tex_unit update

driver/tests/tex_demo/kernel.c

@@ -52,8 +52,8 @@ int main() {
 	targ.karg        = *arg;
 	targ.tile_width  = arg->dst_width;
 	targ.tile_height = (arg->dst_height + arg->num_tasks - 1) / arg->num_tasks;    
-	targ.deltaX      = 1.0f / arg->dst_width;
+	targ.deltaX      = 1.0f / (((float)arg->src_width) / arg->dst_width);
-	targ.deltaY      = 1.0f / arg->dst_height;
+	targ.deltaY      = 1.0f / (((float)arg->src_height) / arg->dst_height);
 	
 	vx_spawn_tasks(arg->num_tasks, kernel_body, &targ);
 }

driver/tests/tex_demo/main.cpp

@@ -124,7 +124,7 @@ int main(int argc, char *argv[]) {
   RT_CHECK(vx_dev_caps(device, VX_CAPS_MAX_WARPS, &max_warps));
   RT_CHECK(vx_dev_caps(device, VX_CAPS_MAX_THREADS, &max_threads));
 
-  uint32_t num_tasks = max_cores * max_warps * max_threads / 4;
+  uint32_t num_tasks = max_cores * max_warps * max_threads;
 
   std::cout << "number of tasks: " << std::dec << num_tasks << std::endl;
   std::cout << "source buffer: width=" << src_width << ", heigth=" << src_height << ", size=" << src_bufsize << " bytes" << std::endl;
@@ -170,8 +170,8 @@ int main(int argc, char *argv[]) {
     RT_CHECK(vx_copy_to_dev(buffer, KERNEL_ARG_DEV_MEM_ADDR, sizeof(kernel_arg_t), 0));
   }
 
-  // upload source buffer0
+  // upload source buffer
-  std::cout << "upload source buffer0" << std::endl;      
+  std::cout << "upload source buffer" << std::endl;      
   {    
     auto buf_ptr = (int8_t*)vx_host_ptr(buffer);
     for (uint32_t i = 0; i < src_bufsize; ++i) {

hw/dpi/float_dpi.cpp

@@ -26,9 +26,9 @@ extern "C" {
   void dpi_utof(int a, int frm, int* result, int* fflags);
 
   void dpi_fclss(int a, int* result);
-  void dpi_fsgnj(int a, int* result);
+  void dpi_fsgnj(int a, int b, int* result);
-  void dpi_fsgnjn(int a, int* result);
+  void dpi_fsgnjn(int a, int b, int* result);
-  void dpi_fsgnjx(int a, int* result);
+  void dpi_fsgnjx(int a, int b, int* result);
 
   void dpi_flt(int a, int b, int* result, int* fflags);
   void dpi_fle(int a, int b, int* result, int* fflags);
@@ -244,21 +244,53 @@ void dpi_fmax(int a, int b, int* result, int* fflags) {
 }
 
 void dpi_fclss(int a, int* result) {
-  // TODO
+
-  *result = 0;
+  int r = 0; // clear all bits
+
+  bool fsign = (a >> 31);
+  uint32_t expo = (a >> 23) & 0xFF;
+  uint32_t fraction = a & 0x7FFFFF;
+
+  if ((expo == 0) && (fraction == 0)) {
+    r = fsign ? (1 << 3) : (1 << 4); // +/- 0
+  } else if ((expo == 0) && (fraction != 0)) {
+    r = fsign ? (1 << 2) : (1 << 5); // +/- subnormal
+  } else if ((expo == 0xFF) && (fraction == 0)) {
+    r = fsign ? (1<<0) : (1<<7); // +/- infinity
+  } else if ((expo == 0xFF ) && (fraction != 0)) { 
+    if (!fsign && (fraction == 0x00400000)) {
+      r = (1 << 9);               // quiet NaN
+    } else { 
+      r = (1 << 8);               // signaling NaN
+    }
+  } else {
+    r = fsign ? (1 << 1) : (1 << 6); // +/- normal
+  }
+
+  *result = r;
 }
 
-void dpi_fsgnj(int a, int* result) {
+void dpi_fsgnj(int a, int b, int* result) {
-  // TODO
+  
-  *result = 0;
+  int sign = b & 0x80000000;
+  int r = sign | (a & 0x7FFFFFFF);
+
+  *result = r;
 }
 
-void dpi_fsgnjn(int a, int* result) {
+void dpi_fsgnjn(int a, int b, int* result) {
-  // TODO
+  
-  *result = 0;
+  int sign = ~b & 0x80000000;
+  int r = sign | (a & 0x7FFFFFFF);
+
+  *result = r;
 }
 
-void dpi_fsgnjx(int a, int* result) {
+void dpi_fsgnjx(int a, int b, int* result) {
-  // TODO
+  
-  *result = 0;
+  int sign1 = a & 0x80000000;
+  int sign2 = b & 0x80000000;
+  int r = (sign1 ^ sign2) | (a & 0x7FFFFFFF);
+
+  *result = r;
 }

hw/dpi/float_dpi.vh

@@ -18,9 +18,9 @@ import "DPI-C" context function void dpi_itof(input int a, input bit[2:0] frm, o
 import "DPI-C" context function void dpi_utof(input int a, input bit[2:0] frm, output int result, output bit[4:0] fflags);
 
 import "DPI-C" context function void dpi_fclss(input int a, output int result);
-import "DPI-C" context function void dpi_fsgnj(input int a, output int result);
+import "DPI-C" context function void dpi_fsgnj(input int a, input int b, output int result);
-import "DPI-C" context function void dpi_fsgnjn(input int a, output int result);
+import "DPI-C" context function void dpi_fsgnjn(input int a, input int b, output int result);
-import "DPI-C" context function void dpi_fsgnjx(input int a, output int result);
+import "DPI-C" context function void dpi_fsgnjx(input int a, input int b, output int result);
 
 import "DPI-C" context function void dpi_flt(input int a, input int b, output int result, output bit[4:0] fflags);
 import "DPI-C" context function void dpi_fle(input int a, input int b, output int result, output bit[4:0] fflags);

hw/rtl/VX_lsu_unit.v

@@ -78,7 +78,7 @@ module VX_lsu_unit #(
     wire [`NUM_THREADS-1:0] rsp_rem_mask_n;
 
     reg [`NUM_THREADS-1:0] req_sent_mask;
-    wire req_sent_all;
+    wire sent_all_ready;
 
     wire [`LSUQ_ADDR_BITS-1:0] mbuf_waddr, mbuf_raddr;
     wire mbuf_full;
@@ -116,13 +116,13 @@ module VX_lsu_unit #(
         .full         (mbuf_full)
     );
 
-    assign req_sent_all = (&(dcache_req_if.ready | req_sent_mask | ~req_tmask))
+    assign sent_all_ready = (&(dcache_req_if.ready | req_sent_mask | ~req_tmask))
-                       || (req_is_dup & dcache_req_if.ready[0]);
+                      || (req_is_dup & dcache_req_if.ready[0]);
 
     always @(posedge clk) begin
-        if (reset || req_sent_all) begin
+        if (reset || sent_all_ready) begin
             req_sent_mask <= 0;
-        end else if (!req_sent_all) begin
+        end else begin
             req_sent_mask <= req_sent_mask | dcache_req_fire;            
         end
     end      
@@ -193,11 +193,11 @@ module VX_lsu_unit #(
     assign dcache_req_if.tag = {`NUM_THREADS{req_tag}};
 `endif
     
-    assign ready_in = req_ready_dep && req_sent_all;
+    assign ready_in = req_ready_dep && sent_all_ready;
 
     // send store commit
 
-    wire is_store_rsp = req_valid && ~req_wb && req_sent_all;
+    wire is_store_rsp = req_valid && ~req_wb && sent_all_ready;
 
     assign st_commit_if.valid = is_store_rsp;
     assign st_commit_if.wid   = req_wid;

hw/rtl/fp_cores/VX_fpu_dpi.v

@@ -330,9 +330,9 @@ module VX_fpu_dpi #(
                 dpi_feq    (dataa[i], datab[i], result_feq[i], fflags_feq[i]);
                 dpi_fmin   (dataa[i], datab[i], result_fmin[i], fflags_fmin[i]);
                 dpi_fmax   (dataa[i], datab[i], result_fmax[i], fflags_fmax[i]);            
-                dpi_fsgnj  (dataa[i], result_fsgnj[i]);
+                dpi_fsgnj  (dataa[i], datab[i], result_fsgnj[i]);
-                dpi_fsgnjn (dataa[i], result_fsgnjn[i]);
+                dpi_fsgnjn (dataa[i], datab[i], result_fsgnjn[i]);
-                dpi_fsgnjx (dataa[i], result_fsgnjx[i]);
+                dpi_fsgnjx (dataa[i], datab[i], result_fsgnjx[i]);
                 result_fmv[i] = dataa[i];
             end
         end

hw/rtl/tex_unit/VX_tex_memory.v

@@ -99,7 +99,7 @@ module VX_tex_memory #(
     ///////////////////////////////////////////////////////////////////////////
 
     wire req_texel_valid;
-    wire req_texel_sent, last_texel_sent;
+    wire sent_all_ready, last_texel_sent;
     wire req_texel_dup;
     wire [`NUM_THREADS-1:0][29:0] req_texel_addr;
     reg [1:0] req_texel_idx;
@@ -108,7 +108,7 @@ module VX_tex_memory #(
     always @(posedge clk) begin
         if (reset || last_texel_sent) begin
             req_texel_idx <= 0;
-        end else if (req_texel_sent) begin
+        end else if (req_texel_valid && sent_all_ready) begin
             req_texel_idx <= req_texel_idx + 1;
         end
     end
@@ -126,7 +126,7 @@ module VX_tex_memory #(
     assign req_texel_dup   = q_dup_reqs[req_texel_idx];
 
     wire is_last_texel = (req_texel_idx == (q_req_filter ? 3 : 0));
-    assign last_texel_sent = req_texel_sent && is_last_texel;
+    assign last_texel_sent = req_texel_valid && sent_all_ready && is_last_texel;
 
     // DCache Request
 
@@ -136,11 +136,11 @@ module VX_tex_memory #(
 
     assign dcache_req_fire = dcache_req_if.valid & dcache_req_if.ready;    
 
-    assign req_texel_sent = (&(dcache_req_if.ready | texel_sent_mask | ~q_req_tmask))
+    assign sent_all_ready = (&(dcache_req_if.ready | texel_sent_mask | ~q_req_tmask))
-                     || (req_texel_dup & dcache_req_if.ready[0]);
+                         || (req_texel_dup & dcache_req_if.ready[0]);
 
     always @(posedge clk) begin
-        if (reset || req_texel_sent) begin
+        if (reset || sent_all_ready) begin
             texel_sent_mask <= 0;
         end else begin
             texel_sent_mask <= texel_sent_mask | dcache_req_fire;            

simX/execute.cpp

@@ -561,20 +561,18 @@ void Warp::execute(const Instr &instr, Pipeline *pipeline) {
 
         // FSGNJ.S, FSGNJN.S, FSGNJX.S
         case 0x10: {
-          bool     fsign1 = rsdata[0] & 0x80000000;
+          bool     fsign1 = (rsdata[0] >> 31);
           uint32_t fdata1 = rsdata[0] & 0x7FFFFFFF;
-          bool     fsign2 = rsdata[1] & 0x80000000;
+          bool     fsign2 = (rsdata[1] >> 31);
           switch (func3) {            
           case 0: // FSGNJ.S
             rddata = (fsign2 << 31) | fdata1;
             break;          
           case 1: // FSGNJN.S
-            fsign2 = !fsign2;
+            rddata = (!fsign2 << 31) | fdata1;
-            rddata = (fsign2 << 31) | fdata1;
             break;          
           case 2: { // FSGNJX.S
-            bool sign = fsign1 ^ fsign2;
+            rddata = ((fsign1 ^ fsign2) << 31) | fdata1;
-            rddata = (sign << 31) | fdata1;
             } break;
           }
         } break;