Merge pull request #16 from lukel97/fractional-lmul

add support for fractional LMUL in instruction benchmarks
camel-cdr · Nov 2, 2024 · c9e4b6a · c9e4b6a
2 parents 7f6dcbf + 8ab1568
commit c9e4b6a
Show file tree

Hide file tree

Showing 3 changed files with 63 additions and 26 deletions.
diff --git a/instructions/rvv/config.h b/instructions/rvv/config.h
@@ -4,22 +4,22 @@
 #define RUNS 500
 
 // processor specific configs
-//               m8  m4  m2  m1
-//          SEW: 6310    6310
-//               4268... 4268...
-#define T_A    0b1111111111111111 // all
-#define T_W    0b0000011101110111 // widen
-#define T_WR   0b0111011101110111 // widen reduction
-#define T_N    0b0000011101110111 // narrow
-#define T_F    0b1100110011001100 // float
-#define T_FW   0b0000010001000100 // float widen
-#define T_FWR  0b0100010001000100 // float widen reduction
-#define T_FN   0b0000010001000100 // float narrow
+//               m8  m4  m2  m1  mf2 mf4 mf8
+//          SEW: 6310    6310    6310    6410
+//               4268... 4268... 4268... 4268...
+#define T_A    0b1111111111111111111111111111 // all
+#define T_W    0b0000011101110111011101110111 // widen
+#define T_WR   0b0111011101110111011101110111 // widen reduction
+#define T_N    0b0000011101110111011101110111 // narrow
+#define T_F    0b1110111011101110111011101110 // float
+#define T_FW   0b0000011001100110011001100110 // float widen
+#define T_FWR  0b0110011001100110011001100110 // float widen reduction
+#define T_FN   0b0000011001100110011001100110 // float narrow
 
-#define T_E2   0b1110111011101110 // extend 2
-#define T_E4   0b1100110011001100 // extend 4
-#define T_E8   0b1000100010001000 // extend 8
-#define T_ei16 0b1110111111111111 // no m8
+#define T_E2   0b1110111011101110111011101110 // extend 2
+#define T_E4   0b1100110011001100110011001100 // extend 4
+#define T_E8   0b1000100010001000100010001000 // extend 8
+#define T_ei16 0b1110111111111111111111111111 // no m8
 
 // special:
-#define T_m1 ((1 << 16) | T_A) // emul<=1
+#define T_m1 ((1 << 28) | T_A) // emul<=1
diff --git a/instructions/rvv/gen.S b/instructions/rvv/gen.S
@@ -107,8 +107,8 @@ define(`m_bench_wxi',
 
 define(`m_ifmx_t', `m_$3(shift(shift(shift($*))))')
 define(`m_ifmx_f', `m_$3($4, 0, m_nop, m_unimpl)')
-define(`m_ifmx',`ifelse(eval($1 <= m_LMUL && $2 >= m_LMUL),1,`m_ifmx_t($*)',`m_ifmx_f($*)')')
-
+define(`m_ifmx_lmul_to_i', `ifelse(regexp($1,f), -1, $1, 0)')
+define(`m_ifmx',`ifelse(eval($1 <= m_ifmx_lmul_to_i(m_LMUL) && $2 >= m_ifmx_lmul_to_i(m_LMUL)),1,`m_ifmx_t($*)',`m_ifmx_f($*)')')
 
 # calls $1 with: name type setup code:vararg
 define(`m_bench_all',`
@@ -344,32 +344,56 @@ define(`m_bench_all',`
 #define defptr .dword
 #endif
 
-define(`m_gen_function_pointers', `defptr $1_m`'m_LMUL')
+define(`m_gen_function_pointers', `defptr $1_m`'m_benchLMUL')
 define(`m_gen_types', `defptr $2')
 define(`m_format',`$2 shift(shift($*))')
 define(`m_gen_strings', `.string "m_format(,shift(shift(shift($*))))"')
 
+.balign 8
+define(`m_LMUL', f8)
+define(`m_benchLMUL', 1)
+.global bench_m`'m_LMUL
+	bench_m`'m_LMUL:
+	m_bench_all(`gen_function_pointers')
+
+.balign 8
+define(`m_LMUL', f4)
+define(`m_benchLMUL', 1)
+.global bench_m`'m_LMUL
+	bench_m`'m_LMUL:
+	m_bench_all(`gen_function_pointers')
+
+.balign 8
+define(`m_LMUL', f2)
+define(`m_benchLMUL', 1)
+.global bench_m`'m_LMUL
+	bench_m`'m_LMUL:
+	m_bench_all(`gen_function_pointers')
 
 .balign 8
 define(`m_LMUL', 1)
+define(`m_benchLMUL', 1)
 .global bench_m`'m_LMUL
 	bench_m`'m_LMUL:
 	m_bench_all(`gen_function_pointers')
 
 .balign 8
 define(`m_LMUL', 2)
+define(`m_benchLMUL', 2)
 .global bench_m`'m_LMUL
 	bench_m`'m_LMUL:
 	m_bench_all(`gen_function_pointers')
 
 .balign 8
 define(`m_LMUL', 4)
+define(`m_benchLMUL', 4)
 .global bench_m`'m_LMUL
 	bench_m`'m_LMUL:
 	m_bench_all(`gen_function_pointers')
 
 .balign 8
 define(`m_LMUL', 8)
+define(`m_benchLMUL', 8)
 .global bench_m`'m_LMUL
 	bench_m`'m_LMUL:
 	m_bench_all(`gen_function_pointers')

diff --git a/instructions/rvv/main.c b/instructions/rvv/main.c
@@ -7,8 +7,8 @@ typedef ux (*BenchFunc)(void);
 extern size_t bench_count;
 extern char bench_names;
 extern ux bench_types;
-extern BenchFunc bench_m1, bench_m2, bench_m4, bench_m8;
-static BenchFunc *benches[] = { &bench_m1, &bench_m2, &bench_m4, &bench_m8 };
+extern BenchFunc bench_mf8, bench_mf4, bench_mf2, bench_m1, bench_m2, bench_m4, bench_m8;
+static BenchFunc *benches[] = { &bench_mf8, &bench_mf4, &bench_mf2, &bench_m1, &bench_m2, &bench_m4, &bench_m8 };
 
 extern ux run_bench(ux (*bench)(void), ux type, ux vl, ux seed);
 
@@ -29,22 +29,35 @@ run_all_types(char const *name, ux bIdx, ux vl, int ta, int ma)
 	print("<tr><td>")(s,name)("</td>");
 	ux mask = bIdx[&bench_types];
 
+	ux lmuls[] = { 5, 6, 7, 0, 1, 2, 3 };
+
 	for (ux sew = 0; sew < 4; ++sew)
-	for (ux lmul = 0; lmul < 4; ++lmul) {
+	for (ux lmul_idx = 0; lmul_idx < 7; ++lmul_idx) {
+		ux lmul = lmuls[lmul_idx];
 		ux vtype = lmul | (sew<<3) | (!!ta << 6) | (!!ma << 7);
 
-		if (!(mask >> (lmul*4 + sew) & 1)) {
+		if (!(mask >> (lmul_idx*4 + sew) & 1)) {
+			print("<td></td>");
+			continue;
+		}
+
+		ux lmul_val = 1 << lmul_idx; // fixed-point, denum 8
+		ux sew_val = 1 << (sew + 3);
+		// > For a given supported fractional LMUL setting,
+		// > implementations must support SEW settings between SEWMIN
+		// > and LMUL * ELEN, inclusive.
+		if (sew_val * 8 > lmul_val * __riscv_v_elen) {
 			print("<td></td>");
 			continue;
 		}
 
-		ux emul = lmul;
+		ux emul = lmul_idx;
 		if (mask == T_W || mask == T_FW || mask == T_N || mask == T_FN)
 			emul += 1;
 		if (mask == T_ei16 && sew == 0)
-			emul = emul < 3 ? emul+1 : 3;
+			emul = emul < 7 ? emul+1 : 7;
 		if (mask == T_m1)
-			emul = 1;
+			emul = 4; // m2
 		BenchFunc bench = benches[emul][bIdx];
 
 		for (ux i = 0; i < RUNS; ++i) {