|
732 | 732 | "name": "stdout",
|
733 | 733 | "output_type": "stream",
|
734 | 734 | "text": [
|
735 |
| - "ALPAKA_STATIC_ACC_MEM_GLOBAL const float bias_layer1[32] = {\n", |
| 735 | + "HOST_DEVICE_CONSTANT const float bias_layer1[32] = {\n", |
736 | 736 | "-1.3837075f, -0.0653152f, -0.7900129f, 0.0714758f, -1.1574365f, -1.4634879f, -0.9317133f, -0.1455518f, -0.0459635f, -0.2055620f, 0.0586231f, -0.8943899f, -0.1009487f, 0.0166031f, -0.5451909f, -0.1384538f, 1.2664700f, -1.8996916f, -0.0025585f, -0.1647783f, -1.9019107f, 0.0707104f, -0.2373025f, 0.0357050f, -0.0048417f, 2.3127339f, -0.0508943f, -0.1116435f, -0.1610904f, -1.6463890f, -1.0739423f, -0.0962902f };\n",
|
737 | 737 | "\n",
|
738 |
| - "ALPAKA_STATIC_ACC_MEM_GLOBAL const float wgtT_layer1[23][32] = {\n", |
| 738 | + "HOST_DEVICE_CONSTANT const float wgtT_layer1[23][32] = {\n", |
739 | 739 | "{ -0.1881404f, -0.0534256f, 1.6563641f, 0.0401664f, 2.8318353f, 1.5049738f, 1.4111555f, -0.2339872f, 0.0431970f, 0.1220361f, -0.0450153f, -1.6025578f, 0.0394025f, -0.3051167f, 1.9442217f, 0.1599094f, 0.1376955f, 2.4181051f, -0.0226484f, -0.1801709f, -0.4861264f, -0.0268545f, 0.5463807f, 0.2420150f, -0.1238829f, 0.2916382f, 0.1507791f, 0.7952659f, 0.2736979f, 3.2790639f, 1.2062043f, -0.0884467f },\n",
|
740 | 740 | "{ -0.0469924f, 0.2013927f, 0.0307775f, -0.1241788f, -0.0100412f, 0.0422375f, 0.0211071f, -0.0359304f, 0.0451861f, 0.0291862f, -0.2094866f, -0.0013007f, 0.1191471f, 0.0750159f, 0.0184378f, 0.0419437f, -0.0207304f, -0.0444109f, 0.0013400f, -0.0699210f, -0.0668742f, -0.0880825f, -0.0107244f, 0.0363424f, 0.1391699f, -0.0112885f, -0.0060098f, -0.0073863f, -0.0566143f, -0.0224207f, 0.0103718f, -0.0015193f },\n",
|
741 | 741 | "{ 0.4520382f, 0.1227609f, -1.3887709f, -0.0542129f, -3.2003114f, -0.8354173f, -1.3173198f, 0.3292131f, -0.1657729f, -0.1982902f, 0.1599589f, -0.0417666f, -0.1461042f, -1.3237997f, -5.3609071f, -0.0981676f, 0.2922535f, -1.8692241f, -0.0345302f, 0.1810613f, 0.4473544f, -0.0159401f, -0.7293931f, -1.4816793f, -0.1431545f, -0.0955672f, -0.2370718f, -0.7204540f, 0.8451244f, -3.4310548f, -1.3518151f, 0.1551731f },\n",
|
|
761 | 761 | "{ 0.0340557f, -0.1659652f, -0.0042457f, 0.0010229f, -2.1550148f, -0.4728722f, -1.3667214f, 0.2625635f, -0.0302200f, -0.0322885f, 0.0227866f, 0.6977839f, 0.0050141f, -1.6183628f, 0.0869662f, -0.0775411f, 0.4754244f, 0.4596581f, 2.1509945f, -0.0313832f, 0.0336208f, -0.1547154f, -0.6017126f, 0.0369996f, -0.1102583f, -0.5788267f, 0.0017006f, 2.6352038f, -1.7847317f, 1.7510574f, 2.1478791f, -0.2251654f },\n",
|
762 | 762 | "};\n",
|
763 | 763 | "\n",
|
764 |
| - "ALPAKA_STATIC_ACC_MEM_GLOBAL const float bias_layer2[32] = {\n", |
| 764 | + "HOST_DEVICE_CONSTANT const float bias_layer2[32] = {\n", |
765 | 765 | "-0.2689391f, 1.5461178f, -0.2424639f, 0.4424149f, -0.0411816f, -4.1070848f, 1.4709516f, -0.2439820f, -0.1750926f, 2.8802166f, -0.1573734f, -1.3724055f, 0.3671952f, 1.8267332f, 1.5655776f, -0.7323843f, 1.6318209f, 2.2198663f, -1.5951139f, -0.0870247f, 0.2806863f, -0.2407108f, 0.1310665f, -0.5246177f, 0.1914421f, -0.3386542f, -0.6310596f, 3.2995102f, 0.7519229f, -0.1565450f, -0.1496341f, 1.0073272f };\n",
|
766 | 766 | "\n",
|
767 |
| - "ALPAKA_STATIC_ACC_MEM_GLOBAL const float wgtT_layer2[32][32] = {\n", |
| 767 | + "HOST_DEVICE_CONSTANT const float wgtT_layer2[32][32] = {\n", |
768 | 768 | "{ -0.1731049f, 1.7775618f, -0.2532010f, -0.2902778f, -0.1392802f, 4.2428946f, -0.1866968f, -0.1800365f, -0.0634398f, 0.0763313f, 0.0472901f, -0.8030146f, 0.3161853f, -1.0713238f, -4.6514492f, -0.3908085f, 1.1607268f, 0.8834935f, -0.1194544f, -0.0785166f, 0.4967587f, -0.0558136f, -0.9601135f, -0.1001592f, 3.4427991f, -0.2144053f, -0.3632556f, 0.0117088f, 0.1742481f, -0.2540179f, -0.1705156f, -0.2627344f },\n",
|
769 | 769 | "{ -0.1478276f, -0.1659575f, 0.1602777f, -0.0758106f, 0.1067696f, -0.0247068f, -0.1123443f, -0.1724832f, -0.0013103f, -0.0685904f, 0.1537329f, 0.1042632f, -0.0360880f, -0.0679077f, 0.0672719f, 0.1597116f, -0.0150259f, 0.0367102f, -0.0545881f, -0.0693004f, -0.1008447f, -0.0672846f, -0.1395939f, -0.0324785f, -0.1051702f, -0.0530534f, -0.1019061f, -0.0921245f, 0.1195077f, 0.0453448f, 0.0257045f, -0.0622537f },\n",
|
770 | 770 | "{ -0.0363173f, -0.1990481f, -0.0452148f, 0.4074381f, -0.0731660f, -0.0823270f, 0.3154473f, -0.1909118f, -0.0165690f, 0.1325824f, -0.0760181f, 0.7768906f, -0.2702211f, -0.6023573f, 1.5904741f, 0.2384946f, 0.7610655f, -2.8705251f, 0.5754877f, -0.1587478f, -0.5708794f, -0.3421216f, 0.5023443f, 1.2806857f, 0.2158970f, -0.1364033f, -0.3398291f, 0.9066412f, -1.2935438f, 0.0273695f, -0.1850613f, -0.9301611f },\n",
|
|
799 | 799 | "{ -0.0246332f, 0.1165779f, 0.0255498f, -0.0601489f, 0.1545041f, -0.0977981f, 0.1242626f, -0.1533627f, -0.1294386f, -0.0231293f, -0.1460808f, 0.1763088f, 0.0953614f, -0.0716483f, -0.1003436f, 0.0804519f, 0.1373295f, -0.0686773f, 0.1198382f, 0.1519430f, 0.1640775f, -0.1675753f, 0.0790529f, -0.1521838f, 0.0378523f, 0.1039687f, -0.0701027f, 0.0509319f, 0.1355647f, 0.0978021f, 0.0391430f, 0.0241266f },\n",
|
800 | 800 | "};\n",
|
801 | 801 | "\n",
|
802 |
| - "ALPAKA_STATIC_ACC_MEM_GLOBAL const float bias_output_layer[1] = {\n", |
| 802 | + "HOST_DEVICE_CONSTANT const float bias_output_layer[1] = {\n", |
803 | 803 | "-0.7420582f };\n",
|
804 | 804 | "\n",
|
805 |
| - "ALPAKA_STATIC_ACC_MEM_GLOBAL const float wgtT_output_layer[32][1] = {\n", |
| 805 | + "HOST_DEVICE_CONSTANT const float wgtT_output_layer[32][1] = {\n", |
806 | 806 | "{ 0.0381968f },\n",
|
807 | 807 | "{ 1.0667214f },\n",
|
808 | 808 | "{ 0.0505496f },\n",
|
|
843 | 843 | "source": [
|
844 | 844 | "def print_formatted_weights_biases(weights, biases, layer_name):\n",
|
845 | 845 | " # Print biases\n",
|
846 |
| - " print(f\"ALPAKA_STATIC_ACC_MEM_GLOBAL const float bias_{layer_name}[{len(biases)}] = {{\")\n", |
| 846 | + " print(f\"HOST_DEVICE_CONSTANT const float bias_{layer_name}[{len(biases)}] = {{\")\n", |
847 | 847 | " print(\", \".join(f\"{b:.7f}f\" for b in biases) + \" };\")\n",
|
848 | 848 | " print()\n",
|
849 | 849 | "\n",
|
850 | 850 | " # Print weights\n",
|
851 |
| - " print(f\"ALPAKA_STATIC_ACC_MEM_GLOBAL const float wgtT_{layer_name}[{len(weights[0])}][{len(weights)}] = {{\")\n", |
| 851 | + " print(f\"HOST_DEVICE_CONSTANT const float wgtT_{layer_name}[{len(weights[0])}][{len(weights)}] = {{\")\n", |
852 | 852 | " for row in weights.T:\n",
|
853 | 853 | " formatted_row = \", \".join(f\"{w:.7f}f\" for w in row)\n",
|
854 | 854 | " print(f\"{{ {formatted_row} }},\")\n",
|
|
0 commit comments