Signed-off-by: Max Chou <[email protected]>
---
tests/tcg/riscv64/system/meson.build | 12 +
tests/tcg/riscv64/test-zvdota.S | 343 +++++++++++++++++++++++++++
2 files changed, 355 insertions(+)
create mode 100644 tests/tcg/riscv64/test-zvdota.S
diff --git a/tests/tcg/riscv64/system/meson.build
b/tests/tcg/riscv64/system/meson.build
index 3c8b018d9e9..01370badb0d 100644
--- a/tests/tcg/riscv64/system/meson.build
+++ b/tests/tcg/riscv64/system/meson.build
@@ -89,6 +89,18 @@ tests += {
}
}
+dot_cpu = ('rv64,v=true,vlen=512' +
+ ',x-zvqwdota8i=true,x-zvqwdota16i=true' +
+ ',x-zvqwbdota8i=true,x-zvqwbdota16i=true' +
+ ',x-zvfwdota16bf=true,x-zvfwbdota16bf=true' +
+ ',x-zvfqwdota8f=true,x-zvfqwbdota8f=true,x-zvfbdota32f=true')
+tests += {
+ 'test-zvdota.S': {
+ 'cflags': cflags,
+ 'qemu_args': ['-cpu', dot_cpu, qemu_args],
+ }
+}
+
if 'qemu-system-riscv64' in emulators
tcg_tests += {
'riscv64-softmmu': {
diff --git a/tests/tcg/riscv64/test-zvdota.S b/tests/tcg/riscv64/test-zvdota.S
new file mode 100644
index 00000000000..eefb615088f
--- /dev/null
+++ b/tests/tcg/riscv64/test-zvdota.S
@@ -0,0 +1,343 @@
+/*
+ * Test the Zvdota/Zvbdota extensions
+ *
+ * Each dot has four input pairs. Integer inputs distinguish all four
+ * signedness combinations; the -7 accumulator also exercises wraparound.
+ * Floating inputs represent A = [1, -2, 0.5, 4] and column n of B =
+ * [n + 1, -2, 3, 0.5]. With accumulator 1.5, C[n] = n + 10 exactly.
+ * Separate BF16 cases check round-to-odd in the final accumulator addition.
+ *
+ * Poisoned source tails must not contribute. Destination prestart/tail
+ * elements and masked-off batched outputs must remain undisturbed (tu/mu).
+ * The test requires VLEN 512, which also exercises ci = 8.
+ *
+ * Exit status is zero on success or the numbered check on failure, including
+ * unexpected traps.
+ *
+ * SPDX-License-Identifier: GPL-2.0-or-later
+ */
+
+ .option arch, +v
+ .option norvc
+ .option norelax
+
+/* ci is an element index; its encoding is the index divided by eight. */
+#define DOT(f6, f3, vm, ci) \
+ (((f6) << 26) | ((vm) << 25) | ((8 | ((ci) / 8)) << 20) | \
+ (1 << 15) | ((f3) << 12) | (16 << 7) | 0x77)
+
+ .text
+
+/* Load four elements per source, leaving all remaining elements poisoned. */
+.macro LOAD_INPUT sew, a, b
+ vsetvli t0, zero, e\sew, m1, tu, mu
+ vmv.v.i v1, -1
+ .irp reg, 8, 9, 10, 11, 12, 13, 14, 15
+ vmv.v.i v\reg, -1
+ .endr
+ li t0, 4
+ vsetvli t0, t0, e\sew, m1, tu, mu
+ la t1, \a
+ vle\sew\().v v1, (t1)
+ la t1, \b
+ .irp reg, 8, 9, 10, 11, 12, 13, 14, 15
+ vle\sew\().v v\reg, (t1)
+ addi t1, t1, \sew / 2
+ .endr
+.endm
+
+/*
+ * Initialize an m4 destination at its output width, execute with the input
+ * vtype, and compare the whole m4 group (including untouched registers).
+ * v0 contains repeating 0xaa55: scalar dots mask inputs, batched dots outputs.
+ * Nonzero-ci cases give prestart elements a different accumulator seed.
+ */
+.macro RUN id, f6, f3, type, eew, exp, batch=0, vm=1, vl=4, ci=0
+ li t6, \id
+ vsetvli s0, zero, e\eew, m4, tu, mu
+ .if \f3
+ li s1, 0x3fc00000 /* FP32 accumulator = 1.5. */
+ .else
+ li s1, -7
+ .endif
+ vmv.v.x v16, s1
+ .if \ci
+ li t0, \ci
+ vsetvli t0, t0, e\eew, m4, tu, mu
+ addi t0, s1, 1
+ vmv.v.x v16, t0
+ .endif
+ li t1, \vl
+ li t2, \type
+ vsetvl t0, t1, t2
+ bne t0, t1, fail
+ .word DOT(\f6, \f3, \vm, \ci)
+ csrr t0, vstart
+ bnez t0, fail
+ csrr t0, fflags
+ bnez t0, fail
+ vsetvli t0, zero, e\eew, m4, tu, mu
+ la a0, result
+ vse\eew\().v v16, (a0)
+ la a1, \exp
+ mv a2, s0
+ li a3, \ci
+ .if \vl == 0
+ li a4, 0
+ .elseif \batch
+ li a4, 8
+ .else
+ li a4, 1
+ .endif
+ li a5, \eew / 8
+ mv a6, s1
+ .if \batch && !\vm
+ li a7, (0xaa55 >> \ci) & 0xff
+ .else
+ li a7, 0xff
+ .endif
+ call check_result
+.endm
+
+/* The trap handler verifies the cause, faulting PC and instruction bits. */
+.macro ILLEGAL id, f6, f3, type, start=0
+ li t6, \id
+ li t1, 4
+ li t2, \type
+ vsetvl t0, t1, t2
+ li t0, \start
+ csrw vstart, t0
+ la s2, 1f
+ li s3, DOT(\f6, \f3, 1, 0)
+1: .word DOT(\f6, \f3, 1, 0)
+ bnez s2, fail /* Must have taken the expected trap. */
+.endm
+
+ .global _start
+_start:
+ li t6, 44 /* Setup failure. */
+ li s2, 0 /* No trap expected outside ILLEGAL. */
+ la t0, trap_handler
+ csrw mtvec, t0
+ li t0, (1 << 9) | (1 << 13)
+ csrs mstatus, t0
+ csrw fcsr, zero
+ csrr t0, vlenb
+ li t1, 64 /* VLEN 512 only. */
+ bne t0, t1, fail
+ vsetvli t0, zero, e16, m1, tu, mu
+ li t0, 0xaa55
+ vmv.v.x v0, t0
+
+ /* Integer scalar dots: vs2 signedness in opcode, vs1 in altfmt. */
+ LOAD_INPUT 8, a_i8, b_i8
+ RUN 1, 0x26, 0, 0x000, 32, exp_i8_uu
+ RUN 2, 0x27, 0, 0x000, 32, exp_i8_us
+ RUN 3, 0x26, 0, 0x100, 32, exp_i8_su
+ RUN 4, 0x27, 0, 0x100, 32, exp_i8_ss
+ RUN 5, 0x26, 0, 0x000, 32, exp_i8_mask, vm=0
+ RUN 6, 0x26, 0, 0x000, 32, exp_i8_uu, vl=0
+ LOAD_INPUT 16, a_i16, b_i16
+ RUN 7, 0x26, 0, 0x008, 64, exp_i16_uu
+ RUN 8, 0x27, 0, 0x008, 64, exp_i16_us
+ RUN 9, 0x26, 0, 0x108, 64, exp_i16_su
+ RUN 10, 0x27, 0, 0x108, 64, exp_i16_ss
+
+ /* Batched dots use all eight columns, even though vl is only four. */
+ LOAD_INPUT 8, a_i8, b_i8
+ RUN 11, 0x2e, 0, 0x000, 32, exp_i8_uu, batch=1
+ RUN 12, 0x2f, 0, 0x000, 32, exp_i8_us, batch=1
+ RUN 13, 0x2e, 0, 0x100, 32, exp_i8_su, batch=1
+ RUN 14, 0x2f, 0, 0x100, 32, exp_i8_ss, batch=1
+ RUN 15, 0x2e, 0, 0x000, 32, exp_i8_uu, batch=1, vm=0
+ RUN 16, 0x2e, 0, 0x000, 32, exp_i8_uu, batch=1, vl=0
+ LOAD_INPUT 16, a_i16, b_i16
+ RUN 17, 0x2e, 0, 0x008, 64, exp_i16_uu, batch=1
+ RUN 18, 0x2f, 0, 0x008, 64, exp_i16_us, batch=1
+ RUN 19, 0x2e, 0, 0x108, 64, exp_i16_su, batch=1
+ RUN 20, 0x2f, 0, 0x108, 64, exp_i16_ss, batch=1
+
+ /* BF16 requires SEW=16, altfmt=1. */
+ LOAD_INPUT 16, a_bf16, b_bf16
+ RUN 21, 0x24, 1, 0x108, 32, exp_fp
+ RUN 22, 0x24, 1, 0x108, 32, exp_fp_mask, vm=0
+ RUN 23, 0x2c, 1, 0x108, 32, exp_fp, batch=1
+ RUN 24, 0x2c, 1, 0x108, 32, exp_fp, batch=1, vm=0
+
+ /* OFP8: ordinary/alt opcode selects vs2, altfmt selects vs1. */
+ LOAD_INPUT 8, a_e4m3, b_e4m3
+ RUN 25, 0x26, 1, 0x000, 32, exp_fp
+ RUN 29, 0x26, 1, 0x000, 32, exp_fp_mask, vm=0
+ RUN 30, 0x2e, 1, 0x000, 32, exp_fp, batch=1
+ RUN 34, 0x2e, 1, 0x000, 32, exp_fp, batch=1, vm=0
+ LOAD_INPUT 8, a_e4m3, b_e5m2
+ RUN 26, 0x27, 1, 0x000, 32, exp_fp
+ RUN 31, 0x2f, 1, 0x000, 32, exp_fp, batch=1
+ LOAD_INPUT 8, a_e5m2, b_e4m3
+ RUN 27, 0x26, 1, 0x100, 32, exp_fp
+ RUN 32, 0x2e, 1, 0x100, 32, exp_fp, batch=1
+ LOAD_INPUT 8, a_e5m2, b_e5m2
+ RUN 28, 0x27, 1, 0x100, 32, exp_fp
+ RUN 33, 0x2f, 1, 0x100, 32, exp_fp, batch=1
+
+ LOAD_INPUT 32, a_fp32, b_fp32
+ RUN 35, 0x2b, 1, 0x010, 32, exp_fp, batch=1
+ RUN 36, 0x2b, 1, 0x010, 32, exp_fp, batch=1, vm=0
+
+ /* 1.5 + 2^-24 rounds to odd (0x3fc00001), without setting NX. */
+ LOAD_INPUT 16, a_tiny, b_tiny
+ RUN 37, 0x24, 1, 0x108, 32, exp_odd
+ RUN 38, 0x2c, 1, 0x108, 32, exp_odd, batch=1
+
+ LOAD_INPUT 8, a_i8, b_i8
+ RUN 39, 0x2e, 0, 0x000, 32, exp_i8_uu, batch=1, ci=8
+ RUN 40, 0x2e, 0, 0x000, 32, exp_i8_uu, batch=1, vm=0, ci=8
+ ILLEGAL 41, 0x26, 0, 0x000, start=1 /* Nonzero vstart. */
+ ILLEGAL 42, 0x2e, 0, 0x001 /* Batched LMUL=2. */
+ ILLEGAL 43, 0x24, 1, 0x008 /* BF16 altfmt=0. */
+ li a0, 0
+ j _exit
+
+ .balign 4
+trap_handler:
+ beqz s2, fail
+ csrr t0, mcause
+ li t1, 2 /* Illegal instruction. */
+ bne t0, t1, fail
+ csrr t0, mtval
+ bne t0, s3, fail
+ csrr t0, mepc
+ bne t0, s2, fail
+ addi t0, t0, 4
+ csrw mepc, t0
+ csrw vstart, zero
+ li s2, 0
+ mret
+
+/*
+ * a0=result, a1=expected body, a2=total elements, a3=first body element,
+ * a4=body count, a5=element bytes, a6=seed, a7=active body mask.
+ * Prestart elements use seed + 1; masked-off body and tail elements use seed.
+ */
+check_result:
+ li t0, 0
+1: mv t3, a6
+ bltu t0, a3, 6f
+ sub t1, t0, a3
+ bgeu t1, a4, 3f
+ srl t2, a7, t1
+ andi t2, t2, 1
+ beqz t2, 3f
+ mul t2, t1, a5
+ add t2, a1, t2
+ li t3, 8
+ beq a5, t3, 2f
+ lw t3, 0(t2)
+ j 3f
+2: ld t3, 0(t2)
+3: li t2, 8
+ beq a5, t2, 4f
+ lw t2, 0(a0)
+ j 5f
+4: ld t2, 0(a0)
+5: bne t2, t3, fail
+ add a0, a0, a5
+ addi t0, t0, 1
+ bltu t0, a2, 1b
+ ret
+6: addi t3, a6, 1
+ j 3b
+
+fail:
+ mv a0, t6
+_exit:
+ la a1, semiargs
+ li t0, 0x20026
+ sd t0, 0(a1)
+ sd a0, 8(a1)
+ li a0, 0x20
+ .balign 16
+ slli zero, zero, 0x1f
+ ebreak
+ srai zero, zero, 0x7
+ j .
+
+ .data
+ .balign 8
+semiargs:
+ .space 16
+result:
+ .space 256 /* m4 at VLEN 512. */
+
+a_i8:
+ .byte 0xfe, 0x82, 2, 3
+b_i8:
+ .irp n, 0, 1, 2, 3, 4, 5, 6, 7
+ .byte 0xff - \n, 0x80 + \n, 3, 7
+ .endr
+ .balign 2
+a_i16:
+ .half 0xfffe, 0x8002, 2, 3
+b_i16:
+ .irp n, 0, 1, 2, 3, 4, 5, 6, 7
+ .half 0xffff - \n, 0x8000 + \n, 3, 7
+ .endr
+
+a_bf16:
+ .half 0x3f80, 0xc000, 0x3f00, 0x4080
+b_bf16:
+ .irp bits,0x3f80,0x4000,0x4040,0x4080,0x40a0,0x40c0,0x40e0,0x4100
+ .half \bits, 0xc000, 0x4040, 0x3f00
+ .endr
+a_e4m3:
+ .byte 0x38, 0xc0, 0x30, 0x48
+b_e4m3:
+ .irp bits, 0x38, 0x40, 0x44, 0x48, 0x4a, 0x4c, 0x4e, 0x50
+ .byte \bits, 0xc0, 0x44, 0x30
+ .endr
+a_e5m2:
+ .byte 0x3c, 0xc0, 0x38, 0x44
+b_e5m2:
+ .irp bits, 0x3c, 0x40, 0x42, 0x44, 0x45, 0x46, 0x47, 0x48
+ .byte \bits, 0xc0, 0x42, 0x38
+ .endr
+ .balign 4
+a_fp32:
+ .float 1.0, -2.0, 0.5, 4.0
+b_fp32:
+ .irp n, 1, 2, 3, 4, 5, 6, 7, 8
+ .float \n, -2.0, 3.0, 0.5
+ .endr
+a_tiny:
+ .half 0x3f80, 0, 0, 0
+b_tiny:
+ .rept 8
+ .half 0x3380, 0, 0, 0
+ .endr
+
+/* Suffixes describe vs1 then vs2; -7 + A dot B, modulo output width. */
+.macro EXPECT_INT name, width, a0, a1, b0, b1
+\name:
+ .irp n, 0, 1, 2, 3, 4, 5, 6, 7
+ .\width -7 + (\a0) * ((\b0) - \n) + (\a1) * ((\b1) + \n) + 2 * 3 + 3 * 7
+ .endr
+.endm
+ .balign 8
+ EXPECT_INT exp_i8_uu, word, 254, 130, 255, 128
+ EXPECT_INT exp_i8_us, word, 254, 130, -1, -128
+ EXPECT_INT exp_i8_su, word, -2, -126, 255, 128
+ EXPECT_INT exp_i8_ss, word, -2, -126, -1, -128
+ EXPECT_INT exp_i16_uu, dword, 65534, 32770, 65535, 32768
+ EXPECT_INT exp_i16_us, dword, 65534, 32770, -1, -32768
+ EXPECT_INT exp_i16_su, dword, -2, -32766, 65535, 32768
+ EXPECT_INT exp_i16_ss, dword, -2, -32766, -1, -32768
+exp_i8_mask:
+ .word -7 + 254 * 255 + 2 * 3
+exp_fp:
+ .float 10.0, 11.0, 12.0, 13.0, 14.0, 15.0, 16.0, 17.0
+exp_fp_mask:
+ .float 4.0
+exp_odd:
+ .rept 8
+ .word 0x3fc00001
+ .endr