Implements context write tracking.

src/nv2a_vsh_emulator.c

@@ -37,8 +37,7 @@ static inline void set_register(float *out, const float *in,
   out[3] = mult * in[swizzle[3]];
 }
 
-static inline void fetch_value(float *out,
-                               const Nv2aVshExecutionState *state,
+static inline void fetch_value(float *out, const Nv2aVshExecutionState *state,
                                const Nv2aVshInput *input) {
   const float *in;
 
@@ -125,12 +124,11 @@ static inline void apply_operation(Nv2aVshExecutionState *state,
   }
 }
 
-static inline void apply(Nv2aVshExecutionState *state,
-                         const Nv2aVshStep *step) {
+static inline void prepare_inputs(float *mac_inputs, float *ilu_input,
+                                  Nv2aVshExecutionState *state,
+                                  const Nv2aVshStep *step) {
   // Copy the inputs for both operations first to prevent introducing order
   // dependent behavior.
-  float mac_inputs[3 * 4];
-  float ilu_input[4];
   if (step->mac.opcode) {
     for (uint32_t i = 0; i < 3; ++i) {
       if (step->mac.inputs[i].type == NV2ART_NONE) {
@@ -142,6 +140,13 @@ static inline void apply(Nv2aVshExecutionState *state,
   if (step->ilu.opcode) {
     fetch_value(ilu_input, state, &step->ilu.inputs[0]);
   }
+}
+
+static inline void apply(Nv2aVshExecutionState *state,
+                         const Nv2aVshStep *step) {
+  float mac_inputs[3 * 4];
+  float ilu_input[4];
+  prepare_inputs(mac_inputs, &ilu_input, state, step);
 
   if (step->mac.opcode) {
     apply_operation(state, &step->mac, mac_inputs);
@@ -151,6 +156,31 @@ static inline void apply(Nv2aVshExecutionState *state,
   }
 }
 
+static inline void apply_track_context_writes(Nv2aVshExecutionState *state,
+                                              const Nv2aVshStep *step,
+                                              bool *context_dirty) {
+  float mac_inputs[3 * 4];
+  float ilu_input[4];
+  prepare_inputs(mac_inputs, &ilu_input, state, step);
+
+  if (step->mac.opcode) {
+    apply_operation(state, &step->mac, mac_inputs);
+    if (step->mac.outputs[0].type == NV2ART_CONTEXT) {
+      context_dirty[step->mac.outputs[0].index] = true;
+    } else if (step->mac.outputs[1].type == NV2ART_CONTEXT) {
+      context_dirty[step->mac.outputs[1].index] = true;
+    }
+  }
+  if (step->ilu.opcode) {
+    apply_operation(state, &step->ilu, &ilu_input);
+    if (step->ilu.outputs[0].type == NV2ART_CONTEXT) {
+      context_dirty[step->ilu.outputs[0].index] = true;
+    } else if (step->ilu.outputs[1].type == NV2ART_CONTEXT) {
+      context_dirty[step->ilu.outputs[1].index] = true;
+    }
+  }
+}
+
 void nv2a_vsh_emu_execute(Nv2aVshExecutionState *state,
                           const Nv2aVshProgram *program) {
   assert(state);
@@ -166,6 +196,23 @@ void nv2a_vsh_emu_execute(Nv2aVshExecutionState *state,
   }
 }
 
+void nv2a_vsh_emu_execute_track_context_writes(Nv2aVshExecutionState *state,
+                                               const Nv2aVshProgram *program,
+                                               bool *context_dirty) {
+  assert(state);
+  assert(program && program->steps);
+  assert(context_dirty);
+
+  Nv2aVshStep *step = program->steps;
+  while (true) {
+    apply_track_context_writes(state, step, context_dirty);
+    if (step->is_final) {
+      break;
+    }
+    ++step;
+  }
+}
+
 void nv2a_vsh_emu_apply(Nv2aVshExecutionState *state, const Nv2aVshStep *step) {
   apply(state, step);
 }

src/nv2a_vsh_emulator.h

@@ -11,6 +11,12 @@ extern "C" {
 void nv2a_vsh_emu_execute(Nv2aVshExecutionState *state,
                           const Nv2aVshProgram *program);
 
+// context_dirty is an array of 192 bools that will be set when writing to
+// entries in context_regs.
+void nv2a_vsh_emu_execute_track_context_writes(Nv2aVshExecutionState *state,
+                                               const Nv2aVshProgram *program,
+                                               bool *context_dirty);
+
 // Emulates the given step by applying it to the given state.
 void nv2a_vsh_emu_apply(Nv2aVshExecutionState *state, const Nv2aVshStep *step);
 

src/nv2a_vsh_emulator_execution_state.c

@@ -8,19 +8,17 @@ Nv2aVshExecutionState nv2a_vsh_emu_initialize_full_execution_state(
   Nv2aVshExecutionState ret = {
       (float *)state->input_regs, (float *)state->output_regs,
       (float *)state->temp_regs, (float *)state->context_regs,
-      (float *)&state->address_reg, NULL};
+      (float *)&state->address_reg};
   return ret;
 }
 
 Nv2aVshExecutionState nv2a_vsh_emu_initialize_xss_execution_state(
-    Nv2aVshCPUXVSSExecutionState *state, float *context_regs,
-    bool *context_dirty) {
+    Nv2aVshCPUXVSSExecutionState *state, float *context_regs) {
   memset(state, 0, sizeof(*state));
   state->context_regs = context_regs;
-  state->context_dirty = context_dirty;
   Nv2aVshExecutionState ret = {
       (float *)state->input_regs, (float *)state->output_regs,
       (float *)state->temp_regs, (float *)state->context_regs,
-      (float *)&state->address_reg, state->context_dirty};
+      (float *)&state->address_reg};
   return ret;
 }

src/nv2a_vsh_emulator_execution_state.h

@@ -69,10 +69,6 @@ typedef struct Nv2aVshCPUXVSSExecutionState_ {
   // Context regs should be initialized to a flat array of 192 registers.
   float *context_regs;
 
-  // Optional array of 192 bools that will be set when writing to entries in
-  // context_regs.
-  bool *context_dirty;
-
   float address_reg[4];
 } Nv2aVshCPUXVSSExecutionState;
 
@@ -84,8 +80,7 @@ Nv2aVshExecutionState nv2a_vsh_emu_initialize_full_execution_state(
 // Initializes the given Nv2aVshCPUXVSSExecutionState and returns an
 // Nv2aVshExecutionState appropriate for use with nv2a_vsh_cpu_* functions.
 Nv2aVshExecutionState nv2a_vsh_emu_initialize_xss_execution_state(
-    Nv2aVshCPUXVSSExecutionState *state, float *context_regs,
-    bool *context_dirty);
+    Nv2aVshCPUXVSSExecutionState *state, float *context_regs);
 
 #ifdef __cplusplus
 }  // extern "C"

test/emulator/test_basic.cpp

@@ -2,9 +2,9 @@
 
 #include "nv2a_vsh_emulator.h"
 
-#define CHECK_REGISTER(bank, index, actual) \
-  do {                                      \
-    float *expected = bank + index * 4;     \
+#define CHECK_REGISTER(bank, index, actual)   \
+  do {                                        \
+    float *expected = bank + index * 4;       \
     BOOST_TEST((expected)[0] == (actual)[0]); \
     BOOST_TEST((expected)[1] == (actual)[1]); \
     BOOST_TEST((expected)[2] == (actual)[2]); \
@@ -84,4 +84,41 @@ BOOST_AUTO_TEST_CASE(step_trivial) {
   CHECK_REGISTER(state.output_regs, NV2AOR_TEX2, &full_state.input_regs[reg]);
 }
 
+BOOST_AUTO_TEST_CASE(program_context_tracked) {
+  Nv2aVshCPUFullExecutionState full_state;
+  Nv2aVshExecutionState state =
+      nv2a_vsh_emu_initialize_full_execution_state(&full_state);
+  uint32_t reg = 11 * 4;
+  full_state.input_regs[reg + 0] = 123.0f;
+  full_state.input_regs[reg + 1] = -456.0f;
+  full_state.input_regs[reg + 2] = 0.789f;
+  full_state.input_regs[reg + 3] = 32.64f;
+
+  // MOV c1.xyzw, v11
+  Nv2aVshStep steps[1];
+  Nv2aVshStep *step = &steps[0];
+  clear_step(step);
+  step->mac.opcode = NV2AOP_MOV;
+  step->mac.outputs[0].type = NV2ART_CONTEXT;
+  step->mac.outputs[0].index = 1;
+  step->mac.outputs[0].writemask = NV2AWM_XYZW;
+  step->mac.inputs[0].type = NV2ART_INPUT;
+  step->mac.inputs[0].index = 11;
+  step->is_final = true;
+
+  Nv2aVshProgram program;
+  program.steps = steps;
+
+  bool context_dirty[192] = {false};
+  nv2a_vsh_emu_execute_track_context_writes(&state, &program, context_dirty);
+
+  CHECK_REGISTER(state.context_regs, 1, &full_state.input_regs[reg]);
+  BOOST_TEST(!context_dirty[0]);
+  BOOST_TEST(context_dirty[1]);
+  for (uint32_t i = 2; i < 192; ++i) {
+    BOOST_TEST_INFO(i);
+    BOOST_TEST(!context_dirty[i]);
+  }
+}
+
 BOOST_AUTO_TEST_SUITE_END()