nihui
/
ncnn
Not watched
1
0
Fork 0
Code
Releases 46 Wiki evaluate Activity
Issues 0 Pull Requests 0 Datasets Model Cloudbrain HPC
You can not select more than 25 topics Topics must start with a chinese character,a letter or number, can include dashes ('-') and can be up to 35 characters long.
Tree: cfcb1cffa9
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from 'cfcb1cffa9'
${ noResults }
ncnn/src/layer/vulkan/shader/deconvolution_pack8_gemm.comp

203 lines
8.8 kB
Raw Blame History 

  1. // Tencent is pleased to support the open source community by making ncnn available.

  2. //

  3. // Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.

  4. //

  5. // Licensed under the BSD 3-Clause License (the "License"); you may not use this file except

  6. // in compliance with the License. You may obtain a copy of the License at

  7. //

  8. // https://opensource.org/licenses/BSD-3-Clause

  9. //

  10. // Unless required by applicable law or agreed to in writing, software distributed

  11. // under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR

  12. // CONDITIONS OF ANY KIND, either express or implied. See the License for the

  13. // specific language governing permissions and limitations under the License.

  14. 

  15. #version 450

  16. 

  17. #if NCNN_fp16_storage

  18. #extension GL_EXT_shader_16bit_storage: require

  19. struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };

  20. #endif

  21. #if NCNN_fp16_arithmetic

  22. #extension GL_EXT_shader_explicit_arithmetic_types_float16: require

  23. #endif

  24. 

  25. #define shape_constant_id_offset 0

  26. layout (constant_id = shape_constant_id_offset + 0) const int w = 0;

  27. layout (constant_id = shape_constant_id_offset + 1) const int h = 0;

  28. layout (constant_id = shape_constant_id_offset + 2) const int c = 0;

  29. layout (constant_id = shape_constant_id_offset + 3) const int cstep = 0;

  30. 

  31. layout (constant_id = shape_constant_id_offset + 4) const int outw = 0;

  32. layout (constant_id = shape_constant_id_offset + 5) const int outh = 0;

  33. 

  34. #if NCNN_image_shader

  35. layout (binding = 0) uniform unfp sampler3D bottom_blob;

  36. layout (binding = 1, imfmtc4) writeonly uniform unfp image3D col_blob;

  37. layout (binding = 2) uniform unfp sampler3D weight_blob;

  38. #else

  39. layout (binding = 0) readonly buffer bottom_blob { sfpvec8 bottom_blob_data[]; };

  40. layout (binding = 1) writeonly buffer col_blob { sfpvec8 col_blob_data[]; };

  41. layout (binding = 2) readonly buffer weight_blob { sfpvec8 weight_data[]; };

  42. #endif

  43. 

  44. layout (push_constant) uniform parameter

  45. {

  46.     int w;

  47.     int h;

  48.     int c;

  49.     int cstep;

  50. 

  51.     int outw;

  52.     int outh;

  53. } p;

  54. 

  55. void main()

  56. {

  57.     int gx = int(gl_GlobalInvocationID.x) * 4;

  58.     int gy = int(gl_GlobalInvocationID.y);

  59. 

  60.     if (gx >= psc(outw) || gy >= psc(outh))

  61.         return;

  62. 

  63.     afpvec8 sum0 = afpvec8(afpvec4(0.f), afpvec4(0.f));

  64.     afpvec8 sum1 = afpvec8(afpvec4(0.f), afpvec4(0.f));

  65.     afpvec8 sum2 = afpvec8(afpvec4(0.f), afpvec4(0.f));

  66.     afpvec8 sum3 = afpvec8(afpvec4(0.f), afpvec4(0.f));

  67. 

  68. #if NCNN_image_shader

  69.     ivec4 gx4 = gx + ivec4(0, 1, 2, 3);

  70. 

  71.     ivec4 sy4 = gx4 / psc(w);

  72.     ivec4 sx4 = gx4 % psc(w);

  73. 

  74.     for (int z = 0; z < psc(c); z++)

  75.     {

  76.         afpvec8 v0 = image3d_ld8(bottom_blob, ivec3(sx4.r, sy4.r, z));

  77.         afpvec8 v1 = image3d_ld8(bottom_blob, ivec3(sx4.g, sy4.g, z));

  78.         afpvec8 v2 = image3d_ld8(bottom_blob, ivec3(sx4.b, sy4.b, z));

  79.         afpvec8 v3 = image3d_ld8(bottom_blob, ivec3(sx4.a, sy4.a, z));

  80. 

  81.         afpvec8 k0 = image3d_ld8(weight_blob, ivec3(z * 8 + 0, gy, 0));

  82.         afpvec8 k1 = image3d_ld8(weight_blob, ivec3(z * 8 + 1, gy, 0));

  83.         afpvec8 k2 = image3d_ld8(weight_blob, ivec3(z * 8 + 2, gy, 0));

  84.         afpvec8 k3 = image3d_ld8(weight_blob, ivec3(z * 8 + 3, gy, 0));

  85.         afpvec8 k4 = image3d_ld8(weight_blob, ivec3(z * 8 + 4, gy, 0));

  86.         afpvec8 k5 = image3d_ld8(weight_blob, ivec3(z * 8 + 5, gy, 0));

  87.         afpvec8 k6 = image3d_ld8(weight_blob, ivec3(z * 8 + 6, gy, 0));

  88.         afpvec8 k7 = image3d_ld8(weight_blob, ivec3(z * 8 + 7, gy, 0));

  89. 

  90.         // sum += v * k

  91.         sum0[0].r += dot(v0[0], k0[0]) + dot(v0[1], k0[1]);

  92.         sum0[0].g += dot(v0[0], k1[0]) + dot(v0[1], k1[1]);

  93.         sum0[0].b += dot(v0[0], k2[0]) + dot(v0[1], k2[1]);

  94.         sum0[0].a += dot(v0[0], k3[0]) + dot(v0[1], k3[1]);

  95.         sum0[1].r += dot(v0[0], k4[0]) + dot(v0[1], k4[1]);

  96.         sum0[1].g += dot(v0[0], k5[0]) + dot(v0[1], k5[1]);

  97.         sum0[1].b += dot(v0[0], k6[0]) + dot(v0[1], k6[1]);

  98.         sum0[1].a += dot(v0[0], k7[0]) + dot(v0[1], k7[1]);

  99. 

  100.         sum1[0].r += dot(v1[0], k0[0]) + dot(v1[1], k0[1]);

  101.         sum1[0].g += dot(v1[0], k1[0]) + dot(v1[1], k1[1]);

  102.         sum1[0].b += dot(v1[0], k2[0]) + dot(v1[1], k2[1]);

  103.         sum1[0].a += dot(v1[0], k3[0]) + dot(v1[1], k3[1]);

  104.         sum1[1].r += dot(v1[0], k4[0]) + dot(v1[1], k4[1]);

  105.         sum1[1].g += dot(v1[0], k5[0]) + dot(v1[1], k5[1]);

  106.         sum1[1].b += dot(v1[0], k6[0]) + dot(v1[1], k6[1]);

  107.         sum1[1].a += dot(v1[0], k7[0]) + dot(v1[1], k7[1]);

  108. 

  109.         sum2[0].r += dot(v2[0], k0[0]) + dot(v2[1], k0[1]);

  110.         sum2[0].g += dot(v2[0], k1[0]) + dot(v2[1], k1[1]);

  111.         sum2[0].b += dot(v2[0], k2[0]) + dot(v2[1], k2[1]);

  112.         sum2[0].a += dot(v2[0], k3[0]) + dot(v2[1], k3[1]);

  113.         sum2[1].r += dot(v2[0], k4[0]) + dot(v2[1], k4[1]);

  114.         sum2[1].g += dot(v2[0], k5[0]) + dot(v2[1], k5[1]);

  115.         sum2[1].b += dot(v2[0], k6[0]) + dot(v2[1], k6[1]);

  116.         sum2[1].a += dot(v2[0], k7[0]) + dot(v2[1], k7[1]);

  117. 

  118.         sum3[0].r += dot(v3[0], k0[0]) + dot(v3[1], k0[1]);

  119.         sum3[0].g += dot(v3[0], k1[0]) + dot(v3[1], k1[1]);

  120.         sum3[0].b += dot(v3[0], k2[0]) + dot(v3[1], k2[1]);

  121.         sum3[0].a += dot(v3[0], k3[0]) + dot(v3[1], k3[1]);

  122.         sum3[1].r += dot(v3[0], k4[0]) + dot(v3[1], k4[1]);

  123.         sum3[1].g += dot(v3[0], k5[0]) + dot(v3[1], k5[1]);

  124.         sum3[1].b += dot(v3[0], k6[0]) + dot(v3[1], k6[1]);

  125.         sum3[1].a += dot(v3[0], k7[0]) + dot(v3[1], k7[1]);

  126.     }

  127. #else

  128.     int v_offset = gx;

  129.     int w_offset = gy * psc(c) * 8;

  130. 

  131.     for (int z = 0; z < psc(c); z++)

  132.     {

  133.         afpvec8 v0 = buffer_ld8(bottom_blob_data, v_offset + 0);

  134.         afpvec8 v1 = buffer_ld8(bottom_blob_data, v_offset + 1);

  135.         afpvec8 v2 = buffer_ld8(bottom_blob_data, v_offset + 2);

  136.         afpvec8 v3 = buffer_ld8(bottom_blob_data, v_offset + 3);

  137. 

  138.         afpvec8 k0 = buffer_ld8(weight_data, w_offset + 0);

  139.         afpvec8 k1 = buffer_ld8(weight_data, w_offset + 1);

  140.         afpvec8 k2 = buffer_ld8(weight_data, w_offset + 2);

  141.         afpvec8 k3 = buffer_ld8(weight_data, w_offset + 3);

  142.         afpvec8 k4 = buffer_ld8(weight_data, w_offset + 4);

  143.         afpvec8 k5 = buffer_ld8(weight_data, w_offset + 5);

  144.         afpvec8 k6 = buffer_ld8(weight_data, w_offset + 6);

  145.         afpvec8 k7 = buffer_ld8(weight_data, w_offset + 7);

  146. 

  147.         // sum += v * k

  148.         sum0[0].r += dot(v0[0], k0[0]) + dot(v0[1], k0[1]);

  149.         sum0[0].g += dot(v0[0], k1[0]) + dot(v0[1], k1[1]);

  150.         sum0[0].b += dot(v0[0], k2[0]) + dot(v0[1], k2[1]);

  151.         sum0[0].a += dot(v0[0], k3[0]) + dot(v0[1], k3[1]);

  152.         sum0[1].r += dot(v0[0], k4[0]) + dot(v0[1], k4[1]);

  153.         sum0[1].g += dot(v0[0], k5[0]) + dot(v0[1], k5[1]);

  154.         sum0[1].b += dot(v0[0], k6[0]) + dot(v0[1], k6[1]);

  155.         sum0[1].a += dot(v0[0], k7[0]) + dot(v0[1], k7[1]);

  156. 

  157.         sum1[0].r += dot(v1[0], k0[0]) + dot(v1[1], k0[1]);

  158.         sum1[0].g += dot(v1[0], k1[0]) + dot(v1[1], k1[1]);

  159.         sum1[0].b += dot(v1[0], k2[0]) + dot(v1[1], k2[1]);

  160.         sum1[0].a += dot(v1[0], k3[0]) + dot(v1[1], k3[1]);

  161.         sum1[1].r += dot(v1[0], k4[0]) + dot(v1[1], k4[1]);

  162.         sum1[1].g += dot(v1[0], k5[0]) + dot(v1[1], k5[1]);

  163.         sum1[1].b += dot(v1[0], k6[0]) + dot(v1[1], k6[1]);

  164.         sum1[1].a += dot(v1[0], k7[0]) + dot(v1[1], k7[1]);

  165. 

  166.         sum2[0].r += dot(v2[0], k0[0]) + dot(v2[1], k0[1]);

  167.         sum2[0].g += dot(v2[0], k1[0]) + dot(v2[1], k1[1]);

  168.         sum2[0].b += dot(v2[0], k2[0]) + dot(v2[1], k2[1]);

  169.         sum2[0].a += dot(v2[0], k3[0]) + dot(v2[1], k3[1]);

  170.         sum2[1].r += dot(v2[0], k4[0]) + dot(v2[1], k4[1]);

  171.         sum2[1].g += dot(v2[0], k5[0]) + dot(v2[1], k5[1]);

  172.         sum2[1].b += dot(v2[0], k6[0]) + dot(v2[1], k6[1]);

  173.         sum2[1].a += dot(v2[0], k7[0]) + dot(v2[1], k7[1]);

  174. 

  175.         sum3[0].r += dot(v3[0], k0[0]) + dot(v3[1], k0[1]);

  176.         sum3[0].g += dot(v3[0], k1[0]) + dot(v3[1], k1[1]);

  177.         sum3[0].b += dot(v3[0], k2[0]) + dot(v3[1], k2[1]);

  178.         sum3[0].a += dot(v3[0], k3[0]) + dot(v3[1], k3[1]);

  179.         sum3[1].r += dot(v3[0], k4[0]) + dot(v3[1], k4[1]);

  180.         sum3[1].g += dot(v3[0], k5[0]) + dot(v3[1], k5[1]);

  181.         sum3[1].b += dot(v3[0], k6[0]) + dot(v3[1], k6[1]);

  182.         sum3[1].a += dot(v3[0], k7[0]) + dot(v3[1], k7[1]);

  183. 

  184.         v_offset += psc(cstep);

  185.         w_offset += 8;

  186.     }

  187. #endif

  188. 

  189. #if NCNN_image_shader

  190.     image3d_st8(col_blob, ivec3(gx4.r, gy, 0), sum0);

  191.     image3d_st8(col_blob, ivec3(gx4.g, gy, 0), sum1);

  192.     image3d_st8(col_blob, ivec3(gx4.b, gy, 0), sum2);

  193.     image3d_st8(col_blob, ivec3(gx4.a, gy, 0), sum3);

  194. #else

  195.     const int gi = gy * psc(outw) + gx;

  196. 

  197.     buffer_st8(col_blob_data, gi, sum0);

  198.     if (gx + 1 < psc(outw)) buffer_st8(col_blob_data, gi + 1, sum1);

  199.     if (gx + 2 < psc(outw)) buffer_st8(col_blob_data, gi + 2, sum2);

  200.     if (gx + 3 < psc(outw)) buffer_st8(col_blob_data, gi + 3, sum3);

  201. #endif

  202. }