da4ml 0.3.2__py3-none-any.whl → 0.3.3__py3-none-any.whl

da4ml/_version.py

@@ -28,7 +28,7 @@ version_tuple: VERSION_TUPLE
 commit_id: COMMIT_ID
 __commit_id__: COMMIT_ID
 
-__version__ = version = '0.3.2'
-__version_tuple__ = version_tuple = (0, 3, 2)
+__version__ = version = '0.3.3'
+__version_tuple__ = version_tuple = (0, 3, 3)
 
 __commit_id__ = commit_id = None

da4ml/codegen/cpp/cpp_codegen.py

@@ -46,7 +46,7 @@ def ssa_gen(sol: Solution, print_latency: bool, typestr_fn: Callable[[bool | int
         match op.opcode:
             case -1:
                 # Input marker
-                val = f'inp[{ops[op.id0].id0}]'
+                val = f'inp[{op.id0}]'
             case 0 | 1:
                 # Common a+/-b<<shift op
                 ref1 = f'bit_shift<{op.data}>(v{op.id1})' if op.data != 0 else f'v{op.id1}'

da4ml/codegen/cpp/source/binder_util.hh

@@ -9,31 +9,26 @@ constexpr bool _openmp = true;
 constexpr bool _openmp = false;
 #endif
 
-template <typename CONFIG_T, typename T> void _inference(T *c_inp, T *c_out, size_t n_samples)
-{
+template <typename CONFIG_T, typename T> void _inference(T *c_inp, T *c_out, size_t n_samples) {
     typename CONFIG_T::inp_t in_fixed_buf[CONFIG_T::N_inp];
     typename CONFIG_T::out_t out_fixed_buf[CONFIG_T::N_out];
 
-    for(size_t i = 0; i < n_samples; ++i)
-        {
-            size_t offset_in = i * CONFIG_T::N_inp;
-            size_t offset_out = i * CONFIG_T::N_out;
-            for(size_t j = 0; j < CONFIG_T::N_inp; ++j)
-                {
-                    in_fixed_buf[j] = c_inp[offset_in + j];
-                }
+    for (size_t i = 0; i < n_samples; ++i) {
+        size_t offset_in = i * CONFIG_T::N_inp;
+        size_t offset_out = i * CONFIG_T::N_out;
+        for (size_t j = 0; j < CONFIG_T::N_inp; ++j) {
+            in_fixed_buf[j] = c_inp[offset_in + j];
+        }
 
-            CONFIG_T::f(in_fixed_buf, out_fixed_buf);
+        CONFIG_T::f(in_fixed_buf, out_fixed_buf);
 
-            for(size_t j = 0; j < CONFIG_T::N_out; ++j)
-                {
-                    c_out[offset_out + j] = out_fixed_buf[j];
-                }
+        for (size_t j = 0; j < CONFIG_T::N_out; ++j) {
+            c_out[offset_out + j] = out_fixed_buf[j];
         }
+    }
 }
 
-template <typename CONFIG_T, typename T> void batch_inference(T *c_inp, T *c_out, size_t n_samples)
-{
+template <typename CONFIG_T, typename T> void batch_inference(T *c_inp, T *c_out, size_t n_samples) {
 #ifdef _OPENMP
     size_t n_max_threads = omp_get_max_threads();
     size_t n_samples_per_thread = std::max<size_t>(n_samples / n_max_threads, 32);
@@ -41,15 +36,14 @@ template <typename CONFIG_T, typename T> void batch_inference(T *c_inp, T *c_out
     n_thread += (n_samples % n_samples_per_thread) ? 1 : 0;
 
 #pragma omp parallel for num_threads(n_thread) schedule(static)
-    for(size_t i = 0; i < n_thread; ++i)
-        {
-            size_t start = i * n_samples_per_thread;
-            size_t end = std::min<size_t>(start + n_samples_per_thread, n_samples);
-            size_t n_samples_this_thread = end - start;
-            size_t offset_in = start * CONFIG_T::N_inp;
-            size_t offset_out = start * CONFIG_T::N_out;
-            _inference<CONFIG_T, T>(&c_inp[offset_in], &c_out[offset_out], n_samples_this_thread);
-        }
+    for (size_t i = 0; i < n_thread; ++i) {
+        size_t start = i * n_samples_per_thread;
+        size_t end = std::min<size_t>(start + n_samples_per_thread, n_samples);
+        size_t n_samples_this_thread = end - start;
+        size_t offset_in = start * CONFIG_T::N_inp;
+        size_t offset_out = start * CONFIG_T::N_out;
+        _inference<CONFIG_T, T>(&c_inp[offset_in], &c_out[offset_out], n_samples_this_thread);
+    }
 #else
     _inference<CONFIG_T, T>(c_inp, c_out, n_samples);
 #endif

da4ml/codegen/cpp/source/vitis_bitshift.hh

@@ -1,14 +1,16 @@
 #pragma once
-#include "ap_types/ap_fixed.h"
+#include "ap_fixed.h"
 
-template <int s, int b, int i, ap_q_mode Q, ap_o_mode O, int N> ap_fixed<b, i + s> bit_shift(ap_fixed<b, i, Q, O, N> x) {
+template <int s, int b, int i, ap_q_mode Q, ap_o_mode O, int N>
+ap_fixed<b, i + s> bit_shift(ap_fixed<b, i, Q, O, N> x) {
 #pragma HLS INLINE
     ap_fixed<b, i + s> r;
     r.range() = x.range();
     return r;
 };
 
-template <int s, int b, int i, ap_q_mode Q, ap_o_mode O, int N> ap_ufixed<b, i + s> bit_shift(ap_ufixed<b, i, Q, O, N> x) {
+template <int s, int b, int i, ap_q_mode Q, ap_o_mode O, int N>
+ap_ufixed<b, i + s> bit_shift(ap_ufixed<b, i, Q, O, N> x) {
 #pragma HLS INLINE
     ap_ufixed<b, i + s> r;
     r.range() = x.range();

da4ml/converter/hgq2/parser.py

@@ -6,8 +6,8 @@ import keras
 import numpy as np
 from keras import KerasTensor, Operation
 
-from ...trace import FixedVariableArray, HWConfig, comb_trace
-from ...trace.fixed_variable_array import FixedVariableArrayInput
+from ...trace import FixedVariableArray, FixedVariableArrayInput, HWConfig, comb_trace
+from ...trace.fixed_variable import FixedVariable
 from .replica import _registry
 
 
@@ -55,6 +55,8 @@ def replace_tensors(tensor_map: dict[KerasTensor, FixedVariableArray], obj: Any)
 def _flatten_arr(args: Any) -> FixedVariableArray:
     if isinstance(args, FixedVariableArray):
         return np.ravel(args)  # type: ignore
+    if isinstance(args, FixedVariable):
+        return FixedVariableArray(np.array([args]))
     if not isinstance(args, Sequence):
         return None  # type: ignore
     args = [_flatten_arr(a) for a in args]

da4ml/trace/fixed_variable.py

@@ -265,6 +265,10 @@ class FixedVariable:
     def __sub__(self, other: 'FixedVariable|int|float|Decimal'):
         return self + (-other)
 
+    def __truediv__(self, other: 'int|float|Decimal'):
+        assert not isinstance(other, FixedVariable), 'Division by variable is not supported'
+        return self * (1 / other)
+
     def __mul__(self, other: 'FixedVariable|int|float|Decimal') -> 'FixedVariable':
         if other == 0:
             return FixedVariable(0, 0, 1, hwconf=self.hwconf, opr='const')

da4ml/trace/fixed_variable_array.py

@@ -164,6 +164,10 @@ class FixedVariableArray:
         latency: NDArray[np.floating] | float = 0.0,
         solver_options: dict[str, Any] | None = None,
     ):
+        mask = k + i + f <= 0
+        k = np.where(mask, 0, k)
+        i = np.where(mask, 0, i)
+        f = np.where(mask, 0, f)
         step = 2.0**-f
         _high = 2.0**i
         high, low = _high - step, -_high * k

da4ml/trace/ops/reduce_utils.py

@@ -99,7 +99,7 @@ def reduce(operator: Callable[[T, T], T], x: TA, axis: int | Sequence[int] | Non
     r = _arr.reshape(target_shape)  # type: ignore
 
     if isinstance(x, FixedVariableArray):
-        ret = FixedVariableArray(r, solver_config)
-        if ret.size == 1 and not keepdims:
-            return ret.ravel()[0]  # type: ignore
+        r = FixedVariableArray(r, solver_config)
+        if r.size == 1 and not keepdims:
+            return r.ravel()[0]  # type: ignore
     return r if r.size > 1 or keepdims else r.ravel()[0]  # type: ignore

{da4ml-0.3.2.dist-info → da4ml-0.3.3.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: da4ml
-Version: 0.3.2
+Version: 0.3.3
 Summary: Digital Arithmetic for Machine Learning
 Author-email: Chang Sun <chsun@cern.ch>
 License: GNU Lesser General Public License v3 (LGPLv3)

{da4ml-0.3.2.dist-info → da4ml-0.3.3.dist-info}/RECORD

@@ -1,5 +1,5 @@
 da4ml/__init__.py,sha256=IETRRvzsJvPMLu1kzzi8UN5FYaM5MhNaXH2A_ZKr2_w,469
-da4ml/_version.py,sha256=e8NqPtZ8fggRgk3GPrqZ_U_BDV8aSULw1u_Gn9NNbnk,704
+da4ml/_version.py,sha256=lemL_4Kl75FgrO6lVuFrrtw6-Dcf9wtXBalKkXuzkO4,704
 da4ml/cmvm/__init__.py,sha256=4Tbt913k9zP0w8R1p6Oss06v5jrManbUhskyHl6e-U0,154
 da4ml/cmvm/api.py,sha256=JpecMt6g8zutGh_uWT61_0iX8TuXct7-jq7N7HMIsgA,9626
 da4ml/cmvm/types.py,sha256=O8BuBZ2SyucxoXt_KbulAuHNgim7Ls3M6Ovw8prLgXM,21340
@@ -11,11 +11,11 @@ da4ml/cmvm/util/bit_decompose.py,sha256=SUco70HRYf4r1JU6BXwcgabDrhm_yAmucae5FC67
 da4ml/cmvm/util/mat_decompose.py,sha256=eSJNlXwx_jxgqt5vLJrSLQaeq2ZXu8j9mC4d-eq883M,4094
 da4ml/codegen/__init__.py,sha256=Chdh3oO_vLR4saLbT9VxBPz_0wlEzxJldFSZaVUJo7U,331
 da4ml/codegen/cpp/__init__.py,sha256=SIePoi_T4iJph50OQUosAnaVuLCckukYjLxp91Y8xQs,134
-da4ml/codegen/cpp/cpp_codegen.py,sha256=ot293c8aHBx7wy1R7hnB9IVI22jYMO0476ghYKD8ECA,6162
+da4ml/codegen/cpp/cpp_codegen.py,sha256=I3YcxK524_oJ7jebxOlRGuYbN2uCY5mpKACoQShqZxs,6153
 da4ml/codegen/cpp/hls_model.py,sha256=J5lnB8sAvMy0Bo5MSJOpgyUm1tzEJqBxgPTlOd38Gbg,8978
-da4ml/codegen/cpp/source/binder_util.hh,sha256=pBVmhXIDvdCr8n2wwYehc3Fpp60sWYrrZaDoP3x9JZE,1880
+da4ml/codegen/cpp/source/binder_util.hh,sha256=ClECVxcEynE_9i4jWCV4y1dnadG3wFqLZfjxg4qHFQQ,1752
 da4ml/codegen/cpp/source/build_binder.mk,sha256=RLu4TP28aJsveyMOHxuDRGEJVoIPMo9T8WyPtqnmtbQ,584
-da4ml/codegen/cpp/source/vitis_bitshift.hh,sha256=yFpYCVJ8gof-EzPjkIWWZYmdFh_wk133Pxzs7f61IQo,774
+da4ml/codegen/cpp/source/vitis_bitshift.hh,sha256=u8wjT_cRn7bXcbC5pH3-rS76ekRbwv-VWAAdaP52-dw,765
 da4ml/codegen/cpp/source/ap_types/ap_binary.h,sha256=yOcafu2IofstDqxn0wDq8vY3JIwZQ9H5z6IY1dEqMr0,2764
 da4ml/codegen/cpp/source/ap_types/ap_common.h,sha256=1hJY9uvKOdwRSSll5uehUISZR4tsSsQ1z4PNRUc44KU,10180
 da4ml/codegen/cpp/source/ap_types/ap_decl.h,sha256=z1HsH-2RSvSoofTZR7RHeqIfAnEYVuHcIu_ute9gjEg,6473
@@ -48,19 +48,19 @@ da4ml/codegen/verilog/source/shift_adder.v,sha256=qrpXBX9bhHI-o75v5zshOfq0giEATv
 da4ml/codegen/verilog/source/template.xdc,sha256=GlSRy8tw_orohSuUwUSNEYJLLkAAHttGTfLTcQqRQDg,1262
 da4ml/converter/__init__.py,sha256=x7J2PEXYZsVWffRAkucLxbwzzU404eaijMdLwdhBxtY,57
 da4ml/converter/hgq2/__init__.py,sha256=-gnT_7zXY-KQtPLxsqngwDKZ2TUIynn996pUjjB03B8,59
-da4ml/converter/hgq2/parser.py,sha256=O55QTrlkev0lvxiIweXlTGG9RPcfjdrJgpkZc-rwetg,5472
+da4ml/converter/hgq2/parser.py,sha256=Yc5V-B_aEslqIXXJihRi3GMjF9vMkmUQ2_yHMGHMPVo,5573
 da4ml/converter/hgq2/replica.py,sha256=aKi6BF2x4s3VUF1Q-__GE4-is9eSC3H8TGFDT05vTWc,16292
 da4ml/trace/__init__.py,sha256=dv-rti3t8iE0RqeThfOb40mAg8FZB2WkkGQq3enJft0,282
-da4ml/trace/fixed_variable.py,sha256=samW_xChnERsMaXVQz7aKUQJsIrnSHu2ox4x9dMzhR0,20918
-da4ml/trace/fixed_variable_array.py,sha256=1gGSc-ZmRG59sUXvgdN7pulG4XhacAGmgSmzq7nAhJ4,12846
+da4ml/trace/fixed_variable.py,sha256=7vaXFZToCVzPtUZcHv4aoqpqJp46SHUzSWTQijVT0os,21101
+da4ml/trace/fixed_variable_array.py,sha256=mJj9aU-jLCPVkFXrTbcRQndtUKEuhVwiFUGVSGX7PHE,12975
 da4ml/trace/pipeline.py,sha256=AVeO9BNpQlo_WO6S1nQl7RxiHs5VFRR10tWMg_36C2o,5354
 da4ml/trace/tracer.py,sha256=xnaVO4oTWwasfiEBqqeY9o60Lek3eX65IIbvB7JtVKQ,6099
 da4ml/trace/ops/__init__.py,sha256=fz5Cg7ZQqPkZlUj4bIOKY6aaoA1fX_G22TeA8I1n4qY,2166
 da4ml/trace/ops/conv_utils.py,sha256=Yn73t4F6Tcs1hBwK08L1DPOin2HYVcng4PSkU4vuZFo,8245
 da4ml/trace/ops/einsum_utils.py,sha256=ODofbvR98FwKBTDZsJ0ObbMjU9_GjPu5AbGuWX6sdCY,11453
-da4ml/trace/ops/reduce_utils.py,sha256=9bi-fizhl1BPy9quQzaWMs83eCDSRMFag2PuvqlVFgI,3500
-da4ml-0.3.2.dist-info/licenses/LICENSE,sha256=46mU2C5kSwOnkqkw9XQAJlhBL2JAf1_uCD8lVcXyMRg,7652
-da4ml-0.3.2.dist-info/METADATA,sha256=zZnCaLH3ndDuURdIXAZD37A06L0ommMlBzfuL93lG-E,4055
-da4ml-0.3.2.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-da4ml-0.3.2.dist-info/top_level.txt,sha256=N0tnKVwRqFiffFdeAzCgFq71hUNySh5-ITbNd6-R58Q,6
-da4ml-0.3.2.dist-info/RECORD,,
+da4ml/trace/ops/reduce_utils.py,sha256=vQjEUUbvnW8inAYJWHDzgy-PbgwIdHlH-uzPzSEvrSc,3494
+da4ml-0.3.3.dist-info/licenses/LICENSE,sha256=46mU2C5kSwOnkqkw9XQAJlhBL2JAf1_uCD8lVcXyMRg,7652
+da4ml-0.3.3.dist-info/METADATA,sha256=C3NAvObpQ5xNOmQQ-cE77AJMFevKJ0gCCO-BrlQpAeA,4055
+da4ml-0.3.3.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+da4ml-0.3.3.dist-info/top_level.txt,sha256=N0tnKVwRqFiffFdeAzCgFq71hUNySh5-ITbNd6-R58Q,6
+da4ml-0.3.3.dist-info/RECORD,,