chipsalliance · Avimitin · Nov 18, 2024 · Nov 18, 2024 · Nov 18, 2024 · Nov 19, 2024
diff --git a/.github/designs/benchmark_dlen1024_vlen1024_fp/t1rocketemu.json b/.github/designs/benchmark_dlen1024_vlen1024_fp/t1rocketemu.json
@@ -0,0 +1,6 @@
+{
+  "rvv_bench.ascii_to_utf32": 226749,
+  "rvv_bench.byteswap": 391077,
+  "rvv_bench.memcpy": 589350,
+  "rvv_bench.memset": 285597
+}
diff --git a/.github/designs/benchmark_dlen128_vlen1024_fp/t1rocketemu.json b/.github/designs/benchmark_dlen128_vlen1024_fp/t1rocketemu.json
@@ -0,0 +1,6 @@
+{
+  "rvv_bench.ascii_to_utf32": 229892,
+  "rvv_bench.byteswap": 415921,
+  "rvv_bench.memcpy": 592239,
+  "rvv_bench.memset": 295276
+}
diff --git a/.github/designs/benchmark_dlen256_vlen1024_fp/t1rocketemu.json b/.github/designs/benchmark_dlen256_vlen1024_fp/t1rocketemu.json
@@ -0,0 +1,6 @@
+{
+  "rvv_bench.ascii_to_utf32": 228089,
+  "rvv_bench.byteswap": 401959,
+  "rvv_bench.memcpy": 590978,
+  "rvv_bench.memset": 289402
+}
diff --git a/.github/designs/benchmark_dlen512_vlen1024_fp/t1rocketemu.json b/.github/designs/benchmark_dlen512_vlen1024_fp/t1rocketemu.json
@@ -0,0 +1,6 @@
+{
+  "rvv_bench.ascii_to_utf32": 227173,
+  "rvv_bench.byteswap": 394703,
+  "rvv_bench.memcpy": 589878,
+  "rvv_bench.memset": 286745
+}
diff --git a/.github/designs/blastoise/t1emu.json b/.github/designs/blastoise/t1emu.json
@@ -13,11 +13,11 @@
   "intrinsic.softmax": 7032,
   "rvv_bench.ascii_to_utf16": 11282,
   "rvv_bench.ascii_to_utf32": 4694,
-  "rvv_bench.byteswap": 19954,
+  "rvv_bench.byteswap": 20017,
   "rvv_bench.mandelbrot": 230851,
-  "rvv_bench.memcpy": 34534,
+  "rvv_bench.memcpy": 34718,
   "rvv_bench.memset": 11501,
   "rvv_bench.mergelines": 24842,
   "rvv_bench.strlen": 22697,
-  "rvv_bench.utf8_count": 151155
+  "rvv_bench.utf8_count": 151165
 }
diff --git a/.github/designs/blastoise/t1rocketemu.json b/.github/designs/blastoise/t1rocketemu.json
@@ -1,9 +1,9 @@
 {
-  "asm.memcpy": 751,
-  "asm.mmm": 51749,
-  "asm.smoke": 8001,
-  "asm.strlen": 7986,
-  "asm.utf8_count": 205,
+  "asm.memcpy": 774,
+  "asm.mmm": 51794,
+  "asm.smoke": 8051,
+  "asm.strlen": 8009,
+  "asm.utf8_count": 232,
   "codegen.vaadd_vv": 170452,
   "codegen.vaadd_vx": 539363,
   "codegen.vaaddu_vv": 170452,
@@ -499,30 +499,30 @@
   "codegen.vxor_vx": 136518,
   "codegen.vzext_vf2": 14020,
   "codegen.vzext_vf4": 4333,
-  "intrinsic.conv2d_less_m2": 2498,
-  "intrinsic.linear_normalization": 3350,
-  "intrinsic.matmul": 65866,
-  "intrinsic.softmax": 6795,
-  "mlir.axpy_masked": 4048,
-  "mlir.conv": 125859,
-  "mlir.hello": 131,
-  "mlir.matmul": 56059,
-  "mlir.maxvl_tail_setvl_front": 700,
-  "mlir.rvv_vp_intrinsic_add": 466,
-  "mlir.rvv_vp_intrinsic_add_scalable": 807,
-  "mlir.stripmining": 8882,
-  "mlir.vectoradd": 13236,
-  "pytorch.demo": 31521,
-  "pytorch.matmul": 69793,
-  "rvv_bench.ascii_to_utf16": 677090,
-  "rvv_bench.ascii_to_utf32": 226918,
-  "rvv_bench.byteswap": 399524,
-  "rvv_bench.chacha20": 39957,
-  "rvv_bench.mandelbrot": 512683,
-  "rvv_bench.memcpy": 671955,
-  "rvv_bench.memset": 290725,
-  "rvv_bench.mergelines": 564159,
-  "rvv_bench.poly1305": 39957,
-  "rvv_bench.strlen": 219139,
-  "rvv_bench.utf8_count": 2283382
+  "intrinsic.conv2d_less_m2": 2550,
+  "intrinsic.linear_normalization": 3373,
+  "intrinsic.matmul": 65918,
+  "intrinsic.softmax": 6824,
+  "mlir.axpy_masked": 4098,
+  "mlir.conv": 125881,
+  "mlir.hello": 175,
+  "mlir.matmul": 56109,
+  "mlir.maxvl_tail_setvl_front": 749,
+  "mlir.rvv_vp_intrinsic_add": 517,
+  "mlir.rvv_vp_intrinsic_add_scalable": 858,
+  "mlir.stripmining": 8931,
+  "mlir.vectoradd": 13281,
+  "pytorch.demo": 31524,
+  "pytorch.matmul": 69835,
+  "rvv_bench.ascii_to_utf16": 690179,
+  "rvv_bench.ascii_to_utf32": 228538,
+  "rvv_bench.byteswap": 399715,
+  "rvv_bench.chacha20": 40010,
+  "rvv_bench.mandelbrot": 512811,
+  "rvv_bench.memcpy": 589522,
+  "rvv_bench.memset": 290522,
+  "rvv_bench.mergelines": 563328,
+  "rvv_bench.poly1305": 40010,
+  "rvv_bench.strlen": 219124,
+  "rvv_bench.utf8_count": 2230881
 }
diff --git a/.github/designs/rookidee/t1emu.json b/.github/designs/rookidee/t1emu.json
@@ -8,8 +8,8 @@
   "mlir.stripmining": 27810,
   "rvv_bench.ascii_to_utf16": 16070,
   "rvv_bench.ascii_to_utf32": 6057,
-  "rvv_bench.byteswap": 43274,
-  "rvv_bench.memcpy": 46320,
+  "rvv_bench.byteswap": 43109,
+  "rvv_bench.memcpy": 46538,
   "rvv_bench.memset": 19350,
   "rvv_bench.mergelines": 38541,
   "rvv_bench.strlen": 34517,

diff --git a/.github/designs/rookidee/t1rocketemu.json b/.github/designs/rookidee/t1rocketemu.json
@@ -1,6 +1,6 @@
 {
-  "asm.mmm": 56295,
-  "asm.smoke": 7780,
+  "asm.mmm": 56340,
+  "asm.smoke": 7830,
   "codegen.vaadd_vv": 129499,
   "codegen.vaadd_vx": 393229,
   "codegen.vaaddu_vv": 129499,
@@ -430,19 +430,19 @@
   "codegen.vxor_vx": 98738,
   "codegen.vzext_vf2": 11522,
   "codegen.vzext_vf4": 3497,
-  "intrinsic.conv2d_less_m2": 2540,
-  "mlir.hello": 130,
-  "mlir.rvv_vp_intrinsic_add": 462,
-  "mlir.rvv_vp_intrinsic_add_scalable": 706,
-  "mlir.stripmining": 27825,
-  "rvv_bench.ascii_to_utf16": 678943,
-  "rvv_bench.ascii_to_utf32": 226997,
-  "rvv_bench.byteswap": 422035,
-  "rvv_bench.chacha20": 39957,
-  "rvv_bench.memcpy": 677231,
-  "rvv_bench.memset": 295653,
-  "rvv_bench.mergelines": 580218,
-  "rvv_bench.poly1305": 39957,
-  "rvv_bench.strlen": 235252,
-  "rvv_bench.utf8_count": 2346912
+  "intrinsic.conv2d_less_m2": 2592,
+  "mlir.hello": 174,
+  "mlir.rvv_vp_intrinsic_add": 513,
+  "mlir.rvv_vp_intrinsic_add_scalable": 757,
+  "mlir.stripmining": 27874,
+  "rvv_bench.ascii_to_utf16": 692008,
+  "rvv_bench.ascii_to_utf32": 228630,
+  "rvv_bench.byteswap": 422042,
+  "rvv_bench.chacha20": 40010,
+  "rvv_bench.memcpy": 594872,
+  "rvv_bench.memset": 295446,
+  "rvv_bench.mergelines": 579448,
+  "rvv_bench.poly1305": 40010,
+  "rvv_bench.strlen": 235255,
+  "rvv_bench.utf8_count": 2294504
 }
diff --git a/difftest/dpi_t1rocketemu/src/bus.rs b/difftest/dpi_t1rocketemu/src/bus.rs
@@ -36,7 +36,7 @@ impl ShadowBus {
   pub fn new() -> Self {
     const DDR_SIZE: usize = 0x80000000;
     const SCALAR_SIZE: usize = 0x20000000;
-    const SRAM_SIZE: usize = 0x00400000;
+    const SRAM_SIZE: usize = 0x32000000;
 
     Self {
       devices: [

diff --git a/difftest/dpi_t1rocketemu/src/dpi.rs b/difftest/dpi_t1rocketemu/src/dpi.rs
@@ -245,6 +245,9 @@ unsafe extern "C" fn axi_write_loadStoreAXI(
         info!("driver is ready to quit");
         driver.success = true;
         driver.quit = true;
+      } else {
+        // hack
+        driver.record_mmio_hack(exit_data);
       }
     }
   });

diff --git a/difftest/dpi_t1rocketemu/src/drive.rs b/difftest/dpi_t1rocketemu/src/drive.rs
@@ -11,6 +11,8 @@ use elf::{
   ElfStream,
 };
 use std::collections::HashMap;
+use std::fs::File;
+use std::io::Write as _;
 use std::os::unix::fs::FileExt;
 use std::{fs, path::Path};
 use tracing::{debug, error, info, trace};
@@ -37,6 +39,8 @@ pub(crate) struct Driver {
 
   shadow_bus: ShadowBus,
 
+  mmio_hack_file: Option<File>,
+
   pub(crate) quit: bool,
   pub(crate) success: bool,
 }
@@ -47,6 +51,8 @@ impl Driver {
     let (e_entry, shadow_bus, _fn_sym_tab) =
       Self::load_elf(&args.elf_file).expect("fail creating simulator");
 
+    let mmio_hack_file = File::create("mmio_hack.jsonl").unwrap();
+
     Self {
       scope,
 
@@ -58,6 +64,8 @@ impl Driver {
 
       shadow_bus,
 
+      mmio_hack_file: Some(mmio_hack_file),
+
       quit: false,
       success: false,
     }
@@ -130,6 +138,13 @@ impl Driver {
     Ok((elf.ehdr.e_entry, mem, fn_sym_tab))
   }
 
+  pub fn record_mmio_hack(&mut self, value: u32) {
+    let time = get_t();
+    let mmio_hack = self.mmio_hack_file.as_mut().unwrap();
+    mmio_hack.write(format!("{{\"cycle\": {time}, \"value\": {value}}}\n").as_bytes()).unwrap();
+    mmio_hack.flush().unwrap();
+  }
+
   pub fn update_commit_cycle(&mut self) {
     self.last_commit_cycle = get_t();
   }

diff --git a/nix/overlay.nix b/nix/overlay.nix
@@ -3,7 +3,7 @@
 final: prev:
 
 let
-  llvmForRVV_attrName = "llvmPackages_17"; # brand new clang with v0.12 rvv intrinsic support
+  llvmForRVV_attrName = "llvmPackages_18"; # brand new clang with v0.12 rvv intrinsic support
   rv32_pkgs = final.pkgsCross.riscv32-embedded;
   rv32_buildPkgs = rv32_pkgs.buildPackages;
 in

diff --git a/nix/t1/run/run-vcs-emu.nix b/nix/t1/run/run-vcs-emu.nix
@@ -81,6 +81,9 @@ stdenvNoCC.mkDerivation (finalAttr: {
       mv perf.json $out/
     fi
 
+    # TODO: remove the hack
+    cp -v mmio_hack.jsonl "$out"
+
     cp -v cm.log "$out"
     cp -vr cm.vdb "$out"
 

diff --git a/tests/rvv_bench/t1_runtime.patch b/tests/rvv_bench/t1_runtime.patch
@@ -2,7 +2,7 @@ diff --git a/bench/bench.h b/bench/bench.h
 index af1c839..661e8df 100644
 --- a/bench/bench.h
 +++ b/bench/bench.h
-@@ -120,45 +120,11 @@ static void
+@@ -120,45 +120,13 @@ static void
  bench_run(Bench *benches, size_t nBenches)
  {
  	for (Bench *b = benches; b != benches + nBenches; ++b) {
@@ -43,7 +43,9 @@ index af1c839..661e8df 100644
 -				print(f,bench_time(n, *i, *b))(",")(flush,);
 -			}
 -			print("],\n")(flush,);
++           *(int*)0x40000000 = 1;
 +			bench_time(N, *i, *b);
++           *(int*)0x40000000 = 2;
  		}
 -		print("]\n},\n");
  	}

diff --git a/tests/t1.ld b/tests/t1.ld
@@ -5,7 +5,7 @@ MEMORY {
   SCALAR (RWX) : ORIGIN = 0x20000000, LENGTH = 512M /* put first to set it as default */
   MMIO   (RW)  : ORIGIN = 0x00000000, LENGTH = 512M
   DDR    (RW)  : ORIGIN = 0x40000000, LENGTH = 2048M
-  SRAM   (RW)  : ORIGIN = 0xc0000000, LENGTH = 4M /* TODO: read from config */
+  SRAM   (RW)  : ORIGIN = 0xc0000000, LENGTH = 800M /* TODO: read from config */
 }
 
 SECTIONS {

diff --git a/tests/t1_main.S b/tests/t1_main.S
@@ -4,6 +4,38 @@ _start:
     csrs mstatus, a0
     csrwi vcsr, 0
 
+    li x1, 0
+    li x2, 0
+    li x3, 0
+    li x4, 0
+    li x5, 0
+    li x6, 0
+    li x7, 0
+    li x8, 0
+    li x9, 0
+    li x10, 0
+    li x11, 0
+    li x12, 0
+    li x13, 0
+    li x14, 0
+    li x15, 0
+    li x16, 0
+    li x17, 0
+    li x18, 0
+    li x19, 0
+    li x20, 0
+    li x21, 0
+    li x22, 0
+    li x23, 0
+    li x24, 0
+    li x25, 0
+    li x26, 0
+    li x27, 0
+    li x28, 0
+    li x29, 0
+    li x30, 0
+    li x31, 0
+
     la sp, __stacktop
 
     // no ra to save