feat: add pipelined fp8 scalar tmem softmax unit

hw/dpi/float_dpi.cpp

@@ -12,6 +12,7 @@
 // limitations under the License.
 
 #include <stdio.h>
+#include <stdint.h>
 #include <math.h>
 #include <unordered_map>
 #include <vector>
@@ -37,6 +38,11 @@ extern "C" {
 
   void dpi_fdiv(bool enable, int dst_fmt, int64_t a, int64_t b, const svBitVecVal* frm, int64_t* result, svBitVecVal* fflags);
   void dpi_fsqrt(bool enable, int dst_fmt, int64_t a, const svBitVecVal* frm, int64_t* result, svBitVecVal* fflags);
+  void dpi_fexp(bool enable, int dst_fmt, int64_t a, int64_t* result, svBitVecVal* fflags);
+  int dpi_f32_max(bool enable, int a_bits, int b_bits);
+  int dpi_softmax_exp_acc(bool enable, int score_bits, int max_bits, int accum_bits);
+  int dpi_softmax_prob_to_f16x2(bool enable, int score_bits, int max_bits, int denom_bits);
+  int dpi_softmax_prob_to_fp8e4m3x4(bool enable, int score_bits, int max_bits, int denom_bits);
 
   void dpi_ftoi(bool enable, int dst_fmt, int src_fmt, int64_t a, const svBitVecVal* frm, int64_t* result, svBitVecVal* fflags);
   void dpi_ftou(bool enable, int dst_fmt, int src_fmt, int64_t a, const svBitVecVal* frm, int64_t* result, svBitVecVal* fflags);
@@ -83,6 +89,89 @@ inline int64_t check_boxing(int64_t a) {
   return a;
 }
 
+inline float bits_to_float(uint32_t value) {
+  union {
+    uint32_t u;
+    float f;
+  } bits = {value};
+  return bits.f;
+}
+
+inline uint32_t float_to_bits(float value) {
+  union {
+    float f;
+    uint32_t u;
+  } bits = {value};
+  return bits.u;
+}
+
+int dpi_f32_max(bool enable, int a_bits, int b_bits) {
+  if (!enable)
+    return 0;
+  float a = bits_to_float(static_cast<uint32_t>(a_bits));
+  float b = bits_to_float(static_cast<uint32_t>(b_bits));
+  return static_cast<int>(float_to_bits(a > b ? a : b));
+}
+
+int dpi_softmax_exp_acc(bool enable, int score_bits, int max_bits, int accum_bits) {
+  if (!enable)
+    return 0;
+  float score = bits_to_float(static_cast<uint32_t>(score_bits));
+  float row_max = bits_to_float(static_cast<uint32_t>(max_bits));
+  float accum = bits_to_float(static_cast<uint32_t>(accum_bits));
+  return static_cast<int>(float_to_bits(accum + expf(score - row_max)));
+}
+
+int dpi_softmax_prob_to_f16x2(bool enable, int score_bits, int max_bits, int denom_bits) {
+  if (!enable)
+    return 0;
+  float score = bits_to_float(static_cast<uint32_t>(score_bits));
+  float row_max = bits_to_float(static_cast<uint32_t>(max_bits));
+  float denom = bits_to_float(static_cast<uint32_t>(denom_bits));
+  float prob = denom == 0.0f ? 0.0f : expf(score - row_max) / denom;
+  half_float::half half_prob(prob);
+  uint16_t half_bits;
+  static_assert(sizeof(half_bits) == sizeof(half_prob), "unexpected half size");
+  __builtin_memcpy(&half_bits, &half_prob, sizeof(half_bits));
+  return static_cast<int>(half_bits | (uint32_t(half_bits) << 16));
+}
+
+static float fp8e4m3_positive_to_float(uint8_t bits) {
+  const int exp = (bits >> 3) & 0xf;
+  const int frac = bits & 0x7;
+  if (exp == 0) {
+    return frac == 0 ? 0.0f : ldexpf(static_cast<float>(frac) / 8.0f, -6);
+  }
+  return ldexpf(1.0f + static_cast<float>(frac) / 8.0f, exp - 7);
+}
+
+static uint8_t float_to_fp8e4m3_positive(float value) {
+  if (!(value > 0.0f))
+    return 0;
+  uint8_t best = 0;
+  float best_diff = fabsf(value);
+  for (uint32_t bits = 1; bits < 0x80; ++bits) {
+    const float candidate = fp8e4m3_positive_to_float(static_cast<uint8_t>(bits));
+    const float diff = fabsf(candidate - value);
+    if (diff < best_diff || (diff == best_diff && ((bits & 1u) == 0u))) {
+      best = static_cast<uint8_t>(bits);
+      best_diff = diff;
+    }
+  }
+  return best;
+}
+
+int dpi_softmax_prob_to_fp8e4m3x4(bool enable, int score_bits, int max_bits, int denom_bits) {
+  if (!enable)
+    return 0;
+  float score = bits_to_float(static_cast<uint32_t>(score_bits));
+  float row_max = bits_to_float(static_cast<uint32_t>(max_bits));
+  float denom = bits_to_float(static_cast<uint32_t>(denom_bits));
+  float prob = denom == 0.0f ? 0.0f : expf(score - row_max) / denom;
+  uint32_t fp8 = float_to_fp8e4m3_positive(prob);
+  return static_cast<int>(fp8 | (fp8 << 8) | (fp8 << 16) | (fp8 << 24));
+}
+
 void dpi_fadd(bool enable, int dst_fmt, int64_t a, int64_t b, const svBitVecVal* frm, int64_t* result, svBitVecVal* fflags) {
   if (!enable) 
     return;
@@ -173,6 +262,38 @@ void dpi_fsqrt(bool enable, int dst_fmt, int64_t a, const svBitVecVal* frm, int6
   }
 }
 
+void dpi_fexp(bool enable, int dst_fmt, int64_t a, int64_t* result, svBitVecVal* fflags) {
+  if (!enable)
+    return;
+  *fflags = 0;
+  if (dst_fmt) {
+    double input;
+    static_assert(sizeof(input) == sizeof(a), "unexpected double size");
+    __builtin_memcpy(&input, &a, sizeof(input));
+    double output = exp(input);
+    __builtin_memcpy(result, &output, sizeof(output));
+    if (isinf(output) && isfinite(input)) {
+      *fflags |= 0x05; // OF | NX
+    } else if (output == 0.0 && input < 0.0 && isfinite(input)) {
+      *fflags |= 0x03; // UF | NX
+    } else if (isfinite(input)) {
+      *fflags |= 0x01; // NX
+    }
+  } else {
+    uint32_t boxed = static_cast<uint32_t>(check_boxing(a));
+    float input = bits_to_float(boxed);
+    float output = expf(input);
+    *result = nan_box(float_to_bits(output));
+    if (isinf(output) && isfinite(input)) {
+      *fflags |= 0x05; // OF | NX
+    } else if (output == 0.0f && input < 0.0f && isfinite(input)) {
+      *fflags |= 0x03; // UF | NX
+    } else if (isfinite(input)) {
+      *fflags |= 0x01; // NX
+    }
+  }
+}
+
 void dpi_ftoi(bool enable, int dst_fmt, int src_fmt, int64_t a, const svBitVecVal* frm, int64_t* result, svBitVecVal* fflags) {
   if (!enable) 
     return;

hw/dpi/float_dpi.vh

@@ -26,6 +26,11 @@ import "DPI-C" function void dpi_fnmsub(input logic enable, input int dst_fmt, i
 
 import "DPI-C" function void dpi_fdiv(input logic enable, input int dst_fmt, input longint a, input longint b, input bit[2:0] frm, output longint result, output bit[4:0] fflags);
 import "DPI-C" function void dpi_fsqrt(input logic enable, input int dst_fmt, input longint a, input bit[2:0] frm, output longint result, output bit[4:0] fflags);
+import "DPI-C" function void dpi_fexp(input logic enable, input int dst_fmt, input longint a, output longint result, output bit[4:0] fflags);
+import "DPI-C" function int dpi_f32_max(input logic enable, input int a_bits, input int b_bits);
+import "DPI-C" function int dpi_softmax_exp_acc(input logic enable, input int score_bits, input int max_bits, input int accum_bits);
+import "DPI-C" function int dpi_softmax_prob_to_f16x2(input logic enable, input int score_bits, input int max_bits, input int denom_bits);
+import "DPI-C" function int dpi_softmax_prob_to_fp8e4m3x4(input logic enable, input int score_bits, input int max_bits, input int denom_bits);
 
 import "DPI-C" function void dpi_ftoi(input logic enable, input int dst_fmt, input int src_fmt, input longint a, input bit[2:0] frm, output longint result, output bit[4:0] fflags);
 import "DPI-C" function void dpi_ftou(input logic enable, input int dst_fmt, input int src_fmt, input longint a, input bit[2:0] frm, output longint result, output bit[4:0] fflags);

hw/rtl/VX_config.vh

@@ -422,6 +422,11 @@
 `endif
 `endif
 
+// FEXP Latency
+`ifndef LATENCY_FEXP
+`define LATENCY_FEXP 8
+`endif
+
 // FCVT Latency
 `ifndef LATENCY_FCVT
 `define LATENCY_FCVT 5

hw/rtl/VX_core_wrapper.sv

@@ -104,6 +104,15 @@ module Vortex import VX_gpu_pkg::*; #(
     output [NUM_TENSOR_CORES*9-1:0]                       tc_tmem_C_waddr,
     output [NUM_TENSOR_CORES*`NUM_THREADS*`XLEN-1:0]      tc_tmem_C_wdata,
     output [NUM_TENSOR_CORES*`NUM_THREADS*`XLEN/8-1:0]    tc_tmem_C_mask,
+    output                                                sc_tmem_ren,
+    input                                                 sc_tmem_rready,
+    output [8:0]                                          sc_tmem_raddr,
+    input  [`NUM_THREADS*`XLEN-1:0]                       sc_tmem_rdata,
+    output                                                sc_tmem_wen,
+    input                                                 sc_tmem_wready,
+    output [8:0]                                          sc_tmem_waddr,
+    output [`NUM_THREADS*`XLEN-1:0]                       sc_tmem_wdata,
+    output [`NUM_THREADS*`XLEN/8-1:0]                     sc_tmem_mask,
 
     // gbar ------------------------------------------------
 
@@ -514,6 +523,15 @@ module Vortex import VX_gpu_pkg::*; #(
         .tensor_tmem_C_waddr(tc_tmem_C_waddr),
         .tensor_tmem_C_wdata(tc_tmem_C_wdata),
         .tensor_tmem_C_mask(tc_tmem_C_mask),
+        .scalar_tmem_ren(sc_tmem_ren),
+        .scalar_tmem_rready(sc_tmem_rready),
+        .scalar_tmem_raddr(sc_tmem_raddr),
+        .scalar_tmem_rdata(sc_tmem_rdata),
+        .scalar_tmem_wen(sc_tmem_wen),
+        .scalar_tmem_wready(sc_tmem_wready),
+        .scalar_tmem_waddr(sc_tmem_waddr),
+        .scalar_tmem_wdata(sc_tmem_wdata),
+        .scalar_tmem_mask(sc_tmem_mask),
         .tensor_smem_B_if (tc_p2_bus_if),
     `else
         .tensor_tmem_A_ren(tc_tmem_A_ren),
@@ -529,6 +547,15 @@ module Vortex import VX_gpu_pkg::*; #(
         .tensor_tmem_C_waddr(tc_tmem_C_waddr),
         .tensor_tmem_C_wdata(tc_tmem_C_wdata),
         .tensor_tmem_C_mask(tc_tmem_C_mask),
+        .scalar_tmem_ren(sc_tmem_ren),
+        .scalar_tmem_rready(sc_tmem_rready),
+        .scalar_tmem_raddr(sc_tmem_raddr),
+        .scalar_tmem_rdata(sc_tmem_rdata),
+        .scalar_tmem_wen(sc_tmem_wen),
+        .scalar_tmem_wready(sc_tmem_wready),
+        .scalar_tmem_waddr(sc_tmem_waddr),
+        .scalar_tmem_wdata(sc_tmem_wdata),
+        .scalar_tmem_mask(sc_tmem_mask),
         .tensor_smem_B_if (tc_p2_bus_if),
     `endif
 

hw/rtl/VX_define.vh

@@ -208,6 +208,9 @@
 `define INST_LSU_SH          4'b1001
 `define INST_LSU_SW          4'b1010
 `define INST_LSU_SD          4'b1011 // new for RV64I SD
+`define INST_LSU_TMEM_LD     4'b1100
+`define INST_LSU_TMEM_ST     4'b1101
+`define INST_LSU_TMEM_SOFTMAX 4'b1110
 `define INST_LSU_FENCE       4'b1111
 `define INST_LSU_BITS        4
 `define INST_LSU_FMT(op)     op[2:0]
@@ -238,6 +241,7 @@
 `define INST_FPU_IS_W(mod)   (mod[4])
 `define INST_FPU_IS_CLASS(op, mod) (op == `INST_FPU_MISC && mod == 3)
 `define INST_FPU_IS_MVXW(op, mod) (op == `INST_FPU_MISC && mod == 4)
+`define INST_FPU_IS_EXP(op, mod) (op == `INST_FPU_MISC && mod == 8)
 
 `define INST_SFU_TMC         4'h0
 `define INST_SFU_WSPAWN      4'h1 

hw/rtl/core/VX_core.sv

@@ -54,6 +54,15 @@ module VX_core import VX_gpu_pkg::*; #(
     output logic [NUM_TENSOR_CORES*9-1:0]                       tensor_tmem_C_waddr,
     output logic [NUM_TENSOR_CORES*`NUM_THREADS*`XLEN-1:0]      tensor_tmem_C_wdata,
     output logic [NUM_TENSOR_CORES*`NUM_THREADS*`XLEN/8-1:0]    tensor_tmem_C_mask,
+    output logic                                                scalar_tmem_ren,
+    input  logic                                                scalar_tmem_rready,
+    output logic [8:0]                                          scalar_tmem_raddr,
+    input  logic [`NUM_THREADS*`XLEN-1:0]                       scalar_tmem_rdata,
+    output logic                                                scalar_tmem_wen,
+    input  logic                                                scalar_tmem_wready,
+    output logic [8:0]                                          scalar_tmem_waddr,
+    output logic [`NUM_THREADS*`XLEN-1:0]                       scalar_tmem_wdata,
+    output logic [`NUM_THREADS*`XLEN/8-1:0]                     scalar_tmem_mask,
     VX_tc_bus_if.master     tensor_smem_B_if[NUM_TENSOR_CORES],
 
 `ifdef GBAR_ENABLE
@@ -410,6 +419,15 @@ module VX_core import VX_gpu_pkg::*; #(
         .tensor_tmem_C_waddr(tensor_tmem_C_waddr),
         .tensor_tmem_C_wdata(tensor_tmem_C_wdata),
         .tensor_tmem_C_mask(tensor_tmem_C_mask),
+        .scalar_tmem_ren(scalar_tmem_ren),
+        .scalar_tmem_rready(scalar_tmem_rready),
+        .scalar_tmem_raddr(scalar_tmem_raddr),
+        .scalar_tmem_rdata(scalar_tmem_rdata),
+        .scalar_tmem_wen(scalar_tmem_wen),
+        .scalar_tmem_wready(scalar_tmem_wready),
+        .scalar_tmem_waddr(scalar_tmem_waddr),
+        .scalar_tmem_wdata(scalar_tmem_wdata),
+        .scalar_tmem_mask(scalar_tmem_mask),
         .tensor_smem_B_if  (tensor_smem_B_if),
     `endif
     `endif

hw/rtl/core/VX_decode.sv

@@ -508,9 +508,24 @@ module VX_decode  #(
             end
             `INST_EXT2: begin                
                 case (func3)
+                    3'h0: begin
+                        if (func7 == 7'h30) begin
+                            ex_type = `EX_LSU;
+                            op_type = `INST_LSU_TMEM_LD;
+                            use_rd = 1;
+                            `USED_IREG (rd);
+                            `USED_IREG (rs1);
+                        end
+                    end
                     3'h1: begin
-                        case (func2)                       
-                            2'h0: begin // CMOV
+                        if (func7 == 7'h30) begin
+                            ex_type = `EX_LSU;
+                            op_type = `INST_LSU_TMEM_ST;
+                            `USED_IREG (rs1);
+                            `USED_IREG (rs2);
+                        end else begin
+                            case (func2)
+                                2'h0: begin // CMOV
                                 ex_type = `EX_SFU;
                                 op_type = `INST_OP_BITS'(`INST_SFU_CMOV);
                                 use_rd = 1;
@@ -518,9 +533,32 @@ module VX_decode  #(
                                 `USED_IREG (rs1);
                                 `USED_IREG (rs2);
                                 `USED_IREG (rs3);
-                            end
-                            default:;
-                        endcase
+                                end
+                                default:;
+                            endcase
+                        end
+                    end
+                `ifdef EXT_F_ENABLE
+                    3'h2: begin
+                        if (func7 == 7'h30 && rs2 == 5'd0) begin
+                            ex_type = `EX_FPU;
+                            op_type = `INST_OP_BITS'(`INST_FPU_MISC);
+                            op_mod  = `INST_MOD_BITS'(8); // FEXP.S
+                            use_rd  = 1;
+                            `USED_FREG (rd);
+                            `USED_FREG (rs1);
+                        end
+                    end
+                `endif
+                    3'h3: begin
+                        if (func7 == 7'h30) begin
+                            ex_type = `EX_LSU;
+                            op_type = `INST_LSU_TMEM_SOFTMAX;
+                            use_rd = 1;
+                            `USED_IREG (rd);
+                            `USED_IREG (rs1);
+                            `USED_IREG (rs2);
+                        end
                     end
                     default:;
                 endcase

hw/rtl/core/VX_execute.sv

@@ -12,6 +12,9 @@
 // limitations under the License.
 
 `include "VX_define.vh"
+`ifdef SV_DPI
+`include "float_dpi.vh"
+`endif
 
 module VX_execute import VX_gpu_pkg::*; #(
     parameter CORE_ID = 0,
@@ -92,6 +95,16 @@ module VX_execute import VX_gpu_pkg::*; #(
 `endif
 `endif
 
+    output logic                                                scalar_tmem_ren,
+    input  logic                                                scalar_tmem_rready,
+    output logic [8:0]                                          scalar_tmem_raddr,
+    input  logic [`NUM_THREADS*`XLEN-1:0]                       scalar_tmem_rdata,
+    output logic                                                scalar_tmem_wen,
+    input  logic                                                scalar_tmem_wready,
+    output logic [8:0]                                          scalar_tmem_waddr,
+    output logic [`NUM_THREADS*`XLEN-1:0]                       scalar_tmem_wdata,
+    output logic [`NUM_THREADS*`XLEN/8-1:0]                     scalar_tmem_mask,
+
     // simulation helper signals
     output wire             sim_ebreak,
 
@@ -286,8 +299,39 @@ module VX_execute import VX_gpu_pkg::*; #(
 
     `SCOPE_IO_SWITCH (1)
 
+    VX_dispatch_if scalar_mem_lsu_dispatch_if[`ISSUE_WIDTH]();
+    VX_commit_if scalar_mem_lsu_commit_if[`ISSUE_WIDTH]();
+    VX_commit_if scalar_tmem_commit_if[`ISSUE_WIDTH]();
     VX_commit_if lsu_scalar_commit_if[`ISSUE_WIDTH]();
 
+    wire [`ISSUE_WIDTH-1:0] scalar_tmem_dispatch;
+    wire [`ISSUE_WIDTH-1:0] scalar_tmem_ld_dispatch;
+    wire [`ISSUE_WIDTH-1:0] scalar_tmem_st_dispatch;
+    wire [`ISSUE_WIDTH-1:0] scalar_tmem_softmax_dispatch;
+    wire [`ISSUE_WIDTH-1:0] scalar_tmem_dispatch_ready;
+    wire [`ISSUE_WIDTH-1:0] scalar_tmem_dispatch_fire;
+    wire [`ISSUE_WIDTH-1:0] scalar_tmem_commit_ready;
+    `UNUSED_VAR (scalar_tmem_dispatch_fire)
+
+    for (genvar i = 0; i < `ISSUE_WIDTH; ++i) begin : g_scalar_lsu_split
+        assign scalar_tmem_ld_dispatch[i] = lsu_dispatch_if[i].valid
+                                         && (lsu_dispatch_if[i].data.op_type == `INST_LSU_TMEM_LD);
+        assign scalar_tmem_st_dispatch[i] = lsu_dispatch_if[i].valid
+                                         && (lsu_dispatch_if[i].data.op_type == `INST_LSU_TMEM_ST);
+        assign scalar_tmem_softmax_dispatch[i] = lsu_dispatch_if[i].valid
+                                             && (lsu_dispatch_if[i].data.op_type == `INST_LSU_TMEM_SOFTMAX);
+        assign scalar_tmem_dispatch[i] = scalar_tmem_ld_dispatch[i]
+                                      || scalar_tmem_st_dispatch[i]
+                                      || scalar_tmem_softmax_dispatch[i];
+
+        assign scalar_mem_lsu_dispatch_if[i].valid = lsu_dispatch_if[i].valid && !scalar_tmem_dispatch[i];
+        assign scalar_mem_lsu_dispatch_if[i].data = lsu_dispatch_if[i].data;
+        assign lsu_dispatch_if[i].ready = scalar_tmem_dispatch[i] ? scalar_tmem_dispatch_ready[i]
+                                                                  : scalar_mem_lsu_dispatch_if[i].ready;
+
+        assign scalar_tmem_dispatch_fire[i] = scalar_tmem_dispatch[i] && scalar_tmem_dispatch_ready[i];
+    end
+
     VX_mem_bus_if #(
         .DATA_SIZE (DCACHE_WORD_SIZE),
         .TAG_WIDTH (DCACHE_TAG_WIDTH)
@@ -301,10 +345,300 @@ module VX_execute import VX_gpu_pkg::*; #(
         .reset          (lsu_reset),
         .downstream_mem_busy (downstream_mem_busy),
         .cache_bus_if   (scalar_lsu_bus_if),
-        .dispatch_if    (lsu_dispatch_if),
-        .commit_if      (lsu_scalar_commit_if)
+        .dispatch_if    (scalar_mem_lsu_dispatch_if),
+        .commit_if      (scalar_mem_lsu_commit_if)
     );
 
+    wire scalar_tmem_pending;
+    reg  [`ISSUE_WIDTH-1:0] scalar_tmem_grant;
+    reg  [`ISSUE_WIDTH-1:0] scalar_tmem_grant_r;
+    reg                     scalar_tmem_load_pending;
+    reg                     scalar_tmem_store_pending;
+    reg                     scalar_tmem_commit_pending;
+    reg                     scalar_tmem_softmax_active;
+    reg [5:0]               scalar_tmem_softmax_read_index;
+    reg                     scalar_tmem_softmax_reads_issued;
+    reg                     scalar_tmem_softmax_read_response_valid;
+    reg [5:0]               scalar_tmem_softmax_read_response_index;
+    reg [8:0]               scalar_tmem_softmax_score_base_r;
+    reg [8:0]               scalar_tmem_softmax_p_base_r;
+    reg  [`UUID_WIDTH-1:0]  scalar_tmem_uuid_r;
+    reg  [`NW_WIDTH-1:0]    scalar_tmem_wid_r;
+    reg  [`NUM_THREADS-1:0] scalar_tmem_tmask_r;
+    reg  [`XLEN-1:0]        scalar_tmem_pc_r;
+    reg  [`NR_BITS-1:0]     scalar_tmem_rd_r;
+    reg  [`NUM_THREADS*`XLEN-1:0] scalar_tmem_rdata_r;
+    reg                     scalar_tmem_rdata_valid_r;
+    wire [`ISSUE_WIDTH-1:0][8:0] scalar_tmem_req_addr;
+    wire [`ISSUE_WIDTH-1:0][`NUM_THREADS*`XLEN-1:0] scalar_tmem_req_wdata;
+    wire [`ISSUE_WIDTH-1:0][`UUID_WIDTH-1:0] scalar_tmem_req_uuid;
+    wire [`ISSUE_WIDTH-1:0][`NW_WIDTH-1:0] scalar_tmem_req_wid;
+    wire [`ISSUE_WIDTH-1:0][`NUM_THREADS-1:0] scalar_tmem_req_tmask;
+    wire [`ISSUE_WIDTH-1:0][`XLEN-1:0] scalar_tmem_req_pc;
+    wire [`ISSUE_WIDTH-1:0][`NR_BITS-1:0] scalar_tmem_req_rd;
+    wire [`ISSUE_WIDTH-1:0][8:0] scalar_tmem_req_softmax_p_addr;
+
+    localparam [5:0] SCALAR_TMEM_SOFTMAX_ROW_LAST  = 6'd31;
+    localparam [5:0] SCALAR_TMEM_SOFTMAX_TILE_LAST = 6'd63;
+
+    assign scalar_tmem_pending = |scalar_tmem_dispatch;
+
+    always @(*) begin
+        scalar_tmem_grant = '0;
+        for (integer i = `ISSUE_WIDTH-1; i >= 0; --i) begin
+            if (scalar_tmem_dispatch[i]) begin
+                scalar_tmem_grant = '0;
+                scalar_tmem_grant[i] = 1'b1;
+            end
+        end
+    end
+
+    wire scalar_tmem_grant_valid = scalar_tmem_pending && !scalar_tmem_load_pending
+                                && !scalar_tmem_store_pending && !scalar_tmem_commit_pending
+                                && !scalar_tmem_softmax_active;
+    wire scalar_tmem_grant_is_load = |(scalar_tmem_grant & scalar_tmem_ld_dispatch);
+    wire scalar_tmem_grant_is_store = |(scalar_tmem_grant & scalar_tmem_st_dispatch);
+    wire scalar_tmem_grant_is_softmax = |(scalar_tmem_grant & scalar_tmem_softmax_dispatch);
+    wire scalar_tmem_req_ready = scalar_tmem_grant_is_softmax ? 1'b1 :
+                                 scalar_tmem_grant_is_load ? scalar_tmem_rready : scalar_tmem_wready;
+    wire scalar_tmem_req_fire = scalar_tmem_grant_valid && scalar_tmem_req_ready;
+    wire scalar_tmem_softmax_read_valid = scalar_tmem_softmax_active
+                                       && !scalar_tmem_softmax_reads_issued;
+    wire scalar_tmem_softmax_read_issue = scalar_tmem_softmax_read_valid && scalar_tmem_rready;
+    wire [8:0] scalar_tmem_softmax_read_addr = scalar_tmem_softmax_score_base_r
+                                             + 9'(scalar_tmem_softmax_read_index[4:0]);
+
+    wire scalar_tmem_softmax_vector_busy;
+    wire scalar_tmem_softmax_vector_out_valid;
+    wire [5:0] scalar_tmem_softmax_vector_out_index;
+    wire [`NUM_THREADS*`XLEN-1:0] scalar_tmem_softmax_vector_out_data;
+    wire scalar_tmem_softmax_vector_done;
+    wire scalar_tmem_softmax_vector_out_ready = scalar_tmem_softmax_active && scalar_tmem_wready;
+    wire scalar_tmem_softmax_write_valid = scalar_tmem_softmax_active
+                                          && scalar_tmem_softmax_vector_out_valid;
+    wire scalar_tmem_softmax_write_issue = scalar_tmem_softmax_write_valid && scalar_tmem_wready;
+    wire [8:0] scalar_tmem_softmax_write_addr = scalar_tmem_softmax_p_base_r
+                                              + 9'(scalar_tmem_softmax_vector_out_index);
+    wire [`NUM_THREADS*`XLEN-1:0] scalar_tmem_softmax_wdata =
+        scalar_tmem_softmax_vector_out_data;
+    wire scalar_tmem_softmax_commit = scalar_tmem_commit_pending
+                                    && !scalar_tmem_load_pending
+                                    && !scalar_tmem_store_pending;
+    wire [`XLEN-1:0] scalar_tmem_softmax_token = `XLEN'(scalar_tmem_softmax_p_base_r);
+
+    VX_tmem_softmax_unit #(
+        .NUM_LANES (`NUM_THREADS),
+        .ROW_SIZE  (32)
+    ) tmem_softmax_unit (
+        .clk        (clk),
+        .reset      (reset),
+        .start      (scalar_tmem_req_fire && scalar_tmem_grant_is_softmax),
+        .load_valid (scalar_tmem_softmax_read_response_valid),
+        .load_data  (scalar_tmem_rdata),
+        .load_last  (scalar_tmem_softmax_read_response_index == SCALAR_TMEM_SOFTMAX_ROW_LAST),
+        .busy       (scalar_tmem_softmax_vector_busy),
+        .out_valid  (scalar_tmem_softmax_vector_out_valid),
+        .out_ready  (scalar_tmem_softmax_vector_out_ready),
+        .out_index  (scalar_tmem_softmax_vector_out_index),
+        .out_data   (scalar_tmem_softmax_vector_out_data),
+        .done       (scalar_tmem_softmax_vector_done)
+    );
+    `UNUSED_VAR (scalar_tmem_softmax_vector_busy)
+
+    for (genvar i = 0; i < `ISSUE_WIDTH; ++i) begin : g_scalar_tmem_ready
+        assign scalar_tmem_dispatch_ready[i] = scalar_tmem_grant_valid && scalar_tmem_grant[i] && scalar_tmem_req_ready;
+        assign scalar_tmem_req_addr[i] = lsu_dispatch_if[i].data.rs1_data[0][8:0];
+        assign scalar_tmem_req_wdata[i] = lsu_dispatch_if[i].data.rs2_data;
+        assign scalar_tmem_req_uuid[i] = lsu_dispatch_if[i].data.uuid;
+        assign scalar_tmem_req_wid[i] = wis_to_wid(lsu_dispatch_if[i].data.wis, ISSUE_ISW_W'(i));
+        assign scalar_tmem_req_tmask[i] = lsu_dispatch_if[i].data.tmask;
+        assign scalar_tmem_req_pc[i] = lsu_dispatch_if[i].data.PC;
+        assign scalar_tmem_req_rd[i] = lsu_dispatch_if[i].data.rd;
+        assign scalar_tmem_req_softmax_p_addr[i] = lsu_dispatch_if[i].data.rs2_data[0][8:0];
+
+        assign scalar_tmem_commit_if[i].valid = scalar_tmem_commit_pending && scalar_tmem_grant_r[i];
+        assign scalar_tmem_commit_if[i].data.uuid = scalar_tmem_uuid_r;
+        assign scalar_tmem_commit_if[i].data.wid = scalar_tmem_wid_r;
+        assign scalar_tmem_commit_if[i].data.tmask = scalar_tmem_tmask_r;
+        assign scalar_tmem_commit_if[i].data.PC = scalar_tmem_pc_r;
+        assign scalar_tmem_commit_if[i].data.wb = scalar_tmem_load_pending || scalar_tmem_softmax_commit;
+        assign scalar_tmem_commit_if[i].data.rd = (scalar_tmem_load_pending || scalar_tmem_softmax_commit) ? scalar_tmem_rd_r : '0;
+        assign scalar_tmem_commit_if[i].data.data =
+            scalar_tmem_load_pending
+                ? (scalar_tmem_rdata_valid_r ? scalar_tmem_rdata_r : scalar_tmem_rdata)
+                : {`NUM_THREADS{scalar_tmem_softmax_token}};
+        assign scalar_tmem_commit_if[i].data.tensor = 1'b0;
+        assign scalar_tmem_commit_if[i].data.pid = '0;
+        assign scalar_tmem_commit_if[i].data.sop = 1'b1;
+        assign scalar_tmem_commit_if[i].data.eop = 1'b1;
+        assign scalar_tmem_commit_ready[i] = scalar_tmem_commit_if[i].ready;
+    end
+
+    function automatic [`NUM_THREADS*`XLEN/8-1:0] scalar_tmem_expand_tmask;
+        input [`NUM_THREADS-1:0] tmask;
+        begin
+            scalar_tmem_expand_tmask = '0;
+            for (integer lane = 0; lane < `NUM_THREADS; ++lane) begin
+                scalar_tmem_expand_tmask[lane * (`XLEN / 8) +: (`XLEN / 8)] =
+                    {(`XLEN / 8){tmask[lane]}};
+            end
+        end
+    endfunction
+
+`ifdef DBG_TRACE_CORE_PIPELINE_VCS
+    always @(posedge clk) begin
+        if (!reset && ($time > `TRACE_STARTTIME) && (CORE_ID == 0)) begin
+            if (scalar_tmem_req_fire && scalar_tmem_grant_is_softmax) begin
+                `TRACE(1, ("%d: core%0d-scalar-tmem-softmax-start: grant=%b wid=%0d PC=0x%0h rd=%0d score=%0d prob=%0d tmask=%b (#%0d)\n",
+                    $time, CORE_ID, scalar_tmem_grant, scalar_tmem_req_wid[0], scalar_tmem_req_pc[0],
+                    scalar_tmem_req_rd[0], scalar_tmem_req_addr[0], scalar_tmem_req_softmax_p_addr[0],
+                    scalar_tmem_req_tmask[0], scalar_tmem_req_uuid[0]));
+            end
+            if (scalar_tmem_softmax_read_issue && (scalar_tmem_softmax_read_index == 6'd0 || scalar_tmem_softmax_read_index == SCALAR_TMEM_SOFTMAX_ROW_LAST)) begin
+                `TRACE(1, ("%d: core%0d-scalar-tmem-softmax-read: index=%0d addr=%0d rready=%b PC=0x%0h rd=%0d\n",
+                    $time, CORE_ID, scalar_tmem_softmax_read_index,
+                    scalar_tmem_softmax_read_addr, scalar_tmem_rready, scalar_tmem_pc_r, scalar_tmem_rd_r));
+            end
+            if (scalar_tmem_softmax_write_issue && (scalar_tmem_softmax_vector_out_index == 6'd0 || scalar_tmem_softmax_vector_out_index == SCALAR_TMEM_SOFTMAX_TILE_LAST)) begin
+                `TRACE(1, ("%d: core%0d-scalar-tmem-softmax-write: index=%0d addr=%0d wready=%b PC=0x%0h rd=%0d\n",
+                    $time, CORE_ID, scalar_tmem_softmax_vector_out_index, scalar_tmem_softmax_write_addr,
+                    scalar_tmem_wready, scalar_tmem_pc_r, scalar_tmem_rd_r));
+            end
+            if (scalar_tmem_commit_pending && scalar_tmem_pc_r == 32'h80000318) begin
+                `TRACE(1, ("%d: core%0d-scalar-tmem-softmax-commit-pending: grant_r=%b ready=%b valid0=%b wb0=%0d rd=%0d PC=0x%0h token=0x%0h\n",
+                    $time, CORE_ID, scalar_tmem_grant_r, scalar_tmem_commit_ready,
+                    scalar_tmem_commit_if[0].valid, scalar_tmem_commit_if[0].data.wb,
+                    scalar_tmem_commit_if[0].data.rd, scalar_tmem_commit_if[0].data.PC,
+                    scalar_tmem_softmax_token));
+            end
+        end
+    end
+`endif
+
+    always @(*) begin
+        scalar_tmem_ren = (scalar_tmem_grant_valid && scalar_tmem_grant_is_load)
+                       || scalar_tmem_softmax_read_valid;
+        scalar_tmem_wen = (scalar_tmem_grant_valid && scalar_tmem_grant_is_store)
+                       || scalar_tmem_softmax_write_valid;
+        scalar_tmem_raddr = '0;
+        scalar_tmem_waddr = '0;
+        scalar_tmem_wdata = '0;
+        scalar_tmem_mask = '0;
+        for (integer i = 0; i < `ISSUE_WIDTH; ++i) begin
+            if (scalar_tmem_grant[i]) begin
+                scalar_tmem_raddr = scalar_tmem_req_addr[i];
+                scalar_tmem_waddr = scalar_tmem_req_addr[i];
+                scalar_tmem_wdata = scalar_tmem_req_wdata[i];
+                scalar_tmem_mask = scalar_tmem_expand_tmask(scalar_tmem_req_tmask[i])
+                                 & {(`NUM_THREADS * (`XLEN / 8)){scalar_tmem_grant_is_store}};
+            end
+        end
+        if (scalar_tmem_softmax_read_valid) begin
+            scalar_tmem_raddr = scalar_tmem_softmax_read_addr;
+        end
+        if (scalar_tmem_softmax_write_valid) begin
+            scalar_tmem_waddr = scalar_tmem_softmax_write_addr;
+            scalar_tmem_wdata = scalar_tmem_softmax_wdata;
+            scalar_tmem_mask = scalar_tmem_expand_tmask(scalar_tmem_tmask_r);
+        end
+    end
+
+    always @(posedge clk) begin
+        if (reset) begin
+            scalar_tmem_grant_r <= '0;
+            scalar_tmem_load_pending <= 1'b0;
+            scalar_tmem_store_pending <= 1'b0;
+            scalar_tmem_commit_pending <= 1'b0;
+            scalar_tmem_softmax_active <= 1'b0;
+            scalar_tmem_softmax_read_index <= '0;
+            scalar_tmem_softmax_reads_issued <= 1'b0;
+            scalar_tmem_softmax_read_response_valid <= 1'b0;
+            scalar_tmem_softmax_read_response_index <= '0;
+            scalar_tmem_softmax_score_base_r <= '0;
+            scalar_tmem_softmax_p_base_r <= '0;
+            scalar_tmem_uuid_r <= '0;
+            scalar_tmem_wid_r <= '0;
+            scalar_tmem_tmask_r <= '0;
+            scalar_tmem_pc_r <= '0;
+            scalar_tmem_rd_r <= '0;
+            scalar_tmem_rdata_r <= '0;
+            scalar_tmem_rdata_valid_r <= 1'b0;
+        end else begin
+            if (scalar_tmem_req_fire) begin
+                scalar_tmem_grant_r <= scalar_tmem_grant;
+                scalar_tmem_load_pending <= scalar_tmem_grant_is_load;
+                scalar_tmem_store_pending <= scalar_tmem_grant_is_store;
+                scalar_tmem_commit_pending <= !scalar_tmem_grant_is_softmax;
+                scalar_tmem_rdata_valid_r <= 1'b0;
+                for (integer i = 0; i < `ISSUE_WIDTH; ++i) begin
+                    if (scalar_tmem_grant[i]) begin
+                        scalar_tmem_uuid_r <= scalar_tmem_req_uuid[i];
+                        scalar_tmem_wid_r <= scalar_tmem_req_wid[i];
+                        scalar_tmem_tmask_r <= scalar_tmem_req_tmask[i];
+                        scalar_tmem_pc_r <= scalar_tmem_req_pc[i];
+                        scalar_tmem_rd_r <= scalar_tmem_req_rd[i];
+                        if (scalar_tmem_grant_is_softmax) begin
+                            scalar_tmem_softmax_active <= 1'b1;
+                            scalar_tmem_softmax_read_index <= '0;
+                            scalar_tmem_softmax_reads_issued <= 1'b0;
+                            scalar_tmem_softmax_read_response_valid <= 1'b0;
+                            scalar_tmem_softmax_read_response_index <= '0;
+                            scalar_tmem_softmax_score_base_r <= scalar_tmem_req_addr[i];
+                            scalar_tmem_softmax_p_base_r <= scalar_tmem_req_softmax_p_addr[i];
+                        end
+                    end
+                end
+            end else if (scalar_tmem_load_pending && scalar_tmem_commit_pending && !scalar_tmem_rdata_valid_r) begin
+                scalar_tmem_rdata_r <= scalar_tmem_rdata;
+                scalar_tmem_rdata_valid_r <= 1'b1;
+            end
+
+            scalar_tmem_softmax_read_response_valid <= scalar_tmem_softmax_read_issue;
+            if (scalar_tmem_softmax_read_issue) begin
+                scalar_tmem_softmax_read_response_index <= scalar_tmem_softmax_read_index;
+                if (scalar_tmem_softmax_read_index == SCALAR_TMEM_SOFTMAX_ROW_LAST) begin
+                    scalar_tmem_softmax_reads_issued <= 1'b1;
+                end else begin
+                    scalar_tmem_softmax_read_index <= scalar_tmem_softmax_read_index + 6'd1;
+                end
+            end
+
+            if (scalar_tmem_softmax_vector_done) begin
+                scalar_tmem_softmax_active <= 1'b0;
+                scalar_tmem_commit_pending <= 1'b1;
+            end
+
+            if (scalar_tmem_commit_pending && (|(scalar_tmem_grant_r & scalar_tmem_commit_ready))) begin
+                scalar_tmem_grant_r <= '0;
+                scalar_tmem_load_pending <= 1'b0;
+                scalar_tmem_store_pending <= 1'b0;
+                scalar_tmem_commit_pending <= 1'b0;
+                scalar_tmem_rdata_valid_r <= 1'b0;
+            end
+        end
+    end
+
+    localparam SCALAR_LSU_COMMIT_DATAW = `UUID_WIDTH + `NW_WIDTH + `NUM_THREADS + `XLEN + 1 + `NR_BITS + (`NUM_THREADS * `XLEN) + 1 + 1 + 1 + 1;
+
+    for (genvar i = 0; i < `ISSUE_WIDTH; ++i) begin : g_scalar_lsu_commit
+        VX_stream_arb #(
+            .NUM_INPUTS (2),
+            .DATAW      (SCALAR_LSU_COMMIT_DATAW),
+            .ARBITER    ("R"),
+            .OUT_REG    (1)
+        ) scalar_lsu_commit_arb (
+            .clk       (clk),
+            .reset     (reset),
+            .valid_in  ({scalar_tmem_commit_if[i].valid, scalar_mem_lsu_commit_if[i].valid}),
+            .ready_in  ({scalar_tmem_commit_if[i].ready, scalar_mem_lsu_commit_if[i].ready}),
+            .data_in   ({scalar_tmem_commit_if[i].data, scalar_mem_lsu_commit_if[i].data}),
+            .data_out  (lsu_scalar_commit_if[i].data),
+            .valid_out (lsu_scalar_commit_if[i].valid),
+            .ready_out (lsu_scalar_commit_if[i].ready),
+            `UNUSED_PIN (sel_out)
+        );
+    end
+
 `ifdef EXT_T_ENABLE
     VX_commit_if lsu_tensor_commit_if[`ISSUE_WIDTH]();
 

hw/rtl/core/VX_fpu_unit.sv

@@ -145,6 +145,7 @@ module VX_fpu_unit import VX_fpu_pkg::*; #(
 
             .valid_in   (fpu_req_valid),
             .op_type    (execute_if[block_idx].data.op_type),
+            .op_mod     (execute_if[block_idx].data.op_mod),
             .lane_mask  (execute_if[block_idx].data.tmask),
             .fmt        (fpu_fmt),
             .frm        (fpu_req_frm),
@@ -174,6 +175,7 @@ module VX_fpu_unit import VX_fpu_pkg::*; #(
 
             .valid_in   (fpu_req_valid),
             .op_type    (execute_if[block_idx].data.op_type),
+            .op_mod     (execute_if[block_idx].data.op_mod),
             .lane_mask  (execute_if[block_idx].data.tmask),
             .fmt        (fpu_fmt),
             .frm        (fpu_req_frm),
@@ -204,6 +206,7 @@ module VX_fpu_unit import VX_fpu_pkg::*; #(
             .valid_in   (fpu_req_valid),
             .lane_mask  (execute_if[block_idx].data.tmask),
             .op_type    (execute_if[block_idx].data.op_type),
+            .op_mod     (execute_if[block_idx].data.op_mod),
             .fmt        (fpu_fmt),
             .frm        (fpu_req_frm),
             .dataa      (execute_if[block_idx].data.rs1_data),

hw/rtl/core/VX_tensor_blackwell_core.sv

@@ -28,6 +28,9 @@ module VX_tensor_blackwell_core_block import VX_gpu_pkg::*; #(
     VX_tc_bus_if.master smem_B_if,
     VX_commit_if.master commit_if
 );
+    `STATIC_ASSERT((`NUM_THREADS == 4),
+        ("4-lane Blackwell tensor core wrapper requires NUM_THREADS == 4"))
+
     localparam NUM_LANES = `NUM_THREADS;
     localparam METADATA_QUEUE_DEPTH = 2;
 
@@ -144,10 +147,6 @@ module VX_tensor_blackwell_core_block import VX_gpu_pkg::*; #(
         .io_writeback_bits_data_1(writeback_data[1]),
         .io_writeback_bits_data_2(writeback_data[2]),
         .io_writeback_bits_data_3(writeback_data[3]),
-        .io_writeback_bits_data_4(writeback_data[4]),
-        .io_writeback_bits_data_5(writeback_data[5]),
-        .io_writeback_bits_data_6(writeback_data[6]),
-        .io_writeback_bits_data_7(writeback_data[7]),
 
         .io_respA_ready(tmem_if.rsp_ready),
         .io_respA_valid(tmem_if.rsp_valid),

hw/rtl/core/VX_tmem_softmax_unit.sv

@@ -0,0 +1,383 @@
+// Streaming, synthesizable softmax engine for the Blackwell TMEM path.
+// Four independent lane streams are processed in parallel; each stream has
+// ROW_SIZE elements. The exponential datapath is pipelined with II=1.
+module VX_tmem_softmax_unit #(
+    parameter integer NUM_LANES = 4,
+    parameter integer ROW_SIZE = 32
+) (
+    input  wire                               clk,
+    input  wire                               reset,
+    input  wire                               start,
+    input  wire                               load_valid,
+    input  wire [NUM_LANES-1:0][31:0]         load_data,
+    input  wire                               load_last,
+    output wire                               busy,
+    output reg                                out_valid,
+    input  wire                               out_ready,
+    output reg  [5:0]                         out_index,
+    output reg  [NUM_LANES-1:0][31:0]         out_data,
+    output reg                                done
+);
+    localparam [2:0] STATE_IDLE  = 3'd0;
+    localparam [2:0] STATE_LOAD  = 3'd1;
+    localparam [2:0] STATE_EXP   = 3'd2;
+    localparam [2:0] STATE_RECIP = 3'd3;
+    localparam [2:0] STATE_NORM  = 3'd4;
+
+    reg [2:0] state;
+    reg [$clog2(ROW_SIZE)-1:0] load_index;
+    reg [$clog2(ROW_SIZE)-1:0] exp_feed_index;
+    reg                         exp_feed_done;
+    reg [1:0]                   reciprocal_lane;
+    reg [5:0]                   norm_index;
+
+    reg [31:0] scores [0:ROW_SIZE-1][0:NUM_LANES-1];
+    reg [31:0] max_bits [0:NUM_LANES-1];
+    reg [23:0] exp_values [0:ROW_SIZE-1][0:NUM_LANES-1];
+    reg [29:0] exp_sums [0:NUM_LANES-1];
+    reg [31:0] reciprocals [0:NUM_LANES-1];
+
+    reg [2:0] exp_pipe_valid;
+    reg [$clog2(ROW_SIZE)-1:0] exp_pipe_index [0:2];
+    reg [3:0] exp_int_stage0 [0:NUM_LANES-1];
+    reg [3:0] exp_frac_stage0 [0:NUM_LANES-1];
+    reg [7:0] exp_rem_stage0 [0:NUM_LANES-1];
+    reg [3:0] exp_int_stage1 [0:NUM_LANES-1];
+    reg [23:0] exp_frac_stage1 [0:NUM_LANES-1];
+    reg [23:0] exp_stage2 [0:NUM_LANES-1];
+
+    assign busy = state != STATE_IDLE;
+
+    function automatic f32_greater;
+        input [31:0] a;
+        input [31:0] b;
+        begin
+            if ((a[30:0] == 0) && (b[30:0] == 0)) begin
+                f32_greater = 1'b0;
+            end else if (a[31] != b[31]) begin
+                f32_greater = !a[31];
+            end else if (!a[31]) begin
+                f32_greater = a[30:0] > b[30:0];
+            end else begin
+                f32_greater = a[30:0] < b[30:0];
+            end
+        end
+    endfunction
+
+    // Convert finite FP32 to signed Q12, saturating outside the supported
+    // softmax score range. Subnormals are treated as zero.
+    function automatic signed [23:0] f32_to_q12;
+        input [31:0] value;
+        reg [7:0] exponent;
+        reg [23:0] mantissa;
+        reg [63:0] magnitude;
+        integer shift;
+        begin
+            exponent = value[30:23];
+            mantissa = {1'b1, value[22:0]};
+            magnitude = 0;
+            if (exponent == 0) begin
+                f32_to_q12 = 24'sd0;
+            end else if (exponent == 8'hff) begin
+                f32_to_q12 = value[31] ? -24'sh7fffff : 24'sh7fffff;
+            end else begin
+                shift = integer'(exponent) - 138;
+                if (shift >= 0) begin
+                    magnitude = {40'd0, mantissa} << shift;
+                end else if (shift > -64) begin
+                    magnitude = {40'd0, mantissa} >> (-shift);
+                end
+                if (magnitude > 64'h7fffff)
+                    magnitude = 64'h7fffff;
+                f32_to_q12 = value[31]
+                    ? -$signed({1'b0, magnitude[22:0]})
+                    :  $signed({1'b0, magnitude[22:0]});
+            end
+        end
+    endfunction
+
+    // {integer_part, fractional_lut_index, interpolation_remainder} for
+    // exp(score - max). Inputs below -8 are clamped to exp(-8).
+    function automatic [15:0] exp_range_info;
+        input signed [23:0] score_q12;
+        input signed [23:0] max_q12;
+        reg signed [24:0] distance;
+        reg [15:0] clamped;
+        begin
+            distance = max_q12 - score_q12;
+            if (distance <= 0)
+                clamped = 16'd0;
+            else if (distance >= 25'sd32768)
+                clamped = 16'd32768;
+            else
+                clamped = distance[15:0];
+
+            if (clamped == 16'd32768)
+                exp_range_info = {4'd8, 4'd0, 8'd0};
+            else
+                exp_range_info = {clamped[15:12], clamped[11:8], clamped[7:0]};
+        end
+    endfunction
+
+    function automatic [23:0] exp_frac_lut;
+        input [4:0] index;
+        begin
+            case (index)
+                5'd0:  exp_frac_lut = 24'd1048576;
+                5'd1:  exp_frac_lut = 24'd985046;
+                5'd2:  exp_frac_lut = 24'd925365;
+                5'd3:  exp_frac_lut = 24'd869300;
+                5'd4:  exp_frac_lut = 24'd816632;
+                5'd5:  exp_frac_lut = 24'd767155;
+                5'd6:  exp_frac_lut = 24'd720675;
+                5'd7:  exp_frac_lut = 24'd677012;
+                5'd8:  exp_frac_lut = 24'd635993;
+                5'd9:  exp_frac_lut = 24'd597461;
+                5'd10: exp_frac_lut = 24'd561262;
+                5'd11: exp_frac_lut = 24'd527257;
+                5'd12: exp_frac_lut = 24'd495312;
+                5'd13: exp_frac_lut = 24'd465303;
+                5'd14: exp_frac_lut = 24'd437112;
+                5'd15: exp_frac_lut = 24'd410628;
+                default: exp_frac_lut = 24'd385750;
+            endcase
+        end
+    endfunction
+
+    function automatic [23:0] exp_int_lut;
+        input [3:0] index;
+        begin
+            case (index)
+                4'd0: exp_int_lut = 24'd1048576;
+                4'd1: exp_int_lut = 24'd385750;
+                4'd2: exp_int_lut = 24'd141909;
+                4'd3: exp_int_lut = 24'd52206;
+                4'd4: exp_int_lut = 24'd19205;
+                4'd5: exp_int_lut = 24'd7065;
+                4'd6: exp_int_lut = 24'd2599;
+                4'd7: exp_int_lut = 24'd956;
+                default: exp_int_lut = 24'd352;
+            endcase
+        end
+    endfunction
+
+    // Convert a positive unsigned Q15 probability to FP8 E4M3 using
+    // round-to-nearest-even. Softmax probabilities are in [0, 1], so the
+    // positive subnormal and exponent ranges through 1.0 are sufficient.
+    function automatic [7:0] q15_to_fp8e4m3;
+        input [15:0] value;
+        integer msb;
+        integer i;
+        integer shift;
+        reg [3:0] exponent;
+        reg [4:0] rounded;
+        reg [15:0] remainder;
+        reg [15:0] halfway;
+        begin
+            if (value < 16'd512) begin
+                // E4M3 subnormals and the minimum normal are spaced by 2^-9,
+                // which is 64 in unsigned Q15.
+                rounded = value[10:6];
+                remainder = value & 16'h003f;
+                if ((remainder > 16'd32)
+                 || ((remainder == 16'd32) && rounded[0]))
+                    rounded = rounded + 1'b1;
+                q15_to_fp8e4m3 = {3'b000, rounded};
+            end else begin
+                msb = 0;
+                for (i = 0; i < 16; ++i) begin
+                    if (value[i]) msb = i;
+                end
+                exponent = 4'(msb - 8);
+                shift = integer'(exponent) + 5;
+                rounded = 5'(value >> shift);
+                remainder = value & ((16'd1 << shift) - 1'b1);
+                halfway = 16'd1 << (shift - 1);
+                if ((remainder > halfway)
+                 || ((remainder == halfway) && rounded[0]))
+                    rounded = rounded + 1'b1;
+                if (rounded == 5'd16)
+                    q15_to_fp8e4m3 = {1'b0, exponent + 1'b1, 3'b000};
+                else
+                    q15_to_fp8e4m3 = {1'b0, exponent, rounded[2:0]};
+            end
+        end
+    endfunction
+
+    function automatic [7:0] normalize_to_fp8e4m3;
+        input [23:0] exp_value;
+        input [31:0] reciprocal;
+        reg [55:0] product;
+        reg [55:0] probability;
+        begin
+            product = exp_value * reciprocal;
+            probability = (product + 56'd524288) >> 20;
+            if (probability > 56'd32768)
+                normalize_to_fp8e4m3 = 8'h38;
+            else
+                normalize_to_fp8e4m3 = q15_to_fp8e4m3(probability[15:0]);
+        end
+    endfunction
+
+    function automatic [23:0] interpolate_exp_frac;
+        input [3:0] frac_index;
+        input [7:0] remainder;
+        reg [23:0] upper;
+        reg [23:0] lower;
+        reg [31:0] product;
+        begin
+            upper = exp_frac_lut({1'b0, frac_index});
+            lower = exp_frac_lut({1'b0, frac_index} + 5'd1);
+            product = (upper - lower) * remainder;
+            interpolate_exp_frac = upper - ((product + 32'd128) >> 8);
+        end
+    endfunction
+
+    function automatic [23:0] scale_exp_frac;
+        input [23:0] frac_value;
+        input [3:0] int_index;
+        reg [47:0] product;
+        begin
+            product = frac_value * exp_int_lut(int_index);
+            scale_exp_frac = (product + 48'd524288) >> 20;
+        end
+    endfunction
+
+    function automatic [31:0] reciprocal_q15;
+        input [29:0] sum_value;
+        reg [63:0] numerator;
+        reg [63:0] quotient;
+        begin
+            numerator = (64'd1 << 35) + ({34'd0, sum_value} >> 1);
+            quotient = (sum_value == 0) ? 64'd0 : numerator / {34'd0, sum_value};
+            reciprocal_q15 = quotient[31:0];
+        end
+    endfunction
+
+    wire exp_feed = (state == STATE_EXP) && !exp_feed_done;
+    wire [15:0] exp_feed_range [0:NUM_LANES-1];
+    for (genvar lane = 0; lane < NUM_LANES; ++lane) begin : g_exp_feed_range
+        assign exp_feed_range[lane] = exp_range_info(
+            f32_to_q12(scores[exp_feed_index][lane]),
+            f32_to_q12(max_bits[lane]));
+    end
+
+    integer lane;
+
+    always @(posedge clk) begin
+        if (reset) begin
+            state <= STATE_IDLE;
+            load_index <= '0;
+            exp_feed_index <= '0;
+            exp_feed_done <= 1'b0;
+            reciprocal_lane <= '0;
+            norm_index <= '0;
+            exp_pipe_valid <= '0;
+            out_valid <= 1'b0;
+            out_index <= '0;
+            out_data <= '0;
+            done <= 1'b0;
+        end else begin
+            done <= 1'b0;
+
+            if ((state == STATE_IDLE) && start) begin
+                state <= STATE_LOAD;
+                load_index <= '0;
+                exp_pipe_valid <= '0;
+                out_valid <= 1'b0;
+            end
+
+            if ((state == STATE_LOAD) && load_valid) begin
+                for (lane = 0; lane < NUM_LANES; ++lane) begin
+                    scores[load_index][lane] <= load_data[lane];
+                    if ((load_index == 0) || f32_greater(load_data[lane], max_bits[lane]))
+                        max_bits[lane] <= load_data[lane];
+                end
+                if (load_last) begin
+                    state <= STATE_EXP;
+                    exp_feed_index <= '0;
+                    exp_feed_done <= 1'b0;
+                    exp_pipe_valid <= '0;
+                    for (lane = 0; lane < NUM_LANES; ++lane)
+                        exp_sums[lane] <= '0;
+                end else begin
+                    load_index <= load_index + 1'b1;
+                end
+            end
+
+            exp_pipe_valid[0] <= exp_feed;
+            exp_pipe_valid[1] <= exp_pipe_valid[0];
+            exp_pipe_valid[2] <= exp_pipe_valid[1];
+            if (exp_feed) begin
+                exp_pipe_index[0] <= exp_feed_index;
+                for (lane = 0; lane < NUM_LANES; ++lane) begin
+                    exp_int_stage0[lane] <= exp_feed_range[lane][15:12];
+                    exp_frac_stage0[lane] <= exp_feed_range[lane][11:8];
+                    exp_rem_stage0[lane] <= exp_feed_range[lane][7:0];
+                end
+                if (exp_feed_index == $clog2(ROW_SIZE)'(ROW_SIZE - 1))
+                    exp_feed_done <= 1'b1;
+                else
+                    exp_feed_index <= exp_feed_index + 1'b1;
+            end
+
+            if (exp_pipe_valid[0]) begin
+                exp_pipe_index[1] <= exp_pipe_index[0];
+                for (lane = 0; lane < NUM_LANES; ++lane) begin
+                    exp_frac_stage1[lane] <= interpolate_exp_frac(
+                        exp_frac_stage0[lane], exp_rem_stage0[lane]);
+                    exp_int_stage1[lane] <= exp_int_stage0[lane];
+                end
+            end
+
+            if (exp_pipe_valid[1]) begin
+                exp_pipe_index[2] <= exp_pipe_index[1];
+                for (lane = 0; lane < NUM_LANES; ++lane) begin
+                    exp_stage2[lane] <= scale_exp_frac(
+                        exp_frac_stage1[lane], exp_int_stage1[lane]);
+                end
+            end
+
+            if (exp_pipe_valid[2]) begin
+                for (lane = 0; lane < NUM_LANES; ++lane) begin
+                    exp_values[exp_pipe_index[2]][lane] <= exp_stage2[lane];
+                    exp_sums[lane] <= exp_sums[lane] + exp_stage2[lane];
+                end
+                if (exp_pipe_index[2] == $clog2(ROW_SIZE)'(ROW_SIZE - 1)) begin
+                    state <= STATE_RECIP;
+                    reciprocal_lane <= '0;
+                end
+            end
+
+            if (state == STATE_RECIP) begin
+                reciprocals[reciprocal_lane] <= reciprocal_q15(exp_sums[reciprocal_lane]);
+                if (reciprocal_lane == 2'(NUM_LANES - 1)) begin
+                    state <= STATE_NORM;
+                    norm_index <= '0;
+                    out_valid <= 1'b0;
+                end else begin
+                    reciprocal_lane <= reciprocal_lane + 1'b1;
+                end
+            end
+
+            if (state == STATE_NORM) begin
+                if (out_valid && out_ready && (out_index == 6'd63)) begin
+                    out_valid <= 1'b0;
+                    done <= 1'b1;
+                    state <= STATE_IDLE;
+                end else if (!out_valid || out_ready) begin
+                    for (lane = 0; lane < NUM_LANES; ++lane) begin
+                        out_data[lane] <= {
+                            4{normalize_to_fp8e4m3(
+                                exp_values[norm_index[4:0]][lane], reciprocals[lane])}
+                        };
+                    end
+                    out_index <= norm_index;
+                    out_valid <= 1'b1;
+                    if (norm_index != 6'd63)
+                        norm_index <= norm_index + 1'b1;
+                end
+            end
+        end
+    end
+endmodule

hw/rtl/core/VX_trace.vh

@@ -327,6 +327,7 @@ task trace_ex_op(input int level,
             `INST_FPU_MISC: begin
                 if  (fdst_d) begin
                     case (op_mod)
+                        8: `TRACE(level, ("FEXP.D"));
                         0: `TRACE(level, ("FSGNJ.D"));
                         1: `TRACE(level, ("FSGNJN.D"));
                         2: `TRACE(level, ("FSGNJX.D"));
@@ -338,6 +339,7 @@ task trace_ex_op(input int level,
                     endcase
                 end else begin
                     case (op_mod)
+                        8: `TRACE(level, ("FEXP.S"));
                         0: `TRACE(level, ("FSGNJ.S"));
                         1: `TRACE(level, ("FSGNJN.S"));
                         2: `TRACE(level, ("FSGNJX.S"));

hw/rtl/fpu/VX_fpu_dpi.sv

@@ -31,6 +31,7 @@ module VX_fpu_dpi import VX_fpu_pkg::*; #(
     input wire [TAGW-1:0] tag_in,
     
     input wire [`INST_FPU_BITS-1:0] op_type,
+    input wire [`INST_MOD_BITS-1:0] op_mod,
     input wire [`INST_FMT_BITS-1:0] fmt,
     input wire [`INST_FRM_BITS-1:0] frm,
 
@@ -51,7 +52,8 @@ module VX_fpu_dpi import VX_fpu_pkg::*; #(
     localparam FPU_DIVSQRT = 1;
     localparam FPU_CVT     = 2;
     localparam FPU_NCP     = 3;
-    localparam NUM_FPC     = 4;
+    localparam FPU_EXP     = 4;
+    localparam NUM_FPC     = 5;
     localparam FPC_BITS    = `LOG2UP(NUM_FPC);
 
     localparam RSP_DATAW = (NUM_LANES * `XLEN) + 1 + $bits(fflags_t) + TAGW;
@@ -133,6 +135,7 @@ module VX_fpu_dpi import VX_fpu_pkg::*; #(
             `INST_FPU_I2F:   begin core_select = FPU_CVT; is_itof = 1; end
             `INST_FPU_U2F:   begin core_select = FPU_CVT; is_utof = 1; end
             `INST_FPU_F2F:   begin core_select = FPU_CVT; is_f2f  = 1; end            
+            `INST_FPU_MISC:  begin core_select = `INST_FPU_IS_EXP(op_type, op_mod) ? FPU_EXP : FPU_NCP; end
             default:         begin core_select = FPU_NCP; end
         endcase
     end
@@ -437,6 +440,45 @@ module VX_fpu_dpi import VX_fpu_pkg::*; #(
     end
     endgenerate
 
+    generate
+    begin : fexp
+
+        reg [NUM_LANES-1:0][`XLEN-1:0] result_fexp_r;
+        reg [NUM_LANES-1:0][63:0] result_fexp;
+        fflags_t [NUM_LANES-1:0] fflags_fexp;
+
+        wire fexp_valid = (valid_in && core_select == FPU_EXP);
+        wire fexp_ready = per_core_ready_out[FPU_EXP] || ~per_core_valid_out[FPU_EXP];
+        wire fexp_fire  = fexp_valid && fexp_ready;
+
+        always @(*) begin
+            for (integer i = 0; i < NUM_LANES; ++i) begin
+                dpi_fexp(fexp_fire, int'(dst_fmt), operands[0][i], result_fexp[i], fflags_fexp[i]);
+                result_fexp_r[i] = result_fexp[i][`XLEN-1:0];
+            end
+        end
+
+        fflags_t fflags_merged;
+        `FPU_MERGE_FFLAGS(fflags_merged, fflags_fexp, lane_mask, NUM_LANES);
+
+        VX_shift_register #(
+            .DATAW  (1 + TAGW + NUM_LANES * `XLEN + $bits(fflags_t)),
+            .DEPTH  (`LATENCY_FEXP),
+            .RESETW (1)
+        ) shift_reg (
+            .clk      (clk),
+            .reset    (reset),
+            .enable   (fexp_ready),
+            .data_in  ({fexp_valid, tag_in, result_fexp_r, fflags_merged}),
+            .data_out ({per_core_valid_out[FPU_EXP], per_core_tag_out[FPU_EXP], per_core_result[FPU_EXP], per_core_fflags[FPU_EXP]})
+        );
+
+        assign per_core_has_fflags[FPU_EXP] = 1;
+        assign per_core_ready_in[FPU_EXP] = fexp_ready;
+
+    end
+    endgenerate
+
     ///////////////////////////////////////////////////////////////////////////
 
     assign per_core_ready_in[FPU_DIVSQRT] = is_div ? div_ready_in : sqrt_ready_in;

hw/rtl/fpu/VX_fpu_dsp.sv

@@ -31,6 +31,7 @@ module VX_fpu_dsp import VX_fpu_pkg::*; #(
     input wire [TAGW-1:0] tag_in,
     
     input wire [`INST_FPU_BITS-1:0] op_type,
+    input wire [`INST_MOD_BITS-1:0] op_mod,
     input wire [`INST_FMT_BITS-1:0] fmt,
     input wire [`INST_FRM_BITS-1:0] frm,
 
@@ -51,7 +52,8 @@ module VX_fpu_dsp import VX_fpu_pkg::*; #(
     localparam FPU_DIVSQRT = 1;
     localparam FPU_CVT     = 2;
     localparam FPU_NCP     = 3;
-    localparam NUM_FPC     = 4;
+    localparam FPU_EXP     = 4;
+    localparam NUM_FPC     = 5;
     localparam FPC_BITS    = `LOG2UP(NUM_FPC);
 
     localparam RSP_DATAW = (NUM_LANES * 32) + 1 + $bits(fflags_t) + TAGW;
@@ -98,6 +100,7 @@ module VX_fpu_dsp import VX_fpu_pkg::*; #(
             `INST_FPU_F2U:    begin core_select = FPU_CVT; end
             `INST_FPU_I2F:    begin core_select = FPU_CVT; is_itof = 1; is_signed = 1; end
             `INST_FPU_U2F:    begin core_select = FPU_CVT; is_itof = 1; end
+            `INST_FPU_MISC:   begin core_select = `INST_FPU_IS_EXP(op_type, op_mod) ? FPU_EXP : FPU_NCP; end
             default:          begin core_select = FPU_NCP; end
         endcase
     end
@@ -107,6 +110,7 @@ module VX_fpu_dsp import VX_fpu_pkg::*; #(
     `RESET_RELAY (sqrt_reset, reset);
     `RESET_RELAY (cvt_reset, reset);
     `RESET_RELAY (ncp_reset, reset);
+    `RESET_RELAY (exp_reset, reset);
 
     wire [NUM_LANES-1:0][31:0] dataa_s;
     wire [NUM_LANES-1:0][31:0] datab_s;
@@ -243,6 +247,25 @@ module VX_fpu_dsp import VX_fpu_pkg::*; #(
         .ready_out  (per_core_ready_out[FPU_NCP])
     );
 
+    VX_fpu_exp #(
+        .NUM_LANES (NUM_LANES),
+        .TAGW      (TAGW)
+    ) fpu_exp (
+        .clk        (clk),
+        .reset      (exp_reset),
+        .valid_in   (valid_in && (core_select == FPU_EXP)),
+        .ready_in   (per_core_ready_in[FPU_EXP]),
+        .lane_mask  (lane_mask),
+        .tag_in     (tag_in),
+        .dataa      (dataa_s),
+        .has_fflags (per_core_has_fflags[FPU_EXP]),
+        .fflags     (per_core_fflags[FPU_EXP]),
+        .result     (per_core_result[FPU_EXP]),
+        .tag_out    (per_core_tag_out[FPU_EXP]),
+        .valid_out  (per_core_valid_out[FPU_EXP]),
+        .ready_out  (per_core_ready_out[FPU_EXP])
+    );
+
     ///////////////////////////////////////////////////////////////////////////
 
     assign per_core_ready_in[FPU_DIVSQRT] = is_div ? div_ready_in : sqrt_ready_in;

hw/rtl/fpu/VX_fpu_exp.sv

@@ -0,0 +1,92 @@
+// Copyright © 2019-2023
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+// http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+`include "VX_fpu_define.vh"
+`ifdef SV_DPI
+`include "float_dpi.vh"
+`else
+`ERROR(("VX_fpu_exp requires SV_DPI; replace dpi_fexp with synthesizable exp RTL for synthesis"))
+`endif
+
+module VX_fpu_exp import VX_fpu_pkg::*; #(
+    parameter NUM_LANES = 1,
+    parameter TAGW = 1
+) (
+    input wire clk,
+    input wire reset,
+
+    output wire ready_in,
+    input wire  valid_in,
+
+    input wire [NUM_LANES-1:0] lane_mask,
+
+    input wire [TAGW-1:0] tag_in,
+
+    input wire [NUM_LANES-1:0][31:0] dataa,
+    output wire [NUM_LANES-1:0][31:0] result,
+
+    output wire has_fflags,
+    output wire [`FP_FLAGS_BITS-1:0] fflags,
+
+    output wire [TAGW-1:0] tag_out,
+
+    input wire  ready_out,
+    output wire valid_out
+);
+    wire stall  = ~ready_out && valid_out;
+    wire enable = ~stall;
+
+    fflags_t [NUM_LANES-1:0] per_lane_fflags;
+    wire [NUM_LANES-1:0] lane_mask_out;
+
+    VX_shift_register #(
+        .DATAW  (1 + NUM_LANES + TAGW),
+        .DEPTH  (`LATENCY_FEXP),
+        .RESETW (1)
+    ) shift_reg (
+        .clk      (clk),
+        .reset    (reset),
+        .enable   (enable),
+        .data_in  ({valid_in, lane_mask, tag_in}),
+        .data_out ({valid_out, lane_mask_out, tag_out})
+    );
+
+    assign ready_in = enable;
+
+    for (genvar i = 0; i < NUM_LANES; ++i) begin
+        reg [63:0] r;
+        `UNUSED_VAR (r)
+
+        fflags_t f;
+
+        always @(*) begin
+            dpi_fexp(enable && valid_in, int'(0), {32'hffffffff, dataa[i]}, r, f);
+        end
+
+        VX_shift_register #(
+            .DATAW  (32 + $bits(fflags_t)),
+            .DEPTH  (`LATENCY_FEXP)
+        ) shift_req_dpi (
+            .clk      (clk),
+            `UNUSED_PIN (reset),
+            .enable   (enable),
+            .data_in  ({r[31:0], f}),
+            .data_out ({result[i], per_lane_fflags[i]})
+        );
+    end
+
+    assign has_fflags = 1;
+
+    `FPU_MERGE_FFLAGS(fflags, per_lane_fflags, lane_mask_out, NUM_LANES);
+
+endmodule

hw/rtl/fpu/VX_fpu_fpnew.sv

@@ -34,6 +34,7 @@ module VX_fpu_fpnew
     input wire [TAGW-1:0] tag_in,
     
     input wire [`INST_FPU_BITS-1:0] op_type,
+    input wire [`INST_MOD_BITS-1:0] op_mod,
     input wire [`INST_FMT_BITS-1:0] fmt,
     input wire [`INST_FRM_BITS-1:0] frm,
 
@@ -104,6 +105,7 @@ module VX_fpu_fpnew
     reg fpu_has_fflags, fpu_has_fflags_out;
     fpnew_pkg::fp_format_e fpu_src_fmt, fpu_dst_fmt;
     fpnew_pkg::int_format_e fpu_int_fmt;
+    wire is_fexp = `INST_FPU_IS_EXP(op_type, op_mod);
 
     `UNUSED_VAR (fmt)
 
@@ -183,7 +185,6 @@ module VX_fpu_fpnew
     end
 
 `ifdef XLEN_64
-    `UNUSED_VAR (lane_mask)
     for (genvar i = 0; i < NUM_LANES; ++i) begin
         wire [(TAGW+1)-1:0] fpu_tag;        
         wire fpu_valid_out_uq;
@@ -248,8 +249,8 @@ module VX_fpu_fpnew
         .int_fmt_i      (fpu_int_fmt),
         .vectorial_op_i (1'b1),
         .simd_mask_i    (lane_mask),
-        .tag_i          ({fpu_tag_in, fpu_has_fflags}),        
-        .in_valid_i     (fpu_valid_in),
+            .tag_i          ({fpu_tag_in, fpu_has_fflags}),
+            .in_valid_i     (fpu_valid_in),
         .in_ready_o     (fpu_ready_in),
         .flush_i        (reset),
         .result_o       (fpu_result),
@@ -261,10 +262,70 @@ module VX_fpu_fpnew
     );
 `endif
 
-    assign fpu_valid_in = valid_in;
-    assign ready_in = fpu_ready_in;
+    wire exp_ready_in;
+    wire exp_valid_out;
+    wire [NUM_LANES-1:0][31:0] exp_dataa;
+    wire [NUM_LANES-1:0][31:0] exp_result_s;
+    wire [NUM_LANES-1:0][`XLEN-1:0] exp_result;
+    wire exp_has_fflags;
+    fflags_t exp_fflags;
+    wire [TAGW-1:0] exp_tag_out;
+    wire exp_ready_out;
+
+    for (genvar i = 0; i < NUM_LANES; ++i) begin
+        assign exp_dataa[i] = dataa[i][31:0];
+    `ifdef FPU_RV64F
+        assign exp_result[i] = {32'hffffffff, exp_result_s[i]};
+    `else
+        assign exp_result[i] = exp_result_s[i];
+    `endif
+    end
+
+    VX_fpu_exp #(
+        .NUM_LANES (NUM_LANES),
+        .TAGW      (TAGW)
+    ) fpu_exp (
+        .clk        (clk),
+        .reset      (reset),
+        .valid_in   (valid_in && is_fexp),
+        .ready_in   (exp_ready_in),
+        .lane_mask  (lane_mask),
+        .tag_in     (tag_in),
+        .dataa      (exp_dataa),
+        .has_fflags (exp_has_fflags),
+        .fflags     (exp_fflags),
+        .result     (exp_result_s),
+        .tag_out    (exp_tag_out),
+        .valid_out  (exp_valid_out),
+        .ready_out  (exp_ready_out)
+    );
+
+    assign fpu_valid_in = valid_in && !is_fexp;
+    assign ready_in = is_fexp ? exp_ready_in : fpu_ready_in;
     assign fpu_tag_in = tag_in;
 
+    wire [RSP_DATAW-1:0] exp_rsp_data;
+    wire exp_rsp_valid;
+    wire exp_rsp_ready;
+    wire [RSP_DATAW-1:0] fpu_rsp_data;
+    wire fpu_rsp_valid;
+    wire fpu_rsp_ready;
+
+    VX_elastic_buffer #(
+        .DATAW   (RSP_DATAW),
+        .SIZE    (`OUT_REG_TO_EB_SIZE(OUT_REG)),
+        .OUT_REG (`OUT_REG_TO_EB_REG(OUT_REG))
+    ) exp_rsp_buf (
+        .clk       (clk),
+        .reset     (reset),
+        .valid_in  (exp_valid_out),
+        .ready_in  (exp_ready_out),
+        .data_in   ({exp_result, exp_has_fflags, exp_fflags, exp_tag_out}),
+        .data_out  (exp_rsp_data),
+        .valid_out (exp_rsp_valid),
+        .ready_out (exp_rsp_ready)
+    );
+
     VX_elastic_buffer #(
         .DATAW   (RSP_DATAW),
         .SIZE    (`OUT_REG_TO_EB_SIZE(OUT_REG)),
@@ -275,9 +336,26 @@ module VX_fpu_fpnew
         .valid_in  (fpu_valid_out),
         .ready_in  (fpu_ready_out),
         .data_in   ({fpu_result, fpu_has_fflags_out, fpu_status, fpu_tag_out}),
+        .data_out  (fpu_rsp_data),
+        .valid_out (fpu_rsp_valid),
+        .ready_out (fpu_rsp_ready)
+    );
+
+    VX_stream_arb #(
+        .NUM_INPUTS (2),
+        .DATAW      (RSP_DATAW),
+        .ARBITER    ("R"),
+        .OUT_REG    (0)
+    ) rsp_arb (
+        .clk       (clk),
+        .reset     (reset),
+        .valid_in  ({exp_rsp_valid, fpu_rsp_valid}),
+        .ready_in  ({exp_rsp_ready, fpu_rsp_ready}),
+        .data_in   ({exp_rsp_data, fpu_rsp_data}),
         .data_out  ({result, has_fflags, fflags, tag_out}),
         .valid_out (valid_out),
-        .ready_out (ready_out)
+        .ready_out (ready_out),
+        `UNUSED_PIN (sel_out)
     );
 
 endmodule

kernel/include/VX_config.h

@@ -402,6 +402,11 @@
 #endif
 #endif
 
+// FEXP Latency
+#ifndef LATENCY_FEXP
+#define LATENCY_FEXP 8
+#endif
+
 // FCVT Latency
 #ifndef LATENCY_FCVT
 #define LATENCY_FCVT 5