summaryrefslogtreecommitdiff
path: root/platform
diff options
context:
space:
mode:
Diffstat (limited to '')
-rw-r--r--platform/wavelet3d/gfx_fpint.sv103
-rw-r--r--platform/wavelet3d/gfx_pkg.sv2
-rw-r--r--platform/wavelet3d/main.cpp238
3 files changed, 215 insertions, 128 deletions
diff --git a/platform/wavelet3d/gfx_fpint.sv b/platform/wavelet3d/gfx_fpint.sv
index babc916..e5c457c 100644
--- a/platform/wavelet3d/gfx_fpint.sv
+++ b/platform/wavelet3d/gfx_fpint.sv
@@ -1,34 +1,39 @@
module gfx_fpint
(
- input logic clk,
+ input logic clk,
+ rst_n,
- input gfx::word a,
- b,
- input logic setup_mul_float,
- setup_unit_b,
- mnorm_put_hi,
- mnorm_put_lo,
- mnorm_put_mul,
- mnorm_zero_b,
- mnorm_zero_flags,
- minmax_copy_flags,
- shiftr_int_signed,
- addsub_copy_flags,
- addsub_int_operand,
- clz_force_nop,
- shiftl_copy_flags,
- round_copy_flags,
- round_enable,
- encode_enable,
+ input gfx::word a[gfx::SHADER_LANES],
+ b[gfx::SHADER_LANES],
+ input logic in_valid,
+ setup_mul_float,
+ setup_unit_b,
+ mnorm_put_hi,
+ mnorm_put_lo,
+ mnorm_put_mul,
+ mnorm_zero_b,
+ mnorm_zero_flags,
+ minmax_copy_flags,
+ shiftr_int_signed,
+ addsub_copy_flags,
+ addsub_int_operand,
+ clz_force_nop,
+ shiftl_copy_flags,
+ round_copy_flags,
+ round_enable,
+ encode_enable,
- output gfx::word q
+ output logic out_valid,
+ output gfx::word q[gfx::SHADER_LANES]
);
import gfx::*;
+ logic stage_valid[FPINT_STAGES];
fpint_op op, stage_op[FPINT_STAGES];
assign stage_op[0] = op;
+ assign stage_valid[0] = in_valid;
assign op.setup_mul_float = setup_mul_float;
assign op.setup_unit_b = setup_unit_b;
@@ -47,32 +52,44 @@ module gfx_fpint
assign op.round_enable = round_enable;
assign op.encode_enable = encode_enable;
- gfx_fpint_lane lane
- (
- .clk(clk),
- .a(a),
- .b(b),
- .q(q),
- .mul_float_0(stage_op[0].setup_mul_float),
- .unit_b_0(stage_op[0].setup_unit_b),
- .put_hi_2(stage_op[2].mnorm_put_hi),
- .put_lo_2(stage_op[2].mnorm_put_lo),
- .put_mul_2(stage_op[2].mnorm_put_mul),
- .zero_b_2(stage_op[2].mnorm_zero_b),
- .zero_flags_2(stage_op[2].mnorm_zero_flags),
- .copy_flags_3(stage_op[3].minmax_copy_flags),
- .int_signed_5(stage_op[5].shiftr_int_signed),
- .copy_flags_6(stage_op[6].addsub_copy_flags),
- .int_operand_6(stage_op[6].addsub_int_operand),
- .force_nop_7(stage_op[7].clz_force_nop),
- .copy_flags_11(stage_op[11].shiftl_copy_flags),
- .copy_flags_12(stage_op[12].round_copy_flags),
- .enable_12(stage_op[12].round_enable),
- .enable_14(stage_op[14].encode_enable)
- );
+ genvar lane;
+ generate
+ for (lane = 0; lane < SHADER_LANES; ++lane) begin: lanes
+ gfx_fpint_lane unit
+ (
+ .clk(clk),
+ .a(a[lane]),
+ .b(b[lane]),
+ .q(q[lane]),
+ .mul_float_0(stage_op[0].setup_mul_float),
+ .unit_b_0(stage_op[0].setup_unit_b),
+ .put_hi_2(stage_op[2].mnorm_put_hi),
+ .put_lo_2(stage_op[2].mnorm_put_lo),
+ .put_mul_2(stage_op[2].mnorm_put_mul),
+ .zero_b_2(stage_op[2].mnorm_zero_b),
+ .zero_flags_2(stage_op[2].mnorm_zero_flags),
+ .copy_flags_3(stage_op[3].minmax_copy_flags),
+ .int_signed_5(stage_op[5].shiftr_int_signed),
+ .copy_flags_6(stage_op[6].addsub_copy_flags),
+ .int_operand_6(stage_op[6].addsub_int_operand),
+ .force_nop_7(stage_op[7].clz_force_nop),
+ .copy_flags_11(stage_op[11].shiftl_copy_flags),
+ .copy_flags_12(stage_op[12].round_copy_flags),
+ .enable_12(stage_op[12].round_enable),
+ .enable_14(stage_op[14].encode_enable)
+ );
+ end
+ endgenerate
always_ff @(posedge clk)
for (int i = 1; i < FPINT_STAGES; ++i)
stage_op[i] <= stage_op[i - 1];
+ always_ff @(posedge clk or negedge rst_n) begin
+ for (int i = 1; i < FPINT_STAGES; ++i)
+ stage_valid[i] <= !rst_n ? 0 : stage_valid[i - 1];
+
+ out_valid <= !rst_n ? 0 : stage_valid[FPINT_STAGES - 1];
+ end
+
endmodule
diff --git a/platform/wavelet3d/gfx_pkg.sv b/platform/wavelet3d/gfx_pkg.sv
index 5c420cc..75a564a 100644
--- a/platform/wavelet3d/gfx_pkg.sv
+++ b/platform/wavelet3d/gfx_pkg.sv
@@ -1,5 +1,7 @@
package gfx;
+ localparam int SHADER_LANES = 4;
+
typedef logic[31:0] word;
typedef logic[63:0] dword;
typedef logic[7:0] float_exp;
diff --git a/platform/wavelet3d/main.cpp b/platform/wavelet3d/main.cpp
index 49c96c1..2555d1d 100644
--- a/platform/wavelet3d/main.cpp
+++ b/platform/wavelet3d/main.cpp
@@ -2,6 +2,7 @@
#include <cstddef>
#include <cstdio>
#include <cstdlib>
+#include <string>
#include <Python.h>
#include <verilated.h>
@@ -23,99 +24,166 @@ int main(int argc, char **argv)
Py_Initialize();
- float q;
- int a, b;
- const char *op = "int->fp";
-
- std::cin >> a >> b;
-
- // int->fp
- top.setup_mul_float = 0;
- top.setup_unit_b = 1;
- top.mnorm_put_hi = 0;
- top.mnorm_put_lo = 1;
- top.mnorm_put_mul = 0;
- top.mnorm_zero_flags = 1;
- top.mnorm_zero_b = 1;
- top.minmax_copy_flags = 0;
- top.shiftr_int_signed = 1;
- top.addsub_int_operand = 1;
- top.addsub_copy_flags = 1;
- top.clz_force_nop = 1;
- top.shiftl_copy_flags = 0;
- top.round_copy_flags = 0;
- top.round_enable = 1;
- top.encode_enable = 1;
-
- // mul int
- //top.setup_mul_float = 0;
- //top.setup_unit_b = 0;
- //top.mnorm_put_hi = 0;
- //top.mnorm_put_lo = 1;
- //top.mnorm_put_mul = 0;
- //top.mnorm_zero_flags = 1;
- //top.mnorm_zero_b = 1;
- //top.minmax_copy_flags = 1;
- //top.shiftr_int_signed = 0;
- //top.addsub_int_operand = 0;
- //top.addsub_copy_flags = 1;
- //top.clz_force_nop = 0;
- //top.shiftl_copy_flags = 1;
- //top.round_copy_flags = 1;
- //top.round_enable = 0;
- //top.encode_enable = 0;
+ int time = 0;
- // mul fp
- //top.setup_mul_float = 1;
- //top.setup_unit_b = 0;
- //top.mnorm_put_hi = 0;
- //top.mnorm_put_lo = 0;
- //top.mnorm_put_mul = 1;
- //top.mnorm_zero_flags = 0;
- //top.mnorm_zero_b = 1;
- //top.minmax_copy_flags = 1;
- //top.shiftr_int_signed = 0;
- //top.addsub_int_operand = 0;
- //top.addsub_copy_flags = 1;
- //top.clz_force_nop = 1;
- //top.shiftl_copy_flags = 1;
- //top.round_copy_flags = 1;
- //top.round_enable = 1;
- //top.encode_enable = 1;
-
- // suma/resta
- //top.setup_mul_float = 0;
- //top.setup_unit_b = 1;
- //top.mnorm_put_hi = 0;
- //top.mnorm_put_lo = 1;
- //top.mnorm_put_mul = 0;
- //top.mnorm_zero_flags = 0;
- //top.mnorm_zero_b = 0;
- //top.minmax_copy_flags = 0;
- //top.shiftr_int_signed = 0;
- //top.addsub_int_operand = 0;
- //top.addsub_copy_flags = 0;
- //top.clz_force_nop = 1;
- //top.shiftl_copy_flags = 0;
- //top.round_copy_flags = 0;
- //top.round_enable = 1;
- //top.encode_enable = 1;
-
- top.a = *reinterpret_cast<unsigned*>(&a);
- top.b = *reinterpret_cast<unsigned*>(&b);
-
- for (int i = 0; i < 1000; ++i) {
+ auto cycle = [&]()
+ {
top.clk = 0;
top.eval();
top.clk = 1;
top.eval();
- }
- unsigned q_bits = top.q;
- q = *reinterpret_cast<decltype(q)*>(&q_bits);
+ ++time;
+ };
+
+ auto send_op = [&](auto a, const char *op, auto b)
+ {
+ std::printf
+ (
+ "[%03d] %s %s %s\n",
+ time, std::to_string(a).c_str(), op, std::to_string(b).c_str()
+ );
+
+ top.a[0] = *reinterpret_cast<unsigned*>(&a);
+ top.b[0] = *reinterpret_cast<unsigned*>(&b);
+ top.in_valid = 1;
+ cycle();
+
+ top.a[0] = 0;
+ top.b[0] = 0;
+ top.in_valid = 0;
+ };
+
+ auto int_to_fp = [&](int a)
+ {
+ top.setup_mul_float = 0;
+ top.setup_unit_b = 1;
+ top.mnorm_put_hi = 0;
+ top.mnorm_put_lo = 1;
+ top.mnorm_put_mul = 0;
+ top.mnorm_zero_flags = 1;
+ top.mnorm_zero_b = 1;
+ top.minmax_copy_flags = 0;
+ top.shiftr_int_signed = 1;
+ top.addsub_int_operand = 1;
+ top.addsub_copy_flags = 1;
+ top.clz_force_nop = 0;
+ top.shiftl_copy_flags = 0;
+ top.round_copy_flags = 0;
+ top.round_enable = 1;
+ top.encode_enable = 1;
+
+ send_op(a, "fp->int", 0);
+ };
+
+ auto mul_int = [&](unsigned a, unsigned b)
+ {
+ top.setup_mul_float = 0;
+ top.setup_unit_b = 0;
+ top.mnorm_put_hi = 0;
+ top.mnorm_put_lo = 1;
+ top.mnorm_put_mul = 0;
+ top.mnorm_zero_flags = 1;
+ top.mnorm_zero_b = 1;
+ top.minmax_copy_flags = 1;
+ top.shiftr_int_signed = 0;
+ top.addsub_int_operand = 0;
+ top.addsub_copy_flags = 1;
+ top.clz_force_nop = 1;
+ top.shiftl_copy_flags = 1;
+ top.round_copy_flags = 1;
+ top.round_enable = 0;
+ top.encode_enable = 0;
+
+ send_op(a, "*", b);
+ };
- std::cout << a << ' ' << op << ' ' << b << " = " << q << '\n';
+ // mul fp
+ auto mul_fp = [&](float a, float b)
+ {
+ top.setup_mul_float = 1;
+ top.setup_unit_b = 0;
+ top.mnorm_put_hi = 0;
+ top.mnorm_put_lo = 0;
+ top.mnorm_put_mul = 1;
+ top.mnorm_zero_flags = 0;
+ top.mnorm_zero_b = 1;
+ top.minmax_copy_flags = 1;
+ top.shiftr_int_signed = 0;
+ top.addsub_int_operand = 0;
+ top.addsub_copy_flags = 1;
+ top.clz_force_nop = 1;
+ top.shiftl_copy_flags = 1;
+ top.round_copy_flags = 1;
+ top.round_enable = 1;
+ top.encode_enable = 1;
+
+ send_op(a, "*", b);
+ };
+
+ auto add_fp = [&](float a, float b)
+ {
+ top.setup_mul_float = 0;
+ top.setup_unit_b = 1;
+ top.mnorm_put_hi = 0;
+ top.mnorm_put_lo = 1;
+ top.mnorm_put_mul = 0;
+ top.mnorm_zero_flags = 0;
+ top.mnorm_zero_b = 0;
+ top.minmax_copy_flags = 0;
+ top.shiftr_int_signed = 0;
+ top.addsub_int_operand = 0;
+ top.addsub_copy_flags = 0;
+ top.clz_force_nop = 0;
+ top.shiftl_copy_flags = 0;
+ top.round_copy_flags = 0;
+ top.round_enable = 1;
+ top.encode_enable = 1;
+
+ send_op(a, "+", b);
+ };
+
+ top.rst_n = 0;
+ top.in_valid = 0;
+ cycle();
+
+ top.rst_n = 1;
+ cycle();
+
+ int a, b;
+ float a_flt, b_flt;
+
+ std::cout << "a_int: ";
+ std::cin >> a;
+ std::cout << "b_int: ";
+ std::cin >> b;
+ std::cout << "a_flt: ";
+ std::cin >> a_flt;
+ std::cout << "b_flt: ";
+ std::cin >> b_flt;
+
+ int_to_fp(a);
+ mul_int(a, b);
+ mul_fp(a_flt, b_flt);
+ add_fp(a_flt, b_flt);
+
+ while (time < 20) {
+ cycle();
+
+ if (!top.out_valid)
+ continue;
+
+ unsigned q_bits = top.q[0];
+ int q_int = *reinterpret_cast<int*>(&q_bits);
+ float q_flt = *reinterpret_cast<float*>(&q_bits);
+
+ std::printf
+ (
+ "[%03d] q=0x%08x, q_flt=%g, q_int=%d, q_uint=%u\n",
+ time, q_bits, q_flt, q_int, q_bits
+ );
+ }
bool failed = Py_FinalizeEx() < 0;
generated by cgit v1.2.3 (git 2.25.1) at 2026-04-17 10:44:51 +0000