From db9b359b8b2dd0cb47bfef7319d214c7c326774e Mon Sep 17 00:00:00 2001
From: UnknownShadow200 <unknownshadow200@gmail.com>
Date: Fri, 2 May 2025 07:16:29 +1000
Subject: [PATCH] Optimise vertex upload

---
 misc/n64/gpu.c                | 41 +++++++++++++------------
 misc/n64/rsp_gpu.S            | 56 +++++++++++++++--------------------
 misc/n64/rsp_gpu_clipping.inc |  6 ++++
 3 files changed, 50 insertions(+), 53 deletions(-)

diff --git a/misc/n64/gpu.c b/misc/n64/gpu.c
index b3d4399af..92a7fde13 100644
--- a/misc/n64/gpu.c
+++ b/misc/n64/gpu.c
@@ -34,7 +34,7 @@ enum {
     GPU_CMD_SET_LONG         = 0x3,
 
     GPU_CMD_DRAW_QUAD        = 0x4,
-    GPU_CMD_UPLOAD_VTX       = 0x5,
+    GPU_CMD_UPLOAD_QUAD      = 0x5,
     GPU_CMD_MATRIX_LOAD      = 0x6,
 
 	GPU_CMD_PUSH_RDP         = 0x7,
@@ -156,41 +156,40 @@ static inline void put_word(rspq_write_t* s, uint16_t v1, uint16_t v2)
 	rspq_write_arg(s, v2 | (v1 << 16));
 }
 
-static void upload_vertex(uint32_t index, uint8_t cache_index)
+static void upload_vertex(rspq_write_t* s, uint32_t index)
 {
-    rspq_write_t s = rspq_write_begin(gpup_id, GPU_CMD_UPLOAD_VTX, 5);
-    rspq_write_arg(&s, cache_index * PRIM_VTX_SIZE);
 	char* ptr = gpu_pointer + index * gpu_stride;
 
 	float* vtx = (float*)(ptr + 0);
-	put_word(&s, vtx[0] * (1<<VTX_SHIFT),
-				 vtx[1] * (1<<VTX_SHIFT));
-	put_word(&s, vtx[2] * (1<<VTX_SHIFT),
-				 1.0f   * (1<<VTX_SHIFT));
+	put_word(s, vtx[0] * (1<<VTX_SHIFT),
+				vtx[1] * (1<<VTX_SHIFT));
+	put_word(s, vtx[2] * (1<<VTX_SHIFT),
+				1.0f   * (1<<VTX_SHIFT));
 
-	uint32_t* col = (uint32_t*)(ptr + 12); // TODO put_byte ?
-	rspq_write_arg(&s, *col);
+	uint32_t* col = (uint32_t*)(ptr + 12);
+	rspq_write_arg(s, *col);
 
 	if (gpu_texturing) {
 		float* tex = (float*)(ptr + 16);
-		put_word(&s, tex[0] * (1<<TEX_SHIFT),
-					 tex[1] * (1<<TEX_SHIFT));
+		put_word(s, tex[0] * (1<<TEX_SHIFT),
+					tex[1] * (1<<TEX_SHIFT));
 	} else {
-		put_word(&s, 0,
-					 0);
+		put_word(s, 0,
+					0);
     }
-    rspq_write_end(&s);
 }
 
 static void gpuDrawArrays(uint32_t first, uint32_t count)
 {
-    for (uint32_t i = 0; i < count; i++)
+    for (uint32_t i = 0; i < count; i += 4)
     {
-        uint8_t cache_index = i & 3;
-        upload_vertex(first + i, cache_index);
-
-		// Last vertex of quad?
-		if ((i & 3) != 3) continue;
+    	rspq_write_t s = rspq_write_begin(gpup_id, GPU_CMD_UPLOAD_QUAD, 17);
+    	rspq_write_arg(&s, 0);
+       	for (uint32_t j = 0; j < 4; j++)
+		{
+        	upload_vertex(&s, first + i + j);
+		}
+    	rspq_write_end(&s);
 
 		// We pass -1 because the triangle can be clipped and split into multiple
     	// triangles.
diff --git a/misc/n64/rsp_gpu.S b/misc/n64/rsp_gpu.S
index 9b3a57739..732dbbe0a 100644
--- a/misc/n64/rsp_gpu.S
+++ b/misc/n64/rsp_gpu.S
@@ -10,7 +10,7 @@
         RSPQ_DefineCommand GPUCmd_SetLong,       12  # 0x3
 
         RSPQ_DefineCommand GPUCmd_DrawQuad,      4   # 0x4
-        RSPQ_DefineCommand GPUCmd_UploadVertex,  20  # 0x5
+        RSPQ_DefineCommand GPUCmd_UploadQuad,    68  # 0x5
         RSPQ_DefineCommand GPUCmd_MatrixLoad,    68  # 0x6
 
         RSPQ_DefineCommand GPUCmd_PushRDP,       12  # 0x7
@@ -38,15 +38,9 @@ VERTEX_CACHE:   .dcb.b      PRIM_VTX_SIZE * VERTEX_CACHE_SIZE
     RSPQ_EndSavedState
 
     .align 4
-CACHE_OFFSETS:          .half 2,4,6,8,10,12,14,16,18
-
 CLIP_CODE_FACTORS:      .half 1, 1, GUARD_BAND_FACTOR, GUARD_BAND_FACTOR
 DRAW_TRI_RA:            .word 0
 
-#define CLIPPING_PLANE_COUNT  6
-#define CLIPPING_CACHE_SIZE   9
-#define CLIPPING_PLANE_SIZE   8
-
 #define SCREEN_VTX_CS_POSi          0     // X, Y, Z, W (all 32-bit)
 #define SCREEN_VTX_CS_POSf          8     // X, Y, Z, W (all 32-bit)
 #define SCREEN_VTX_X               16
@@ -131,23 +125,13 @@ GPUCmd_MatrixLoad:
 #undef dst
     .endfunc
 
-    ########################################
-    # GPUCmd_UploadVertex
-    #
-    # Arguments:
-    # * 0x00 (a0): offset within VERTEX_CACHE
-    # * 0x04 (a1): object space X, Y (16-bit)
-    # * 0x08 (a2): object space Z, W (16-bit)
-    # * 0x0C (a3): RGBA (8-bit each one)
-    # * 0x10:      S, T (16-bit)
-    #
-    ########################################
     .align 3
-    .func GPUCmd_UploadVertex
-GPUCmd_UploadVertex:
+    .func GPUCmd_UploadQuad
+GPUCmd_UploadQuad:
     #define vtx         a0
     #define mtx_ptr     s0
     #define src_ptr     s4
+	#define vcount      s3
 
     #define v___        $v01
 
@@ -171,17 +155,10 @@ GPUCmd_UploadVertex:
     #define z  e2
     #define w  e3
 
-    addi src_ptr, rspq_dmem_buf_ptr, %lo(RSPQ_DMEM_BUFFER) - 16
+    addi src_ptr, rspq_dmem_buf_ptr, %lo(RSPQ_DMEM_BUFFER) - 64
+    li vtx, %lo(VERTEX_CACHE)
+	li vcount, 4
 
-	ldv vpos,  0, src_ptr # Load X, Y, Z, W
-	llv vcol,  8, src_ptr # Load RGBA
-	llv vtex, 12, src_ptr # Load U, V
-
-    addi vtx, %lo(VERTEX_CACHE)
-    slv vcol, SCREEN_VTX_RGBA, vtx
-    slv vtex, SCREEN_VTX_S_T,  vtx
-
-# == matrix multiply ==
     li mtx_ptr, %lo(GL_MATRIX_MVP)
     ldv vmtx0_i.e0,  0x00,mtx_ptr
     ldv vmtx1_i.e0,  0x08,mtx_ptr
@@ -192,6 +169,12 @@ GPUCmd_UploadVertex:
     ldv vmtx2_f.e0,  0x30,mtx_ptr
     ldv vmtx3_f.e0,  0x38,mtx_ptr
 
+upload_vertex:
+	ldv vpos,  0, src_ptr # Load X, Y, Z, W
+	llv vcol,  8, src_ptr # Load RGBA
+	llv vtex, 12, src_ptr # Load U, V
+
+	# matrix multiply
     vmudn v___,      vmtx0_f, vpos.h0
     vmadh v___,      vmtx0_i, vpos.h0
     vmadn v___,      vmtx1_f, vpos.h1
@@ -201,12 +184,17 @@ GPUCmd_UploadVertex:
     vmadn v___,      vmtx3_f, vpos.h3
     vmadh vcspos_i,  vmtx3_i, vpos.h3
     vmadn vcspos_f,  vzero,   vzero
-# == end matrix multiply ==
+
+    slv vcol, SCREEN_VTX_RGBA, vtx
+    slv vtex, SCREEN_VTX_S_T,  vtx
 
     # 32-bit right shift by 5, to keep the clip space coordinates unscaled
     vmudm vcspos_i, vcspos_i, vshift8.e4
     vmadl vcspos_f, vcspos_f, vshift8.e4
 
+	addi vcount,  -1
+	addi src_ptr, 16
+
     sdv vcspos_i, SCREEN_VTX_CS_POSi,vtx
     sdv vcspos_f, SCREEN_VTX_CS_POSf,vtx
 
@@ -221,9 +209,13 @@ GPUCmd_UploadVertex:
     srl t1, t0, 5
     andi t0, 0x7
     or t0, t1
-    jr ra
     sb t0, PRIM_VTX_TRCODE(vtx)
 
+	bnez vcount, upload_vertex
+	addi vtx, PRIM_VTX_SIZE
+
+	jr ra
+	nop
     #undef src_ptr
     #undef vtx
 
diff --git a/misc/n64/rsp_gpu_clipping.inc b/misc/n64/rsp_gpu_clipping.inc
index 1fa9eb60f..f406650fb 100644
--- a/misc/n64/rsp_gpu_clipping.inc
+++ b/misc/n64/rsp_gpu_clipping.inc
@@ -1,3 +1,6 @@
+#define CLIPPING_PLANE_COUNT  6
+#define CLIPPING_CACHE_SIZE   9
+#define CLIPPING_PLANE_SIZE   8
 
     .section .data.gl_clipping
 
@@ -10,6 +13,9 @@ CLIP_PLANES:
     .half 0, 1, 0, -GUARD_BAND_FACTOR
     .half 0, 0, 1, -1
 
+    .align 4
+CACHE_OFFSETS: .half 2,4,6,8,10,12,14,16,18
+
     .section .bss.gl_clipping
 
 CLIP_CACHE: .dcb.b     SCREEN_VTX_SIZE * CLIPPING_CACHE_SIZE