|
1 | | -<!DOCTYPE html><html> <head><title>gpgpu playground</title><link rel="stylesheet" type="text/css" media="all" href="/learn-gpgpu/styles/index.css"><script type="module" src="/learn-gpgpu/_astro/hoisted.BVkJFRlA.js"></script> |
| 1 | +<!DOCTYPE html><html> <head><title>gpgpu playground</title><link rel="stylesheet" type="text/css" media="all" href="/learn-gpgpu/styles/index.css"><script type="module" src="/learn-gpgpu/_astro/hoisted.ozVI7Nxi.js"></script> |
2 | 2 | <script type="module" src="/learn-gpgpu/_astro/page.BO5rBSqo.js"></script></head> <body> <nav> <ul> <li><a href="/learn-gpgpu/docs">docs/</a></li> </ul> </nav> <div id="talvos"> <div> <main> <textarea class="module" autocorrect="off" autocomplete="off" autocapitalize="off" spellcheck="false" aria-label="SPIR-V module assembly (spv_text format)" wrap="off">; SPIR-V |
3 | | -; Version: 1.3 |
4 | | -OpCapability Shader ; TODO we need this for descriptors? really? |
| 3 | +; Version: 1.5 |
5 | 4 | OpCapability Kernel |
| 5 | +OpCapability BuffersTALVOS |
| 6 | +OpCapability ExecTALVOS |
| 7 | +OpCapability PhysicalStorageBufferAddresses |
| 8 | +OpExtension "SPV_TALVOS_buffers" |
| 9 | +OpExtension "SPV_TALVOS_exec" |
6 | 10 | OpMemoryModel Logical OpenCL |
7 | 11 |
|
8 | 12 | OpEntryPoint Kernel %main_fn "main" %gl_GlobalInvocationID |
9 | 13 |
|
10 | | -; TODO instead of DISPATCH, one of these: |
11 | | -;OpExecutionMode %main_fn LocalSize 16 1 1 |
12 | | -;OpExecutionMode %main_fn GlobalSize 16 1 1 |
| 14 | +OpExecutionGlobalSizeTALVOS %main_fn 16 1 1 |
13 | 15 |
|
14 | | -; TODO[seth]: is this a spec misread in talvos? Shouldn't the DISPATCH map to local size? |
15 | | -; (or is it execution mode dependent?) |
16 | 16 | OpDecorate %gl_GlobalInvocationID BuiltIn GlobalInvocationId |
| 17 | +OpDecorate %_arr_uint32_t ArrayStride 4 |
17 | 18 |
|
18 | | -; TODO instead of descriptors, static allocation via `OpBufferTALVOS 64` (maybe w/ %ty? and/or name?) |
19 | | -; TODO how do people pass data back & forth w/ OpenCL kernels for real? |
20 | | -OpDecorate %buf0 DescriptorSet 0 |
21 | | -OpDecorate %buf0 Binding 0 |
| 19 | +; types |
| 20 | + %void_t = OpTypeVoid |
| 21 | + %void_fn_t = OpTypeFunction %void_t |
| 22 | + %uint32_t = OpTypeInt 32 0 |
| 23 | + %gbl_id_t = OpTypeVector %uint32_t 3 |
22 | 24 |
|
23 | | -; TODO instead of `DUMP`, something like `OpDumpAtEndTALVOS %buffer_id` ? |
24 | | -; Or perhaps that's just implied by `OpBufferTALVOS` ? |
| 25 | + %arr_len = OpConstant %uint32_t 16 |
| 26 | + %_arr_uint32_t = OpTypeArray %uint32_t %arr_len |
25 | 27 |
|
26 | | - ; types |
27 | | - %void_t = OpTypeVoid |
28 | | -%void_fn_t = OpTypeFunction %void_t |
29 | | - %uint32_t = OpTypeInt 32 0 |
30 | | - %gbl_id_t = OpTypeVector %uint32_t 3 |
| 28 | + %_ptr_Input_gbl_id_t = OpTypePointer Input %gbl_id_t |
| 29 | + %_ptr_Input_uint32_t = OpTypePointer Input %uint32_t |
31 | 30 |
|
32 | | -%_arr_uint32_t = OpTypeRuntimeArray %uint32_t |
| 31 | +%_ptr_PhysicalStorageBuffer_uint32_t = OpTypePointer PhysicalStorageBuffer %uint32_t |
| 32 | +%_arr_PhysicalStorageBuffer_uint32_t = OpTypePointer PhysicalStorageBuffer %_arr_uint32_t |
33 | 33 |
|
34 | | -%_ptr_StorageBuffer_uint32_t = OpTypePointer StorageBuffer %uint32_t |
35 | | -%_arr_StorageBuffer_uint32_t = OpTypePointer StorageBuffer %_arr_uint32_t |
36 | | - %_ptr_Input_gbl_id_t = OpTypePointer Input %gbl_id_t |
37 | | - %_ptr_Input_uint32_t = OpTypePointer Input %uint32_t |
38 | 34 |
|
39 | | - |
40 | | - ; global arguments & constants |
| 35 | +; global arguments & constants |
41 | 36 | %n = OpConstant %uint32_t 0 |
42 | 37 | %gl_GlobalInvocationID = OpVariable %_ptr_Input_gbl_id_t Input |
43 | | - %buf0 = OpVariable %_arr_StorageBuffer_uint32_t StorageBuffer |
| 38 | + %buf0 = OpBufferTALVOS %_arr_PhysicalStorageBuffer_uint32_t PhysicalStorageBuffer 64 "a" |
44 | 39 |
|
45 | | - ; FILL_IDX entry point |
| 40 | +; FILL_IDX entry point |
46 | 41 | %main_fn = OpFunction %void_t None %void_fn_t |
47 | 42 | %0 = OpLabel |
48 | 43 | %2 = OpAccessChain %_ptr_Input_uint32_t %gl_GlobalInvocationID %n |
49 | | - %3 = OpLoad %uint32_t %2 |
50 | | - %4 = OpAccessChain %_ptr_StorageBuffer_uint32_t %buf0 %3 |
51 | | - OpStore %4 %3 |
| 44 | + %3 = OpLoad %uint32_t %2 Aligned 4 |
| 45 | + %4 = OpAccessChain %_ptr_PhysicalStorageBuffer_uint32_t %buf0 %3 |
| 46 | + OpStore %4 %3 Aligned 4 |
52 | 47 | OpReturn |
53 | 48 | OpFunctionEnd |
54 | | -</textarea> </main> <aside> <section> <textarea class="commands_todo_make_js_concatenate_all_the_commands inactive" autocorrect="off" autocomplete="off" autocapitalize="off" spellcheck="false" aria-label="Talvos commands (informational)" disabled wrap="off"># MODULE fill.spvasm</textarea> </section> <section> <label>ENTRY |
| 49 | +</textarea> </main> <aside> <section> <textarea class="commands_todo_make_js_concatenate_all_the_commands inactive" autocorrect="off" autocomplete="off" autocapitalize="off" spellcheck="false" aria-label="Talvos commands (informational)" disabled wrap="off"># MODULE fill_idx.spvasm</textarea> </section> <section> <label>ENTRY |
55 | 50 | <select name="entry"></select> </label> </section> <section> <textarea class="commands" style="height: 434px" autocorrect="off" autocomplete="off" autocapitalize="off" spellcheck="false" aria-label="Talvos commands (tcf format)" wrap="off"> |
56 | | -BUFFER a 64 UNINIT |
57 | | -# ^ bytes |
58 | | - |
59 | | -DESCRIPTOR_SET 0 0 0 a |
60 | | - |
61 | | -DISPATCH 16 1 1 |
62 | | -# ^ work items (here, "elements") |
63 | | - |
64 | | -DUMP UINT32 a |
65 | | -# ^ bits |
| 51 | +EXEC |
66 | 52 |
|
67 | | -# NB: a's size and the DISPATCH must agree; |
68 | | -# i.e. 16 elements == (64 bytes / (32 bits per element / 8 bits per byte)) |
69 | | -# a should be filled with the value of %FILL |
70 | 53 | </textarea> </section> </aside> </div> <div id="controls"> <button class="validate exec">validate</button> <button class="run exec">run</button> <button class="debug exec">debug</button> <div hidden> <button class="exception">exception</button> <button class="assertion">assertion</button> <button class="shrubbery">shrubbery</button> </div> <div class="debugger" hidden> <!-- cf. PipelineExecutor::help --> <button class="step">step</button> <button class="switch">switch</button> <button class="print">print</button> <button class="continue">continue</button> <button class="stop">stop</button> <!-- missing: [break, breakpoint, help, quit] --> </div> </div> <!-- TODO |
71 | 54 | can we pull out just this one wodget into a "component"? |
72 | 55 | it'd be nice to write some code for turning `102043253.21231222ms` into a human time |
|
0 commit comments