xref: /aosp_15_r20/external/mesa3d/src/intel/vulkan/grl/gpu/input_dump.grl (revision 6104692788411f58d303aa86923a9ff6ecaded22) 
1//
2// Copyright (C) 2009-2021 Intel Corporation
3//
4// SPDX-License-Identifier: MIT
5//
6//
7
8module input_dump;
9
10kernel_module input_dumper("input_dump.cl")
11{
12    links lsc_intrinsics;
13
14    kernel opencl_kernel_find_max_used_byte_in_buff                  < kernelFunction="find_max_used_byte_in_buff" >;
15    kernel opencl_kernel_allocate_linear_offsets_for_vertex_buffers  < kernelFunction="allocate_linear_offsets_for_vertex_buffers" >;
16    kernel opencl_kernel_allocate_data_space_for_inputs              < kernelFunction="allocate_data_space_for_inputs" >;
17    kernel opencl_kernel_allocate_data_space_for_outputs             < kernelFunction="allocate_data_space_for_outputs" >;
18    kernel opencl_kernel_calc_outputs_data_size                      < kernelFunction="calc_outputs_data_size" >;
19    kernel opencl_kernel_write_output_data_op                        < kernelFunction="write_output_data_op" >;
20    kernel opencl_kernel_write_geo_data                              < kernelFunction="write_geo_data" >;
21    kernel opencl_kernel_write_input_build_op                        < kernelFunction="write_input_build_op" >;
22    kernel opencl_kernel_copy_instance_descriptors_array             < kernelFunction="copy_instance_descriptors_array" >;
23    kernel opencl_kernel_copy_instance_descriptors_array_of_ptrs     < kernelFunction="copy_instance_descriptors_array_of_ptrs" >;
24    kernel opencl_kernel_insert_copy_op                              < kernelFunction="insert_copy_op" >;
25    kernel opencl_kernel_copy_vertex_data                            < kernelFunction="copy_vertex_data" >;
26    kernel opencl_kernel_generate_unique_batch_id                    < kernelFunction="generate_unique_batch_id" >;
27    kernel opencl_kernel_finish_batch_dump_inputs                    < kernelFunction="finish_batch_dump_inputs" >;
28    kernel opencl_kernel_finish_batch_dump_outputs                   < kernelFunction="finish_batch_dump_outputs" >;
29}
30
31
32metakernel find_max_used_byte_in_buff(
33    qword indexBuffPtr,
34    qword vertexBufferUsedByteEnd,
35    dword IndexCount,
36    dword IndexFormat,
37    dword VertexCount,
38    qword VertexBufferByteStride,
39    dword numPhysThreads)
40{
41    dispatch opencl_kernel_find_max_used_byte_in_buff(numPhysThreads, 1, 1)   args(
42        indexBuffPtr,
43        vertexBufferUsedByteEnd,
44        IndexCount,
45        IndexFormat,
46        VertexCount,
47        VertexBufferByteStride);
48}
49
50metakernel allocate_linear_offsets_for_vertex_buffers(
51    qword batchPtrs,
52    qword m_VertexBufferUsedByteEnd,
53    qword m_VertexBufferOffset,
54    dword numVertexBuffers,
55    dword numPhysThreads)
56{
57    dispatch opencl_kernel_allocate_linear_offsets_for_vertex_buffers(numPhysThreads, 1, 1) args(
58        batchPtrs,
59        m_VertexBufferUsedByteEnd,
60        m_VertexBufferOffset,
61        numVertexBuffers);
62}
63
64metakernel allocate_data_space_for_inputs(
65    qword inputDumpMainBuffer,
66    qword batchPtrs,
67    dword nonVertexSize,
68    qword batchIdPtr)
69{
70    dispatch opencl_kernel_allocate_data_space_for_inputs(1, 1, 1) args(
71        inputDumpMainBuffer,
72        batchPtrs,
73        nonVertexSize,
74        batchIdPtr);
75}
76
77metakernel allocate_data_space_for_outputs(
78    qword inputDumpMainBuffer,
79    qword batchPtrs,
80    qword batchIdPtr)
81{
82    dispatch opencl_kernel_allocate_data_space_for_outputs(1, 1, 1) args(
83        inputDumpMainBuffer,
84        batchPtrs,
85        batchIdPtr);
86}
87
88metakernel calc_outputs_data_size(
89    qword pbi,
90    qword destOffsets,
91    qword numOutputs,
92    qword batchPtrs)
93{
94    dispatch opencl_kernel_calc_outputs_data_size(1, 1, 1) args(
95        pbi,
96        destOffsets,
97        numOutputs,
98        batchPtrs);
99}
100
101metakernel write_output_data_op(
102    qword batchPtrs,
103    qword destOffset,
104    qword src,
105    qword pbi)
106{
107    dispatch opencl_kernel_write_output_data_op(1, 1, 1) args(
108        batchPtrs,
109        destOffset,
110        src,
111        pbi);
112}
113
114metakernel write_geo_data(
115    qword batchPtrs,
116    qword srcDesc,
117    qword pVertexBufferOffsetInLinearisedUniqueVertexBuffers,
118    qword pVertexBufferSize,
119    qword dstDescOffset,
120    qword dstDataOffset,
121    dword numThreads)
122{
123    dispatch opencl_kernel_write_geo_data(numThreads, 1, 1) args(
124        batchPtrs,
125        srcDesc,
126        pVertexBufferOffsetInLinearisedUniqueVertexBuffers,
127        pVertexBufferSize,
128        dstDescOffset,
129        dstDataOffset,
130        numThreads);
131}
132
133metakernel write_input_build_op(
134    qword batchPtrs,
135    qword buildOpOffset,
136    qword srcBvh,
137    qword dstBvhAddr,
138    dword offsetToEnd,
139    dword flags,
140    dword numGeometries,
141    dword numInstances,
142    dword instArrayOfPtrs)
143
144{
145    dispatch opencl_kernel_write_input_build_op(1, 1, 1) args(
146        batchPtrs,
147        buildOpOffset,
148        srcBvh,
149        dstBvhAddr,
150        offsetToEnd,
151        flags,
152        numGeometries,
153        numInstances,
154        instArrayOfPtrs);
155}
156
157metakernel copy_instance_descriptors_array(
158    qword batchPtrs,
159    qword instanceDescArr,
160    qword offset,
161    dword numInstances,
162    dword numPhysThreads)
163{
164    dispatch opencl_kernel_copy_instance_descriptors_array(numPhysThreads, 1, 1) args(
165        batchPtrs,
166        instanceDescArr,
167        offset,
168        numInstances);
169}
170
171metakernel copy_instance_descriptors_array_of_ptrs(
172    qword batchPtrs,
173    qword instanceDescArrPtrs,
174    qword offset,
175    dword numInstances,
176    dword numPhysThreads)
177{
178    dispatch opencl_kernel_copy_instance_descriptors_array_of_ptrs(numPhysThreads, 1, 1) args(
179        batchPtrs,
180        instanceDescArrPtrs,
181        offset,
182        numInstances);
183}
184
185metakernel insert_copy_op(
186    qword batchPtrs,
187    qword offset,
188    qword src,
189    qword dst,
190    dword type)
191{
192    dispatch opencl_kernel_insert_copy_op(1, 1, 1) args(
193        batchPtrs,
194        offset,
195        src,
196        dst,
197        type);
198}
199
200metakernel copy_vertex_data(
201    qword desc,
202    qword src,
203    qword offset,
204    qword size)
205{
206    define byteSize REG0;
207    define numGroupsRqd REG1;
208    define shift REG2;
209    define minimum REG3;
210
211    shift = 6;
212    minimum = 1;
213    byteSize = load_dword(size);
214    numGroupsRqd = byteSize >> shift;
215    numGroupsRqd = numGroupsRqd + minimum;
216    DISPATCHDIM_X = numGroupsRqd.lo;
217    DISPATCHDIM_Y = 1;
218    DISPATCHDIM_Z = 1;
219
220    dispatch_indirect opencl_kernel_copy_vertex_data args(
221        desc,
222        src,
223        offset,
224        size);
225}
226
227metakernel generate_unique_batch_id(
228    qword batchIds,
229    dword batchIndex)
230{
231    dispatch opencl_kernel_generate_unique_batch_id(1, 1, 1) args(
232        batchIds,
233        batchIndex);
234}
235
236metakernel finish_batch_dump_inputs(
237    qword batchPtrs,
238    qword dumpMainBuffer)
239{
240    dispatch opencl_kernel_finish_batch_dump_inputs(1, 1, 1) args(
241        batchPtrs,
242        dumpMainBuffer);
243}
244
245metakernel finish_batch_dump_outputs(
246    qword batchPtrs,
247    qword dumpMainBuffer)
248{
249    dispatch opencl_kernel_finish_batch_dump_outputs(1, 1, 1) args(
250        batchPtrs,
251        dumpMainBuffer);
252}
253