From: Brendan Hansen <brendan.f.hansen@gmail.com>
Date: Thu, 10 Sep 2020 18:50:46 +0000 (-0500)
Subject: added simd example; simd bugfixes
X-Git-Url: https://git.brendanfh.com/?a=commitdiff_plain;h=0a696419010cc5aa4089569d46ac1fc98329969a;p=onyx.git

added simd example; simd bugfixes
---

diff --git a/onyx b/onyx
index efe0a3be..699262fd 100755
Binary files a/onyx and b/onyx differ
diff --git a/progs/particle_sym.onyx b/progs/particle_sym.onyx
new file mode 100644
index 00000000..c694db38
--- /dev/null
+++ b/progs/particle_sym.onyx
@@ -0,0 +1,75 @@
+package main
+
+#include_file "core/std/wasi"
+#include_file "core/simd_intrinsics"
+
+use package core
+use package simd
+
+main :: proc (args: [] cstring) {
+    init_positions();
+    init_velocities();
+
+    print("Beginning simulation.\n");
+
+    for i: 0 .. 20 {
+        update();
+        print(cast(i64) (avg_motion() * 100000000000000000.0f));
+        print("\n");
+    }
+}
+
+OBJECT_COUNT :: 10000
+
+positions  : [OBJECT_COUNT] f32x4
+velocities : [OBJECT_COUNT] f32x4
+
+init_positions :: proc () {
+    for ^p: positions {
+        *p = f32x4_replace_lane(*p, 0, random_float(-127.0f, 127.0f));
+        *p = f32x4_replace_lane(*p, 1, random_float(-127.0f, 127.0f));
+        *p = f32x4_replace_lane(*p, 2, random_float(-127.0f, 127.0f));
+        *p = f32x4_replace_lane(*p, 3, random_float(-127.0f, 127.0f));
+    }
+}
+
+init_velocities :: proc () {
+    for ^v: velocities do *v = f32x4_const(0.0f, 0.0f, 0.0f, 0.0f);
+}
+
+horizontal_add :: proc (f: f32x4) -> f32 {
+    return f32x4_extract_lane(f, 0)
+         + f32x4_extract_lane(f, 1)
+         + f32x4_extract_lane(f, 2)
+         + f32x4_extract_lane(f, 3);
+}
+
+distance :: proc (x: f32x4, y: f32x4) -> f32 {
+    d := f32x4_sub(x, y);
+    return horizontal_add(f32x4_mul(d, d));
+}
+
+update :: proc () {
+    for i: 0 .. OBJECT_COUNT {
+        for j: 0 .. OBJECT_COUNT {
+            dist := distance(positions[i], positions[j]);
+            if dist == 0.0f do continue;
+
+            velocities[i] = f32x4_add(velocities[i], f32x4_div(f32x4_sub(positions[j], positions[i]), f32x4_splat(dist)));
+        }
+    }
+
+    for i: 0 .. OBJECT_COUNT {
+        positions[i] = f32x4_add(positions[i], velocities[i]);
+    }
+}
+
+avg_motion :: proc () -> f32 {
+    avg_motion := f32x4_splat(0.0f);
+    obj_vec := f32x4_splat(cast(f32) OBJECT_COUNT);
+    for v: velocities {
+        avg_motion = f32x4_add(avg_motion, f32x4_div(v, obj_vec));
+    }
+
+    return distance(f32x4_splat(0.0f), avg_motion);
+}
diff --git a/src/onyxwasm.c b/src/onyxwasm.c
index b9312e9e..f5286ec6 100644
--- a/src/onyxwasm.c
+++ b/src/onyxwasm.c
@@ -1387,7 +1387,7 @@ EMIT_FUNC(call, AstCall* call) {
         WIP(WI_V128_CONST, byte_buffer); \
     }
 
-#define SIMD_LANE_INSTR(instr, arg) \
+#define SIMD_EXTRACT_LANE_INSTR(instr, arg) \
     emit_expression(mod, &code, arg->value);\
     arg = (AstArgument *) arg->next; \
     if (arg->value->kind != Ast_Kind_NumLit) { \
@@ -1397,6 +1397,20 @@ EMIT_FUNC(call, AstCall* call) {
     } \
     WID(instr, (u8) ((AstNumLit *) arg->value)->value.i);
 
+#define SIMD_REPLACE_LANE_INSTR(instr, arg) { \
+        emit_expression(mod, &code, arg->value);\
+        arg = (AstArgument *) arg->next; \
+        if (arg->value->kind != Ast_Kind_NumLit) { \
+            onyx_report_error(arg->token->pos, "SIMD lane instructions expect a compile time lane number."); \
+            *pcode = code; \
+            return; \
+        } \
+        u8 lane = (u8) ((AstNumLit *) arg->value)->value.i; \
+        arg = (AstArgument *) arg->next; \
+        emit_expression(mod, &code, arg->value); \
+        WID(instr, lane); \
+    }
+
 
 EMIT_FUNC(intrinsic_call, AstIntrinsicCall* call) {
     bh_arr(WasmInstruction) code = *pcode;
@@ -1544,20 +1558,20 @@ EMIT_FUNC(intrinsic_call, AstIntrinsicCall* call) {
             break;
         }
 
-        case ONYX_INTRINSIC_I8X16_EXTRACT_LANE_S: SIMD_LANE_INSTR(WI_I8X16_EXTRACT_LANE_S, call->arguments); break;
-        case ONYX_INTRINSIC_I8X16_EXTRACT_LANE_U: SIMD_LANE_INSTR(WI_I8X16_EXTRACT_LANE_U, call->arguments); break;
-        case ONYX_INTRINSIC_I8X16_REPLACE_LANE:   SIMD_LANE_INSTR(WI_I8X16_REPLACE_LANE, call->arguments); break;
-        case ONYX_INTRINSIC_I16X8_EXTRACT_LANE_S: SIMD_LANE_INSTR(WI_I16X8_EXTRACT_LANE_S, call->arguments); break;
-        case ONYX_INTRINSIC_I16X8_EXTRACT_LANE_U: SIMD_LANE_INSTR(WI_I16X8_EXTRACT_LANE_U, call->arguments); break;
-        case ONYX_INTRINSIC_I16X8_REPLACE_LANE:   SIMD_LANE_INSTR(WI_I16X8_REPLACE_LANE, call->arguments); break;
-        case ONYX_INTRINSIC_I32X4_EXTRACT_LANE:   SIMD_LANE_INSTR(WI_I32X4_EXTRACT_LANE, call->arguments); break;
-        case ONYX_INTRINSIC_I32X4_REPLACE_LANE:   SIMD_LANE_INSTR(WI_I32X4_REPLACE_LANE, call->arguments); break;
-        case ONYX_INTRINSIC_I64X2_EXTRACT_LANE:   SIMD_LANE_INSTR(WI_I64X2_EXTRACT_LANE, call->arguments); break;
-        case ONYX_INTRINSIC_I64X2_REPLACE_LANE:   SIMD_LANE_INSTR(WI_I64X2_REPLACE_LANE, call->arguments); break;
-        case ONYX_INTRINSIC_F32X4_EXTRACT_LANE:   SIMD_LANE_INSTR(WI_F32X4_EXTRACT_LANE, call->arguments); break;
-        case ONYX_INTRINSIC_F32X4_REPLACE_LANE:   SIMD_LANE_INSTR(WI_F32X4_REPLACE_LANE, call->arguments); break;
-        case ONYX_INTRINSIC_F64X2_EXTRACT_LANE:   SIMD_LANE_INSTR(WI_F64X2_EXTRACT_LANE, call->arguments); break;
-        case ONYX_INTRINSIC_F64X2_REPLACE_LANE:   SIMD_LANE_INSTR(WI_F64X2_REPLACE_LANE, call->arguments); break;
+        case ONYX_INTRINSIC_I8X16_EXTRACT_LANE_S: SIMD_EXTRACT_LANE_INSTR(WI_I8X16_EXTRACT_LANE_S, call->arguments); break;
+        case ONYX_INTRINSIC_I8X16_EXTRACT_LANE_U: SIMD_EXTRACT_LANE_INSTR(WI_I8X16_EXTRACT_LANE_U, call->arguments); break;
+        case ONYX_INTRINSIC_I8X16_REPLACE_LANE:   SIMD_REPLACE_LANE_INSTR(WI_I8X16_REPLACE_LANE, call->arguments); break;
+        case ONYX_INTRINSIC_I16X8_EXTRACT_LANE_S: SIMD_EXTRACT_LANE_INSTR(WI_I16X8_EXTRACT_LANE_S, call->arguments); break;
+        case ONYX_INTRINSIC_I16X8_EXTRACT_LANE_U: SIMD_EXTRACT_LANE_INSTR(WI_I16X8_EXTRACT_LANE_U, call->arguments); break;
+        case ONYX_INTRINSIC_I16X8_REPLACE_LANE:   SIMD_REPLACE_LANE_INSTR(WI_I16X8_REPLACE_LANE, call->arguments); break;
+        case ONYX_INTRINSIC_I32X4_EXTRACT_LANE:   SIMD_EXTRACT_LANE_INSTR(WI_I32X4_EXTRACT_LANE, call->arguments); break;
+        case ONYX_INTRINSIC_I32X4_REPLACE_LANE:   SIMD_REPLACE_LANE_INSTR(WI_I32X4_REPLACE_LANE, call->arguments); break;
+        case ONYX_INTRINSIC_I64X2_EXTRACT_LANE:   SIMD_EXTRACT_LANE_INSTR(WI_I64X2_EXTRACT_LANE, call->arguments); break;
+        case ONYX_INTRINSIC_I64X2_REPLACE_LANE:   SIMD_REPLACE_LANE_INSTR(WI_I64X2_REPLACE_LANE, call->arguments); break;
+        case ONYX_INTRINSIC_F32X4_EXTRACT_LANE:   SIMD_EXTRACT_LANE_INSTR(WI_F32X4_EXTRACT_LANE, call->arguments); break;
+        case ONYX_INTRINSIC_F32X4_REPLACE_LANE:   SIMD_REPLACE_LANE_INSTR(WI_F32X4_REPLACE_LANE, call->arguments); break;
+        case ONYX_INTRINSIC_F64X2_EXTRACT_LANE:   SIMD_EXTRACT_LANE_INSTR(WI_F64X2_EXTRACT_LANE, call->arguments); break;
+        case ONYX_INTRINSIC_F64X2_REPLACE_LANE:   SIMD_REPLACE_LANE_INSTR(WI_F64X2_REPLACE_LANE, call->arguments); break;
 
         case ONYX_INTRINSIC_I8X16_SWIZZLE: WI(WI_I8X16_SWIZZLE); break;
         case ONYX_INTRINSIC_I8X16_SPLAT:   WI(WI_I8X16_SPLAT); break;