bwidth = 4
+        elif 4 < b_width <= 8:
+            promoted_bwidth = 8
+        elif 8 < b_width <= 16:
+            promoted_bwidth = 16
+        else:
+            promoted_bwidth = 32
+    elif "FLOAT" in dtype:
+        promoted_dtype = "FLOAT"
+        if b_width <= 16:
+            promoted_bwidth = 16
+        else:
+            promoted_bwidth = 32
+    else:
+        raise Exception("Unsupported data type")
+
+    return promoted_dtype, promoted_bwidth
+
+
+def dtype_casting(dtype1, dtype2, b_width1, b_width2):
+    """Implementing datatype promotion."""
+
+    promoted_dtype1, promoted_bwidth1 = promoting_datatype(dtype1, b_width1)  # either INT or FLOAT
+    promoted_dtype2, promoted_bwidth2 = promoting_datatype(dtype2, b_width2)
+
+    if promoted_dtype1 == promoted_dtype2:  # same datatype
+        if promoted_bwidth1 == promoted_bwidth2:  # same precision.
+            dtype = promoted_dtype1 + str(promoted_bwidth1)  # can also use dtype_2 + new_bwidth2
+        else:  # different precision.
+            if promoted_bwidth1 >= promoted_bwidth2:
+                dtype = promoted_dtype1 + str(promoted_bwidth1)
+            else:
+                dtype = promoted_dtype2 + str(promoted_bwidth2)
+    else:  # dtype_1 != dtype_2 (Different datatype and same/different precision)
+        if promoted_dtype1 == "FLOAT":  # with different datatypes, using float and it's respective precision.
+            dtype = promoted_dtype1 + str(promoted_bwidth1)
+        else:
+            dtype = promoted_dtype2 + str(promoted_bwidth2)
+
+    return dtype
+
+
+def core_resources(inf_cost, dsp_type, bwidth_lower_limit, bwidth_upper_limit):
+    """Provide estimate resources required for the processing ("CORE"), assuming maximum unfolding.
+    Args:
+        inf_cost (dict): Inference cost dict.
+        dsp_type (str): None OR "DSP48" OR "DSP58". Default to None.
+        bwidth_lower_limit (int): Default to 8. It indicates bit values less than 8 will be processed using LUTs.
+        bwidth_upper_limit (int): Default to 32. It indicates bit values less than 32 will be processed using LUTs.
+    Returns:
+        A dictionary containing CORE resource estimates."""
+
+    dsp_res_mac = 0
+    lut_res_mac = 0
+    for i in inf_cost.keys():
+        if "op_mac" in i:
+            mac_count = inf_cost[i]
+            detail_list = i.split("_")
+            dtype1, dtype2 = detail_list[-1], detail_list[-2]
+            b_width1, b_width2 = DataType[dtype1].bitwidth(), DataType[dtype2].bitwidth()
+            if dsp_type is None:  # Computing everything in LUTs.
+                lut_res_mac += 1.1 * b_width1 * b_width2 * mac_count
+                dsp_comp = "DSP"  # default name for DSP and dsp_res_mac = 0
+            else:  # dsp_type == "DSP48" or dsp_type == "DSP58"
+                if (b_width1 < bwidth_lower_limit or b_width2 < bwidth_lower_limit) or (
+                    b_width1 > bwidth_upper_limit or b_width2 > bwidth_upper_limit
+                ):  # Computing everything in LUTs.
+                    lut_res_mac += 1.1 * b_width1 * b_width2 * mac_count  # dsp_res_mac = 0
+                else:
+                    casted_dtype = dtype_casting(dtype1, dtype2, b_width1, b_width2)
+                    casted_bwidth = DataType[casted_dtype].bitwidth()
+                    if casted_bwidth > bwidth_upper_limit:  # Computing everything in LUTs.
+                        lut_res_mac += (
+                            1.1 * b_width1 * b_width2 * mac_count
+                        )  # original bwidth values are used, since dsp_res_mac = 0.
+                    else:
+                        dsp_res_mac += (
+                            resource_table[casted_dtype][dsp_type][0] * mac_count
+                        )  # at index zero, we expect to have dsp factor.
+                        lut_res_mac += (
+                            resource_table[casted_dtype][dsp_type][1] * mac_count
+                        )  # at index one, we expect to have lut factor.
+                dsp_comp = dsp_type  # assigning name as per dsp type.
+        else:
+            continue
+
+    core_res = {"LUT": lut_res_mac, dsp_comp: dsp_res_mac}
+
+    return core_res
+
+
+def l0_resource_estimates(inf_cost, dsp_type=None, bwidth_lower_limit=8, bwidth_upper_limit=32, d_factor=None):
+    """Provide estimate resources required for the processing ("CORE") and memory ("OCM"), assuming maximum unfolding.
+    Args:
+        inf_cost (dict): Inference cost dict.
+        dsp_type (str): None OR "DSP48" OR "DSP58". Default to None.
+        bram_type (str): Default to "BRAM". It can be BRAM, BRAM36, BRAM_36K, BRAM_18K.
+        bwidth_lower_limit (int): Default to 8. It indicates bit values less than 8 will be processed using LUTs.
+        bwidth_upper_limit (int): Default to 32. It indicates bit values less than 32 will be processed using LUTs.
+        d_factor (float): Default to 1. It can have values between 0 and 1.
+    Returns:
+        A dictionary containing CORE and OCM resource estimates."""
+
+    core_res = core_resources(inf_cost, dsp_type, bwidth_lower_limit, bwidth_upper_limit)
+
+    num_mem_bits = inf_cost["total_mem_w_bits"]
+    ocm_res = ocm_resources(num_mem_bits, d_factor)
+
+    est_res_req = {"CORE": core_res, "OCM": ocm_res}
+
+    return est_res_req