Read indirect dispatch counts from GPU buffers

Co-authored-by: Mac <me@itsmemac.com>
This commit is contained in:
nmzikandMac committed 2026-09-23 15:41:02 +04:00
1 parent d8a4412498
commit 79530b4c5e
7 files changed
+256 -31

No files matched your search

@@ -73,6 +73,9 @@ public:
void SetIndexBufferSize(uint32_t index_buffer_size);
void SetDrawIndirectArgsBaseAddress(uint64_t draw_indirect_args_base_addr);
void SetDispatchIndirectArgsBaseAddress(uint64_t dispatch_indirect_args_base_addr);
[[nodiscard]] uint64_t GetDispatchIndirectArgsBaseAddress() const {
return m_dispatch_indirect_args_base_addr;
}
void SetNumInstances(uint32_t num_instances);
void DrawIndex(DrawIndexArgs args);
void DrawIndexOffset(uint32_t index_offset, uint32_t index_count);
@@ -99,7 +102,7 @@ public:
void TriggerEopEventAtEndOfPipe(uint32_t interrupt_context_id);
void DispatchDirect(uint32_t thread_group_x, uint32_t thread_group_y, uint32_t thread_group_z,
uint32_t mode);
void DispatchIndirect(uint32_t data_offset, uint32_t mode);
void DispatchIndirect(uint64_t args_addr, uint32_t mode);
void WaitFlipDone(uint32_t video_out_handle, uint32_t display_buffer_index);
void TriggerEvent(uint32_t event_type, uint32_t event_index, uint64_t event_address = 0);
@@ -1331,26 +1331,13 @@ KYTY_CP_OP_PARSER(CpOpDispatchIndirect) {
EXIT_NOT_IMPLEMENTED(cmd_id != 0xc0011600 && cmd_id != 0xc0021600);
if (cmd_id == 0xc0021600) {
struct DispatchIndirectArgs {
uint32_t thread_group_x;
uint32_t thread_group_y;
uint32_t thread_group_z;
};
auto* args = reinterpret_cast<const DispatchIndirectArgs*>(
buffer[0] | (static_cast<uint64_t>(buffer[1]) << 32u));
uint32_t mode = buffer[2];
EXIT_NOT_IMPLEMENTED(args == nullptr);
cp.DispatchDirect(args->thread_group_x, args->thread_group_y, args->thread_group_z, mode);
cp.DispatchIndirect(buffer[0] | (static_cast<uint64_t>(buffer[1]) << 32u), buffer[2]);
return 3;
}
uint32_t data_offset = buffer[0];
uint32_t mode = buffer[1];
cp.DispatchIndirect(data_offset, mode);
const auto base_addr = cp.GetDispatchIndirectArgsBaseAddress();
EXIT_NOT_IMPLEMENTED(base_addr == 0);
cp.DispatchIndirect(base_addr + buffer[0], buffer[1]);
return 2;
}
+10 -13
View File
@@ -1090,19 +1090,16 @@ void CommandProcessor::DispatchDirect(uint32_t thread_group_x, uint32_t thread_g
}*/
}
void CommandProcessor::DispatchIndirect(uint32_t data_offset, uint32_t mode) {
struct DispatchIndirectArgs {
uint32_t thread_group_x;
uint32_t thread_group_y;
uint32_t thread_group_z;
};
EXIT_NOT_IMPLEMENTED(m_dispatch_indirect_args_base_addr == 0);
const auto args_addr = m_dispatch_indirect_args_base_addr + data_offset;
auto* args = reinterpret_cast<const DispatchIndirectArgs*>(args_addr);
DispatchDirect(args->thread_group_x, args->thread_group_y, args->thread_group_z, mode);
void CommandProcessor::DispatchIndirect(uint64_t args_addr, uint32_t mode) {
EXIT_NOT_IMPLEMENTED(args_addr == 0 || (args_addr & 3u) != 0);
if ((mode & Pm4::COMPUTE_DISPATCH_INITIATOR_USE_THREAD_DIMENSIONS) != 0) {
const auto* args = reinterpret_cast<const vk::DispatchIndirectCommand*>(args_addr);
DispatchDirect(args->x, args->y, args->z, mode);
return;
}
m_sh_ctx.SetCsWaveSize(Pm4::ComputeWaveSize(mode));
CheckBuffer();
m_renderer.GetRenderExecutor().DispatchIndirect(m_submit_id, CurrentBuffer(), args_addr, mode);
}
void CommandProcessor::DrawIndexAuto(DrawAutoArgs args) {
+1
View File
@@ -910,6 +910,7 @@ constexpr uint32_t COMPUTE_PGM_RSRC1_WGP_MODE_MASK = 0x1;
constexpr uint32_t COMPUTE_PGM_RSRC1_FWD_PROGRESS_SHIFT = 31;
constexpr uint32_t COMPUTE_PGM_RSRC1_FWD_PROGRESS_MASK = 0x1;
constexpr uint32_t COMPUTE_DISPATCH_INITIATOR_USE_THREAD_DIMENSIONS = 1u << 5u;
constexpr uint32_t COMPUTE_DISPATCH_INITIATOR_CS_W32_EN_SHIFT = 15;
constexpr uint32_t COMPUTE_DISPATCH_INITIATOR_CS_W32_EN_MASK = 0x1;