yess

Examples/Software/main.c

@@ -47,7 +47,7 @@ int main(void)
 		PulseCommandList cmd = PulseRequestCommandList(device, PULSE_COMMAND_LIST_GENERAL);
 
 		PulseComputePass pass = PulseBeginComputePass(cmd);
-			//PulseBindStorageBuffers(pass, &buffer, 1);
+			PulseBindStorageBuffers(pass, &buffer, 1);
 			PulseBindComputePipeline(pass, pipeline);
 			PulseDispatchComputations(pass, 16, 1, 1);
 		PulseEndComputePass(pass);

Examples/Software/shader.nzsl

@@ -3,23 +3,23 @@ module;
 
 struct Input
 {
-    [builtin(global_invocation_indices)] indices: vec3[u32]
+	[builtin(global_invocation_indices)] indices: vec3[u32]
 }
 
 [layout(std430)]
 struct SSBO
 {
-    data: dyn_array[i32]
+	data: dyn_array[i32]
 }
 
 external
 {
-    //[set(1), binding(0)] ssbo: storage[SSBO],
+	[set(1), binding(0)] ssbo: storage[SSBO],
 }
 
 [entry(compute)]
 [workgroup(16, 16, 1)]
 fn main(input: Input)
 {
-    //ssbo.data[input.indices.x * input.indices.y] = i32(input.indices.x * input.indices.y);
+	ssbo.data[input.indices.x * input.indices.y] = i32(input.indices.x * input.indices.y);
 }

Sources/Backends/Software/SoftCommandList.c

@@ -17,8 +17,6 @@
 #include "SoftComputePipeline.h"
 #include "SoftBuffer.h"
 
-#include <stdio.h>
-
 static void SoftCommandCopyBufferToBuffer(SoftCommand* cmd)
 {
 	const PulseBufferRegion* src = cmd->CopyBufferToBuffer.src;
@@ -32,16 +30,26 @@ static void SoftCommandCopyBufferToBuffer(SoftCommand* cmd)
 
 static int SoftCommandDispatchCore(void* arg)
 {
-	SoftComputePipeline* soft_pipeline = (SoftComputePipeline*)arg;
+	SoftCommand* cmd = (SoftCommand*)arg;
-	spvm_state_t state = spvm_state_create(soft_pipeline->program);
+	SoftComputePipeline* soft_pipeline = SOFT_RETRIEVE_DRIVER_DATA_AS(cmd->Dispatch.pipeline, SoftComputePipeline*);
+
+	mtx_lock(&cmd->Dispatch.dispatch_mutex);
+		spvm_state_t state = spvm_state_create(soft_pipeline->program);
+	mtx_unlock(&cmd->Dispatch.dispatch_mutex);
+
 	spvm_ext_opcode_func* glsl_ext_data = spvm_build_glsl450_ext();
 	spvm_result_t glsl_std_450 = spvm_state_get_result(state, "GLSL.std.450");
 	if(glsl_std_450)
 		glsl_std_450->extension = glsl_ext_data;
 	spvm_word main = spvm_state_get_result_location(state, (spvm_string)soft_pipeline->entry_point);
+
+	spvm_word mem_count = 0;
+	spvm_member_t local_invocation_id = spvm_state_get_builtin(state, SpvBuiltInLocalInvocationId, &mem_count);
+
 	spvm_state_prepare(state, main);
 	spvm_state_call_function(state);
 	spvm_state_delete(state);
+	atomic_fetch_sub(&cmd->cmd_list->commands_running, 1);
 	return 0;
 }
 
@@ -62,15 +70,13 @@ static void SoftCommandDispatch(SoftCommand* cmd)
 			{
 				for(uint32_t i = 0; i < local_size; i++)
 				{
-					thrd_create(&invocations[invocation_index], SoftCommandDispatchCore, soft_pipeline);
+					atomic_fetch_add(&cmd->cmd_list->commands_running, 1);
-					//thrd_join(invocations[invocation_index], NULL);
+					thrd_create(&invocations[invocation_index], SoftCommandDispatchCore, cmd);
 					invocation_index++;
 				}
 			}
 		}
 	}
-	for(uint32_t i = 0; i < invocations_count; i++)
-		thrd_join(invocations[i], PULSE_NULLPTR);
 	free(invocations);
 }
 
@@ -97,11 +103,8 @@ static int SoftCommandsRunner(void* arg)
 		}
 	}
 
-	if(soft_cmd->fence != PULSE_NULL_HANDLE)
+	atomic_fetch_sub(&soft_cmd->commands_running, 1); // Remove fence safety
-	{
+
-		SoftFence* fence = SOFT_RETRIEVE_DRIVER_DATA_AS(soft_cmd->fence, SoftFence*);
-		atomic_store(&fence->signal, true);
-	}
 	cmd->state = PULSE_COMMAND_LIST_STATE_READY;
 	return 0;
 }
@@ -126,6 +129,7 @@ PulseCommandList SoftRequestCommandList(PulseDevice device, PulseCommandListUsag
 	cmd->pass = SoftCreateComputePass(device, cmd);
 	cmd->state = PULSE_COMMAND_LIST_STATE_RECORDING;
 	cmd->is_available = false;
+	atomic_store(&soft_cmd->commands_running, 0);
 
 	return cmd;
 }
@@ -133,6 +137,7 @@ PulseCommandList SoftRequestCommandList(PulseDevice device, PulseCommandListUsag
 void SoftQueueCommand(PulseCommandList cmd, SoftCommand command)
 {
 	SoftCommandList* soft_cmd = SOFT_RETRIEVE_DRIVER_DATA_AS(cmd, SoftCommandList*);
+	command.cmd_list = soft_cmd;
 	PULSE_EXPAND_ARRAY_IF_NEEDED(soft_cmd->commands, SoftCommand, soft_cmd->commands_count, soft_cmd->commands_capacity, 8);
 	soft_cmd->commands[soft_cmd->commands_count] = command;
 	soft_cmd->commands_count++;
@@ -150,6 +155,7 @@ bool SoftSubmitCommandList(PulseDevice device, PulseCommandList cmd, PulseFence
 		fence->cmd = cmd;
 		atomic_store(&soft_fence->signal, false);
 	}
+	atomic_fetch_add(&soft_cmd->commands_running, 1); // Fence safety to avoid fence being signaled before first command being sumitted
 	return thrd_create(&soft_cmd->thread, SoftCommandsRunner, cmd) == thrd_success;
 }
 
@@ -157,6 +163,20 @@ void SoftReleaseCommandList(PulseDevice device, PulseCommandList cmd)
 {
 	SoftCommandList* soft_cmd = SOFT_RETRIEVE_DRIVER_DATA_AS(cmd, SoftCommandList*);
 	SoftDestroyComputePass(device, cmd->pass);
+
+	for(uint32_t i = 0; i < soft_cmd->commands_count; i++)
+	{
+		SoftCommand* command = &soft_cmd->commands[i];
+		switch(command->type)
+		{
+			// Lock/Unlock to make sure the mutex is not in use
+			case SOFT_COMMAND_DISPATCH: mtx_lock(&command->Dispatch.dispatch_mutex); mtx_unlock(&command->Dispatch.dispatch_mutex); break;
+			case SOFT_COMMAND_DISPATCH_INDIRECT: mtx_lock(&command->DispatchIndirect.dispatch_mutex); mtx_unlock(&command->DispatchIndirect.dispatch_mutex); break;
+
+			default: break;
+		}
+	}
+	free(soft_cmd->commands);
 	free(soft_cmd);
 	free(cmd);
 }

Sources/Backends/Software/SoftCommandList.h

@@ -9,6 +9,7 @@
 #ifndef PULSE_SOFTWARE_COMMAND_LIST_H_
 #define PULSE_SOFTWARE_COMMAND_LIST_H_
 
+#include <stdatomic.h>
 #include <tinycthread.h>
 
 #include "Soft.h"
@@ -17,6 +18,7 @@
 
 typedef struct SoftCommand
 {
+	struct SoftCommandList* cmd_list;
 	SoftCommandType type;
 	union
 	{
@@ -50,6 +52,7 @@ typedef struct SoftCommand
 			uint32_t groupcount_x;
 			uint32_t groupcount_y;
 			uint32_t groupcount_z;
+			mtx_t dispatch_mutex;
 		} Dispatch;
 
 		struct
@@ -57,8 +60,20 @@ typedef struct SoftCommand
 			PulseComputePipeline pipeline;
 			PulseBuffer buffer;
 			uint32_t offset;
+			mtx_t dispatch_mutex;
 		} DispatchIndirect;
 	};
+	union
+	{
+		struct
+		{
+			uint32_t global_invocation_id[3];
+			uint32_t local_invocation_id[3];
+			uint32_t workgroup_count[3];
+			uint32_t workgroup_index[3];
+			uint32_t local_invocation_index;
+		} Dipsatch;
+	} Private;
 } SoftCommand;
 
 typedef struct SoftCommandList
@@ -68,6 +83,7 @@ typedef struct SoftCommandList
 	SoftCommand* commands;
 	uint32_t commands_count;
 	uint32_t commands_capacity;
+	atomic_ullong commands_running;
 } SoftCommandList;
 
 PulseCommandList SoftRequestCommandList(PulseDevice device, PulseCommandListUsage usage);

Sources/Backends/Software/SoftComputePass.c

@@ -14,11 +14,7 @@ PulseComputePass SoftCreateComputePass(PulseDevice device, PulseCommandList cmd)
 	PulseComputePass pass = (PulseComputePass)calloc(1, sizeof(PulseComputePassHandler));
 	PULSE_CHECK_ALLOCATION_RETVAL(pass, PULSE_NULL_HANDLE);
 
-	SoftComputePass* soft_pass = (SoftComputePass*)calloc(1, sizeof(SoftComputePass));
-	PULSE_CHECK_ALLOCATION_RETVAL(soft_pass, PULSE_NULL_HANDLE);
-
 	pass->cmd = cmd;
-	pass->driver_data = soft_pass;
 
 	return pass;
 }
@@ -42,6 +38,30 @@ void SoftEndComputePass(PulseComputePass pass)
 
 void SoftBindStorageBuffers(PulseComputePass pass, const PulseBuffer* buffers, uint32_t num_buffers)
 {
+	PulseBufferUsageFlags usage = buffers[0]->usage;
+	bool is_readwrite = (usage & PULSE_BUFFER_USAGE_STORAGE_WRITE) != 0;
+	PulseBuffer* array = is_readwrite ? pass->readwrite_storage_buffers : pass->readonly_storage_buffers;
+
+	for(uint32_t i = 0; i < num_buffers; i++)
+	{
+		if(is_readwrite && (buffers[i]->usage & PULSE_BUFFER_USAGE_STORAGE_WRITE) == 0)
+		{
+			if(PULSE_IS_BACKEND_LOW_LEVEL_DEBUG(pass->cmd->device->backend))
+				PulseLogError(pass->cmd->device->backend, "cannot bind a read only buffer with read-write buffers");
+			PulseSetInternalError(PULSE_ERROR_INVALID_BUFFER_USAGE);
+			return;
+		}
+		else if(!is_readwrite && (buffers[i]->usage & PULSE_BUFFER_USAGE_STORAGE_WRITE) != 0)
+		{
+			if(PULSE_IS_BACKEND_LOW_LEVEL_DEBUG(pass->cmd->device->backend))
+				PulseLogError(pass->cmd->device->backend, "cannot bind a read-write buffer with read only buffers");
+			PulseSetInternalError(PULSE_ERROR_INVALID_BUFFER_USAGE);
+			return;
+		}
+		if(array[i] == buffers[i])
+			continue;
+		array[i] = buffers[i];
+	}
 }
 
 void SoftBindUniformData(PulseComputePass pass, uint32_t slot, const void* data, uint32_t data_size)
@@ -66,5 +86,6 @@ void SoftDispatchComputations(PulseComputePass pass, uint32_t groupcount_x, uint
 	command.Dispatch.groupcount_y = groupcount_y;
 	command.Dispatch.groupcount_z = groupcount_z;
 	command.Dispatch.pipeline = pass->current_pipeline;
+	mtx_init(&command.Dispatch.dispatch_mutex, mtx_plain);
 	SoftQueueCommand(pass->cmd, command);
 }

Sources/Backends/Software/SoftComputePass.h

@@ -11,11 +11,6 @@
 
 #include "Soft.h"
 
-typedef struct SoftComputePass
-{
-	int dummy;
-} SoftComputePass;
-
 PulseComputePass SoftCreateComputePass(PulseDevice device, PulseCommandList cmd);
 void SoftDestroyComputePass(PulseDevice device, PulseComputePass pass);
 

Sources/Backends/Software/SoftComputePipeline.c

@@ -10,6 +10,105 @@
 #include "SoftDevice.h"
 #include "SoftComputePipeline.h"
 
+void SoftAllocateWorkgroupMemory(struct spvm_state* state, spvm_word result_id, spvm_word type_id)
+{
+	SoftComputePipeline* pipeline = (SoftComputePipeline*)state->owner->user_data;
+	mtx_lock(&pipeline->workgroup_memory_allocations_mutex);
+		spvm_result_t result = &state->results[result_id];
+		PULSE_EXPAND_ARRAY_IF_NEEDED(pipeline->workgroup_memory_allocations, SoftSharedMemoryEntry, pipeline->workgroup_memory_allocations_size, pipeline->workgroup_memory_allocations_capacity, 16);
+		SoftSharedMemoryEntry* entry = &pipeline->workgroup_memory_allocations[pipeline->workgroup_memory_allocations_size];
+		entry->destination = &state->results[result_id];
+		entry->slot = result_id;
+		memcpy(&entry->data, entry->destination, sizeof(spvm_result));
+		spvm_result_allocate_typed_value(result, state->results, type_id);
+		pipeline->workgroup_memory_allocations_size++;
+	mtx_unlock(&pipeline->workgroup_memory_allocations_mutex);
+}
+/*
+void SoftWriteWorkgroupMemory(struct spvm_state* state, spvm_word result_id, spvm_word val_id)
+{
+	spvm_result_t ptr = &state->results[result_id];
+	if(ptr->source_location != PULSE_NULLPTR)
+	{
+		spvm_word word_count = ptr->source_word_count;
+		spvm_source code = ptr->source_location;
+
+		spvm_word var_type = SPVM_READ_WORD(code);
+		spvm_word id = SPVM_READ_WORD(code);
+		spvm_word memory_id = SPVM_READ_WORD(code);
+
+		spvm_result_t sharedData = PULSE_NULLPTR;
+
+		ed::DebugInformation* dbgr = (ed::DebugInformation*)state->owner->user_data;
+
+		for(int i = 0; i < dbgr->SharedMemory.size(); i++)
+		{
+			if(dbgr->SharedMemory[i].Slot == memory_id)
+			{
+				sharedData = &dbgr->SharedMemory[i].Data;
+				break;
+			}
+		}
+
+		if(sharedData == PULSE_NULLPTR)
+			return;
+
+		spvm_word index_count = word_count - 4;
+
+		spvm_word index_id = SPVM_READ_WORD(code);
+		spvm_word index = state->results[index_id].members[0].value.s;
+
+		spvm_member_t result = sharedData->members + MIN(index, sharedData->member_count - 1);
+
+		while(index_count)
+		{
+			index_id = SPVM_READ_WORD(code);
+			index = state->results[index_id].members[0].value.s;
+
+			result = result->members + MIN(index, result->member_count - 1);
+
+			index_count--;
+		}
+
+		spvm_member* members = PULSE_NULLPTR;
+		spvm_word member_count = 0;
+		if(result->member_count != 0)
+		{
+			member_count = result->member_count;
+			members = result->members;
+		}
+		else
+		{
+			member_count = 1;
+			members = result;
+		}
+		spvm_member_memcpy(members, state->results[val_id].members, member_count);
+	}
+}
+
+void SoftControlBarrier(struct spvm_state* state, spvm_word exec, spvm_word mem, spvm_word sem)
+{
+	ed::DebugInformation* dbgr = (ed::DebugInformation*)state->owner->user_data;
+
+	// copy memory
+	for(int i = 0; i < dbgr->SharedMemory.size(); i++)
+	{
+		const ed::DebugInformation::SharedMemoryEntry& entry = dbgr->SharedMemory[i];
+		spvm_member_memcpy(entry.Data.members, entry.Destination->members, entry.Data.member_count);
+	}
+
+	// synchronize threads
+	dbgr->SyncWorkgroup();
+
+	// copy memory
+	for(int i = 0; i < dbgr->SharedMemory.size(); i++)
+	{
+		const ed::DebugInformation::SharedMemoryEntry& entry = dbgr->SharedMemory[i];
+		spvm_member_memcpy(entry.Destination->members, entry.Data.members, entry.Destination->member_count);
+	}
+}
+*/
+
 PulseComputePipeline SoftCreateComputePipeline(PulseDevice device, const PulseComputePipelineCreateInfo* info)
 {
 	SoftDevice* soft_device = SOFT_RETRIEVE_DRIVER_DATA_AS(device, SoftDevice*);
@@ -32,12 +131,18 @@ PulseComputePipeline SoftCreateComputePipeline(PulseDevice device, const PulseCo
 
 	soft_pipeline->program = spvm_program_create(soft_device->spv_context, (spvm_source)info->code, info->code_size / sizeof(spvm_word));
 	soft_pipeline->entry_point = calloc(1, strlen(info->entrypoint));
+	PULSE_CHECK_ALLOCATION_RETVAL(soft_pipeline->entry_point, PULSE_NULL_HANDLE);
 	strcpy((char*)soft_pipeline->entry_point, info->entrypoint);
 
+	soft_pipeline->program->user_data = soft_pipeline;
+	soft_pipeline->program->allocate_workgroup_memory = SoftAllocateWorkgroupMemory;
+
 	// Create dummy state to retrieve informations from the spirv
 	spvm_state_t state = spvm_state_create(soft_pipeline->program);
 	spvm_state_delete(state);
 
+	mtx_init(&soft_pipeline->workgroup_memory_allocations_mutex, mtx_plain);
+
 	pipeline->driver_data = soft_pipeline;
 
 	if(PULSE_IS_BACKEND_HIGH_LEVEL_DEBUG(device->backend))
@@ -56,7 +161,9 @@ void SoftDestroyComputePipeline(PulseDevice device, PulseComputePipeline pipelin
 	PULSE_UNUSED(device);
 	SoftComputePipeline* soft_pipeline = SOFT_RETRIEVE_DRIVER_DATA_AS(pipeline, SoftComputePipeline*);
 	spvm_program_delete(soft_pipeline->program);
-	free(soft_pipeline->entry_point);
+	mtx_destroy(&soft_pipeline->workgroup_memory_allocations_mutex);
+	free(soft_pipeline->workgroup_memory_allocations);
+	free((void*)soft_pipeline->entry_point);
 	free(soft_pipeline);
 	if(PULSE_IS_BACKEND_HIGH_LEVEL_DEBUG(device->backend))
 		PulseLogInfoFmt(device->backend, "(Soft) destroyed compute pipeline %p", pipeline);

Sources/Backends/Software/SoftComputePipeline.h

@@ -9,14 +9,27 @@
 #ifndef PULSE_SOFTWARE_COMPUTE_PIPELINE_H_
 #define PULSE_SOFTWARE_COMPUTE_PIPELINE_H_
 
+#include <tinycthread.h>
+
 #include "Soft.h"
 #include <spvm/state.h>
 #include <spvm/program.h>
 
+typedef struct SoftSharedMemoryEntry
+{
+	spvm_result data;
+	spvm_result_t destination;
+	spvm_word slot;
+} SoftSharedMemoryEntry;
+
 typedef struct SoftComputePipeline
 {
 	spvm_program_t program;
 	const char* entry_point;
+	SoftSharedMemoryEntry* workgroup_memory_allocations;
+	uint32_t workgroup_memory_allocations_size;
+	uint32_t workgroup_memory_allocations_capacity;
+	mtx_t workgroup_memory_allocations_mutex;
 } SoftComputePipeline;
 
 PulseComputePipeline SoftCreateComputePipeline(PulseDevice device, const PulseComputePipelineCreateInfo* info);

Sources/Backends/Software/SoftFence.c

@@ -6,6 +6,7 @@
 #include "../../PulseInternal.h"
 #include "Soft.h"
 #include "SoftFence.h"
+#include "SoftCommandList.h"
 
 PulseFence SoftCreateFence(PulseDevice device)
 {
@@ -35,6 +36,9 @@ bool SoftIsFenceReady(PulseDevice device, PulseFence fence)
 {
 	PULSE_UNUSED(device);
 	SoftFence* soft_fence = SOFT_RETRIEVE_DRIVER_DATA_AS(fence, SoftFence*);
+	SoftCommandList* soft_cmd = SOFT_RETRIEVE_DRIVER_DATA_AS(fence->cmd, SoftCommandList*);
+	if(atomic_load(&soft_cmd->commands_running) == 0)
+		atomic_store(&soft_fence->signal, true);
 	return atomic_load(&soft_fence->signal);
 }
 

Sources/PulseDevice.c

@@ -25,7 +25,6 @@ PULSE_API void PulseDestroyDevice(PulseDevice device)
 	free(device->allocated_buffers);
 	free(device->allocated_images);
 	device->PFN_DestroyDevice(device);
-	device->driver_data = PULSE_NULLPTR;
 }
 
 PULSE_API PulseBackendBits PulseGetBackendInUseByDevice(PulseDevice device)

Tests/Image.c

@@ -39,7 +39,6 @@ void TestImageCreation()
 	 * This test may crash some Nouveau NVK drivers (wtf ???).
 	 * It seems to be comming exclusively from 3D read-only images
 	 */
-	if(false)
 	{
 		PulseImageCreateInfo image_create_info = { 0 };
 		image_create_info.type = PULSE_IMAGE_TYPE_3D;

xmake.lua

@@ -33,7 +33,7 @@ local backends = {
 	Software = {
 		option = "software",
 		default = true,
-		packages = { "spirv-vm", "cpuinfo" }
+		packages = { "spirv-vm", "cpuinfo", "spirv-reflect" }
 	},
 	OpenGL = {
 		option = "opengl",