Merge pull request #18580 from hrydgard/sleepy-describe-vertexdec

DescribePtr: Individually describe vertex decoders

Core/HW/MediaEngine.cpp

@@ -29,6 +29,18 @@
 
 #include <algorithm>
 
+#ifdef _M_SSE
+#include <emmintrin.h>
+#endif
+
+#if PPSSPP_ARCH(ARM_NEON)
+#if defined(_MSC_VER) && PPSSPP_ARCH(ARM64)
+#include <arm64_neon.h>
+#else
+#include <arm_neon.h>
+#endif
+#endif
+
 #ifdef USE_FFMPEG
 
 extern "C" {
@@ -744,9 +756,34 @@ inline void writeVideoLineRGBA(void *destp, const void *srcp, int width) {
 	u32_le *dest = (u32_le *)destp;
 	const u32_le *src = (u32_le *)srcp;
 
-	const u32 mask = 0x00FFFFFF;
-	for (int i = 0; i < width; ++i) {
-		dest[i] = src[i] & mask;
+	int count = width;
+
+#if PPSSPP_ARCH(SSE2)
+	__m128i mask = _mm_set1_epi32(0x00FFFFFF);
+	while (count >= 8) {
+		__m128i pixels1 = _mm_and_si128(_mm_loadu_si128((const __m128i *)src), mask);
+		__m128i pixels2 = _mm_and_si128(_mm_loadu_si128((const __m128i *)src + 1), mask);
+		_mm_storeu_si128((__m128i *)dest, pixels1);
+		_mm_storeu_si128((__m128i *)dest + 1, pixels2);
+		src += 8;
+		dest += 8;
+		count -= 8;
+	}
+#elif PPSSPP_ARCH(ARM_NEON)
+	int32x4_t mask = vdupq_n_u32(0x00FFFFFF);
+	while (count >= 8) {
+		int32x4_t pixels1 = vandq_u32(vld1q_u32(src), mask);
+		int32x4_t pixels2 = vandq_u32(vld1q_u32(src + 4), mask);
+		vst1q_u32(dest, pixels1);
+		vst1q_u32(dest + 4, pixels2);
+		src += 8;
+		dest += 8;
+		count -= 8;
+	}
+#endif
+	const u32 mask32 = 0x00FFFFFF;
+	while (count--) {
+		*dest++ = *src++ & mask32;
 	}
 }
 

GPU/Common/DrawEngineCommon.cpp

@@ -1036,3 +1036,33 @@ void TessellationDataTransfer::CopyControlPoints(float *pos, float *tex, float *
 		}
 	}
 }
+
+bool DrawEngineCommon::DescribeCodePtr(const u8 *ptr, std::string &name) const {
+	if (!decJitCache_ || !decJitCache_->IsInSpace(ptr)) {
+		return false;
+	}
+
+	// Loop through all the decoders and see if we have a match.
+	VertexDecoder *found = nullptr;
+	u32 foundKey;
+
+	decoderMap_.Iterate([&](u32 key, VertexDecoder *value) {
+		if (!found) {
+			if (value->IsInSpace(ptr)) {
+				foundKey = key;
+				found = value;
+			}
+		}
+	});
+
+	if (found) {
+		char temp[256];
+		found->ToString(temp, false);
+		name = temp;
+		snprintf(temp, sizeof(temp), "_%08X", foundKey);
+		name += temp;
+		return true;
+	} else {
+		return false;
+	}
+}

GPU/Common/DrawEngineCommon.h

@@ -138,11 +138,7 @@ class DrawEngineCommon {
 		everUsedExactEqualDepth_ = v;
 	}
 
-	bool IsCodePtrVertexDecoder(const u8 *ptr) const {
-		if (decJitCache_)
-			return decJitCache_->IsInSpace(ptr);
-		return false;
-	}
+	bool DescribeCodePtr(const u8 *ptr, std::string &name) const;
 	int GetNumDrawCalls() const {
 		return numDrawVerts_;
 	}

GPU/Common/VertexDecoderArm.cpp

@@ -257,7 +257,7 @@ JittedVertexDecoder VertexDecoderJitCache::Compile(const VertexDecoder &dec, int
 			// Reset the code ptr and return zero to indicate that we failed.
 			ResetCodePtr(GetOffset(start));
 			char temp[1024] = {0};
-			dec.ToString(temp);
+			dec.ToString(temp, true);
 			INFO_LOG(G3D, "Could not compile vertex decoder: %s", temp);
 			return 0;
 		}
@@ -285,7 +285,7 @@ JittedVertexDecoder VertexDecoderJitCache::Compile(const VertexDecoder &dec, int
 	/*
 	DisassembleArm(start, GetCodePtr() - start);
 	char temp[1024] = {0};
-	dec.ToString(temp);
+	dec.ToString(temp, true);
 	INFO_LOG(G3D, "%s", temp);
 	*/
 

GPU/Common/VertexDecoderArm64.cpp

@@ -253,7 +253,7 @@ JittedVertexDecoder VertexDecoderJitCache::Compile(const VertexDecoder &dec, int
 			// Reset the code ptr (effectively undoing what we generated) and return zero to indicate that we failed.
 			ResetCodePtr(GetOffset(start));
 			char temp[1024] = {0};
-			dec.ToString(temp);
+			dec.ToString(temp, true);
 			ERROR_LOG(G3D, "Could not compile vertex decoder, failed at step %d: %s", i, temp);
 			return nullptr;
 		}
@@ -288,7 +288,7 @@ JittedVertexDecoder VertexDecoderJitCache::Compile(const VertexDecoder &dec, int
 
 	if (log) {
 		char temp[1024] = { 0 };
-		dec.ToString(temp);
+		dec.ToString(temp, true);
 		INFO_LOG(JIT, "=== %s (%d bytes) ===", temp, (int)(GetCodePtr() - start));
 		std::vector<std::string> lines = DisassembleArm64(start, (int)(GetCodePtr() - start));
 		for (auto line : lines) {

GPU/Common/VertexDecoderCommon.cpp

@@ -1276,7 +1276,7 @@ void VertexDecoder::SetVertexType(u32 fmt, const VertexDecoderOptions &options,
 
 	if (reportNoPos) {
 		char temp[256]{};
-		ToString(temp);
+		ToString(temp, true);
 		ERROR_LOG_REPORT(G3D, "Vertices without position found: (%08x) %s", fmt_, temp);
 	}
 
@@ -1402,7 +1402,7 @@ void VertexDecoder::CompareToJit(const u8 *startPtr, u8 *decodedptr, int count,
 		jittedReader.Goto(i);
 		if (!DecodedVertsAreSimilar(controlReader, jittedReader)) {
 			char name[512]{};
-			ToString(name);
+			ToString(name, true);
 			ERROR_LOG(G3D, "Encountered vertexjit mismatch at %d/%d for %s", i, count, name);
 			if (morphcount > 1) {
 				printf("Morph:\n");
@@ -1456,8 +1456,9 @@ static const char *idxnames[4] = { "-", "u8", "u16", "?" };
 static const char *weightnames[4] = { "-", "u8", "u16", "f" };
 static const char *colnames[8] = { "", "?", "?", "?", "565", "5551", "4444", "8888" };
 
-int VertexDecoder::ToString(char *output) const {
-	char * start = output;
+int VertexDecoder::ToString(char *output, bool spaces) const {
+	char *start = output;
+
 	output += sprintf(output, "P: %s ", posnames[pos]);
 	if (nrm)
 		output += sprintf(output, "N: %s ", nrmnames[nrm]);
@@ -1474,15 +1475,24 @@ int VertexDecoder::ToString(char *output) const {
 	if (throughmode)
 		output += sprintf(output, " (through)");
 
-	output += sprintf(output, " (size: %i)", VertexSize());
+	output += sprintf(output, " (%ib)", VertexSize());
+
+	if (!spaces) {
+		size_t len = strlen(start);
+		for (int i = 0; i < len; i++) {
+			if (start[i] == ' ')
+				start[i] = '_';
+		}
+	}
+
 	return output - start;
 }
 
 std::string VertexDecoder::GetString(DebugShaderStringType stringType) {
 	char buffer[256];
 	switch (stringType) {
 	case SHADER_STRING_SHORT_DESC:
-		ToString(buffer);
+		ToString(buffer, true);
 		return std::string(buffer);
 	case SHADER_STRING_SOURCE_CODE:
 		{

GPU/Common/VertexDecoderCommon.h

@@ -431,7 +431,11 @@ class VertexDecoder {
 	// output must be big for safety.
 	// Returns number of chars written.
 	// Ugly for speed.
-	int ToString(char *output) const;
+	int ToString(char *output, bool spaces) const;
+
+	bool IsInSpace(const uint8_t *ptr) const {
+		return ptr >= (const uint8_t *)jitted_ && ptr < ((const uint8_t *)jitted_ + jittedSize_);
+	}
 
 	// Mutable decoder state
 	mutable u8 *decoded_ = nullptr;
@@ -507,6 +511,7 @@ class VertexDecoderJitCache : public VERTEXDECODER_JIT_BACKEND {
 
 	// Returns a pointer to the code to run.
 	JittedVertexDecoder Compile(const VertexDecoder &dec, int32_t *jittedSize);
+	bool DescribeCodePtr(const u8 *ptr, std::string &name) const;
 	void Clear();
 
 	void Jit_WeightsU8();

GPU/Common/VertexDecoderRiscV.cpp

@@ -318,7 +318,7 @@ JittedVertexDecoder VertexDecoderJitCache::Compile(const VertexDecoder &dec, int
 			// Reset the code ptr (effectively undoing what we generated) and return zero to indicate that we failed.
 			ResetCodePtr(GetOffset(start));
 			char temp[1024]{};
-			dec.ToString(temp);
+			dec.ToString(temp, true);
 			ERROR_LOG(G3D, "Could not compile vertex decoder, failed at step %d: %s", i, temp);
 			return nullptr;
 		}
@@ -351,7 +351,7 @@ JittedVertexDecoder VertexDecoderJitCache::Compile(const VertexDecoder &dec, int
 
 	if (log) {
 		char temp[1024]{};
-		dec.ToString(temp);
+		dec.ToString(temp, true);
 		INFO_LOG(JIT, "=== %s (%d bytes) ===", temp, (int)(GetCodePtr() - start));
 		std::vector<std::string> lines = DisassembleRV64(start, (int)(GetCodePtr() - start));
 		for (auto line : lines) {

GPU/GPUCommon.cpp

@@ -1938,11 +1938,9 @@ bool GPUCommon::GetCurrentSimpleVertices(int count, std::vector<GPUDebugVertex>
 }
 
 bool GPUCommon::DescribeCodePtr(const u8 *ptr, std::string &name) {
-	if (drawEngineCommon_->IsCodePtrVertexDecoder(ptr)) {
-		name = "VertexDecoderJit";
-		return true;
-	}
-	return false;
+	// The only part of GPU emulation (other than software) that jits is the vertex decoder, currently,
+	// which is owned by the drawengine.
+	return drawEngineCommon_->DescribeCodePtr(ptr, name);
 }
 
 void GPUCommon::UpdateUVScaleOffset() {