Merge pull request #30 from sspenst/dev

shfl instruction implemented

1 files changed, 77 insertions, 0 deletions

diff --git a/src/cuda-sim/instructions.cc b/src/cuda-sim/instructions.cc
index b7a4a99..011c285 100644
--- a/src/cuda-sim/instructions.cc
+++ b/src/cuda-sim/instructions.cc
@@ -50,8 +50,10 @@ unsigned ptx_instruction::g_num_ptx_inst_uid=0;
 
 const char *g_opcode_string[NUM_OPCODES] = {
 #define OP_DEF(OP,FUNC,STR,DST,CLASSIFICATION) STR,
+#define OP_W_DEF(OP,FUNC,STR,DST,CLASSIFICATION) STR,
 #include "opcodes.def"
 #undef OP_DEF
+#undef OP_W_DEF
 };
 
 void inst_not_implemented( const ptx_instruction * pI ) ;
@@ -3537,6 +3539,81 @@ void set_impl( const ptx_instruction *pI, ptx_thread_info *thread )
 
 }
 
+void shfl_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst )
+{
+	unsigned i_type = pI->get_type();
+	int tid = inst.warp_id() * core->get_warp_size();
+	ptx_thread_info *thread = core->get_thread_info()[tid];
+	ptx_warp_info *warp_info = thread->m_warp_info;
+	int lane = warp_info->get_done_threads();
+	thread = core->get_thread_info()[tid + lane];
+
+	const operand_info &dst = pI->dst();
+	const operand_info &src1 = pI->src1();
+	const operand_info &src2 = pI->src2();
+	const operand_info &src3 = pI->src3();
+	int bval = (thread->get_operand_value(src2, dst, i_type, thread, 1)).u32;
+	int cval = (thread->get_operand_value(src3, dst, i_type, thread, 1)).u32;
+	int mask = cval >> 8;
+	bval &= 0x1F;
+	cval &= 0x1F;
+
+	int maxLane = (lane & mask) | (cval & ~mask);
+	int minLane = lane & mask;
+
+	int src_idx;
+	unsigned p;
+	switch(pI->shfl_op()) {
+	case UP_OPTION:
+		src_idx = lane - bval;
+		p = (src_idx >= maxLane);
+		break;
+	case DOWN_OPTION:
+		src_idx = lane + bval;
+		p = (src_idx <= maxLane);
+		break;
+	case BFLY_OPTION:
+		src_idx = lane ^ bval;
+		p = (src_idx <= maxLane);
+		break;
+	case IDX_OPTION:
+		src_idx = minLane | (bval & ~mask);
+		p = (src_idx <= maxLane);
+		break;
+	default:
+		printf("GPGPU-Sim PTX: ERROR: Invalid shfl option\n");
+		assert(0);
+		break;
+	}
+	// copy from own lane
+	if (!p) src_idx = lane;
+
+	// copy input from lane src_idx
+	ptx_reg_t data;
+	if (inst.active(src_idx)) {
+		ptx_thread_info *source = core->get_thread_info()[tid + src_idx];
+		data = source->get_operand_value(src1, dst, i_type, source, 1);
+	} else {
+		printf("GPGPU-Sim PTX: WARNING: shfl input value unpredictable for inactive threads in a warp\n");
+		data.u32 = 0;
+	}
+	thread->set_operand_value(dst, data, i_type, thread, pI);
+
+	/*
+	TODO: deal with predicates appropriately using the following pseudocode:
+	if (!isGuardPredicateTrue(src_idx)) {
+		printf("GPGPU-Sim PTX: WARNING: shfl input value unpredictable for predicated-off threads in a warp\n");
+	}
+	if (dest predicate selected) data.pred = p;
+	*/
+
+	// keep track of the number of threads that have executed in the warp
+	warp_info->inc_done_threads();
+	if (warp_info->get_done_threads() == inst.active_count()) {
+		warp_info->reset_done_threads();
+	}
+}
+
 void shl_impl( const ptx_instruction *pI, ptx_thread_info *thread ) 
 {
    ptx_reg_t a, b, d;